Research On Personalization Recommendations In Large Scale Internet Service Systems

With the development of science and technology,the Internet has influenced our daily life.However,Internet service systems are facing an information overload problem.A personalized recommendation system is a key method to solve it.Nowadays,a real world personalized recommendation system mainly faces the following three problems: user access records from mobile devices are difficult to be identified and categorized;item-based recommendations are affected by short-term hotspots;it is hard to develop a new collaborative filtering recommendation which is suitable for various applications.The first problem is at the user level.Data preprocessing is an important step in data mining and machine learning.Nowadays,mobile phones have overtaken desktop computers as the most widely used devices.However,it is difficult to identify and group users' mobile access records.To solve the problem of mobile access record resolution,this thesis proposes a graph-based parallel algorithm,which utilizes computing clusters to efficiently and effectively identify and group more than two billion access records.The second problem is at the item level.The existing recommendation algorithms rarely consider the influence of time series factors.Social hotspots have deeply affected recommendations.Therefore,short-term forecasts are even more significant.To solve the short-term personalized recommendation problem for hotspots,this thesis constructs a deep neural network for item representation learning for short-term activities.It uses short-term time series information and inherent features to obtain short-term item representations.Based on results,systems can predict and recommend for multiple tasks according to the current hotspot.The third problem is at the user-item interaction level.Plenty of works have studied on personalized recommendation algorithms.It is difficult to improve the quality of personalized recommendations for different tasks through a single algorithm.To develop a universal and transferable algorithm to improve personalized recommendation,this thesis proposes a collaborative filtering framework based on user-item subgroups.The algorithm utilizes multiclass co-clustering to discover user-item subgroups.Most existing collaborative filtering recommendation algorithms can be improved by our method.All experiments in this thesis use real word industrial data sets to enhance the credibility of our algorithmic results.A large number of experimental results show that all three problems have been solved effectively and efficiently.The personalization recommendations in large scale Internet service systems have been improved.