Researches On Hybrid Collaborative Filtering In Recommendation Algorithm Based On Relational Metric Learning Of Multi-entity Database

Different kinds of data has been accumulated with the great development of internet and information technology. Big data technology has come into being in many fields. As an important technology in data mining area, how to find out the useful information in the target dataset has become a hot issue. Metric learning aims to find out a best metric for a target dataset and measure the similarity between samples. A variety of metric learning approaches are applied to image retrieval, text indexing and facial recognition. Furthermore, A good metric can project the samples onto a better metric space and improve the accuracy of clustering and classification.This thesis mainly contains three parts:1.Metric learning algorithms may have a long learning time and an bad influence on clustering and classification when applied to high dimension dataset. Thus, I propose an acceleration framework for the metric learning for high dimension data in the part to solve this problem in this part. This framework uses Markov blanket and entropy theory to divide the features into groups and we learn low dimension metric based on these features groups, then we combine these low dimension metric into a high dimension metric.2.Metric learning is mostly limited in single dataset and it needs many manually labeled samples. In this part, I propose a sample labeling algorithm by using the information in relational table in relational database. After that, we can use these labeled sample to learn relational metric.Relational metric is a new application of metric learning in relational database and it provides a feasible way for learning metric in database with few labeled sample.3. Most collaborative filtering can not reach a high recommendation accuracy because of the high sparsity of rating scores and the lack of consideration on time characteristics of rating scores as well as the database structure. In this part, I propose an improved hybrid collaborative filtering recommendation system based on relational metric as well as the reuse of early effective rating scores. This recommendation system can better use the recent rating scores,early effective rating scores and database structure to raise the accuracy.