1 00:00:11,700 --> 00:00:16,440 In this section of the course we are going to look at how deep learning can be used for recommender 2 00:00:16,440 --> 00:00:17,920 systems. 3 00:00:17,940 --> 00:00:21,700 This lecture will discuss the theory behind how deep learning can be applied. 4 00:00:21,840 --> 00:00:25,330 And following that we will look at the code to start. 5 00:00:25,350 --> 00:00:29,070 I want to talk about recommender systems as a general concept. 6 00:00:29,250 --> 00:00:32,880 Where can we find recommender systems and what is their purpose. 7 00:00:32,880 --> 00:00:38,010 I think you'll be surprised to find out that recommender systems are one of the most applicable machine 8 00:00:38,010 --> 00:00:39,970 learning concepts ever. 9 00:00:39,990 --> 00:00:45,060 It's a type of machine learning that can be used by nearly any consumer facing business and you most 10 00:00:45,060 --> 00:00:53,440 certainly encounter recommender systems multiple times daily. 11 00:00:53,630 --> 00:00:55,100 Let's start with the first Web site. 12 00:00:55,100 --> 00:00:57,080 Everybody goes to when they go online. 13 00:00:57,110 --> 00:01:01,440 Google Google search is in fact a recommendation engine. 14 00:01:01,640 --> 00:01:07,070 Actually when you perform a search multiple recommendation engines are working to build the page that 15 00:01:07,070 --> 00:01:08,390 you see. 16 00:01:08,390 --> 00:01:13,940 Not only do you have the search results which are recommendations based on your query your geographical 17 00:01:13,940 --> 00:01:16,470 location and maybe even your search history. 18 00:01:16,670 --> 00:01:20,450 But don't forget that Google is also an advertising company. 19 00:01:20,450 --> 00:01:26,450 The online advertisements you see when you go to Google dot com are also recommendations based on your 20 00:01:26,450 --> 00:01:27,740 previous browsing history 21 00:01:32,900 --> 00:01:35,320 another great example is Amazon. 22 00:01:35,540 --> 00:01:40,790 When you search for a product on Amazon it's more likely that you'll see advertisements for similar 23 00:01:40,790 --> 00:01:42,590 products in the future. 24 00:01:42,590 --> 00:01:46,130 You might also see recommendations for related products. 25 00:01:46,130 --> 00:01:51,920 For example if you're buying an iPod you might also be in the market for new headphones a case and a 26 00:01:51,920 --> 00:01:52,730 screen protector 27 00:01:57,840 --> 00:02:04,560 another great but seriously unfortunate example is the news news sites employ recommendation systems 28 00:02:04,620 --> 00:02:09,720 and they even do a b testing on their article headlines to improve click rate. 29 00:02:09,720 --> 00:02:15,570 This is very unfortunate because this means that news has become more biased and more polarizing over 30 00:02:15,570 --> 00:02:16,590 time. 31 00:02:16,590 --> 00:02:18,520 People don't care about boring news. 32 00:02:18,570 --> 00:02:24,000 They want to read something exciting and news companies are more than happy to give that to you in place 33 00:02:24,000 --> 00:02:26,490 of real factual and unbiased news 34 00:02:31,570 --> 00:02:35,670 there are really too many examples of recommendation systems to mention. 35 00:02:35,670 --> 00:02:41,910 Here are a few more examples to help you internalize what exactly it means to recommend we have Netflix 36 00:02:41,910 --> 00:02:44,600 which recommends movies and TV shows. 37 00:02:44,640 --> 00:02:51,090 We have YouTube which recommends videos we have read it which recommends a mix of user generated content 38 00:02:51,330 --> 00:02:54,400 and third party content posted by users. 39 00:02:54,450 --> 00:03:00,330 We have Facebook which sorts your news feed based on what will make you stay on it the longest. 40 00:03:00,330 --> 00:03:05,490 Closely related to that we have Instagram which is owned by Facebook and it does largely the same thing 41 00:03:06,220 --> 00:03:10,820 you're shown a feed of Instagram posts that are recommended in order to keep you engaged. 42 00:03:15,800 --> 00:03:19,830 In this lecture we'll be focusing on a specific kind of recommender. 43 00:03:20,000 --> 00:03:25,580 This recommender works on data that comes in the form of triples the three items that form a sample 44 00:03:25,580 --> 00:03:29,680 or the user the item and the rating that the user gave that item. 45 00:03:34,920 --> 00:03:38,400 As an example suppose you're creating a movie recommender. 