Weighted Factorization Machines

This is an implementation of Weighted Factorization Machines (WFMs) with tensorflow based on tffm.

Documentation

Weighted Factorization Machines (WFM) is an extension to Factorization Machines where additional weight parameters are learned for groups of features (e.g. users, items and context).

This repository is an implementation of WFM with python and tensorflow. The implementation supports reading of multiple datasets, training and evalation. WFM can be used by its command line interface or by calling its API. Next section describes the usage of command line interface. This section will be completed with more details. For questions please contact: b DOT loni AT tudelf DOT nl.

Usage

usage: wfm.py [-h] [--dataset ml1m/frappe/kassandra/msd/goodbooks]
              [--model bpr/fm/fmp/wfm/wfmp/wmf] [--load_model FOLDER_PATH]
              [--save_model FOLDER_PATH] [--epochs N] [--eval_freq N]
              [--eval_results FILE_PATH] [--order N] [--k K]
              [--hyper_params LR,REG,STD] [--batch_size N]
              [--has_context true/false] [--implicit true/false]
              [--weights all-one/all-diff/c-one/c-diff]
              [--all-conf true/false]

Weighted Factorization Machines

optional arguments:
  -h, --help            show this help message and exit
  --dataset ml1m/frappe/kassandra/msd/goodbooks
                        name of dataset (default: )
  --model bpr/fm/fmp/wfm/wfmp/wmf
                        name of the model (default: fmp)
  --load_model FOLDER_PATH
                        the path to the model tf model folder (default: None)
  --save_model FOLDER_PATH
                        folder path to save the tf model (default: None)
  --epochs N            number of epochs (default: 10)
  --eval_freq N         evaluate every N epochs (default: 10)
  --eval_results FILE_PATH
                        name of the file to save evaluation results to
                        (default: None)
  --order N             order of FM (default: 2)
  --k K                 number of latent factors (default: 10)
  --hyper_params LR,REG,STD
                        comma-separated list of hyper-parameters:
                        LearningRate, Regularization, InitStd. (default:
                        0.01,0.01,0.01)
  --batch_size N        number of samples in each mini-batch (default: 10000)
  --has_context true/false
                        Whether data has context or not (default: True)
  --implicit true/false
                        Whether dataset is implicit or explicit (default:
                        True)
  --weights all-one/all-diff/c-one/c-diff
                        how the weights should be initialized (default: c-dif)
  --all-conf true/false
                        Run experiments with pre-defined configurartions
                        (default: False)