Research On Online C Ollaborative Filtering Recommender Algorithms

With the rapid development of mobile Internet,the exponential growth of information intensifies the occurrence of "Information Labyrinth".How to provide users with information and services,which are most likely to be interested in on the limited display page,is the core for current commercial application.Recommendation system,as an effective tool to solve the "Information Labyrinth",has been widely used in various business fields,such as Alibaba's Taobao,Google News,Netflix and Coursera.And Recommendation system has achieved great commercial value in business fields.Traditional collaborative filtering recommendation algorithm,as one of the most widely used technologies in recommender system,has achieved great success.But there are some drawbacks in traditional collaborative filtering.For example,when new training data arrives,traditional collaborative filtering algorithms work in a batch learning or offline learning fashion which results in low efficiency and updates not in time.When traditional collaborative filtering algorithms learn from continuous data streams,they are poor scalable for large-scale applications because the models often have to be re-trained from scratch for new training data.Online collaborative filtering algorithm overcomes the drawbacks in that the model can be updated instantly and efficiently by an online learner when new training data arrives.Based on the online collaborative filtering algorithm,this dissertation proposed three improved online collaborative filtering algorithms from three aspects:traditional machine learning,deep learning in Euclidean space and deep learning in non-Euclidean space(hyperbolic space).The main research work of this dissertation is outlined as follows:1)Concerning the change of user preferences and item popularities of the online collaborative algorithm,this dissertation proposed two inproved methods of online collaborative filtering from traditional machine learning perspective,including online collaborative filtering with dynamic regularization(OCF-DR)and online collaborative filtering with dynamic regularization and neighborhood factor(OCF-DRNF).Firstly,OCF-DR added the user dynamic average rating regularization and user bias regularization and item bias regularization.In addition,OCF-DR updated the weights of the user feature vector and item feature vector in each round.Secondly,this dissertation took the neighborhood factor into account to track the change of user preferences.The related experiments show that performance of OCF-DRNF is better than baselines in low dimension features space.Meanwhile,we can find that OCF-DR and OCF-DRNF converge faster than all the baseline approaches,especially when the algorithms started running.2)Concerning the "cold start" and data sparsity of online collaborative filtering algorithm,this dissertation proposed two inproved methods of online collaborative filtering from Euclidean deep learning perspective,including a deep bias probabilistic matrix factorization algorithm(OCF-DBPMF)and a deep constrain bias probabilistic matrix factorization algorithm(OCF-DCBPMF).Firstly,OCF-DBPMF utilized the convolutional neural network to extract latent user/item features and adding the bias into probabilistic matrix factorization to track user rating behavior and item popularity.Secondly,to solve the problem of the users or items feature vectors will close to the prior mean in the process of training,we constrain user-specific and item-specific feature vectors.The related experiments show that performance of OCF-DBPMF and OCF-DCBPMF are better than baselines in low dimension features space,especially when latent features' dimension is less than 8.3)Concerning the latent hierarchical information of knowledge and the real time of students' learning data in educational application,this dissertation proposed two inproved methods of online collaborative filtering from hyperbolic deep learning perspective,including online collaborative filtering based on hyperbolic deep knowledge tracing(OCF-HDKT)and improved OCF-HDKT algorithm(OCFHDKT+).Firstly,OCF-HDKT utilized hyperbolic gate recurrent unit(Hyperbolic GRU)to build hyperbolic deep knowledge tracing algorithm(HDKT).Secondly,considering the real-time of learning resource recommendation in education scenario,this dissertation proposed to combine the HDKT and online collaborative filtering for embedding the latent hierarchical information and meeting real-time requirements of elearning.The related experiments show that performance of OCF-HDKT+is better than other baselines,especially in terms of model prediction stability and consistency.In general,this dissertation proposed three improved online collaborative filtering algorithms from three aspects:traditional machine learning,deep learning in Euclidean space and deep learning in hyperbolic space.Numerous comparison experiments have proved the efficiency,as well as the guiding significance and application value,of the proposed solutions and algorithms.