Intelligent recommendation system and application enhanced by knowledge graph

With the rapid development of Internet technology, especially the rise of mobile Internet, newly generated information has shown explosive growth. In order to better solve the information overload (Information Overload) and the long tail problem (Long Tail) in information acquisition, the recommendation system came into being and has basically become a standard feature of various products. Recommendation systems are a branch of information filtering systems that can automatically mine connections between users and items. Specifically, it attempts to be based on the user's own multi-dimensional attribute data (such as age, region, gender, etc.) and behavioral data feedback (such as clicks, collections, likes, purchases, etc.), combined with the item's own attribute data (such as title, tags, categories, text, etc.) to predict users' ratings or preferences for recommended items. From the user's perspective, the recommendation system makes automatic recommendations for thousands of people based on the user's personal interests and preferences, which helps alleviate the problem of information overload. From the perspective of items, differences in their own attributes and corresponding interactive behaviors can be reached through various recommendation methods to user groups who are more interested in them, alleviating the long-tail problem caused by insufficient exposure. From an enterprise perspective, the recommendation system brings better product interaction methods and achieves an immersive experience, thereby further increasing user stickiness and ultimately significantly increasing conversion revenue.

Figure 1 Daguan Intelligent Recommendation System

In the field of intelligent recommendation ToB enterprise services, Daguan Data has accumulated more than 10 years of recommendation technology and practical experience in industry application with thousands of customers. As early as 2012, the team led by Dr. Chen Yunwen, the founder of Daguan Data, participated in the EMI Data Hacking Competition held in London and won the international championship. The competition mainly focused on music recommendation scenarios and how to perform based on user listening behavior and other data. Analyze and mine to predict user interests and preferences and make song recommendations. After a fierce battle, the intelligent recommendation system developed by them modeled the data of 5 million listening users, and generated a thousand-faceted recommendation for each user from hundreds of thousands of song libraries based on each user's personalized interests and preferences. The song recommendation results and the accuracy of the recommendation beat more than 300 participating teams from the University of Cambridge, University of Oxford, University of Michigan, etc., and won the championship in one fell swoop. Daguan Intelligent Recommendation is based on cutting-edge artificial intelligence and big data analysis and mining technology. After years of product polishing and continuous industry application exploration, the cumulative number of customers it has served has reached thousands. （https://www.datagrand.com/products/recommend/）

Recommendation method based on filtering ideas

After years of theoretical development of recommendation systems, three generations of major recommendation systems have been produced. The first generation recommendation system (1995-2005) mainly includes three methods: content filtering-based method, collaborative filtering-based method and hybrid method. Technically, it is mainly rule statistics and machine learning. The second generation recommendation system (2003-2014) is mainly based on feature context such as time, location, user group ratings, etc. Research on this generation of recommendation system is still in progress. Research on the third generation of recommendation systems focuses more on semantic models based on representation learning and the use of more knowledge components in the recommendation process.

01Recommendation method based on collaborative filtering

Collaborative Filtering (CF) is a traditional recommendation method that embodies the wisdom of the crowd. It makes recommendations based on the user's interest preferences and historical interaction behaviors with items. This method can be divided into memory-based methods and model-based methods. Memory-based methods can be divided into two categories: user-based (User-based CF) and item-based (Item-based CF). The most popular algorithm for memory-based methods is the KNN algorithm, which uses some traditional similarity measures, such as Pearson, Spearman, Cosine, Jaccard, etc. On the other hand, among model-based methods, the most commonly used are matrix factorization (MF) and its variants (NMF, SVD). At present, some new model-based collaborative filtering methods have emerged, such as Bayesian, clustering-based, rule-based and graph-based recommendation methods.

There are two main problems with collaborative filtering: the sparsity of user data when there is little interaction between users and items, and the cold start problem (new users and new items). In addition, traditional recommendation technology does not take advantage of the many semantic information, keyword relationships and hierarchical structures in recommendation scenarios.

02Recommendation method based on content filtering

The recommendation system based on this method recommends new items by learning new items that are similar in content characteristics to the items the user has preferred in the past. Such methods can be divided into case-based reasoning and attribute-based techniques. Case-based reasoning technology mainly recommends items that are highly related to items that the user liked in the past. In contrast, attribute-based techniques generate recommendation results based on matching item attributes with user attributes. The models used by most recommendation systems based on content filtering include: keyword matching or vector space model (VSM), term frequency-inverse document frequency (TF-IDF) weighting, topic modeling, etc.

Based on the content filtering recommendation method, the recommended items have high textual relevance and can well explain the recommendation results. However, the recommended results are often less surprising and the text Sparse features will also affect the calculation of correlation.

