ArnetMiner: extraction and mining of academic social networks

Jie TangJing ZhangLimin YaoJuan-Zi LiLi ZhangZhong Su

article2008KDD2,586 citationsTest of Time Award for Applied Science

Presents ArnetMiner, an academic social network mining system that integrates conditional random fields for automated researcher profiling, hidden Markov random fields for publication name disambiguation, and joint author-conference-topic models to enable effective expertise and association search.

Listen

The article describes the development of the ArnetMiner system to extract researcher profiles from the web, integrate publication records from digital libraries, model academic social networks in a unified way, and deliver search services such as expertise and association search. These tasks address the growing need for semantics-rich tools that go beyond simple keyword matching in large scholarly databases, where incomplete profiles and name ambiguities hinder effective discovery.

The article set out to evaluate methods for automatic profile extraction, name disambiguation during data integration, and simultaneous topic modeling of papers, authors, and venues. It also aimed to demonstrate how the resulting models support practical search applications.

The authors built the system through a sequence of experiments on real web data and publication collections. They used conditional random fields for unified tagging of researcher homepages, a hidden Markov random field framework incorporating publication relationships for disambiguation, and three variants of an author-conference-topic model for joint topic estimation. Performance was measured on nearly 900 annotated homepages and a disambiguation dataset of 14 names, with search evaluations drawn from frequent user queries against pooled judgments from similar systems.

The unified extraction approach achieved an average F1 score of 83.37 percent, substantially above rule-based and classification baselines. The disambiguation framework reached 91.39 percent average F1, improving over hierarchical clustering by more than 10 points when relationships among papers were included. The author-conference-topic models outperformed language-model, LDA, and author-topic baselines on expertise search, with the first variant delivering the highest mean average precision of 71 percent across papers, authors, and conferences. Roughly 448,000 researcher profiles and over one million papers were integrated into the live system.

These results show that holistic modeling captures dependencies across data types that separate methods miss, leading to more accurate profiles, cleaner author identities, and more relevant search results. For decision makers, the gains translate into reduced manual curation effort, lower risk of missed expertise, and faster discovery of collaborators or venues.

The article recommends extending the topic models with citation and temporal links, automating the choice of the number of distinct persons in disambiguation, and refining extraction rules for greater coverage. Further work on these fronts is needed before the system can scale without manual intervention to all author names.

The main limitations are reliance on manually supplied person counts for disambiguation, evaluations confined to computer science data, and fixed hyperparameter settings in the topic models. Confidence is high for the reported extraction and disambiguation tasks on the tested collections, but lower for broader domains or fully automatic operation.

IBM Tsinghua University digital libraries libraries network architecture academic social networks researcher profile extraction ArnetMiner system

Tang et al (2008).pdf

Paper: The Author-Topic Model for Authors and Documents, Michal Rosen-Zvi et al. (2004). Its author-topic model is a direct methodological precursor to ArnetMiner’s joint modeling of authors and research topics.
Paper: Unsupervised Learning by Probabilistic Latent Semantic Analysis, Thomas Hofmann (2001). Its probabilistic latent-topic framework supplies foundational context for the topic-modeling component ArnetMiner adapts to authors and venues.
Paper: Early results for Named Entity Recognition with Conditional Random Fields, Feature Induction and Web-Enhanced Lexicons, A. McCallum et al. (2003). Its CRF-based sequence-labeling work provides a relevant methodological foundation for ArnetMiner’s CRF extraction of researcher profiles.

Paper: PathSim, Yizhou Sun et al. (2011). It carries academic-network analysis toward typed, path-based similarity search, extending the kind of researcher and publication relationships ArnetMiner brings together.

Cover for ArnetMiner: extraction and mining of academic social networks

Abstract

This paper addresses several key issues in the ArnetMiner system, which aims at extracting and mining academic social networks. Specifically, the system focuses on: 1) Extracting researcher profiles automatically from the Web; 2) Integrating the publication data into the network from existing digital libraries; 3) Modeling the entire academic network; and 4) Providing search services for the academic network. So far, 448,470 researcher profiles have been extracted using a unified tagging approach. We integrate publications from online Web databases and propose a probabilistic framework to deal with the name ambiguity problem. Furthermore, we propose a unified modeling approach to simultaneously model topical aspects of papers, authors, and publication venues. Search services such as expertise search and people association search have been provided based on the modeling results. In this paper, we describe the architecture and main features of the system. We also present the empirical evaluation of the proposed methods.

