Recommender System Seminar

Slide 1

Slide 1 text

RECOMMENDER SYSTEM

Slide 2

Slide 2 text

Building Recommender System in Industries

Slide 3

Slide 3 text

1. Background: Deﬁnition, Problem Construction and Possible Solutions. 2. SVD-like for Recommender System: Matrix Factorization to model matrix completion problem 3. Implicit and Explicit signal: How to model Explicit and Implicit Signal with SVD-like RecSys 4. Recommender System workﬂow: Estimate, Filtering, Ranking and Randomization 5. Recommender System Cases Content

Slide 4

Slide 4 text

Background: Recommender System in Industries

Slide 5

Slide 5 text

● Chris Anderson in “The Long Tail ○ “We are leaving the age of information and entering the age of recommendation” ● CNN Money, “The race to create a 'smart' Google”: ○ “The Web, they say, is leaving the era of search and entering one of discovery. What's the difference? Search is what you do when you're looking for something. Discovery is when something wonderful that you didn't know existed, or didn't know how to ask for, ﬁnds you.” Source: compiled from Xavier Amatriain MLSS slides (2014) Background

Slide 6

Slide 6 text

Background

Slide 7

Slide 7 text

Background

Slide 8

Slide 8 text

● Value of Recommendation ○ Netﬂix: 2/3 of the movies watched are recommended ○ Google News: recommendations generate 38% more click through ○ Amazon: 35% sales from recommendations ● Metrics which Affected by Recommendation ○ Activity Metrics: Increase in user retention ○ Financial Metrics: Increase in sales ○ Product Activity: Increase on number of unique items bought Source: compiled from Xavier Amatriain MLSS slides (2014) Background

Slide 9

Slide 9 text

● The “Recommender problem” ○ Estimate a utility function that automatically predicts how a user will like an item. ○ Based on: ■ Past behavior ■ Relations to other users ■ Item similarity ■ Context Source: compiled from Xavier Amatriain MLSS slides (2014) Background

Slide 10

Slide 10 text

The “Recommender problem” ● Let C be set of all users and let S be set of all possible recommendable items ● Let u be a utility function measuring the usefulness of item s to user c, i.e., u : C X S→R, where R is a totally ordered set ● For each user c є C, we want to choose items s є S that maximize u. Utility is usually represented by rating but can be any function Source: compiled from Xavier Amatriain MLSS slides (2014) Background

Slide 11

Slide 11 text

Possible Solutions Source: compiled from Xavier Amatriain MLSS slides (2014) Background

Slide 12

Slide 12 text

Problems in Recommendation System: Long Tail Sales Background

Slide 13

Slide 13 text

Problems in Recommendation System ● Some users never use some items Background

Slide 14

Slide 14 text

SVD-like for Recommender System: Matrix Factorization to model RecSys

Slide 15

Slide 15 text

Non-Personalized Recommendation

Slide 16

Slide 16 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Non-Personalized Recommendation a. For example, Content based ﬁltering, recommend similar items. Product name embedding with Word2Vec SVD-like Recommendation

Slide 17

Slide 17 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Non-Personalized Recommendation SVD-like Recommendation

Slide 18

Slide 18 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Non-Personalized Recommendation a. Recommend items based on similarity: SVD-like Recommendation

Slide 19

Slide 19 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Non-Personalized Recommendation a. Pros: i. Can be used if you don’t have any transaction history in the beginning ii. Sometimes can beat popular items benchmark. b. Cons: i. Low diversity metrics 1. Buy (mie-ayam), next recommendation (mie-ayam) SVD-like Recommendation

Slide 20

Slide 20 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Non-Personalized Recommendation SVD-like Recommendation

Slide 21

Slide 21 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Non-Personalized Recommendation SVD-like Recommendation arg max similarity(item1, item2)

Slide 22

Slide 22 text

Personalized Recommendation

Slide 23

Slide 23 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 2. Personalized Recommendation SVD-like Recommendation arg max Probability(item, user)

Slide 24

Slide 24 text

Before dwelling to SVD-like RecSys CF, let’s focus on other approach: 1. Personalized Recommendation a. Based on items, users, and items-users interaction SVD-like Recommendation Alex Smola slides form Berkley ML class (2012)

