Author name disambiguation literature review with consolidated meta-analytic approach

Analyzing the reviews presented, we note they classify AND approaches, describe the main characteristics, and propose a taxonomy for classification. In contrast, our review work imposes questions focusing on the most cited works, most relevant authors, most-used approaches, and pursued lines of research. To answer such questions, we use meta-analysis based on statistics to summarize the work results. The Theory of Consolidated Meta-analytic Approach (Teoria do Enfoque Meta-analítico Consolidado - TEMAC [14], in Portuguese) is used to collect data on AND area with quantitative information available in bibliographic repositories, such as the Web of Science (WoS) and Scopus. However, a systematic review answers research questions by collecting and summarizing empirical evidence that fits pre-specified eligibility criteria as suggested by [15, 16]. We think a systematic review would add investigation threads to our work in a complementary way but not replaceable. In addition, a meta-analytical AND review has not been conducted to date, creating a foundation for forthcoming research endeavors that will contribute to the existing body of knowledge built upon prior review studies [5, 11, 12].

The following sections present the literature review methodology, results and analysis, and conclusions with future work.

The TEMAC meta-analytical overall framework includes quantitative techniques and bibliometric aspects based on three laws.

The data presentation and interrelation using the consolidated meta-analytic approach of TEMAC allows for a review of the most relevant authors and citations, journals, countries, organizations, or universities, and knowledge areas most related to the research field. To perform the data analysis review, we used the VOSviewer bibliometric tool [24], and the BiblioTools [25, 26].

The VOSviewer tool for visualizing and analyzing bibliometric networks provides insights into patterns, relationships, and trends in the research literature. The VOSviewer allowed the production of visual representations of bibliometric data. The co-authorship and co-citation networks’ visualization with research clusters, influential authors, and new research directions help reveal relationships between authors and documents in review works. In summary, the VOSviewer tool helps to explore large bibliometric datasets, contributing to understanding the landscape of research fields.

BiblioTools is a suite of Python scripts for bibliometric analysis integrable to different digital repositories, with numerous functions, such as data mining, data processing, data analysis, keyword visualization, and automated report generation. BiblioTools makes it possible to refine and clean raw data with a preprocessing script preparing the dataset for analysis. We explored our data, producing a variety of co-occurrence networks, such as co-words, co-authors, and co-citations. The BiblioTools allows the visualization of bibliographic coupling networks and clusters, providing information about publications, authors, and research topic connections.

Using the BiblioTools and VOSviewer, we extracted information as follows.

In the co-citation analysis of Fig. 8, we identify a similarity between the authors’ contributions and their areas of study interest. There are four dark red spots representing co-citation cores. Below, we will detail the leading studies of each cluster and classify them according to the taxonomy used in this literature review (Fig. 1).

This cluster consists of two works, Shin et al. [38] and Ferreira et al. [40], which use co-authorship information for disambiguation. This similarity indicates the proximity in the heat map and justifies the high co-citation of the cluster. However, the computational approach for disambiguation is different. The work of [38] has an approach based on graphs constructed with co-authorship relations to solve the AND problem using the DBLP and Arnetminer databases. According to the taxonomy, we can classify the type of approach as Author Grouping and the explored evidence as Citation Information and Web Information.

The work presented by Ferreira et al. [40] uses a three-step approach for AND. First, using a heuristic based on co-authorship makes the citations clustered. Using similarities, some of these clusters will be selected to become training data in the second step. In the third step, the selected clusters are added into an associative name disambiguator with self-training capabilities. This work is classified according to the taxonomy in the type of approach as Author Assignment and explored evidence as Citation Information and Web Information (i.e., data extracted from DBLP and BDBComp).

While neither of the two works in this cluster suggests a direct solution for the AND problem, they do provide strategies that help resolution approaches. The cluster appearance is justified by their high co-citation in the literature, serving as a basis for other studies. Kim et al. [42] introduce a method for generating labeled data to compose machine learning approaches. With test runs, the proposal achieved high performance compared to works in the literature. Kim [43] implements a framework integrating five validation measures for AND approaches using clustering. This integration may help scholars in the AND area to compare the similarities and differences of the various validation measures before selecting the ones that best characterize the clustering performances of their AND methods.

