Research On Session-besed Recommendation Algorithm

The recommender system is an important technology in the field of the computer science.It is becoming more and more widely used in many other fields in the era of big data.And meanwhile,it is faced with many new challenges,such as how to reduce the response time of recommendation in the context of large data sets,and how to take advantage of all data of the user to make more accurate recommendations,and how to realize the online model update mechanism to capture the users'interests in real time,and so on.As the core of the recommender system,the recommendation algorithm also needs to be improved.The session-based recommendation methods are a kind of recommendation methods which don't need user information.Therefore,they are widely used in the scenario in which the user information is transparent to the system.Normally,they are divided into two types:the deep learning technology-based method and the traditional neighborhood-based method.Due to the different ways of realization,each of the two types has an irreplaceable advantage.Thereby,in this thesis,Firstly,we conduct in-depth research of these two types of methods,and analyze the defects of them.Then,aiming at the defects mentioned before,we propose improved methods based on different type of methods.The main research contents of this thesis are as follows:Firstly,for deep learning technology-based method,because the item implicit factor vector extracted in the classic GRU4REC model can't reflect the intrinsic relationship between items very well,we put forward the Recurrent Neural Network with Item to Vector model,also called RI2V.First of all,we utilize the distributed representation model to mine the"encode" function and the "decode" function between the implicit factor vector space and the item score vector space.Then to improve the robustness of the model,we added noise to the input data.In the last,we train the whole model in the implicit vector space but not in the way of end-to-end to get the best model.Secondly,for the problems of poor scalability and pool performance in the traditional neighbor-hood-based method,we proposed the session-based k-nearest-neighbor algorithm which combines the user's interest attenuation factors,also called SCIV-kNN in this thesis.Sessions are sampled to reduce algorithm time complexity and different time effect curves are proposed to simulate the user's interest decay on different items.In the last,we design the online model update mechanism for this method to meet the needs of capturing the user's interest in real time.Finally,to verify the proposed models' performance,we conduct experiments on two different public datasets and compare our methods with many kinds of traditional methods.The result of the experiments demonstrated that the proposed methods improved the recommendation performance.Besides,the SCIV-kNN method also significantly reduces the recommendation response time and the RI2V method further shortens the training time.