Topics

digital libraries libraries network architecture social networks modeling and analysis faculty

Keywords

academic social networks researcher profile extraction ArnetMiner system author-topic model expert finding association search Latent Dirichlet Allocation academic search

1. INTRODUCTION
2. RELATED WORK
2.1 Person Profile Extraction
2.2 Name Disambiguation
2.3 Topic Modeling
2.4 Academic Search
3. OVERVIEW OF ARNETMINER
4. RESEARCHERPROFILEEXTRACTION
4.1 Problem Definition
4.2 A Unified Approach to Profiling
4.2.1 Process
4.2.2 CRF model and Features
4.3 Profile Extraction Performance
5. NAMEDISAMBIGUATION
5.1 Problem Definition
5.2 A Unified Probabilistic Framework
5.2.1 Formalization using HMRF
5.2.2 EM framework
5.3 Name Disambiguation Performance
6. MODELING ACADEMIC NETWORK
6.1 Our Proposed Topic Models
6.2 ACT Model 1
6.3 ACT Model 2
6.4 ACT Model 3
7. ACADEMIC SEARCH SERVICES
7.1 Applying ACT Models to Expertise Search
7.1.1 Process
7.1.2 Expertise Search Performance
7.2 Applying ACT Models to Association Search
7.3 Other Applications
8. CONCLUSION
9. ACKNOWLEDGMENTS
10. REFERENCES

Knowls

Knowl 1 — Author-Conference-Topic Model 1 (ACT1)
model/method
The Author-Conference-Topic Model 1 (ACT1) is a generative probabilistic topic model that simultaneously models the topical distributions of papers, authors, and publication venues (conferences, journals, or books). In ACT1, each author is associated with a multinomial distribution over topics, and each word token in a paper alongside the publication venue stamp is generated conditioned on a sampled latent topic.

Given a collection of $D$ papers where paper $d \in \{1, \dots, D\}$ has word vector $\mathbf{w}_d = (w_{d1}, \dots, w_{d N_d})$ drawn from vocabulary $V$ , author subset $\mathbf{a}_d \subseteq \{1, \dots, A\}$ , and publication venue $c_d \in \{1, \dots, C\}$ :
1. For each topic $z \in \{1, \dots, T\}$ , draw word distribution $\boldsymbol{\phi}_z \sim \text{Dirichlet}(\boldsymbol{\beta})$ and venue distribution $\boldsymbol{\psi}_z \sim \text{Dirichlet}(\boldsymbol{\mu})$ .
2. For each author $x \in \{1, \dots, A\}$ , draw topic distribution $\boldsymbol{\theta}_x \sim \text{Dirichlet}(\boldsymbol{\alpha})$ .
3. For each word $w_{di}$ $w_{d i}$ ( $i \in \{1, \dots, N_d\}$ $i \in {1, \dots, N_{d}}$ ) in paper $d$ $d$ :
  - Draw author $x_{di} \sim \text{Uniform}(\mathbf{a}_d)$ ;
  - Draw topic $z_{di} \sim \text{Multinomial}(\boldsymbol{\theta}_{x_{di}})$ ;
  - Draw word $w_{di} \sim \text{Multinomial}(\boldsymbol{\phi}_{z_{di}})$ ;
  - Draw venue stamp $c_{di} \sim \text{Multinomial}(\boldsymbol{\psi}_{z_{di}})$ .
In Gibbs sampling inference, the joint posterior distribution of assigning the $i$ -th token in paper $d$ to author $x_{di}$ and topic $z_{di}$ , conditioned on all other variables, is given by:
$P(z_{di}, x_{di} \mid \mathbf{z}_{-di}, \mathbf{x}_{-di}, \mathbf{w}, \mathbf{c}, \boldsymbol{\alpha}, \boldsymbol{\beta}, \boldsymbol{\mu}) \propto \frac{m_{x_{di} z_{di}}^{-di} + \alpha_{z_{di}}}{\sum_{z=1}^T (m_{x_{di} z}^{-di} + \alpha_z)} \cdot \frac{n_{z_{di} w_{di}}^{-di} + \beta_{w_{di}}}{\sum_{v=1}^V (n_{z_{di} v}^{-di} + \beta_v)} \cdot \frac{n_{z_{di} c_d}^{-d} + \mu_{c_d}}{\sum_{c=1}^C (n_{z_{di} c}^{-d} + \mu_c)}$
where superscript $-di$ denotes counts excluding the current token, $m_{xz}$ is the number of times topic $z$ is assigned to author $x$ , $n_{zw}$ is the number of times word $w$ is generated by topic $z$ , and $n_{zc}$ is the number of times venue $c$ is generated by topic $z$ . Hyperparameters are configured as $\alpha = 50/T$ , $\beta = 0.01$ , and $\mu = 0.1$ .

