Современные подходы к проблемам в рекомендательных системах. Александр Тужилин New York University, Сбербанк AI Lab

Transcription

1 Современные подходы к проблемам в рекомендательных системах Александр Тужилин New York University, Сбербанк AI Lab

2 Outline of the Talk Traditional matrix completion paradigm of recommender systems Going beyond the traditional matrix completion paradigm Example of a non-traditional approach recommending products with the most valuable aspects based on customer reviews [KDD 2017] 2

3 Recommender Systems (RS) 3

4 The Traditional Paradigm of Recommender Systems Matrix R of known ratings r ij : rating user c i assigns to item s j Matrix X of user attributes x ij : attribute x j of user c i Matrix Y of item attributes y ij : attribute y j of item s i A matrix completion problem: estimate unknown ratings in R Solutions [AT05]: Collaborative Filtering (CF) Content-Based Hybrid Example: Netflix Prize Competition c 1 c 2 c M c 1 c 2 c M s 1 s 2 s N s 1 s 2 s N R x 1 x 2 x P X y 1 y 2 y Q Y Rˆ c 1 c 2 c M f ( R, X, Y) s 1 s 2 s N Rˆ 4

5 Netflix Competition Recommendation Matrix K-PAX Life of Brian Memento Notorious U U U U The Users Items Matrix of Ratings (w/ timestamps) Key issue: accurate estimation of unknown ratings (RMSE) 5

6 Characteristics of the Traditional (Matrix Completion) Paradigm Two-dimensional (2D) paradigm: Users and Items 3 matrices: R, X and Y Utility of an item to a user revealed by a single rating binary or multi-scaled Recommendations of individual items provided to individual users Many solution via estimation of unknown ratings 6

7 performance Going Beyond the Traditional Matrix Completion Paradigm Remove limiting assumptions of the traditional paradigm: Go beyond the characteristics of (a) User Item matrix, (b) single ratings, (c) recommending individual items to individual users [JRTZ16] Recommender Systems Beyond Matrix Completion Traditional New time 7

8 Going Beyond the Traditional Matrix Completion Paradigm Context-Aware RSes (CARS) Including spatiotemporal and mobile RSes Multi-criteria ratings Aggregate ratings and recommendations to groups Flexible and constraint-based recommendations Non-rating-based approaches (e.g., ranking-based) New performance measures: novelty, serendipity, Social RSes User interactions/feedback, e.g., conversational RSes Trust and privacy Manipulation-resistant RSes Additional data sources, such as customer reviews 8

9 Recommending Products with Valuable Aspects Based on Customer Reviews Joint work with K. Bauman (Temple) and B. Liu (UIC) 9

10 Research Idea Idea: Recommending not only a product but also the most important (positive or negative) aspects that can enhance customer experience with the product. Examples: Positive: visit Aquagrill and order FISH there. Negative: visit Cafe X but do not order DESSERT there. Aspects come from customer reviews (e.g. Yelp). 10

11 Recommending Products and Aspects Try Goat Cheese Try Sweet Mango Sauce Try Lamb Curry or Chicken Masala Avoid Thai Tea 11

12 Importance of such Recommendations Why is it important? Novel approach to RSes that provides more tangible recommendations that enhance customer experience with the products. 12

13 Method of Recommending the Most Valuable Aspects Input: set of historical reviews with ratings. Output: product recommendations with the most valuable aspects enhancing customer experiences. 13

14 Method of Recommending the Most Valuable Aspects: An Overview 1. Extracting aspects from the reviews 2. Training Sentiment Utility Logistic Model (SULM) aspect sentiments overall satisfaction 3. Calculating aspect impact on rating 4. Recommending products and aspects 14

15 1. Extracting Aspects and Sentiments from the Reviews Determine set A of aspects in an application Aspects are characterized by a set of terms, e.g. MEAT - {meat, pork, bbq, lamb,..} For each review r determine set of aspects Ar discussed in r and corresponding set of sentiments Use Opinion Parser [Liu, 2010]. 15

16 1. Extracting Aspects and Sentiments from the Reviews Example: (1) Had lunch in Taqueria today. (2) Ordered the taco with rice and beans and it was great. (3) The service was quick. (4) The atmosphere was dark and soothing. FOOD positive SERVICE positive ATMOSPHERE positive 16

17 2. Sentiment Utility Logistic Model (SULM) Main purpose is to estimate the overall customer experience Simultaneously fits ratings and sentiments Identifies relative importance of the aspects in the potential customer experience 17

18 2. (a) Training SULM: Aspect Sentiments Expressed sentiment (OP output): Sentiment utility - level of satisfaction of user with aspect of item, latent variable. Use Matrix Factorization (Koren et al. 2009) to estimate: 18

