Distinguish User Long-term And Short-term Interest Of Personalized Dynamic Recommended Model

Face to the information explosion caused by Internet, personalized recommendation does better than search engine in helping users to find their own interest. Modeling based on users' interest which is the basis of personalized recommendation has become a research hotspot. Meanwhile, researchers have proved that the user's interest tend to be affected by the environment. At this point, researches on the users' interest change in a dynamic process has become the focus under this background. Previously interest model concerning time factor usually sees the users' interest as a whole and makes little distinction between long-term and short-term interest. And the research contents is always associated with time window and the law of forgetfulness. The main point of this paper is about the recognition of long-term and short-term interest, as well as the recommendation mechanism based on the research result. Details of this paper can be expressed as the following:(1) It is generally believed that the user's long-term interest usually is stable, not easy to change and catches more attention while the user's short-term interest usually is unstable, easy to change and catches less attention. In order to build an interest model that could distinguish user's short-term and long-term interest, firstly, this paper collects varies kinds of data from T-mall including users' query records and classification of goods. Secondly, the paper processes the user's query words and commodity category with Word Segmentation, then convert them to the corresponding TF-IDF values and calculate their similarity. Afterwards, a mapping relationship between the query words and the commodity category under the maximum similarity. is established. To distinguish users long-term and short-term interest, a threshold is calculated based on the ratio of the amount of words querying under a specific commodity classification and the amount of words querying under all kinds of commodity classification. Lastly, users' short-term interest can be obtained by the judgment of query words under undetermined interest. This paper also proposes the concept of interest acceleration to describe the changing process of short-term interest and generate the image of interest acceleration.(2) In order to realize the automatic identification of users' short-term interest model, firstly, the paper gives six patterns of the short-term interest(each model reflects a kind of short-term interest)including normal, rising gradually, decline gradually, jumping up, jumping down and cycle model according to the change of interest acceleration image. Secondly, the paper makes linear fitting according to the maximum data dimension of interest acceleration, fill value automatically and generate the training data of six different patterns. Lastly, three layers' BP network isdesigned to identify the model of short-term interest automatically.(3) In order to realize the recommendation mechanism based on users' short-term and long-term interest. Firstly, the paper will temporarily see the http as a specific commodity and build the user-http matrix table. If the http under the user's long-term interest appears in this table, mark 1 in the corresponding position and mark 0otherwise. Repeat the same operation to the short-term interest of the same model.Secondly, the paper realize recommendation mechanism according to users' long-term and short-term interest by the user collaborative filtering algorithm. Lastly,the paper use key string matching to acquire specific commodity links, craw the specific commodity by using Web page analysis algorithm, recommend specific commodity for users eventually.The paper uses T-mall user's query behavior dataset from Data tang, the model could better distinguish the users' short-term and long-term interest. All in all, this paper gets better recommendation results and has better value of application in E-commerce.