Abstract
Heterogeneous graph representation learning is designed to learn meaningful representation vectors from heterogeneous networks in few dimensions to extract the structure and features of the attributes of these networks. The embedding vector is the basis of and crucial to complex network analysis, and can be used in such downstream tasks as classification, clustering, link prediction, and recommendation. Key issues in heterogeneous graph neural networks pertain to ways to define heterogeneous neighbors and ways to aggregate them. Although considerable research has been devoted to homogeneous and heterogeneous network representation, the effective combination of information on the network structure and the attributes of nodes, especially effective use of meta-paths containing specific semantic information, remains rare. Here a meta-path-based heterogeneous graph neural network model is proposed. The meta-path is applied to sample the heterogeneous neighbors of each node in the network, and aggregate features of the same types of nodes to form type-related embedding. A multi-head attention mechanism is then applied to aggregate information on neighbors of different types of nodes and the model is trained by reducing context loss. Experiments on classification, clustering, link prediction, and recommendation tasks verified the validity of this model, which significantly improved the results of baseline methods.
Similar content being viewed by others
References
Barabasi A, Albert R (1999) Emergence of scaling in random networks. Science 286(5439):509
Lu L, Zhou T (2011) Link prediction in complex networks: a survey. Physica A 390(6):1150
Aiello LM, Barrat A, Schifanella R, Cattuto C, Markines B, Menczer F (2012) Friendship prediction and homophily in social media. ACM Trans Web 6(2):9
Zitnik M, Leskovec J (2017) Predicting multicellular function through multi-layer tissue networks. Bioinformatics 33:i190
Ma Y, Cheng G, Liu Z, Liang X (2017) Clustering-based link prediction in scientific coauthorship networks. Int J Mod Phys C 28:1750082
Ma Y, Liang X, Huang J, Cheng G (2017) Intercity transportation construction based on link prediction. In: 2017 IEEE 29th international conference on tools with artificial intelligence (ICTAI), pp 1135–1138
Piras A, Germond A (1996) Heterogeneous artificial neural network for short term electrical load forecasting. Power Syst IEEE Trans 11(1):397
Cui P, Wang X, Pei J, Zhu W (2019) A survey on network embedding. IEEE Trans Knowl Data Eng 31(5):833
Zhang D, Yin J, Zhu X, Zhang C (2020) Network representation learning: a survey. IEEE Trans Big Data 6(1):3
Hamilton WL, Ying Z, Leskovec J (2017) Inductive representation learning on large graphs, neural information processing systems, pp 1024–1034
Ji M, Han J, Danilevsky M (2011) Ranking-based classification of heterogeneous information networks, knowledge discovery and data mining, pp 1298–1306
Zhou J, Cui G, Zhang Z, Yang C, Liu Z, Wang L, Li C, Sun M (2018) Graph neural networks: a review of methods and applications. http://arxiv.org/abs/Learning
Yin R, Li K, Zhang G, Lu J (2019) A deeper graph neural network for recommender systems. Knowl Based Syst 185:105020
Li Y, Wang H, Li J, Hong G (2013) Efficient community detection with additive constrains on large networks. Knowl Based Syst 52:268
Amiri B, Hossain L, Crawford JW, Wigand RT (2013) Community detection in complex networks: multi-objective enhanced firefly algorithm. Knowl-Based Syst 46(1):1
Kipf T, Welling M (2016) Semi-supervised classification with graph convolutional networks, international conference on learning representations
Velickovic P, Cucurull G, Casanova A, Romero A, Lio P, Bengio Y (2018) Graph attention networks, international conference on learning representations
Liao L, He X, Zhang H, Chua T (2018) Attributed social network embedding. IEEE Trans Knowl Data Eng 30(12):2257
Sun Y, Norick B, Han J, Yan X, Yu PS, Yu X (2013) PathSelClus: integrating meta-path selection with user-guided object clustering in heterogeneous information networks. ACM Trans Knowl Discov Data 7(3):11
Zhang C, Swami A, Chawla N.V (2019) SHNE: representation learning for semantic-associated heterogeneous networks. In: Proceedings of the twelfth ACM international conference on web search and data mining, pp 690–698
Zhang C, Song D, Huang C, Swami A, Chawla NV (2019) Heterogeneous graph neural network. In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp 793–803
Wang X, Ji H, Shi C, Wang B, Ye Y, Cui P, Yu PS (2019) Heterogeneous graph attention network. In: The world wide web conference, pp 2022–2032
