metapath2vec: Scalable Representation Learning for Heterogeneous Networks

Yuxiao DongNitesh V. ChawlaAnanthram Swami

article2017KDD2,509 citations

Develops scalable heterogeneous network embedding frameworks that combine meta-path-guided random walks with type-specific negative sampling to capture both structural and semantic relations across diverse node types.

Listen

Heterogeneous networks, which contain multiple types of nodes and relationships, are common in real-world settings such as academic collaboration graphs, yet conventional network embedding methods like DeepWalk and node2vec treat all nodes uniformly and therefore fail to preserve both structural proximity and semantic distinctions. The article addresses this limitation by introducing two models, metapath2vec and metapath2vec++, that learn low-dimensional node representations while respecting network heterogeneity.

The work set out to maximize the likelihood of preserving heterogeneous neighborhoods so that the resulting embeddings could improve downstream tasks including node classification, clustering, and similarity search. The approach relies on meta-path-guided random walks to generate context sequences that encode semantic relations, followed by a heterogeneous skip-gram objective; metapath2vec++ further refines negative sampling to respect node types. Experiments were conducted on two large academic networks—the AMiner computer-science graph with millions of authors and papers and the smaller DBIS dataset—using standard parameter settings and multiple evaluation splits.

The models consistently outperformed baselines. With only 5 % labeled data, venue classification improved by 35–319 % relative to DeepWalk, LINE, and PTE; author clustering gains reached 13–16 % over the strongest baselines. Visualization of the embeddings showed that metapath2vec++ automatically grouped venues and authors by research area and aligned related pairs along consistent directions, a property absent from prior methods. Parameter studies indicated that performance remains high under cost-effective choices of walk length and neighborhood size.

These results imply that organizations managing heterogeneous data can obtain richer features for search, recommendation, and anomaly detection without hand-crafted meta-path features for every task. The approach scales to networks of millions of nodes when parallelized across dozens of cores, completing training in minutes.

Next steps supported by the article include automatic discovery of useful meta-paths, reduction of intermediate walk data, incorporation of temporal dynamics, and extension to other network genres. The main limitations are dependence on user-specified meta-path schemes and the generation of large intermediate path sets; results are demonstrated primarily on academic networks, so caution is warranted when generalizing to domains with markedly different heterogeneity patterns. Overall, the evidence for the core performance claims is strong and reproducible on the released data and code.

Microsoft United States Army Research Laboratory University of Notre Dame anomaly detection classification heterogeneous networks Metapath2vec Heterogeneous Network Meta‑Path

Dong et al (2017).pdf

Paper: DeepWalk: online learning of social representations, Bryan Perozzi et al. (2014). DeepWalk establishes the foundational framework of using random walks on graphs combined with the Word2Vec skip-gram objective, which metapath2vec directly adapts and extends to heterogeneous networks.
Paper: node2vec: Scalable Feature Learning for Networks, Aditya Grover et al. (2016). Node2vec formalizes parameterized, flexible random walk strategies for network representation learning, serving as the immediate homogeneous baseline and conceptual precursor to meta-path-guided walks.
Paper: LINE: Large-scale Information Network Embedding, Jian Tang et al. (2015). LINE introduces scalable first- and second-order proximity optimization with negative sampling for large-scale information networks, providing a core baseline and architectural point of comparison for metapath2vec.
Paper: Efficient Estimation of Word Representations in Vector Space, Tomáš Mikolov et al. (2013). This work introduces the Skip-gram architecture and efficient negative sampling formulations that metapath2vec customizes into heterogeneous skip-gram models.
Paper: Distributed Representations of Words and Phrases and their Compositionality, Tomas Mikolov et al. (2013). Mikolov et al. formulate the negative sampling optimization and subsampling strategies that underpin the heterogeneous negative sampling techniques designed in metapath2vec++.
Paper: Translating Embeddings for Modeling Multi-relational Data, Antoine Bordes et al. (2013). TransE provides seminal principles for multi-relational and heterogeneous graph embedding by modeling distinct relational semantics in latent vector spaces.