03Recommendation method based on demographic information filtering

The main idea of ​​this method is the fact that users who share certain personal attributes (gender, age, country, etc.) also share common preferences. Based on this, these systems can generate recommendations by classifying users based on demographic attributes. These methods are particularly useful when the amount of information about the item is limited. An advantage of this approach is that it does not require users to rate items or have interactive feedback as is necessary for content-based and collaborative filtering approaches.

However, the main problems with this type of recommendation method are: first, it is impractical to collect complete information for users due to security and privacy issues; second, this method targets relevant demographics Users in a group recommend the same products, and the degree of personalization is limited.

04Recommendation method based on context-aware filtering

This type of recommendation system combines scene context information to make recommendations. This approach assumes that the context of the current recommendation scenario is defined with a predefined set of observable attributes, whose structure does not change significantly over time. The so-called contextual information mainly includes time, location or other people (such as friends, relatives or colleagues). This contextual information provides additional information for the generation of recommendation results, which is more complementary than only considering the user or item's own information.

05Recommendation method based on knowledge filtering

This type of recommendation system mainly considers how to recommend to meet the user's interests and preferences based on domain knowledge. These systems should use three types of knowledge: knowledge about users, knowledge about items, and knowledge about correspondences between items and user needs. Generally speaking, this method mainly relies on knowledge graphs to provide more auxiliary information for the recommendation system to improve recommendation accuracy. It will be introduced in detail later.

06Recommended method for hybrid filtering

These systems usually combine collaborative filtering with content filtering or collaborative filtering with any other recommendation method for recommendations. The goal of the combination is to leverage the strengths of each approach to improve overall system performance and recommendation effectiveness. Currently, some works on hybrid methods include methods based on deep learning, Bayesian networks, clustering, latent features, graph structures, etc.

In recent years, methods based on deep neural networks, such as DNN, Wide & Deep, and DeepFM, have achieved impressive performance in learning to rank (LTR). These methods follow the Embedding and Multilayer Perceptron (MLP) paradigms, in which large-scale sparse features are first embedded into low-dimensional vectors and then concatenated together into a multilayer perceptron to learn the nonlinearity between features. relation. Advanced LTR methods discover the effectiveness of extracting user interests from their historical behavior for ranking. Specifically, DIN (Deep Interest Network) uses the attention mechanism to learn the representation of user interest from the user's historical behavior on candidate items. DIEN (Deep Interest Evolution Network) uses a recurrent neural network to capture the evolution of user interests. DMT (Method Deep Multifaceted Transformers) uses multiple transformers to model different behavior sequences of users.

Generally speaking, recommendation algorithms are the core elements of recommendation systems. The recommendation method based on collaborative filtering models user interests and preferences based on the similarity of users or items in interactive data, while the recommendation method based on content filtering mainly uses the content characteristics of items. Recommendation systems based on collaborative filtering have been widely used because they can effectively capture user preferences and can be implemented quickly and conveniently in a variety of scenarios without the need to extract various features like in content filtering-based recommendation systems. However, recommendation methods based on collaborative filtering suffer from data sparse and cold start problems. To solve these problems, many types of hybrid recommendation systems have been proposed to unify interaction-level similarity and content-level similarity. In this process, various types of auxiliary information were also explored, such as item attributes, review data, users' social networks, etc. Practice has proven that hybrid recommendation systems can usually achieve better recommendation results and have become increasingly popular in recent years.
 

Knowledge Graph Overview

Knowledge Graph (KG) is a structure that describes entities or concepts and connects them using different types of semantic relationships. In 2012, Google proposed the term "knowledge graph" to refer to the use of semantic knowledge in web searches, with the purpose of improving the capabilities of search engines and enhancing users' search experience. Before the term "knowledge graph" became popular, DBPedia and other linked data sets were generated by Semantic Web technologies and the linked data design problem posed by Berners-Lee. Today, KG has gained widespread attention in the industry and has been applied in large-scale systems.

Over the past few years, an increasing amount of semantic data has followed Linked Data principles, by combining a variety of information from different subject areas such as people, books, music, movies, and geolocations. ) is connected to a unified global data space for publishing. These heterogeneous data are interconnected and form a huge information resource library called a knowledge base. Several typical knowledge bases have been built, including academic projects such as YAGO, NELL, DBpedia, and DeepDive, as well as commercial projects such as Microsoft's Satori and Google's Knowledge Graph. Using heterogeneous connection information from knowledge bases can help provide insights into issues that are difficult to uncover with data from a single domain.

