data-preparation.md

Data Preparation Guideline

Movielens 1M (ML-1M)

1. Download the raw atomic files from RecBole's dataset hub. [link]

2. Sample 6,040 (all) users and corresponding candidate items for experiments.

   # generate ml-1m-full.random
   cd llmrank/
   python sample_candidates.py -d ml-1m-full -u 6040 -s random

   Note that ml-1m-full.random has the format user_id<\t>candidate_item_1 candidate_item_2 ...:

   5077	1116 884 1081 ...
   5075	3913 749 445 ...
   

   For each user in ml-1m-full.random, there are 100 randomly selected candidate items. We store these candidates for fair comparison between different variants.

   When we do experiments using python evaluate.py -m Rank -d ml-1m-full, the number of items used to construct the candidate set depends on the hyperparameter recall_budget. For example, when recall_budget=20, the first 20 items of each line in ml-1m-full.random will be used.

   Then if has_gt=True (ground truth items are guaranteed to appear in the candidate set), the ground truth item will be appended into the candidate set (implemented in trainer.py, lines 70-90).

3. For different candidate retrieval strategies, please replace -s random to -s bm25, -s bert, -s pop, -s bpr, -s gru4rec, or -s sasrec in step 2.

4. For generating candidates which are retrieved by multiple strategies (.rabmbepobpgrsa_3, denotes for concatenating the first two letters of each strategy name).

   python mix_sources.py

Amazon Review - Games (Games)

1. Download raw datasets from Amazon review data 2018, including the metadata and ratings only data of each category. [link] Here is an example.

   dataset/
     raw/
       Metadata/
         meta_Video_Games.json.gz
       Ratings/
         Video_Games.csv
   

2. Process downstream datasets.

   cd llmrank/
   python data_process_amazon.py --dataset Games
   

   Note that following [UniSRec], we split the data into separate files for training, validation and evaluation: .train.inter, .valid.inter, .test.inter .

3. Repeat the second step in processing ML-1M datasets, for example:

   # generate Games-6k.random
   cd llmrank/
   python sample_candidates.py -d Games-6k -u 6000 -s random
   

Auxiliary Files of UniSRec / VQ-Rec Preparation Guideline

UniSRec

1. Generate item embedding from item related text (title).

   cd llmrank/
   python dataset/unisrec_auxiliary_files_process.py
   

VQ-Rec

1. Run step 1 of UniSRec to obtain *.feat1CLS.
2. Generate the item index file.

   python dataset/vqrec_auxiliary_files_process.py --dataset ml-1m