3.2.1. Profiling knowledge landscapes

While a topic is defined a collection of terms representing similar semantic meanings, the knowledge landscape of a given domain is described as a set of topics and their relationships. Targeting to the list of cleaned terms, co-occurrence analysis is applied for initially measuring the relationships between terms [38], with a hypothesis that if two words frequently appear together, they are similar [39]. The corresponding algorithm is described below:

• Given that T=t1,…,ti,…,tn is the list of cleaned terms and term ti could represented by a vector C1={c1,i,…,ci,i…,ci,j,…,cn,i}, in which ci,j represents the frequency of the co-occurrence between terms ti and tj in the dataset;

• The similarity Simti,tj between terms ti and tj could be calculated by the Salton's cosine [40], i.e.,

  Simti,tj=Ci⋅Cj|Ci||Cj|

• A triangle similarity matrix ST is then generated, which records the similarities between all pairs of terms in list T;

• A network GtT,Et is constructed based on matrix ST, in which each term is represented by a node while Et represents the set of edges in the network and is filled with the similarities Simti,tj.

• Regarding to the map of science [41], an approach of community detection is applied to visualize the network GtT,Et through certain communities, which are then identified as topics in a given domain.

Note that despite numerous text similarity algorithms in the literature, our pilot studies have examined that the cosine similarity algorithm could achieve the best performance in bibliometric datasets [42], and thus, the entire study of this paper exploits Salton's cosine algorithm for similarity measurements.

The science map as well as the network GtT,Et are considered as the outcomes of this function, which profile knowledge landscapes of a given domain in a vivid manner. Additionally, the structure of communities and terms provides a hierarchical solution to represent and organize knowledge and helps understand a domain at a macro level.

3.2.2. Identifying key players and detecting potential collaborations

The identification of key players follows the algorithm of co-occurrence analysis presented above but using lists of cleaned entities such as authors A, research institutions O, and countries/regions R. Thus, the generated networks are denoted as GaA,Ea, GoO,Eo and GrR,Er respectively. Such science maps and networks visualize key players and their existing collaborations, as well as their research groups that are identified as communities in the maps.

A link prediction approach [43] is applied for analyzing the topological structure of the generated networks to detect potential collaborations for authors and research institutions. The basic assumption of link prediction approaches is that: if A and B links with C respectively, A will link with B in the near future, and thus the task of link prediction approaches is to fulfill those missing links with weights.

Considering two types of collaborative strategies—i.e., maintaining existing collaborations and establishing new collaborations, two lists of recommendation pairs will be provided, in which one is for existing collaborators and the other is for potential collaborators with whom they have never collaborated before.

3.2.3. Tracking knowledge trends and predicting emergent topics

A general concern for knowledge trends is the potential change underlying certain given knowledge over time. For example, in the early 2000s and earlier, “data mining” mostly referred to studies in database management and data warehouse, but now “data mining” closely interacts with machine learning techniques which were not common in those early days. Given this evolutionary history occurs, our functional aim is to track knowledge trends and predict emergent topics by considering such changes. The main function is refined from the approach of SEP [10], for which related definitions are given below:

The stepwise algorithm is described as follows:

Step 1: Set S0 as the initial time slice, and group all involved articles as one topic Tpc,r,S0 and consider it as the initial topic of an evolutionary pathway.

Step 2: Iteratively process the data corpus Φ as a simulated stream—i.e., process one time slice once by analyzing its articles one by one.

Step 3: Measure the similarity sima,Tp between a forthcoming article a and centroids of all alive topics via Salton's cosine [40]—i.e.,

Step 4: Assign article a to the most similar topic Tp, and calculate the Euclidean distance Ea,c between a and the topic's centroid c. If Ea,c<r, the article will be directly assigned to the topic, or else, the article will be labeled as “drift.” Then, return to Step 3 and analyze the next article.

Step 5: At the end of each iteration, set a topic as “dead” if it does not receive any articles in two continuous time slices, which means since this topic is generated, there is not any new knowledge accumulated to it in the following time slice. Then, logically “live” topics continue, given that it is in time slice Sx, we iteratively apply an unsupervised K-means approach [3] to each topic and group their assigned articles labeled with “drift” into certain sub-topics Tp′c′,r′,Sx.

Step 6: Measure the cosine similarity between Tp′ and two sets of topics—the assigned topic Tp and all dead topics Tpd. If SimTp′,Tp>SimTp′,Tpd, the relationship between Tp and Tp′ is defined as “descendent-predecessor,” or else, the “dead” topic Tpd will be resurged and set as “alive,” and then, becomes the predecessor of topic Tp′.

Step 7: Label topic Tp′ via the term with the highest similarities with all other terms in this topic—if the term has already been used by existing topics, choose its following terms.

Step 8: Update all alive topics by recalculating their centroid and boundary and return to Step 2 until the stream ends.

The outcome of the SEP approach is a list of topics with information such as labels, descriptions, involved terms and articles, and “sleeping beauties”-related indicators (e.g., born time, dead time, and resurgence). These topics could be visualized in a direct network, in which each topic is represented by a node and weighted edges represent the similarities between their connected nodes. It is clear that this network provides a solution of tracing knowledge trends by detecting such predecessor–descendant relationships and predicting emerging topics by identifying topics of “sleeping beauties.”