The following is an introduction to part of the knowledge base:

1. Freebase is a very practical and extensible tuple database system designed to become a public repository of the world's knowledge. Its design is inspired by widely used information communities such as the Semantic Web and Wikipedia. Data in Freebase is structured and generated through collaborative creation. It supports highly diverse and heterogeneous data and is highly scalable. Freebase currently contains 125,000,000+ tuples, 4000+ types, and 7000+ attributes. MQL (Metaweb Query Language) is a language for querying and operating data. Read and write operations on Freebase can be realized through the graph query (graph-query) API based on the HTTP protocol. MQL provides an easy-to-use object-oriented interface for tuple data in Freebase and was created to facilitate the collaborative creation of web-based data-oriented applications.

2. DBpedia is a large-scale multilingual knowledge base constructed by extracting structured data from Wikipedia versions in 111 languages. The largest DBpedia knowledge base extracted from the English version of Wikipedia contains more than 400 million factual data describing 3.7 million things. The DBpedia knowledge base, drawn from 110 other Wikipedia versions, contains a total of 1.46 billion facts describing 10 million additional things. DBpedia maps Wikipedia information boxes (infoboxes) in 27 different language versions into a single shared ontology, which consists of 320 classes and 1650 attributes. These maps are created through a worldwide crowdsourcing effort to blend knowledge from different Wikipedia versions. The project regularly publishes all DBpedia knowledge bases for download and provides access to SPARQL queries in 14 of the 111 language versions through a global network of local DBpedia chapters. In addition to regular releases, the project maintains a live knowledge base that is updated when pages in Wikipedia change. DBpedia has 27 million RDF links in place to more than 30 external data sources, enabling data from these sources to be used with DBpedia data.

3. YAGO is a link database developed by the Max Planck Institute in Germany. YAGO mainly integrates data from three sources: Wikipedia, WordNet and GeoNames. YAGO is built on entities and relationships. It currently contains more than 1 million entities and 5 million facts, 120 million triples of knowledge, including Is-A hierarchies and non-categorical relationships between entities. Facts have been automatically extracted from Wikipedia Extracted from and unified with WordNet. YAGO integrates WordNet's vocabulary definitions and Wikipedia's classification system, giving YAGO a richer entity classification system. YAGO also considers time and space knowledge and adds attribute descriptions of time and space dimensions to many knowledge items.
    

Knowledge graph is essentially a graph-based data structure and a semantic network that reveals the relationships between entities. Generally speaking, it is a semantic relationship network obtained by connecting different types of information. Knowledge graph describes the objective world in a structured way, precipitates background knowledge, and expresses information knowledge into a form that is closer to human understanding of the world. It has been It is widely used in search engines, intelligent recommendations, intelligent question and answer, language understanding, decision analysis and other fields.

Figure 2 Functional display of Daguan Knowledge Graph

Daguan Knowledge Graph is a new generation product launched by Daguan Data Company for knowledge graph applications in various industries. It integrates the full life cycle implementation of knowledge graph design, construction, editing, management, and application, and is based on customers' multi-source heterogeneous data. Integrating and building a knowledge middle platform can start from business scenarios, generate maps, and then implement map-based applications, which significantly improves the efficiency and effectiveness of knowledge maps in various industries.
 

Knowledge graph and recommendation system

Traditional recommendation systems take more explicit or implicit feedback between users and items as input, which brings two problems:

1. In actual scenarios, the interaction information between users and items is particularly sparse. For example, an online shopping application may contain hundreds of thousands of items, but the number of items actually purchased by users may only be in the hundreds. Using such a small amount of behavioral feedback data to predict a large amount of unknown information significantly increases the risk of algorithm overfitting.
    

2. For recommendations of new users and new items, due to the lack of historical interaction information, the accuracy of system recommendations will be greatly negatively affected. A common way to solve the sparsity and cold-start problem is to introduce additional auxiliary information into the input of the recommendation algorithm, such as user attributes, item attributes, contextual information, etc.

In recent years, introducing knowledge graphs as auxiliary information into recommendation systems has become a research hotspot in industry and academia. On the one hand, KG can provide rich domain knowledge as supplementary information to overcome the problems faced by collaborative filtering and content filtering-based recommendation methods; on the other hand, the recommendation system can use the semantic relationships existing in KG to improve its accuracy and increase recommendations. Variety of items. Specifically, KG recommendation exploits the connection between entities representing users, items to be recommended, and their interactions. Recommender systems use various connections to identify collections of items that may be of interest to a target user. Therefore, complex relationship representations provide additional valuable information for KG-based recommender systems to apply reasoning between nodes to discover new connections. On the contrary, in general, classic recommendation methods based on feature vectors ignore this connection, which may lead to poor overall recommendation performance, especially in the case of sparse data.

