en_Collaborative Filtering.srt

0
00:00:00,729 --> 00:00:05,200
Hello, and welcome! In this video, we’ll be covering a recommender

1
00:00:05,200 --> 00:00:10,839
system technique called, Collaborative filtering. So let’s get started.

2
00:00:10,839 --> 00:00:15,629
Collaborative filtering is based on the fact that relationships exist between products

3
00:00:15,629 --> 00:00:21,039
and people’s interests. Many recommendation systems use Collaborative

4
00:00:21,039 --> 00:00:26,660
filtering to find these relationships and to give an accurate recommendation of a product

5
00:00:26,660 --> 00:00:31,050
that the user might like or be interested in.

6
00:00:31,050 --> 00:00:37,510
Collaborative filtering has basically two approaches: User-based and Item-based.

7
00:00:37,510 --> 00:00:43,329
User-based collaborative filtering is based on the user’s similarity or neighborhood.

8
00:00:43,329 --> 00:00:47,930
Item-based collaborative filtering is based on similarity among items.

9
00:00:47,930 --> 00:00:53,039
Let’s first look at the intuition behind the “user-based” approach.

10
00:00:53,039 --> 00:00:58,230
In user-based collaborative filtering, we have an active user for whom the recommendation

11
00:00:58,230 --> 00:01:02,460
is aimed. The collaborative filtering engine, first

12
00:01:02,460 --> 00:01:07,841
looks for users who are similar, that is, users who share the active user’s rating

13
00:01:07,841 --> 00:01:12,270
patterns. Collaborative filtering bases this similarity

14
00:01:12,270 --> 00:01:18,770
on things like history, preference, and choices that users make when buying, watching, or

15
00:01:18,770 --> 00:01:23,350
enjoying something. For example, movies that similar users have

16
00:01:23,350 --> 00:01:27,660
rated highly. Then, it uses the ratings from these similar

17
00:01:27,660 --> 00:01:34,250
users to predict the possible ratings by the active user for a movie that she had not previously

18
00:01:34,250 --> 00:01:39,120
watched. For instance, if 2 users are similar or are

19
00:01:39,120 --> 00:01:45,360
neighbors, in terms of their interest in movies, we can recommend a movie to the active user

20
00:01:45,360 --> 00:01:50,830
that her neighbor has already seen. Now, let’s dive into the algorithm to see

21
00:01:50,830 --> 00:01:53,830
how all of this works.

22
00:01:53,830 --> 00:02:00,580
Assume that we have a simple user-item matrix, which shows the ratings of 4 users for 5 different

23
00:02:00,580 --> 00:02:04,580
movies. Let’s also assume that our active user has

24
00:02:04,580 --> 00:02:11,549
watched and rated 3 out of these 5 movies. Let’s find out which of the two movies that

25
00:02:11,549 --> 00:02:16,239
our active user hasn’t watched, should be recommended to her.

26
00:02:16,239 --> 00:02:21,870
The first step is to discover how similar the active user is to the other users.

27
00:02:21,870 --> 00:02:25,849
How do we do this? Well, this can be done through several different

28
00:02:25,849 --> 00:02:32,159
statistical and vectorial techniques such as distance or similarity measurements, including

29
00:02:32,159 --> 00:02:38,640
Euclidean Distance, Pearson Correlation, Cosine Similarity, and so on.

30
00:02:38,640 --> 00:02:44,420
To calculate the level of similarity between 2 users, we use the 3 movies that both the

31
00:02:44,420 --> 00:02:49,799
users have rated in the past. Regardless of what we use for similarity measurement,

32
00:02:49,799 --> 00:02:59,969
let’s say, for example, the similarity, could be 0.7, 0.9, and 0.4 between the active

33
00:02:59,969 --> 00:03:05,439
user and other users. These numbers represent similarity weights,

34
00:03:05,439 --> 00:03:11,569
or proximity of the active user to other users in the dataset.

35
00:03:11,569 --> 00:03:15,730
The next step is to create a weighted rating matrix.

36
00:03:15,730 --> 00:03:21,120
We just calculated the similarity of users to our active user in the previous slide.

