Source code and datasets of ACL2017 paper: "CANE: Context-Aware Network Embedding for Relation Modeling"
This folder "datasets" contains three datasets used in CANE, including Cora, HepTh and Zhihu. In each dataset, there are two files named "data.txt" and "graph.txt".
- data.txt: Each line represents the text information of a vertex.
- graph.txt: The edgelist file of current social network.
Besides, there is an additional "group.txt" file in Cora.
- group.txt: Each vertex in Cora has been annotated with a label. This file can be used for vertex classification.
Run the following command for training CANE:
python3 run.py --dataset [cora,HepTh,zhihu] --gpu gpu_id --ratio [0.15,0.25,...] --rho rho_value
For example, you can train like:
python3 run.py --dataset zhihu --gpu 0 --ratio 0.55 --rho 1.0,0.3,0.3
The experimental results are generated by the newest version of codes:
| 0.15 | 0.25 | 0.35 | 0.45 | 0.55 | 0.65 | 0.75 | 0.85 | 0.95 | |
|---|---|---|---|---|---|---|---|---|---|
| cora | 85.2 | 90.5 | 92.2 | 93.5 | 93.4 | 93.6 | 94.4 | 95 | 92.5 |
| HepTh | 85 | 89.7 | 91.7 | 95 | 94.4 | 94.2 | 95.1 | 95.8 | 93.1 |
| zhihu | 64.5 | 67.1 | 69.2 | 69.9 | 72 | 72.2 | 72.5 | 72.8 | 73.3 |
Computation time for keyword/keyphrase extraction methods given the entire graph:
| T: 5 | Time (sec) |
|---|---|
| TFIDF | 2.622457 |
| YAKE | 24.542249 |
| PRank | 41.415699 |
| TxtRank | 40.290922 |
| ToRank | 34.672437 |
| T: 10 | Time (sec) |
|---|---|
| TFIDF | 2.669563 |
| YAKE | 62.481739 |
| PRank | 37.410711 |
| TxtRank | 35.84592 |
| ToRank | 32.635109 |
The experimental results for the cora citation graph with keyword/keyphrase extraction techniques:
| T: 5 | 0.15 | 0.45 | 0.75 |
|---|---|---|---|
| Abst | 84.6 | 92.9 | 94.1 |
| TFIDF | 64.6 | 77.2 | 88.2 |
| YAKE | 71.5 | 80.6 | 86.2 |
| PRank | 76.8 | 90 | 93.9 |
| KBERT3 | 74.4 | 87.4 | 90.5 |
| KBERT3G | 77.3 | 88 | 93 |
| T: 10 | 0.15 | 0.45 | 0.75 |
|---|---|---|---|
| Abst | 84.6 | 92.9 | 94.1 |
| TFIDF | 66.8 | 81.8 | 89.8 |
| YAKE | 70.2 | 86.6 | 91.7 |
| PRank | 85.3 | 92.2 | 94.6 |
Time in seconds:
| T: 5 | 0.15 | 0.45 | 0.75 |
|---|---|---|---|
| Abst | 42.1 | 113.4 | 190 |
| TFIDF | 23.4 + 0.852434 | 36.4 + 1.83892 | 59.6 + 2.249102 |
| YAKE | 13.3 + 10.146231 | 35.6 + 18.81623 | 59.6 + 21.892501 |
| PRank | 17 + 16.420615 | 37.7 + 29.095505 | 63.6 + 33.532091 |
| KBERT3 | 13.4 | 34.5 | 59.4 |
| KBERT3G | 12.7 | 35.6 | 57 |
| T: 10 | 0.15 | 0.45 | 0.75 |
|---|---|---|---|
| Abst | 42.1 | 113.4 | 190 |
| TFIDF | 20.8 + 0.833934 | 34.8 + 1.96583 | 59.8 + 2.243566 |
| YAKE | 13.5 + 25.919926 | 38 + 48.326221 | 59.7 + 56.408093 |
| PRank | 17 + 15.118391 | 47.4 + 28.553312 | 80.2 + 33.022757 |
- Tensorflow == 1.11.0
- Scipy == 1.1.0
- Numpy == 1.16.2
If you use the code, please cite this paper:
Cunchao Tu, Han Liu, Zhiyuan Liu, Maosong Sun. CANE: Context-Aware Network Embedding for Relation Modeling. The 55th Annual Meeting of the Association for Computational Linguistics (ACL 2017).
For more related works on network representation learning, please refer to my homepage.