DS504/CS586: Big Data Analytics Recommender System

Transcription

1 Welcome to DS0/CS86: Big Data Analytics Recommender System Prof. Yanhua Li Time: 6:00pm 8:0pm Thu. Location: AK Fall 06

2 Example: Recommender Systems v Customer X Star War I Star War II v Customer Y Does search on Star War I Recommender system suggests Star War II from data collected about customer X J. Leskovec, A. Rajaraman, J. Ullman:

3 Recommendations Examples: Search Recommendations Items Products, web sites, blogs, news items, J. Leskovec, A. Rajaraman, J. Ullman:

4 From Scarcity to Abundance v Shelf space is a scarce commodity for traditional retailers Also: TV networks, movie theaters, v Web enables near-zero-cost dissemination of information about products From scarcity to abundance, e.g., Amazon, Target online, ebay, etc. v More choices necessitates better filters Recommendation engines J. Leskovec, A. Rajaraman, J. Ullman:

5 Types of Recommendations v Editorial and hand curated List of favorites Lists of essential items v Simple aggregates Top 0, Most Popular, Recent Uploads v Tailored to individual users Amazon, Netflix, J. Leskovec, A. Rajaraman, J. Ullman:

6 Formal Model v X = set of Customers v S = set of Items v Utility function u: X S à R R = set of ratings R is a totally ordered set e.g., 0- stars, real number in [0,] 6 J. Leskovec, A. Rajaraman, J. Ullman:

7 Utility Matrix Avatar LOTR Matrix Pirates Alice 0. Bob Carol 0. David 0. 7 J. Leskovec, A. Rajaraman, J. Ullman:

8 Key Problems v () Gathering known ratings for matrix How to collect the data in the utility matrix v () Estimate unknown ratings from the known ones Mainly interested in high unknown ratings We are not interested in knowing what you don t like but what you like v () Evaluating estimation methods How to measure success/performance of recommendation methods 8 J. Leskovec, A. Rajaraman, J. Ullman:

9 () Gathering Ratings v Explicit Ask people to rate items Doesn t work well in practice people can t be bothered v Implicit Learn ratings from user actions E.g., purchase implies high rating What about low ratings? 9 J. Leskovec, A. Rajaraman, J. Ullman:

10 () Estimating Utilities v Key problem: Utility matrix U is sparse Most people have not rated most items Cold start: New items have no ratings New users have no history v Approaches to recommender systems: ) Content-based ) Collaborative filtering 0 J. Leskovec, A. Rajaraman, J. Ullman:

11 Content-based Recommender Systems

12 Content-based Recommendations v Main idea: Recommend items to customer x similar to previous items rated highly by x Look at x s items vs all items Example: v Movie recommendations Recommend movies with same actor(s), director, genre, v Websites, blogs, news Recommend other sites with similar content J. Leskovec, A. Rajaraman, J. Ullman:

13 Plan of Action Item profiles likes recommend build match Red Circles Triangles User profile J. Leskovec, A. Rajaraman, J. Ullman:

14 Item Profiles v For each item, create an item profile v Profile is a set (vector) of features Movies: author, title, actor, director, Text: Set of important words in document v How to pick important features? Usual heuristic from text mining is TF-IDF (Term frequency * Inverse Doc Frequency) Term Feature Document Item J. Leskovec, A. Rajaraman, J. Ullman:

15 Sidenote: TF-IDF f ij = frequency of term (feature) i in doc j n i = number of docs that mention term i N = total number of docs Note: we normalize TF by the frequency of the most frequent term to discount for longer documents TF-IDF score: w ij = TF ij IDF i Doc profile = set of words with highest TF- IDF scores, together with their scores wj = (w j,..., wij,..., wkj) J. Leskovec, A. Rajaraman, J. Ullman:

16 User Profiles and Prediction v User profile possibilities: Weighted average of rated item profiles Variations: weight by difference from average rating for item wx = wj(rxj r x ) j=...nx v Prediction heuristic: Given user profile w x and item profile w j, estimate rxj = cos(wx, wj) = wxwj/ wj wx J. Leskovec, A. Rajaraman, J. Ullman: 6

17 Pros: Content-based Approach v +: No need for data on other users v +: Able to recommend to users with unique tastes v +: Able to recommend new & unpopular items No item cold-start v +: Able to provide explanations Can provide explanations of recommended items by listing content-features that caused an item to be recommended J. Leskovec, A. Rajaraman, J. Ullman: 7

18 Cons: Content-based Approach v : Finding the appropriate features is hard E.g., images, movies, music v : Recommendations for new users How to build a user profile? User code-start problem v : Overspecialization Never recommends items outside user s content profile People might have multiple interests Unable to exploit quality judgments of other users J. Leskovec, A. Rajaraman, J. Ullman: 8

19 Collaborative Filtering Harnessing quality judgments of other users

20 Collaborative Filtering v Consider user x v Find set N of other users whose ratings are similar to x s ratings v Estimate x s ratings based on ratings of users in N x N J. Leskovec, A. Rajaraman, J. Ullman: 0