37
00:03:21,120 --> 00:03:27,730
Now we can use it to calculate the possible opinion of the active user about our 2 target

38
00:03:27,730 --> 00:03:31,290
movies. This is achieved by multiplying the similarity

39
00:03:31,290 --> 00:03:37,810
weights to the user ratings. It results in a weighted ratings matrix, which

40
00:03:37,810 --> 00:03:45,079
represents the user’s neighbour’s opinion about our 2 candidate movies for recommendation.

41
00:03:45,079 --> 00:03:50,609
In fact, it incorporates the behaviour of other users and gives more weight to the ratings

42
00:03:50,609 --> 00:03:55,180
of those users who are more similar to the active user.

43
00:03:55,180 --> 00:04:01,430
Now we can generate the recommendation matrix by aggregating all of the weighted rates.

44
00:04:01,430 --> 00:04:08,290
However, as 3 users rated the first potential movie, and 2 users rated the second movie,

45
00:04:08,290 --> 00:04:14,959
we have to normalize the weighted rating values. We do this by dividing it by the sum of the

46
00:04:14,959 --> 00:04:20,350
similarity index for users. The result is the potential rating that our

47
00:04:20,350 --> 00:04:26,940
active user will give to these movies, based on her similarity to other users.

48
00:04:26,940 --> 00:04:32,190
It is obvious that we can use it to rank the movies for providing recommendation to our

49
00:04:32,190 --> 00:04:33,530
active user.

50
00:04:33,530 --> 00:04:39,660
Now, let’s examine what’s different between “User-based” and “Item-based” Collaborative

51
00:04:39,660 --> 00:04:43,721
filtering: In the User-based approach, the recommendation

52
00:04:43,721 --> 00:04:50,200
is based on users of the same neighborhood, with whom he or she shares common preferences.

53
00:04:50,200 --> 00:04:58,720
For example, as User1 and User3 both liked Item 3 and Item 4, we consider them as similar

54
00:04:58,720 --> 00:05:08,070
or neighbor users, and recommend Item 1, which is positively rated by User1 to User3.

55
00:05:08,070 --> 00:05:13,420
In the item-based approach, similar items build neighborhoods on the behavior of users.

56
00:05:13,420 --> 00:05:18,780
(Please note, however, that it is NOT based on their content).

57
00:05:18,780 --> 00:05:24,790
For example, Item 1 and Item 3 are considered neighbors, as they were positively rated by

58
00:05:24,790 --> 00:05:31,771
both User1 and User2. So, Item 1 can be recommended to User 3 as

59
00:05:31,771 --> 00:05:37,810
he has already shown interest in Item3. Therefore, the recommendations here are based

60
00:05:37,810 --> 00:05:42,960
on the items in the neighborhood that a user might prefer.

61
00:05:42,960 --> 00:05:47,650
Collaborative filtering is a very effective recommendation system, however, there are

62
00:05:47,650 --> 00:05:53,110
some challenges with it as well. One of them is Data Sparsity.

63
00:05:53,110 --> 00:05:57,910
Data sparsity happens when you have a large dataset of users, who generally, rate only

64
00:05:57,910 --> 00:06:03,430
a limited number of items. As mentioned, collaborative-based recommenders

65
00:06:03,430 --> 00:06:08,300
can only predict scoring of an item if there are other users who have rated it.

66
00:06:08,300 --> 00:06:14,080
Due to sparsity, we might not have enough ratings in the user-item dataset, which makes

67
00:06:14,080 --> 00:06:20,410
it impossible to provide proper recommendations. Another issue to keep in mind is something

68
00:06:20,410 --> 00:06:25,720
called ‘cold start.’ Cold start refers to the difficulty the recommendation

69
00:06:25,720 --> 00:06:32,110
system has when there is a new user and, as such, a profile doesn’t exist for them yet.

70
00:06:32,110 --> 00:06:38,140
Cold start can also happen when we have a new item, which has not received a rating.

71
00:06:38,140 --> 00:06:44,000
Scalability can become an issue, as well. As the number of users or items increases

72
00:06:44,000 --> 00:06:49,750
and the amount of data expands, Collaborative filtering algorithms will begin to suffer

73
00:06:49,750 --> 00:06:55,500
drops in performance, simply due to growth in the similarity computation.

74
00:06:55,500 --> 00:07:01,000
There are some solutions for each of these challenges, such as using hybrid-based recommender

75
00:07:01,000 --> 00:07:04,350
systems, but they are out of scope of this course.

76
00:07:04,350 --> 00:07:05,720
Thanks for watching!