Parameter estimates are obtained as:
$\phi_{zw} = \frac{n_{zw} + \beta_w}{\sum_{v=1}^V (n_{zv} + \beta_v)}, \quad \psi_{zc} = \frac{n_{zc} + \mu_c}{\sum_{c'=1}^C (n_{zc'} + \mu_{c'})}, \quad \theta_{xz} = \frac{m_{xz} + \alpha_z}{\sum_{z'=1}^T (m_{xz'} + \alpha_{z'})}$
Knowl 2 — Author-Conference-Topic Models 2 and 3 (ACT2 and ACT3)
model/method
ACT2 and ACT3 are alternative formulations for simultaneously modeling text, authorship, and venues:

ACT2 (Author-Venue Pair Conditioned Model): Assumes authors select a venue prior to writing and generate topics conditioned on the author-venue pair $(x_{di}, c_d)$ . For each word token $w_{di}$ in paper $d$ with venue $c_d$ :
1. Draw author-venue pair $(x_{di}, c_d)$ uniformly from $\{(a, c_d) \mid a \in \mathbf{a}_d\}$ .
2. Draw topic $z_{di} \sim \text{Multinomial}(\boldsymbol{\theta}_{(x_{di}, c_d)})$ , with $\boldsymbol{\theta}_{(x, c)} \sim \text{Dirichlet}(\boldsymbol{\alpha})$ .
3. Draw word $w_{di} \sim \text{Multinomial}(\boldsymbol{\phi}_{z_{di}})$ .
The Gibbs sampling update is:
$P(z_{di}, (xc)_{di} \mid \mathbf{z}_{-di}, \mathbf{x}_{-di}, \mathbf{c}_{-d}, \mathbf{w}, \boldsymbol{\alpha}, \boldsymbol{\beta}) \propto \frac{m_{(xc)_{di} z_{di}}^{-di} + \alpha_{z_{di}}}{\sum_{z=1}^T (m_{(xc)_{di} z}^{-di} + \alpha_z)} \cdot \frac{n_{z_{di} w_{di}}^{-di} + \beta_{w_{di}}}{\sum_{v=1}^V (n_{z_{di} v}^{-di} + \beta_v)}$
ACT3 (Continuous Venue Normal Linear Model): Treats the venue index $c_d$ as a continuous response generated after all paper words are sampled, drawn from a normal linear model $c_d \sim \mathcal{N}(\boldsymbol{\eta}^\top \boldsymbol{\tau}_d, \sigma^2)$ , where $\boldsymbol{\tau}_d \in \mathbb{R}^T$ is the normalized topic frequency vector of paper $d$ (with $\tau_d[k] = \frac{1}{N_d} \sum_{i=1}^{N_d} I[z_{di} = k]$ ).

Inference uses a Gibbs EM algorithm. In the E-step, the posterior topic assignment is:
$P(z_{di}, x_{di} \mid \mathbf{z}_{-di}, \mathbf{x}_{-di}, c_d, \mathbf{w}, \boldsymbol{\alpha}, \boldsymbol{\beta}, \boldsymbol{\eta}, \sigma^2) \propto \frac{1}{\sqrt{2\pi\sigma^2}} \exp\left(-\frac{(c_d - \boldsymbol{\eta}^\top \boldsymbol{\tau}_d)^2}{2\sigma^2}\right) \cdot \frac{m_{x_{di} z_{di}}^{-di} + \alpha_{z_{di}}}{\sum_z (m_{x_{di} z}^{-di} + \alpha_z)} \cdot \frac{n_{z_{di} w_{di}}^{-di} + \beta_{w_{di}}}{\sum_v (n_{z_{di} v}^{-di} + \beta_v)}$
In the M-step, regression parameters are updated as:
$\boldsymbol{\eta}_{\text{new}} = (\mathbb{E}[A^\top A])^{-1} \mathbb{E}[A]^\top \mathbf{c}, \quad \sigma^2_{\text{new}} = \frac{1}{D} \left(\mathbf{c}^\top \mathbf{c} - \mathbf{c}^\top \mathbb{E}[A] (\mathbb{E}[A^\top A])^{-1} \mathbb{E}[A]^\top \mathbf{c}\right)$
where $A$ is a $D \times T$ matrix whose $d$ -th row is $\mathbb{E}[\boldsymbol{\tau}_d] = \frac{1}{N_d} \sum_{i=1}^{N_d} \boldsymbol{\phi}_{di}$ , and $\mathbb{E}[A^\top A] = \sum_{d=1}^D \frac{1}{N_d^2} \left(\sum_{i=1}^{N_d} \sum_{j \neq i} \boldsymbol{\phi}_{di} \boldsymbol{\phi}_{dj}^\top + \sum_{i=1}^{N_d} \text{diag}\{\boldsymbol{\phi}_{di}\}\right)$ with $\boldsymbol{\phi}_{di}$ denoting the topic probability vector generating word $w_{di}$ .
Knowl 3 — Hidden Markov Random Field Formulation for Name Disambiguation
model/method
Author name disambiguation is formulated as assigning each publication $p_i \in P = \{p_1, \dots, p_n\}$ sharing an ambiguous author name $a$ to one of $k$ true underlying researcher entities $y_h \in \{1, \dots, k\}$ . Each paper $p_i$ is represented by a feature vector $\mathbf{x}_i$ of word counts from its title, venue, abstract, coauthors, and references.