Slide 25

Slide 25 text

How to model matrix completion: ● Matrix Factorization SVD-like Recommendation Alex Smola slides form Berkley ML class (2012)

Slide 26

Slide 26 text

Matrix Factorization ● Latent Variable SVD-like Recommendation Alex Smola slides form Berkley ML class (2012)

Slide 27

Slide 27 text

Matrix Factorization ● Unfortunately we can’t user Matrix Factorization for sparse data SVD-like Recommendation Alex Smola slides form Berkley ML class (2012)

Slide 28

Slide 28 text

Funk SVD Alex Smola slides form Berkley ML class (2012)

Slide 29

Slide 29 text

Funk SVD Source: compiled from Alex Lin slides, CF with MF (2011)

Slide 30

Slide 30 text

SVD-like Recommender System Source: compiled from Alex Lin slides, CF with MF (2011)

Slide 31

Slide 31 text

SVD-like Recommender System Source: compiled from Alex Lin slides, CF with MF (2011)

Slide 32

Slide 32 text

SVD-like Recommender System Source: compiled from Alex Lin slides, CF with MF (2011)

Slide 33

Slide 33 text

SVD-like Recommender System Source: compiled from Alex Lin slides, CF with MF (2011)

Slide 34

Slide 34 text

Modeling Implicit and Explicit Feedback

Slide 35

Slide 35 text

Two types of Feedback Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)

Slide 36

Slide 36 text

In rating systems, such as MovieLens and Allmusic, the value of an entry denotes the rating of the item given by the user. Ratings can explicitly reﬂect the preference of an individual. Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)

Slide 37

Slide 37 text

For instance, in a music podcast service, the value of an entry may denote the subscription. Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)

Slide 38

Slide 38 text

Or play count Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)

Slide 39

Slide 39 text

Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)

Slide 40

Slide 40 text

SVD-like model is so easy to ﬁt any feedback type 1. Explicit a. Rating (Netﬂix Competition) Change the loss function to RMSE, regression task. Explicit and Implicit Feedback Source: compiled from Li Yen Kuo slides, Implicit RecSys (2019)

Slide 41

Slide 41 text

SVD-like model is so easy to ﬁt any feedback type 1. Explicit b. Like/Dislike (Netﬂix Now) Change the loss function to log-loss, sigmoid activation. Explicit and Implicit Feedback

Slide 42

Slide 42 text

SVD-like model is so easy to ﬁt any feedback type 2. Implicit a. Read/Not Read (Quora) Change loss function to BPR Loss Explicit and Implicit Feedback

Slide 43

Slide 43 text

SVD-like model is so easy to ﬁt any feedback type 2. Implicit b. Frequency of Buy (Retail) Change loss function to Regression loss with Alpha, Koren (2009) Explicit and Implicit Feedback

Slide 44

Slide 44 text

Recommender System workﬂow: Estimate, Ranking and Randomization

Slide 45

Slide 45 text

Recommender System Workﬂow Source: compiled from Xavier Amatriain MLSS slides (2014)

Slide 46

Slide 46 text

Modeling 1. Build Recommender System based on Watch/Not a. Implicit model, maximize click through rate 2. Build Recommender System based on Rating a. Explicit model, maximize preference 3. Build Recommender System based on Churn a. Implicit model, maximize retention Recommender System Workﬂow Source: compiled from Xavier Amatriain MLSS slides (2014)

Slide 47

Slide 47 text

Recommender System Workﬂow Ranking ● Most recommendations are presented in a sorted list ● Recommendation can be understood as a ranking problem ● Popularity is the obvious baseline ● Ratings prediction is a clear secondary data input that allows for personalization ● Many other features can be added

Slide 48

Slide 48 text

Recommender System Workﬂow Ranking Source: compiled from Xavier Amatriain MLSS slides (2014)

Slide 49

Slide 49 text

Randomization Randomization ● We will focus on Diversity Metrics vs Accuracy Metrics. ● More diverse recommendation will increase Netﬂix CTR ● More accurate recommendation will increase Netﬂix CTR ● Diversity and Accuracy are negatively correlated.