21 Finding Similar Users r x = [*, _, _, *, ***] r y = [*, _, **, **, _] v Let r x be the vector of user x s ratings v Jaccard similarity measure r x, r y as sets: r x = {,, } r y = {,, } Problem: Ignore the value of the ratings: v Cosine Similarity measure Sim(x,y)=cos(r x, r y )=r x r y / r x r y Problem: Treading missing ratings as negatives v Pearson correlation coefficient r x, r y as points: r x = {, 0, 0,, } r y = {, 0,,, 0} v Sim(x,y)=(r x -r x,ave )(r y -r y,ave )/ r x -r x,ave r y -r y,ave

22 Cosine sim: Similarity Metric v Intuitively we want: sim(a, B) > sim(a, C) v Jaccard similarity: / < / v Cosine similarity: 0.86 > 0. Considers missing ratings as negative Solution: subtract the (row) mean Notice cosine sim. is correlation when data is centered at 0

23 User-User Collaborative Filtering For user u, find other similar users Estimate rating for item i based on ratings from similar users pred(u,i) = n neighbors(u) n neighbors(u) sim(u, n) rni sim(u, n) Sim(u,n) similarity of user u and n r ui rating of user u on item i neighbor(u) set of users similar to user u J. Leskovec, A. Rajaraman, J. Ullman:

24 Item-Item Collaborative Filtering v So far: User-user collaborative filtering v Another view: Item-item For item i, find other similar items Estimate rating for item i based on ratings for similar items Can use same similarity metrics and prediction functions as in user-user model r xi = j N ( i; x) s ij j N ( i; x) s r ij xj s ij similarity of items i and j r xj rating of user x on item j N(i;x) set items rated by x similar to i J. Leskovec, A. Rajaraman, J. Ullman:

25 Item-Item CF ( N =) users movies - unknown rating - rating between to J. Leskovec, A. Rajaraman, J. Ullman:

26 Item-Item CF ( N =) ? 6 users - estimate rating of movie by user 6 J. Leskovec, A. Rajaraman, J. Ullman: movies

27 Item-Item CF ( N =) users? sim(,m) movies Neighbor selection: Identify movies similar to movie, rated by user Here we use Pearson correlation as similarity: ) Subtract mean rating m i from each movie i m = (++++)/ =.6 row : [-.6, 0, -0.6, 0, 0,., 0, 0,., 0, 0., 0] ) Compute cosine similarities between rows 7

28 Item-Item CF ( N =) users? sim(,m) movies Compute similarity weights: s, =0., s,6 =0.9 J. Leskovec, A. Rajaraman, J. Ullman: 8

29 Item-Item CF ( N =) users movies 6 Predict by taking weighted average: r. = (0.* + 0.9*) / (0.+0.9) =.6 J. Leskovec, A. Rajaraman, J. Ullman: 9

30 Item-Item vs. User-User Avatar LOTR Matrix Pirates Alice 0.8 Bob Carol David 0. In prac(ce, it has been observed that item-item oen works be8er than user-user Why? Items are simpler, users have mul0ple tastes J. Leskovec, A. Rajaraman, J. Ullman: 0

31 Pros/Cons of Collaborative Filtering v + Works for any kind of item No feature selection needed v - Cold Start: Need enough users in the system to find a match v - Sparsity: The user/ratings matrix is sparse Hard to find users that have rated the same items v - First rater: Cannot recommend an item that has not been previously rated New items, Esoteric items v - Popularity bias: Cannot recommend items to someone with unique taste Tends to recommend popular items J. Leskovec, A. Rajaraman, J. Ullman:

32 Hybrid Methods v Implement two or more different recommenders and combine predictions Perhaps using a linear model v Add content-based methods to collaborative filtering Item profiles for new item problem Demographics to deal with new user problem J. Leskovec, A. Rajaraman, J. Ullman:

33 Evaluation movies users J. Leskovec, A. Rajaraman, J. Ullman:

34 Evaluation users movies????? Test Data Set J. Leskovec, A. Rajaraman, J. Ullman:

35 Collaborative Filtering: Complexity v Expensive step is finding k most similar customers: O( X ) v Too expensive to do at runtime Could pre-compute v Naïve pre-computation takes time O(k X ) X set of customers v We already know how to do this! Near-neighbor search in high dimensions Clustering Dimensionality reduction J. Leskovec, A. Rajaraman, J. Ullman:

36 Location-based & Preference-Aware Recommendation Using Sparse Geo-Social Networking Data Jie Bao Yu Zheng Mohamed F. Mokbel Microsoft Research Asia Beijing, China Department of Computer Science &Engineering University of Minnesota

37 Background Loopt Foursquare Facebook Places v Location-based Social Networks Dianping Users share photos, comments or check-ins at a location Expanded rapidly, e.g., Foursquare gets over million check-ins every day

