The traditional recommendation system only uses the historical interaction information (explicit or implicit feedback) between users and items as input, but this will lead to two problems:

Sparsity of (1) data: In practical application scenarios, the interaction information between users and items is often very sparse. For example, a movie APP may contain tens of thousands of movies, but a user may only have dozens of movies on average. Using such a small amount of observation data to predict a large number of unknown information will greatly increase the risk of over-fitting;

(2) Cold start problem: For newly added users or items, because the system has no historical interactive information, it cannot accurately model and recommend.

Generally speaking, the solution to sparsity and cold start is to introduce some additional auxiliary information into the recommendation algorithm as input. These auxiliary information can enrich the description of users and items, thus effectively making up for the sparseness or deficiency of interactive information. Among all kinds of auxiliary information, knowledge map, as a new type of auxiliary information, has been studied a lot in recent years.

Knowledge map is a kind of semantic network, whose nodes represent entities and edges represent various semantic relationships between entities. The knowledge map consists of several triples, in which sum represents the head node and tail node of a relationship and the relationship between nodes.

Knowledge map contains rich semantic associations between entities, which provides a potential source of auxiliary information for recommendation system. Knowledge maps have potential applications in many recommended scenarios, such as movies, news, scenic spots, restaurants, shopping and so on. Compared with other kinds of auxiliary information, the introduction of knowledge map can make the recommendation results have the following characteristics:

(1) accuracy

Knowledge map introduces more semantic relations to the article, which can deeply discover users' interests. For example, as shown in the figure below, users like Farewell My Concubine starring Leslie Cheung, and Leslie Cheung has just starred in The Story of Teddy Boy, so users may also like the movie The Story of Teddy Boy.

Based on the different ways of using KG information, the combination methods of knowledge map and recommendation system can be divided into three categories: embedding-based method, path-based method and unified method.

1. Method based on embedded system

Embedding-based methods usually use information from KG directly to enrich the representation of projects or users. In order to utilize the information of knowledge graph, it is necessary to encode knowledge graph into low-rank embedding by using knowledge graph embedding (KGE) algorithm. According to whether the user is included in the KG, the methods based on embedding can be divided into two categories: project-based graph and user-project graph.

(1) Based on the project diagram

The graph consists of items extracted from data sets or external knowledge bases and their related attributes, and does not contain user information. This method uses the knowledge graph embedding (KGE) algorithm to encode the graph, which can obtain a more comprehensive representation of the project, and then integrates the side information of the project into the recommendation framework. Specifically, the potential vector of the project can be obtained through various information, including KG, user-project interaction matrix, project content, project attributes and so on. Then, using the preference score function, the probability of users choosing items is calculated through the obtained potential vectors of users and items, and the preference ranking of users is obtained according to the probability results.

(2) Based on User-Project Diagram

In this diagram, users, projects and their related attributes act as nodes; Their attribute level relationship (brand, category, etc. ) and user-related relationships (* * * and purchase, * * and view, etc. ) as an edge. This method can get the entity embedding from the constructed atlas, and then get the result according to the preference score function. Different from the project-based graph, this preference score function can be calculated by embedding relationship.

The method based on embedding mainly includes two modules: graph embedding module, which mainly uses graph embedding method to learn the representation of entities and relationships in knowledge map; And a recommendation module for modeling the user's preference for the project. According to the combination of these two modules, the work in this direction can be divided into three categories, namely, sequential learning, joint learning and alternating learning.

(1) Learn in turn

This method first obtains entity vectors and relationship vectors by learning the characteristics of knowledge map, and then introduces these low-dimensional vectors into the recommendation system to learn user vectors and item vectors.

At present, the recommendation system based on knowledge map is still in the primary learning stage, and the specific models of various methods are not well understood. A few days ago, I read an article about the combination of He Xiangnan and knowledge map, which should be better in the current recommendation system based on knowledge map. In this paper, the multi-task learning strategy is adopted, and considering the fact that KG may be missing, the completion module and recommendation module are jointly trained. According to the current understanding, the recommendation system based on knowledge map has a good research prospect in dynamic recommendation, multi-task learning and cross-domain recommendation.

blogs.com/niuxichuan/p/93 177 1 1.html

Summary of recommendation system based on knowledge graph