Paper: Modeling Relational Data with Graph Convolutional Networks, Michael Schlichtkrull et al. (2018). Relational Graph Convolutional Networks (R-GCNs) extend message-passing graph neural networks to multi-relational and heterogeneous graphs, moving beyond random-walk-based embeddings like metapath2vec to end-to-end deep relational learning.
Paper: KGAT: Knowledge Graph Attention Network for Recommendation, Xiang Wang et al. (2019). KGAT extends heterogeneous and relational network representation into recommendation systems by replacing predefined meta-paths with end-to-end attentive propagation across multi-hop entity pathways.
Paper: Graph Convolutional Neural Networks for Web-Scale Recommender Systems, Rex Ying et al. (2018). PinSage advances scalable random-walk-based graph convolution methods to web-scale bipartite graphs with rich node attributes, building on principles of scalable walk sampling seen in metapath2vec.
Paper: Graph Neural Networks: A Review of Methods and Applications, Jie Zhou et al. (2018). This comprehensive survey contextualizes random-walk embedding approaches like metapath2vec within the broader evolution of spatial, spectral, and relational Graph Neural Network frameworks.

Cover for metapath2vec: Scalable Representation Learning for Heterogeneous Networks

Abstract

We study the problem of representation learning in heterogeneous networks. Its unique challenges come from the existence of multiple types of nodes and links, which limit the feasibility of the conventional network embedding techniques. We develop two scalable representation learning models, namely metapath2vec and metapath2vec++. The metapath2vec model formalizes meta-path-based random walks to construct the heterogeneous neighborhood of a node and then leverages a heterogeneous skip-gram model to perform node embeddings. The metapath2vec++ model further enables the simultaneous modeling of structural and semantic correlations in heterogeneous networks. Extensive experiments show that metapath2vec and metapath2vec++ are able to not only outperform state-of-the-art embedding models in various heterogeneous network mining tasks, such as node classification, clustering, and similarity search, but also discern the structural and semantic correlations between diverse network objects.

1 INTRODUCTION
2 PROBLEM DEFINITION
3 THE METAPATH2VEC FRAMEWORK
3.1 Homogeneous Network Embedding
3.2 Heterogeneous Network Embedding: metapath2vec
3.3 metapath2vec++
4 EXPERIMENTS
4.1 Experimental Setup
4.2 Multi-Class Classification
4.3 Node Clustering
4.4 Case Study: Similarity Search
4.5 Case Study: Visualization
4.6 Scalability
5 RELATED WORK
6 CONCLUSION
REFERENCES

Knowls

Knowl 1 — Heterogeneous Network Representation Learning Problem Formulation
definition
A heterogeneous network is defined as a graph $G = (V, E, \mathcal{T})$ associated with an object type mapping function $\phi(v): V \to T_V$ and a relation type mapping function $\varphi(e): E \to T_E$ , where $T_V$ and $T_E$ denote the sets of object types and relation types, respectively, with $|T_V| + |T_E| > 2$ .

Given a heterogeneous network $G = (V, E, \mathcal{T})$ , the problem of heterogeneous network representation learning is to learn a low-dimensional latent embedding matrix $X \in \mathbb{R}^{|V| \times d}$ with $d \ll |V|$ , where the $v^{\text{th}}$ row vector $X_v \in \mathbb{R}^d$ is the latent representation of node $v \in V$ . The learned embeddings for all nodes regardless of their type are mapped into the same $d$ -dimensional latent feature space such that structural and semantic relationships between diverse node and relation types are preserved.
Knowl 2 — Meta-Path-Guided Heterogeneous Random Walks
model/method
To prevent random walks in heterogeneous networks from being biased toward highly visible or dominant node types, walk transitions are constrained by a pre-defined meta-path scheme $\mathcal{P}$ denoted by $V_1 \xrightarrow{R_1} V_2 \xrightarrow{R_2} \dots V_t \xrightarrow{R_t} V_{t+1} \dots \xrightarrow{R_{l-1}} V_l$ , which defines a composite relationship $R = R_1 \circ R_2 \circ \dots \circ R_{l-1}$ between node types $V_1$ and $V_l$ .