The authors in [35] present a brief literature review focusing on the definition and challenges of the AND problem. Ferreira et al. [11] conducted a literature review with approaches to AND resolution, suggesting a taxonomy for classifying these approaches. We observed that the two reviews are close compared to the whole heatmap, evidencing a large co-citation of these papers in the studied databases.

Two other works propose approaches to AND. First, Levin et al. [45] present a self-supervised algorithm that uses bootstrap techniques for clustering and a supervised training algorithm. The work uses information from the authors’ citations and other attributes such as email, authors’ names, and language. We classify this work in the type of approach as Author Grouping and the explored evidence as Citation Information.

The work presented by [46] uses a heuristic-based approach for AND with similarity functions of authorship evidence records extracted from DBLP and BDBComp. According to the taxonomy, the type of approach is Author Grouping and explored evidence Citation Information.

This cluster contains two papers by the same authors. The first one of [36], a probabilistic approach, named Authority, to solve the AND problem in the MEDLINE [32] database using information such as title, journal name, co-authorship, language, and other features. Authority computes the similarity between two articles by analyzing the authors’ names and emails. The model also presents ways of automatically generating training sets, methods to estimate the probability between author names, and an agglomerative clustering algorithm based on maximum likelihood to compute clusters of articles that represent the authors studied.

The second work of [34] also uses a probabilistic model for AND, but it only used authors’ names, discarding other information such as email addresses and affiliations. Thus, it is evident that the 2009 work is an evolution of the 2005 one. According to the taxonomy, both papers use the approach type as Author Grouping and as explored evidence Citation Information. The other clusters of co-citation presented by the heat map in Fig. 8 did not present patterns detected by this study.

Figure 9 presents another co-citation analysis using cluster density with clustering among all the co-citation papers and allows insight into other similarities among the documents in each group. Thus, we can analyze the other works not so evident in Fig. 8. Note there are three general clusters: green, blue, and red. A common characteristic is a space-time between the documents. The red has papers from 2005 to 2010 and the blue from 2009 to 2015. This space-time feature does not appear in the green cluster, as it is more diverse with documents from 2004 to 2019. Note there are works on the cluster edges, uniting groups based on the date characteristic, such as [11] that link the blue and red clusters.

The green cluster is quite diverse as there are various types of approaches, such as cognitive maps and network analysis [47], probabilistic models [48], heuristic-based models [49], agglomerative hierarchical clustering [50], and supervised learning [51, 52].

The red cluster cover works using clustering approaches [53‐55]. The study conducted by [56] investigates the influence of co-authorship attributes for solving the AND problem. The blue cluster presents a similarity with research using cluster similarity and agglomerative clustering for AND [57, 58]. In contrast, Strotmann and Zhao [59] presents research that indicates the influence of AND on citation and bibliographic base analysis studies.

The bibliographic coupling analysis allows insights into the current state of the AND research area. We present in this section the AND research fronts, including works from 2019 to 2022. The works are classified considering the AND approach using the taxonomy presented in Fig. 1.

Figure 10 presents a bibliographic coupling of works using a heat map for the merged WoS and Scopus merged databases. Note there are three clusters explicitly numbered, highlighting the current AND research fronts. In the sequence, we present a summary of the works included in the three clusters.

The authors in [39] used a graph node embedding approach to solve the AND problem. This type of solution is inspired by the word embedding model but adapted for a graph structure solution. A graph is constructed with co-authorship relationships, using the random walk method for learning graphs and assigning clusters to unique people in the real world. The approach used CiteSeerX data with results improved compared to similar approaches.

The authors in [41] propose a hybrid pairwise classification method for estimating the probability that an author record is correct in a bibliographic repository. This solution uses global features extracted from text using supervised training on a dataset of an author’s citations. This text classification and the supervised training use word embedding methods such as Bag of Words and TF-IDF with data from PubMed and ArnetMiner. According to the taxonomy, [39, 41] can be classified as Author Grouping approaches because they use similarity calculation with training and machine learning. Authors use word embedding as basis for the AND method, justifying the proximity of the studies observed in the heat map presented in Fig. 10.

