CSE 258. Web Mining and Recommender Systems. Advanced Recommender Systems

Transcription

1 CSE 258 Web Mining and Recommender Systems Advanced Recommender Systems

2 This week Methodological papers Bayesian Personalized Ranking Factorizing Personalized Markov Chains Personalized Ranking Metric Embedding Translation-based Recommendation

3 This week Goals:

4 This week Application papers (Wednesday) Recommending Product Sizes to Customers Playlist prediction via Metric Embedding Efficient Natural Language Response Suggestion for Smart Reply Personalized Itinerary Recommendation with Queuing Time Awareness Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences

5 This week We (hopefully?) know enough by now to Read academic papers on Recommender Systems Understand most of the models and evaluations used See also CSE291

6 Bayesian Personalized Ranking

7 Bayesian Personalized Ranking Goal: Estimate a personalized ranking function for each user

8 Bayesian Personalized Ranking Why? Compare to traditional approach of replacing missing values by 0: But! 0 s aren t necessarily negative!

9 Bayesian Personalized Ranking Why? Compare to traditional approach of replacing missing values by 0: This suggests a possible solution based on ranking

10 Bayesian Personalized Ranking Defn: AUC (for a user u) ( ) scoring function that compares an item i to an item j for a user u The AUC essentially counts how many times the model correctly identifies that u prefers the item they bought (positive feedback) over the item they did not

11 Bayesian Personalized Ranking Defn: AUC (for a user u) AUC = 1: AUC = 0.5: We always guess correctly among two potential items i and j We guess no better than random

12 Bayesian Personalized Ranking Defn: AUC = Area Under Precision Recall Curve

13 Bayesian Personalized Ranking Summary: Goal is to count how many times we identified i as being more preferable than j for a user u

14 Bayesian Personalized Ranking Summary: Goal is to count how many times we identified i as being more preferable than j for a user u

15 Bayesian Personalized Ranking Idea: Replace the counting function by a smooth function is any function that compares the compatibility of i and j for a user u e.g. could be based on matrix factorization:

16 Bayesian Personalized Ranking Idea: Replace the counting function by a smooth function

17 Bayesian Personalized Ranking Idea: Replace the counting function by a smooth function

18 Bayesian Personalized Ranking Experiments: RossMann (online drug store) Netflix (treated as a binary problem)

19 Bayesian Personalized Ranking Experiments:

20 Bayesian Personalized Ranking Morals of the story: Given a one-class prediction task (like purchase prediction) we might want to optimize a ranking function rather than trying to factorize a matrix directly The AUC is one such measure that counts among a users u, items they consumed i, and items they did not consume, j, how often we correctly guessed that i was preferred by u We can optimize this approximately by maximizing where

21 Factorizing Personalized Markov Chains for Next-Basket Recommendation

22 Factorizing Personalized Markov Chains for Next-Basket Recommendation Goal: build temporal models just by looking at the item the user purchased previously (or )

23 Factorizing Personalized Markov Chains for Next-Basket Recommendation Assumption: all of the information contained by temporal models is captured by the previous action this is what s known as a first-order Markov property

24 Factorizing Personalized Markov Chains for Next-Basket Recommendation Is this assumption realistic?

25 Factorizing Personalized Markov Chains for Next-Basket Recommendation Data setup: Rossmann basket data

26 Factorizing Personalized Markov Chains for Next-Basket Recommendation Prediction task:

27 Factorizing Personalized Markov Chains for Next-Basket Recommendation Could we try and compute such probabilities just by counting? Seems okay, as long as the item vocabulary is small (I^2 possible item/item combinations to count) But it s not personalized

28 Factorizing Personalized Markov Chains for Next-Basket Recommendation What if we try to personalize? Now we would have U*I^2 counts to compare Clearly not feasible, so we need to try and estimate/model this quantity (e.g. by matrix factorization)

29 Factorizing Personalized Markov Chains for Next-Basket Recommendation What if we try to personalize?

30 Factorizing Personalized Markov Chains for Next-Basket Recommendation What if we try to personalize?

31 Factorizing Personalized Markov Chains for Next-Basket Recommendation Prediction task:

32 Factorizing Personalized Markov Chains for Next-Basket Recommendation Prediction task:

33 Factorizing Personalized Markov Chains for Next-Basket Recommendation FMC: not personalized MF: personalized, but not sequentially-aware

34 Factorizing Personalized Markov Chains for Next-Basket Recommendation Morals of the story: Can improve performance by modeling third order interactions between the user, the item, and the previous item This is simpler than temporal models but makes a big assumption Given the blowup in the interaction space, this can be handled by tensor decomposition techniques

35 Personalized Ranking Metric Embedding for Next New POI Recommendation

36 Factorizing Personalized Markov Chains for Next-Basket Recommendation Goal: Can we build better sequential recommendation models by using the idea of metric embeddings vs.

