| Nowadays,rich and diverse information is full of everyone’s life,and the problem of information overload is particularly serious.For the majority of users,a major problem caused by information overload is that a huge amount of redundant information will greatly interfere with the accurate selection of the information they need.How to quickly obtain more useful information from such a large amount of data is a common concern at present.Therefore,recommendation technology comes into being,which can predict the needs of users and recommend the content that users are most likely to like,effectively alleviating the anxiety of users who are difficult to choose when dealing with a large amount of information.In the recommendation system,the core problem is to model the user’s preference.The so-called user preference is the degree of interest of the user to the item(music,movie,commodity,etc.).How to accurately predict user preferences based on existing user behaviors and information is a very important issue.In response to the above problems,this paper conducts a full investigation of the research status in the field of recommendation,analyzes the shortcomings and improvements of the current work,and optimizes the model on the basis of previous work.Finally,the recommendation method is applied to the prototype system.The following is the main work of this paper:(1)Since the end-to-end model ignores the association and fusion between context and user behavior data,this paper proposes a two-stage sequential recommendation framework.By combining the non-intrusive multi-head attention mechanism and the masked language model to fuse the context,the user sequence obtains rich semantic information and feature representation.Aiming at the problem that users’ long-term and short-term interests are significantly different,this paper captures users’ long-term stable preferences and users’ immediate intentions in different ways.Finally,the adaptive fusion module is used to dynamically assign different weights to long-term and short-term preferences,so as to obtain an overall preference representation that is more in line with the user’s personalized characteristics.The proposed model achieves good performance on Movie Lens and Amazon Beauty/Books datasets.(2)There are many types of user behaviors in real scenarios.To solve this problem,this paper proposes a multi-behavior recommendation model based on graph neural network.By adopting a custom meta-learning paradigm to encode the feature representations of different types of behaviors,the learned meta-knowledge is used to generate transformation weights,and the semantics of specific types of behaviors are injected into the initial embedding.At the same time,a multi-behavior relationship learning function based on attention network is used to solve the interweaving problem of different types of behaviors.Finally,the high-order behavior information in the graph structure is captured by the graph neural network.The effectiveness of the model is verified by a large number of experiments on three e-commerce data sets.(3)Based on the recommendation model proposed in this paper,a prototype system for shopping recommendation is designed and constructed.Firstly,the system used traditional recall algorithm based on collaborative filtering and content similarity to reduce the calculation range of goods.Then,the recommendation model proposed in this paper is used to calculate the prediction value of the recall result list and sort it.Finally,the page is rendered to the user according to the recommendation list.Tests and verification are carried out for the functions designed in the system,such as user login,order management,and recommendation for you.The results prove that the prototype system can run stably. |
