Personalized Recommendation Through User Behavior Relationship Mining

Recommender system is one of the most popular research problems in the field of data mining.It becomes an urgent matter that how to mine the personalized behavior from the large-scale user behavior data,which is high-dimensional and sparse,and employ it to generate personalized recommendation with high accuracy for users in real time.Generally,the performance of recommender system is heavily dependent on the mined behavior relationship.A useful behavior relationship can greatly uncover the underlying information of observed user historical data,and also can clearly improve the performance of recommender system.This dissertation is based on deep neural network and link prediction,and concentrates on mining meaningful behavior relationship to bridge the relationships between low-level observed data and high-level user behavior relationship for effective recommendations.Moreover,the users of recommender system is able to be categorized into three kinds,namely active users,inactive users and unregistered users.Due to the difference in interaction,recommender system should take different recommendation strategies for different kinds of users.Therefore,it is critical to mine useful behavior relationship from different user behavior data for different recommendation strategies.To solve this problem,this dissertation proposes some novel personalized recommendation models and algorithms which are performed on different recommendation tasks showing their superior results.Specifically,the major research innovations of this dissertation are summarized as follows:(1)To solve the problem of capturing feature interaction relationship in click through rate prediction,Deep FM is proposed in this dissertation.Specifically,Deep FM is an endto-end framework,combines the power of factorization machines(FM)for recommendation and deep learning for feature learning in a novel neural network architecture.To make the feature representation of deep component more accurate,we design the embedding sharing strategy.Specifically,the low-rank vector for 2-order feature interactions of FM is shared with the embedding layer of deep component.Compared with the related models,Deep FM has following advantages: no need for pre-training;no need for feature engineering;capturing low-and high-order feature interactions simultaneously.As a general learning framework,Deep FM is able to incorporate various network architectures in its deep component.This dissertation studies two instances of Deep FM where its “deep” component is DNN and PNN respectively,for which we denote as Deep FM-D and Deep FM-P.Comprehensive experiments are conducted to demonstrate the effectiveness of Deep FM models over the state-of-the-art models for CTR prediction,on both benchmark data and commercial data.We conduct online A/B test in Huawei App Market,which reveals that Deep FM-D leads to higher click-through-rate in the production environment,compared with a well-engineered LR model.(2)To capture the partial-order relationship from the observed user historical behavior data in learning to rank problem,we propose Deep-BLM.Specifically,we use a neural network to score samples and model the pairwise preference of samples relying on their scores under a Bayesian-Personlized-Rank framework.A gradient approach is adopted to minimize the cross-entropy between the predicted probability and the groundtruth.It should be noted that the models based on pairwise preference learning can learn the partial-order relationship from both explicit and implicit feedback.Compared with the Beyesian linear model BLM-Rank,Deep-BLM is better at learning the partial-order relationship hidden in the data with more complex distribution.Moreover,we propose the parallel algorithm based on GPU for accelerating the procedure of training and testing.The experiment results on benchmark datasets demonstrate that Deep-BLM achieves higher performance in terms of NDCG than compared models.(3)To solve the problem of target user mining in push recommendation,we explore various algorithmic choices for mining target user group,and highlight one based on recent advance in graph mining,the Partially Absorbing Random Walk(namely PARW).Particularly,we propose and implement an approximate partially absorbing random walk algorithm(A-PARW)for both single server and distributed cluster that can support very large-scale problems and can efficiently respond to a multitude of push services simultaneously.Moreover,A-PARW is able to capture community structure when the regularized parameter is set as an identity matrix,which is named as A-PARW-I.Giving several nodes as seed,A-PARW-I prefer to capture the related nodes in the community where the seed is located,rather than tend to high degree nodes.Based on public dataset and commercial dataset,we conduct off-line and on-line experiments to verify the performance of A-PARW-I for push recommendation.(4)To solve the problem of session-based recommendation,we propose a contextual K nearest neighbor(CKNN)algorithm,namely CKNN-DSM-EPCSR.The KNN approach is widely used in recommender systems because of its efficiency,robustness and interpretability.In order to enhance the performance of CKNN,we conduct two extensions.At first,we propose the diffusion-based similarity method(DSM)to utilize the graph structure of session-item bipartite network,two parameters ? and ? are introduced to control the influence of session length and item popularity,respectively.Then,we propose the candidate selection method EPCSR for balancing the influence of different relevant historical items in current session.Incorporating EPCSR and DSM in our CKNN algorithm,we propose CKNN-DSM-EPCSR for session-based recommendation.Extensive experiments are performed on different publicly session datasets.The experiment results demonstrate that our algorithm outperforms other KNN approaches for sessionbased recommendation.