Given a walker currently at node $v_t^i$ of type $V_t$ at step $i$ , the transition probability to the next node $v^{i+1}$ is governed by:

$p(v^{i+1} \mid v_t^i, \mathcal{P}) = \begin{cases} \frac{1}{|N_{t+1}(v_t^i)|} & \text{if } (v^{i+1}, v_t^i) \in E \text{ and } \phi(v^{i+1}) = t+1 \\ 0 & \text{if } (v^{i+1}, v_t^i) \in E \text{ and } \phi(v^{i+1}) \neq t+1 \\ 0 & \text{if } (v^{i+1}, v_t^i) \notin E \end{cases}$

where $N_{t+1}(v_t^i)$ is the set of neighbors of node $v_t^i$ that are of type $V_{t+1}$ . Meta-path schemes are commonly chosen to be symmetric ( $V_1 = V_l$ ), enabling recursive execution such that $p(v^{i+1} \mid v_t^i) = p(v^{i+1} \mid v_1^i)$ when $t = l$ .
Knowl 3 — The metapath2vec Embedding Model
model/method
The metapath2vec model learns $d$ -dimensional node representations $X \in \mathbb{R}^{|V| \times d}$ for a heterogeneous network $G = (V, E, \mathcal{T})$ by maximizing the log-probability of observing heterogeneous context neighborhoods $N_t(v)$ of type $t \in T_V$ for every node $v \in V$ :

$\arg \max_X \sum_{v \in V} \sum_{t \in T_V} \sum_{c_t \in N_t(v)} \log p(c_t \mid v; X)$

where $c_t \in N_t(v)$ is a context node of type $t$ generated from meta-path-guided random walks, and the conditional probability $p(c_t \mid v; X)$ is defined via softmax:

$p(c_t \mid v; X) = \frac{e^{X_{c_t} \cdot X_v}}{\sum_{u \in V} e^{X_u \cdot X_v}}$

Using negative sampling with $M$ negative samples drawn from a global node distribution $P(u)$ , the objective per context pair $(v, c_t)$ is approximated as:

$\mathcal{O}_{\text{m2v}}(X) = \log \sigma(X_{c_t} \cdot X_v) + \sum_{m=1}^M \mathbb{E}_{u^m \sim P(u)} [\log \sigma(-X_{u^m} \cdot X_v)]$

where $\sigma(x) = \frac{1}{1 + e^{-x}}$ is the sigmoid function. Negative samples $u^m$ are drawn across the entire node set $V$ irrespective of node types.
Knowl 4 — The metapath2vec++ Model with Heterogeneous Negative Sampling
model/method
The metapath2vec++ model extends heterogeneous network embedding by conditioning both the softmax normalizer and negative sampling on the node type $t$ of the target context node $c_t$ . The conditional probability $p(c_t \mid v; X)$ is normalized strictly over nodes of the matching type $V_t$ :

$p(c_t \mid v; X) = \frac{e^{X_{c_t} \cdot X_v}}{\sum_{u_t \in V_t} e^{X_{u_t} \cdot X_v}}$

This creates distinct type-specific multinomial output distributions in the skip-gram architecture. With heterogeneous negative sampling, $M$ negative samples $u_t^m$ are drawn specifically from a type-dependent distribution $P_t(u_t)$ defined over $V_t$ , yielding the objective:

$\mathcal{O}(X) = \log \sigma(X_{c_t} \cdot X_v) + \sum_{m=1}^M \mathbb{E}_{u_t^m \sim P_t(u_t)} [\log \sigma(-X_{u_t^m} \cdot X_v)]$

where $\sigma(x) = \frac{1}{1 + e^{-x}}$ is the sigmoid function, $X_v$ is the embedding of target node $v$ , $X_{c_t}$ is the embedding of context node $c_t \in V_t$ , and $u_t^m \in V_t$ are negative node samples of type $t$ .
Knowl 5 — Gradient Formulation for metapath2vec++ Stochastic Optimization
equation
For a target node $v \in V$ , a positive context node $c_t \in V_t$ , and $M$ type-specific negative samples $\{u_t^m\}_{m=1}^M \subset V_t$ , let $u_t^0 = c_t$ . The objective is:

$\mathcal{O}(X) = \log \sigma(X_{c_t} \cdot X_v) + \sum_{m=1}^M \log \sigma(-X_{u_t^m} \cdot X_v)$

where $\sigma(z) = \frac{1}{1 + e^{-z}}$ . The partial derivatives of $\mathcal{O}(X)$ with respect to the negative/context node embeddings $X_{u_t^m}$ and the target node embedding $X_v$ are given by:

$\frac{\partial \mathcal{O}(X)}{\partial X_{u_t^m}} = \left(\sigma(X_{u_t^m} \cdot X_v) - \mathbb{I}_{c_t}[u_t^m]\right) X_v$

$\frac{\partial \mathcal{O}(X)}{\partial X_v} = \sum_{m=0}^M \left(\sigma(X_{u_t^m} \cdot X_v) - \mathbb{I}_{c_t}[u_t^m]\right) X_{u_t^m}$

where $\mathbb{I}_{c_t}[u_t^m]$ is an indicator function equal to $1$ if $u_t^m = c_t$ (i.e., for $m = 0$ ) and $0$ otherwise.
Knowl 6 — The metapath2vec++ Representation Learning Algorithm
algorithm
The metapath2vec++ algorithm generates node representations by executing meta-path-guided random walks and optimizing node embeddings via heterogeneous skip-gram with stochastic gradient descent.
Input: Heterogeneous information network $G = (V, E, \mathcal{T})$ , meta-path scheme $\mathcal{P}$ , number of walks per node $w$ , walk length $l$ , embedding dimension $d$ , neighborhood window size $k$ , learning rate $\eta$
Output: Latent node embedding matrix $X \in \mathbb{R}^{|V| \times d}$
initialize $X$ randomly
for step = 1 to $w$ do
    for each node $v \in V$ do
        MP = MetaPathRandomWalk( $G$ , $\mathcal{P}$ , $v$ , $l$ )
         $X$ = HeterogeneousSkipGram( $X$ , $k$ , MP)
    end for
end for
return $X$
Procedure MetaPathRandomWalk( $G$ , $\mathcal{P}$ , $v$ , $l$ )
    MP[1] = $v$
    for $i = 1$ to $l - 1$ do
        draw node $u$ with probability $p(u \mid \text{MP}[i], \mathcal{P})$ according to the meta-path transition rule
        MP[ $i + 1$ ] = $u$
    end for
    return MP
Procedure HeterogeneousSkipGram( $X$ , $k$ , MP)
    for $i = 1$ to $l$ do
         $v$ = MP[ $i$ ]
        for $j = \max(1, i - k)$ to $\min(i + k, l)$ with $j \neq i$ do
             $c_t$ = MP[ $j$ ]
            sample $M$ negative nodes $\{u_t^m\}_{m=1}^M$ from type-specific distribution $P_t(u_t)$
            compute gradients $\frac{\partial \mathcal{O}(X)}{\partial X_{u_t^m}}$ and $\frac{\partial \mathcal{O}(X)}{\partial X_v}$
             $X = X - \eta \cdot \frac{\partial \mathcal{O}(X)}{\partial X}$
        end for
    end for
    return $X$

Knowl 7 — Multi-Class Node Classification Performance on AMiner Dataset

data/table

The quality of learned representations ( $d=128$ , $w=1000$ , $l=100$ , $k=7$ ) was evaluated by training a logistic regression classifier on the AMiner CS network across 8 venue categories (133 labeled venues) and 8 author research categories (246,678 labeled authors), varying the training split from 5% to 90%.

Task / Metric	Method	5% Train	10% Train	30% Train	50% Train	90% Train
Venue Macro-F1	DeepWalk/node2vec	0.0723	0.1396	0.2795	0.3911	0.4457
	LINE (1st+2nd)	0.2245	0.4629	0.8473	0.9203	0.9466
	PTE	0.1702	0.3388	0.8304	0.9210	0.9489
	metapath2vec	0.3033	0.5247	0.8971	0.9532	0.9670
	metapath2vec++	0.3090	0.5444	0.8995	0.9580	0.9503
Venue Micro-F1	DeepWalk/node2vec	0.1701	0.2142	0.3266	0.4090	0.5286
	LINE (1st+2nd)	0.3000	0.5167	0.8457	0.9209	0.9571
	PTE	0.2512	0.4267	0.8372	0.9239	0.9571
	metapath2vec	0.4173	0.5975	0.9011	0.9522	0.9857
	metapath2vec++	0.4331	0.6192	0.9032	0.9582	0.9786
Author Macro-F1	DeepWalk/node2vec	0.7153	0.7222	0.7270	0.7274	0.7275
	LINE (1st+2nd)	0.8849	0.8886	0.8921	0.8929	0.8934
	PTE	0.8898	0.8940	0.8982	0.8990	0.9005
	metapath2vec	0.9216	0.9262	0.9303	0.9314	0.9320
	metapath2vec++	0.9107	0.9156	0.9199	0.9207	0.9212
Author Micro-F1	DeepWalk/node2vec	0.7312	0.7372	0.7414	0.7420	0.7425
	LINE (1st+2nd)	0.8936	0.8969	0.9002	0.9010	0.9017
	PTE	0.8986	0.9023	0.9061	0.9068	0.9082
	metapath2vec	0.9279	0.9319	0.9356	0.9365	0.9369
	metapath2vec++	0.9173	0.9217	0.9254	0.9261	0.9266