19 2. (b) Training SULM: Explicit Sentiment vs. Actual Value of an Aspect Logistic function: Estimation of explicit sentiment: 1 0,5 Maximize log-likelihood: Estimate θ s so that estimated values of sentiments fit the real binary sentiments o k u i extracted by OP 19

20 2. (c) Training SULM: Overall Satisfaction Overall rating (binary): Overall level of satisfaction (latent): 20

21 2. (d) Training SULM: Explicit Rating vs. Actual Value of an Experience Binary rating estimation: 1 0,5 Maximize log-likelihood:

22 2. Scheme of SULM Estimate parameters of SULM such that 1) 2) 1 0, ,

23 2. Training SULM Simultaneously fits ratings and sentiments: Use Stochastic Gradient Descent to fit the model. 23

24 3. Aspect Impact on Rating For a new potential review compute the impact of each aspect on the predicted rating as the corresponding summand from the rating prediction part of the model 24

25 4. Recommending Products and Aspects identify group of aspects over which the customer has control (e.g. Dish vs. Decor); same for the management identify the most valuable aspects of the potential experience recommend a product and its corresponding suggestions to experience (positive) or do not experience (negative) a particular aspect 25

26 4. Recommending Aspects to Managers Provide complimentary drink Recommend foot massage Manager of a Spa Salon Do not talk too much during the procedure 26

27 Experimental Settings Restaurants Hotels Beauty&Spa Initial 1,344,405 96, ,199 Filtered 602,112 5,669 5,065 Users 23,

28 Examples of Aspects (Restaurants) Meat Fish Dessert Money Service Decor beef cod tiramisu price bartender design meat salmon cheesecake dollars waiter ceiling bbq catfish chocolate cost service décor ribs tuna dessert budget hostess lounge veal shark ice cream charge manager window pork fish macaroons check staff space 28

29 Numbers of Aspects Restaurant Hotel Beauty & Spa Total number Customer can control Management can control

30 Performance Measures 1. Recommendations of Aspects: how much the average rating is changed for those customers who followed the recommendations of aspects 2. Rating prediction: AUC, 3. Aspect ranking: 30

31 Baselines Recommending the most popular aspect of a product Recommending highly rated aspect of a product Hidden Factors as Topics (HFT) (McAuley and Leskovec 2013) - state-ofthe-art rating prediction method based on customer reviews Learning to Rank User Preferences Based on Phrase-Level Sentiment Analysis across Multiple Categories (LRPPM) (Chen et al. 2016) - the latest method for predicting the list of aspects appearing in the review 31

32 Recommendations of Aspects (Restaurants) 72.3% Average 65.1% Followed Positive 62.9% Not Followed Positive Conclusion: our recommendations help to get better customer experience as captured by the ratings 32

33 Recommendations of Aspects (Restaurants) Followed positive recommendations 74,0 % 72,0 % 70,0 % 68,0 % 66,0 % 64,0 % 62,0 % 60,0 % Customers Managers Average Most Popular Highly Rated SULM Conclusion: our recommendations help to get better customer experience vs. baselines. 33

34 Recommendations of Aspects (Restaurants) Did not follow positive recommendations 66,0 % 65,0 % 64,0 % Average Lowest sentiment SULM 63,0 % 62,0 % 61,0 % Customers Managers Conclusion: our recommendations help to avoid negative customer experience better than baselines 34

35 Recommendations of Aspects (Hotels) Followed positive recommendations 70,0 % 65,0 % 60,0 % 55,0 % 50,0 % Customers Managers Average Most Popular Highly Rated SULM Conclusion: our recommendations help to get better customer experience vs. baselines. 35

36 Recommendations of Aspects (Beauty & Spa) Followed positive recommendations 76,5 % 75,0 % 73,5 % 72,0 % 70,5 % 69,0 % 67,5 % Customers Managers Average Most Popular Highly Rated SULM Conclusion: our recommendations help to get better customer experience vs. baselines. 36

37 Rating Prediction Performance AUC Application R H B&S R H B&S HFT LRPPM SULM R - restaurant, H - hotel, B&S - beauty&spa Conclusion: our rating prediction performance is comparable to the baseline performances. 37

38 Aspect Ranking Performance Application R H B&S R H B&S LRPPM SULM R - restaurant, H - hotel, B&S - beauty&spa Conclusion: our aspect ranking performance is comparable to the baseline performance. 38

39 Conclusion Proposed a new method (SULM) of recommending not only products but also the most valuable aspects enhancing customer experiences. Tested on 3 applications (restaurant, hotel and beauty & spas) and showed that our recommendations lead to better customer experience as captured by the ratings. The proposed method Provides more detailed recommendations Helps customers make more informed decisions Helps customers get better experiences with products. 39

40 Future Work on Recommender Systems In search of the next paradigm of RSes My vision: current methods will be enhanced by novel approaches from economics psychology 40

41 THANK YOU! Alex Tuzhilin