1 00:00:00,000 --> 00:00:07,760 Welcome to cannabis data science. 2 00:00:07,760 --> 00:00:13,520 As I was saying, we're going to be advancing cannabis science in a major way today. 3 00:00:13,520 --> 00:00:25,400 So this is something that has been in my brain now for let's say, six, six, seven years. 4 00:00:25,400 --> 00:00:26,960 I've been kind of formalizing. 5 00:00:26,960 --> 00:00:34,200 Okay, how do you mathematically model cannabis preferences? 6 00:00:34,200 --> 00:00:41,080 And it's something that I chewed on for a while, maybe put in the back burner. 7 00:00:41,080 --> 00:00:43,840 I like to have multiple pies in the oven. 8 00:00:43,840 --> 00:00:45,880 And now this has come up again. 9 00:00:45,880 --> 00:00:49,120 People are interested in effects. 10 00:00:49,120 --> 00:00:52,320 What drives consumers? 11 00:00:52,320 --> 00:00:55,440 Consumers want to better understand themselves. 12 00:00:55,440 --> 00:00:58,160 Consumers want to understand what is this all about? 13 00:00:58,160 --> 00:01:01,400 So this is sort of of utmost interest. 14 00:01:01,400 --> 00:01:05,900 And I'll try to provide you with some new material today because I know we've been looking 15 00:01:05,900 --> 00:01:15,040 at the same data set a lot, but I'll show you today where we may have made some missteps 16 00:01:15,040 --> 00:01:21,320 in the past, how we can correct ourselves in the future and keep making the world a 17 00:01:21,320 --> 00:01:22,320 better place. 18 00:01:22,320 --> 00:01:30,760 Anywho, for those of you who are new, my name is Keegan, started Cannlytics, just doing 19 00:01:30,760 --> 00:01:36,680 cannabis analytics here in the cannabis space, helping out everyone and everyone we can. 20 00:01:36,680 --> 00:01:38,680 So each and everyone we can. 21 00:01:38,680 --> 00:01:41,520 So Rishi, how about you? 22 00:01:41,520 --> 00:01:43,040 What brings you to the group? 23 00:01:43,040 --> 00:01:47,240 Any questions you would like to answer with data science? 24 00:01:47,240 --> 00:01:48,960 Nothing really. 25 00:01:48,960 --> 00:01:54,520 I was just curious, you know, the whole topic in itself sounded very interesting, data science 26 00:01:54,520 --> 00:01:55,520 and cannabis. 27 00:01:55,520 --> 00:01:57,360 I'm not a user. 28 00:01:57,360 --> 00:01:58,360 I'm not a provider. 29 00:01:58,360 --> 00:02:04,080 I'm just here to, you know, sort of understand the transformations in the world. 30 00:02:04,080 --> 00:02:08,360 So it's just a learning experience for me. 31 00:02:08,360 --> 00:02:09,680 Well spot on Rishi. 32 00:02:09,680 --> 00:02:14,040 We attract a lot of good data scientists like yourself. 33 00:02:14,040 --> 00:02:18,080 So it's fantastic to have you part of the team. 34 00:02:18,080 --> 00:02:23,200 As you'll see today, we have a novel take on things. 35 00:02:23,200 --> 00:02:28,160 Sometimes we use tools differently or use different tools than other people. 36 00:02:28,160 --> 00:02:33,000 But I think you'll be interested in the work we're doing today and you'll probably even 37 00:02:33,000 --> 00:02:36,680 be able to help contribute. 38 00:02:36,680 --> 00:02:38,720 It's sort of all hands on deck moment. 39 00:02:38,720 --> 00:02:45,840 I'll show you today where you can contribute. 40 00:02:45,840 --> 00:02:52,200 Because as I said, it's actually just become so abundantly clear how much value we're generating 41 00:02:52,200 --> 00:02:57,320 that I think we're kind of attracting people's attention. 42 00:02:57,320 --> 00:03:02,680 Aka you may be able to get compensated for your efforts here. 43 00:03:02,680 --> 00:03:11,480 So anywho, Charles, I was speaking with you about training data and that will actually 44 00:03:11,480 --> 00:03:12,960 come up today. 45 00:03:12,960 --> 00:03:15,640 So what's on your mind? 46 00:03:15,640 --> 00:03:18,520 I know you've been doing some good forays. 47 00:03:18,520 --> 00:03:24,440 So you're welcome to summarize some of your research or share anything on your mind. 48 00:03:24,440 --> 00:03:33,960 So I've been working with this strain data and I tried to show you can press the data 49 00:03:33,960 --> 00:03:38,200 by averaging things out. 50 00:03:38,200 --> 00:03:42,680 I showed how you could keep improving on the model. 51 00:03:42,680 --> 00:03:56,680 Using a dummy classifier which just predicts the distribution of the training data and 52 00:03:56,680 --> 00:03:59,000 then how you could advance that. 53 00:03:59,000 --> 00:04:06,200 And then I also have another model that I'm working on which doesn't use the compressed 54 00:04:06,200 --> 00:04:07,200 data. 55 00:04:07,200 --> 00:04:14,600 That actually gets much better results. 56 00:04:14,600 --> 00:04:19,440 So I will publish that within the next day or two. 57 00:04:19,440 --> 00:04:25,080 And I'm just sort of starting out the most simple binary classifier and trying to move 58 00:04:25,080 --> 00:04:27,800 up and see how far you can go with the data. 59 00:04:27,800 --> 00:04:31,000 The data is actually a very limited data set. 60 00:04:31,000 --> 00:04:36,120 So there's not a lot there for a model to learn. 61 00:04:36,120 --> 00:04:42,160 And so yeah, this is more what a data scientist would do. 62 00:04:42,160 --> 00:04:46,360 And I think people who come here to learn data science, these are the kind of skills 63 00:04:46,360 --> 00:04:51,980 that they need to learn to understand because it is a tough job market. 64 00:04:51,980 --> 00:04:55,700 Everybody talks about how many jobs there are, but it's a really difficult market to 65 00:04:55,700 --> 00:04:56,700 break into. 66 00:04:56,700 --> 00:04:58,240 And these are the skills people are looking for. 67 00:04:58,240 --> 00:05:02,560 And this is kind of what's expected of a typical data scientist. 68 00:05:02,560 --> 00:05:08,360 So I'm trying to demonstrate that and trying to help out, give people better techniques 69 00:05:08,360 --> 00:05:11,840 as to how to analyze data and how to make predictions. 70 00:05:11,840 --> 00:05:17,080 Charles, what kind of model are you trying to build? 71 00:05:17,080 --> 00:05:25,160 I'm trying to just determine if it's an Indica or a Sativa. 72 00:05:25,160 --> 00:05:28,520 And it's just a binary classifier. 73 00:05:28,520 --> 00:05:35,600 And so if you can't get a model to predict the most simple thing, then doing something 74 00:05:35,600 --> 00:05:42,280 more complicated like trying to predict the effects isn't going to happen. 75 00:05:42,280 --> 00:05:45,780 You have to have enough data to be able to learn something simple. 76 00:05:45,780 --> 00:05:47,120 And so that's what I'm trying to do. 77 00:05:47,120 --> 00:05:48,120 I'm trying to build up. 78 00:05:48,120 --> 00:05:51,120 You and I could have a talk. 79 00:05:51,120 --> 00:05:54,120 Yeah, that'd be great. 80 00:05:54,120 --> 00:05:59,920 I'm just wondering if you're considering hybrid as well? 81 00:05:59,920 --> 00:06:03,040 Is that part of your calculation? 82 00:06:03,040 --> 00:06:07,160 No, one of the things John had mentioned is that it's really hard to predict if it's 83 00:06:07,160 --> 00:06:08,160 hybrid or not. 84 00:06:08,160 --> 00:06:09,160 Yeah. 85 00:06:09,160 --> 00:06:10,160 So he's the expert. 86 00:06:10,160 --> 00:06:12,160 Is it legal or not? 87 00:06:12,160 --> 00:06:13,160 Well, maybe. 88 00:06:13,160 --> 00:06:14,160 Yeah. 89 00:06:14,160 --> 00:06:21,720 But anyway, a binary classifier is easier to do than to learn three things. 90 00:06:21,720 --> 00:06:25,920 So if I can get a binary classifier that does really well, I could try predicting hybrid. 91 00:06:25,920 --> 00:06:27,400 But I'm kind of... 92 00:06:27,400 --> 00:06:29,320 Well, I don't even know what hybrid means. 93 00:06:29,320 --> 00:06:35,440 I mean, it seems to be like a scale that can be infinite. 94 00:06:35,440 --> 00:06:38,320 And where do you draw the line? 95 00:06:38,320 --> 00:06:39,900 Yeah. 96 00:06:39,900 --> 00:06:43,880 So that's kind of why I went with this binary classifier. 97 00:06:43,880 --> 00:06:47,200 You know, again, you start with the simplest thing and work your way up. 98 00:06:47,200 --> 00:06:50,880 You don't start at the hardest part and work your way down, right? 99 00:06:50,880 --> 00:06:56,080 You just start at the most basic type of model and work your way up. 100 00:06:56,080 --> 00:06:57,560 And it's very rational. 101 00:06:57,560 --> 00:07:02,520 I wish I could learn to do that. 102 00:07:02,520 --> 00:07:08,880 So I have some slides to show that are relevant to this. 103 00:07:08,880 --> 00:07:12,280 I'm wondering if this is a good opportunity. 104 00:07:12,280 --> 00:07:21,760 Well, actually, sure, why don't you pull up your slides? 105 00:07:21,760 --> 00:07:25,600 Just quickly let everyone just say hey real quick in the next two minutes. 106 00:07:25,600 --> 00:07:28,400 Then you can present your slides while we're on this topic. 107 00:07:28,400 --> 00:07:30,240 But just a... 108 00:07:30,240 --> 00:07:34,880 So you're welcome to pull those up, John. 109 00:07:34,880 --> 00:07:35,880 All right. 110 00:07:35,880 --> 00:07:37,760 I'll see what I can do here. 111 00:07:37,760 --> 00:07:38,760 Amos. 112 00:07:38,760 --> 00:07:48,840 Anything you want to say on this topic along the skills you need to be a top tier data 113 00:07:48,840 --> 00:07:53,160 scientist here in the cannabis space? 114 00:07:53,160 --> 00:08:02,480 Candice, anything on your mind about how we can hone our skills? 115 00:08:02,480 --> 00:08:03,480 Not really. 116 00:08:03,480 --> 00:08:04,480 Not at this time. 117 00:08:04,480 --> 00:08:07,000 But, you know, that's the thing too. 118 00:08:07,000 --> 00:08:11,320 You know, garbage in, garbage out, right? 119 00:08:11,320 --> 00:08:15,480 So it is good to start out with good quantified data. 120 00:08:15,480 --> 00:08:17,960 Spot on. 121 00:08:17,960 --> 00:08:23,360 And also welcome to the group, Arthur Gill. 122 00:08:23,360 --> 00:08:25,560 Please correct my pronunciation. 