Note that the SEP approach exploited in this study mostly follows the version presented in [10] but aiming to further omit human intervention and adapt to practical needs by consulting with certain domain experts, we modified the algorithm from the following aspects: 1) in Step 1, only one initial topic is grouped rather than applying a K-means approach to group records in a given number of topics; and 2) an unsupervised K-means approach is applied to take the place of the hierarchical clustering approach in Step 5.

3.2.4. Criteria-based knowledge searching and ranking

The function of criteria-based knowledge searching and ranking is designed to conduct data argumentation with limited expert knowledge. The input of the function is a core collection of scientific documents, which may also be manually collected by domain experts and indicates their existing knowledge base,² and the task of this function is to extensively collect relevant articles from the entire dataset based on the core collection and return a set of relevant articles.

Two ranking lists are initially generated—one is based on the cosine similarities between the vectors of individual articles and the core collection, which are created by a Doc2Vec model [45]; and the other is based on the grade of membership that an individual article belongs to the core collection, in which fuzzy sets are involved. Then, the nondominated sorting genetic algorithm II (NSGA-II) [46] is exploited to identify nonnominated solutions (i.e., considering a multi-objective optimization problem, solutions that are superior to the rest of solutions when all objectives are considered, but are inferior to other solutions in one or several objectives [47]), considering both ranking criteria. The stepwise algorithm of this function is described as follows:

Step 1: A Doc2Vec model based on the Word2Vec model [45] is applied to the entire data corpus (including the core collection Δ provided by domain experts), and each article a is represented by an abstract vector.

Step 2: A search strategy consisting of a set of search terms is provided by users, and a combinative search is conducted to return a set Ρ of articles a′ that coincide with the search strategy.

Step 3-1: The mean Mc of the core collection is calculated as Mc=∑a′, and then measure the cosine similarity between Mc and a′. Return a ranking list R1 based on the similarities.

Step 3-2: Given that Ψa′ and ΨΔ is the set of specific tags (e.g., Mesh terms and keywords) of one searched article and the core collection respectively, and FΔa′ is the membership grade that article a′ belongs to the core collection. The membership function FΔ is defined as follows. Then, rank articles a′ based on the membership grades and get a ranking list R2.

Step 4: Exploit the fast-nondominated-sort approach of NSGA-II [46] to calculate the number of articles da′ that each article dominates (i.e., inferior to article a′ in both ranking lists), and then rank all the searched articles based on da′—i.e., the larger the higher, the integrated calculation procedure is given in Algorithm 1.

The function provides a solution of augmenting a specified knowledge base with limited expert support, in which semantic similarities between scientific articles are exploited.

4.2.1. Profiling knowledge landscapes

A co-term network GtT,Et was constructed, including the top 5000 high frequency terms and 1,735,189 edges, with the aid of VoSViewer [41],⁶ a smart local moving algorithm was applied for visualizing the network, as given in Figure 2.

Figure 2

Co-term network for profiling the knowledge landscape of gene-related cardiovascular diseases.

According to Figure 2, the knowledge landscapes of gene-related cardiovascular diseases are illustrated by four core groups: coronary artery diseases (blue nodes), atrial fibrillation (green nodes), molecular mechanisms and heart failure (yellow nodes), and protective effects (red nodes). If targeting one specific node (e.g., atrial fibrillation), its co-occurrent relationships with other nodes could be observed as well, as given in Figure 3.

It is clear that this function provides a vivid way to profile the knowledge landscapes of a given domain by identifying core topics and their relationships.

Figure 3

Co-occurrence relationships of atrial fibrillation.

4.2.2. Identifying key players and detecting potential collaborations

Three coauthorship networks GaA,Ea, GoO,Eo and GrR,Er were generated respectively for individual researchers, research institutions, and countries/regions of gene-related cardiovascular diseases. As an example, the country-based coauthorship network GrR,Er is given in Figure 4.

Figure 4

Country-based coauthorship network for identifying key players in gene-related cardiovascular diseases.

As shown in Figure 4, the United States, the United Kingdoms, France, the Netherlands, Australia, and China are leading the research of gene-related cardiovascular diseases, and the strengths of their collaborations with other countries could be observed as well. Similarly, based on the coauthorships, active researchers, institutions, and countries with strong collaborations with others could be identified as the key players in a given domain.

Furthermore, since a coauthorship network could meet with the requirements of complex networks [49], such as scale free and small world, network analytics could also be applied to further analyze the topological structure of the network. Considering an institution-based coauthorship network GoO,Eo as an example and choosing St Vincent's Hospital (Australia)⁷ as a target institution, a resource allocation approach (as is has been proved that this approach achieves better performance than other algorithms like Adamic/Adar index and common neighbors [50]) was applied to fulfill missing links in GoO,Eo, which then could be facilitated to detect potential collaborations between unconnected nodes. The predicted further collaborators and potential collaborators in the area of gene-related cardiovascular diseases for St. Vincent's Hospital are given in Table 2.