Christou V, Tsipouras MG, Giannakeas N, Tzallas AT, Brown G (2019) Hybrid extreme learning machine approach for heterogeneous neural networks. Neurocomputing 361:137
Sun Y, Han J, Yan X, Yu PS, Wu T (2011) Pathsim: meta path-based top-k similarity search in heterogeneous information networks. Proc VLDB Endowment 4(11):992
Dong Y, Chawla NV, Swami A (2017) metapath2vec: scalable representation learning for heterogeneous networks, knowledge discovery and data mining, pp 135–144
Mikolov T, Sutskever I, Chen K, Corrado G, Dean J (2013) Distributed representations of words and phrases and their compositionality. Adv Neural Inf Process Syst 26:3111
Le QV, Mikolov T (2014) Distributed representations of sentences and documents, international conference on machine learning, pp 1188–1196
Sun Y, Yu Y, Han J (2009) Ranking-based clustering of heterogeneous information networks with star network schema. In: Proceedings of the 15th ACM SIGKDD international conference on knowledge discovery and data mining, pp 797–806
Sun Y, Han J (2012) Mining heterogeneous information networks: principles and methodologies. Synth Lect Data Mining Knowl Discov 3(2):1
Perozzi B, Al-Rfou R, Skiena S (2014) Deepwalk: Online learning of social representations. In: Proceedings of the 20th ACM SIGKDD international conference on knowledge discovery and data mining, pp 701–710
Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735
Maas AL, Hannun AY, Ng AY (2013) Rectifier nonlinearities improve neural network acoustic models. Proc ICML 30:3
Jie T, Jing Z, Yao L, Li J, Zhong S (2008) ArnetMiner: extraction and mining of academic social networks. In: Proceedings of the ACM SIGKDD international conference on knowledge discovery and data mining
Kingma DP, Ba J (2015) Adam: a method for stochastic optimization, international conference on learning representations
Acknowledgements
This work is supported by National Natural Science Foundation of China under Grant No. 62073333.
Author information
Authors and Affiliations
Corresponding authors
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix: detailed experimental data
Appendix: detailed experimental data
-
MP_1H meta-path-based sampling, 1 head attention, 128 dimension, use structural and textual embedding;
-
MP_2H meta-path-based sampling, 2 head attention, 128 dimension, use structural and textual embedding;
-
MP_3H meta-path-based sampling, 3 head attention, 128 dimension, use structural and textual embedding;
-
MP_4H meta-path-based sampling, 4 head attention, 128 dimension, use structural and textual embedding;
-
MP_S meta-path-based sampling, 4 head attention, 128 dimension, use structural embedding;
-
MP_T meta-path-based sampling, 4 head attention, 128 dimension, use textual embedding;
-
MP2 meta-path-based sampling (combination 2), 4 head attention, 128 dimension, use structural and textual embedding;
-
MP3 meta-path-based sampling (combination 3), 4 head attention, 128 dimension, use structural and textual embedding;
-
MP4 meta-path-based sampling (combination 4), 4 head attention, 128 dimension, use structural and textual embedding;
-
MP_W meta-path-based sampling, assign weight to neighbors, 4 head attention, 128 dimension, use structural and textual embedding;
-
MP2_W meta-path-based sampling (combination 2), assign weight to neighbors, 4 head attention, 128 dimension, use structural and textual embedding;
-
RWR random-walk-with-restart-based sampling, 4 head attention, 128 dimension, use structural and textual embedding;
-
Mix half neighbors sampled by MP and half neighbors sampled by RWR, 4 head attention, 128 dimension, use structural and textual embedding;
-
MP_E64 meta-path-based sampling, 4 head attention, 64 dimension, use structural and textual embedding;
-
MP_E32 meta-path-based sampling, 4 head attention, 32 dimension, use structural and textual embedding;
-
MP_E16 meta-path-based sampling, 4 head attention, 16 dimension, use structural and textual embedding;
-
MP2V three meta-paths are set: APA, APVPA and APPA, 128 dimension, walk length 30, number of walk per node 10;
-
HAN textual embedding, 128 dimension, three meta-paths: APA, APPA and APVPA;
-
ASNE concatenate structural and textual embedding, 128 dimension;
-
SHNE 128 dimension, walk length 30, number of walk per node 10, use textual embedding;
-
GSAGE concatenate structural and textual embedding, 128 dimension;
-
GATE concatenate structural and textual embedding, 128 dimension;
-
HetGNN combine structural and textual embedding, 128 dimension.