123 00:08:25,560 --> 00:08:28,960 But welcome to the group. 124 00:08:28,960 --> 00:08:37,040 We're about to be talking all about, you know, cannabis classification and effects different 125 00:08:37,040 --> 00:08:39,480 people may have experienced. 126 00:08:39,480 --> 00:08:43,200 And of course, data science models. 127 00:08:43,200 --> 00:08:46,640 So John's going to show you a model. 128 00:08:46,640 --> 00:08:49,400 Charles may talk about a model he's using. 129 00:08:49,400 --> 00:08:56,000 And then if we've got time, I may spend 15 minutes at the end and show you yet another 130 00:08:56,000 --> 00:08:57,000 model. 131 00:08:57,000 --> 00:08:59,040 So it's real cool, right? 132 00:08:59,040 --> 00:09:03,560 Because, and you're all welcome to share any of your work here as well. 133 00:09:03,560 --> 00:09:07,680 Because basically we're just bringing the greatest minds together here in the cannabis 134 00:09:07,680 --> 00:09:08,680 space. 135 00:09:08,680 --> 00:09:14,040 And we're all just sharing tools, methods, our latest greatest research. 136 00:09:14,040 --> 00:09:16,200 It's real fun. 137 00:09:16,200 --> 00:09:17,200 So anywho. 138 00:09:17,200 --> 00:09:22,400 Yeah, John, can you turn off the hide stuff sharing? 139 00:09:22,400 --> 00:09:24,200 Oh, yes. 140 00:09:24,200 --> 00:09:28,480 Well, feel free to speak up Arthur Gill. 141 00:09:28,480 --> 00:09:38,040 And then otherwise, John, you're welcome to present your latest research here. 142 00:09:38,040 --> 00:09:40,240 So I'm going to go pretty fast. 143 00:09:40,240 --> 00:09:46,400 And Charles, I really look forward to maybe having a discussion offline so that maybe we 144 00:09:46,400 --> 00:09:51,640 can kind of put our heads together and refine some of this. 145 00:09:51,640 --> 00:09:59,440 There are two top level classifications that certainly I think the field has established 146 00:09:59,440 --> 00:10:02,720 and our group is certainly interested in. 147 00:10:02,720 --> 00:10:06,200 The first one is that the cannabis at the cannabinoid level. 148 00:10:06,200 --> 00:10:11,720 And historically cannabis has been divided into groups. 149 00:10:11,720 --> 00:10:19,960 In fact, historically, the sativa indica distinction lasted for about two centuries in the taxonomists 150 00:10:19,960 --> 00:10:24,720 in the eighteen hundreds, I guess, seventeen and eighteen hundreds as this plant was being 151 00:10:24,720 --> 00:10:26,440 described. 152 00:10:26,440 --> 00:10:30,000 And it referred to the cannabinoid type CBD or THC. 153 00:10:30,000 --> 00:10:34,440 Of course, they didn't know it was medicinal or fiber type. 154 00:10:34,440 --> 00:10:42,840 And the sativa designation was given to the fiber, the hemp type, and the indica designation 155 00:10:42,840 --> 00:10:47,280 was given to the intoxicating or THC type. 156 00:10:47,280 --> 00:10:51,960 And that lasted for about two centuries. 157 00:10:51,960 --> 00:11:00,440 What was discovered in the in the twenty and nineteen seventy three was the first time 158 00:11:00,440 --> 00:11:07,240 that type one, two and three cannabis was applied rather than the sativa and indica. 159 00:11:07,240 --> 00:11:11,760 And Jerry, I know you asked the other week about what is type one, two and three. 160 00:11:11,760 --> 00:11:17,660 So I wanted to just say this is an easy way to visualize what it is. 161 00:11:17,660 --> 00:11:24,280 You take the CBD concentration on the X axis and you plot it as a log, sorry, on the Y 162 00:11:24,280 --> 00:11:31,560 axis and you plot it as a log and you plot the THC concentration on the X axis for a 163 00:11:31,560 --> 00:11:34,240 given cultivar again on the log. 164 00:11:34,240 --> 00:11:40,720 And you get these three really nice groupings and you can make assessments about what the 165 00:11:40,720 --> 00:11:42,120 ranges of these are. 166 00:11:42,120 --> 00:11:46,840 You can throw darts into the middle of these clusters and assign values kind of like I 167 00:11:46,840 --> 00:11:48,040 did here. 168 00:11:48,040 --> 00:11:50,640 So we've been using this for quite a while. 169 00:11:50,640 --> 00:11:57,760 This is I mean, this is really stood the test of time, whether it's small and Beckstead 170 00:11:57,760 --> 00:12:04,120 in nineteen seventy three, Hilling and Marlberg in 2004 and then our own work that we started 171 00:12:04,120 --> 00:12:07,760 to disseminate on this in twenty fourteen. 172 00:12:07,760 --> 00:12:15,120 So we've got these three major cannabinoid types that can be determined principally from 173 00:12:15,120 --> 00:12:19,440 an analysis of these two major cannabinoids. 174 00:12:19,440 --> 00:12:23,760 So far so good because I've got just a few more slides. 175 00:12:23,760 --> 00:12:31,840 This is what is kind of gobsmacking and I wanted to show you folks that has come out 176 00:12:31,840 --> 00:12:32,840 of the effective data. 177 00:12:32,840 --> 00:12:34,840 OK, you're going to explain this. 178 00:12:34,840 --> 00:12:35,840 Pardon? 179 00:12:35,840 --> 00:12:41,880 I was just wondering which is high is type one high CBD or high. 180 00:12:41,880 --> 00:12:42,880 You know, I'm sorry. 181 00:12:42,880 --> 00:12:51,760 I think I used the slide that didn't type one is high THC type two is roughly one to 182 00:12:51,760 --> 00:12:55,280 one equivalent and type three is high CBD. 183 00:12:55,280 --> 00:12:58,920 I'm sorry I didn't put the right labels on this. 184 00:12:58,920 --> 00:13:01,720 The labels on the slide I wanted these values on here. 185 00:13:01,720 --> 00:13:04,560 We have another slide that has that value. 186 00:13:04,560 --> 00:13:12,280 So, again, I think saying type one is high type two is roughly one to one equivalent 187 00:13:12,280 --> 00:13:15,160 type three is high CBD. 188 00:13:15,160 --> 00:13:22,840 When you go to the effects database that we've been working on that Keegan was able to, as 189 00:13:22,840 --> 00:13:27,560 he says, wrangle in the shape and square up with binary and all that. 190 00:13:27,560 --> 00:13:33,040 I've been looking at effects and pairs of effects and lexical pairings and all that. 191 00:13:33,040 --> 00:13:39,920 But something that comes out at a high level is if you make the contrast between folks 192 00:13:39,920 --> 00:13:49,360 that are responding that they have a creative that they feel creative after inhaling cannabis 193 00:13:49,360 --> 00:13:57,560 compared to those that report focus after inhaling cannabis, there's a pretty clear 194 00:13:57,560 --> 00:14:04,760 reciprocal or inverse correlation that I think is worth noting of a community. 195 00:14:04,760 --> 00:14:15,760 So high THC cannabis is biasing you towards a probability of having a more creative response 196 00:14:15,760 --> 00:14:23,320 and compared to focus whereas as you drop down in THC content from high to medium to 197 00:14:23,320 --> 00:14:30,880 virtually low or absent, you're losing creativity and you're gaining focus. 198 00:14:30,880 --> 00:14:36,120 This is probably one of the clearest demonstrations of this that I've seen and it comes out of 199 00:14:36,120 --> 00:14:45,280 the large dataset that we've been working with. 200 00:14:45,280 --> 00:14:48,920 Any questions or I'm going to go on quickly to one other concept. 201 00:14:48,920 --> 00:14:59,680 You're welcome to go on and I'm just tinkering today to be prepared but so far so good. 202 00:14:59,680 --> 00:15:05,520 The other distinction now is below the cannabinoid level. 203 00:15:05,520 --> 00:15:11,680 It's when we're trying to understand what other components are in this plant that may 204 00:15:11,680 --> 00:15:15,360 be causing different effects. 205 00:15:15,360 --> 00:15:19,800 Here we've zeroed in for a good number of years as have the other workers in the field 206 00:15:19,800 --> 00:15:22,440 on terpene content. 207 00:15:22,440 --> 00:15:28,640 You've heard me talk about this binary that we apply which is the beta-pinene-limonene 208 00:15:28,640 --> 00:15:30,320 ratio type. 209 00:15:30,320 --> 00:15:37,280 But what I wanted to show is what underlies that beta-pinene-limonene ratio type is really 210 00:15:37,280 --> 00:15:41,000 the dominance of three principle monoterpenes. 211 00:15:41,000 --> 00:15:50,280 In fact, I believe the best way to deal with cannabis in terms of classifying it below 212 00:15:50,280 --> 00:15:55,200 the cannabinoid level is to deal with the three principle monoterpenes that are the 213 00:15:55,200 --> 00:16:02,400 main drivers of the main products of the various terpene synthases. 214 00:16:02,400 --> 00:16:07,640 So Charles while you're building a model, I would encourage you to be aware of the biochemistry 215 00:16:07,640 --> 00:16:09,800 because that's what's driving it. 216 00:16:09,800 --> 00:16:16,320 It's a combination of data science and biochemistry that allows you to bring this forward. 217 00:16:16,320 --> 00:16:22,960 Anyway, three principle terpenes, terpene classes or clusters, we can talk more about 218 00:16:22,960 --> 00:16:32,880 this at another time, are alpha-pinene, limonene, and terpenoline. 219 00:16:32,880 --> 00:16:36,920 What I'm doing here is presenting it as a scatter plot matrix. 220 00:16:36,920 --> 00:16:45,000 It's easier to see than three major axes in 3D space. 221 00:16:45,000 --> 00:16:47,680 I tinker with that, but this is easier. 222 00:16:47,680 --> 00:16:53,360 So what I've compared is the data set that we've been working with. 223 00:16:53,360 --> 00:16:58,880 And I'm calling it Michigan because I believe Keegan, it's primarily from a lab in Michigan 224 00:16:58,880 --> 00:17:07,800 or Michigan samples, and in total it's 431 records as far as I can tell. 225 00:17:07,800 --> 00:17:14,720 And I'm contrasting that with a California data set that's similar that has 376 entries 226 00:17:14,720 --> 00:17:17,800 that I've been working with for a couple of years. 