Dependencies and relational constraints between publications are modeled using a Hidden Markov Random Field (HMRF) maximizing the a-posteriori conditional probability:
$P(\mathbf{y} \mid \mathbf{x}) = \frac{1}{Z} \exp\left( -\sum_{i=1}^n D(\mathbf{x}_i, y_h) - \sum_{i=1}^n \sum_{j \neq i} \left( D(\mathbf{x}_i, \mathbf{x}_j) \sum_{k=1}^5 w_k r_k(\mathbf{x}_i, \mathbf{x}_j) \right) \right)$
where $Z$ is a normalization partition factor, $D(\mathbf{x}_i, y_h)$ is the distance from paper vector $\mathbf{x}_i$ to the centroid of researcher entity $y_h$ , $D(\mathbf{x}_i, \mathbf{x}_j)$ is the pairwise distance between papers, and $r_k(\mathbf{x}_i, \mathbf{x}_j) \in \{0, 1\}$ indicates the presence of relationship type $k$ with assigned weight $w_k$ .

Five relationship types are defined between papers $p_i$ and $p_j$ :
1. CoPubvenue ( $r_1$ , $w_1 = 0.2$ ): Both papers are published in the same venue ( $p_i.\text{pubvenue} = p_j.\text{pubvenue}$ ).
2. CoAuthor ( $r_2$ , $w_2 = 0.7$ ): Papers share at least one secondary coauthor ( $a_i^{(r)} = a_j^{(s)}$ for $r, s > 0$ ).
3. Citation ( $r_3$ , $w_3 = 0.3$ ): $p_i$ cites $p_j$ or $p_j$ cites $p_i$ .
4. Constraints ( $r_4$ , $w_4 = 1.0$ ): Explicit constraints supplied via user feedback.
5. $\tau$ -CoAuthor ( $r_5$ , $w_5 = 0.7^\tau$ for $\tau > 1$ ): $\tau$ -step extended co-authorship chain (e.g., a secondary author of $p_i$ coauthors a separate paper with a secondary author of $p_j$ ).
Knowl 4 — Expectation-Maximization Algorithm for HMRF Name Disambiguation
algorithm
An Expectation-Maximization (EM) algorithm performs iterative cluster reassignment, centroid updating, and distance metric parameter learning for HMRF-based name disambiguation.