46 00:03:38,480 --> 00:03:41,610 Here are what a few rows of your data set might look like. 47 00:03:41,790 --> 00:03:42,440 Alice rates. 48 00:03:42,450 --> 00:03:44,800 Avatar 5 Bob rates. 49 00:03:44,820 --> 00:03:49,490 Star Wars of 4.5 and Carol rates the godfather of 4. 50 00:03:49,560 --> 00:03:51,920 So all of our samples will look like this. 51 00:03:52,080 --> 00:03:53,760 The first item is the user. 52 00:03:53,760 --> 00:03:55,560 The second item is the movie. 53 00:03:55,560 --> 00:03:57,120 And the third item is the rating 54 00:04:02,260 --> 00:04:06,970 one important feature of this type of dataset is that it must be incomplete. 55 00:04:06,970 --> 00:04:08,200 Why. 56 00:04:08,200 --> 00:04:08,980 Well think about it. 57 00:04:09,400 --> 00:04:15,220 If all users have already watched and rated all existing movies then what would there be to recommend 58 00:04:15,910 --> 00:04:18,820 if I've already watched and rated every movie in existence. 59 00:04:18,820 --> 00:04:21,280 There is no need to recommend anything to me. 60 00:04:21,280 --> 00:04:24,280 You can't recommend a movie to me because I've already seen it. 61 00:04:25,330 --> 00:04:27,070 Luckily for any real data set. 62 00:04:27,100 --> 00:04:28,870 This won't be the case. 63 00:04:28,870 --> 00:04:34,660 Imagine you have 1 million users and one hundred thousand movies which is not unrealistic. 64 00:04:34,660 --> 00:04:39,220 You can probably count the number of movies you've seen in your life and it's probably a lot less than 65 00:04:39,220 --> 00:04:40,770 one hundred thousand. 66 00:04:40,930 --> 00:04:44,800 The number of movies you've actually bother to rate is probably even smaller. 67 00:04:49,940 --> 00:04:56,050 The question now is once we have such a dataset how does it help us make recommendations. 68 00:04:56,060 --> 00:05:03,350 Let's take a very simple example consider Bob Bob R rated every movie in the Star Wars series of 5 Bob 69 00:05:03,380 --> 00:05:06,270 also rated a few of the Star Trek movies of five. 70 00:05:06,500 --> 00:05:09,510 And Bob also rated Avatar five. 71 00:05:09,530 --> 00:05:13,310 Common sense tells us that Bob is probably a fan of sci fi. 72 00:05:13,490 --> 00:05:17,720 Bob likes movies that have space travel aliens and so forth. 73 00:05:17,780 --> 00:05:22,330 It is probable then that Bob would like movies that have the same features. 74 00:05:22,550 --> 00:05:28,190 A pretty obvious recommendation for Bob would be a star trek movie or a TV show that he hasn't seen 75 00:05:28,190 --> 00:05:28,430 yet. 76 00:05:33,520 --> 00:05:35,990 Let's try to generalize this concept. 77 00:05:36,400 --> 00:05:42,130 As you know we're given a dataset that consists of triples containing user's items and the user's ratings 78 00:05:42,130 --> 00:05:43,760 for those items. 79 00:05:43,780 --> 00:05:45,960 Suppose we can fit a model to this data. 80 00:05:46,060 --> 00:05:49,510 Let's just call it F F takes in two arguments. 81 00:05:49,510 --> 00:05:54,540 A user you and a movie M and it outputs a predicted rating. 82 00:05:54,580 --> 00:05:57,730 There are two things that we would like this function to do. 83 00:05:57,730 --> 00:06:03,520 Number one obviously if the user movie and rating appear in our dataset then we would like the predicted 84 00:06:03,520 --> 00:06:07,510 rating to be close to the real rating and number two. 85 00:06:07,510 --> 00:06:13,150 We would like this function to be able to output a predicted rating even if the rating did not appear 86 00:06:13,150 --> 00:06:19,850 in our training center luckily a neuron that work is a perfect model for this use case. 87 00:06:19,850 --> 00:06:25,160 As you know a neuron that work is a function approximate how it fits to the data set so that when it 88 00:06:25,160 --> 00:06:31,130 makes predictions those predictions are close to the training targets but not only that it can also 89 00:06:31,190 --> 00:06:39,870 make new predictions for data which did not appear in the training set. 90 00:06:39,910 --> 00:06:43,090 So how might we use such a model to make recommendations. 91 00:06:43,690 --> 00:06:45,510 Well here's one strategy. 92 00:06:45,730 --> 00:06:52,000 Once we have a model that can predict ratings this becomes easy for a given user predicts a rating for 93 00:06:52,000 --> 00:06:58,930 every movie that they have not yet seen then sort these movies by predicted rating the movies you want 94 00:06:58,930 --> 00:07:03,910 to recommend are simply the movies with the highest predicted rating for this user. 95 00:07:03,970 --> 00:07:04,630 Pretty simple 96 00:07:09,710 --> 00:07:13,760 The next question is how can we go about building such a model. 