Sequence or Pseudo-Sequence?

An Analysis of Sequential Recommendation Datasets

This repo contains our re-implementation and extension of SASRec [1] as well as code to download, process and format the datasets we use in our paper.

In order to get the data, run the notebook and processed datasets will be saved into a data/ directory.

This code a rebuild of this repo https://github.com/kang205/SASRec

Paper

Link to paper

Teaser Video

How to Run

Install the requirements.

Run the notebook to download and process the datasets.

In the environment, run the experiments with the commands below.

Params

Param	Description	Example
--f-name	File name of interaction dataset	ml-1m
--shuffle-sequence	Whether to shuffle the input sequences	True or False
--batch-size	Batch size	1024
--lr	Learning Rate	0.001
--maxlen	Maximum length of sequences	200
--hidden-dim	Hidden Dimension	100
--num-blocks	Number of attention blocks	2
--num-epochs	Number of epochs	200
--num-heads	Number of heads	2
--dropout-rate	Dropout Rate	0.2
--seed	Random Seed	123
--output-metrics-path	Path to save results	'results/experiment_results.txt'

Shuffled / Unshuffled Experiments

python -m sas_rec.experiment --f-name 'ml_1m' --shuffle-sequence False --batch-size 128 --seed 101 --num-epochs 1 --output-metrics-path 'results/experiment_results.txt'

Rating Experiments

Modified Experiments

python -m sas_rec.modified_experiment --f-name 'ml_1m_rating' --shuffle-sequence False --batch-size 128 --seed 101 --num-epochs 1 --output-metrics-path 'results/experiment_results.txt'

Joint loss experiment on movielens

python -m sas_rec.joint_loss_experiment --f-name 'ml-1m-rating' --shuffle-sequence False --batch-size 128 --seed 101 --num-epochs 20 --output-metrics-path 'results/experiment_results.txt'

References

[1] SASRec https://arxiv.org/abs/1808.09781