Identifying experts is an important challenge in many contexts. In this paper, we present a method to build a knowledge graph by integrating data from Google Scholar and Wikipedia to help students nd a research advisor or thesis committee member. This knowledge graph is used to power the exploratory search interface to recommend similar
Identifying experts is an important challenge in many contexts. The nature of this challenge is to nd a knowledgeable person with an advance expertise in one or more target topics among a large number of potential candidates. A wellexplored example of this task is nding an expert for a speci c project within a large company or nding a doctor with advance knowledge of a speci c disease in a large city. While in these two contexts, large companies and hospitals use knowledge management techniques to catalogue key areas of expertise and use it to represent information about each expert, nding experts in other contexts could be more challenging.
The context that we target in this paper involves students nding a research advisor. Each year, undergraduates, master-level and doctorate students face the di cult challenge of nding a research advisor. While large universities have many highly knowledgeable faculty, nding one with the expertise that matches the student's interests, requirements, and preparation is a challenging task. Whether the task is nding an advisor for a summer research project, a faculty sponsor for an independent study, or a committee member for a doctoral thesis, online sources frequently fail the students, and they resort to 'word of mouth' within a limited circle of instructors, classmates, and university sta . One problem in using online sources is the wide variety of sources with relevant information that can exist (e.g., department directories, publication sites, funding agency pages, personal home pages, etc.) Each of these sources covers only some aspects of the faculty member's expertise and frequently represent only a subset of available advisors. Despite these di erent sources, there is typically a lack of "expertise catalogs". A university usually o ers a catalog of courses and majors, but not a ne-grained catalog of expertise areas covered by faculty. As a result, students frequently cannot even properly name their target area of interest or formulate a Web search query when looking for advisors.
The focus of our project is to o er a single-access-point exploratory search system, which allows students to discover their target areas of interest and nd relevant advisors within these areas. In its core, the platform uses a knowledge expertise graph, which represents multiple connections between research topics and prospective research advisors within a large university or a large research eld. We built this graph by processing several knowledge sources about faculty and their research interests. This paper brie y reviews the type of knowledge graph we built, the process of extracting information for its development, and the information exploration system powered by this knowledge. 2
In the past, there have been attempts to build \a map of science" representing
most important areas of research expertise and their connections with experts;
however, the lack of proper information sources makes it hard to produce maps
that are suitable for nding advisors. Examples of this attempt to build a map
of science is presented in [
Google Scholar has been long recognized as one of the best freely accessible
academic information sources in terms of coverage and accessibility [
Wikipedia is commonly used by researchers to compute the semantic
relatedness of concepts between and within documents [
To support students in nding advisors, we created a knowledge graph using data from Google Scholar and enriched it semantically using Wikipedia. In turn, this graph was used to power an interactive exploratory recommendation interface, which makes the task of advisor- nding easier, especially for students with a limited level of knowledge and familiarity with the subject of research. To support several advisor- nding scenarios, we built several versions of the graph. The graph presented in this paper is focused on the task of nding a top expert in a speci c topic of interest within some broad eld of research (such as Articial Intelligence) across many universities. This is a typical task for a student selecting a doctoral program to join or for a senior doctoral student looking for an external thesis committee member. 3.1
Google Scholar We utilized the information of 1000 active scholars in two popular elds of computer science: Arti cial Intelligent and Computer Architecture (focusing on the top 500 scholars in each eld). For each individual, we extracted the following information (see Table 1): { Name: Full name of the scholar. { A liation: The university or research institution the scholar is a liated with. { Veri ed Email Domain: Used to check the validity of the scholar pro le. { Self-De ned Keywords: A list of up to ve keywords de ned by scholars to describe their research interests. { Citations: The total number of citations received by all of the scholar's publications. { h-index: The h-index measures the citation impact and productivity of a scholar's publications. We use this measure alongside other quantitative scores to re-order the results of the recommendations. { i10-Index: i10-Index describes the total number of the scholar's publications with 10 citations or more. This score, which is only used by Google Scholar, was also used to re-rank the results of the recommendations. { Recent publications (20): We used the 20 most recent publications to generate additional keywords representing the current interests of each scholar. The keywords were extracted from the titles of recent publications as follows: After removing stop-words, we generated all of the possible keyword candidates as uni-grams, bi-grams, and tri-grams. Next, we only kept the keywords that have an entry in Wikipedia (see keyword veri cation below). { Top Co-Authors (10): For each scholar, we extracted a list of the top 10 co-authors from their Google Scholar pro le. Wikipedia We used Wikipedia to add a semantic layer to pro les extracted from Google Scholar. Throughout this process, we also obtained useful information that led to a stronger connection between keywords and enables us to add weight to each scholar-keyword relation. The Wikipedia API has been used for the following purposes: { Keyword veri cation: As mentioned before, we collected two sets of keywords for each scholar: self-de ned keywords and keywords extracted from recent publications. The Wikipedia API has been used to verify the validity of these keywords by using fuzzy match techniques to nd the Wikipedia entry describing the keyword. We removed all keywords that did not match with any article in Wikipedia. While Wikipedia might miss articles for some less popular research topics, we need to have all topic keywords explained for the student audience and a match to a Wikipedia article was the best way to assure it. For all remaining keywords, we calculated the association weight between a keyword and a scholar as cosine similarity between the full-text Wikipedia entry of each keyword and concatenated text from the scholar's recent publications. { Entry Summary: To o er student users a short description of each topic keyword, we collected page summaries for all keywords using the Wikipedia API. { Top relevant keywords (10): Most (if not all) Wikipedia pages have multiple links to similar or related articles. We collected the top 10 links based on the number of their occurrences in each page. We employ these links to create a highly connected network of keywords. { Entry Categories: Wikipedia uses categories to group similar articles. We extracted all categories associated with a page and used a full Wikipedia category hierarchy schema to nd relationships between categories in our data set.
