This repo contains the code for DITL.

Usage

1. To run the code, you need to

   • download the Wiki data from https://portals.mdi.georgetown.edu/public/demographic-inference

2. We use some dummy examples here to help you start. This includes

   • a random generated IMDB dataset
     • imdb_gt.csv -- IMDB ground truth with fake username and gender
     • imdb_embeddings -- the embeddings for each user from the IMDB dataset
   • a random generated Wiki dataset
     • wiki_gt.csv -- wiki ground truth with fake username and gender
     • wiki_embeddings -- the embeddings for each user from the IMDB dataset
   • a random generated model
     • under the directory model/model

3. classifier.py

   • Python script to train a model from scratch.
   • Command: python classifier.py --inference_type xxx --bin_type xxx --epoch xxx --fix_seq_len xxx
     • inference_type: gender/age
     • bin_type: for age, you can choose two/three/four. For gender, this attribute will be ignored whatever you type
     • epoch: number of epochs to run
     • fix_seq_len: number of posts to use
   • To run the dummy example, execute
     • python classifier.py --inference_type gender --bin_type three --epoch 5 --fix_seq_len 200

4. classfier_reg.py

   • Python script to train a model with regularization
   • Command: python classifier.py --inference_type age --bin_type three --epoch 5 --fix_seq_len 200
   • inference_type
     • As above
   • To run the dummy example, execute
     • python classifier_reg.py --inference_type gender --bin_type three --epoch 5 --fix_seq_len 200