Slide 50

Slide 50 text

Randomization

Slide 51

Slide 51 text

Recommender System Cases

Slide 52

Slide 52 text

Questions Pair: Similarity Task

Slide 53

Slide 53 text

Questions Pair: Similarity Task Question 1 Question 2 {0, 1}

Slide 54

Slide 54 text

Questions Pair: Similarity Task ● Is Item A similar to Item B?

Slide 55

Slide 55 text

Questions Pair: Similarity Task Input 1 Input 2 Neural Networks {0, 1}

Slide 56

Slide 56 text

Temporal Recommendation Task

Slide 57

Slide 57 text

Temporal Recommendation Task Item 1 Item 4 Item 2 Item 1 Item 2 Item 3 Item 2 Item 1 Item 5 RNN

Slide 58

Slide 58 text

Recommendation Task

Slide 59

Slide 59 text

Recommendation Task Click data Hotel 33 Hotel 77 . . . Hotel 44

Slide 60

Slide 60 text

Searching Task 60 Pseudo code BPR For K times: 1. A = Sample random label positive 2. B = Sample random label negative 3. Dist = Model predict distance(A,B) 4. Update, make the distance farther.

Slide 61

Slide 61 text

Searching Task Loss Function: 61

Slide 62

Slide 62 text

Searching Task ● Pointwise ○ for each items, train a classiﬁer / regressor on it to predict how relevant it is ● Pairwise ○ given a pair of documents, compare which one has the highest rank. ● Listwise ○ sort the entire list of documents ■ Direct optimization of IR measures such as NDCG

Slide 63

Slide 63 text

Searching Task: 63

Slide 64

Slide 64 text

Cowok 1 Cowok 3 Cowok 2 Cewek A Cewek C Cewek B Cowok 4 Cewek D Matching Task

Slide 65

Slide 65 text

Matching Task Cowok 1 Cewek C Cewek D Cewek A Cewek B Cewek A Cowok 4 Cowok 2 Cowok 1 Cowok 3

Slide 66

Slide 66 text

Matching Task 66 Cewek 1 Cowok 2 {0, 1}

Slide 67

Slide 67 text

Matching Task Loss Function: 67

Slide 68

Slide 68 text

Matching Task Matching Algorithm: Gale-Shapley 68

Slide 69

Slide 69 text

cowok cowok cowok cewek cewek cewek cowok cewek Matching Task

Slide 70

Slide 70 text

Matching Task 70

Slide 71

Slide 71 text

You can learn all of these in our Advance ML CLass Contact: business@pacmannai.com

Slide 72

Slide 72 text

We believe everyone can build cool Recommendation Systems Contact: business@pacmannai.com

Slide 73

Slide 73 text

Upcoming Advance ML CLass Contact: business@pacmannai.com

Slide 74

Slide 74 text

No content

Slide 75

Slide 75 text

Check our website: www.pacmann.ai Contact: business@pacmannai.com

Slide 76

Slide 76 text

Check our website: www.pacmann.ai Contact: business@pacmannai.com

Slide 77

Slide 77 text

Check our website: www.pacmann.ai Contact: business@pacmannai.com

Slide 78

Slide 78 text

Past Classes Contact: business@pacmannai.com

Slide 79

Slide 79 text

Pacmann AI Classes Quality State of the Art of Machine Learning Research Practical Skills Theoretical Understanding > 50 institutions 400++ alumni 6 Classes in the past business@pacmannai.com https://pacmann.ai

Slide 80

Slide 80 text

Previous Classes business@pacmannai.com https://pacmann.ai 61 participants 8 weeks 48 institutions

Slide 81

Slide 81 text

Previous Classes business@pacmannai.com https://pacmann.ai 59 participants 8 weeks 48 institutions

Slide 82

Slide 82 text

Previous Classes business@pacmannai.com https://pacmann.ai 44 participants 2 weeks 33 institutions

Slide 83

Slide 83 text

Previous Classes Facts & Figure business@pacmannai.com https://pacmann.ai Field of Work

Slide 84

Slide 84 text

Previous Participants business@pacmannai.com https://pacmann.ai

Slide 85

Slide 85 text

Contact Email: business@pacmannai.com Whatsapp Business: +62 812-8122-1707