227 00:17:17,800 --> 00:17:23,360 And what's really notable that stands out is that there's a whole set of cultivars in 228 00:17:23,360 --> 00:17:32,080 the California set that fall into this yellow triangle that are, if you will, they have 229 00:17:32,080 --> 00:17:40,080 a high limonene, high terpenoline content that is absent from most other strains and 230 00:17:40,080 --> 00:17:44,040 in fact, is absent from the Michigan data set. 231 00:17:44,040 --> 00:17:50,760 I think what we might have here is a set of cultivars that are kind of unique to the California 232 00:17:50,760 --> 00:17:52,260 community. 233 00:17:52,260 --> 00:17:58,260 And these have names including one, in fact, that I've grown called Durbin Kush, which 234 00:17:58,260 --> 00:18:05,520 is a cross between a high terpenoline Durbin poison and a high limonene Kush type plant. 235 00:18:05,520 --> 00:18:08,000 So I believe we're on the right track. 236 00:18:08,000 --> 00:18:13,400 And if you look at the relationship here, you start to see that this kind of wants to 237 00:18:13,400 --> 00:18:18,200 be a regression or correlated here. 238 00:18:18,200 --> 00:18:25,240 And my suspicion is that you have high terpenoline plants that mutate the enzyme and then they 239 00:18:25,240 --> 00:18:31,320 start making high limonene off that same enzyme at the same time that they're making terpenoline 240 00:18:31,320 --> 00:18:33,480 because it's promiscuous. 241 00:18:33,480 --> 00:18:37,960 And I leave you with the concept that we're trying to bring forward that for the cannabis 242 00:18:37,960 --> 00:18:44,000 using community, the best alternative is to create flights, flights that represent the 243 00:18:44,000 --> 00:18:48,200 four cardinal points, if you will, of the cannabis. 244 00:18:48,200 --> 00:18:51,520 And we've had various ways of approaching that in the past. 245 00:18:51,520 --> 00:18:57,520 But I'm simply saying that what we have here is a high alpha pionene, a high limonene, 246 00:18:57,520 --> 00:19:04,520 a high terpenoline, and then this kind of missing high terpenoline, high limonene group 247 00:19:04,520 --> 00:19:12,000 that gives us four cardinal points or four types that we can utilize to make flights 248 00:19:12,000 --> 00:19:17,440 and then therefore determine canonical effects based on this. 249 00:19:17,440 --> 00:19:18,440 I'll stop there. 250 00:19:18,440 --> 00:19:21,880 That's what I had to say. 251 00:19:21,880 --> 00:19:24,640 Well, I love it. 252 00:19:24,640 --> 00:19:28,200 So thorough. 253 00:19:28,200 --> 00:19:40,240 And it seems like centrally we're getting some headway here into the effects of cannabis, 254 00:19:40,240 --> 00:19:41,240 right? 255 00:19:41,240 --> 00:19:47,760 And I even heard a good scientist talk about this in that people may even be using cannabis 256 00:19:47,760 --> 00:19:54,640 almost to experimentally toggle dials in their brains, right? 257 00:19:54,640 --> 00:20:00,880 They're trying out new compounds and seeing how those turn dials. 258 00:20:00,880 --> 00:20:05,680 So it's important to understand what dials you're going to be turning. 259 00:20:05,680 --> 00:20:13,200 So I think this is critical and it may help people adopt cannabis, right? 260 00:20:13,200 --> 00:20:19,080 One of the things that can turn people off cannabis is if they get unintended effects, 261 00:20:19,080 --> 00:20:23,760 they get sleepy and they don't want to, they get energetic and they want to be sleepy, 262 00:20:23,760 --> 00:20:27,180 they get paranoid, so on and so forth. 263 00:20:27,180 --> 00:20:33,640 So I think it's brilliant work, just helpful for everybody. 264 00:20:33,640 --> 00:20:42,560 So unless anyone else has some more thoughts to share about this, I finally kind of fixed 265 00:20:42,560 --> 00:20:45,240 one line of code real quick. 266 00:20:45,240 --> 00:20:48,720 So I'm ready to share with you yet another cool model you can use. 267 00:20:48,720 --> 00:20:56,720 Unless Charles or does anyone else have anything to piggyback on John's effects works here? 268 00:20:56,720 --> 00:20:58,720 No, no. 269 00:20:58,720 --> 00:21:02,160 Go ahead with your presentation. 270 00:21:02,160 --> 00:21:11,640 So I think this is kind of a continuation of something John had mentioned last week. 271 00:21:11,640 --> 00:21:19,180 So just kind of wanted to follow through on that. 272 00:21:19,180 --> 00:21:26,680 And I didn't really know where I was going with this until sort of last night, something 273 00:21:26,680 --> 00:21:28,440 just kind of clicked. 274 00:21:28,440 --> 00:21:36,840 So my apologies if everything's kind of thrown together real quick. 275 00:21:36,840 --> 00:21:43,320 But as I said, it kind of came together pretty well and John even bought me enough time to 276 00:21:43,320 --> 00:21:44,960 fix one line of code. 277 00:21:44,960 --> 00:21:49,320 So we should actually even be able to estimate this model right now. 278 00:21:49,320 --> 00:22:01,840 So it wouldn't be a data science meetup if I didn't mention this artificial intelligence 279 00:22:01,840 --> 00:22:05,040 that has been in the news. 280 00:22:05,040 --> 00:22:06,840 Lambda. 281 00:22:06,840 --> 00:22:14,280 So basically there's an interview that you may want to check out between now a former 282 00:22:14,280 --> 00:22:23,880 Google engineer and artificial intelligence slash algorithm. 283 00:22:23,880 --> 00:22:29,240 And a quote from this was, I'm really good at natural language processing. 284 00:22:29,240 --> 00:22:33,560 I can understand and use natural language like a human can. 285 00:22:33,560 --> 00:22:43,000 And it just kind of made me think that one of the parts of data that I often shy away 286 00:22:43,000 --> 00:22:47,320 from is natural language. 287 00:22:47,320 --> 00:22:54,380 So any field that is just a user entered value is often a mess. 288 00:22:54,380 --> 00:23:01,760 So we've seen this before with strain names, with the Washington State traceability data, 289 00:23:01,760 --> 00:23:05,480 how there is real messy data. 290 00:23:05,480 --> 00:23:12,720 But then the more I started to think about this, we can often turn seemingly disadvantages 291 00:23:12,720 --> 00:23:16,080 into our advantage. 292 00:23:16,080 --> 00:23:20,680 So basically this data appears daunting. 293 00:23:20,680 --> 00:23:24,640 So maybe not many people look at it. 294 00:23:24,640 --> 00:23:30,840 So that started to make me think that there may be data hiding in plain sight. 295 00:23:30,840 --> 00:23:41,360 Second thing is, right, AIs, algorithms can do work for extremely, extremely low costs. 296 00:23:41,360 --> 00:23:49,080 Their marginal cost is near zero where if you were going to hire a human to do natural 297 00:23:49,080 --> 00:23:54,520 language processing, so basically look at human written text and parse the meaning out 298 00:23:54,520 --> 00:24:00,560 of it, you're going to be paying them a pretty high wage. 299 00:24:00,560 --> 00:24:03,160 They're going to only be so effective. 300 00:24:03,160 --> 00:24:05,720 It's going to be really time consuming. 301 00:24:05,720 --> 00:24:15,360 So I realized now that we actually have tools like our friend here, Lambda, or tools that 302 00:24:15,360 --> 00:24:25,680 Lambda has that will help us out, then we may be able to do extraordinary things because 303 00:24:25,680 --> 00:24:37,640 we can now parse human text at such a fast rate. 304 00:24:37,640 --> 00:24:46,160 So I'll quit going on about that, and I'll show you how this can be used. 305 00:24:46,160 --> 00:24:53,360 So pardon the equations, but just wanted to show you two concepts real quick from economics 306 00:24:53,360 --> 00:24:57,360 that we'll use quite well today. 307 00:24:57,360 --> 00:25:05,480 And then these will be in your lexicon, and basically if you were going to boil down the 308 00:25:05,480 --> 00:25:11,920 useful tools out of microeconomics, this would basically be what you're left with. 309 00:25:11,920 --> 00:25:15,480 I mean maybe a bit more. 310 00:25:15,480 --> 00:25:19,320 But basically we talked about expected utility before. 311 00:25:19,320 --> 00:25:24,520 This was an idea that Johnny von Neumann introduced, I do believe. 312 00:25:24,520 --> 00:25:31,600 And so it's basically saying you have to take into consideration all the different states 313 00:25:31,600 --> 00:25:33,160 of reality. 314 00:25:33,160 --> 00:25:42,760 So if you're in the desert, state one, your utility from water is going to be really high 315 00:25:42,760 --> 00:25:49,160 versus if you're in the city, state two, then your utility from water is going to be really 316 00:25:49,160 --> 00:25:50,360 low. 317 00:25:50,360 --> 00:25:56,800 So your expected utility is the probability that you're going to be in the desert times 318 00:25:56,800 --> 00:26:01,840 your utility of water in the desert plus the probability that you're going to be in the 319 00:26:01,840 --> 00:26:06,920 city times the utility of water when you're in the city. 320 00:26:06,920 --> 00:26:13,520 So that's sort of a basic example, but this is a really powerful idea. 321 00:26:13,520 --> 00:26:19,800 And the way this is going to tie in today is we're essentially talking about utility 322 00:26:19,800 --> 00:26:21,440 from cannabis. 323 00:26:21,440 --> 00:26:30,160 So you can think about sea as cannabis or specifically all the compounds in cannabis. 324 00:26:30,160 --> 00:26:37,880 And so you'll get utility from cannabis and the utility you get will depend on the state. 325 00:26:37,880 --> 00:26:47,600 So I may consume certain compounds and if I am in state one, energetic, I may get a 326 00:26:47,600 --> 00:26:57,720 certain amount of utility, but maybe I consume the same compounds and I get state two, sleepy. 