Recommendation system integrated with knowledge graph

KG is a heterogeneous graph, where nodes represent entities and edges represent relationships between entities. Items and their attributes can be mapped into KG to characterize the interrelationships between items. In addition, users and their information can also be integrated into KG, which allows the relationship between users and items as well as user preferences to be captured more accurately.

Generally speaking, the KG-based recommendation method requires building a KG in the first step, which can be an Item Knowledge Graph (IKG) or a User-Item Knowledge Graph (UIKG).
 

1. About IKG. In IKG, items and their associated entities (such as item attributes) are used as nodes, and edges can represent attribute-level relationships of items (such as brands, categories, etc.), or user-related relationships (such as "all browsed", "Buy all").

2. About UIKG. In UIKG, users, items and their related entities are nodes, and edges can represent the relationship between users and items (such as clicks, collections, purchases, etc.).

Taking the construction of IKG as an example, items are first mapped to the external KG to find their associated entities, and then the multi-hop neighbors of the associated entities are extracted from the KG and form a subgraph of the recommendation system. Of course, there is no need to rely on external KG, and KG can be constructed based on the auxiliary information in the provided data.

Explainable recommendation systems are another popular research direction in recent years. On the one hand, if the implementation of recommendation result presentation can provide users with appropriate recommendation explanations, users can relatively better accept the recommendation results. On the other hand, it is also possible to gain a deeper understanding of recommendation algorithms. Compared with traditional recommendation systems, recommendation systems based on knowledge graphs present a variety of entities and relationships connecting users and items, and can better demonstrate the reasoning process.

Recommendation methods based on knowledge graphs can be divided into three categories according to how to apply knowledge graph data, namely embedding-based methods, connection-based methods and propagation-based methods.

01Embedding-based method

The main idea of ​​the embedding-based method is to use a large amount of factual knowledge in KG to further enrich the multi-dimensional representation of users and items. It mainly includes two basic modules, one is the graph embedding module, For learning the representation of entities and relationships in KG, it is necessary to apply the Knowledge Graph Embedding (KGE) algorithm to encode KG into a low-rank embedding. The KGE algorithm can be divided into two categories: translation distance models, such as TransE, TransH , TransR, TransD, etc., as well as semantic matching models such as DistMult.

The other is the recommendation module, which is used to predict user preferences for items based on learned features. Based on the difference in how these two modules are related in the entire recommendation framework, embedding-based methods can be further subdivided into two-stage learning methods, joint learning methods, and multi-task learning methods. The challenges faced by this type of method include how to use a suitable KGE method to obtain the embedding representation of entities and how to integrate the learned entity embedding representation into the recommendation module.

Figure 3 DKN framework

(1) Two-stage learning method

The two-stage learning method refers to training the graph embedding module and recommendation module separately. In the first step, the KGE algorithm is used to learn the embedding representation of entities and relationships. Then, the pre-trained graph-related embeddings, together with other user features and item features, are input into the recommendation model for user interest prediction. Figure 3 is a two-stage learning framework diagram of DKN (Deep Knowledge-aware Network) used for news recommendation. In the first stage, entities in news headlines are extracted and mapped to Satori KG to mine knowledge-level relationships between news. DKN models news by combining text embedding representations of sentences learned with KCNN and knowledge-level embeddings of entities in news content via TransD. In order to capture the user's dynamic interest in news, an attention mechanism is introduced to learn the user's representation by aggregating the user's historical clicked news embeddings.

The two-stage learning approach is easy to implement, where KG embeddings are often considered as additional features for subsequent recommendation modules. Another benefit is that KG embeddings can be learned without interaction data, so large-scale interaction datasets do not increase computational complexity. In addition, since KG is usually stable, once the embedding representation is learned, there is no need to update the embedding representation frequently. However, entity embeddings optimized through KGE models are more suitable for in-graph applications, such as KG completion. Since the KGE module and the recommendation module are loosely coupled, the learned embeddings may also be unsuitable for subsequent recommendation tasks.

Figure 4 CKE recommendation system process

(2) Joint Learning Method
Another trend is joint learning (Joint Learning) graph embedding module and recommendation module in an end-to-end training method . In this way, the recommendation module can guide the feature learning process in the graph embedding module. CKE (Collaborative Knowledge Base Embedding) unifies various types of auxiliary information in the CF framework, including attribute-level features, text features, and visual features of items. Attribute-level features are encoded with TransR to learn structural knowledge from KG, while text features and visual features are extracted with autoencoders. The objective functions of these three feature learning modules plus the recommendation module jointly learn model parameters.

Federated learning methods enable end-to-end training and can use KG structures to regularize recommender systems. However, in practical applications, the combination of different objective functions needs to be fine-tuned.

