Introduction to Recommender Systems

Transcription

1 1era. Jornada de Inteligencia Geoespacial Colectiva Introduction to Recommender Systems Luis Terán Information Systems Research Group University of Fribourg

2 Outline Introduction Something about Recommender Systems Challenges of RS Collaborative Filtering RS Approaches Taxonomy for the development of RSs SmartParticipation project

3 Introduction Finding information in the World Wide Web General Public Internet Data Not trained Hard to please Almost infinite Non structured Evolving Not an easy task! Heterogeneous Multilingual

4 Introduction Internet exponential growing Global monthly Internet traffic 21 exabytes (21 billion G bytes) World s total digital content 500 exabytes (May 2009) Source: netcraft.com [16]

5 Introduction How to deal with information overload? Search Engines Recommender Systems Semantic Web

6 Introduction Search Engines answers related to RS despite powerful search engines No guaranty to find the best answer

7 Introduction Why Recommender Systems? Web is leaving the era of search and entering the one of discovery.

8 Something About Recommender Systems Recommender Systems for ecommerce RS are computer-based techniques used to reduce information overload and to provide recommendations of products likely to interest a user given some information about the user s profile. The more widely used techniques in recommender systems are based on Collaborative Filtering (CF) methods [1]. Item A Item A Item A? Item B

9 why service providers may want to exploit this technology? Increase the number of items sold Sell more diverse items Increase the user satisfaction Increase user fidelity Better understanding of what the user wants

10 Recommender Systems Applications Entertainment recommendations for movies, music, games, and IPTV. Content personalized newspapers, recommendation for documents, recommendations of webpages, e-learning applications, and filters. E-commerce recommendations of products to buy such as books, cameras, PCs, etc. for consumers. Services recommendations of travel services, recommendation of experts for consultation, recommendation of houses to rent, or matchmaking services. Social recommendation of people in social networks, and recommendations of content social media content such as tweets, Facebook feeds, LinkedIn updates, and others (VAAs).

11 The value of recommendations Netflix: 2/3 of the movies watched are recommended Google News: recommendations generate 38% more clickthroughs Amazon: 35% sales from recommendations Choicestream: 28% of the people would buy more music if they found what they liked. Source: Alexandros Karatzoglou

12 What is a good recommendation? Personalized -> relevant for a user. Diverse -> shows all diverse interests of a user. Recommends unknown items. Expands user tests into new areas -> Serendipity (pleasant surprise)

13 Recommender Systems for ecommerce Problems and goals of RS Quality of Recommendations: reliable information, minimize the number of false positive results Sparsity: number of items rated is small compared to the number of items, which leads to a week recommendations Scalability: increasing the number of users and products increases complexity Lost of Neighbor Transitivity: correlations can not be expressed unless they have rated common items. Synonymy: RS can not link products with different names but belongs to the same type of products. First Rater Problem: a product can not be recommended unless another customer has rated it in previously. Unusual User Problem: refers to users which can not define their opinion about a product, this will cause inconsistent recommendations.

14 Challenges of Recommender Systems Collecting Opinion and Experience Data Finding the Relevant Data for a Purpose Presenting the Data in a Useful Way

15 To recommend, we need to data (users likes, etc.) Basic Concepts of RS Data comes from users -> collected somehow. We can observe: - What users tell us (ratings), What users do (actions) - These are noisy measurements of preference Preferences Rating Review Vote Explicit Implicit Click Purchase Follow

16 Twitter Mute Twitter: Learn how mute, block and report can help you take control of your experience on Twitter. Source: ?lang=en Fed up with your friend's annoying updates? Twitter launches mute button to temporarily silence annoying tweeters. Source: Twitter-launches-mute-button-annoying-tweeters.html

17 RS Approaches Content-based Recommendations Collaborative Filtering (mostly used) Demographic (Context-aware Recommendations) Knowledge-based - Case-based/constraint-based (based on specific domain knowledge about how certain item features meet users needs) - Community-based (based on the preferences of the user s friends) Hybrid Recommender Systems

18 Recommender Systems Framework Consists of 8 Dimensions of Analysis - Domain - Purpose - Recommendation Context - Whose Opinions - Personalization Level - Privacy and Trustworthiness - Interfaces - Recommendation Algorithms

19 Dimension I - Domains Content to commerce and beyond - News, information - Products - Matchmaking (meet people) - Sequences (e.g., music playlists, videos on YouTube) One particularly interesting property - New items (e.g., movies, books,...) - Re-recommend old ones (e.g., groceries, music)

20

21

22 Dimension II - Purposes The recommendations themselves - Increase sales - Provide information Provide information to user/customer Build a community of users/customers around products or content

23

24

25 Dimension III - Context What is the user doing at the time of recommendation - Browsing eshop, shopping - Listening to Music - Social networks How does the context constrain RS? - Groups, automatic consumption (vs. suggestion), level of attention, level of interruption, etc.