In [33], the authors create a knowledge graph with information from the PubMed repository extracting bio-entities from abstracts. In this work, the authors do not propose a new approach to solving the AND problem. However, approaches already known in the literature were used together, such as Authority (uses a graph approach) and Semantic Scholar (uses a binary training classifier to join pairs of author names and create author clusters). The constructed knowledge graph allowed the creation of links between biological entities, articles, authors, and affiliations. In the AND step, the results achieved F1 scores of 98.09%. We can classify the approach of this work as Author Grouping and the explored evidence as Citation Information and Web Information.

The work of [44] created an automatic system for data availability declarations in bibliographic repositories using PubMed. The authors compare first and last names with string similarity techniques. It appears in the heat map of the literature because it cites several influential AND works.

With the literature coupling analysis, we note a current use of grounded techniques for AND with Author Grouping and Clustering methods. The first article found in the dataset of this review dates back to 2003. Since the bibliographic coupling seeks to obtain current research fronts, works from 2019 to 2022 were selected for the bibliographic coupling analysis. Analysis was conducted to classify and present the most recent articles (2020–2022), classifying them according to the taxonomy in Table 8 and Fig. 11.

Based on the co-citation analysis performed across the studied time-space (2003–2022), the works present mainly the use of author grouping approaches. This analysis result is consistent with the findings of [4]. Moreover, the conducted coupling bibliographic analysis points to viable alternatives to address AND-related problems, indicating a wide range of research in the area.

In this section, we present an overview of recent AND works published between 2020 and 2022, arranged by the taxonomy of Fig. 1 [4, 11]. The taxonomy is also used in the co-citation (Sect. 3.2.1) and coupling analysis (Sect. 3.2.2) sections. This time frame was chosen as earlier reviews covered works until 2019 [11, 12]. The coupling analysis includes some works presented in this section.

The overview of current research works is essential to complete the meta-analytical analysis by citing the particular techniques, strategies, and emerging themes within the AND research area. To organize the works overview analysis in a concise way, we guided it by the book devoted to AND study in bibliographic repositories [4]. We present a synopsis of each work included in Table 8. The work synopsis presents the AND approaches, including author grouping, mainly through agglomerative clustering, standing out as prevalent methods in recent studies.

The author grouping approach, as defined in Sect. 1, is especially appropriate for datasets with lots of co-authorship data. Large bibliographic datasets can benefit from this approach, unlike the author assignment approach, as it does not depend on time-consuming manual labeling author annotations. Agglomerative clustering in author grouping is a common approach as it is simple to use and can produce hierarchical clusters to be analyzed at different granularities. This clustering method provides flexibility when creating author groups. Figure 11 shows how the works relate to each other in the used taxonomy.

We identified a set of five works using tree-based learning models, such as Gradient Boosting, Random Forest, and Decision Tree [60‐64]. The works may use other Machine Learning techniques in conjunction with the mentioned, such as Naive Bayes [62], Logistic Regression [60, 61], and Network Graphs [63].

Some works address distinct supervised techniques, such as [10] that use transfer learning. The authors in [65] show that ORCID can validate the performance of supervised AND methods that use large-scale bibliographic data. The authors in [66] used the DBLP database with a neural network to learn the representations of coauthors and titles so AND could consider the similarity between these attributes.

Li et al. [67] present an algorithm with multiple similarity strategies for AND implementing using collaboration network calculations, affiliation, and publications attributes of authors. Another multi-strategy approach is presented by [68] using string comparison with Jaccard similarity, Levenshtein distance, and co-authorship network comparison. Waqas and Qadir [69] propose a multilayer heuristic with a clustering approach. The clustering uses attributes inherent to the author and publication, such as title, abstract, keywords, email, and affiliation. Word embedding Word2Vec is used to extract the attributes.

D’angelo and van Eck [70] use a rule-based scoring approach with author, publication, citation, and institution attributes. Clusters with meta-data allow indexing of a particular author to disambiguate. Zhang et al. [71] use heuristic rules combined with neural networks to analyze publication attributes, such as title and affiliation. An advantage of this method is the possibility of extending the method’s application to other datasets.

