Knowledge Graphs: Opportunities and Challenges

Knowledge graph embedding is one of the central research issues. This task aims to map entities and relations of a knowledge graph to a low-dimensional vector space so that it captures the semantics and the structure of the knowledge graph efficiently (Dai et al. 2020b). Then, the obtained feature vectors can be effectively learned by machine learning models. Three main triplet fact-based embedding methods are as follows: (a) tensor factorization-based, (b) translation-based, and (c) neural network-based methods (Dai et al. 2020b).

Knowledge acquisition, which focuses on modeling and constructing knowledge graphs, is another crucial research direction of knowledge graph study. Typically, the knowledge is imported from structured sources by employing mapping languages, such as R2RML (Rodriguez-Muro and Rezk 2015). Furthermore, the knowledge could be extracted from unstructured documents (e.g., news, research papers, and patents) by adopting relation, entity, or attribute extraction methods (Liu et al. 2020; Yu et al. 2020; Yao et al. 2019).

Although there are many methods for constructing knowledge graphs, it is still unfeasible to create comprehensive representations of all the knowledge in a field. Most knowledge graphs still lack a good number of entities and relationships. Thereby, significant efforts have been made for completing knowledge graphs. Knowledge graph completion aims to improve the quality of knowledge graphs by predicting additional relationships and entities. The first task typically adopts link prediction techniques to generate triplets and then assigns the triplets plausibility scores (Ji et al. 2021). The second task employs entity prediction methods for obtaining and integrating further information from external sources.

Knowledge fusion is also an important research direction that focuses on capturing knowledge from different sources and integrating it into a knowledge graph (Nguyen et al. 2020). The knowledge fusion approaches are useful for both generating and completing knowledge graphs. Recently, entity alignment has been the primary method for implementing knowledge fusion tasks.

Tremendous research efforts have focused on reasoning to enrich the knowledge graphs, which aims to infer new facts based on existing data (Minervini et al. 2020). In particular, new relations between two unconnected entities are inferred, forming new triplets. Also, by reasoning out the false facts, knowledge reasoning has the ability to identify erroneous knowledge. The main methods for knowledge reasoning include logic rule-based, distributed representation-based, and neural network-based methods (Chen et al. 2020b).

Nowadays, knowledge graphs are widely utilized by AI systems (Liang et al. 2022), such as recommenders, question-answering systems, and information retrieval tools. Typically, the richness of information within knowledge graphs enhances the performance of these solutions. Therefore, many studies have focused on taking advantage of knowledge graphs to improve AI systems’ performance.

Knowledge graphs have numerous applications in various fields, including education, scientific research, social media, and medical care (Li et al. 2020b). A variety of intelligent applications are required to improve the standard of human life.

Differing from other works, this paper focuses on surveying the opportunities and challenges of knowledge graphs. In particular, knowledge graphs meet great opportunities by improving the quality of AI services and being applied in various fields. On the contrary, this paper regards the limitations of knowledge graph technologies as the challenges. Therefore, we will discuss the technical limitations regarding knowledge graph embeddings, knowledge acquisition, knowledge graph completion, knowledge fusion, and knowledge reasoning.

There are two traditional methods for developing recommender systems, including content-based and collaborative filtering-based (CF-based) methods. Sun et al. (2019) and Guo et al. (2020) have compared and summarised these two approaches.

To address inherent problems of traditional approaches, the community has produced several hybrid recommender systems, which consider both item features and the distribution of user scores. Most of these solutions adopt knowledge graphs for representing and interlinking items (Palumbo et al. 2020). Specifically, Knowledge graph-based recommender systems integrate knowledge graphs as auxiliary information and leverage users and items networks to learn the relationships of items-users, items-items, and users-users (Palumbo et al. 2018).

Fig 3 presents an example of knowledge graph-based movie recommendation. Here we can see that the movies “Once Upon A Time in Hollywood" and “Interstellar" are recommended to three users according to a knowledge graph that contains the nodes of users, films, directors, actors, and genres. The knowledge graph is thus used to infer latent relations between the user and the recommended movies.

