ICWSM'25

Influence Maximization in Temporal Social Networks with a Cold-start Problem: A Supervised Approach

[Paper]

Preparation

1. Install dependencies:

pip install -r requirements.txt

DGL installation:

pip install dgl-cu116 dglgo -f https://data.dgl.ai/wheels/repo.html

2. Data downloading and run preprocessing

cd data_processing && python3 read_tweet.py

Directory tree:

.
├── checkpoint
├── data_processing
│   ├── DGLgraph
│   ├── netease_week
│   └── sampled_week
├── inference
├── labeling
└── models
    ├── cold_start
    ├── graphmae
    └── tgn_raw

2. Labeling

python3 labeling/label.py

see README in labeling

3. Running Cold-start algorithm

python3 models/cold_start/cold_start.py

Offline Training

Task: node classification

Metrics: accuracy

Data:
./data_processing/netease_week/_graphs.csv, ./data_processing/netease_week/_features.csv, ./data_processing/netease_week/*_labels.csv

Output:
./checkpoint/netease_graphmae_best_model.ckpt

python run_all_model.py --task train_offline --supervised

Online Training

Same with the offline training, except that no test data is splited

python run_all_model.py --task train_online --supervised

Online Inference

Task: Influence Maximization
Given K users at first that will be recommended in priority, how many users will undergo invitation/adoption at least once after T times.

Data:
./data_processing/netease_week/_graphs.csv, ./data_processing/netease_week/_features.csv

Model checkpoint:
./checkpoint/netease_graphmae_best_model.ckpt

Metrics:
daily number of users that have invitation or adoption

Output:
./inference/netease_graphmae_emb.npy ./inference/netease_graphmae_pred.npy

python run_all_model.py --task inference_online --supervised --inference_data netease_week

train TGN

python models/tgn_raw/train_model.py --use_memory --prefix tgn-attn --n_runs 1 --n_epoch 10 --bs 64 --message_dim 50 --memory_dim 86 --platform netease