1 00:00:11,100 --> 00:00:16,420 So in this video, I'm going to introduce you to the Human Activity Recognition, a data set. 2 00:00:17,100 --> 00:00:19,900 This is a multiclass classification problem. 3 00:00:20,700 --> 00:00:23,230 The input is a multivariate time series. 4 00:00:23,730 --> 00:00:27,530 So this is a new data format that we haven't yet dealt with in this course. 5 00:00:28,050 --> 00:00:33,000 This lecture will discuss some basic facts about the data set so that you have some sense of the problem 6 00:00:33,000 --> 00:00:34,230 we are trying to solve. 7 00:00:34,830 --> 00:00:39,970 In fact, you may even want to implement the code yourself before looking at the course code. 8 00:00:40,950 --> 00:00:46,200 Note that the data can be found either at the UCI Machine Learning Repository or Kagle. 9 00:00:47,100 --> 00:00:51,360 So I've linked you both of these in the upcoming notebook in case you want to check them out. 10 00:00:56,080 --> 00:01:02,290 OK, so the basic setup for this problem goes like this, this data comes from real experiments on real 11 00:01:02,290 --> 00:01:07,960 people who performs a variety of activities while wearing a smartphone attached to their waist. 12 00:01:08,500 --> 00:01:14,680 Specifically, they perform to six different activities, walking, walking upstairs, walking down 13 00:01:14,680 --> 00:01:17,290 stairs, sitting, standing and laying down. 14 00:01:17,890 --> 00:01:21,850 OK, so those are the six classes that we are going to try and predict. 15 00:01:22,420 --> 00:01:29,170 The input data was collected from the smartphone sensors, specifically the accelerometer and the gyroscope. 16 00:01:33,890 --> 00:01:37,800 So this gives us a linear acceleration and angular velocity. 17 00:01:38,450 --> 00:01:44,180 Note that because physical space has three dimensions, each of these sensors gives us three separate 18 00:01:44,180 --> 00:01:44,870 channels. 19 00:01:45,230 --> 00:01:52,400 So, for example, the acceleration time series would actually be three time series activity, a YFC 20 00:01:52,400 --> 00:01:53,510 and AZT. 21 00:01:54,770 --> 00:01:59,400 Now, one curious fact about this data set is that there are two kinds of acceleration. 22 00:02:00,110 --> 00:02:04,040 One, they call total acceleration and one they call body acceleration. 23 00:02:05,240 --> 00:02:11,570 According to the README, the body acceleration was computed by subtracting gravity from the total acceleration. 24 00:02:12,860 --> 00:02:19,250 So in total, we actually end up having a time series with nine components, three for a total acceleration, 25 00:02:19,430 --> 00:02:23,060 three for body acceleration and three for angular velocity. 26 00:02:27,650 --> 00:02:33,470 OK, so here are some more details about the Time series, each time series was recorded at 50 hertz, 27 00:02:33,770 --> 00:02:40,620 meaning 50 measurements per second, the total duration of each sample was two point five six seconds. 28 00:02:41,090 --> 00:02:45,050 This means that we have one hundred twenty eight measurements for each time series. 29 00:02:45,740 --> 00:02:51,230 That is to say, the length of each time series corresponding to some activity is one twenty eight. 30 00:02:55,810 --> 00:03:01,150 So the Time series we will see in the data set is not the right time series, but the pre-process Time 31 00:03:01,150 --> 00:03:07,560 series, the researchers performed operations such as noise removal, Lopez filtering and scaling. 32 00:03:08,440 --> 00:03:10,940 You can think of Lopez filtering like smoothing. 33 00:03:11,560 --> 00:03:14,050 So basically it removes any fast movements. 34 00:03:14,050 --> 00:03:20,080 In the TIME series, scaling was done so that the main value is minus one and the max value is plus 35 00:03:20,080 --> 00:03:20,470 one. 36 00:03:21,310 --> 00:03:24,850 Because of this, we won't bother to scale the Time series ourselves. 37 00:03:29,500 --> 00:03:35,500 OK, so, again, one important skill you should have is being able to visualize what this Time series 38 00:03:35,500 --> 00:03:36,250 looks like. 39 00:03:36,760 --> 00:03:40,630 The input data, which is a multivariate time series, has the shape. 40 00:03:40,630 --> 00:03:48,250 And by TBD in this case, we know that T is equal to one twenty eight and D is equal to nine and is 41 00:03:48,250 --> 00:03:51,250 somewhere in the thousands for the targets. 42 00:03:51,430 --> 00:03:57,550 We might have a one dimensional array of length then containing the integers zero up to five representing 43 00:03:57,550 --> 00:03:58,690 the six classes. 44 00:03:59,860 --> 00:04:05,470 So knowing how the data should be formatted will be very useful, since as you'll soon see the way the 45 00:04:05,470 --> 00:04:07,470 data is organized is pretty messy. 46 00:04:07,930 --> 00:04:12,910 Everything is in multiple files, so it takes some effort to keep things organized in your mind. 47 00:04:17,370 --> 00:04:22,050 The final topic of this lecture is to consider the question, why should we bother to work with Time 48 00:04:22,050 --> 00:04:22,810 series at all? 49 00:04:23,550 --> 00:04:28,260 As you recall, one important aspect of feature engineering is domain knowledge. 50 00:04:29,040 --> 00:04:35,400 That is, you use your expertise in some domain to engineer useful features for machine learning models. 51 00:04:36,060 --> 00:04:38,280 Well, Time series is one such domain. 52 00:04:38,910 --> 00:04:44,190 In fact, it's not necessary to work with Raw Time series as we have been during this course. 53 00:04:44,610 --> 00:04:50,660 It's also possible to simply compute features from the TIME series and then use tabular machine learning. 54 00:04:51,600 --> 00:04:57,180 In this way, you can treat a time series data set like any other tabular data set you might use. 55 00:04:57,630 --> 00:05:00,900 So I encourage you to check the read me for a full set of features. 56 00:05:01,110 --> 00:05:02,480 But here are some highlights. 57 00:05:03,150 --> 00:05:06,270 We all know about the mean min max and standard deviation. 58 00:05:06,510 --> 00:05:08,570 OK, so these are all features. 59 00:05:09,090 --> 00:05:13,590 We also have Skewness and keratosis, which are more statistical features. 60 00:05:14,640 --> 00:05:19,740 We also have some frequency based features which can be found after taking the Fourier transform. 61 00:05:20,370 --> 00:05:22,350 OK, so lots of interesting stuff. 62 00:05:23,160 --> 00:05:27,460 In fact, you might want to have a look at these in case they are useful in your own work. 63 00:05:28,530 --> 00:05:33,960 What we'll do later in this section is compare models that only use these features with models that 64 00:05:33,960 --> 00:05:35,570 use the original time series. 65 00:05:36,030 --> 00:05:40,890 In addition, using what we know about how to build neural networks and sensor flow, we'll build a 66 00:05:40,890 --> 00:05:44,940 hybrid model that makes use of both the Time series and their features. 67 00:05:45,270 --> 00:05:48,690 So it'll be interesting to see which kind of approach works best.