Download as PDF, PPTX
![§ This actually means to take the user’s
history hp and compare it to rows of the co-
occurrence matrix (PtP)
§ TF-IDF weighting of co-occurrence would
be nice to mitigate the undue influence
of popular items
§ Find items nearest to the user’s history
§ Sort these by similarity strength and
keep only the highest —you have
recommendations
§ hp
§ user1: [item2, item3]
§ (PtP)
§ item1: [item2, item3]
§ item2: [item1, item3, item95]
§ item3: […]
§ find item that most closely matches the
user’s history
§ item1 !
25.08.2017Recommender Systems 30](https://image.slidesharecdn.com/recommender-170720152835/75/Recommendation-Systems-30-2048.jpg)
Uploaded byCenk Bircanoğlu
Recommendation Systems
- Recommender systems use algorithms to predict items (e.g. movies, music, news articles) that users may like based on their preferences and behaviors. Common techniques include collaborative filtering, content-based filtering, and hybrid approaches. - Factors like ratings data, item metadata, user demographics, and social networks can be used as inputs to recommender algorithms like matrix factorization, deep learning, and similarity metrics. - The effectiveness of recommender systems is demonstrated by services like Netflix, where 2/3 of movies watched are recommended, and Amazon, where 35% of sales come from recommendations.
![[系列活動] 人工智慧與機器學習在推薦系統上的應用](https://cdn.slidesharecdn.com/ss_thumbnails/merged-161217165734-thumbnail.jpg?width=640&height=640&fit=bounds)
Recommendation Systems
- 2. § Netflix: 2/3of the movies watched are recommended § Google News: recommendations generate 38% more clickthrough § Amazon: 35% sales from recommendations § Choicestream: 28% of the people would buy more music if they found what they liked. 25.08.2017Recommender Systems 2
- 3. § In search,user know about item § In recommendation, user even do not know whether item exists or not 25.08.2017Recommender Systems 3
- 4. § Content BasedFiltering § Collabrative Filtering 25.08.2017Recommender Systems 4
- 5. § Learning toRank § Context-Aware Recommendations § Tensor Factorization § Factorization Machines § Deep Learning § Similarity § Social Approaches 25.08.2017Recommender Systems 5
- 6. § Collaborative Filtering:Recommend items based only on the users past behavior § User-based: Find similar users to me and recommend what they liked § Item-based: Find similar items to those that I have previously liked § Content-based: Recommend based on item features § Personalized Learning to Rank:Treat recommendation as a ranking problem § Demographic: Recommend based on user features § Social recommendations (trust-based) § Hybrid: Combine any of the above 25.08.2017Recommender Systems 6
- 7. 25.08.2017Recommender Systems 7 Item1Item2 Item3 Item4 Item5 Alice 5 3 4 4 ? User1 3 1 2 3 3 User2 4 3 4 3 5 User3 3 3 1 5 4 User4 1 5 5 2 1 USER-MOVIE RATINGS
- 8. § List ofm Users and a list of n Items § Each user has a list of items with associated opinion § Explicit opinion - a rating score § Sometime the rating is implicitly – purchase records or listen to tracks § Active user for whom the CF prediction task is performed § Metric for measuring similarity between users § Method for selecting a subset of neighbors § Method for predicting a rating for items not currently rated by the active user. 25.08.2017Recommender Systems 8
- 9. § Identify setof ratings for the target/active user § Identify set of users most similar to the target/active user according to a similarity function (neighborhood formation) § Identify the products these similar users liked § Generate a prediction - rating that would be given by the target user to the product - for each one of these products § Based on this predicted rating recommend a set of top N products 25.08.2017Recommender Systems 9
- 10. § Pros: § Requiresminimal knowledge engineering efforts § Users and products are symbols without any internal structure or characteristics § Produces good-enough results in most cases § Cons: § Requires a large number of reliable “user feedback data points” to bootstrap § Requires products to be standardized (users should have bought exactly the same product) § Assumes that prior behavior determines current behavior without taking into account “contextual” knowledge (session-level) 25.08.2017Recommender Systems 10
- 11. § CF recommendationsare personalized since the “prediction” is based on the ratings expressed by similar users § Those neighbors are different for each target user § A non-personalized collaborative-based recommendation can be generated by averaging the recommendations of ALL the users 25.08.2017Recommender Systems 11
- 12.
- 13.
- 14.
- 15. 25.08.2017Recommender Systems 15 Item1Item2 Item3 Item4 Item5 Alice 5 3 4 4 ? User1 3 1 2 3 3 User2 4 3 4 3 5 User3 3 3 1 5 4 User4 1 5 5 2 1 sim = 0,85 sim = 0,70 sim = -0,79
- 16.
- 17.