At a 5% training ratio, metapath2vec and metapath2vec++ achieve 35% to 319% relative improvements in Macro-F1 and 39% to 145% relative improvements in Micro-F1 over DeepWalk/node2vec, LINE, and PTE for venue classification. On author classification, they achieve a consistent ~2–3% improvement over LINE and PTE and ~20% improvement over DeepWalk/node2vec across all training ratios.

Knowl 8 — Node Clustering Performance on AMiner Dataset

data/table

Node clustering was evaluated by applying $k$ -means clustering to the learned 128-dimensional representations of 8-category venue and author nodes from the AMiner dataset, measured by Normalized Mutual Information (NMI) averaged over 10 runs.

For venue clustering, metapath2vec and metapath2vec++ outperform the strongest baselines (LINE and PTE) by 2% to 3%. For the more challenging author clustering task, metapath2vec and metapath2vec++ outperform LINE and PTE by 13% to 16% in NMI and outperform DeepWalk/node2vec by over 149%.

Knowl 9 — Semantic Relationship and Structural Organization in Latent Embedding Space
empirical result
When 128-dimensional node embeddings of top computer science venues and their corresponding representative researchers are projected via 2D PCA:
1. metapath2vec++ organizes different node types into structurally distinct, parallel columns while capturing semantic relationships as consistent vector offsets: the shift vectors connecting authors to their primary venues (such as J. Dean $\to$ OSDI, C. D. Manning $\to$ ACL, R. E. Tarjan $\to$ FOCS, and M. I. Jordan $\to$ NIPS) exhibit near-identical spatial directions and lengths.
2. metapath2vec groups each author–venue pair into close geometric proximity (e.g., R. E. Tarjan with FOCS, R. Agrawal with SIGMOD).
3. In contrast, homogeneous embeddings (DeepWalk/node2vec) and text-based semi-supervised embeddings (PTE) fail to discern between object types, producing unstructured overlap and failing to preserve author–venue translational semantics.
Knowl 10 — Scalability and Multi-Threaded Speedup of metapath2vec and metapath2vec++
empirical result
The C/C++ multi-threaded implementations of metapath2vec and metapath2vec++ achieve near-linear parallel speedup when tested on the full AMiner computer science dataset (comprising 1,693,531 authors, 3,194,405 papers, 3,883 venues, and 9,323,739 author-paper links) on a 48-core 2.3 GHz Intel Xeon server:
- The models achieve an 11–12 $\times$ speedup using 16 threads/cores and a 24–32 $\times$ speedup using 40 threads/cores relative to single-threaded execution.
- Using 40 threads, the complete representation learning process for metapath2vec++ takes approximately 9 minutes to embed the full multi-million-node AMiner CS network.

Coverage note — Qualitative top-10 similarity search query rankings for specific venues in the AMiner and DBIS datasets were omitted as standalone tables because they serve as illustrative case studies reinforcing the quantitative classification, clustering, and embedding space geometry results.

