Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval

This repository is a fork of the original gSASRec-pytorch implementation. We provide additional configuration files and launch scripts for training SASRec under various regimes (in-batch, uniform, full softmax, etc.). This repository is dedicated to our paper: "Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval".

Compared to the original gSASRec repository, we introduce the following additions:

• Extra training scripts tailored to various objective functions (full softmax, sampled softmax, mns, log-q correction, etc.)
• Configuration files for the ML1M, Steam, and Gowalla datasets
• Preprocessed data for ML1M, Steam, and Gowalla, prepared using both leave-one-out and time-split strategies

Citing

If this repository or its results are useful in academic or industrial work, please cite both our paper and the original gSASRec:

• Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval (RecSys ’25)

@inproceedings{Khrylchenko_2025,
title = {Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval},
booktitle = {Proceedings of the Nineteenth ACM Conference on Recommender Systems},
publisher = {ACM},
author = {Khrylchenko, Kirill and Baikalov, Vladimir and Makeev, Sergei and Matveev, Artem and Liamaev, Sergei},
year = {2025},
month = sep,
pages = {545--550},
doi = {10.1145/3705328.3748033},
series = {RecSys ’25}
}

• gSASRec: Reducing Overconfidence in Sequential Recommendation Trained with Negative Sampling (RecSys ’23)

@inproceedings{petrov2023gsasrec,
title = {gSASRec: Reducing Overconfidence in Sequential Recommendation Trained with Negative Sampling},
author = {Petrov, Aleksandr Vladimirovich and Macdonald, Craig},
booktitle = {Proceedings of the 17th ACM Conference on Recommender Systems},
pages = {116--128},
year = {2023}
}

Getting Started

To run code several packages should be installed:

pip install -r requirements.txt

To train or evaluate the model with our configurations, use the following commands:

Training Example

SASRec (Original)

Implementation of theoriginal SASRec with BPR loss:

python train_sasrec.py --config=configs/ml1m_sasrec.py

gSASRec

Implementaion of the model from "gSASRec: Reducing Overconfidence in Sequential Recommendation Trained with Negative Sampling":

python train_gsasrec.py --config=configs/ml1m_gsasrec.py

SASRec with full softmax

Implemenation which utilizes cross-entropy loss over the entire item set:

python train_full_softmax.py --config=configs/ml1m_other.py 

SASRec with uniformly sampled negatives

Implementation which uniformly subsamples random items from the corpus and computes sampled softmax loss:

python train_uniform.py --config=configs/ml1m_other.py 

SASRec with in-batch negatives

Implementation which subsamples negatives for each positive interaction from the current mini-batch:

python train_in_batch.py --config=configs/ml1m_other.py 

SASRec with in-batch negatives with original log-q correction

Implementation which subsamples negatives for each positive interaction from the current mini-batch and applies the original logQ correction:

python train_in_batch_logq_old.py --config=configs/ml1m_other.py 

SASRec with in-batch negatives with our implementation of log-q correction

Implementation which subsamples negatives for each positive interaction from the current mini-batch and applies our implementation of logQ correction:

python train_in_batch_logq_new.py --config=configs/ml1m_other.py 

SASRec with Mixed Negative Sampling (MNS)

Implementation which subsamples negatives both uniformly from the corpus and from the current mini-batch (even split):

python train_mns.py --config=configs/ml1m_other.py 

SASRec with mixed negative sampling (MNS) with original log-q correction

Implementation which leverages mixed negative sampling with the original logQ correction:

python train_mns_logq_old.py --config=configs/ml1m_other.py 

SASRec with mixed negative sampling (MNS) with our implementation of log-q correction

Implementation which leverages mixed negative sampling with our implementation of log-q correction:

python train_mns_logq_new.py --config=configs/ml1m_other.py 

Using Other Datasets

To use a different dataset, replace ml1m in the config filename with steam or gowalla. Configs with a _time suffix after the dataset name are dedicated to time-based splits.

You can download datasets from here. Unzip datasets.tar in the root of the repository by running tar -xvf datasets.tar.

Evaluation Example

To evaluate any model, use the same script with the appropriate config and checkpoint:

Example

To evaluate SASRec with full softmax on the Steam dataset:

python evaluate.py --config=configs/steam_other.py --checkpoint=your_checkpoint.pt

Replace your_checkpoint.pt with the checkpoint produced after training.