- 微博 QQ QQ空间 贴吧
Coupling Everything: A Universal Guideline for Building State-of-The-Art Recommender Systems
- Introduction (overview, challenges, preliminary)
- RS with User couplings
- RS with Item couplings
- RS with Multifold couplings
- RS with Comprehensive couplings
- Conclusion remarks
1 . Coupling Everything: A Universal Guideline for Building State-of-The-Art Recommender Systems Presenter: Liang Hu, Shoujin Wang and Longbing Cao Authors: Liang Hu, Shoujin Wang, Songlei Jian, Longbing Cao, Jian Cao
2 .Related tutorial • Non-IID Learning of Complex Data and Behaviors (T14) • Non-iid recommender systems • Non-iid customer behaviors • Non-idd relations • Non-iid learning methods • …… • Time : 14:00-18:00, Sunday 11, August 2019 • Venue: L • Presenter: Prof. Longbing Cao
3 .Special issues on recommender systems • IEEE Intelligent Systems Special issue: Intelligent Recommendation with Advanced AI and Learning • Due on September 30, 2019. • Link: https://sites.google.com/view/shoujin-wang/home/activities/intelligent- recommendation-with-advanced-ai-and-learning • International Journal of Data Science and Analytics (JDSA) Special issue: Data Science for Next-Generation Recommender Systems • Due on November 30, 2019. • Link: https://sites.google.com/view/shoujin-wang/home/activities/jdsa- special-issue
4 .Goal Providing a comprehensive understanding of how to apply the state-of-the- art machine learning approaches to model the universal couplings in the advanced recommender systems.
5 .Agenda • Introduction (overview, challenges, preliminary) (45 mins) • RS with User couplings (45 mins + 30 mins break) • RS with Item couplings (45 mins) • RS with Multifold couplings (45 mins) • RS with Comprehensive couplings (45 mins) • Conclusion remarks (10 mins)
6 . Outline Learning Learning Learning Learning Introduction user item multifold comprehensi Conclusion couplings couplings couplings ve couplings • Overview • Social RS • Cross-domain • Multi-modal • Context- • Conclusion • Challenges • Group RS RS RS aware RS • Summary • Preliminaries • Session-based • Multi-criteria • Fashion AI in RS RS RS • Heterogenous couplings in RS
7 . Outline Learning Learning Introduction • Overview user of recommender item Learning systems multifold Learning comprehensi Conclusion couplings couplings couplings ve couplings • Challenges of recommender systems • Couplings in recommender system • Overview • Social RS • Cross-domain • Multi-modal • Context- • Conclusion • Challenges •• Group AI-related RS preliminaries: RS from RS data representation aware RS perspective • Summary • Session-based • Preliminaries • RepresentingRS attributes • Multi-criteria • Fashion AI in RS RS • Representing review • Heterogenous • Representing rating table couplings in RS • Representing image • Representing network • Representing sequence
8 .What Are Recommender Systems • Recommender systems (push information) are the evolution of information retrieval systems (pull information). Information Age Recommendation Age Pull mode (IRS): Query Matched Results Manual Filtering Push mode (RS): Potential Requirement Machine Filtering Recommendation Anderson, C. (2006). The long tail: Why the future of business is selling less of more
9 . Recommender Systems have occupied our life What to eat Which to dress Where to go
10 .Personalized Recommendation
11 .Classic Recommender Systems Recommender Systems Collaborative Content-based Demographic Knowledge- Filtering Filtering RS based RS User-based Item-based
12 . Collaborative Filtering (CF) • Intuition (user-based filtering): If user A A x related to user B and A bought x and y, then B bought x tend to buy y. • Famous examples(item-based filtering): Amazon.com's recommender system B y • Facebook, MySpace, LinkedIn use collaborative filtering to recommend new friends, groups, and other social connections. Sarwar, B., Karypis, G., Konstan, J., & Riedl, J. (2001). Item-based collaborative filtering recommendation algorithms. Paper presented at the Proceedings of the 10th international conference on World Wide Web, Hong Kong, Hong Kong.