The parameterized distance between feature vectors $\mathbf{x}_i$ and $\mathbf{x}_j$ employs a learned diagonal weight matrix $A = \text{diag}(a_{11}, \dots, a_{mm}, \dots)$ :
$D(\mathbf{x}_i, \mathbf{x}_j) = 1 - \frac{\mathbf{x}_i^\top A \mathbf{x}_j}{\|\mathbf{x}_i\|_A \|\mathbf{x}_j\|_A}, \quad \text{where } \|\mathbf{x}\|_A = \sqrt{\mathbf{x}^\top A \mathbf{x}}$
Input: Paper vectors $\mathbf{x}_1, \dots, \mathbf{x}_n$ , number of true persons $k$ , relationship weights $w_k$ , learning rate $\eta$
Output: Researcher assignments $y_1, \dots, y_n$ , centroids $y_1, \dots, y_k$ , metric matrix $A$
Initialize centroids $y_1, \dots, y_k$ and metric matrix $A$
repeat
    // E-step: Paper Reassignment
    repeat
        for each paper $i = 1$ to $n$ do
            Assign paper $\mathbf{x}_i$ to entity $y_h$ ( $h \in \{1, \dots, k\}$ ) minimizing:
             $f(y_h, \mathbf{x}_i) = D(\mathbf{x}_i, y_h) + \sum_{j \neq i} \left( D(\mathbf{x}_i, \mathbf{x}_j) \sum_{k} w_k r_k(\mathbf{x}_i, \mathbf{x}_j) \right)$
        end for
    until no paper changes assignment
    // M-step: Representative and Metric Parameter Update
    for each entity $h = 1$ to $k$ do
         $y_h \leftarrow \frac{\sum_{i: y_i = h} \mathbf{x}_i}{\|\sum_{i: y_i = h} \mathbf{x}_i\|_A}$
    end for
    for each diagonal element $a_{mm}$ in $A$ do
         $a_{mm} \leftarrow a_{mm} + \eta \sum_i \frac{\partial f(y_h, \mathbf{x}_i)}{\partial a_{mm}}$
    end for
until convergence
The gradient with respect to diagonal parameter $a_{mm}$ is:
$\frac{\partial D(\mathbf{x}_i, \mathbf{x}_j)}{\partial a_{mm}} = \frac{x_{im} x_{jm} \|\mathbf{x}_i\|_A \|\mathbf{x}_j\|_A - \mathbf{x}_i^\top A \mathbf{x}_j \frac{x_{im}^2 \|\mathbf{x}_i\|_A^2 + x_{jm}^2 \|\mathbf{x}_j\|_A^2}{2 \|\mathbf{x}_i\|_A \|\mathbf{x}_j\|_A}}{\|\mathbf{x}_i\|_A^2 \|\mathbf{x}_j\|_A^2}$
Knowl 5 — Combined Language Model and Topic Model Scoring for Expertise Search
model/method
To combine query term matching with broad topical semantics, academic expertise search scores the relevance of an academic object (paper $d$ , author $a$ , or conference $c$ ) to a query $q$ by multiplying a Dirichlet-smoothed unigram language model (LM) probability by an Author-Conference-Topic (ACT) model probability.

For a paper $d$ , the probability of generating word $w$ is defined as:
$P(w \mid d) = P_{\text{LM}}(w \mid d) \times P_{\text{ACT}}(w \mid d)$
where $P_{\text{LM}}(w \mid d)$ is the Dirichlet-smoothed language model:
$P_{\text{LM}}(w \mid d) = \frac{N_d}{N_d + \lambda} \frac{\text{tf}(w, d)}{N_d} + \left(1 - \frac{N_d}{N_d + \lambda}\right) \frac{\text{tf}(w, D)}{N_D}$
with $N_d$ the length of document $d$ , $\text{tf}(w, d)$ the frequency of word $w$ in $d$ , $N_D$ the total word count in corpus $D$ , $\text{tf}(w, D)$ the corpus frequency of $w$ , and $\lambda$ the Dirichlet prior smoothing parameter set based on average paper length.

Under ACT1, the document word generation probability is computed as:
$P_{\text{ACT1}}(w \mid d, \boldsymbol{\theta}, \boldsymbol{\phi}) = \sum_{z=1}^T \sum_{x=1}^{A_d} P(w \mid z, \boldsymbol{\phi}_z) P(z \mid x, \boldsymbol{\theta}_x) P(x \mid d)$
with $P(x \mid d) = 1 / |\mathbf{a}_d|$ .

Given a multi-term query $q$ , the overall relevance score for a document is $P(q \mid d) = \prod_{w \in q} P(w \mid d)$ . Analogous likelihoods $P(q \mid a)$ and $P(q \mid c)$ are computed for authors and conferences by aggregating their published papers.
Knowl 6 — Social Network Association Search via Topic-Distribution KL Divergence
algorithm
Given an academic social network $G = (V, E)$ and an association query $(a_i, a_j)$ between source author $a_i$ and target author $a_j$ , association search discovers and ranks referral paths connecting the two individuals.