Figure 5 Example of MKR framework and cross compression unit

(3) Multi-task learning method
A recent research direction is to use the multi-task learning (Multi-Task Learning) strategy to train recommendation tasks under the guidance of KG related tasks . The motivation is that items in a user-item interaction bipartite graph and their associated entities in the KG are likely to share similar structures. Therefore, the transfer of low-level features between items and entities helps facilitate the improvement of recommendation systems. MKR (Multi-task feature learning approach for Knowledge graph enhanced Recommendation) consists of a recommendation module and a KGE module. Instead of feeding KG embeddings into the recommendation module, the two modules are independent and connected through cross-compression units to share knowledge. The recommendation module is trained to estimate user preferences for candidate items, while the KGE module is trained to estimate the tail entity representation in a given head entity and triplet. Specifically, the recommendation module is based on MLP to obtain end-user representation. The final item representation is refined by L-layer cross-compression units and their related entities in the KG. Use nonlinear functions to estimate user preferences for candidate items.

By applying multi-task learning strategies, it helps prevent over-fitting of the recommendation system and improves the generalization ability of the model. However, similar to federated learning methods, it requires efforts to integrate different tasks under one framework.
In summary, although the two-stage learning method is easy to implement, the learned entity embeddings may not be suitable for recommendation tasks, the joint learning method learns optimized entity embeddings through end-to-end training, and the multi-task learning method The generalization ability of the model is further improved by transferring knowledge from KG-related tasks. However, it requires extensive experiments to find the best combination of different objective functions.

02Connection-based method

Connection-based methods use the connection patterns in the graph to guide recommendations. Most related works use UIKG to mine relationships between entities in graphs. There are two main ways to explore connection information in a KG. The first direction is to use the meta-structure in the graph, including meta-paths and meta-graphs, to calculate the similarity between entities. Meta-structure-based similarity can be used as a constraint on user and item representation, and can also be used to predict users' interest preferences for similar users or similar items in the interaction history. The second solution is to encode the connection patterns between user-item pairs or item-item pairs into vectors that can be integrated into the recommendation framework. This method is also called path embedding based method. The challenges of this approach include: 1) how to design appropriate meta-paths for different tasks; 2) how to model the connection patterns between entities.

(1) Method based on metastructure

One implementation of the meta-structure based method is to use the connection similarity of entities in different meta-paths as the graph regularization term to constrain the representation of users and items. The motivation is that the higher the similarity of entities based on meta-paths, the closer they are in the latent space.
The objective function is shown in formula (1):

where LRec represents the objective function of the recommendation system, and a common choice is matrix decomposition. Similarity constrained LSim guides the learning of user embeddings and item embeddings. In order to measure the connection similarity between entities in the graph, PathSim is usually used, as shown in Equation (2):

where Pm~n is a path between entities m and n. Three types of entity similarities are usually used, as follows: (a) User-user similarity, the objective function is as shown in Equation (3):

Where ||Ui-Uj||F represents the Frobenius norm of the matrix, =[1,2,...L] represents the weight of each element path, and U=[u1, u2,...,um] represents The latent vector of all users, S[1-(i,j)] represents the similarity score of users i and j in the meta-path. If users share high similarity based on meta-paths, user-user similarity forces users' embeddings to be close in the latent space.

(b) Item-item similarity, the objective function is shown in equation (4):

where V=[v1,v2,...,vn] represents the latent vector of all items. Similar to user-user similarity, if the meta-path-based similarity of items is high, the low-rank representation of the items should be close to of.

(c) User-item similarity, the objective function is shown in Equation (5):

If the meta-path based similarity is high, the user-item similarity term will force the user and item latent vectors to be close to each other.

In addition to the above three similarity methods, meta-structure-based methods can also use entity similarity to predict users' interest in unrated items, which can be regarded as preference fusion in KG.

In summary, the above methods first learn the latent vectors of users and items from the interaction matrix and their mutual similarities based on meta-structures, and then make predictions based on the enhanced representation. It is also possible to directly use a weighted set of similar user ratings to predict preferences for unrated items. Metastructure-based methods are interpretable because these manually designed metastructures provide more reference information for the recommendation system by matching metastructures between candidate items and interactive items or target users.

Metastructure-based methods are easy to implement, and most works are based on MF techniques with relatively low model complexity. However, the selection of metapaths or metagraphs requires domain knowledge, and these metastructures can vary significantly for different datasets. Furthermore, it may not be appropriate to apply metastructure-based methods in some specific scenarios. For example, in news recommendation tasks, entities belonging to one news may belong to different domains, which makes meta-path design difficult.

