Research On Streaming Recommendation Approaches Based On User Behaviors

Recommender systems have been widely integrated into our daily lives and played an increasingly important role in improving the experience of users and increasing the profit of enterprises.However,conventional offline recommendation approaches commonly leverage large-volume historical interaction data between users and items to train recommender models periodically,and thus cannot well capture the latest preferences of users embedded in their recent interactions.To address this issue,streaming recommender systems have emerged in recent years,which commonly train recommendation models with newly received interaction data to capture the latest user preferences for accurate streaming recommendations.However,streaming recommendations confront several key challenges,including simultaneously capturing short-term and long-term user preferences,effectively handling underload scenarios and overload scenarios,well addressing the heterogeneity issues of user preferences and item characteristics,and well leveraging the multi-behavior interactions like purchase and view to improve the accuracy of streaming recommendations.Additionally,the workload intensity in streaming scenarios varies continuously,the computational resources required by streaming recommendation approaches change over time.Thus,resource scheduling algorithms with short response time are required to timely scheduling resources for streaming recommender systems.For scalability reasons,recommender systems are usually deployed in cloud environment.However,less research has been conducted on the resource scheduling approaches aiming at optimizing response time in the cloud environment.In order to address the above challenges,this dissertation conducts research in various aspects such as data sampling,recommendation model,and resource scheduling.This dissertation contains the following work:(1)A time-stratified sampling based self-adaptive ensemble learning approach.To simultaneously capture the short-term and long-term user preferences while effectively handling underload scenarios and overload scenarios,a time-stratified sampling approach is studied to extract training data from both newly received data and historical data.This approach contributes to simultaneously capturing the short-term and long-term user preferences.After that,targeting the overload problem,an adaptive ensemble learning method is studied to first leverage multiple individual recommendation models to concurrently learn from the prepared training data,and then fuse the results of these individual models through a sequential self-adaptive fusion method for accurate streaming recommendations.Experiments have been conducted on the simulated data stream with real-world interactive data between users and items,which has verified the effectiveness of the proposed recommendation approach in overload scenarios by comparing the recommendation accuracies.(2)A reservoir-enhanced sampling based double-wing mixture-of-expert recommendation approach.Targeting the issue of insufficient new data in underload scenarios,a reservoir enhanced sampling approach is proposed.This sampling approach wisely complement newly received data with representative historical data,which contributes to achieving better training effectiveness.After that,targeting the heterogeneity issue of user preferences and item characteristics,a double-wing mixture-of-expert model is studied.This model effectively learns the heterogeneous user preferences and item characteristics with two mixture-of-experts models,respectively,where each individual expert model specializes in one underlying type of users or items.Experiments have been conducted on the simulated data stream with real-world interactive data between users and items,which has verified the effectiveness of the proposed recommendation approach in underload scenarios by comparing the recommendation accuracies.(3)A multi-behavior streaming recommendation approach.Targeting the issue of insufficient single-behavior interaction data,this dissertation proposes the multi-behavior streaming recommendation approach to sufficiently exploit the multi-behavior interactions for delivering more accurate streaming recommendations.Specifically,a multi-behavior learning method is studied to accurately learn the short-term user preferences and stable item characteristics.Then,an attentive memory network is studied to effectively maintain the long-term user preferences.After that,these obtained short-term and long-term user preferences are merged by an elaborate user preference merging method.In addition,a multi-behavior training process is studied to capture the mutual influence among different behavior types for further increasing the accuracies of streaming recommendations.Experiments have been conducted on the simulated data stream with the interactive data collected on the online shopping platforms,which has verified the effectiveness of the proposed recommendation approach in the scenarios with multiple types of user behaviors by comparing the recommendation accuracies.(4)Response time optimized resource scheduling approaches for streaming recommendations.Targeting the problem that the load intensity in streaming recommendation scenarios is prone to fluctuations and the amount of computing resources required by the recommendation method changes over time,response time optimized resource scheduling approaches oriented to the cloud environment with limited and relative abundant computational resources are studied respectively,in order to satisfy the requirements of streaming recommendation approaches for the computational resources.Specifically,oriented to scenarios with limited computational resources,the branch-and-bound approach based on time-aware divide-and-conquer strategy is studied,which is used to optimize the response time of resource scheduling while saving resources.Oriented to scenarios with more sufficient computational resources,the shuffling process enhanced multi-dimensional first-fit approach is studied to schedule computational resources for the streaming recommendation approaches in real-time.Experiments have been conducted on the real-world and simulated workload,which has verified the effectiveness of the proposed approaches in reducing the response time and the required computational resources.