We define the schema of the researcher profile by extending the FOAF ontology [10], as shown in Figure 2. In the schema, 24 properties and two relations are defined [2, 3].

It is certainly not a trivial task to extract the research network from the Web. The researchers from different universities, institutes or companies have disparate page and profile templates and data feeds. So an ideal extraction method should consider processing all kinds of templates and formats. The approach we proposed consists of three steps:

1) Relevant page identification. Given a researcher name, we first get a list of Web pages by a search engine (Google API is used) and then identify the homepage or introducing page using a classifier. We define a set of features, such as whether the title of the page contains the person name and whether the URL address (partly) contains the person name, and employ SVM [11] for the classification.

2) Preprocessing. We separate the text into tokens and assign possible tags to each token. The tokens form the basic units and the pages form the sequences of units in the following tagging step.

3) Tagging. Given a sequence of units, we determine the most likely corresponding sequence of tags by using a trained tagging model. The type of tag corresponds with the property defined in Figure 2. We define five types of tokens (i.e., standard word, special word, image token, term, and punctuation mark) and use heuristics to identify tokens on the Web. After that, we assign several possible tags to each token based on the token type, and then a trained CRF model [12] is used to find the best tag assignment having the highest likelihood.

Recently, we revisit the problem of Web user profiling in the big data and propose a simple but very effective approach, referred to as MagicFG [4], for profiling Web users by leveraging the power of big data. To avoid error propagation, the approach integrates page identification and profile extraction in an unified framework. To improve the profiling performance, we present the concept of contextual credibility. The proposed framework also supports the incorporation of human knowledge. It defines human knowledge as Markov logics statements and formalizes them into a factor graph model. The MagicFG method has been deployed in AMiner system for profiling millions of researchers.

Figure 3 gives an example of researcher profile.

We have collected more than 200 million publications from existing online data libraries, including DBLP, ACM DL, CiteSeerX and others. In each data source, authors are identified by their names. For integrating the researcher profiles and the publication data, we use researcher name and the author name as the identifier. This process inevitably has the ambiguous problem.

A few years ago, we had proposed a probabilistic framework [5] based on Hidden Markov Random Fields (HMRF) [13] which is able to capture dependencies between observations (here each paper is viewed as an observation). The disambiguation problem is cast as assigning a tag to each paper with each tag representing an actual researcher.

More recently we proposed an additional comprehensive framework [6] to address the name disambiguation problem. The overview of the framework is as shown in Figure 4. A novel representation learning method is proposed, which incorporates both global and local information. In addition, an end-to-end cluster size estimation method is presented in the framework. To improve the accuracy, we involve human annotators into the disambiguation process. The method has now been deployed in AMiner to deal with the name disambiguation problem at the billion scale, which demonstrates its effectiveness and efficiency.

In academic search, representation of the content of text documents, authors interests and conferences themes is a critical issue of any approach. Traditionally, documents are represented based on the “bag of words” (BOW) assumption. However, this representation cannot utilize the “semantic” dependencies between words. In addition, in the course of an academic search there are different types of information sources, thus how to capture the dependencies between them becomes a challenging issue. Unfortunately, existing topic models such as probabilistic Latent Semantic Indexing (pLSI) [14], Latent Dirichlet Allocation (LDA) [15] and Author-Topic model [16, 17] cannot be directly applied to the context of academic search. This is because they simply cannot capture all intrinsic dependencies between papers and conferences.

A unified topic modeling approach [1] is proposed for simultaneously modeling characteristics of documents, authors, conferences and dependencies among them. (For simplicity, we use conference to denote conference, journal and book in the model.) The proposed model is called Author-Conference-Topic (ACT) model. More specifically, different strategies can be employed to model the topic distributions (as shown in Figure 5) and consequently the implemented models can have different knowledge representation capacities. In Figure 5(a) each author is associated with a mixture of weights over topics. For example, each word token correlated to a paper, and likewise a conference stamp associated to each word token, is generated from a sampled topic. In Figure 5(b) each author-conference pair is associated with a mixture of weights over the topics, and word tokens are then generated from the sampled topics. In Figure 5(c), each author is associated with topics, each word token is generated from a sampled topic, and then the conference is generated from the sampled topics of all word tokens in a paper.