(2) Method based on path embedding

One problem with metastructure-based approaches is that connection patterns are not explicitly modeled, which makes it difficult to learn the interplay between user-item pairs and connection patterns. However, path embedding-based methods can explicitly learn embeddings of connection patterns. By learning explicit embeddings of paths connecting user-item pairs in UIKG or item-item pairs in IKG in order to directly model user-item or item-item relationships. Taking the relationship modeling in UIKG as an example, assuming that there are K paths connecting ui and Vj in KG, and the embedding of path p is expressed as hp, the final representation of the interaction between ui and Vj can be obtained through Equation (6):

where g() is a function that summarizes information from each path embedding. Common choices are max pooling operations or weighted sum operations. Then, the preferences of ui and Vj can be modeled by Equation (7):

where f() is a function mapping the interaction representation between user-item pairs and the embedding of user-item pairs to preference scores.

The recommendation results can be explained by checking the weight of each meta-path. A higher meta-path weight means that this relationship between the target user and the candidate item is more important when making decisions.

Path embedding-based methods encode the connection patterns of user-item pairs or item-item pairs into latent vectors, so that the mutual influence of target users, candidate items and connection patterns can be considered. In addition, large Most models are able to automatically mine connection patterns by calculating suitable paths and selecting salient paths without the help of predefined meta-structures. Therefore, it is likely to capture expressive connectivity patterns. However, if the relationships in the graph are complex, the number of possible paths in the graph can grow to be very large. In practice, it is impossible to exploit all paths for every entity pair in a large-scale KG, which may hinder the performance of the model.

In summary, connection-based methods rely heavily on connection patterns. But the representation ability of meta-paths is limited, which hinders the performance of traditional meta-structure-based methods. The path embedding-based method further overcomes another shortcoming of the meta-structure-based method, which is the need for domain knowledge and manual configuration of paths. These methods enumerate possible paths and explicitly model relationships between user-item pairs or item-item pairs. However, path embedding-based methods sacrifice scalability to a certain extent because these models are relatively complex and require more computation when enumerating paths and learning representations.

03Propagation-based method

Embedding-based methods utilize semantic relationships in knowledge graphs to enrich user and item representations, but have difficulty capturing high-order relationships between entities. Connection-based methods use connection information in the graph to guide recommendations, but information is inevitably lost by decomposing complex user-item connection patterns into separate linear paths. To fully utilize the information in KG, propagation-based methods integrate the representation of entities and relationships as well as high-order connection patterns to achieve more personalized recommendations. The main idea of ​​propagation-based methods is embedding propagation, where a common implementation is based on GNN technology. These methods refine the entity representation by aggregating the embedding representations of multi-hop neighbors in the KG. A rich representation of the user and potential items can then be used to predict the user's preferences.

According to the differences in entity types refined in the message dissemination process, it can be further subdivided into three categories. Challenges with this approach include:

1. How to assign appropriate weights to different neighbors

2. How to spread the news across different relationships

3. How to improve model scalability
    

(1) Refining of user embedding representation
Refining user embedding representation based on the user’s interaction history. First, an IKG is constructed to connect interactive items and candidate items using multiple relationships. Then users can be represented as a combination of the items they interact with and their multi-hop neighbors. Specifically, items in the interaction history are selected as seeds for the propagation process. Then, extract the multi-hop triplet set S[k-ui](k=1,2,...,H) along the links in the graph, where S[1-ui] is the triplet set (eh, r ,et), the head entity is the list of items that the user ui has interacted with. The process of learning user representation UI can be expressed as the following two steps:

(a) Compute the user’s embedding representation o[k-u] by aggregating entities in each layer of the triplet set S[k-ui] (k=1,2,...,H).

(b) Merge o[k-u](k=1,2,...,H) to obtain the final user embedding representation ou.
 

Since the propagation process starts from the items that the user has interacted with and ends with distant neighbors, this process can be seen as spreading the user's preferences layer by layer in IKG. Therefore, these methods can be interpreted as propagating users' preferences from historical interests along paths in the KG.

In these methods, edge weights are explicit in IKG. Therefore, significant paths connecting candidate items and interaction items can be selected and used as explanations for the recommendation results. Although these works exploit both entity embedding and higher-order connection information, only the user embedding representation is updated during the propagation process.

(2) Refinement of item representation
The above introduces the optimization of user embedding representation by aggregating entities outward in the graph. Another way is to learn a high-order representation of the candidate item Vj by aggregating the inward embedding representations of item Vj's multi-hop neighbors N[k-u] (k=1,2,...,H) in IKG. In the inward propagation process, a graph attention mechanism is adopted, where the weights of different neighbors are determined by users and relationships. Mainly considering that users have different preferences for different relationships, the information flow of KG can be determined.