The directed distance from author $a_i$ to author $a_j$ is defined by the Kullback-Leibler (KL) divergence between their estimated ACT author-topic distributions $\boldsymbol{\theta}_{a_i}$ and $\boldsymbol{\theta}_{a_j}$ :
$\text{KL}(a_i, a_j) = \sum_{z=1}^T \theta_{a_i z} \log \frac{\theta_{a_i z}}{\theta_{a_j z}}$
The score of a referral chain is the accumulated sum of KL divergences along its constituent edges.
Input: Graph $G = (V, E)$ , topic distributions $\boldsymbol{\theta}_a$ for all $a \in V$ , query pair $(a_i, a_j)$ , relaxation factor $\gamma$ , maximum path length $M$
Output: Ranked list of near-shortest referral chains from $a_i$ to $a_j$
// Stage 1: Shortest Association Finding
Compute edge weights $w(u, v) = \text{KL}(u, v)$ for all $(u, v) \in E$
Run Dijkstra's algorithm with a min-heap to compute the minimum path score $L_{\min}$ from $a_i$ to $a_j$
// Stage 2: Near-Shortest Associations Finding
Initialize empty path collection $\mathcal{P}$
Execute depth-first search (DFS) starting from $a_i$ towards $a_j$ :
    Prune any branch whose accumulated path score exceeds $(1 + \gamma) L_{\min}$
    Prune any branch whose hop count exceeds $M$
    If path reaches $a_j$ with score $L \le (1 + \gamma) L_{\min}$ , add path to $\mathcal{P}$
Sort paths in $\mathcal{P}$ in ascending order of their accumulated scores
return $\mathcal{P}$
Knowl 7 — Unified Researcher Profile Extraction via Conditional Random Fields
model/method
Automated extraction of structured researcher profiles from Web pages is formulated as a unified sequence tagging task using Conditional Random Fields (CRFs) based on an extended FOAF ontology schema capturing 19 profile properties (including Name, Affiliation, Position, Email, Phone, Address, Photo, Degree institutions, Dates, Majors, and Research Interests).

The extraction pipeline consists of three phases:
1. Relevant Page Identification: Given a researcher name, candidate Web pages retrieved via the Google API are classified into homepages/introducing pages versus irrelevant pages using a Support Vector Machine (SVM) binary classifier (achieving 92.39% F1-measure).
2. Preprocessing and Tag Assignment: Web page contents are segmented into five unit types: standard unigram words, special words (URLs, emails, dates, numbers, special symbols identified via regular expressions), <image> tags, base noun phrase terms, and punctuation marks. Unit-specific candidate property tags are assigned to restrict the tag space.
3. CRF Sequence Labeling: A linear-chain CRF is trained via Maximum Likelihood Estimation to find the optimal tag sequence $\mathbf{Y}^* = \arg\max_{\mathbf{Y}} P(\mathbf{Y} \mid \mathbf{X})$ $Y^{*} = ar g max_{Y} P (Y ∣ X)$ using 108,409 Boolean feature functions spanning three categories:
  - Content features: Token text, morphology, image dimensions, aspect ratio, color count, file format, filename tokens, ALT text, and face detection outputs.
  - Pattern features: Predefined positive/negative keywords, regex patterns, researcher name matches, and preceding line break counts.
  - Term features: Base noun phrases and vocabulary dictionary matches.
Knowl 8 — Empirical Performance of Researcher Profile Extraction
empirical result
The unified CRF-based profile extraction framework was evaluated on 898 annotated researcher Web pages (obtained by filtering an initial random sample of 1,000 researcher names from ArnetMiner, labeled by 7 annotators with majority voting) and compared against rule induction via Amilcare ( $LP^2$ ) and an SVM classifier.
Could not parse LaTeX table
The unified CRF model achieved an average F1-score of 83.37%, outperforming Amilcare by +29.93% and SVM by +9.80%. Removing label transition features from the CRF caused performance to drop by 11.28% in F1-score (to 72.09%), confirming that modeling sequential dependencies between adjacent profile properties is critical. Feature ablation demonstrated that combining content, term, and pattern features yielded the highest performance (~83.4% F1), compared to content features alone (~71% F1), content+term (~75% F1), or content+pattern (~79% F1).
Knowl 9 — Empirical Evaluation of HMRF Name Disambiguation and Relationship Types
empirical result
The HMRF name disambiguation method was evaluated on a benchmark dataset of 14 ambiguous person names spanning 861 publications and 161 actual distinct individuals, labeled by 5 human annotators using author homepages, affiliations, and emails. Performance was compared against a hierarchical clustering baseline.
Could not parse LaTeX table
The HMRF approach achieved an average F1-score of 91.39%, outperforming the baseline by +10.75%. Incremental ablation of the relationship graph showed:
- Removing all relational edges dropped disambiguation F1 by 44.72% (to ~46.7%).
- Adding CoPubvenue, Citation, CoAuthor, and $\tau$ -CoAuthor sequentially improved F1 at each step, with the CoAuthor relationship providing the largest individual gain (+24.38% F1).
Knowl 10 — Comparative Retrieval Performance for Academic Expertise Search
empirical result
Expertise search across papers, authors, and conferences was evaluated on a subset of ArnetMiner data (14,134 persons, 10,716 papers, 1,434 conferences) using frequent queries from system search logs, pooled top-30 results from ArnetMiner, Libra, and Rexa, and 4-grade human relevance judgments ( $T = 80$ topics for all topic models).
Could not parse LaTeX table
ACT1 achieved the highest performance across all evaluation metrics, reaching an overall Mean Average Precision (MAP) of 71.0% (compared to 61.0% for LM, 63.7% for Author-Topic, 63.9% for ACT2, and 63.8% for ACT3). For author search, ACT1 achieved 89.6% MAP and 91.4% P@5. Benchmark search engines Libra and Rexa achieved average MAP scores of 48.3% and 45.0% on the same query set.
Knowl 11 — Limitations in Cluster Count Determination and Graph/Temporal Modeling
limitation
The ArnetMiner system architecture and underlying models have three main limitations identified by the authors:
1. Preset Cluster Count in Name Disambiguation: The HMRF name disambiguation framework assumes that the true number of individuals $k$ associated with an ambiguous name is known and supplied empirically in advance, which cannot scale automatically to large-scale uncurated digital libraries.
2. Absence of Link and Network Structure in Topic Models: The ACT probabilistic topic models do not incorporate explicit network structures, such as citation links or co-authorship graphs, directly into the generative topical process.
3. Static Temporal Assumption: The topic models treat publication records statically, ignoring temporal dynamics such as the evolution of author research interests or the drifting themes of conferences over time.

