1 00:00:11,060 --> 00:00:16,730 So in this lecture, we'll be looking at the notebook to demonstrate an AMF note that the format of 2 00:00:16,730 --> 00:00:19,220 this notebook will be very similar to LDA. 3 00:00:19,250 --> 00:00:25,580 So we'll go through most of it fairly quickly as before we'll begin by downloading the BBC news dataset. 4 00:00:34,530 --> 00:00:36,480 The next step is to do our imports. 5 00:00:37,080 --> 00:00:40,320 Note that this time we'll be using animal with TFI Taf. 6 00:00:41,280 --> 00:00:44,160 So why is it OK to use TFI off with enemies? 7 00:00:44,910 --> 00:00:50,240 As you recall, the reason we use simple counts with LDA is because that it corresponds with the multi 8 00:00:50,580 --> 00:00:53,460 distribution, which is part of the LDA model. 9 00:00:59,090 --> 00:01:01,130 The next step is to download or stop words. 10 00:01:06,720 --> 00:01:09,930 The next step is to convert our list of stop words into a set. 11 00:01:14,780 --> 00:01:17,150 The next step is to add additional stop words. 12 00:01:22,450 --> 00:01:25,810 The next step is to load in our data using PDF that reads Feed. 13 00:01:30,340 --> 00:01:33,250 The next step is to remind ourselves what this data looks like. 14 00:01:40,030 --> 00:01:45,430 The next step is to create our TF IDF factories are objects passing in the Stoppard's we define above. 15 00:01:50,260 --> 00:01:54,040 The next step is to transform our text into a 240 of matrix. 16 00:01:59,630 --> 00:02:06,320 The next step is to create our nymph object again, we'll choose 10 components, note that for the loss, 17 00:02:06,320 --> 00:02:12,440 we use the kale divergence, which when you do corresponds with probabilistic latent semantic analysis. 18 00:02:13,220 --> 00:02:18,830 Another way of looking at this is the paper on probabilistic lean semantic analysis really just describes 19 00:02:18,830 --> 00:02:22,820 non-negative matrix factorization with kale divergence as the laws. 20 00:02:23,840 --> 00:02:29,270 Note that when you use this loss, it's required that you specify the solver as AMU, which stands for 21 00:02:29,270 --> 00:02:30,650 multiplicative update. 22 00:02:31,880 --> 00:02:36,740 Note that it's also possible to regularize the weights, but I've commented this out since it's not 23 00:02:36,740 --> 00:02:37,250 needed. 24 00:02:38,900 --> 00:02:43,580 Finally, note that I've set the random state to zero so that we obtain a consistent results. 25 00:02:49,010 --> 00:02:50,570 The next step is to fit our model. 26 00:03:00,530 --> 00:03:03,710 The next step is due to find the plot, top words function once again. 27 00:03:09,460 --> 00:03:12,550 The next step is to call the plot top where its function once again. 28 00:03:20,520 --> 00:03:24,120 OK, so let's have a look at these topics to make sure that they make sense. 29 00:03:25,350 --> 00:03:31,260 So the top words for the first topic are people, UK, mobile, U.S. music and so forth. 30 00:03:31,740 --> 00:03:35,520 So probably mobile technology is a good way to summarize this topic. 31 00:03:36,780 --> 00:03:41,250 The top words for the second topic are Labour Election, The Blair Party and so forth. 32 00:03:41,760 --> 00:03:43,620 So clearly, this is about politics. 33 00:03:45,580 --> 00:03:51,130 Four topic three, we have England win, Wales, Ireland injury, rugby coach and so forth. 34 00:03:51,670 --> 00:03:54,430 Clearly, this is about sports and specifically rugby. 35 00:03:56,160 --> 00:03:59,670 For topic for we have film a best awards and so forth. 36 00:03:59,790 --> 00:04:05,760 So this is entertainment for Topic five, we have growth economy bank and so forth. 37 00:04:06,000 --> 00:04:07,320 So this is economics. 38 00:04:07,950 --> 00:04:11,010 And again, check out all the other topics as an exercise. 39 00:04:16,880 --> 00:04:21,470 The next step is to transform our data to get back the documents by topics matrix. 40 00:04:28,250 --> 00:04:31,790 The next step is to plot the topics for a randomly chosen input sample. 41 00:04:38,770 --> 00:04:44,770 As you can see, the true label is sport in the strongest topic is topic three, by the way. 42 00:04:44,800 --> 00:04:48,190 Note that unlike LDA, these values do not sum to one. 43 00:04:49,840 --> 00:04:53,680 So let's scroll back up and remind ourselves what Typekit three represents. 44 00:05:00,150 --> 00:05:03,930 OK, so topic three is related to sports and specifically rugby. 45 00:05:09,520 --> 00:05:13,000 So let's prints out this article to make sure that our topic makes sense. 46 00:05:19,350 --> 00:05:22,350 So the article is Chavez said, to lose fitness bit. 47 00:05:23,010 --> 00:05:27,840 So these results make perfect sense, since Chavez is in fact a former rugby player.