26

27

28 Dimension IV - Whos Opinion? Experts Ordinary people People like you

29

30 Dimension V - Personalization Generic / Non-Personalized - Same recommendations for everyone Demographic - Matches a target group (age, region, etc.) Ephemeral - Matches current activity Persistent - Matches long-term interests

31

32 Dimension VI - Privacy and Trustworthiness Who knows what about me? - Personal information - Identity - Deniability of preferences Is the recommendation honest? - Biases built-in by operator business rules - Vulnerability to external manipulation - Transparency of recommenders -> reputation

33 Dimension VII - Interfaces Types of Output - Predictions - Recommendations - Filtering - Organic vs. explicit presentation Agent/Discussion Interface Types of Input - Explicit - Implicit

34 Dimension VIII - Algorithms Non-Personalized Collaborative Filtering Content-based Recommendations Context-aware Recommendations Other Approaches Hybrid Recommender Systems

35 User-User Collaborative Filtering I1 I2 I3 I4 I5 I6 Find Items rated by active user (au) Find other users that rated same items -> neighbourhood formation U1 U2 U3 U4 U5 (au) U

36 User-User Collaborative Filtering Compute how similarity of au compared to other users Select n closest neighbours (if possible) au Objectives Estimate missing rankings (prediction) Recommend items with bigger ratings (recommendation)

37 Item-Item Collaborative Filtering I1 I2 I3 I4 I5 I6 (Ti) Identify target item (Ti) Find set of users who rated the target item Identify which other items (neighbours) were rated by the users set U1 U2 U3 U4 U5 U

38 Item-Item Collaborative Filtering Compute similarity between each neighbour & target item (similarity function) In case, select k most similar neighbours Predict ratings for the target item (prediction function) Ti

39 SmartParticipation - Motivation Citizen Society Community Building User specifies himself as Green (profile) How to find Green citizens? 49

40 (A.1) Sharp vs. Fuzzy Set ϰ Sharp set Fuzzy Sets: Assuming U = {x1, x2,..., xn} as the universe of discourse, then a fuzzy set A (A U) is defined as a set of ordered pairs: {(xi, μa(xi))} where xi U, μa : U [0,1] is the membership function of A and μa(x) [0,1] is the degree of membership of x in A. μ Fuzzy set 51

41 (A.3) Sharp vs. Fuzzy Clustering - I Execution, C-men Algorithm (MacQueen [1967]), in two dimensions Execution, Fuzzy C-Mean Algorithm (Bezdec [1981]), in one dimension Sources: - MacQueen, J.B. (1967). Some Methods for classification and Analysis of Multivariate Observations. In Proceedings of 5th Berkeley Symposium on Mathematical Statistics and Probability, , University of California Press, Berkeley, USA. - Bezdek, J.C. (1981). Pattern Recognition with Fuzzy Objective Function Algorithms. Advanced Applications in Pattern 55 Recognition, Springer US

42 Motivation - SmartParticipation Citizen Society Fuzzy-based Community Building User specifies himself as Green (profile) Fuzzy profiles 59

43 60 SmartParticipation

44 Why Using Fuzzy Clustering on RS? People thinking is not sharp (black and white) Better definition of profiles using membership degrees Center of Political Parties are the center of fuzzy clusters Community creation based on fuzzy clusters 61

45 Increase Citizens Trustworthiness eparticipation Evaluation Framework eempowerment eparticipation ediscussion econsulting einforming 62

46 Current Application Develop a platform to promote discussion and participation for the 2017 Ecuador Elections Integration of the first Dynamic VAA Creation of profiles (voters / candidates) Discussion of political issues Analysis of data Development and improvement of algorithms for recommendations Evaluation 64

47 Director of Project - Luis Terán (UniFR y ESPE) Profiles Generation - Andrea Balda (UCG), Fernando Mendez (Uni Zurich), Iria Puyosa (UDLA), Ivan Rivera (ESPOL) Communication - Iria Puyosa (UDLA), Gabriela Baquerizo (UCG), Daniel Pastor (UCG), Adriana Illingworth, (UCG), Irene Gavilanes (UDLA) Web and Recommender Systems Development - Luis Terán, (UniFR y ESPE), Carmen Vaca (ESPOL), Jonathan Mendieta (ESPOL), Lorena Recalde (UPF), Aigul Kaskina (UniFR), José Mancera (UniFR), Alexander Sosa (ESPE), Jorge Flores (ESPE) 65 Team and Institutions

48 Social Network for discussion and participation Tools Developed Timeline: Twitter Candidate Feeds (users do not require Twitter account) VAA Candidates: President, Vice President, Assembly (coming soon). Content Creation: Articles and Community Fact-Check Profiles: Candidates, users Pi Creation of thematic groups (public and private) Community Content Control User Reputation Recommendations: articles, questions, users, groups, and much more. Statistics: Twitter trends and surveys. 66

49 Current Application Dynamic Profiles include three main elements: Experts + Candidates + Tweeter Feeds Experts: SmartCoding programming (University of Zurich) Candidates: media and communication group Twitter: Dynamic approach with Twitter feeds of candidates

50 SmartParticipation Springer: Amazon: cipation-fuzzy-based- Recommender-Community- Building- Management/dp/ ebay: Participation-A-Fuzzy-based- Recommender-System-for-Political- Community-bu-/