When searching for academic resources and formulating a query, a user endeavors to find authors with specific expertise, and papers and conferences related to the research areas of interest.

In the AMiner system we present a topic level expertise search framework [18]. Different from the traditional Web search engines that perform retrieval and ranking at document level, we study the expertise search problem at topic level over disparate heterogenous networks. A unified topic model, namely Citation-Tracing-Topic (CTT), is proposed to simultaneously model topical aspects of different objects in the academic network. Based on the learned topic models, we investigate the expertise search problem from three dimensions: ranking, citation tracing analysis and topic graph search. Specifically, we propose a topic level random walk method for ranking different objects. In citation tracing analysis we seek to uncover how a study influences its follow-up study. Finally, we have developed a topical graph search function, based on the topic modeling and citation tracing analysis.

Figure 6 gives an example result of experts found for the query “Data Mining”.

Based on the AMiner system, this set of researcher-centric academic social network mining functions includes social influence analysis, relationship mining, collaboration recommendation, similarity analysis and community evolution.

Social Influence Analysis. In large social networks, persons are influenced by others for various reasons. We propose a Topic Affinity Propagation (TAP) model [19] to differentiate and quantify the social influence. TAP can take results of any topic modeling and the existing network structure to perform topic-level influence propagation. Recently we design an end-to-end framework that we call DeepInf for feature representation learning and to predict social influence [20]. Each user is represented with a local subnetwork which he or she is embedded in. A graph neural network is used to learn the representation of the sub-network which in turn effectively integrates the user-specific features and network structures. The framework of DeepInf is shown in Figure 7.

Social Relationship Mining. Inferring the type of social relationships between two users is a very important task in social relationship mining. We propose a two-stage framework named Time-constrained Probabilistic Factor Graph model (TPFG) [21] for inferring advisor-advisee relationships in the co-author network. The main idea is to leverage a time-constrained probabilistic factor graph model to decompose the joint probability of the unknown advisors over all the authors. Furthermore, we develop a framework named TranFG for classifying the type of social relationships across disparate heterogeneous resources [22]. The framework incorporates social theories into a factor graph model, which effectively improves the accuracy of predicting the types of social relationships in a target network by borrowing knowledge from another source network.

Similarity Analysis. Estimating similarity between vertices is a fundamental issue in social network analysis. We propose a sampling-based method to estimate the top-k similar vertices [23]. The method is based on the novel idea of the random path sampling method known as Panther. Given a particular network as a starting point, Panther randomly generates a number of paths of a pre-defined length, and then the similarity between two vertices can be modeled as estimating the possibility that the two vertices appear on the same paths.

Collaboration Recommendation. Interdisciplinary collaborations have generated a huge impact on society. However, it is usually hard for researchers to establish such cross-domain collaborations. We analyze the cross-domain collaboration data from research publications and propose a Cross-domain Topic Learning (CTL) model [24] for collaboration recommendation. For handling sparse connections, CTL consolidates the existing cross-domain collaborations through topic layers as opposed to utilizing author layers. This alleviates the sparseness issue. For handling complementary expertise, CTL models topic distributions from source and target domains separately, as well as the correlation across domains. For handling topic skewness, CTL only models relevant topics to the cross-domain collaboration.

Community Evolution. Since social networks are rather dynamic, it is interesting to study how persons in the networks form different clusters and how the various clusters evolve over time. We study mining co-evolution of multi-typed objects in a special type of heterogeneous network, called a star network. We subsequently examine how the multi-typed objects influence each other in the network evolution [25]. A Hierarchical Dirichlet Process Mixture Model-based evolution is proposed which detects the co-evolution of multi-typed objects in the form of a multi-typed cluster evolution in dynamic star networks. An efficient inference algorithm is provided to learn the proposed model.