13 .Content-based Filtering (CBF) • CBF is based on the features of items • Attributes of items • Description of items • Text of an article User Profile • User profile is built with the features of historical items • Recommend items according to user profile
14 .Data Characteristics in Recommender Systems • Power law or Long tail distribution • Data associated with the majority of users are insufficient and even absent in real world. • In most recommender systems, the majority of users/items only associated with very few data while only minority of users/items have sufficient data
15 .Challenges in Collaborative Filtering • Data Sparsity • In real-world recommender systems, the … user-item matrix is very sparse. • Cold Start 1 4 5 3 • When new users or new items are added, the system cannot recommend to these … 1 2 users and these items. 5 • Scalability 5 • There are millions of users and products in real systems. • Large amount of computation • Large storage
16 .Challenges in Content-based Filtering • Limited Content Analysis • System has a limited amount of information on its users or the content of its items. • Over-specialization • The system can only recommend items that highly similar with user’s profile, the user is limited to be recommended items similar to those already rated.
17 .Question: what’s the main cause of these challenges? • Data Sparsity • Cold Start Insufficiency of data • Limited Content Analysis • Over-specialization
18 .Data complexity challenges existing theories and systems What Ad Irrelevant and would Damaging to Brand you place here?
19 .Coupling in Complex Data • Coupling relationships: • Within and between values, attributes, objects, sources, aspects, … • Structures, distributions, relations, … • Methods, models, … • Outcomes, impact, … Longbing Cao. Non-IIDness Learning in Behavioral and Social Data, The Computer Journal, 57(9): 1358-1370 (2014).
20 . Address the RS from multi-coupling perspectives A. Coupling on users: • Social RS: user mutual influence • Group RS: group joint decision B. Coupling on items: • Cross-domain RS: domain coupling • Session-based RS: temporal coupling C. Coupling on implicit interaction: • Context-aware RS: contextual dependency • Multi-objective RS: multi-aspect ratings • Attraction RS: subjective attention Cao, L. (2016). Non-IID Recommender Systems: A Review and Framework of Recommendation Paradigm Shifting. Engineering, 2(2), 212-224.
21 .Coupling modeling in advanced RS • Coupling modeling : • The coupling between users social RS, group-based RS • The coupling between items session-based RS, cross-domain RS • The coupling between data types multi-modal RS • The coupling between domains cross-domain RS • The coupling between objectives multi-objective RS
22 . Universal coupling modeling in this tutorial Universal coupling RS User coupling Item coupling Multifold coupling Cross- Session- Multimodal Multi- Social RS Group RS domain RS based RS RS objective RS
23 . Outline Learning Learning Introduction • Overview user of recommender item Learning systems multifold Learning comprehensi Conclusion couplings couplings couplings ve couplings • Challenges of recommender systems • Couplings in recommender system • Overview • Social RS • Cross-domain • Multi-modal • Context- • Conclusion • Challenges •• Group AI-related RS preliminaries: RS from RS data representation aware RS perspective • Summary • Session-based • Preliminaries • RepresentingRS attributes • Multi-criteria • Fashion AI in RS RS • Representing review • Heterogenous • Representing rating table couplings in RS • Representing image • Representing network • Representing sequence
24 .AI, Machine Learning and Deep Learning
25 .Machine Learning Methods Dominate RS Competitions Alibaba Competitions
26 .Prospects: modeling RS with more complex data • Built on More Complex Data • Multiple data types • Ratings • Images Data complexity • Text • Multisource • Multiple domains • Multiple systems • Social data • Acquire data from user social media • Multiple criteria • Multi-objectives: accuracy, novelty…
27 .Machine Learning: Tell Truths from Data Recommender Systems: Recommend Truths from Data Data with Machine Learning Methods Data with Recommender Systems • Attributes • User/Item features • Regression • Clustering • Factor Analysis • Labels • Ratings • Classification • Learning to Rank • Images, Videos • Item pictures • Computer Vision Approach
28 .Machine Learning: Tell Truths from Data (Cont.) Recommender Systems: Recommend Truths from Data Data with Machine Learning Methods Data with Recommender Systems • Text • Reviews • Natural Language Processing (NLP) • Sentiment Analysis • Sequence • Transaction • Time Series Analysis • Network • Link Prediction, Network Embedding • User/Item Network
29 .In a word • Data is the matchmaker to bring machine learning to recommender systems