Low rank field-weighted FM model in PyTorch (paper)

The code was forked and modified from https://github.com/rixwew/pytorch-fm, written by rixwew. The API Documentation of the original code is: https://rixwew.github.io/pytorch-fm

This package provides a PyTorch implementation of low rank factorization machine models, the factorization model baselines and the common datasets in CTR prediction. The code is running on the following datasets: Avazu, Criteo, and Movielens.

The instructions contain three main steps:

• Step 1: Data Preprocessing and
• Step 2: Running the models on the preprocessed data.
• Step 3: Analyzing the results

Moreover, this repository also contains a notebook notebooks/inference_timing.ipynb that demonstrates inference speedups attained by low-rank models over pruned models, for different ad auction sizes and amounts of context fields.

Available Datasets

• MovieLens Dataset
• Criteo Display Advertising Challenge
• Avazu Click-Through Rate Prediction

Data Processing

Example of preprocessed dataset that will be obtained after running the preprocessing scripts on the initial datasets:

label, user_id, item_id, C1, C2, …
<label>,10,2,3,17,11,15
<label>,11,2,4,16,12,14

How to preprocess Avazu dataset

1. Download the initial file (train, 6.31G) should be called data_avazu.csv https://www.kaggle.com/datasets/atirpetkar/avazu-ctr

2. put it to pytorch-fm/torchfm/test-datasets/avazu/

3. To create train-validation-test datatsets, from the python shell run:

   from torchfm.torch_utils.parsing_datasets.avazu.avazu_parsing import process_data process_data()

4. Check that now train.csv, test.csv, validation.csv and stored under /pytorch-fm/torchfm/test-datasets/avazu/

5. Check you have enough (5G) available space and proceed to run the ML models on the train-validation-test datasets.

How to preprocess Criteo dataset

1. Download the initial file (train.txt, 11.15G) should be called data_criteo.csv https://www.kaggle.com/datasets/mrkmakr/criteo-dataset

2. put it to pytorch-fm/torchfm/test-datasets/criteo/

3. To create train-validation-test datatsets, from the python shell run:

   from torchfm.torch_utils.parsing_datasets.criteo.criteo_parsing import CriteoParsing CriteoParsing.do_preprocessing()

4. Check that now train.csv, test.csv, validation.csv and stored under /pytorch-fm/torchfm/test-datasets/criteo/

5. Check you have enough (5G) available space and proceed to run the ML models on the train-validation-test datasets.

How to preprocess MovieLens dataset

1. Download the initial file users.dat, movies.dat, ratings.dat from https://www.kaggle.com/datasets/sherinclaudia/movielens

2. put it to pytorch-fm/torchfm/test-datasets/movielens/

3. To create train-validation-test datatsets, from the python shell run:

   from torchfm.torch_utils.parsing_datasets.movielens.movielens_parsing import MovielensParsing MovielensParsing.process_data()

4. Check that now train.csv, test.csv, validation.csv and stored under /pytorch-fm/torchfm/test-datasets/movielens/

5. Check you have enough (5G) available space and proceed to run the ML models on the train-validation-test datasets.

Instructions: How to run a dataset on a local environment

1. Copy the code to the dedicated folder. Install all requirements.

2. In the shell redefine PYTHONPATH to point to your project root, .e.g, export PYTHONPATH=$PYTHONPATH:/home/${USER}/pytorch-fm/src:/home/${USER}/pytorch-fm/src/main_functions

3. Edit pytorch-fm/torchfm/torch_utils/constants.py file to have

   • base_path_project pointing to your project root
   • Edit dataset_name to have the dataset you run on (avazu, criteo, movielens) E.g., dataset_name = movielens

   Edit pytorch-fm/src/main_functions/run_processes.py file in examples folder, to refer to the list of options to run: currently, as example, it contains lst_tmp in "for tpl in lst_tmp:".

4. Copy train.csv/validation.csv/test.csv splitted datasets to be under pytorch-fm/data/test-datasets// (dataset is either criteo or avazu or movielens) Then, open a python shell by running just: "python" command from the shell.

5. Check that you have enough space (at least 5G available) after all these steps in your local environment, e.g., by: df -h /home/default/your_location Otherwise remove non-required data (e.g., datasets you don’t use for the current run)

6. Check that you have pytorch-fm/data/tmp_save_dir (if not create this folder)

7. If you are rerunning, check you don’t have a previous run. results stored (especially .log files - locking the next run), otherwise remove: rm pytorch-fm/data/tmp_save_dir/*

8. Run by python ./pytorch-fm/src/main_functions/run_processes.py

9. After the run is done, the results are saved in /pytorch-fm/data/tmp_save_dir/optuna_results.txt. Also, debug info is saved in /pytorch-fm/data/tmp_save_dir/debug_info.txt

Instructions: How to analyze the results

Open the notebook in notebooks/analysis.ipynb, modify the line

files = glob.glob('../data/optuna_results*.txt')

to point to the appropriate paths with the result files saved in the previous stage, and run the notebook. It produces a Pandas table with the metrics and corresponding lifts of each dataset, and a LaTeX code snippet you can put in a paper to share the results.