Author Name Disambiguation on Heterogeneous Information Network with Adversarial Representation Learning

Given an author name reference $a$, let $P^{a}=\{p^{a}_{1},\dots,p^{a}_{N}\}$ be a set of $N$ papers written by the authors with name $a$. Each paper $p_{i}^{a}\in P^{a}$ has the content feature set $I_{i}^{a}$ including title, abstract, publish date, and etc. And it has the relation feature set $R_{i}^{a}$, which contains the relation of paper $p^{a}_{i}$ to the entities in the academic domain including co-author, institute, field of study and venue. Given this, we define the problem of author name disambiguation as follows.

We omit the superscript $a$ in the following description if there is no ambiguity.

Furthermore, let $y_{ij}$ denote the homogeneity of papers $p_{i}$ and $p_{j}$, where $y_{ij}=1$ if $p_{i}$ and $p_{j}$ are homogeneous, and $y_{ij}=0$ otherwise. We denote the generated negative samples as $S_{generated}$ consisting of $(p_{i},p_{j},y_{ij}=0)$, and the pseudo-positive samples for self-training as $S_{pseudo}$ consisting of $(p_{i},p_{j},y_{ij}=1)$.

We solve this task with the help of academic heterogeneous information network (HIN), thus content information and relation information can be efficiently processed. We define the HIN as follows:

The core part of our model is shown in Figure 2, integrating content information and relation information of the papers in an adversarial way. A self-training strategy is added to our discriminative model, which uses top-relevant papers as positive samples iteratively. And to make the generative module aware of relation information, following (?), we design a random walk based generating strategy. Given a paper $p_{k}$, we design two modules as follows:

Discriminative module. $D(p,p_{k};\theta_{D})$, which outputs a scalar possibility of whether two papers are written by the same author, i.e., $\text{Pr}(y_{ik}=1|p_{i},p_{k})$, where $p_{i}\in P$.

Generative module. $G(p|p_{k};\theta_{G})$, which learns to select the possible homogeneous papers under the guidance of reward. It will iteratively approximate the true underlying homogeneity distribution $\text{Pr}_{true}(p|p_{k})$.

Two modules are combined by playing a minimax game: the generative module will try to choose the papers possibly written by the same author as the given paper $p_{k}$, and therefore can fool the discriminative module; the discriminative module will distinguish between the selected papers and the ground truth papers. Formally, generative module $G$ and discriminative module $D$ are playing the following two-player minimax game with value function $V(G,D)$:

The trainable parameters are the representation of all papers. They are learned by alternately minimizing and maximizing the value function in Eq.  (2) until the training procedure converges.

Based on the final representation $\vec{d}$, $\vec{g}$ of papers, we perform hierarchy agglomerative clustering (HAC) to partition $N$ papers into disjoint homogeneous sets.

The process for the author name disambiguation is summarized in Algorithm 1.

To evaluate the proposed method, we collect two real-world author name disambiguation datasets for experiments:

• •

  AMiner-AND¹¹1https://www.aminer.cn/na-data. The dataset is released by (?), which contains 500 author names for training and 100 author names for testing. We construct the heterogeneous network including papers, co-authors, author affiliations (which are referred as institutes in our model), keywords (which are referred as fields of study in our model) and venues. However, there is no abstract in this dataset, so we can only use the titles as our content information in the experiment on this dataset. To illustrate our model’s ability to combine content information and relation information and to support the researches which study the author name disambiguation task using content information, we construct a new dataset collected from AceKG (?). The benchmark dataset consists of 130,655 papers from 17,816 distinguished authors. Each sample has the relation information and content information required by the proposed model. The labeling process is carried out comprehensively based on the e-mail address of authors, the co-author information and the institute information.

We compare our model against three state-of-the-art name disambiguation methods. We perform the hierarchical agglomerative clustering algorithm based on the results from these models and compare them by the pairwise Precision, Recall, and F1-score.

? (?): This model constructs three networks under each name reference. The vertices are authors and papers. The weights of edges represent the connections among them. A designed network embedding is learned with an aim to preserve the connectivity of the constructed networks.

? (?): This model trains a function to measure the similarity between each pair of papers using the carefully designed pairwise features, including author names, titles, institute names etc.

AMiner (?): This model designs a supervised global stage to fine-tune the word2vec result, and designs an unsupervised local stage based on the first stage. In the local stage, it constructs a paper network, where the weight of edge reflects the similarity among papers. Then it uses graph convolutional network to preserve the connectivity of the paper network and learn the representation of papers.

To further evaluate the performance of each module, we also compare our performance at different stages.

Con. This is the result based on the content representation result produced by Doc2vec module. This module represents the abstract and title information by a vector.

Rel. This is the result based on the relation representation results, which maps the nodes in heterogeneous information network into low-dimension representation space.

Dis. The result is from discriminator which aims to distinguish whether two papers are homogeneous based on information representation and relation representation.

Gen. The result is from the generator which aims to approximate the underlying homogeneity distribution and to extract the high-order connections on HIN.

We examine our model with several state-of-the-art models on AMiner-AND and AceKG-AND. In the experiment on AMiner-AND, we use 100 names for testing and compare the result with the results of other models reported in (?). In the experiment on AceKG-AND, we sample 85 names for testing. Since ? and AMiner are supervised algorithms, the results from 5-folds cross-validation are reported. Hierarchy agglomerative clustering is performed on the results produced, where the number of clusters is given in advance.

Table 2 shows the overall performances of different models on two datasets. All the reported metrics are the macro-averaged scores of each metric of all test names. Our model outperforms all the other baselines by at least 5.13% and 13.00% in F1 score on the two datasets respectively. On AMiner-AND, our model outperforms the baselines in terms of F1-score (+10.11% over Zhang et al., +9.82% over ? and +5.13% over AMiner relatively). On AceKG-AND, the superiority is the same. As shown in Table 1, almost all the metrics of 15 random name references are improved by our model, which demonstrates the significant superiority of our proposed method.

To evaluate the performance of each module, we also present our performance at different stages in Table 3. It can be seen that the generative module achieves the most significant result. It can mine some high-order connections among papers and thus covers more homogeneous papers as candidate set.

Content representation module achieves a good result on AceKG-AND, while the result on AMiner-AND is low. Because this dataset only provides title as content information. The experiment has illustrated that content information like abstract is valuable for this task.

The discriminative module achieves the highest Prec on two datasets. Because it mainly measures the pairwise similarity, the papers written by the same author can be discovered precisely. The problem is that it solves the problem from a local perspective, which leads to a low Rec result.

Experiments show that relation representation results achieve F1-scores 56.34% and 53.65% on two datasets respectively. For those homogeneous papers which are connected tightly by the relations, they are close in the relation representation space, which works for the clustering stage. However, for those which are content related but have few relations, this module can not group them together.

To dig into how each module works, we visualize the results of each stage in a 2-D way, which is presented in Figure 5. We analyze the layout of blue points in feature space. After a global measurement by content representation module, the papers in the same $C^{a}$ are preliminarily clustered together in Figure 5(a). Figure 5(b) shows the results of relation representation module. It can be seen that homogeneous papers are grouped much better. The clustering results of discriminator and generator are much better, for they consider both of the content information and relation information. The blue points are grouped into one cluster successfully. And clusters in Figure 5(d) have clearer boundary than clusters in Figure 5(c), which corresponds to the fact that the generator achieves a better result than discriminator.