1 00:00:11,120 --> 00:00:15,290 So in this lecture, you'll be given a few exercises to practice using as we do. 2 00:00:16,309 --> 00:00:20,990 Previously, you weren't that served is helpful for reducing redundant dimensions. 3 00:00:21,500 --> 00:00:25,910 So for instance, run and sprint might be combined into a single dimension. 4 00:00:26,510 --> 00:00:32,570 And therefore, if you used a search engine, a user querying for the word run might also get back documents 5 00:00:32,570 --> 00:00:34,010 that contain the word sprint. 6 00:00:35,210 --> 00:00:41,390 So your first exercise is this We built a recommender system using TFI Taf earlier in the course. 7 00:00:41,930 --> 00:00:48,650 It works by doing a nearest neighbor search on a TFI to effectors directly, but we know that TFI effectors 8 00:00:48,650 --> 00:00:53,180 can have large dimensionality and further might contain redundant terms. 9 00:00:53,840 --> 00:00:58,910 In our case, when we were looking at movie reviews, one example might be action and adventure. 10 00:00:59,180 --> 00:01:01,910 Typically, these both occur together for the same movie. 11 00:01:02,840 --> 00:01:06,170 In other words, they are redundant for this exercise. 12 00:01:06,170 --> 00:01:11,900 You should modify that script to use latent semantic analysis on the TF RDF vectors. 13 00:01:12,500 --> 00:01:17,630 So now your nearest neighbor search should be done on the Z vectors or the latent semantic analysis 14 00:01:17,630 --> 00:01:21,440 vectors instead of the X vectors or the two Jovita effectors. 15 00:01:26,050 --> 00:01:31,720 The next exercise will involve revisiting tax classification, which could be anything from classifying 16 00:01:31,720 --> 00:01:35,410 the news to spam detection to sentiment analysis. 17 00:01:35,950 --> 00:01:38,650 Note that we did these examples earlier in the course. 18 00:01:39,340 --> 00:01:45,460 So your second task will be to take one of these notebooks and modify them to use lean semantic analysis 19 00:01:45,790 --> 00:01:47,290 as a pre processing step. 20 00:01:47,950 --> 00:01:53,350 That is, instead of training your model on X train and Y train and then testing on X tested and Y test, 21 00:01:53,860 --> 00:01:56,470 you were going to transform the X matrix first. 22 00:01:57,190 --> 00:02:02,950 As such, your classifier will be trained on Z Train and Y Train and tested on Z Test and Y Test. 23 00:02:04,470 --> 00:02:10,020 It would be useful to do a comparison between the performance of the model with and without this step 24 00:02:10,050 --> 00:02:11,910 using different numbers of features. 25 00:02:13,380 --> 00:02:19,620 As you recall, the Count Vector Riser and TFI Taf can be told to keep only the most frequent words. 26 00:02:20,160 --> 00:02:26,700 So for a fair comparison, if you're Z Matrix has Dimension K, then also keep K words in your TFT f 27 00:02:26,700 --> 00:02:30,960 matrix and then plot your model's performance for different values of K. 28 00:02:31,590 --> 00:02:38,070 So you'll have K on the horizontal axis and model performance, such as accuracy on the vertical axis. 29 00:02:42,800 --> 00:02:50,000 Your next exercise is to revisit topic modeling, as you recall, topic modeling gives us back to matrices, 30 00:02:50,390 --> 00:02:54,350 one with documents by topics and another with topics by words. 31 00:02:54,980 --> 00:03:02,150 For this exercise, you're going to use lean semantic analysis instead of LDA or an AMF now to convince 32 00:03:02,150 --> 00:03:05,060 you that speed gives us back the same matrices. 33 00:03:05,390 --> 00:03:08,510 I have to discuss a little bit about how SVT works. 34 00:03:09,260 --> 00:03:16,790 Essentially, speed splits up your input matrix into three matrices you multiplied by s multiplied by 35 00:03:16,790 --> 00:03:17,870 v transpose. 36 00:03:18,500 --> 00:03:23,090 Now, you don't really have to know what these matrices mean or why they're relevant, but you do have 37 00:03:23,090 --> 00:03:24,080 to know their shapes. 38 00:03:24,740 --> 00:03:28,730 Suppose that are x matrix has the shape and by D then are you? 39 00:03:28,730 --> 00:03:36,500 Matrix will have the shape and by K s matrix will be a diagonal matrix with the shape K by K and R v 40 00:03:36,500 --> 00:03:37,580 transpose matrix. 41 00:03:37,790 --> 00:03:39,470 We'll have the shape K by D. 42 00:03:40,160 --> 00:03:46,820 Because of this, v not transposed has the shape d by K, but we typically work with the K by D version 43 00:03:46,820 --> 00:03:47,240 anyway. 44 00:03:48,050 --> 00:03:49,850 But this should remind you a lot of enemies. 45 00:03:49,890 --> 00:03:57,650 F enemy is simply another way to do a matrix decomposition where we say X is equal to W times h. 46 00:03:58,220 --> 00:04:02,900 In this case, W has the shape and by K and H has the shape K by D. 47 00:04:03,710 --> 00:04:09,380 So the V transpose matrix in CVD is analogous to the H matrix in an MF. 48 00:04:09,650 --> 00:04:11,630 They are both topics by words. 49 00:04:12,200 --> 00:04:18,680 In fact, inside you learn the V transpose matrix can be accessed by calling ASCVD components with an 50 00:04:18,680 --> 00:04:21,110 underscore, just like an MF and LDA. 51 00:04:21,980 --> 00:04:27,680 Furthermore, when you call Model Di Transform, you get back your Z matrix, which is of shape and 52 00:04:27,680 --> 00:04:30,620 by K or, in other words, documents by topics. 53 00:04:31,160 --> 00:04:37,040 So exactly the same as an MF and LDA as a side note for SVD. 54 00:04:37,580 --> 00:04:42,770 This is equivalent to U times s, so in fact, no code changes are even required. 55 00:04:43,100 --> 00:04:47,510 You simply need to plug in CVD wherever you see an MF for LDA. 56 00:04:48,590 --> 00:04:53,840 Now there's one slight difference between us and the other methods which you will see for yourself when 57 00:04:53,840 --> 00:04:55,160 you do this exercise. 58 00:04:55,820 --> 00:05:01,250 Unlike LDA in an MF, there are no constraints on the values that we get back from CVD. 59 00:05:01,490 --> 00:05:07,160 They can be either positive or negative, so you might find documents that contain negative amounts 60 00:05:07,160 --> 00:05:07,970 of topics. 61 00:05:08,570 --> 00:05:14,990 So that's why LDA and an MF would be conceptually better for a topic modeling compared to SVT. 62 00:05:19,830 --> 00:05:26,490 So the final exercise is perhaps the most difficult, as you saw earlier in this course, latent semantic 63 00:05:26,490 --> 00:05:31,770 analysis can also be applied to text summarization exactly how it does. 64 00:05:31,770 --> 00:05:35,130 This is a bit non-trivial for this exercise. 65 00:05:35,130 --> 00:05:40,470 I've attached two papers to extra reading Detox Tea, which explains how these methods work. 66 00:05:41,070 --> 00:05:42,690 The first paper is the simplest. 67 00:05:43,260 --> 00:05:47,760 The second paper references the first and tries to show that their method is better. 68 00:05:48,450 --> 00:05:52,440 So the exercise is to read these two papers and implement the LSA. 69 00:05:52,440 --> 00:05:55,470 Summarize you yourself using both of these methods.