Each round of propagation process is expressed as the following two steps:

(a) Aggregate the nearest neighbors of entity ei through equation (8):

(b) Use h-1 order neighbor embedding and self-embedding to update the h-order representation of the entity, as shown in Equation (9):

where e[0-i] represents the initial representation of the entity, and e[h-i] represents the h-order representation of the entity, which is a mixture of the initial representation of the entity and the representation from h-hop neighbors. The aggregation function maps N neighbors to the vector ∈ Rd, and the update function g() is a nonlinear function: Rd Rd → Rd. By iteratively repeating this process H times, the candidate item representation contains information from H hop neighbors.

In summary, the embedded representation of items is refined through inward propagation in IKG. However, similar to user refinement for outward aggregation in KG, only one type of entity is refined.

(3) Refinement of user and item representation
During the propagation process in UIKG, users, items and their associated entities are all connected in a graph. Interaction between user-item pairs as a relationship. User embeddings and item embeddings can be refined using their corresponding neighbors during the propagation process, as shown in Equations (8) and (9).

Similar to propagation in IKG, the weight of edges in UIKG is also determined by the user. Therefore, these models can provide explanations for recommendation results by examining the salient paths connecting target users and candidate items. Since users are combined into one type of node, the interpretation is more intuitive since the contribution of each interactive item is available. By incorporating users into the KG, higher-order connection patterns can be explored to a greater extent. The disadvantage is that more relationships in the graph will bring irrelevant entities, which may mislead the user's preferences in the aggregation process.

In summary, propagation-based methods are usually computationally expensive. As the graph gets larger, the model becomes difficult to converge. To improve efficiency, faster graph convolution operations can be used, and neighborhood sampling is typically applied in each layer. However, random sampling inevitably leads to information loss and cannot fully mine the knowledge in the graph.
 

04Summary of recommended methods based on KG

Through the above introduction of the embedding-based method, the connection-based method and the propagation-based method, it can be seen that the embedding-based method is the most flexible method. On the one hand, it is relatively easy to encode KG using KGE modules, and the learned embeddings can be naturally integrated into user representations or item representations. While in connection-based approaches, defining meta-paths or meta-graphs in the graph can be tedious. For propagation-based methods, the aggregation and update parts need to be carefully designed. On the other hand, embedding-based methods are suitable for most application scenarios since external knowledge is usually available in different tasks. In contrast, in metastructure-based methods, meta-paths are usually diverse for different application scenarios and cannot be generalized to new datasets. Furthermore, for specific scenarios, such as news recommendation, it is difficult to define meta-paths and apply meta-structure-based methods. At the same time, both path embedding-based methods and propagation-based methods are not suitable for recommendation scenarios with large-scale datasets because the computational complexity may become large when enumerating paths and neighbors. Furthermore, the quality and quantity of paths are crucial for connection-based methods, and therefore sparse datasets may not provide enough paths to mine relationships and model interest for such methods. However, both embedding-based methods and connection-based methods fail to fully explore the information in KG. In recent years, with the development of GNN technology, propagation-based methods have become a new research trend. Furthermore, both connection-based methods and propagation-based methods can be explained by paths in KG, while embedding-based methods are less intuitive to explain.
 

Interpretability based on KG recommendations

KG contains a large amount of auxiliary information that can be used to interpret recommendation results. There are mainly the following methods:
 

01Attention mechanism of relational embedding

This method is mainly used in embedding-based methods. The attention mechanism is applied to the embedding of relationships between entities in KG. From the attention weights of different relationships, the significance of each type of item attribute to the target user can be obtained. Therefore, this technique can provide preference-level explanations for recommendations.
 

02Define meta path or meta graph

The relationship between the selected item and the target user or interactive item can be decomposed into a combination of several meta-paths or meta-graphs. By converting metapaths or metagraphs into understandable rules, the system can provide explanations.
 

03Attention mechanism of path embedding

For the path embedding method, the weight of a specific path connecting the target user and candidate items can be obtained through the attention mechanism. The weight of each path can represent the relative importance of each path to the user. Therefore, explanations can be provided based on the significant paths in the graph.
 

04Reinforcement Learning in UIKG

By using reinforcement learning techniques to train agents in UIKG, the actual paths connecting user-item pairs can be mined. It can directly display the reasoning process in KG instead of looking for post hoc explanations for the already selected recommendation results. Therefore, the reasoning process is accurate and trustworthy for the target users.
 

