Learning The Association Of User Behavior Factors For Prediction

The location-based services have attracted more and more attention in recent years.These service platforms record a lot of users' behavioral data,which reflect user lifestyles and behavioral patterns.Predicting users' future behavior with these data will help improve user satisfaction,enable service providers to appeal new customers,and promote economic benefits.This paper investigates the problem of learning the associations of user behavior factors for prediction,where exists many challenges.Firstly,users' behavioral data contain some noisy data,it is difficult to learn individual preferences from the collective behaviors.Secondly,user behavior is influenced by the combination of external factors and personalized preferences,it is characterized by randomness and dynamics.Thirdly,user behavior involves action,location,time and other factors,it is difficult to learn the mutual influence among these factors in a unified model.To address the above challenges,this paper makes following contributions:(1)We propose a joint representation learning model.By considering the temporal and spatial characteristics of user behavior,the proposed model learns the representations of user,location,and action in the same latent space.The functional and spatial characteristics of a location are learned from collective behaviors.To solve the problem of stochastic time intervals in user behaviors,temporal pattern is introduced to model users' temporal preference,which describes the frequency of periodic behaviors.The adoption of temporal patterns will help decrease the influence of randomness in user behaviors.Then the interaction between the temporal-spatial context and action is considered and the model is used to learn the representations of users,temporal patterns and actions.(2)We propose two prediction models based on behavior representations.Firstly,a probabilistic inference model is given,which considers the regularity,periodicity,the relevance between location and action in user behaviors.The model infers the probabilistic distribution of future possible behaviors and makes prediction.The second one is the attention based recurrent neural network model.Based on the representations of user and behaviors,the dynamic effects of historical behaviors on user's preference are computed.Then the model predicts user future behavior based on user's preference.(3)The model optimization and prediction based on real datasets.We adopt a multi-objective optimization method base on stochastic gradient descent.In order to better understand the results of learning,this paper analyzes the semantics from the perspective of the location and user respectively.From the perspective of location,it proves that the learned location representations are semantically comparable.From the perspective of user,it finds some user personality characteristics,and helps understand the individual differences.This paper uses the two real representative datasets from Gas and Koubei platform to verify the validity of the proposed method.The results show that our approach outperforms others.