An efficient PyTorch implementation of the evaluation metrics in recommender systems

Overview

Highlights

• Efficient (vectorized) implementations over mini-batches
• Standard RecSys metrics: precision, recall, map, mrr, hr, ndcg
• Beyond-accuracy metrics: e.g. coverage, diversity, novelty, etc.
• All metrics support a top-k argument.

Why do we need recsys_metrics?

Installation

You can install recsys_metrics from PyPI:

pip install recsys_metrics

Or you can also install the latest version from source:

pip install git+https://github.com/zuoxingdong/recsys_metrics.git@master

Note that we support Python 3.7+ only.

How to use

Let us take Hit Rate (HR) to illustrate how to use this library:

preds = torch.tensor([
    [.5, .3, .1],
    [.3, .4, .5]
])
target = torch.tensor([
    [0, 0, 1],
    [0, 1, 1]
])
hit_rate(preds, target, k=1, reduction='mean')

>> tensor(0.5000)

The one example in the batch does not have a hit (i.e. top-1 item is not a relevant item) and second example in the batch gets a hit (i.e. top-1 item is a relevant item). Thus, we have a hit-rate of 0.5.

The API of other metrics are of the same format.

Benchmark

Metrics	Single Example	Mini-Batch
Precision
Recall
MAP
MRR
HR
NDCG

Citation

This work is inspired by Torchmetrics from PyTorchLightning Team.

Please use this bibtex if you want to cite this repository in your publications:

@misc{recsys_metrics,
      author = {Zuo, Xingdong},
      title = {recsys_metrics: An efficient PyTorch implementation of the evaluation metrics in recommender systems.},
      year = {2021},
      publisher = {GitHub},
      journal = {GitHub repository},
      howpublished = {\url{https://github.com/zuoxingdong/recsys_metrics}},
    }

GitHub

View Github