37 Personalized Ranking Metric Embedding for Next New POI Recommendation Why would we expect this to work (or not)?

38 Personalized Ranking Metric Embedding for Next New POI Recommendation Otherwise, goal is the same as the previous paper:

39 Personalized Ranking Metric Embedding for Next New POI Recommendation Data

40 Personalized Ranking Metric Embedding for Next New POI Recommendation Qualitative analysis

41 Personalized Ranking Metric Embedding for Next New POI Recommendation Qualitative analysis

42 Personalized Ranking Metric Embedding for Next New POI Recommendation Basic model (not personalized)

43 Personalized Ranking Metric Embedding for Next New POI Recommendation Basic model (not personalized)

44 Personalized Ranking Metric Embedding for Next New POI Recommendation Personalized version

45 Personalized Ranking Metric Embedding for Next New POI Recommendation Personalized version

46 Personalized Ranking Metric Embedding for Next New POI Recommendation Learning

47 Personalized Ranking Metric Embedding for Next New POI Recommendation Results

48 Personalized Ranking Metric Embedding for Next New POI Recommendation Morals of the story: In some applications, metric embeddings might be better than inner products Examples could include geographical data, but also others (e.g. playlists?)

49 Translation-based Recommendation

50 Translation-based Recommendation Goal: (e.g) which movie is this user going to watch next? viewing history of Want models that consider characteristics/preferences of each user local context, i.e., the last consumed item(s)

51 Translation-based Recommendation Goal: (e.g) which movie is this user going to watch next? viewing history of Option 1: Matrix Factorization

52 Translation-based Recommendation Goal: (e.g) which movie is this user going to watch next? viewing history of Option 2: Markov Chains

53 Translation-based Recommendation Idea: Considering the two simultaneously means modeling the interactions between a user and adjacent items user transition previous item next item

54 Translation-based Recommendation Compare: Factorized Personalized Markov Chains (earlier today) user preference local context

55 Translation-based Recommendation Compare: Personalized Ranking Metric Embedding (earlier today) an additional hyperpara. to balance the two components

56 Translation-based Recommendation Compare: Hierarchical Representation Model (HRM) Wang et al., 2015 (earlier today) average/max pooling, etc.

57 Translation-based Recommendation Compare: Hierarchical Representation Model (HRM) Wang et al., 2015 (earlier today) average/max pooling, etc. Goal: Try and get the best of both worlds, by modeling third-order interactions and using metric embeddings

58 Translation-based Recommendation Detour: Translation models in Knowledge Bases Data: entities; links (multiple types of relationships) State-of-the-art method: relationships as translations Goal: Predict unseen links Training example: entity h relation r entity t Basic idea: E.g. [Bordes et al., 2013], [Wang et al., 2014], [Lin et al., 2015]

59 Translation-based Recommendation Embedding space Items as points Users as translation vectors user Training triplet: previous item next item Objective:

60 Translation-based Recommendation Embedding space Items as points Users as translation vectors

61 Translation-based Recommendation bias Benefit from using metric embeddings Model (u, i, j) with a single component Recommendations can be made by a simple NN search

62 Translation-based Recommendation

63 Translation-based Recommendation Automotives Office Products Toys & Games Video Games Cell Phones & Accessories Clothing, Shoes, and Jewelry Electronics May July 2014

64 Translation-based Recommendation check-ins at different venues Dec Apr movie ratings Nov Nov user reviews Jan Nov (all available online)

65 Translation-based Recommendation 11.4M reviews & ratings of 4.5M users on 3.1M local businesses restaurants, hotels, parks, shopping malls, movie theaters, schools, military recruiting offices, bird control, mediation services... Characteristics: vast vocabulary of items, variability, and sparsity

66 Translation-based Recommendation

67 Translation-based Recommendation varying sparsity

68 Translation-based Recommendation Unified

69 Translation-based Recommendation TransRec

70 Translation-based Recommendation Works well with Doesn t work well with

71 Overview Morals of the story: Today we looked at two main ideas that extend the recommender systems we saw in class: 1. Sequential Recommendation: Most of the dynamics due to time can be captured purely by knowing the sequence of items 2. Metric Recommendation: In some settings, using inner products may not be the correct assumption