- 18. 25.08.2017Recommender Systems 18 Item1Item2 Item3 Item4 Item5 Alice 5 3 4 4 ? User1 3 1 2 3 3 User2 4 3 4 3 5 User3 3 3 1 5 4 User4 1 5 5 2 1
- 19. § Methods ofdimensionality reduction § Matrix Factorization § Clustering § Projection (PCA, SVD …) 25.08.2017Recommender Systems 19
- 20. § Memory based §Use the entire user-item database to generate a prediction. § Usage of statistical techniques to find the neighbors – e.g. nearest- neighbor. § Memory based § First develop a model of user § Type of model: § Probabilistic (e.g. Bayesian Network) § Clustering § Rule-based approaches (e.g. Association Rules) § Classification, Regression § LDA 25.08.2017Recommender Systems 20
- 21.
- 22. § Common forrecommending text-based products (web pages, usenet news messages, ) § Items to recommend are “described” by their associated features (e.g. keywords) § User Model structured in a “similar” way as the content: features/keywords more likely to occur in the preferred documents (lazy approach) § Text documents recommended based on a comparison between their content (words appearing) and user model (a set of preferred words) § The user model can also be a classifier based on whatever technique (Neural Networks, Naïve Bayes...) 25.08.2017Recommender Systems 22
- 23. § Let Content(s)be an item profile, i.e. a set of attributes characterizing item s. • Content usually described with keywords. • “Importance” (or “informativeness”) of word kj in document dj is determined with some weighting • measure wij • One of the best-known measures in IR is the term frequency/inverse document frequency (TF- IDF) 25.08.2017Recommender Systems 23
- 24. § Linear Regression §SVD § SVD++ § Funk SVD § Matrix Factorization § Neural Networks § Locality-Sensitive Hashing (LSH) § Clustering § Association rules 25.08.2017Recommender Systems 24
- 25. Vk T Dim1 -0.44 -0.570.06 0.38 0.57 Dim2 0.58 -0.66 0.26 0.18 -0.36 Uk Dim1 Dim2 Alice 0.47 -0.30 Bob -0.44 0.23 Mary 0.70 -0.06 Sue 0.31 0.93 Dim1 Dim2 Dim1 5.63 0 Dim2 0 3.23 T kkkk VUM ´S´= kS • SVD: • Prediction: = 3 + 0.84 = 3.84 )()(ˆ EPLVAliceUrr T kkkuui ´S´+= 25.08.2017Recommender Systems 25
- 26.
- 27. § Personalized recommendations—user-based §Similar items—item-based § Viewed this bought that—item-based cross-action § Popular Items and User-defined ranking § Item-set recommendations for complimentary purchases or shopping carts—item-set-based § Hybrid collaborative filtering and content based recommendations—limited content-based § Business rules 25.08.2017Recommender Systems 27
- 28. r = recommendations hp= a user’s history of some action (purchase for instance) P = the history of all users’ primary action rows are users, columns are items (PtP) = compares column to column using log-likelihood based correlation test 25.08.2017Recommender Systems 28 r = (PtP)hp
- 29. § Let’s call(PtP) an indicator matrix for some primary action like purchase § Rows = items, columns = items, element = similarity/correlation score § The score is row compared to column using a “similarity” or “correlation” metric § Log-Likelihood Ratio (LLR) finds important/correlating co- occurrences and filters out the rest—a major improvement in quality over simple co-occurrence or other similarity metrics. § Experiments on real-world data show LLR is significantly better than other similarity metrics 25.08.2017Recommender Systems 29
- 30. § This actuallymeans to take the user’s history hp and compare it to rows of the co- occurrence matrix (PtP) § TF-IDF weighting of co-occurrence would be nice to mitigate the undue influence of popular items § Find items nearest to the user’s history § Sort these by similarity strength and keep only the highest —you have recommendations § hp § user1: [item2, item3] § (PtP) § item1: [item2, item3] § item2: [item1, item3, item95] § item3: […] § find item that most closely matches the user’s history § item1 ! 25.08.2017Recommender Systems 30
- 31. Single User Historyof Multi-modal Behavior buy views terms in search users products products categories terms ... A B C E input categorypref products D share user-i 25.08.2017Recommender Systems 31
- 32. All User’s BuysCooccurrence users products A users products At X = cooccurrence products products product-j product-j had 2 other products that were bought in common, we replace cooccurrence magnitude with LLR score, it adds the “correlation test” to simple cooccurrence 25.08.2017Recommender Systems 32
- 33.
- 34. How often doitems co-occur? // compute co-occurrence matrix val C = A.t %*% A 25.08.2017Recommender Systems 34
- 35. Which cooccurences areinteresting? // compute some statistics val interactionsPerItem = drmBroadcast(A.colSums) // convert to indicator matrix val I = C.mapBlock() { // compute LLR scores from // cooccurrences and statistics ... // only keep interesting cooccurrences ... } // save indicator matrix I.writeDrm(...); 25.08.2017Recommender Systems 35
- 36.