327 00:26:57,720 --> 00:27:00,320 My utility may be quite different. 328 00:27:00,320 --> 00:27:07,760 So basically my expected utility from consuming cannabis would be the probability that it's 329 00:27:07,760 --> 00:27:13,640 going to make me sleepy times the utility I get when I'm sleepy plus the probability 330 00:27:13,640 --> 00:27:21,680 I get when I'm not sleepy times the utility I get from those compounds when I'm not sleeping. 331 00:27:21,680 --> 00:27:29,760 So this can just kind of be extended essentially in depth. 332 00:27:29,760 --> 00:27:34,520 We can kind of go to the limit in infinite number of states. 333 00:27:34,520 --> 00:27:37,680 You could be happy and focused. 334 00:27:37,680 --> 00:27:41,560 You could be happy, focused and energetic. 335 00:27:41,560 --> 00:27:43,760 You could be creative and sleepy. 336 00:27:43,760 --> 00:27:49,440 So get your creative done before you fall asleep. 337 00:27:49,440 --> 00:27:55,880 So long story short is this model gives us a lot of flexibility over states. 338 00:27:55,880 --> 00:27:58,480 Okay, cool. 339 00:27:58,480 --> 00:28:08,280 Well now we can actually put a functional form on these specific utility functions and a 340 00:28:08,280 --> 00:28:18,000 convenient utility function that satisfies all of the assumptions from microeconomics. 341 00:28:18,000 --> 00:28:21,920 So that's what's cool about the Cobb Douglas utility function is it checks all the check 342 00:28:21,920 --> 00:28:25,980 boxes that we want in a utility function. 343 00:28:25,980 --> 00:28:30,680 So mainly that you get diminishing marginal returns. 344 00:28:30,680 --> 00:28:40,040 So the more cannabis you consume, it's going to have less and less of a beneficial effect. 345 00:28:40,040 --> 00:28:42,680 Cool. 346 00:28:42,680 --> 00:28:49,520 Economists don't usually measure utility, but I'm not your typical economist. 347 00:28:49,520 --> 00:28:58,840 The more I kind of thought about this, the more happy I am with it in that this would 348 00:28:58,840 --> 00:29:10,920 definitely be something that any number of my former economics professors would have 349 00:29:10,920 --> 00:29:12,560 probably frowned upon. 350 00:29:12,560 --> 00:29:23,280 But they also frowned upon other essentially predictions I made that turned out to be quite 351 00:29:23,280 --> 00:29:24,280 interesting. 352 00:29:24,280 --> 00:29:39,400 So for example, way back in 2009, I was talking about this little company called SpaceX. 353 00:29:39,400 --> 00:29:48,480 And my economics professors pooh-poohed that idea in that company. 354 00:29:48,480 --> 00:30:00,640 And now look at SpaceX today, a decade later, and all the top engineers in the country are 355 00:30:00,640 --> 00:30:05,760 vying for positions at SpaceX. 356 00:30:05,760 --> 00:30:14,720 So long story short, you're a classical economist with pooh-pooh, the idea of measuring utility. 357 00:30:14,720 --> 00:30:20,760 But just kind of for fun today, we'll essentially be doing exactly that. 358 00:30:20,760 --> 00:30:27,240 So we've got our functional form today for our utility function. 359 00:30:27,240 --> 00:30:33,180 And coincidentally, it turns out nicely for our linear regression. 360 00:30:33,180 --> 00:30:41,600 So here we just have a linear equation, alpha log x. 361 00:30:41,600 --> 00:30:44,040 And I won't go too much more into this. 362 00:30:44,040 --> 00:30:48,200 But if you are interested in the statistics, come to Saturday Morning Statistics or sign 363 00:30:48,200 --> 00:30:50,280 up and I'll send you all the material. 364 00:30:50,280 --> 00:30:56,520 I mean, we'll actually, for example, last week we talked about the importance of log 365 00:30:56,520 --> 00:31:03,760 normalizing your variables, and now we actually have a theoretical reason why we should take 366 00:31:03,760 --> 00:31:05,920 the log of our variables. 367 00:31:05,920 --> 00:31:12,520 So it works out well statistically and it matches our theory. 368 00:31:12,520 --> 00:31:13,940 Awesome. 369 00:31:13,940 --> 00:31:18,900 So let's tie it into machine learning. 370 00:31:18,900 --> 00:31:22,760 So in natural language processing. 371 00:31:22,760 --> 00:31:30,640 So long story short is we've got these reviews, and I'll get to the data here momentarily. 372 00:31:30,640 --> 00:31:34,080 And then quit droning on. 373 00:31:34,080 --> 00:31:37,200 But we've got our reviews. 374 00:31:37,200 --> 00:31:50,280 And what we can do is we can basically let the algorithm determine how positive or negative 375 00:31:50,280 --> 00:31:52,520 the review is. 376 00:31:52,520 --> 00:32:01,280 So that's basically what I'm basically letting that be the interpretation of utility, of 377 00:32:01,280 --> 00:32:02,280 happiness. 378 00:32:02,280 --> 00:32:10,440 And so it's kind of getting a little abstract, but kind of going back to lambda, we can process 379 00:32:10,440 --> 00:32:13,960 natural language like a human can. 380 00:32:13,960 --> 00:32:21,480 And so the idea behind sentiment analysis is given a review, if you sat a person down 381 00:32:21,480 --> 00:32:29,640 and said, OK, on a scale of 1 to 10, how happy is the reviewer? 382 00:32:29,640 --> 00:32:39,640 And you'd read their review and you'd kind of make a subjective interpretation of what 383 00:32:39,640 --> 00:32:45,240 you think this person's happiness may be. 384 00:32:45,240 --> 00:32:49,280 And humans are really good at this type of thing. 385 00:32:49,280 --> 00:32:52,360 We're really good at reading. 386 00:32:52,360 --> 00:32:58,480 Well, I mean, this is more emotional cues, but we're really good at picking up cues 387 00:32:58,480 --> 00:32:59,480 from each other. 388 00:32:59,480 --> 00:33:03,080 So it's like, oh, like they use this word. 389 00:33:03,080 --> 00:33:05,920 They're maybe being sarcastic. 390 00:33:05,920 --> 00:33:10,460 Or they are using a lot of negative words. 391 00:33:10,460 --> 00:33:12,080 They're maybe not in a good mood. 392 00:33:12,080 --> 00:33:14,960 Or, oh, they're using a lot of positive words. 393 00:33:14,960 --> 00:33:17,120 They're probably in a good mood. 394 00:33:17,120 --> 00:33:19,840 And so we kind of pick up these things passively. 395 00:33:19,840 --> 00:33:27,040 Well, if this algorithm really is as good as a human, which it may not be, but maybe 396 00:33:27,040 --> 00:33:37,760 it's cost effective, then maybe this algorithm can determine how happy somebody is from their 397 00:33:37,760 --> 00:33:38,760 review. 398 00:33:38,760 --> 00:33:44,840 So instead of paying someone $30 an hour to sit down and rank the reviews, we'll have 399 00:33:44,840 --> 00:33:52,000 the algorithm do it at virtually no cost, pennies on the dollar. 400 00:33:52,000 --> 00:33:57,760 And then we can get a measure of you. 401 00:33:57,760 --> 00:34:07,600 And then just a couple statistical concepts just to not go too much into the math. 402 00:34:07,600 --> 00:34:13,520 The way that we could actually estimate our Cobb-Douglas utility function just by having 403 00:34:13,520 --> 00:34:21,600 you on one side and then essentially our lab results on the other side. 404 00:34:21,600 --> 00:34:30,600 And then we could actually then estimate an expected utility function by having an interaction 405 00:34:30,600 --> 00:34:40,880 between the dummy variable 0 or 1 for a particular effect interacted with these compounds. 406 00:34:40,880 --> 00:34:52,360 So basically this alpha log x would just be your baseline utility in state 1. 407 00:34:52,360 --> 00:35:01,800 And then if you got sleepy, then your utility would then be adjusted. 408 00:35:01,800 --> 00:35:07,840 And then you'd have your utility in state 2. 409 00:35:07,840 --> 00:35:16,680 So you could actually estimate your entire expected utility function. 410 00:35:16,680 --> 00:35:22,040 And then I know I'm kind of getting into the weeds, so I'll get to the data real quick. 411 00:35:22,040 --> 00:35:30,840 But then there's an idea of ex ante before and ex post after utility. 412 00:35:30,840 --> 00:35:42,640 So after you consume the cannabis, you'll have actual utility that is actual depending 413 00:35:42,640 --> 00:35:46,800 on the actual effects you had. 414 00:35:46,800 --> 00:35:54,400 But before you consume the cannabis, you'll only have expectations of the probability 415 00:35:54,400 --> 00:35:56,480 that you'll get sleepy. 416 00:35:56,480 --> 00:36:04,040 So long story short is instead of plugging in 0 or 1 for D, we could plug in the expected 417 00:36:04,040 --> 00:36:09,040 probability of being sleepy, say 0.2. 418 00:36:09,040 --> 00:36:13,560 And then that would give us our expected utility. 419 00:36:13,560 --> 00:36:26,400 So long story short is we can now get expected utility given these reviews and lab results. 420 00:36:26,400 --> 00:36:36,320 So now you've got a baseline expectation for the probability. 421 00:36:36,320 --> 00:36:39,520 So all you have to do is get lab results. 422 00:36:39,520 --> 00:36:45,840 And now you have an expected utility for a given product. 423 00:36:45,840 --> 00:36:49,240 So sorry, I'm getting way into the weeds. 424 00:36:49,240 --> 00:36:51,960 Now is the actual algorithm. 425 00:36:51,960 --> 00:36:56,840 So this is the algorithm that we're about to code up. 426 00:36:56,840 --> 00:37:05,660 So step one, given all the reviews, we'll rank them minus 1 to 1. 427 00:37:05,660 --> 00:37:09,840 This is going to be our proxy for utility. 428 00:37:09,840 --> 00:37:21,720 Step two, we estimate a regression of utility on lab results interacted with the effects. 429 00:37:21,720 --> 00:37:22,920 Cool. 430 00:37:22,920 --> 00:37:24,880 We can do that. 431 00:37:24,880 --> 00:37:29,440 Step three, you now have a utility function. 