Task | LP/a–a/2013 | LP/a–a/2012 | LP/a–p/2013 | LP/a–p/2012 | ||||
---|---|---|---|---|---|---|---|---|
Model | AUC | F1 | AUC | F1 | AUC | F1 | AUC | F1 |
MP_1H | 76.09 ± 0.11 | 71.35 ± 0.19 | 76.05 ± 0.09 | 71.26 ± 0.15 | 78.41 ± 0.13 | 76.61 ± 0.22 | 78.13 ± 0.11 | 76.27 ± 0.17 |
MP_2H | 77.14 ± 0.08 | 73.10 ± 0.14 | 76.87 ± 0.08 | 72.72 ± 0.14 | 78.82 ± 0.07 | 77.11 ± 0.12 | 78.82 ± 0.11 | 77.09 ± 0.18 |
MP_3H | 77.16 ± 0.10 | 73.16 ± 0.15 | 77.25 ± 0.09 | 72.99 ± 0.15 | 79.16 ± 0.09 | 77.44 ± 0.13 | 78.95 ± 0.11 | 77.25 ± 0.20 |
MP_4H | 77.35 ± 0.10 | 72.61 ± 0.16 | 77.25 ± 0.12 | 72.51 ± 0.19 | 79.22 ± 0.09 | 77.37 ± 0.16 | 78.85 ± 0.09 | 76.99 ± 0.14 |
MP_S | 73.83 ± 0.11 | 65.16 ± 0.22 | 73.65 ± 0.07 | 64.96 ± 0.13 | 78.11 ± 0.11 | 74.27 ± 0.19 | 78.52 ± 0.09 | 74.93 ± 0.15 |
MP_T | 74.57 ± 0.08 | 70.55 ± 0.15 | 74.47 ± 0.10 | 70.41 ± 0.19 | 77.51 ± 0.14 | 75.95 ± 0.24 | 77.23 ± 0.05 | 75.67 ± 0.10 |
MP2 | 76.36 ± 0.11 | 71.12 ± 0.20 | 76.38 ± 0.09 | 71.18 ± 0.15 | 75.98 ± 0.06 | 73.32 ± 0.14 | 76.00 ± 0.09 | 73.40 ± 0.14 |
MP3 | 74.28 ± 0.12 | 67.83 ± 0.21 | 74.44 ± 0.07 | 68.05 ± 0.12 | 74.92 ± 0.09 | 71.27 ± 0.18 | 77.30 ± 0.10 | 74.75 ± 0.15 |
MP4 | 74.11 ± 0.11 | 68.17 ± 0.20 | 76.80 ± 0.09 | 74.21 ± 0.17 | 74.03 ± 0.09 | 68.08 ± 0.17 | 77.75 ± 0.08 | 75.46 ± 0.13 |
MP_W | 73.14 ± 0.09 | 66.23 ± 0.13 | 73.02 ± 0.12 | 66.15 ± 0.23 | 76.42 ± 0.08 | 73.48 ± 0.12 | 76.89 ± 0.08 | 74.20 ± 0.14 |
MP2_W | 75.60 ± 0.13 | 70.43 ± 0.24 | 75.65 ± 0.10 | 70.43 ± 0.17 | 75.62 ± 0.10 | 73.33 ± 0.20 | 76.15 ± 0.09 | 74.05 ± 0.14 |
RWR | 76.26 ± 0.06 | 72.15 ± 0.10 | 76.25 ± 0.07 | 72.10 ± 0.10 | 77.78 ± 0.05 | 76.11 ± 0.07 | 77.78 ± 0.10 | 76.09 ± 0.17 |
Mix | 75.89 ± 0.09 | 71.21 ± 0.15 | 75.71 ± 0.07 | 71.04 ± 0.12 | 76.41 ± 0.08 | 74.88 ± 0.15 | 76.62 ± 0.08 | 75.11 ± 0.16 |
MP_E64 | 74.69 ± 0.09 | 70.13 ± 0.15 | 74.79 ± 0.06 | 70.26 ± 0.08 | 77.86 ± 0.07 | 76.37 ± 0.11 | 77.91 ± 0.08 | 76.38 ± 0.14 |
MP_E32 | 73.17 ± 0.05 | 68.79 ± 0.10 | 73.19 ± 0.05 | 68.82 ± 0.10 | 75.83 ± 0.13 | 73.88 ± 0.25 | 75.49 ± 0.06 | 73.56 ± 0.13 |