- 37.
- 38.
- 39. § Virtually allexisting collaborative filtering type recommenders use only one indicator of preference § Virtually anything we know about the user can be used to improve recommendations—purchase, view, category- preference, location-preference, device-preference, gender… 25.08.2017Recommender Systems 39 r = (PtP)hp + (PtV)hv + (PtC)hc + … CROSS-OCCURRENCE
- 40. All User’s BuysCross-occurrence with Search terms users users products At X = cross- occur- rence products product-j product-j had 3 terms that were searched for in common, we replace cross-occurrence magnitude with LLR score, it adds the “correlation test” to simple cross-occurrence! terms in search terms terms E 25.08.2017Recommender Systems 40
- 41. § Comparing thehistory of the primary action to other actions finds actions that lead to the one you want to recommend § Given strong data about user preferences on a general population we can also use § items clicked § terms searched § categories viewed § items shared § people followed § items disliked (yes dislikes may predict likes) § location § device preference § gender § age bracket § Virtually any anything we know about the population can be tested for correlation and used to predict a particular users preferences 25.08.2017Recommender Systems 41
- 42. § Collaborative TopicFiltering § Use Latent Dirichlet Allocation (LDA) to model topics directly from the textual content § Calculate based on Word2Vec type word vectors instead of bag-of-words analysis to boost quality § Create cross-occurrence indicators from topics the user has preferred § Repeat periodically § Entity Preferences: § Use a Named Entity Recognition (NER) system to find entities in textual content § Create cross-occurrence indicators for these entities § Entities and Topics are long lived and richly describe user interests, these are very good for use in the Universal Recommender. 25.08.2017Recommender Systems 42
- 43. § The finalcalculation uses hp as the query on the Cooccurrence Matrix (PtP), returns a ranked set of items § Query is a “similarity” query, not relational or key based fetch § Uses Search Engine as Cosine-based K-Nearest Neighbor (KNN) Engine with norms and TF-IDF weighting § Highly optimized for serving these queries in realtime § Several (Solr, Elasticsearch) have High Availability, massively scalable clustered auto-sharding features like the best of NoSQL DBs. 25.08.2017Recommender Systems 43
- 44. Indicators can alsobe based on content similarity (TTt) is a calculation that compares every 2 documents to each other and finds the most similar—based upon content alone 25.08.2017Recommender Systems 44 r = (TTt)ht + l*L …
- 45. § Cooccurrence § Findthe best indicator of a user preference for the item type to be recommended: examples are “buy”,“read”,“video_watch”, “share”,“follow”,“like”. § Cross-occurrence § Item metadata as “user” preference, for example: treat item category as a user category-preferences § Calculated from user actions on any data that may give information about user— category-preferences, search terms, gender, location § Create with Mahout-Samsara SimilarityAnalysis.cooccurrence 25.08.2017Recommender Systems 45
- 46. § Content ormetadata § Content text, tags, categories, description text, anything describing an item § Create with Mahout-Samsara SimilarityAnalysis.rowSimilarity § Intrinsic § Popularity rank, geo-location, anything describing an item § Some may be derived from usage data like popularity rank, or hotness § Is a known or specially calculated property of the item 25.08.2017Recommender Systems 46
- 47. “Universal” means onequery on all indicators at once Unified query: purchase-correlator: users-history-of-purchases view-correlator: users-history-of-views category-correlator: users-history-of-categories-viewed tags-correlator: users-history-of-purchases geo-location-correlator: users-location 25.08.2017Recommender Systems 47 r = (PtP)hp + (PtV)hv + (PtC)hc + … (TTt)ht + l*L …
- 48.
- 49.
- 50.
- 51.
- 52.
- 53.
- 54.
- 55. § getBiasedSimilarItems § Getsimilar items for an item,these are already in the action correlators in ES § getBoostedMetadata § Get all metadata fields that potentially have boosts (not filters) § getFilteringMetadata § Get all metadata fields that are filters (not boosts) § getFilteringDateRange § Get part of query for dates and date ranges 25.08.2017Recommender Systems 55
- 56. § getExcludedItems § Createa list of item ids that the user has interacted with or are not to be included in recommendations § getBiasedRecentUserActions § Get recent events of the user on items to create the recommendations query from § getExcludingMetadata § Get all metadata fields that are filters (not boosts) 25.08.2017Recommender Systems 56
- 57.
- 58.
Editor's Notes
- #38 Figure 1: Formula used to calculate geospatial statistic (a modified log-likelihood ratio [LLR]) on the basis of geographic distribution of Mycobacterium tuberculosis genotype clusters, Washington, USA. Variables are classified as follows: a = number of tuberculosis (TB) cases with the genotype of interest in the selected county; b = number of cases with the genotype of interest in the United States; c = number of cases without the genotype of interest in the selected county; d = number of cases without the genotype of interest in the United States; N = total number of TB cases.