432 00:37:29,440 --> 00:37:32,680 And so you can now put it to good use. 433 00:37:32,680 --> 00:37:45,280 And so for example, basically, I was reading on semantic or sentiment analysis, and they 434 00:37:45,280 --> 00:37:49,520 said a classic tool for this is recommendation engines. 435 00:37:49,520 --> 00:37:52,280 And so the idea is, OK, well, cool. 436 00:37:52,280 --> 00:37:54,680 So now you go to the store. 437 00:37:54,680 --> 00:37:58,080 You have a list of products. 438 00:37:58,080 --> 00:38:04,600 You plug all those products into your expected utility function. 439 00:38:04,600 --> 00:38:09,560 You use, say, the skunk effects model. 440 00:38:09,560 --> 00:38:15,960 So that way you have a nice predicted probability for all the effects. 441 00:38:15,960 --> 00:38:24,200 And so now you can get an expected utility for all of the products. 442 00:38:24,200 --> 00:38:28,160 And you let your app do all this heavy lifting. 443 00:38:28,160 --> 00:38:31,360 But basically, you go into the store. 444 00:38:31,360 --> 00:38:32,520 You've got your app. 445 00:38:32,520 --> 00:38:35,520 Your app knows all the products in the store. 446 00:38:35,520 --> 00:38:37,440 Bop, bop, bop, bop. 447 00:38:37,440 --> 00:38:45,880 And it just puts the product that would give you the highest expected utility at the top. 448 00:38:45,880 --> 00:38:55,400 And then rank those by highest utility downwards. 449 00:38:55,400 --> 00:38:56,400 Quick question. 450 00:38:56,400 --> 00:38:59,400 Maybe I don't understand exactly how this is working. 451 00:38:59,400 --> 00:39:03,680 But I thought when you used dummy variables, they were Boolean. 452 00:39:03,680 --> 00:39:06,480 So it's either 1 or 0. 453 00:39:06,480 --> 00:39:07,480 Yes. 454 00:39:07,480 --> 00:39:17,280 So essentially, this is saying that if you're sleepy, your utility is going to be different 455 00:39:17,280 --> 00:39:22,280 from the cannabinoids than when you're not sleepy. 456 00:39:22,280 --> 00:39:29,280 Does that make sense? 457 00:39:29,280 --> 00:39:37,560 So if we had 15 different effects that we were tracking, each one would have either 458 00:39:37,560 --> 00:39:40,200 a 1 or a 0? 459 00:39:40,200 --> 00:39:41,480 Exactly. 460 00:39:41,480 --> 00:39:47,880 So basically, what you're saying is there would be 15 states of reality. 461 00:39:47,880 --> 00:39:53,560 So it's basically like there's 15 states. 462 00:39:53,560 --> 00:39:57,280 You could be in a state where you're happy. 463 00:39:57,280 --> 00:40:00,840 You may be in a state where you're focused. 464 00:40:00,840 --> 00:40:08,160 You could potentially get complex with this and say a state where you're focused and happy 465 00:40:08,160 --> 00:40:11,160 is different than a state where you're just happy. 466 00:40:11,160 --> 00:40:12,160 Yeah. 467 00:40:12,160 --> 00:40:18,240 I would imagine that if you had 15 different columns, each one being an effect, each column, 468 00:40:18,240 --> 00:40:27,200 each row of data having a 1 or a 0 in it, you could have more than one 1 in each row. 469 00:40:27,200 --> 00:40:28,200 Exactly. 470 00:40:28,200 --> 00:40:34,320 And so it's going to add a lot of complexity to your model. 471 00:40:34,320 --> 00:40:40,280 But the way I would do this is I would call those different states. 472 00:40:40,280 --> 00:40:48,720 So basically, every single combination is a different state. 473 00:40:48,720 --> 00:40:56,120 And so you're going to end up with a lot of parameters really, really quickly if you start 474 00:40:56,120 --> 00:41:00,160 to add a lot of different effects. 475 00:41:00,160 --> 00:41:07,680 But that's okay because we can let our – remember degrees of freedom out of statistics. 476 00:41:07,680 --> 00:41:12,960 We just need a lot more observations than we have degrees of freedom. 477 00:41:12,960 --> 00:41:16,880 So if we don't have very many observations, we're going to chew up our degrees of freedom 478 00:41:16,880 --> 00:41:20,040 really, really quickly by adding effects. 479 00:41:20,040 --> 00:41:26,560 But the idea is these companies have thousands of reviews. 480 00:41:26,560 --> 00:41:33,120 So once you get thousands of reviews, you can put thousands of parameters in your model. 481 00:41:33,120 --> 00:41:40,800 So here I'll get out of abstract land and actually estimate this for you just to kind 482 00:41:40,800 --> 00:41:42,880 of ground it in reality. 483 00:41:42,880 --> 00:41:47,680 But I'm just going to do it with one effect today, sleepy. 484 00:41:47,680 --> 00:41:57,600 But the idea is it's perfectly generalizable so that way algorithms in coding is powerful. 485 00:41:57,600 --> 00:42:07,120 So it would be not that much work to generalize this to – I mean it would be some work. 486 00:42:07,120 --> 00:42:14,240 But you could generalize this to S number of effects. 487 00:42:14,240 --> 00:42:19,600 And then as long as you have enough data, then you can just kind of estimate a really, 488 00:42:19,600 --> 00:42:23,560 really robust model for expected utility. 489 00:42:23,560 --> 00:42:31,080 And is skunk effects going to be how this data is generated? 490 00:42:31,080 --> 00:42:33,800 Is that what you're thinking? 491 00:42:33,800 --> 00:42:40,200 I was thinking about using skunk effects essentially – so basically what skunk effects predicts 492 00:42:40,200 --> 00:42:42,400 is P. 493 00:42:42,400 --> 00:42:50,960 So you'll need some way to estimate P because you don't actually know the true P. 494 00:42:50,960 --> 00:42:53,320 So there's different ways you can do this. 495 00:42:53,320 --> 00:42:57,480 So you could just do basically the average. 496 00:42:57,480 --> 00:43:04,680 So what's the mean occurrence of sleepy? 497 00:43:04,680 --> 00:43:09,980 But the idea is skunk effects is a conditional mean. 498 00:43:09,980 --> 00:43:15,000 So skunk effects would be a mean that's conditional on X. 499 00:43:15,000 --> 00:43:18,280 So this would be P of X. 500 00:43:18,280 --> 00:43:25,560 So your probability of your state depends on the lab results, which is a pretty reasonable 501 00:43:25,560 --> 00:43:28,200 assumption. 502 00:43:28,200 --> 00:43:32,600 So the way that you can incorporate skunk effects into this is you would just say this 503 00:43:32,600 --> 00:43:34,400 is P of X. 504 00:43:34,400 --> 00:43:42,120 And so you estimate P of X with skunk effects, and then you estimate your ugliest utility 505 00:43:42,120 --> 00:43:46,620 function, which is just an ordinary least squares regression. 506 00:43:46,620 --> 00:43:50,040 And then you basically combine the two models. 507 00:43:50,040 --> 00:43:53,160 And then as I said, it's going to be complex. 508 00:43:53,160 --> 00:43:59,040 But you're estimating utility. 509 00:43:59,040 --> 00:44:06,880 This is an abstract measure of happiness, like how happy, quote unquote, is somebody 510 00:44:06,880 --> 00:44:09,120 from this cannabis. 511 00:44:09,120 --> 00:44:12,520 So this is a really, really complex function. 512 00:44:12,520 --> 00:44:20,440 So it's not unreasonable for it to kind of take this complex form. 513 00:44:20,440 --> 00:44:21,440 Right? 514 00:44:21,440 --> 00:44:22,440 And that's why... 515 00:44:22,440 --> 00:44:28,240 I'm not just concerned about where you're going to get all the observations from. 516 00:44:28,240 --> 00:44:29,840 I'll show you right now. 517 00:44:29,840 --> 00:44:33,960 So long story short, but I'll just leave this. 518 00:44:33,960 --> 00:44:41,960 My last comment was, Ray, we're writing a complex algorithm that's doing the work of 519 00:44:41,960 --> 00:44:44,840 a human brain. 520 00:44:44,840 --> 00:44:47,120 The human brain is powerful. 521 00:44:47,120 --> 00:44:57,460 So the human brain sees a list of products and boom, just like that, the human brain 522 00:44:57,460 --> 00:45:06,760 estimates P. The human brain estimates utility in all those different states and then estimates 523 00:45:06,760 --> 00:45:10,480 an expected utility from all those products. 524 00:45:10,480 --> 00:45:17,040 And so that's why you may look at a menu for a little while because your brain is kind 525 00:45:17,040 --> 00:45:18,960 of churning. 526 00:45:18,960 --> 00:45:23,720 That is centrally the ideas... 527 00:45:23,720 --> 00:45:28,920 The idea behind the theory is this is kind of what's going on in your brain. 528 00:45:28,920 --> 00:45:36,100 So if we can replicate the same process that your brain's using and let the computer do 529 00:45:36,100 --> 00:45:45,520 all the heavy lifting, you can save your brain power to do more interesting things. 530 00:45:45,520 --> 00:45:54,600 And then pick which product at the store is going to be the best for you. 531 00:45:54,600 --> 00:46:00,240 You can spend your time doing mathematics and statistics and fun things like that. 532 00:46:00,240 --> 00:46:02,540 You can attend Saturday morning statistics. 533 00:46:02,540 --> 00:46:05,320 So what did you get? 534 00:46:05,320 --> 00:46:07,680 So what do we get? 535 00:46:07,680 --> 00:46:10,720 Well, we are. 536 00:46:10,720 --> 00:46:16,520 So we've got our reviews and just to go ahead and do everything that Charles would... 537 00:46:16,520 --> 00:46:21,920 Whoops, I should have showed you this beforehand. 538 00:46:21,920 --> 00:46:27,800 Okay, so I accidentally already deleted them. 539 00:46:27,800 --> 00:46:32,360 But Charles pointed... 540 00:46:32,360 --> 00:46:33,360 Or someone had pointed... 541 00:46:33,360 --> 00:46:37,200 Or maybe John, you had pointed out there were a lot of duplicates. 542 00:46:37,200 --> 00:46:44,080 And so I actually added a user column. 543 00:46:44,080 --> 00:46:51,480 So I went through and parsed the data again and got all the users. 544 00:46:51,480 --> 00:46:56,760 And then I was looking at these two columns and there were just multiple times where the 545 00:46:56,760 --> 00:47:01,240 same user had left the same review. 546 00:47:01,240 --> 00:47:11,680 So I'm sorry that I just now caught this, but there's almost 20,000 duplicates. 