Mozafari [72] proposes a genetic algorithm for determining the similarity coefficient between two authors for AND. The algorithm determines the importance of the attributes in the publications, electing an optimal coefficient for comparison between authors.

Jinqi et al. [73] propose an algorithm to put entities and resources into a network graph to set the resource node capacity-based sharing degree. The network graph uses relationships between the author and publication nodes to calculate the flow capacity between nodes which allows clustering of the graph.

Zhang and Ban [64] use publication relationships to construct the graph, with the strongly related publications grouped, forming atomic clusters and reducing the graph size. At another stage, a rule-based similarity algorithm analyzes and combines the feature information from the publication graph to perform AND.

Zhou et al. [74] present an approach with five graphs formed by publication attributes, co-authorship, location, title, keywords, and affiliation. Each attribute creates the node where the edges are the similarity weights between publication pairs. A fusion graph of the attributes is built. A random walk algorithm is applied to the graph to determine paths that represent the local node structural information. Then, a multilayer perceptron algorithm is applied to the graph structure.

Santini et al. [93] propose a Knowledge Graph Embeddings (KGE) using information from the AMiner database. The KGE has three parts: multimodal information extraction from the KGE, a blocking procedure, and hierarchical agglomerative clustering. Qiping et al. [92] use citation information to construct a heterogeneous information network. Representation learning for clustering the authors and disambiguation is applied, and cluster analysis with rule matching is performed.

Ma et al. [76] propose incorporating a Word2Vec model into a Ghaph-Based approach. The algorithm extracts attributes and the relationships between publications, authors, and co-authors. Word2Vec serves to obtain these features allowing the insertion of other features that may appear in the dataset. Subsequently, a graph with relationships between publications and authors is built. Then, an algorithm for clustering and similarity analysis between nodes and edges is applied.

Pooja et al. work presents solutions using a graph-based approach as a basis associated with other computational techniques (e.g., Clustering). The work [80] uses a graph-based clustering approach for AND. Jaccard and Cosine similarity characterizes the relationships between authors and publications in the graphs, and Web information refines the results. In [85], the authors use graphs with publication attributes and a Word2Vec embedding model to create vectors that will serve as input to an agglomerative hierarchical clustering (HAC), which is widely used in AND studies. In [9], the authors use a graph-based approach configured with multi-hop neighborhoods and apply HAC for AND in the final step of the algorithm. The work in [95] uses graphs to build the network of authors and publications and uses clustering for disambiguation. However, the differential of this new approach is the ability to work with online information from digital bibliographic repositories.

The author in [75, 79, 82, 84] use word embedding, graphs, and clustering respectively. Ma et al. [77] use the same approach applied to robotic literature consultants.

The work of [78] proposes a technique with partial classification in three steps to solve the AND problem. The first uses a probability propagation constraint to infer the distribution of a given author’s name. In the second step, a portion of the author name in the documents is linked to their respective authoring if the model exhibits high confidence. In the last step, the initial classification algorithm parameters are updated.

Firdaus et al. [81] propose two methods for AND. In the first work, the technique uses four steps for disambiguation: data labeled, publication attributes extracted, deep neural network, random forest, naive Bayes, and SVM classification are done, and the validation of the result comparing the classification techniques. In the second work, Firdaus et al. [87] uses classification with deep neural network technique is increased with cost-sensitive learning considering cost variation from unclassified data.

Manzoor et al. [90] use convolutional neural networks for unbalanced and balanced dataset classification. According to the authors, the solution is flexible by learning the attributes without concatenating similarity measures. The same method is also Single Citation Based which preprocesses the dataset efficiently, decreasing computational costs.

Farber and Ao [91] do not propose a new approach but use an unsupervised rule-based classification method. The method does not require data training, adapted for the authors’ proposal.

Correia et al. [83] propose a crowd-systems-based prototype allowing interaction and contribution from the Web for the general public correcting name ambiguities, missing data, and incorrect references in a digital bibliographic repository. Backes and Dietze [89] present a technique for progressive AND with lattice structures for name inclusion. Waqas and Qadir [94] do not propose an AND resolution but present a dataset to assist developers. The “CustAND” labeled dataset with 7886 publication records is presented using data from DBLP and Google Scholar.