05Extract edge weights

Propagation-based methods require assigning user-specific weights to each type of neighbor during the aggregation process. Edge weights control the information flow between entities in the graph and can reflect the importance of each relationship in the KG. In addition, the edge weights between entities in KG can also be obtained from attention weights or learned relationship matrices. Therefore, explanations can be generated by finding salient paths connecting candidate items and target users or interacting items in multi-hop neighbors.
 

future outlook

From the previous introduction, we can know that the recommendation system based on KG has many advantages in terms of recommendation accuracy and interpretability of recommendation results. Excellent models have also been proposed in academia and industry to make full use of the auxiliary information in KG for personalized and accurate recommendations. However, there is still a lot of work worthy of in-depth study in some directions, mainly reflected in:

01 Movement recommendation

Although KG-based recommendation systems with GNN or GCN architecture achieve good performance, the training process is very time-consuming. Therefore such a model can be regarded as a static preference recommendation. However, in some scenarios, such as online shopping, news recommendations, etc., users' interests will be quickly affected by social events. In this case, recommendations using static preference modeling may not be sufficient to understand real-time interests. To capture dynamic preferences, leveraging dynamic graph networks may be a solution.
 

02 Cross-region recommendation

There are also some research progresses in cross-domain recommendation, mainly because the interaction data is unbalanced in various fields. For example, on the Amazon platform, the subset of books is larger than other domains. Through transfer learning technology, interaction data from source domains with relatively rich data can be shared for better recommendations in the target domain.
 

03 Knowledge-enhanced language representation

In order to improve the performance of various NLP tasks, there is a trend to integrate external knowledge into language representation models so that knowledge representation and text representation can refine each other. Applying knowledge-augmented text representation strategies to text-based recommendation tasks allows for better representation learning to provide more accurate recommendations.

References
[1] Bollacker K, Evans C, Paritosh P, et al. Freebase: a collaboratively created graph database for structuring human knowledge[C]/ /Proceedings of the 2008 ACM SIGMOD international conference on Management of data. 2008: 1247-1250.

[2] Lehmann J, Isele R, Jakob M, et al. Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia[J]. Semantic web, 2015, 6(2): 167-195.[3] Suchanek F M, Kasneci G, Weikum G. Yago: a core of semantic knowledge[C]//Proceedings of the 16th international conference on World Wide Web. 2007: 697-706.

[4] Bordes A, Usunier N, Garcia-Duran A, et al. Translating embeddings for modeling multi-relational data[J]. Advances in neural information processing systems, 2013, 26.

[5] Wang Z, Zhang J, Feng J, et al. Knowledge graph embedding by translating on hyperplanes[C]//Proceedings of the AAAI conference on artificial intelligence. 2014, 28(1).

[6] Lin Y, Liu Z, Sun M, et al. Learning entity and relation embeddings for knowledge graph completion[C]//Twenty-ninth AAAI conference on artificial intelligence. 2015.

[7] Ji G, He S, Xu L, et al. Knowledge graph embedding via dynamic mapping matrix[C]//Proceedings of the 53rd annual meeting of the association for computational linguistics and the 7th international joint conference on natural language processing (volume 1: Long papers). 2015: 687-696.

[8] Yang B, Yih W, He X, et al. Embedding entities and relations for learning and inference in knowledge bases[J]. arXiv preprint arXiv:1412.6575, 2014.

[9] Zou X. A survey on application of knowledge graph[C]//Journal of Physics: Conference Series. IOP Publishing, 2020, 1487(1): 012016.

[10] Q. Guo et al., "A Survey on Knowledge Graph-Based Recommender Systems," in IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 8, pp. 3549-3568, 1 Aug. 2022, doi: 10.1109/TKDE.2020.3028705.

[11] Chicaiza J, Valdiviezo-Diaz P. A comprehensive survey of knowledge graph-based recommender systems: Technologies, development, and contributions[J]. Information, 2021, 12(6): 232.

[12] Choudhary S, Luthra T, Mittal A, et al. A survey of knowledge graph embedding and their applications[J]. arXiv preprint arXiv:2107.07842, 2021.

[13] Gao Y, Li Y F, Lin Y, et al. Deep learning on knowledge graph for recommender system: A survey[J]. arXiv preprint arXiv:2004.00387, 2020.

[14] Wang H, Zhang F, Xie X, et al. DKN: Deep knowledge-aware network for news recommendation[C]//Proceedings of the 2018 world wide web conference. 2018: 1835-1844.

[15] Zhang F, Yuan N J, Lian D, et al. Collaborative knowledge base embedding for recommender systems[C]//Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. 2016: 353-362.

[16] Wang H, Zhang F, Zhao M, et al. Multi-task feature learning for knowledge graph enhanced recommendation[C]//The world wide web conference. 2019: 2000-2010.