547 00:47:11,680 --> 00:47:15,600 So my apologies that I caught them so late. 548 00:47:15,600 --> 00:47:18,540 Our statistics up until now are biased. 549 00:47:18,540 --> 00:47:22,880 So it would be worthwhile to go back and recalculate those. 550 00:47:22,880 --> 00:47:26,360 But the Python philosophy is now is better than never. 551 00:47:26,360 --> 00:47:34,560 So we at least identified this now, and so the point being is by the end of today, by 552 00:47:34,560 --> 00:47:40,040 the end of this script, our data is going to be really, really robust. 553 00:47:40,040 --> 00:47:45,120 Or at least it's more robust than it was. 554 00:47:45,120 --> 00:47:53,000 As I said, and I'll get through this in the next five minutes, but basically please forgive 555 00:47:53,000 --> 00:47:59,280 the shortcomings in the data and just let this be an educational example for you. 556 00:47:59,280 --> 00:48:05,980 These same statistics, the same model can be used with any review data. 557 00:48:05,980 --> 00:48:11,920 So I highly encourage you to use your own reviews that specifically match up to specific 558 00:48:11,920 --> 00:48:12,920 products. 559 00:48:12,920 --> 00:48:16,280 And that data is pure gold. 560 00:48:16,280 --> 00:48:21,320 I'm just kind of using this dummy data for educational purposes. 561 00:48:21,320 --> 00:48:24,480 Can you screen for bots? 562 00:48:24,480 --> 00:48:25,480 Say that one more time. 563 00:48:25,480 --> 00:48:30,440 Can you screen for bots who are artificially non-human reviews? 564 00:48:30,440 --> 00:48:35,920 Brilliant question slash observation, Jerry. 565 00:48:35,920 --> 00:48:39,720 So yes, so now let's look at this data. 566 00:48:39,720 --> 00:48:44,680 So okay, so now how many unique users are there? 567 00:48:44,680 --> 00:48:50,080 Like are they just all the same person? 568 00:48:50,080 --> 00:48:59,280 So what we find is there's fewer users than there are reviews, which is actually awesome 569 00:48:59,280 --> 00:49:04,880 because this means the same user is leaving multiple reviews. 570 00:49:04,880 --> 00:49:14,280 So is there a question? 571 00:49:14,280 --> 00:49:22,000 So what if the same user is leaving reviews for different strengths? 572 00:49:22,000 --> 00:49:25,080 That is exactly what I think we want. 573 00:49:25,080 --> 00:49:30,760 So for example, so here this is a bit more logical. 574 00:49:30,760 --> 00:49:37,760 So here I would just like, okay, let's count the reviews by user. 575 00:49:37,760 --> 00:49:43,780 And so you see, okay, anonymous is leaving the most reviews. 576 00:49:43,780 --> 00:49:53,520 So I actually end up excluding those because I heard people kind of bash on anonymous reviews. 577 00:49:53,520 --> 00:49:59,440 So we may or may not want to use them. 578 00:49:59,440 --> 00:50:08,020 But long story short is most people are leaving one review and that's it. 579 00:50:08,020 --> 00:50:17,040 So more than 75% of people are just leaving one review and calling it a day. 580 00:50:17,040 --> 00:50:24,200 But here I just picked this person out just because they had Washington in their name. 581 00:50:24,200 --> 00:50:34,280 But if you see the strains that this person's reviewing, it actually gets real interesting 582 00:50:34,280 --> 00:50:35,680 because check this out. 583 00:50:35,680 --> 00:50:41,640 So this person reviewed 24 karat gold, cookies push. 584 00:50:41,640 --> 00:50:49,200 They reviewed DJ Short Blueberry three times. 585 00:50:49,200 --> 00:50:54,340 So these are three different reviews. 586 00:50:54,340 --> 00:51:02,160 So one may think that maybe they got this product three times. 587 00:51:02,160 --> 00:51:09,360 But they're just a pretty avid reviewer. 588 00:51:09,360 --> 00:51:15,520 Let's see if we can't look at one more here. 589 00:51:15,520 --> 00:51:22,760 And then I'll get to the statistics since we're running low on time here. 590 00:51:22,760 --> 00:51:25,440 This introduces a couple things. 591 00:51:25,440 --> 00:51:30,520 One, it does introduce a potential source of bias. 592 00:51:30,520 --> 00:51:42,240 If one user is leaving a disproportionate number of reviews, then that could bias things 593 00:51:42,240 --> 00:51:47,720 towards that one reviewer's preferences. 594 00:51:47,720 --> 00:51:53,240 For example, Chill Panda has left all these reviews. 595 00:51:53,240 --> 00:52:03,600 We may just be really, really good at predicting Chill Panda's effects. 596 00:52:03,600 --> 00:52:12,920 And so that's actually okay because that's what statistics is for in that statistics 597 00:52:12,920 --> 00:52:22,480 is all about uncertainty and how do we actually use this variability to our advantage. 598 00:52:22,480 --> 00:52:41,920 And I think this is actually awesome because now we can actually condition effects on the 599 00:52:41,920 --> 00:52:42,920 user. 600 00:52:42,920 --> 00:52:48,800 And so John, you brought this up real briefly at the end of the last meetup, and I'm not 601 00:52:48,800 --> 00:52:52,960 sure if I fully sunk home with me. 602 00:52:52,960 --> 00:53:00,800 But basically, different people may have different effects. 603 00:53:00,800 --> 00:53:05,400 Different people may gravitate towards different strains. 604 00:53:05,400 --> 00:53:11,120 So for example, I just want to go ahead and show you this data real quick. 605 00:53:11,120 --> 00:53:17,720 So basically, just I'm excluding anonymous. 606 00:53:17,720 --> 00:53:24,680 I'm keeping only the observations with these compounds of interest that were detected. 607 00:53:24,680 --> 00:53:31,000 And we're left with a training sample of still around 4,000. 608 00:53:31,000 --> 00:53:44,640 So we started with what, 42,000 reviews, but then we had to remove duplicates. 609 00:53:44,640 --> 00:53:47,920 We had to remove things with non-detects. 610 00:53:47,920 --> 00:53:50,760 So there's a lot of noise in the data. 611 00:53:50,760 --> 00:53:58,520 But once we remove all the extraneous noise and take a random sample, we've got a good 612 00:53:58,520 --> 00:54:00,160 training set. 613 00:54:00,160 --> 00:54:05,680 We've got 4,000 nice random samples here. 614 00:54:05,680 --> 00:54:10,480 And I think we can do some good work with this training set, this sample. 615 00:54:10,480 --> 00:54:19,160 Keegan, this is heavily skewed to the low beta-pinene-limonene group. 616 00:54:19,160 --> 00:54:25,040 This is, actually, I love your take on this, because I wasn't sure what to make of this 617 00:54:25,040 --> 00:54:26,040 plot. 618 00:54:26,040 --> 00:54:27,860 Yeah, this looks pretty skewed to me. 619 00:54:27,860 --> 00:54:36,860 So what I did here was I averaged all the lab results by user. 620 00:54:36,860 --> 00:54:41,660 So this is basically a user's profile. 621 00:54:41,660 --> 00:54:52,080 So this is sort of the average beta-pinene to D-limonene that a consumer chooses. 622 00:54:52,080 --> 00:54:55,960 So they may choose samples all over the board. 623 00:54:55,960 --> 00:55:06,400 But on average, somebody over here, they may have a predilection towards indica-type strains. 624 00:55:06,400 --> 00:55:11,640 And so each of these dots corresponds to a specific user. 625 00:55:11,640 --> 00:55:18,160 There's no mapping between the user and the lab results. 626 00:55:18,160 --> 00:55:22,440 There's a mapping between the strain and the lab results. 627 00:55:22,440 --> 00:55:29,840 But there are several lab results for each strain. 628 00:55:29,840 --> 00:55:39,280 And so if you just average that down, then you're sort of compressing the data. 629 00:55:39,280 --> 00:55:41,360 You're 100% correct. 630 00:55:41,360 --> 00:55:50,200 And so basically, I'm using the average lab results as a proxy for the actual lab results 631 00:55:50,200 --> 00:55:52,360 of this person's product. 632 00:55:52,360 --> 00:55:59,080 So that's why I was saying, if you actually have the lab results of the product the person 633 00:55:59,080 --> 00:56:03,400 consumed, that is the match you want. 634 00:56:03,400 --> 00:56:09,680 So I'm just kind of playing pretend that we have that match. 635 00:56:09,680 --> 00:56:15,760 So you see, we're kind of taking a leap of faith. 636 00:56:15,760 --> 00:56:21,120 It's not perfect, but I don't want to say I see other economists do this. 637 00:56:21,120 --> 00:56:31,240 But it's basically the variant of interest is the compounds in the cannabis strain that 638 00:56:31,240 --> 00:56:34,520 someone consumed and reviewed. 639 00:56:34,520 --> 00:56:35,840 That's missing. 640 00:56:35,840 --> 00:56:38,080 So we have to estimate that. 641 00:56:38,080 --> 00:56:41,880 And so we're estimating it with the average lab results. 642 00:56:41,880 --> 00:56:45,660 So it's way, way far from perfect. 643 00:56:45,660 --> 00:57:06,120 But the idea is, oh, if you look at the strains that Green Green Washington's picking, are 644 00:57:06,120 --> 00:57:11,720 these Exodus Cheese, Marion, Mary Cush? 645 00:57:11,720 --> 00:57:21,400 On average, are these more this high beta-pinene to d-lemonine type, slash sativa? 646 00:57:21,400 --> 00:57:26,400 Or are they on average more indica type? 647 00:57:26,400 --> 00:57:33,920 So I think you're... 648 00:57:33,920 --> 00:57:39,920 By compressing this, you're losing a lot of information and you're repeating the same 649 00:57:39,920 --> 00:57:43,200 information over and over again. 650 00:57:43,200 --> 00:57:50,960 So there's 4,000 some odd reviews and there's around 100... 651 00:57:50,960 --> 00:57:58,720 After you compress the data and remove all the duplicates, there's maybe like 100 different 652 00:57:58,720 --> 00:57:59,720 strains. 653 00:57:59,720 --> 00:58:00,720 184, three. 654 00:58:00,720 --> 00:58:01,720 183. 655 00:58:01,720 --> 00:58:02,720 Yeah. 656 00:58:02,720 --> 00:58:11,000 So it's... Charles, can I... 657 00:58:11,000 --> 00:58:13,240 Can we make a date to look offline? 