97 00:07:13,850 --> 00:07:16,070 We know that we need a neuron network of some sort. 98 00:07:16,790 --> 00:07:23,690 But here's a problem both users and movies are categorical objects neural networks are basically a series 99 00:07:23,690 --> 00:07:25,660 of matrix multiplications. 100 00:07:25,910 --> 00:07:30,200 As you know we can't multiply a categorical object by a number. 101 00:07:30,200 --> 00:07:31,850 What is the Star Wars times five. 102 00:07:31,850 --> 00:07:33,110 This does not make sense 103 00:07:38,230 --> 00:07:41,160 luckily we have encountered this problem before. 104 00:07:41,350 --> 00:07:47,830 Strangely natural language processing is the field we draw inspiration from in order to build deep recommended 105 00:07:47,830 --> 00:07:49,390 systems. 106 00:07:49,390 --> 00:07:53,310 As you know in MLP the main input is words. 107 00:07:53,410 --> 00:07:54,550 Words are categorical. 108 00:07:54,550 --> 00:07:56,560 They cannot be multiplied by numbers. 109 00:07:57,040 --> 00:07:58,690 So what do we do. 110 00:07:58,750 --> 00:08:05,350 We use an embedding Recall that an embedding is a mapping from each category to a feature vector which 111 00:08:05,350 --> 00:08:08,290 is essentially a list of numbers that can be multiplied 112 00:08:13,450 --> 00:08:17,090 here's how we might use a neuron that work for recommender systems. 113 00:08:17,140 --> 00:08:23,620 Let's say we have an input user Bob and an input movie star was the first thing we do is map both the 114 00:08:23,620 --> 00:08:27,920 user and the movie to their respective embedding vectors. 115 00:08:28,000 --> 00:08:32,110 So I have a vector for Bob and I have a vector for Star Wars. 116 00:08:32,230 --> 00:08:37,680 Now usually these embedding vectors are the same size but this need not be the case. 117 00:08:37,760 --> 00:08:42,160 So now both Bob and Star Wars are represented by feature vectors. 118 00:08:42,200 --> 00:08:48,090 The next thing we do is concatenate these two feature vectors into a single feature vector. 119 00:08:48,140 --> 00:08:51,340 Once we have a feature vector the rest is obvious. 120 00:08:51,440 --> 00:08:57,720 We just do the same thing we've always done pass this through a neural network since this is a feature 121 00:08:57,720 --> 00:09:01,320 vector and not some special object like an image or a sequence. 122 00:09:01,320 --> 00:09:06,900 We can use simple antennas rather than CNN or Arnold's for recommender systems. 123 00:09:06,900 --> 00:09:12,900 The final output layer will be a linear regression with no activation function since predicting ratings 124 00:09:12,900 --> 00:09:15,900 is a regression task and not a classification task 125 00:09:21,110 --> 00:09:22,580 as a point of interest. 126 00:09:22,580 --> 00:09:28,260 I said earlier that we can use an techniques to inspire models for recommender system. 127 00:09:28,340 --> 00:09:34,020 In fact the opposite is true as well recommender techniques have been used in an LP. 128 00:09:34,070 --> 00:09:40,040 My favorite example of this is a technique called Matrix Factorization Matrix Factorization which we 129 00:09:40,040 --> 00:09:44,960 haven't talked about is an algorithm that has been known for quite a while in the field of recommender 130 00:09:44,960 --> 00:09:49,370 systems around the time word embedding were becoming popular. 131 00:09:49,370 --> 00:09:54,980 Specifically due to the algorithm known as works of EQ another similar algorithm came out called a glove 132 00:09:55,760 --> 00:09:58,360 Glove essentially solves the same task as where to vac. 133 00:09:58,430 --> 00:10:03,710 It's an algorithm for finding word meetings using these words I'm betting you can do all sorts of fun 134 00:10:03,710 --> 00:10:06,950 analogies such as King minus man equals Queen minus woman. 135 00:10:07,790 --> 00:10:12,080 Well it turns out that glove is nothing more than Matrix Factorization. 136 00:10:12,170 --> 00:10:17,540 Just an old technique from recommender systems so it's always good to study techniques from different 137 00:10:17,540 --> 00:10:21,740 fields because you never know when or what you learn that can be applied to your own work.