Temporary recommendation on graphs via long- and short term preference fusion

brief introduction

This article is written by Xiang Liang in KDD10 Papers at the conference , The main focus is temporal recommendation On . In order to model the long-term and short-term preferences of users , And used in timing recommendation , This paper presents a session based sequence diagram that can model both long-term and short-term user preferences （STG） And the embedded preference fusion which extends the personalized random walk （IFA） Algorithm .

primary coverage

1 Different from the traditional concept drift research, it focuses on the change of global patterns over time , Time series recommendation pays more attention to the change of each user's recommendation model . meanwhile , It is worth noting that , The user's overall behavior is determined by his long-term interests , But at some point in time , Users' behavior is often affected by their short-term interests and some sudden events, such as new product release and birthday .

2 Many time related models , Such as tensor decomposition , Add time to the model as one of the dimensions shared by all users . however , In the field of recommendation , It is not appropriate to compare time as a dimension between users , Time is more suitable to be considered as a local factor related to specific users . such as , Users at the same time , You may buy different items for the same external reasons , It is also possible to buy different items for different reasons , Therefore, the time factor should be related to specific users .

3 Koren The proposed timing recommendation algorithm does not improve the performance of implicit feedback ,

4 The uniqueness of the proposed model lies in the introduction of session nodes , Used to capture user related points in time . say concretely , Session based sequence diagram STG Capture long-term interests through the association between users and projects , Capture short-term interests through the association between the session and the project .

5 Progress in timing recommendations ：

1. Reduce the weight of old data
2. take CF Consider it as a univariate time series problem , Then use the decision tree
3. Time dependent iterative prediction problem
4. Automatically allocate the nearest neighbor according to time K The choice of
5. Mining past interactions to define stable preferences , And use the comments to find the user's preference based on the session
6. Time series change is introduced into factor model

front 4 Long term and short-term preferences are not modeled in , And after 2 The difference is that this paper uses graph model to model