MP_E16 | 72.06 ± 0.09 | 68.98 ± 0.18 | 71.75 ± 0.09 | 68.61 ± 0.18 | 75.09 ± 0.05 | 74.42 ± 0.13 | 74.93 ± 0.06 | 74.17 ± 0.12 |
MP_2H_C | 75.98 ± 0.08 | 71.43 ± 0.15 | 76.17 ± 0.07 | 71.66 ± 0.13 | 78.64 ± 0.09 | 77.06 ± 0.14 | 78.59 ± 0.08 | 76.95 ± 0.14 |
MP_4H_C | 75.72 ± 0.07 | 71.29 ± 0.12 | 76.01 ± 0.07 | 72.89 ± 0.13 | 78.17 ± 0.11 | 76.53 ± 0.18 | 78.27 ± 0.14 | 76.50 ± 0.24 |
HAN | 71.11 ± 0.13 | 69.72 ± 0.25 | 71.24 ± 0.14 | 69.73 ± 0.23 | - | – | – | – |
MP2V | 59.6 | 34.8 | 58.6 | 31.8 | 71.2 | 64.7 | 72.4 | 66.4 |
ASNE | 36.9 | 64.3 | 67.1 | 61.5 | 72.1 | 71.3 | 72.6 | 73.7 |
SHNE | 68.3 | 63.9 | 67.2 | 61.2 | 69.5 | 67.4 | 70.6 | 69.2 |
GSAGE | 69.5 | 61.5 | 67.6 | 57.3 | 71.4 | 66.4 | 73.9 | 70.6 |
GAT | 67.8 | 61.3 | 65.5 | 56 | 73.2 | 70.5 | 75 | 71.5 |
HetGNN | 71.7 | 66.9 | 70.1 | 64.2 | 76.7 | 75.4 | 77.5 | 75.7 |
Task | Recommendation/a–v | Classification/10% | Classification/30% | Clustering | |||||
---|---|---|---|---|---|---|---|---|---|
Model | Recall | Pre | f1 | Macro-F1 | Micro-F1 | Macro-F1 | Micro-F1 | NMI | ARI |
MP_1H | 71.15 ± 0.05 | 31.14 ± 0.02 | 43.32 ± 0.03 | 96.41 ± 0.10 | 96.52 ± 0.10 | 96.70 ± 0.07 | 96.80 ± 0.06 | 87.03 | 90.63 |
MP_2H | 72.08 ± 0.04 | 31.63 ± 0.01 | 43.97 ± 0.02 | 96.96 ± 0.09 | 97.07 ± 0.08 | 97.12 ± 0.14 | 97.24 ± 0.15 | 88.36 | 91.57 |
MP_3H | 72.34 ± 0.05 | 31.78 ± 0.02 | 44.16 ± 0.03 | 96.93 ± 0.11 | 97.04 ± 0.12 | 96.94 ± 0.08 | 97.05 ± 0.09 | 87.95 | 91.12 |
MP_4H | 71.61 ± 0.05 | 31.39 ± 0.02 | 43.65 ± 0.03 | 96.95 ± 0.10 | 97.05 ± 0.10 | 97.15 ± 0.15 | 97.26 ± 0.15 | 88.97 | 92.22 |
MP_S | 70.37 ± 0.05 | 30.86 ± 0.02 | 42.90 ± 0.02 | 96.70 ± 0.10 | 96.82 ± 0.10 | 96.84 ± 0.16 | 96.96 ± 0.16 | 86.95 | 90.61 |
MP_T | 68.18 ± 0.05 | 29.90 ± 0.02 | 41.57 ± 0.03 | 96.26 ± 0.11 | 96.36 ± 0.11 | 96.43 ± 0.15 | 96.56 ± 0.16 | 84.49 | 87.44 |
MP2 | 70.35 ± 0.05 | 30.79 ± 0.02 | 42.83 ± 0.03 | 96.43 ± 0.09 | 96.56 ± 0.09 | 96.58 ± 0.13 | 96.71 ± 0.12 | 87.92 | 91.41 |
MP3 | 71.20 ± 0.03 | 31.25 ± 0.01 | 43.44 ± 0.02 | 96.28 ± 0.08 | 96.41 ± 0.08 | 96.35 ± 0.13 | 96.48 ± 0.12 | 86.66 | 89.99 |
