Industrial recommender systems generally have two main stages: matching and ranking. In the first stage, candidate item matching (also known as candidate retrieval) aims for efficient and high-recall retrieval from a large item corpus. MatchBox provides an open source library for candidate item matching, with stunning features in configurability, tunability, and reproducibility.
| Publication | Model | Paper | Benchmark |
|---|---|---|---|
| UAI'09 | MF-BPR | BPR: Bayesian Personalized Ranking from Implicit Feedback | |
| RecSys'16 | YoutubeNet | Deep Neural Networks for YouTube Recommendations | |
| CIKM'21 | MF-CCL/ SimpleX | SimpleX: A Simple and Strong Baseline for Collaborative Filtering |
We suggest to use the following environment where we test MatchBox only.
- CUDA 10.0
- python 3.6
- pytorch 1.0
- PyYAML
- pandas
- scikit-learn
- numpy
- h5py
- tqdm
The code workflow is structured as follows:
# Set the dataset config and model config
feature_cols = [{...}] # define feature columns
label_col = {...} # define label column
params = {...} # set data params and model params
# Set the feature encoding specs
feature_encoder = FeatureEncoder(feature_cols, label_col, ...) # define the feature encoder
datasets.build_dataset(feature_encoder, ...) # fit feature_encoder and build dataset
# Load data generators
train_gen, valid_gen, test_gen = datasets.h5_generator(feature_encoder, ...)
# Define a model
model = SimpleX(...)
# Train the model
model.fit(train_gen, valid_gen, ...)
# Evaluation
model.evaluate(test_gen)For reproducing the experiment results, you can run the benchmarking scripts with the corresponding configs as follows.
- --config: The config directory where dataset config and model config are located.
- --expid: The experiment id defined in a model config file to set a group of hyper-parameters.
- --gpu: The gpu index used for experiment, and -1 for CPU.
cd data/Yelp18/yelp18_m1
python matchbox_convert_data.py
cd model_zoo/SimpleX
python run_expid.py --config ./config/SimpleX_yelp18_m1 --expid SimpleX_yelp18_m1 --gpu 0
...
python run_expid.py --config ./config/SimpleX_amazonbooks_m1 --expid SimpleX_amazonbooks_m1 --gpu 0
python run_expid.py --config ./config/SimpleX_gowalla_m1 --expid SimpleX_gowalla_m1 --gpu 0The running logs are also available in each config directory.