This is my first nuggets blog post, the beginning of nuggets writing.

🐴 recently, of course, or have been doing map, also see some articles, for example, recently saw an article on the arxiv article recommended system of knowledge map and, in fact, before the original also have the knowledge map and recommendation system combined with articles, don’t know if you have read this article, simple to share some today, bosses glad to ~

First, the core summary

It is nothing new to use the information of knowledge graph to enhance the effect of recommendation system. Now there are many research papers frequently appearing, and there are always a few papers whose views can be deeply rooted in your heart.

Recommendation system is now in our daily life everywhere, shopping electric business platform will give you recommend one thousand thousand items, below the APP on the AD above recommended, recommended by the social platform media news, etc., their goal is nothing more than recommended in vast amounts of information to the user is interested in, filter out the useless information. Recommendation system practices may be divided into the following categories.

1. Traditional methods include content-based and Collaborative Filtering (CF) recommendation methods, which mainly rely on content information and user and item similar data.

2. In recent years, there are more and more methods to use knowledge graph as auxiliary information of recommendation system, and many research results have shown remarkable effects. One of the reasons why knowledge graph can achieve such remarkable effects is that it uses graph structure to describe the relationship between entities in the real world. The potential information interaction between user and item can be clearly expressed. Figure 1 below is a simple knowledge graph.

There are several methods for knowledge graph to act on recommendation system as auxiliary information. Among them, the earliest one is knowledge representation, which mainly uses some knowledge representation methods to learn vector representation of some entities and combines vector representation of entities with recommendation system. The knowledge representation methods involved mainly include TransE, TransH, TransR, etc., which are based on a series of translation models and some improved methods proposed in the last two years. In addition to the method of knowledge representation, there were also some methods based on the path information of knowledge graph relationship in the early stage, which mainly used the semantic similarity of paths between entities to refine the representation of user and item.

However, the above two methods are difficult to capture high-level link information and cannot make full use of the rich semantic information and topological structure information of the knowledge graph itself. Recently, the method of using Graph Neural Network (GNN) has aroused the interest of a large number of researchers and industry insiders, including KGAT, KGNN-LS, KNI, AKGE and KGIN. This method can integrate the higher-order semantic information of the knowledge graph into the related feature representation of the recommendation system, thus promoting the effect transformation of the recommendation system.

The above is some background information and conventional mainstream methods combined with knowledge graph and recommendation system.

Two, insufficient improvement

The article to be shared this time is also a work based on the evolution of GNN methods. This paper is a work of the Chinese Academy of Sciences. According to this paper, despite the success of GNN in utilizing multi-hop attribute information, it still has three shortcomings.

1. Different attribute information is independent in terms of semantics and user preferences. Combining different attribute information with weighted sum will cause pollution. For example, in figure 1 above, movie e_1 contains the attribute e_2 of actor and e_8 of singer. Semantic independence means that e_2 and E_8 may not appear at the same time, because for a movie, actor and singer are independent of each other. Behavioral independence means that whether a user likes an actor or not is independent of whether a user likes a singer or not. GNN’s current use of weighted sums makes it difficult to obtain information about attributes of interest to users.

2. The paper puts forward a view that in general, the relationship between higher-order neighbors and central nodes in the graph is small, but this is not absolute. As shown in Figure 1, for some moviegoers who like music, the attribute e_8 of singer is more valuable than the attribute E_7 of director. However, E_8 is farther than E_7 in the figure. Therefore, higher-order attribute nodes are also very important. The weight of higher-order attribute node information decreases gradually.

3, the user’s interest is different from person to person, the user may only interested in partial attributes not interested in all of the properties, different users have different interest expression, so the need for different user preferences processing corresponding attribute information, that is, through geri weis-corbley learn the characteristics of objects to said contains noise signals, You need to extract some of the attributes that the user is personally interested in. However, the existing GNN methods do not fully consider this aspect.

To solve this problem, a novel Network structure (Attentive Knowledge Graph Attribute Network (AKGAN) is proposed. AKGAN network structure mainly consists of two parts.

  • The first part is Knowledge Graph Attribute Network (KGAN), whose core task is to learn Item representation without affecting the semantics and weights of important Attribute nodes. Under the GNN framework, the first and last layers of KGAN are designed differently.
  • The second part is user interest-aware Attention Network(IAAN). The core of this part is to design the Attention layer mechanism of Interest perception to extract the attributes that users are interested in, instead of treating all attribute nodes equally.