Coverage note — None was omitted; all key contributed models (CRF extraction, HMRF name disambiguation, ACT1/ACT2/ACT3 topic models, LM-ACT expertise search, KL-divergence association search), algorithms, empirical evaluations, and stated limitations are fully represented.

References

1.L. A. Adamic and E. Adar. How to search a social network. Social Networks, 27:187–203, 2005.
2.C. Andrieu, N. de Freitas, A. Doucet, and M. I. Jordan. An introduction to mcmc for machine learning. Machine Learning, 50:5–43, 2003.
3.R. Baeza-Yates and B. Ribeiro-Neto. Modern Information Retrieval. ACM Press, 1999.
4.K. Balog, L. Azzopardi, and M. de Rijke. Formal models for expert finding in enterprise corpora. In Proc. of SIGIR’06, pages 43–55, 2006.
5.S. Basu, M. Bilenko, and R. J. Mooney. A probabilistic framework for semi-supervised clustering. In Proc. of KDD’04, pages 59–68, 2004.
6.R. Bekkerman and A. McCallum. Disambiguating web appearances of people in a social network. In Proc. of WWW’05, pages 463–470, 2005.
7.D. M. Blei and J. D. McAuliffe. Supervised topic models. In Proc. of NIPS’07, 2007.
8.D. M. Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3:993–1022, 2003.
9.D. Brickley and L. Miller. Foaf vocabulary specification. In Namespace Document, http://xmlns.com/foaf/0.1/, September 2004.
10.C. Buckley and E. M. Voorhees. Retrieval evaluation with incomplete information. In Proc. of SIGIR’04, pages 25–32, 2004.
11.F. Ciravegna. An adaptive algorithm for information extraction from web-related texts. In Proc. of IJCAI’01 Workshop, August 2001.
12.C. Cortes and V. Vapnikn. Support-vector networks. Machine Learning, 20:273–297, 1995.
13.N. Craswell, A. P. de Vries, and I. Soboroff. Overview of the trec-2005 enterprise track. In TREC’05, pages 199–205, 2005.
14.H. Han, L. Giles, H. Zha, C. Li, and K. Tsioutsiouliklis. Two supervised learning approaches for name disambiguation in author citations. In Proc. of JCDL’04, pages 296–305, 2004.
15.H. Han, H. Zha, and C. L. Giles. Name disambiguation in author citations using a k-way spectral clustering method. In Proc. of JCDL’05, pages 334–343, 2005.
16.T. Hofmann. Collaborative filerting via gaussian probabilistic latent semantic analysis. In Proc.of SIGIR’03, pages 259–266, 1999.
17.T. Hofmann. Probabilistic latent semantic indexing. In Proc.of SIGIR’99, pages 50–57, 1999.
18.H. Kautz, B. Selman, and M. Shah. Referral web: Combining social networks and collaborative filtering. Communications of the ACM, 40(3):63–65, 1997.
19.T. Kristjansson, A. Culotta, P. Viola, and A. McCallum. Interactive information extraction with constrained conditional random fields. In Proc. of AAAI’04, 2004.
20.J. Lafferty, A. McCallum, and F. Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proc. of ICML’01, 2001.
21.A. McCallum. Multi-label text classification with a mixture model trained by em. In Proc. of AAAI’99 Workshop, 1999.
22.D. Mimno and A. McCallum. Expertise modeling for matching papers with reviewers. In Proc. of KDD’07, pages 500–509, 2007.
23.T. Minka. Estimating a dirichlet distribution. In Technique Report, http://research.microsoft.com/ minka/papers/dirichlet/, 2003.
24.Z. Nie, Y. Ma, S. Shi, J.-R. Wen, and W.-Y. Ma. Web object retrieval. In Proc. of WWW’07, pages 81–90, 2007.
25.M. Rosen-Zvi, T. Griffiths, M. Steyvers, and P. Smyth. The author-topic model for authors and documents. In Proc. of UAI’04, 2004.
26.M. Steyvers, P. Smyth, and T. Griffiths. Probabilistic author-topic models for information discovery. In Proc. of SIGKDD’04, 2004.
27.Y. F. Tan, M.-Y. Kan, and D. Lee. Search engine driven author disambiguation. In Proc. of JCDL’06, pages 314–315, 2006.
28.J. Tang, D. Zhang, and L. Yao. Social network extraction of academic researchers. In Proc. of ICDM’07, pages 292–301, 2007.
29.X. Wei and W. B. Croft. Lda-based document models for ad-hoc retrieval. In Proc. of SIGIR’06, pages 178–185, 2006.
30.E. Xun, C. Huang, and M. Zhou. A unified statistical model for the identification of english basenp. In Proc. of ACL’00, 2000.
31.X. Yin, J. Han, and P. Yu. Object distinction: Distinguishing objects with identical names. In Proc. of ICDE’2007, pages 1242–1246, 2007.
32.K. Yu, G. Guan, and M. Zhou. Resume information extraction with cascaded hybrid model. In Proc. of ACL’05, pages 499–506, 2005.

