Predicting User Behaviors From Temporal Click Data

With the rapid development of the Internet,online activities have become pervasive in daily life.The number of users on Internet platform such as Internet TV and online music is growing rapidly,so a large amount of user data with temporal information are produced.This kind of data involve users' implicit interests and is beneficial for mining user-behavior patterns and providing personalized recommendation.As a result,it has become a hot topic in the field of recommender system.There are some major challenges in mining temporal behaviors:(1)The item features are high dimensional and sparse,which makes it difficult to deal with largescale data.(2)Since user's interests are drifting with time,how to capture the dynamic preference for user behavior efficiently is another challenge of user temporal behavior mining.(3)The users' behaviors are not uniformly distributed with the change of time.Reducing the influence caused by the imbalance of active behaviors is the third challenge for the mining of users' temporal behaviors.Aiming at the challenges mentioned above,we proposed a Dynamic Recommendation based on Embedding Learning(DREL)model based on the idea of Word2 Vec.The model's main idea and innovation include:(1)The embedding is used to map the item feature vector to the low-dimensional vector space to deal with the high dimensionality and sparsity of item features.Not only can it learn the user's short-term preferences,but also can reflect the similarity between different items using the embedded vector.(2)As for the dynamics of user preference,we investigate user's click behaviors and find that the correlation would decrease with the increase of time interval by data exploration.Thus,a fixed behavior window is used to mine user's short-term dynamic preference.In order to deal with the imbalance of users' active time,this paper model the users' click behavior with a dynamic context window by taking each action of user actions as the time unit,rather than take time window as the influencing factor of user interest.In this paper,two datasets are used to evaluate the effectiveness of DREL in short-term prediction effect and ranking quality,among which Last.fm is an online music data and TV is an Internet TV data.Extensive experiments on two datasets demonstrate the effectiveness of the proposed model.