Research On Key Technologies Of Attributes-Driven Context-Aware Recommendation

Recent years have witnessed the rapid development of deep learning techniques.With the promotion of deep learning techniques,recommender systems have achieved great success in many fields such as news,e-commerce,music,movies and advertising,and have become one of the hot topics in both academia and industry.However,from the state-of-the-art recommendation models to the earlier ones which are based on Collaborative Filtering,many severe issues have not been solved so far.Among them,we have investigated three outstanding issues:(1)the inadequate modeling of the long-term preference and the short-term intent,(2)the lack of attributes modeling,and(3)the inadequate modeling of the sequential patterns from user's long history sequence.Therefore,we have studied the key technologies of attributes-driven context-aware recommendation based on these three issues.The accomplishments of our research are listed below.(1)Attributes-driven recommendation model on the prediction of user's contextual intent.In the context-aware recommender systems,the user's intent toward the items is diverse among different sessions and different time steps within one session.However,in earlier works the variation of user's intent has not been fully modeled.To deal with this issue,we proposed an attributes-driven recommendation model on the prediction of user's contextual intent.Firstly,the contexts have been defined as the rating-related attributes and have used as the drive to construct the contextual user profile.We have studied the variation of user's short-term intent within sessions with the attention mechanism and bi-directional long short-term memory(Bi-LSTM).Secondly,we modeled the user's long-term preference with the gated recurrent unit(GRU),and integrated it with the user's short-term intent,to build a more comprehensive contextual user profile and generate an accurate next-item recommendation.The experimental study on two real-world datasets and two sub-datasets demonstrates that the proposed approach outperforms the best baseline ?-RNN.For the Reddit dataset,it improves 9.12%and 22.10%in Recall@5 and MRR@5 scores,respectively;for the Last.fm dataset,it improves 4.36%and 7.33%in Recall@5 and MRR@5 scores,respectively.The results demonstrate that,the proposed approach can significantly improve the next-item recommendation,and is capable of coping with the cold-start problem at the beginning of each session.Based on this accomplishment,one SCI paper has been published.(2)Attributes-driven explainable recommendation model on the extraction of latent information for high-order feature interactions.The current recommender systems lack the capability of modeling the high-order feature interactions from attributes and generating the intuitive explainable recommendations,which makes it difficult for the systems to accurately capture the user's general preference and provide convincing recommendations.To deal with this issue,we proposed an attributes-driven explainable recommendation model on the extraction of latent information for high-order feature interactions.We regarded user and item attributes as features which have been used as the drive to model the latent information from high-order feature interactions and user history behavior,and have constructed two attention-based neural layers from the macro and micro perspectives,to predict the probability of user clicking the candidate item and provide explainable recommendations with attributes and user history behavior:(1)the macro layer learns the latent information by modeling the low and high-order of feature interactions,extracts user's general preference and interprets the contribution of each feature interaction;(2)the micro layer investigates the different impact which each history interaction has on the candidate item.The experimental study on two real-world datasets demonstrates that the proposed approach outperforms the best baseline OPNN on click-probability which is appropriately measured in terms of AUC and ACC.For the MovieLens-1M dataset,it improves 1.14%and 1.49%in AUC and ACC scores,respectively;for the Taobao dataset,it improves 2.41%and 2.50%in AUC and ACC scores,respectively;for the JD dataset,it improves 0.49%and 0.43%in AUC and ACC scores,respectively.The results demonstrate considerable improvements on the click-through rate prediction.By plotting two types of heat map as the explanations for the recommendation,the approach interprets the contribution of each feature interaction and different impact which each history interaction has on the candidate item.Based on this accomplishment,one SCI paper has been published.(3)Attributes-driven recommendation model on the extraction and distillation of sequential latent information for user history behavior.The current recommender systems have been faced with the issues of long-distance dependency and noise when modeling user-item interaction sequences and extracting the sequential patterns.In addition,the complexity of current recommendation models'architectures has been increasing recently,which results in the significant increase of computation.Therefore,these models cannot meet the requirement of fast response in the scenarios such as the display advertising.To deal with this issue,we proposed an attributes-driven recommendation model on the extraction and distillation of sequential latent information for user history behavior.The user and item attributes and the contexts from user's history behavior sequence have been utilized as the drive to model the latent information from high-order feature interactions and user's sequential history behavior.With regard to the issues of long-distance dependency and noise,we have adopted the self-attention mechanism to learn the sequential patterns between items;With regard to the issue that some complex models cannot meet the requirement of fast response,the objective of model compression and acceleration has been realized twofold:(a)we constructed the knowledge distilled teacher and student modules,respectively.The complex teacher module extracts user's general preference from high-order feature interactions and sequential patterns from user's long history sequence;(b)we proposed a long sequence sampling method to sample the relatively long-term items and short-term items which resulted in a shortened history sequence for modeling.The experimental study on two real-world datasets demonstrates that the proposed approach outperforms the best baseline ALIEN.For the MovieLens-1M dataset,the teacher module improves 2.00%and 3.11%in AUC and ACC scores,respectively;while the student module improves 0.60%and 1.36%,respectively.For the JD dataset,the teacher module both improves 0.27%in AUC and ACC scores;while the student module improves 0.16%and 0.65%,respectively.The results demonstrate considerable improvements on the click-through rate prediction.With regard to the efficiency,the training times of the student module on MovieLens-1M and JD datasets decrease 36.71%and 34.85%compared with those of the teacher module,and the recommendation times decrease 37.63%and 49.18%,respectively,which shows the effectiveness of the model compression and acceleration.Based on this accomplishment,one SCI paper has been submitted.