AKGAN network

As mentioned above, AKGAN contains two parts of network structure, KGAN and IAAN. Let’s take a brief look respectively.

3.1 KGAN

The figure above is the network structure diagram of KGAN, including AML(Attribute Modeling Layer) and APL(Attribute Propagation Layer).

KGAN is a GNN network structure adopted, and its core task is to enhance Item representation by learning information containing multiple attributes. The first layer and the last layer adopt heterogeneous design, named attribute modeling layer (AML) and attribute propagation layer (APL) respectively.

AML: In the knowledge graph, the same entity may have multiple neighbor nodes and may correspond to multiple relationships. For example, in the graph structure in the figure above, E_8 and its neighbor nodes E_4 and E_5 have different relationship attributes R_6 and R_4. In order to obtain entity representation without semantic pollution, AML embedded each entity into different relational attributes to construct entity feature representation.

As shown in the AML network structure above, entity E_1 has adjacent attribute nodes E_2 e_3 E_4 e_5 e_6 E_7. They correspond to different relationship attributes r_1, R_2, R_3 and AML respectively. Conduct Average pooling operation on neighbor nodes with the same relationship to extract vector expression of entity attributes in different relationships. In order to ensure consistency of length, zero vector is used to fill those without corresponding relationship. Concat operation is adopted to fuse vector expression of entity in different relations into vector information e_i^0 of entity, where I is the entity and 0 is the network layer.

To sum up, the above operation avoids the interaction between different attributes and maintains semantic independence, which will further facilitate the maintenance of high-order significant neighborhood weights in APL and the subsequent extraction of attributes of interest to users.

APL: AML has extracted the feature expression of the entity by using the one-hop neighbor of the entity adjacent to it. Therefore, it is easy to think of using GNN network to propagate them so as to obtain the attribute information of higher-order neighbors and aggregate the higher-order neighbors into the central entity node to obtain the Item representation with more abundant information.

APL is the aggregation of the same relationship-aware neighbors and the combination of relationships. The same relational-aware neighbor contains similar attribute types, so the representation of the same relational-aware neighbor is aggregated first. Secondly, in order to learn a more comprehensive representation, it is necessary to integrate multiple attributes in different relationship perception neighbors. The pooling methods of the above two operations are Average and sum respectively, and the final Item representation can be obtained by summating the same entity representation on each layer.

3.2 IAAN

The KGAN in the previous chapter is mainly to obtain the vector representation of Item. With the vector representation of Item, a typical approach in the recommendation system is to enhance the representation of the user by using the user’s behavior data, such as the Item clicked by the user. But this method does not take into account users’ personal preferences, so in recent years there are a lot of research work of the attributes of the user preferences by attention mechanism modeling, expected to extraction of interest to the user, this article also puts forward a new kind of module on the user’s attention different attribute information extraction and refining, the following figure. IAAN consists of two parts. Firstly, it calculates the interest score f_{att} between user (u) and attribute (r_m), and then combines Item representation and interest score to obtain user representation E_u.

4. Experimental conclusions

1, the data

The data set adopts three public data sets, amazon-Book, Last-fm and Alibaba-Ifashion.

2. Experimental results

Evaluation Method Top-K is used for evaluation in the recommendation. For each user, all items that the user does not interact with are considered negative examples, while items in the test set that interact are considered positive examples. Then these items are sorted and two commonly used evaluation schemes are adopted: Recall@K and NDCG@K. The final experimental report is the average index of all users in the test set.

Recall@K is the proportion of correct predicted correlation results to all relevant results. The value range is [0,1], and the larger the better.

NDCG@K is Normalized Cumulative Gain (NDCG), which is an evaluation index considering return order. The value ranges from 0 to 1. The larger the value, the better the effect.

The following are the experimental results

The above are the experimental results. From the experimental results, three data sets, two evaluation indicators and six comparison methods are all positive. In addition, there are some interesting case studies in the paper, which you can read when you have time.

Share the paper

Paper: Knowledge graph enhanced recommender system

Arxiv: arxiv.org/abs/2112.09…

The resources

  1. Knowledge graph enhanced recommender system
  2. Blog.csdn.net/qq_40006058…
  3. zhuanlan.zhihu.com/p/359528909

PS Welcome to “AI Natural Language Processing and Knowledge Mapping” public account