MP4 | 66.81 ± 0.00 | 29.06 ± 0.00 | 40.50 ± 0.00 | 98.60 ± 0.11 | 98.62 ± 0.10 | 98.69 ± 0.08 | 98.71 ± 0.08 | 92.92 | 95.44 |
MP_W | 68.11 ± 0.03 | 29.80 ± 0.01 | 41.46 ± 0.02 | 97.15 ± 0.11 | 97.25 ± 0.10 | 97.27 ± 0.15 | 97.38 ± 0.16 | 87.18 | 90.45 |
MP2_W | 70.77 ± 0.03 | 30.93 ± 0.01 | 43.05 ± 0.01 | 96.01 ± 0.16 | 96.16 ± 0.16 | 96.14 ± 0.15 | 96.30 ± 0.15 | 84.39 | 88.47 |
RWR | 67.19 ± 0.33 | 29.41 ± 0.12 | 40.91 ± 0.18 | 97.05 ± 0.07 | 97.09 ± 0.07 | 97.14 ± 0.12 | 97.18 ± 0.11 | 88.62 | 92.11 |
Mix | 70.09 ± 0.04 | 30.77 ± 0.02 | 42.77 ± 0.02 | 96.89 ± 0.14 | 96.92 ± 0.14 | 97.00 ± 0.13 | 97.14 ± 0.13 | 87.36 | 90.95 |
MP_E64 | 69.35 ± 0.05 | 30.48 ± 0.02 | 42.35 ± 0.03 | 96.66 ± 0.07 | 96.75 ± 0.07 | 96.70 ± 0.07 | 96.78 ± 0.08 | 88.41 | 91.52 |
MP_E32 | 64.31 ± 0.04 | 27.88 ± 0.02 | 38.89 ± 0.02 | 96.18 ± 0.08 | 96.37 ± 0.10 | 96.30 ± 0.13 | 96.49 ± 0.12 | 86.91 | 90.42 |
MP_E16 | 57.66 ± 0.05 | 25.26 ± 0.02 | 35.13 ± 0.03 | 93.11 ± 0.15 | 93.37 ± 0.16 | 93.71 ± 0.16 | 93.94 ± 0.16 | 71.69 | 76.22 |
MP_2H_C | 72.27 ± 0.03 | 31.81 ± 0.01 | 44.18 ± 0.02 | 96.69 ± 0.07 | 96.81 ± 0.06 | 96.85 ± 0.10 | 96.97 ± 0.10 | 87.32 | 90.66 |
MP_4H_C | 71.55 ± 0.04 | 31.43 ± 0.01 | 46.67 ± 0.02 | 96.51 ± 0.14 | 96.63 ± 0.15 | 96.65 ± 0.11 | 96.76 ± 0.09 | 86.80 | 90.52 |
HAN | – | – | – | 97.71 ± 0.07 | 97.76 ± 0.07 | 97.84 ± 0.13 | 97.88 ± 0.13 | 89.95 | 93.22 |
MP2V | 46.8 | 20.4 | 28.4 | 97.2 | 97.3 | 97.5 | 97.5 | 89.4 | 93.3 |
ASNE | 38.2 | 17.1 | 23.6 | 96.5 | 96.7 | 96.9 | 97 | 85.4 | 89.8 |
SHNE | 55.2 | 23.3 | 32.7 | 93.9 | 94 | 93.9 | 94.1 | 77.6 | 81.3 |
GSAGE | 51.2 | 22.4 | 31.2 | 97.8 | 97.8 | 97.9 | 98 | 91.4 | 94.5 |
GAT | 51.8 | 22.7 | 31.6 | 96.2 | 96.3 | 96.5 | 96.5 | 84.5 | 88.2 |
HetGNN | 60.6 | 26.4 | 36.8 | 97.1 | 97.1 | 97.1 | 97.2 | 88.6 | 92.1 |
Rights and permissions
About this article
Cite this article
Liang, X., Ma, Y., Cheng, G. et al. Meta-path-based heterogeneous graph neural networks in academic network. Int. J. Mach. Learn. & Cyber. 13, 1553–1569 (2022). https://doi.org/10.1007/s13042-021-01465-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13042-021-01465-8