Research On Time-Evolving Based Machine Learning Methods

In classical machine learning setting,it always assumes that data distribution is static.However,in many applications,such as recommender systems,data distribution usually changes along the time,which is also called as time-evolving phenomenon.Traditional machine learning methods are not able to take time-evolving property into account,thus modeling an inaccuracy distribution and making imprecise prediction.As the amount of data continues to grow and the data environment becomes compli-cated,traditional methods pose great weakness to time-evolving problems.More and more researchers focus on how to consider the time-evolving property well to improve prediction precision.This thesis focuses on online-shopping,which is a complex data environment,and tries to take time-evolving property into account.Feature processing and space embed-ding are adopted to transform data dynamics to static state.In this way,we can not only model time-evolving property well,but also ease the difficulty of training a time-evolving based model.Main innovative contributions of this thesis can be summarized as follows:1.We propose a time-evolving based recommender methods with time window based features.This method adopts several time windows with varying width to capture time-evolving features of customers and items from different granularity.Powerful ensemble method is applied to learn these static features,thus building a time-evolving based recommender system.Experiments conducted in a real online-shopping system provided by Tencent shows that our proposed method performs better than traditional methods in recommender precision.2.We propose a life-stage evolving modeling algorithm.This method uses only customers' behavior history to learn a low dimensional customer-manifold space to align stages.And in this static stage space,it would be easier to capture complicated stage dynamics and predict future behavior evolution.Visualization results show that customers' evolving preferences are well embedded and empirical results on a real world data set,which is provided by Tencent demonstrate that life stage modeling is effective in the recommender scenario in comparison with a few baselines.