References

1.Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Je‚rey Dean, MaŠhieu Devin, Sanjay Ghemawat, Geo‚rey Irving, and others. 2016. TensorFlow: A system for large-scale machine learning. In OSDI ’16.
2.Amr Ahmed, Nino Shervashidze, Shravan Narayanamurthy, Vanja Josifovski, and Alexander J. Smola. 2013. Distributed Large-scale Natural Graph Factorization. In WWW ’13. ACM, 37–48.
3.Yoshua Bengio, Aaron Courville, and Pierre Vincent. 2013. Representation learning: A review and new perspectives. IEEE TPAMI 35, 8 (2013), 1798–1828.
4.Shiyu Chang, Wei Han, Jiliang Tang, Guo-Jun Qi, Charu C. Aggarwal, and Œomas S. Huang. 2015. Heterogeneous Network Embedding via Deep Architectures. In KDD ’15. ACM, 119–128.
5.Ting Chen and Yizhou Sun. 2017. Task-Guided and Path-Augmented Heterogeneous Network Embedding for Author Identi€cation. In WSDM ’17. ACM.
6.Yuxiao Dong, Jing Zhang, Jie Tang, Nitesh V. Chawla, and Bai Wang. 2015. CoupledLP: Link Prediction in Coupled Networks. In KDD ’15. ACM, 199–208.
7.Yoav Goldberg and Omer Levy. 2014. word2vec Explained: deriving Mikolov et al.’s negative-sampling word-embedding method. CoRR abs/1402.3722 (2014).
8.Aditya Grover and Jure Leskovec. 2016. Node2Vec: Scalable Feature Learning for Networks. In KDD ’16. ACM, 855–864.
9.Keith Henderson, Brian Gallagher, Tina Eliassi-Rad, Hanghang Tong, Sugato Basu, Leman Akoglu, Danai Koutra, Christos Faloutsos, and Lei Li. 2012. Rolx: structural role extraction & mining in large graphs. In KDD ’12. ACM, 1231–1239.
10.Peter D Ho‚, Adrian E Ra‰ery, and Mark S Handcock. 2002. Latent space approaches to social network analysis. Journal of the American Statistical association 97, 460 (2002), 1090–1098.
11.Xiao Huang, Jundong Li, and Xia Hu. 2017. Label Informed AŠributed Network Embedding. In WSDM ’17. na.
12.Zhipeng Huang, Yudian Zheng, Reynold Cheng, Yizhou Sun, Nikos Mamoulis, and Xiang Li. 2016. Meta structure: Computing relevance in large heterogeneous information networks. In KDD ’16. ACM, 1595–1604.
13.Ming Ji, Jiawei Han, and Marina Danilevsky. 2011. Ranking-based classi€cation of heterogeneous information networks. In KDD ’11. ACM, 1298–1306.
14.Yehuda Koren. 2008. Factorization meets the neighborhood: a multifaceted collaborative €ltering model. In KDD ’08. ACM, 426–434.
15.Yann LeCun, Yoshua Bengio, and Geo‚rey Hinton. 2015. Deep learning. Nature 521, 7553 (2015), 436–444.
16.Hao Ma, Dengyong Zhou, Chao Liu, Michael R Lyu, and Irwin King. 2011. Recommender systems with social regularization. In WSDM ’11. 287–296.
17.Tomas Mikolov, Kai Chen, Greg Corrado, and Je‚rey Dean. 2013. E cient Estimation of Word Representations in Vector Space. CoRR abs/1301.3781 (2013). hŠp://arxiv.org/abs/1301.3781
18.Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Je‚ Dean. 2013. Distributed representations of words and phrases and their compositionality. In NIPS ’13. 3111–3119.
19.Jennifer Neville and David Jensen. 2005. Leveraging relational autocorrelation with latent group models. In Proceedings of the 4th international workshop on Multi-relational mining. ACM, 49–55.
20.Mingdong Ou, Peng Cui, Jian Pei, Ziwei Zhang, and Wenwu Zhu. 2016. Asymmetric Transitivity Preserving Graph Embedding. In KDD ’16. ACM, 1105–1114.
21.Siddharth Pal, Yuxiao Dong, Bishal Œapa, Nitesh V Chawla, Ananthram Swami, and Ram Ramanathan. 2016. Deep learning for network analysis: Problems, approaches and challenges. In Military Communications Conference, MILCOM 2016-2016. IEEE, 588–593.
22.Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. DeepWalk: Online Learning of Social Representations. In KDD ’14. ACM, 701–710.
23.Xiang Ren, Wenqi He, Meng ‹, Clare R Voss, Heng Ji, and Jiawei Han. 2016. Label noise reduction in entity typing by heterogeneous partial-label embedding. In KDD ’16. ACM.
24.Xin Rong. 2014. word2vec Parameter Learning Explained. CoRR abs/1411.2738 (2014). hŠp://arxiv.org/abs/1411.2738
25.Yizhou Sun and Jiawei Han. 2012. Mining Heterogeneous Information Networks: Principles and Methodologies. Morgan & Claypool Publishers.
26.Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S. Yu, and Tianyi Wu. 2011. Pathsim: Meta path-based top-k similarity search in heterogeneous information networks. In VLDB ’11. 992–1003.
27.Yizhou Sun, Brandon Norick, Jiawei Han, Xifeng Yan, Philip S. Yu, and Xiao Yu. 2012. Integrating Meta-path Selection with User-guided Object Clustering in Heterogeneous Information Networks. In KDD ’12. ACM, 1348–1356.
28.Yizhou Sun, Yintao Yu, and Jiawei Han. 2009. Ranking-based Clustering of Heterogeneous Information Networks with Star Network Schema. In KDD ’09. ACM, 797–806.
29.Jian Tang, Meng ‹, and Qiaozhu Mei. 2015. PTE: Predictive Text Embedding Œrough Large-scale Heterogeneous Text Networks. In KDD ’15. ACM, 1165–1174.
30.Jian Tang, Meng ‹, Mingzhe Wang, Ming Zhang, Jun Yan, and Qiaozhu Mei. 2015. LINE: Large-scale Information Network Embedding.. In WWW ’15. ACM.
31.Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, and Zhong Su. 2008. ArnetMiner: Extraction and Mining of Academic Social Networks. In KDD ’08. 990–998.
32.Lei Tang and Huan Liu. 2009. Relational learning via latent social dimensions. In KDD ’09. 817–826.
33.Lei Tang and Huan Liu. 2011. Leveraging social media networks for classi€cation. DMKD 23, 3 (2011), 447–478.
34.Shuicheng Yan, Dong Xu, Benyu Zhang, Hong-Jiang Zhang, Qiang Yang, and Stephen Lin. 2007. Graph embedding and extensions: A general framework for dimensionality reduction. IEEE TPAMI 29, 1 (2007).
35.Jing Zhang, Jie Tang, Cong Ma, Hanghang Tong, Yu Jing, and Juanzi Li. 2015. Panther: Fast top-k similarity search on large networks. In KDD ’15. ACM, 1445–1454.

