| 英文摘要 | Information networks generally refer to directed graphs with specific types of nodes and edges, which are widespread in real life. According to the types of nodes and edges, there are three common forms in information networks: homogeneous information networks, heterogeneous information networks, and knowledge graphs.
Learning representations of information networks and then obtaining potential correlations have a wide range of applications in bioinformatics, recommender systems, information retrieval, and other fields.
In recent years, driven by deep learning and graph neural networks, the representation learning and correlations methods on information networks have achieved outstanding results, but there are still some challenges: (1) homogeneous networks rely heavily on labeled data or negative samples; (2) it is difficult for heterogeneous networks to capture the semantic similarity between subgraphs; (3) high-quality negative samples are relatively sparse in knowledge graphs. To alleviate the above problems, this thesis uses self-supervised learning methods, and takes negative samples as an entry point, then conducts research from three points: learning without negative samples, designing new negative samples for contrastive learning, and generating high-quality hard negative samples based on existing negative samples. The main works and innovations of this thesis can be summarized into the following three aspects:
(1) Research on homogeneous information network without labeled data or negative samples.
The existing learning methods for homogeneous information network mainly rely on the guidance of supervision or the comparison of positive and negative samples, but obtaining supervision or effective negative samples is difficult in real scenes. To deal with the unsupervised scenario without negative samples, this thesis proposes a homogeneous information network learning model based on bootstrapping mechanism. The model contains two parts: the online network and the target network. The core idea is making the online network and the target network learn from each other, which makes the model get rid of supervision and negative samples. Furthermore, considering that the online network and the target network require similar but different inputs, this thesis utilizes two classes of graph-structured data augmentation methods for generating two perspectives of homogeneous networks. Four groups of experiments on three public datasets validate the effectiveness of the model.
(2) Research on heterogeneous information network to capture the similarity between subgraphs.
Since heterogeneous information networks contain rich semantic information, the existing methods generally divide the heterogeneous networks into several homogeneous subgraphs through meta-paths. These subgraphs have strong semantic similarity due to their semantically related topological structures, but the existing methods cannot capture the semantic similarity between subgraphs. To make up for this deficiency, this thesis proposes a contrastive learning model between subgraphs. The model treats two subgraphs with the same node features and semantically related topology as the anchor samples and positive samples, respectively, and then let the representations of the two subgraphs get closer in the vector space. In order to contrast with the close distances of the two subgraphs, this thesis designs negative samples with the same node features but no topology, and makes the anchor samples farther away from the negative samples.
To further highlight the importance of topology, this thesis makes the encoders of the positive and negative samples share the same parameters,
so that the reason why the distances (between anchor and positive samples) are much smaller than the distances (between anchor and negative samples) are only that the positive and anchor samples have semantically related topology.
This thesis conducts six groups of experiments on four datasets to demonstrate the superiority of the model from different aspects.
(3) Research on knowledge graph for mining high-quality hard negative samples.
The core of knowledge graph representation learning is to contrast positive and negative triplets.
Since there are only positive triplets in knowledge graphs, negative triples are generally generated by randomly selecting other entities to replace the entities in positive triplets.
There are two shortcomings in the present methods: first, the negative samples obtained by fixed sampling becomes easier to distinguish with the model's training process, so that the gradients disappear; second, the semantics of the negative samples obtained by replacing the existing entities are single, and the semantic information of different negative samples cannot be integrated. To cope with these deficiencies, this thesis generates high-quality hard negative samples through mixing operation. To achieve dynamic negative sampling, this thesis proposes two criteria for selecting existing hard negatives, which allows the model to select hard negatives in different periods of training. In order to fuse the semantic information of different negatives, this thesis uses virtual entities to generate hard negative samples through mixing operation, which offers more valuable gradients. Four groups of experiments on two datasets and four scoring functions show that the proposed model can generate higher-quality hard negatives, and achieve performances that surpass previous negative sampling methods. |
修改评论