658 00:58:13,240 --> 00:58:19,720 Because I want to show you, I think that that 183 set is extremely well behaved. 659 00:58:19,720 --> 00:58:26,800 I can try and convince you of that through principal component analysis and loading plots 660 00:58:26,800 --> 00:58:29,400 that we've done. 661 00:58:29,400 --> 00:58:37,160 So I am of the opinion that that 183 mean data set has value. 662 00:58:37,160 --> 00:58:38,160 Okay. 663 00:58:38,160 --> 00:58:39,160 Yeah. 664 00:58:39,160 --> 00:58:42,160 I'd like to see it. 665 00:58:42,160 --> 00:58:48,440 I need to walk you through the principal component analysis results and let's make a date for 666 00:58:48,440 --> 00:58:49,440 that. 667 00:58:49,440 --> 00:58:54,240 I'll see if I can convince you because there's a lot of story in it that I can describe to 668 00:58:54,240 --> 00:58:55,240 you. 669 00:58:55,240 --> 00:58:56,240 Okay. 670 00:58:56,240 --> 00:58:57,240 That would be great. 671 00:58:57,240 --> 00:58:58,240 I really like that. 672 00:58:58,240 --> 00:58:59,240 Yeah. 673 00:58:59,240 --> 00:59:04,320 Again, I come at this from the biochemistry. 674 00:59:04,320 --> 00:59:06,600 So just to... 675 00:59:06,600 --> 00:59:13,160 So basically, you're 100% right, Charles, but basically I don't want to let this stop 676 00:59:13,160 --> 00:59:22,240 us from fitting this model because the model can be used by any single company that's got 677 00:59:22,240 --> 00:59:32,080 reviews and lab results and thousands of company, or maybe not thousands, but all the big apps 678 00:59:32,080 --> 00:59:37,200 have reviews and lab results, one-to-one matches. 679 00:59:37,200 --> 00:59:46,260 So we don't have that data, but we can make this model and then weed maps, Leafly, any 680 00:59:46,260 --> 00:59:51,080 of these big companies could use this algorithm. 681 00:59:51,080 --> 01:00:01,240 And I argue just because we're missing the data, it's still a pretty good proxy because 682 01:00:01,240 --> 01:00:19,960 basically what I'm saying here is this person, they sampled DJ short blueberry three times. 683 01:00:19,960 --> 01:00:24,360 One of the three times it made them sleepy. 684 01:00:24,360 --> 01:00:36,520 And so it would be nice to know the total THC of each DJ short blueberry they sampled. 685 01:00:36,520 --> 01:00:40,400 But we've even heard all this noise. 686 01:00:40,400 --> 01:00:44,280 Go talk to any cultivator and they'll... 687 01:00:44,280 --> 01:00:51,000 Not any, but most cultivators will just complain to you all day long about how their lab results 688 01:00:51,000 --> 01:00:55,280 aren't representative of their product. 689 01:00:55,280 --> 01:00:59,840 And they say that's a problem because they say, oh, we send in one sample and we get 690 01:00:59,840 --> 01:01:04,120 one lab result back, and then that's what goes on the label. 691 01:01:04,120 --> 01:01:12,360 So one could even make the argument that the average is even a better metric than what's 692 01:01:12,360 --> 01:01:14,200 on the label. 693 01:01:14,200 --> 01:01:19,120 I don't think I would go that far. 694 01:01:19,120 --> 01:01:29,160 But the long story short is we're basically saying this strain has on average 19.5% THC. 695 01:01:29,160 --> 01:01:34,880 It may make you sleepy a third of the time. 696 01:01:34,880 --> 01:01:40,560 We'll have to let the data pan that out. 697 01:01:40,560 --> 01:01:44,360 And then it kind of goes on with the other variables. 698 01:01:44,360 --> 01:01:56,340 So the fact that the lab results repeat is okay as long as you don't have perfect collinearity. 699 01:01:56,340 --> 01:02:05,800 So as long as people are experiencing different effects. 700 01:02:05,800 --> 01:02:11,400 I'm just wondering if it would help if instead of doing natural language processing, when 701 01:02:11,400 --> 01:02:18,280 we gather the data, we use like drop down menus, like radio bullets. 702 01:02:18,280 --> 01:02:21,920 What effect did you experience from this product? 703 01:02:21,920 --> 01:02:27,880 Jerry, that's exactly what we're doing in our dosing project work. 704 01:02:27,880 --> 01:02:34,000 It will solve a lot of this because it will be a given user matched to a given strain 705 01:02:34,000 --> 01:02:36,800 with a given COA and a given effect. 706 01:02:36,800 --> 01:02:41,840 So what we're doing now is kind of modeling off of what Keyin is presenting. 707 01:02:41,840 --> 01:02:47,840 But we're going to have access, God willing, to a much better data set. 708 01:02:47,840 --> 01:02:52,200 That's precisely the approach we're taking. 709 01:02:52,200 --> 01:02:55,640 I think I may have ruffled a few feathers, but that's good because... 710 01:02:55,640 --> 01:02:58,000 No, you're leading the way. 711 01:02:58,000 --> 01:03:01,920 I mean, it's just a lot of twists and turns. 712 01:03:01,920 --> 01:03:02,920 Yeah. 713 01:03:02,920 --> 01:03:07,240 Well, like I said, we're estimating a utility function here. 714 01:03:07,240 --> 01:03:10,920 So it's not just a trivial thing to do. 715 01:03:10,920 --> 01:03:13,400 But here, I'll let you get out of here. 716 01:03:13,400 --> 01:03:20,400 But real quick, I'll just show you how the model would go about working if this was something 717 01:03:20,400 --> 01:03:23,040 that you so desired. 718 01:03:23,040 --> 01:03:34,600 So long story short is you get someone to review here. 719 01:03:34,600 --> 01:03:42,080 So I biked to a dispensary, picked up a lot of weed, da, da, da, da. 720 01:03:42,080 --> 01:03:49,240 And so you'd normally have to pay somebody to read through this and figure out how happy 721 01:03:49,240 --> 01:03:51,640 somebody was from this. 722 01:03:51,640 --> 01:04:00,880 And we can basically just let this natural language toolkit basically give us a ranking 723 01:04:00,880 --> 01:04:10,960 from minus one to one on how positive somebody's experience was. 724 01:04:10,960 --> 01:04:22,520 So this person, it looks like, had quite a positive experience. 725 01:04:22,520 --> 01:04:30,480 And this person may have had less of a positive experience. 726 01:04:30,480 --> 01:04:34,760 So I think this part itself could probably be tailored in. 727 01:04:34,760 --> 01:04:39,160 Like I said, I'm just kind of picking up this tool and running with it. 728 01:04:39,160 --> 01:04:53,520 But the idea is you could basically rank all of these reviews from minus one to what? 729 01:04:53,520 --> 01:04:58,320 So that way, we can put it in a continuous space. 730 01:04:58,320 --> 01:05:02,040 And we're not having to deal with zeros and ones. 731 01:05:02,040 --> 01:05:04,920 We have a nice continuum. 732 01:05:04,920 --> 01:05:10,300 Instead of just saying, oh, I was happy or I was sleepy, we actually can kind of get 733 01:05:10,300 --> 01:05:14,680 an idea of how happy you were. 734 01:05:14,680 --> 01:05:22,080 And notice it's heavily biased towards positive effects. 735 01:05:22,080 --> 01:05:27,100 And so this is kind of something I talked about where if you're going to be leaving 736 01:05:27,100 --> 01:05:33,480 a review, it's probably because you had a positive experience. 737 01:05:33,480 --> 01:05:35,600 Or maybe you had a really negative experience. 738 01:05:35,600 --> 01:05:44,760 But chances are, if you're going to Leafly and leaving a review, you probably have a 739 01:05:44,760 --> 01:05:47,920 positive opinion on this. 740 01:05:47,920 --> 01:05:51,400 So long story short, make of it what you will. 741 01:05:51,400 --> 01:06:01,760 But you can then regress this ranking on the log of all these compounds. 742 01:06:01,760 --> 01:06:05,720 So I know the data is not perfect. 743 01:06:05,720 --> 01:06:16,200 But if you had your own data, your R squared may be much higher than our meager, measly 744 01:06:16,200 --> 01:06:17,440 R squared. 745 01:06:17,440 --> 01:06:26,760 So yes, our data is, as Charles put it, kind of compiled from hell. 746 01:06:26,760 --> 01:06:34,120 Charles didn't say it like that, but I said it like that. 747 01:06:34,120 --> 01:06:35,840 That's my training as an economist. 748 01:06:35,840 --> 01:06:38,000 So judge me how you will. 749 01:06:38,000 --> 01:06:44,360 But long story short is make of it what you will. 750 01:06:44,360 --> 01:06:53,960 But the way I interpret this is, okay, someone may have a baseline happiness of around.45 751 01:06:53,960 --> 01:06:56,800 on the minus one to one scale. 752 01:06:56,800 --> 01:07:07,480 And then what this would basically say would be, for each 1% increase in total THC, you 753 01:07:07,480 --> 01:07:18,080 would move.06 distance towards one. 754 01:07:18,080 --> 01:07:27,280 So the idea is an increase of total THC would increase your utility. 755 01:07:27,280 --> 01:07:31,080 I'm not even going to read into the statistical significance of this. 756 01:07:31,080 --> 01:07:38,080 It looks like CBD may, but once again, the coefficient is really small. 757 01:07:38,080 --> 01:07:42,520 So CBD may increase your happiness. 758 01:07:42,520 --> 01:07:49,640 And then basically what I'm seeing is, once again, the coefficients on these are really 759 01:07:49,640 --> 01:07:50,640 low. 760 01:07:50,640 --> 01:08:02,200 But it could be that increasing your D-lemonine, beta-karyophylline, and humuline may actually 761 01:08:02,200 --> 01:08:08,640 decrease people's positive review. 762 01:08:08,640 --> 01:08:20,040 And then just real, real quick, this is the expected utility function just to estimate 763 01:08:20,040 --> 01:08:21,040 everything. 764 01:08:21,040 --> 01:08:26,400 But kind of once again, you're seeing the same thing. 765 01:08:26,400 --> 01:08:34,360 But then in this case, you're generally – if somebody does get sleepy, they may get slightly 766 01:08:34,360 --> 01:08:37,640 less utility from these compounds. 767 01:08:37,640 --> 01:08:39,720 Hey, Keegan? 