bandit_reranker

Clone this Project

git clone https://github.com/chihming/bandit_reranker

Prepare Data

Run this:

bash prepare.sh

You'll get train/dev/test data: exp/train.data / exp/dev.data / exp/test.data. Each of them has the format [user_id]\t[item_id]\t[value] likes:

58146 132 4.000000
68889 457 5.000000
68889 653 3.000000

Generate Environments for Reranking

Run this:

bash train.sh

You'll get two .env data: exp/dev.data.env and exp/test.data.env. Each of them has the format [user_id]\t[answer_ids]\t[recommendation_ids] likes:

68889   457 653 356 153 434     457 356 593 589 153 1 364 377 527 434 208 47 50 34 32 185 253 648 367 454
52509   711 76 135 32 708 736   135 9 637 743 88 694 762 61 12 74 736 260 724 653 839 32 3 102 1210 708
...   ...   ...

The reank task (i.e. rerank.py) is to rerank the <recommendation_ids> for matching <answer_ids>.

Train, Rerank and Evaluate

Run this:

bash rerank.sh

You'll get evalation results.

Add NEW Bandit Method

1. inherit the base class in bandit/bandit.py
2. call your created class in rerank.py