Scalable and feasible learning and retrieval from matrix data

Abstract

Matrix data are commonly found in AI applications. They differ from traditional tables in relational database, as they only contain numerical values and the interaction between matrices involves linear algebraic operations instead of relational operations. Matrix data management plays an important role in the age of big data and artificial intelligence, as it is directly related to the performance of many AI systems. In this thesis, we develop a number of techniques that improve the scalability and accuracy of two AI applications: recommender systems and knowledge bases.

We first conduct an experimental study which demonstrates that the bottleneck of matrix factorization in recommender systems is the retrieval phase. Based on this observation, we design an exact inner product retrieval framework FEXIPRO, which can retrieve top-k results from learned factors extremely fast. Then, we extend the inner product retrieval problem to a multi-matrix product retrieval problem, which is related to the evaluation of knowledge base completion. For this problem, we present a sampling based evaluation framework WedgeEval, which can give a fast estimation of model performance. Lastly, we investigate the cross-domain application of sequential recommendation. We present CTransRec, which adapts ideas from knowledge base completion and utilizes auxiliary information to improve the quality of sequential recommendation.

The frameworks presented in this thesis not only assist researchers and developers in fast model selection and hyper-parameter tuning, but also help the systems give quick and accurate feedback to users' queries. Although the learning and retrieval algorithms that we propose in this thesis mainly focus on two applications, recommender systems and knowledge bases, they can be easily extended to apply in other applications where similar search operations exist.