Citation

MLA

Tang, J., et al. “ArnetMiner”. Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2008, pp. 990–98, https://doi.org/10.1145/1401890.1402008.

APA

Tang, J., Zhang, J., Yao, L., Li, J., Zhang, L., & Su, Z. (2008). ArnetMiner. Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 990–998. https://doi.org/10.1145/1401890.1402008

Chicago

Tang, J., J. Zhang, L. Yao, J. Li, L. Zhang, and Z. Su. 2008. “ArnetMiner”. Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 990–98. https://doi.org/10.1145/1401890.1402008.

Harvard

Tang, J. et al. (2008) “ArnetMiner”, Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pp. 990–998. Available at: https://doi.org/10.1145/1401890.1402008.

Vancouver

1. Tang J, Zhang J, Yao L, Li J, Zhang L, Su Z (2008) ArnetMiner. In: Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pp 990–998

BibTeX

@inproceedings{Tang_2008, series={KDD08}, title={ArnetMiner: extraction and mining of academic social networks}, url={http://dx.doi.org/10.1145/1401890.1402008}, DOI={10.1145/1401890.1402008}, booktitle={Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining}, publisher={ACM}, author={Tang, Jie and Zhang, Jing and Yao, Limin and Li, Juanzi and Zhang, Li and Su, Zhong}, year={2008}, month=Aug, pages={990–998}, collection={KDD08} }

Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

ArnetMiner: extraction and mining of academic social networks

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Author-Conference-Topic Model 1 (ACT1)

Knowl 2 — Author-Conference-Topic Models 2 and 3 (ACT2 and ACT3)

Knowl 3 — Hidden Markov Random Field Formulation for Name Disambiguation

Knowl 4 — Expectation-Maximization Algorithm for HMRF Name Disambiguation

Knowl 5 — Combined Language Model and Topic Model Scoring for Expertise Search

Knowl 6 — Social Network Association Search via Topic-Distribution KL Divergence

Knowl 7 — Unified Researcher Profile Extraction via Conditional Random Fields

Knowl 8 — Empirical Performance of Researcher Profile Extraction

Knowl 9 — Empirical Evaluation of HMRF Name Disambiguation and Relationship Types

Knowl 10 — Comparative Retrieval Performance for Academic Expertise Search

Knowl 11 — Limitations in Cluster Count Determination and Graph/Temporal Modeling

References

Citation

BibTeX

Access the Paper