Recently, a great deal of research has been conducted to utilize knowledge graphs for recommendation tasks. For instance, Wang et al. (2019b) introduced KPRN. KPRN is a recommender system that generates entity-relation paths according to the user-item interaction and constructs a knowledge graph that consists of the users, items, and their interaction. It then infers the user preference based on the entity-relation path. The user-item interaction, which is extracted from knowledge graphs, improves the quality of the recommendations and allows the presentation of the recommended results in a more explainable manner. Wang et al. (2019c) also applied multi-task knowledge graph representation (MKR) for recommendation tasks. MKR models knowledge graphs based on the user-item interaction. It is worth noting that MKR focuses on the structural information of knowledge graphs for learning the latent user-item interaction. Sun et al. (2020) proposed a Multi-modal Knowledge Graph Attention Network (MKGAT) for achieving precise recommendations. MKGAT constructs knowledge graphs based on two aspects: (1) it enriches entity information by extracting the information of the neighbor entities; (2) it scores the triplets to construct the reasoning relations. Finally, they applied knowledge graphs that are enriched with structured data to recommender systems.

Wang et al. (2018b) presented the RippleNet model, which incorporates knowledge graphs into recommendation tasks by preference propagation. RippleNet firstly regards users’ historical records as the basis of a knowledge graph. Then, it predicts the user preference list among candidate items based on the knowledge graph links. Based on both RippleNet and MKR models, Wang et al. (2021) applied the Ripp-MKR model. Ripp-MKR combines the advantages of preference propagation and user-item interaction to dig the potential information of knowledge graphs. Shu and Huang (2021) proposed RKG, which achieves recommendation by referring to the user preference-based knowledge graph. RKG first obtains users’ preference lists; then, it analyzes the relations between the user’s preferred items and the items which are to be recommended. Therefore, the model effectively learns the scores of the candidate items according to the relationships between candidate items and the user’s preferred items.

Many studies have utilized ontological knowledge base information to improve retrieving results from various data sources (Farfán et al. 2009). Wu et al. (2013) adopted the ontological knowledge base to extract highly semantically similar sub-graphs in graph databases. Their method effectively recommends semantically relevant sub-graphs according to ontological information. Farfán et al. (2009) proposed the XOntoRank, which adopts the ontological knowledge base to facilitate the data exploration and recommendation on XML medical records.

Compared with the traditional recommender systems, knowledge graph-based recommender systems have the following advantages:

The core idea of tensor factorization-based methods is transforming the triplets in the knowledge graph into a 3D tensor (Balažević et al. 2019). As Fig 5 presents, the tensor \({\mathcal {X}}\in R^{m\times m \times n}\), where m and n indicate the number of entity and relation, respectively, contains n slices, and each slice corresponds to one relation type. If the condition \({\mathcal {X}}_{ijk}=1\) is met, the triplet \((e_i,r_k,e_j)\), where e and r denote entity and relation, respectively, exists in the knowledge graph. Otherwise, if \({\mathcal {X}}_{ijk}=0\), there is no such a triplet in the knowledge graph. Then, the tensor is represented by the embedding matrices that consist of the vectors of entities and relations.

Translation-based methods exploit the scoring function, which is based on translation invariance. Translation invariance interprets the distance between the vectors of the two words, which is represented by the vector of their semantic relationships (Mikolov et al. 2013). Bordes et al. (2013) firstly utilized the translation invariance-based scoring functions to measure the embedding results. They creatively proposed the TransE model, which translates all the entities and relations of a knowledge graph into a continuous and low vector space. Specifically, the vectors of the head and tail entities in a triplet are connected by the vector of their relation. Consequently, in the vector space, the semantic meaning of every triplet is preserved. Formally, given a triplet (head, relation, tail), the embedding vectors of the head entity, relation, and tail entity are \({\textbf{h}}\), \({\textbf{r}}\), and \({\textbf{t}}\), respectively. In the vector space, the plausibility of the triplet \(({\textbf{h}},{\textbf{r}},{\textbf{t}})\) is computed by the translation invariance-based scoring function to ensure it follows the geometric principle: \({\textbf{h}}+{\textbf{r}}\approx {\textbf{t}}\).

After TransE, a lot of related extensions, such as TransH (Wang et al. 2014) and TransR (Lin et al. 2015), are continually proposed to improve the performance of the Translation-based knowledge graph embeddings.