38 Introduction v v Location Recommendations in LBSN Recommend locations using a user s location histories and community opinions Location bridges gap between physical world & social networks Existing Solutions Based on item/user collaborative filtering Similar users gives the similar ratings to similar items users Visit some places User location histories So, what is the PROBLEM here? Build recommendation models Similar Users Similar Items based on the model of co-rating and co-visit Recommendatio n query + user location Why? Mao Ye, Peifeng Yin, Wang-Chien Lee: Location recommendation for location-based social networks. GIS00 stin J. Levandoski, Mohamed Sarwat, Ahmed Eldawy, and Mohamed F. Mokbel: LARS: A Location-Aware Recommender System. ICDE0

39 Motivation (/) v User-item rating/visiting matrix Millions of locations around the world Los Angeles New York City User U 0 U i L L L L m- L m- L m A user visit ~00 locations User location histories are locally clustered U j U n Recommendation queries target an area (very specific subset) Noulas, S. Scellato, C Mascolo and M Pontil An Empirical Study of Geographic User Activity Patterns in Foursquare (ICWSM 0).

40 Motivation (/) v User s activities are very limited in distant locations May NOT get any recommendations in some areas Things can get worse in NEW Areas (small cities and abroad) (Where you need recommendations the most)

41 Key Components in Location Recommendation. User position & locations around Movie Food Shopping. User Personal Interests/Preferences Recommender System. Social/Community Opinions

42 Our Main Ideas Movie Food Shopping User Personal Interests/Preferences Main idea #: Identify user preference using semantic information from the location history User position & locations around Main idea #: Use local experts & user preferences for recommendation Social/Community Opinions Main idea #: Discover local experts for different categories in a specific area

43 Offline Modeling User preferences discovery Movie Food Shopping User Personal Interests/Preferences Main idea #: Identify user preference using semantic information from the location history User position & locations around Main idea #: Use local experts & user preferences for recommendation Social/Community Opinions Main idea #: Discover local experts for different categories in a specific area

44 Users User preference discovery (/) Our Solution v A natural way to express a user s preference Check-ins Venues Categories E.g., Jie likes shopping, football... User preferences is not that spatial-aware. User preferences is more semantic v Can we extract such preferences from user locations? YES!.. Map Category Hierarchy (a) Overview of a location-based social network Category Name Number of sub-categories Arts & Entertainment 7 College & University Food 78 Great Outdoors 8 Home, Work, Other Nightlife Spot 0 Shop Travel Spot (b) Detailed location category hierarchy in FourSquare Millions of locations Hundreds of categories AND NOT limited only to the residence areas

45 User preference discovery (/) Weighted Category Hierarchy v User preferences discovery Location history Semantic information User preference hierarchy Use TF-IDF approach to minimize the bias Food Sport Pizza Coffee Bar Soccer

46 Offline Modeling (/) Social Knowledge Learning Movie Food Shopping User Personal Interests/Preferences Main idea #: Identify user preference using semantic information from the location history User position & locations around Main idea #: Use local experts & user preferences for recommendation Social/Community Opinions Main idea #: Discover local experts for different categories in a specific area

47 Offline Modeling (/) Social Knowledge Learning v Why local experts High quality Less number (Efficiency) v How to discover local experts Local knowledge (in an area) Speciality (in a category) Mutual Inference (HITS) User hub nodes Location authority nodes

48 Online Recommendation Movie Food Shopping User Personal Interests/Preferences Main idea #: Identify user preference using semantic information from the location history User position & locations around Main idea #: Use local experts & user preferences for recommendation Social/Community Opinions Main idea #: Discover local experts for different categories in a specific area

49 Online Recommendations (/)Candidate Selection v Select the candidate locations and local experts Food Sport Pizza Coffee Bar Soccer Candidate Local Experts More local experts are selected for the more preferred category

50 Online Recommendations (/) Location Rating Inference v Similarity Computing Overlaps: Different weights for different levels Diversity of user preferences Based on entropy theory c 0. c 0. c 0. c 0. c 0. c 0. c 0. c 0. c c 6 c 0 c c 6 c 8 c c 6 c 7 c 8 c c v (a) WCH of u (b) WCH of u (c) WCH of u Infer the ratings for the candidate locations H(u, l) is user u s entropy at level l P(c) is the probability that u visited

51 Experiments Data Set v v Data Sets 9,06 users and,8 tips in New York City (NYC), users and 0,78 tips in Los Angels (LA). Statistics v Visualization

52 Evaluation Framework v Evaluation Method v Evaluation Metrics

53 Experimental Results

54 Experimental Results v Efficiency

55 Conclusion v Location Recommendations Data sparsity is a big challenge in recommendation systems Location-awareness amplify the data sparsity challenge v Our Solution Take advantage of category information to overcome the sparsity Using the knowledge from the local experts Dynamically select the local experts for recommendation based on user location

56

57 CF: Common Practice r Before: xi = j N ( i; x) s j N ( i; x) ij s r ij xj v Define similarity s ij of items i and j v Select k nearest neighbors N(i; x) Items most similar to i, that were rated by x v Estimate rating r xi as the weighted average: r xi = b xi + j N ( i; x) s ij ( r j N ( i; x) xj s ij b xj ) baseline estimate μ = overall mean movie ra0ng for r xi b x = ra0ng devia0on of user x = (avg. ra'ng of user x) μ b i = ra0ng devia0on of movie i 7