EAGER: Data Management Systems Support for Personalized Recommendation Applications

点击翻译按钮获取中文摘要

A recommender system helps users to identify useful and interesting items from a considerably large search space. Recommender systems have been widely used in various commercial services. A recommender system exploits the users' opinions in order to extract a set of interesting items for each user. This project conducts research, develop requisite knowledge and build software infrastructure to support efficient, salable, and usable data management for personalized recommendation applications. Recommender systems have already been widely used with a strong broad impact on all web users and the project aims to take personalized recommendation applications recommender systems to its next stage and widening its scope to new applications. The project enhances the research infrastructure by distributing a free and portable software artifact. All proposed ideas will be realized inside an open-source recommendation-aware database system maintained at Arizona State University. It is envisioned that the proposed system will be used by researchers world wide as a vehicle for evaluating their research and exchanging new modules related to recommender systems. It is also envisioned that several commercial database systems will adopt the ideas from this project. The project will have a significant educational component. Researchers in both data management and recommender systems will be trained through the proposed project, through curricular innovations as well as workshops and tutorials. Students will be introduced to career pathways through their participations in research.The project tackles the following system challenges to support recommendation applications: (1) Flexibility and Usability: The user should be able to declaratively define a variety of recommenders using popular recommendation algorithms that fit the application needs. The system must be able to integrate the recommendation functionality with other data attributes/sources as well as performing the recommendation functionality and other data access operations side by side. (2) Efficiency and scalability: The system is expected to produce personalized recommendations to a high number of users concurrently over a large pool of items. Unfortunately, recommender models are not easily updatable, and hence they are rebuilt periodically. As a result, the model loses its accuracy over time till the next rebuild process. This is not acceptable in modern applications (e.g., social media) where new items and ratings are streaming into the system. To tackle these challenges, the project injects the recommendation functionality inside the core functionality of a database system by: (a) indexing the set of recommenders to efficiently answer of ad-hoc recommendation queries, (b) encapsulating the recommendation functionality inside a pipeline-able query operator that integrates well with other database operators, and designing query optimization techniques that include the recommendation functionality. Moreover, since a common operation to train recommendation models is to factorize multi-relational user, item, and attribute data, in the forms of tensors, this proposal develops a scalable parallelizable data processing software framework that provides co-optimization of tensor-algebraic and relational algebraic operations. The project also leverages database systems to support context (e.g., spatial location and social network)-aware recommendations.