Citation

MLA

Dong, Y., et al. “Metapath2vec”. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2017, pp. 135–44, https://doi.org/10.1145/3097983.3098036.

APA

Dong, Y., Chawla, N. V., & Swami, A. (2017). metapath2vec. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 135–144. https://doi.org/10.1145/3097983.3098036

Chicago

Dong, Y., N. V. Chawla, and A. Swami. 2017. “Metapath2vec”. Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 135–44. https://doi.org/10.1145/3097983.3098036.

Harvard

Dong, Y., Chawla, N.V. and Swami, A. (2017) “metapath2vec”, Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp. 135–144. Available at: https://doi.org/10.1145/3097983.3098036.

Vancouver

1. Dong Y, Chawla NV, Swami A (2017) metapath2vec. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, pp 135–144

BibTeX

@inproceedings{Dong_2017, series={KDD ’17}, title={metapath2vec: Scalable Representation Learning for Heterogeneous Networks}, url={http://dx.doi.org/10.1145/3097983.3098036}, DOI={10.1145/3097983.3098036}, booktitle={Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining}, publisher={ACM}, author={Dong, Yuxiao and Chawla, Nitesh V. and Swami, Ananthram}, year={2017}, month=Aug, pages={135–144}, collection={KDD ’17} }

Metadata:Crossref

Access the Paper

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

Open PDF

Method	Venue NMI	Author NMI
DeepWalk / node2vec	0.1952	0.2941
LINE (1st+2nd)	0.8967	0.6423
PTE	0.9060	0.6483
metapath2vec	0.9274	0.7470
metapath2vec++	0.9261	0.7354

metapath2vec: Scalable Representation Learning for Heterogeneous Networks

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — Heterogeneous Network Representation Learning Problem Formulation

Knowl 2 — Meta-Path-Guided Heterogeneous Random Walks

Knowl 3 — The metapath2vec Embedding Model

Knowl 4 — The metapath2vec++ Model with Heterogeneous Negative Sampling

Knowl 5 — Gradient Formulation for metapath2vec++ Stochastic Optimization

Knowl 6 — The metapath2vec++ Representation Learning Algorithm