Nowadays, deep learning has become a popular tool that is utilized for knowledge graph embeddings, and a considerable amount of research proposes to employ neural networks to represent the triplets of knowledge graphs (Dai et al. 2020a). In this section, we discuss three representative works, including SME, ConvKB, and R-GCN, to briefly introduce neural network-based knowledge graph embeddings.

SME (Bordes et al. 2014) designs an energy function to conduct semantic matching, which utilizes neural networks to measure the confidence of each triplet (h, r, t) in knowledge graphs. The scoring function of SME is defined as follows:The scoring function of SME (bilinear) is:Here \({\textbf{W}} \in {\mathbb {R}}^{d\times d}\) denotes the weight matrix, \({\textbf{b}}\) indicates the bias vector. \({\textbf{h}}\), \({\textbf{r}}\), and \({\textbf{t}}\) are the embedding vectors of head entity, relation, and tail entity, respectively.

ConvKB (Nguyen et al. 2017) utilizes a convolutional neural network (CNN) to conduct knowledge graph embeddings. ConvKB represents each triplet (h, r, t) as a three-row matrix \({\textbf{A}}\), which is input to a convolution layer to obtain feature maps. Afterward, the feature maps are concatenated as a vector, and then a score is calculated to estimate the confidence of the triplet. The scoring function is as follows:where O signifies the concatenation operator, \(\textit{g}(\cdot )\) is the ReLU activation function, \({\textbf{A}}*\Omega\) indicates the convolution operation of matrix \({\textbf{A}}\) by using the filters in the set \(\Omega\), \({\textbf{w}}\in {\mathbb {R}}^{3d}\) is a weight vector.

R-GCN (Schlichtkrull et al. 2018) is an improvement of graph neural networks (GNNs). R-GCN represents knowledge graphs by providing relation-specific transformation. Its forward propagation is calculated as follows:where \(h_k^{(l+1)}\) is the hidden state of the entity k in l-th layer, \(N_k^r\) denotes a neighbor collection of entity k and relation \(r \in R\), \(n_{k,r}\) is the normalization process, \(W_i^{(l)}\) and \(W_k^{(l)}\) are the weight matrices.

The existing methods for generating knowledge graph embeddings still suffer several severe limitations. Many established methods only consider surface facts (triplets) of knowledge graphs. However, additional information, such as entity types and relation paths, are ignored, which can further improve the embedding accuracy. The performance of most traditional methods that do not consider the additional information is unsatisfactory. Table 3 lists the embedding methods, which do not consider the additional information. In Table 3, the performance evaluation is based on the link prediction and triplet classification tasks. The metrics that are for evaluation results are hit rate at 10 (Hits@10) and accuracy. As Table 3 presents, only a few models have impressive results, including the results of QuatE (90%), RMNN (89.9%), and KBGAN (89.2%). Recently, some researchers have started to combine additional information with a knowledge graph to improve the efficiency of embedding models. For example, Guo et al. (2015) take advantage of additional entity type information, which is the semantic category of each entity, to obtain the correlation between the entities and to tackle the data sparsity issue. Therefore, knowledge graphs are represented more accurately. Not only entity types, some other information, including relation paths (Li et al. 2021), time information of dynamic graphs (Messner et al. 2022), and textual descriptions of entities (An et al. 2018), are getting the researchers’ attention in recent years. However, it is still a daunting challenge to effectively utilize rich additional information to improve the accuracy of knowledge graph embeddings.

General additional information can not adequately represent the semantic meaning of the triplets. For instance, the entity types are not related to the semantic information of triplets. Furthermore, the types of additional information that can be incorporated into the features of the triplets are now severely limited. Therefore, to improve the performance of existing knowledge graph embedding methods, multivariate information (such as the hierarchical descriptions of relations and the combination of entity types and textual descriptions) needs to be incorporated into the features of the triplets.

To the best of our knowledge, complex relation path remains an open research problem (Peng et al. 2021). For example, the inherent relations, referring to the indirect relationships between two unconnected entities, are not represented effectively. Although the inherent relations between the entities can be explored based on the chain of relationships in knowledge graphs, the inherent relations are complex and multiple. Therefore, it is not straightforward to represent these relations effectively.