We use BERT embeddings (bert-base-multilingual-uncased) instead of the unknown and irreproducible Ma Jing Embeds of Twitter15/16. Since we cannot reproduce those, and half the tweets are missing, and a lot of her embeds have more tweets in the processed version than in the reported tree, we change datasets to PHEME.
(That's the point of this entire github repo)
Keep in mind the original code reports results via batch wise averaged results for F1.
The results are hence inaccurate as compared to overall results.
forced_parser.py can be run on an output file (see it's description string inside) and it will give you a quick comparison between actual and their reported version. The issue is minor but an issue nonetheless
Note that: bigcn_predict_all.py.py has an example within for usage on single instances. (requires a tree) Otherwise it is PHEME only, and requires the actual files.bigcn_predict_all.py
Tian Bian, Xi Xiao, Tingyang Xu, Peilin Zhao, Wenbing Huang, Yu Rong, Junzhou Huang. Rumor Detectionon Social Media with Bi-Directional Graph Convolutional Networks. AAAI 2020.
#READ: HOW TO ACTUALLY INSTALL It will fast track your process for reimplementing this/ running this.
It acts to LOAD PHEME data in a (seemingly) correct (from what i can tell) way for you to get bigcn results.
SHOULD NOT VARY TOO MUCH. (hopefully. I mean i wrote down the link for where to get torch scatter,sparse and cluster too...)
The datasets used in the experiments were based on the three publicly available Weibo and Twitter datasets released by Ma et al. (2016) and Ma et al. (2017):
Jing Ma, Wei Gao, Prasenjit Mitra, Sejeong Kwon, Bernard J Jansen, Kam-Fai Wong, and Meeyoung Cha. Detecting rumors from microblogs with recurrent neural networks. In Proceedings of IJCAI 2016.
Jing Ma, Wei Gao, Kam-Fai Wong. Detect Rumors in Microblog Posts Using Propagation Structure via Kernel Learning. ACL 2017.
In the 'data' folder we provide the pre-processed data files used for our experiments. The raw datasets can be respectively downloaded from https://www.dropbox.com/s/46r50ctrfa0ur1o/rumdect.zip?dl=0. and https://www.dropbox.com/s/7ewzdrbelpmrnxu/rumdetect2017.zip?dl=0.
The Weibo datafile 'weibotree.txt' is in a tab-sepreted column format, where each row corresponds to a weibo. Consecutive columns correspond to the following pieces of information:
1: root-id -- an unique identifier describing the tree (weiboid of the root);
2: index-of-parent-weibo -- an index number of the parent weibo for the current weibo;
3: index-of-the-current-weibo -- an index number of the current weibo;
4: list-of-index-and-counts -- the rest of the line contains space separated index-count pairs, where a index-count pair is in format of "index:count", E.g., "index1:count1 index2:count2" (extracted from the "text" field in the json format from Weibo raw datasets)
For a detailed description of Twitter datafile 'data.TD_RvNN.vol_5000.txt' can be seen at RvNN.
python==3.5.2
numpy==1.18.1
torch==1.4.0
torch_scatter==1.4.0
torch_sparse==0.4.3
torch_cluster==1.4.5
torch_geometric==1.3.2
tqdm==4.40.0
joblib==0.14.1
Make sure that cuda/bin, cuda/include and cuda/lib64 are in your $PATH, $CPATH and $LD_LIBRARY_PATH respectively before the installation, e.g.:
$ echo $PATH
>>> /usr/local/cuda/bin:...
$ echo $CPATH
>>> /usr/local/cuda/include:...
and
$ echo $LD_LIBRARY_PATH
>>> /usr/local/cuda/lib64
on Linux or
$ echo $DYLD_LIBRARY_PATH
>>> /usr/local/cuda/lib
on macOS.
Run script
$ sh main.sh
and choose "model/Weibo/BiGCN_Weibo.py" for BiGCN model on Weibo dataset or "model/Twitter/BiGCN_Twitter.py" on Twitter15/Twitter16 dataset.
In "main.sh", two arguments need to be specified, representing the datasetname and iteration times respectively. E.g.,
python ./model/Twitter/BiGCN_Twitter.py Twitter15 100
will reproduce the average experimental results of 100 iterations of BiGCN model on Twitter15 dataset with 5-fold cross-validation.
If you find this code useful, please let us know and cite our paper.
If you have any question, please contact Tian at: bt18 at mails dot tsinghua dot edu dot cn.