We used the Neo4j graph database to represent information about all scholars and keywords. The overall schema of the knowledge graph is represented in Figure 1. As is shown, there are three distinct node types in the graph:
{ Individual: This node type conveys demographic information about scholars including full name, a liation, veri ed email domain, and URLs of personal homepages and Wikipedia pages (if they exist). Individual nodes are connected via "works with" links which represent the co-authorship relations between scholars. An individual node also connects to several keyword nodes that represent the scholar's research interests and expertise. The individual nodes with a dashed border represent scholars added via co-authorship extraction who are not among the top 500 extracted scholars. These nodes are not considered in the nal recommendations and only used to indicate the connections of the top scholars. { Keyword: There are three types of keyword nodes. Self-De ned keywords, keywords extracted from recent publications, and relevant keywords that represent the connection between two other types (shown by a dashed border) and will not appear in the recommendations. The relationship between keywords represented by "has link" arc is established if the target node has been mentioned in the source node's entry page. Keyword nodes are connected to individual nodes via "has key" and to category nodes via "belongs to" arcs. { Category: We employed a full hierarchical schema of Wikipedia categories to represent the inter-connectivity between categories in our data set. These relations are presented as the "has child" arc in Figure 1. The category nodes are used to nd the semantic relationship between keywords.
https://en.wikipedia.org/wiki/Neo4j
4.1
The knowledge graph is used to power the exploratory search interface for nding advisors. The interface consists of four main sections.
Instant search box Users can use the search box to search for topic keywords or scholars of interest (Figure 2:B). When the user starts typing, a list of matching keywords and scholars appears. When an item is selected from the list, it is automatically added to the proper location on the left or the right side of the interface. At the same time, an updated list of recommendations is presented to the user. Favorite keywords This section (Figure 2:A) shows users' \favorite" keywords. Users add keywords to this list using the instant search box or by clicking on the plus button next to each recommended keyword. Users can interact with three buttons on the right side of each keywords to (1) see more information about that keyword (including its Wikipedia summary, similar keywords, and other scholars with this research interest), (2) remove the keyword from the favorite list, and (3) enable/disable the e ect of this keyword on the list of recommendations. Favorite scholars Similar to favorite keywords, users can assemble a list of favorite scholars (Figure 2:C). A new favorite scholar can be added to the list from the instant search results or a list or recommended scholars by clicking on the plus button next to a recommended scholar. The three buttons on the right side of each favorite scholar can be used to obtain more details about the scholar (a liation, full list of research interests, and similar scholars), remove the scholar from the list, and enable/disable the e ect of this scholar in the nal recommendation. Together with the favorite keywords list explained above, the list of favorite scholars form the users' pro les of interests, which the users gradually assemble while exploring possible areas of interests and scholars. In turn, the pro les of interests are used to generate further recommendations as explained below.
Recommendations This section (Figure 2:D) consists of two subsections. Recommended keywords (Figure 2:D1) shows the list of the three most relevant additional keywords, which are suggested given already selected (and enabled) favorite keywords and scholars. Users can see more information about the keyword (similar to the favorite keywords section) and also add these recommended keywords to their favorite lists using two circular buttons on the right side of each keyword. Recommended scholars (Figure 2:D2) shows a list of recommended scholars, which are most relevant to the active (enabled) favorite topic keywords and most similar to the active favorite scholars. For each recommended scholar, the list shows basic personal and academic information. Users can also see the similarity between the recommended scholars and their pro les of interests represented by favorite keywords and scholars. 4.2
We generate the recommendations using Cypher Query Language in Neo4j. In the following we explain how we generate recommendations for keywords and scholars.
Keyword Recommendations In order to recommend similar keywords, we use the user's favorite keywords and scholars. Each keyword is connected to other keywords in two ways: (1) via the similar research interest between scholars and (2) via similar relevant keywords and categories. We consider both of these relations to nd similar keywords. In the nal list, we sorted the keywords based on the number of occurrences then we chose the top three keywords to be presented to the user.
Scholar Recommendations Similar to keyword recommendations, we use both favorite keywords and scholars. There are three criteria for scholar recommendations: the scholar's weighted research interests, co-authorship relationship between scholars, and connection between the scholar's interests through relevant keywords and categories. After generating the list of candidate scholars, we sort it based on the similarity score (calculated based on weighted similarity score for each of the three criteria) and present the top ten results to the user. 5
We presented a method to build a knowledge graph by integrating data from Google Scholar and Wikipedia to help students with limited knowledge about a subject nd a research advisor or thesis committee member. Although Google scholar covers a variety of publications and patents, additional sources of information (e.g., the scholar's active research projects, funding information, etc.) could make the knowledge graph more connected and provide the users with additional critical information when it comes to nding an advisor. We plan to re ne our keyword extraction techniques. More sophisticated methods of extraction using natural language processing and machine learning could potentially improve the semantic relations between concepts and provide users with a more realistic set of research interests for scholars. We have also designed a series of controlled user studies and eld studies to evaluate the usability and value of the exploratory search interface. We hope that these user studies will provide valuable insights for improving the knowledge graph and the interface.