This function creates a way to investigate the key players of a given domain by identifying who they are, exploring how they collaborate with each other, and predicting how such collaborations will be in the near future.

4.2.3. Tracking knowledge trends and predicting emergent topics

The refined SEP approach was applied to the 142,877 articles, which were simulated into a bibliometric data stream with 12 time slices, covering the time period from 2008 to 2019. The SEP for tracking knowledge trends of gene-related cardiovascular diseases is given in Figure 5.

Figure 5

Scientific evolutionary pathways for tracking knowledge trends of gene-related cardiovascular diseases.

208 nodes and 207 edges—i.e., their descendent-predecessor relationships were generated. Each node represents a specific topic, with its detailed information on the year of born, the number of articles, the number of terms, the value of term frequency inverse document frequency (TFIDF) analysis, and the number of survival batches. Specifically, regarding the design of “sleeping beauty” detection, four types of nodes were classified:

• Always alive, indicating a continuous interest from the community to this topic.

• Resurgence and alive—i.e., the “sleeping beauties” which became dead topics, then were awaken by the involvement of possible new concepts, techniques, materials, and devices, and are still a core interest of the community.

• Dead with resurgence, indicating former “sleeping beauties” but have already become dead topics again.

• Dead without resurgence, indicating topics either which are meaningless to the community or whose potential has not been discovered yet.

The descriptive statistics of nodes in the four types are given in Table 3. As shown, topics in resurgence and alive have the highest average value of TFIDF, indicating its relatively high-quality and emerging potential.

For a reference, the 14 topics in resurgence and alive and their related information are provided in Table 4, and coinciding with the concept of “sleeping beauties,” these topics are considered as emergent topics in the area of gene-related cardiovascular diseases.

This function exploits streaming data analytics and machine learning techniques to draw the knowledge trends of a given domain and the involvement of “sleeping beauty” detection creates a manner to predict emergent research topics in the area.

Aiming to compare the benefits of the proposed method, we list 20 terms with the highest term frequency in Table 5. These 20 terms were collected from the term clumping process (i.e., Step 9 in Table 1) and were normally considered as the outcome of traditional bibliometrics for identifying key research topics.

It is clear that the key outcome of the proposed method is Figure 5, in which all terms in Table 5 could be identified, and then Table 4 provides a supplementary source to identify emerging topics (i.e., significant and within increasing interests to the community). Initially, the overlaps between Tables 4 and 5 (e.g., “risk factors” and “coronary artery disease”) could the potential cross-boundaries in emerging topics and key topics in the area of gene-related cardiovascular diseases. We conducted an extensive consultation with our clinical and research experts, and they agreed with the results based on their expertise and domain knowledge. They also concluded the following observations and may anticipate our further studies: 1) the life of “sleeping beauty” terms may be due to more specific knowledge requiring changes in use of terminology to define the areas of emergent interest in clinical and research areas; 2) they expressed no surprise that these topics and terms are evolving and appear interesting for further research in papers such as this and for validation in clinical settings i.e., the findings from the use of the KBC are not discordant with clinical experts understanding of the topics.

4.2.4. Criteria-based knowledge searching and ranking

A user-friend interface was specifically designed for this function, which consists of 2 sets of auto-completed input textboxes for search terms, an input textbox for the expert knowledge-based core collection, and an output panel for listing ranked relevant documents searched from the entire dataset. The interface is given in Figure 6. Specifically, there are five steps to utilize this interface:

Step 1: Import the full knowledge base—the dataset;

Step 2: Select a group of entities (including diseases, symptoms, drugs, etc.), terms and/or another group of gene terms as inputs, those selected terms are automatically combined by “OR” inside the group and “AND” between the group to form search criteria, representing the co-occurrence of 2 term-groups from a bibliometric perspective.

Step 3: Click button “search,” articles which meet with the search criteria will be retrieved and related bibliographic data is displayed on the bottom panel.

Figure 6

Interface for the function of criteria-based knowledge searching and ranking.

Step 4: Pick out a set of core collection which contains one or more article(s) from the dataset based on expert opinion and type in their PubMed IDs, then click button “Confirm,” the details of picked set will be displayed.

Step 5: Click button “Calculate Semantic Similarity” to calculate the similarity of every single article in search results with the set of core collection based on a pretrained doc2vec model. All the searching results will be redisplayed and ranked by the similarity from high to low in the text field of “Semantic Ranking Results.”

Step 6: Click button “Calculate MeSH Similarity” to calculate the similarity of every single article in search results with the set of core collection based on the num of mutual their MeSH terms. All the searching results will be redisplayed and ranked by the num of mutual MeSH terms from high to low in the text field of “MeSH Ranking Results.”

Step 7: Click button “Comprehensive Ranking” to apply fast-nondominated-sort approach of NSGA-II to the semantic and MeSH ranking lists, a new comprehensive ranking list would be generated. All the searching results will be redisplayed and ranked by the new ranking value from high to low in the text field of “Comprehensive Ranking Results.”

This function directly interacts with users and provides a solution of knowledge search and augmenting an existing knowledge base with limited expert support.