72 Assignment 1

73 Assignment 1

74 CSE 258 Web Mining and Recommender Systems Real-world applications of recommender systems

75 Recommending product sizes to customers

76 Recommending product sizes to customers Goal: Build a recommender system that predicts whether an item will fit :

77 Recommending product sizes to customers Challenges: Data sparsity: people have very few purchases from which to estimate size Cold-start: How to handle new customers and products with no past purchases? Multiple personas: Several customers may use the same account

78 Recommending product sizes to customers Data: Shoe transactions from Amazon.com For each shoe j, we have a reported size c_j (from the manufacturer), but this may not be correct! Need to estimate the customer s size (s_i), as well as the product s true size (t_j)

79 Recommending product sizes to customers Loss function:

80 Recommending product sizes to customers Loss function:

81 Recommending product sizes to customers Loss function:

82 Recommending product sizes to customers

83 Recommending product sizes to customers Loss function:

84 Recommending product sizes to customers Model fitting:

85 Recommending product sizes to customers Extensions: Multi-dimensional sizes Customer and product features User personas

86 Recommending product sizes to customers Experiments:

87 Recommending product sizes to customers Experiments: Online A/B test

88 Playlist prediction via Metric Embedding

89 Playlist prediction via Metric Embedding Goal: Build a recommender system that recommends sequences of songs Idea: Might also use a metric embedding (consecutive songs should be nearby in some space)

90 Playlist prediction via Metric Embedding Basic model: (compare with metric model from last lecture)

91 Playlist prediction via Metric Embedding Basic model ( single point ):

92 Playlist prediction via Metric Embedding Dual-point model

93 Playlist prediction via Metric Embedding Extensions: Popularity biases

94 Playlist prediction via Metric Embedding Extensions: Personalization

95 Playlist prediction via Metric Embedding Extensions: Semantic Tags

96 Playlist prediction via Metric Embedding Extensions: Observable Features

97 Playlist prediction via Metric Embedding Experiments: Yes.com playlists Dec 2010 May 2011 Small dataset: 3,168 songs 134, ,191,279 transitions Large dataset 9,775 songs 172,510 transitions + 1,602,079 transitions

98 Playlist prediction via Metric Embedding Experiments:

99 Playlist prediction via Metric Embedding Experiments: Small Big

100 Efficient Natural Language Response Suggestion for Smart Reply

101 Efficient Natural Language Response Suggestion for Smart Reply Goal: Automatically suggest common responses to s

102 Efficient Natural Language Response Suggestion for Smart Reply Basic setup

103 Efficient Natural Language Response Suggestion for Smart Reply Previous solution (KDD 2016) Based on a seq2seq method

104 Efficient Natural Language Response Suggestion for Smart Reply Idea: Replace this (complex) solution with a simple multiclass classification-based solution

105 Efficient Natural Language Response Suggestion for Smart Reply Idea: Replace this (complex) solution with a simple multiclass classification-based solution

106 Efficient Natural Language Response Suggestion for Smart Reply Model: S(x,y)

107 Efficient Natural Language Response Suggestion for Smart Reply Model: Architecture v1

108 Efficient Natural Language Response Suggestion for Smart Reply Model: Architecture v2

109 Efficient Natural Language Response Suggestion for Smart Reply Model: Extensions

110 Efficient Natural Language Response Suggestion for Smart Reply Model: Extensions

111 Efficient Natural Language Response Suggestion for Smart Reply Experiments: (offline)

112 Efficient Natural Language Response Suggestion for Smart Reply Experiments: (online)

113 Personalized Itinerary Recommendation with Queuing Time Awareness

114 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences

115 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences Goal: Identify items that might be purchased together

116 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences

117 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences browsed together (substitutable) bought together (complementary)

118 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences Four types of relationship: 1) People who viewed X also viewed Y 2) People who viewed X eventually bought Y 3) People who bought X also bought Y 4) People bought X and Y together Substitutes (1 and 2), and Complements (3 and 4)

119 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences 1) Data collection

120 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences 2) Training data generation

121 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences 2) Training (simpler models)

122 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences 2) Training (simpler models)

123 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences 3) Siamese CNNs

124 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences 4) Recommendation

125 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences

126 Learning Visual Clothing Style with Heterogeneous Dyadic Co-occurrences