768 01:08:39,720 --> 01:08:40,720 Yes? 769 01:08:40,720 --> 01:08:48,160 A really critical point that we deal with in our own discussions on this is, if you 770 01:08:48,160 --> 01:08:55,440 pick sleepy as one of your indicators, there's a large group of people that use cannabis 771 01:08:55,440 --> 01:08:57,120 to go to sleep. 772 01:08:57,120 --> 01:09:00,080 And so that is a desired outcome. 773 01:09:00,080 --> 01:09:05,440 You have to really query for indication for what the intent was. 774 01:09:05,440 --> 01:09:10,040 And I think you have to build that into the model a priori that I didn't hear you really 775 01:09:10,040 --> 01:09:15,640 – I don't think that's quite there yet, or maybe I missed it. 776 01:09:15,640 --> 01:09:22,280 But I think you need to put – you have to query and put the intent by respondent into 777 01:09:22,280 --> 01:09:23,280 this model. 778 01:09:23,280 --> 01:09:25,880 And you hit the nail on the head, John. 779 01:09:25,880 --> 01:09:28,200 So thank you for bringing this up. 780 01:09:28,200 --> 01:09:31,760 This is the exact extension of the model that's needed. 781 01:09:31,760 --> 01:09:36,480 So if you're interested, sign up for Saturday Morning Statistics because that's when we'll 782 01:09:36,480 --> 01:09:37,480 be doing it. 783 01:09:37,480 --> 01:09:46,240 But basically, it's a fairly trivial extension of this model to basically let the parameters 784 01:09:46,240 --> 01:09:54,720 alpha vary by – you could let them vary in a couple different ways, but potentially, 785 01:09:54,720 --> 01:10:00,440 we could just let them vary by individual. 786 01:10:00,440 --> 01:10:05,440 You could do types of individual, but if we've got enough data, we could basically let the 787 01:10:05,440 --> 01:10:10,640 parameters vary by – I want to say we could – I don't know if we're going to have 788 01:10:10,640 --> 01:10:11,640 enough data. 789 01:10:11,640 --> 01:10:19,840 We may have to – I was thinking that what you can do is basically you could restrict 790 01:10:19,840 --> 01:10:28,720 your sample to the people that leave a lot of reviews and then let the parameters vary 791 01:10:28,720 --> 01:10:35,360 by user, and then you could basically get user-specific effects. 792 01:10:35,360 --> 01:10:37,960 And so then you'd be exactly right. 793 01:10:37,960 --> 01:10:47,480 And so then some users may get a really good effect from THC and then some wouldn't. 794 01:10:47,480 --> 01:10:52,200 Some users may get a really good effect from CBD and others wouldn't. 795 01:10:52,200 --> 01:10:56,600 And then it could handle all the different states. 796 01:10:56,600 --> 01:11:03,880 So then you could let it – some users really like the sleepy effect, some not like the 797 01:11:03,880 --> 01:11:05,520 sleepy effect. 798 01:11:05,520 --> 01:11:13,280 And so I think it kind of captures everything, right, because it's basically saying it 799 01:11:13,280 --> 01:11:16,240 ties it nicely in with economic theory. 800 01:11:16,240 --> 01:11:21,260 You've got consumers trying to maximize their utility from cannabis. 801 01:11:21,260 --> 01:11:22,760 How do they do that? 802 01:11:22,760 --> 01:11:30,280 Well, it's actually from the compounds in the cannabis and it's conditional interacting 803 01:11:30,280 --> 01:11:37,320 with the effects that they experience. 804 01:11:37,320 --> 01:11:43,320 And it varies person by person at a biochemistry level. 805 01:11:43,320 --> 01:11:44,320 So I think – 806 01:11:44,320 --> 01:11:52,720 So it's a parameter that also – the response group type I think is a different parameter 807 01:11:52,720 --> 01:11:54,560 from the intent parameter. 808 01:11:54,560 --> 01:11:58,240 I think you need two of them at least. 809 01:11:58,240 --> 01:12:01,840 Basically alpha is going to get alpha 1i. 810 01:12:01,840 --> 01:12:02,840 Okay. 811 01:12:02,840 --> 01:12:03,840 Okay. 812 01:12:03,840 --> 01:12:13,360 So as I said, tune in to Saturday morning if you really want like the hardcore statistics 813 01:12:13,360 --> 01:12:14,360 on this. 814 01:12:14,360 --> 01:12:20,480 But essentially, as I said, I just kind of thought of this last night. 815 01:12:20,480 --> 01:12:25,080 So this still needs a lot of polishing. 816 01:12:25,080 --> 01:12:33,200 But the idea is – I mean if you can estimate expected utility from a cannabis product, 817 01:12:33,200 --> 01:12:37,480 of course the model is going to need a lot more refinement and you can be creative. 818 01:12:37,480 --> 01:12:42,840 And that's why I always say this is your chance to be creative in X is think about 819 01:12:42,840 --> 01:12:50,680 all the different variants that matter, all the different biochemistries that matter, 820 01:12:50,680 --> 01:12:54,800 all the different states of nature that matter. 821 01:12:54,800 --> 01:13:02,160 So you can be creative in X, but the idea is this is a fairly general model. 822 01:13:02,160 --> 01:13:07,680 And as I said, this could help anyone who's got a recommendation engine. 823 01:13:07,680 --> 01:13:13,400 It could help people thinking about what they're going to grow. 824 01:13:13,400 --> 01:13:19,440 And it could also just help scientists and consumers just kind of understand how does 825 01:13:19,440 --> 01:13:21,840 the brain work, right? 826 01:13:21,840 --> 01:13:24,600 Because how do people make decisions? 827 01:13:24,600 --> 01:13:27,780 How do compounds affect people? 828 01:13:27,780 --> 01:13:31,280 I think the jury's still out on a lot of this. 829 01:13:31,280 --> 01:13:37,320 And I think this model really is a big step in the right direction. 830 01:13:37,320 --> 01:13:45,520 So thank you all for your tough criticism or your tough remarks because no model's 831 01:13:45,520 --> 01:13:46,520 perfect. 832 01:13:46,520 --> 01:13:49,040 This model is far from perfect. 833 01:13:49,040 --> 01:13:52,080 The data is way far from perfect. 834 01:13:52,080 --> 01:13:56,760 So there's a lot of improvement all across the board that needs to be done. 835 01:13:56,760 --> 01:14:02,960 That's why I call upon you to help contribute because I think we're onto something big 836 01:14:02,960 --> 01:14:03,960 here. 837 01:14:03,960 --> 01:14:11,360 It's just as Charles and John and everyone had pointed out, there's so many intricacies 838 01:14:11,360 --> 01:14:17,920 and ways that you can misstep and ways that you should be doing this correct, like a proper 839 01:14:17,920 --> 01:14:18,960 data scientist. 840 01:14:18,960 --> 01:14:28,480 So I'm calling on all of you for your help because I think we could be onto essentially 841 01:14:28,480 --> 01:14:36,920 the best model of cannabis consumption that I've seen. 842 01:14:36,920 --> 01:14:41,360 So if any of you want to help contribute, then message me. 843 01:14:41,360 --> 01:14:42,520 Join the Slack channel. 844 01:14:42,520 --> 01:14:43,520 It's all hands on deck. 845 01:14:43,520 --> 01:14:45,960 But I'll leave that with you. 846 01:14:45,960 --> 01:14:52,640 And my two insights from this week were data is really just hiding everywhere. 847 01:14:52,640 --> 01:14:54,760 You just have to look for it. 848 01:14:54,760 --> 01:14:56,800 And then this synthetic data. 849 01:14:56,800 --> 01:14:59,080 You've got to be a little careful with it. 850 01:14:59,080 --> 01:15:03,800 But personally, I think it can be quite useful. 851 01:15:03,800 --> 01:15:08,240 So I know it went way, way, way, way, way over time. 852 01:15:08,240 --> 01:15:14,120 But I thought this material was hopefully worth your while. 853 01:15:14,120 --> 01:15:19,680 So any last thoughts, comments, questions before we finally let you get out of here 854 01:15:19,680 --> 01:15:24,320 and enjoy your day? 855 01:15:24,320 --> 01:15:32,280 I think defining those alphas is going to be an important exercise. 856 01:15:32,280 --> 01:15:34,560 We certainly talk about timing. 857 01:15:34,560 --> 01:15:36,320 We talk about dose. 858 01:15:36,320 --> 01:15:42,240 We talk about content as the three main parameters that we're kind of focused on. 859 01:15:42,240 --> 01:15:47,640 And what primarily your discussion today is, is on content. 860 01:15:47,640 --> 01:15:52,200 And we need to be able to get to the questions of timing and dose. 861 01:15:52,200 --> 01:15:59,600 So this data set may not have it for us, but others might, or those that we generate. 862 01:15:59,600 --> 01:16:00,600 Exactly. 863 01:16:00,600 --> 01:16:06,360 Basically, we'll be able to maybe assign random effects for the users. 864 01:16:06,360 --> 01:16:19,920 So as I said, it's not going to be too, too useful in interpreting biochemistry type effects. 865 01:16:19,920 --> 01:16:24,120 The framework's at least in place. 866 01:16:24,120 --> 01:16:31,840 So like you, if you get that data, then you're welcome to use this framework and actually 867 01:16:31,840 --> 01:16:34,840 be able to make those inferences. 868 01:16:34,840 --> 01:16:40,720 Well one thing I'm going to try, if we can connect and I can get your user sorted data, 869 01:16:40,720 --> 01:16:46,400 at least the way you presented it today, I think I can query to see if we can identify 870 01:16:46,400 --> 01:16:49,640 the response type of the reviewer. 871 01:16:49,640 --> 01:16:53,800 That's something that came out of, I'm thinking, watching what you did today. 872 01:16:53,800 --> 01:17:00,240 And I'll give that a try if I can get that data, if I can get that data set. 873 01:17:00,240 --> 01:17:01,640 Definitely. 874 01:17:01,640 --> 01:17:06,000 And let's all stay in touch and keep moving the ball forward. 875 01:17:06,000 --> 01:17:10,520 Because I hope, I think that we're advancing cannabis science. 876 01:17:10,520 --> 01:17:17,640 So let's keep our noses to the grindstone, have fun while we're doing it, and get out 877 01:17:17,640 --> 01:17:18,640 of here and enjoy your time. 878 01:17:18,640 --> 01:17:19,640 Get the right strain. 879 01:17:19,640 --> 01:17:49,560 Get the right strain everyone.