tunazislam / yoga-user-network-yun

IEEE ICSC2021

Yoga-User-Network-YUN

This repository contains code for [Do You Do Yoga? Understanding Twitter Users' Types and Motivations using Social and Textual Information], IEEE ICSC 2021.

Data:

1. Please download the 'data' folder from following link (On request):

[YUN Data]

2. 'data' folder should be kept inside the 'Yoga-User-Network-YUN/code/' folder.

data/yoga_user_name_loc_des_mergetweets_yoga_1300_lb.csv file contains 6 columns: name, location, description, text, utype, umotivation

3. Please download pre-trained Word2Vec from this link and save it inside the 'code/pre-trained/' folder

[Word2Vec]

code/pre-trained/GoogleNews-vectors-negative300.bin

4. Pre-trained emoji2vec.bin is inside 'code/pre-trained/' folder

code/pre-trained/emoji2vec.bin

5. Train and Test data are inside data folder.

data/train.txt

data/test.txt

Computing Machine:

OS: MacBookPro, Processor: 2.5 GHz Dual-Core Intel Core i7, Memory: 16 GB 2133 MHz LPDDR3

Software Packages and libraries:

python 3.6.6
PyTorch 1.1.0
jupiter notebook
pandas
gensim
nltk
spacy
emoji
sklearn
matplotlib
numpy
preprocessor
transformers

Run all codes from 'code' directory.

cd code

Construct User Networks:

1. Create user graphs:

python create_yoga_graph.py  data/yoga_user_mentioned_yoga_1300_lb.txt 

Now user graph will save in data/usergraph folder.

2. Merge all usergraphs to make one:

cat data/usergraph/*.txt > data/yoga_usergraph.txt

Create Embeddings:

1. Create description, location, tweets embeddings using pre-trained Word2Vec and Emoji2Vec. This will take ~2 hours to run in CPU.

create_embeddings.ipynb

This will create following three embeddings:

data/locationEmbeddings.pt

data/descriptionEmbeddings.pt

data/tweetsEmbeddings.pt

2. Create user network embeddings using Node2Vec and input graph for this is data/yoga_usergraph.txt

Please download Node2Vec from this link:

https://github.com/aditya-grover/node2vec

This will create following embedding:

data/userNetworkEmd.emd

Run Models:

1. Run Joint embedding attention-based neural network model Yoga User Network (YUN) model (Dec + Loc + Twt + Net). Each of them will take ~24 hours to run in CPU.

DLTN_utype_2layer.ipynb. 

DLTN_umotivation_2layer.ipynb

2. Run description only baseline model. Each of them will take ~10 minutes to run in CPU.

Description_BiLSTMAttn_utype_2layer_classifier.ipynb

Description_BiLSTMAttn_umotivation_2layer_classifier.ipynb

3. Run location only baseline model. Each of them will take ~7 minutes to run in CPU.

Location_only_utype_2layer_classifier.ipynb

Location_only_umotivation_2layer_classifier.ipynb

4. Run tweets only baseline model. Each of them will take ~15 hours to run in CPU.

Tweets_BiLSTMAttn_utype_2layer_classifier.ipynb

Tweets_BiLSTMAttn_umotivation_2layer_classifier.ipynb

5. Run user network only baseline model. Each of them will take ~5 minutes to run in CPU.

Network_only_utype_2layer_classifier.ipynb

Network_only_umotivation_2layer_classifier.ipynb

6. Run joint Description and Location (Des + Loc) model. Each of them will take ~10 minutes to run in CPU.

DL_utype_2layer_classifier.ipynb

DL_umotivation_2layer_classifier.ipynb

7. Run joint Description, Location, Tweet (Des + Loc + Twt) model. Each of them will take ~18 hours to run in CPU.

DLT_utype_2layer_classifier.ipynb

DLT_umotivation_2layer_classifier.ipynb

8. Run joint Description, Location, Network (Des + Loc + Net) model. Each of them will take ~10 minutes to run in CPU.

DLN_utype_2layer_classifier.ipynb

DLN_umotivation_2layer_classifier.ipynb

9. Run fine-tuned BERT model on Description (Description_BERT). Each of them will take ~10 minutes to run in GPU.

baseline_BERT_finetuned_description_utype_preprocessed.ipynb

baseline_BERT_finetuned_description_umotivation_preprocessed.ipynb

10. Run fine-tuned BERT model on Location (Location_BERT). Each of them will take ~10 minutes to run in GPU.

baseline_BERT_finetuned_location_utype_preprocessed.ipynb

baseline_BERT_finetuned_location_umotivation_preprocessed.ipynb

11. Run fine-tuned BERT model on Tweets (Tweets_BERT). Each of them will take ~15 minutes to run in GPU.

baseline_BERT_finetuned_tweet_utype_preprocessed_split.ipynb

baseline_BERT_finetuned_tweet_umotivation_preprocessed_split.ipynb

Citation:

If you find the paper useful in your work, please cite:

@inproceedings{islam2021you,
  title={Do You Do Yoga? Understanding Twitter Users' Types and Motivations using Social and Textual Information},
  author={Islam, Tunazzina and Goldwasser, Dan},
  booktitle={2021 IEEE 15th International Conference on Semantic Computing (ICSC)},
  pages={362--365},
  year={2021},
  organization={IEEE}
}