Promotion of Semantic Item Tail via Adversarial Learning ( KDD-EAI 2024 )

Introduction

This is the implementation of Paper "Promotion of Semantic Item Tail via Adversarial Learning" published in KDD-EAI 2024.

In this work, we focus on the semantic tail of recommendation model, which is the tail in recommendation system with semantic similarity. Promoting this tail not only improves recommendation diversity but also improves general metrics. To this aim, we devise an adversarial judge to dynamically locate this tail and use adversarial learning to keep tracking it.

Contributions

1. We propose a novel method to derive item-level advantage consistently from the user-level metrics.
2. We devised a learning-based approach to calibrate disadvantaged items without manually designed criteria.
3. We theoretically show that with a Lipschitz-bounded adversary as a judge, our calibration forms semantic tails, which consist of semantically similar but disadvantaged items.
4. We further demonstrate that optimizing on the semantic tails not only reduces popularity bias, but also improves overall performance. It achieves a win-win situation unmatched by other baselines.

Get the preprocessed data

We follow the setting of implicit feedback recommendation setting. This setting is closely related to the industrial recommendation setting. We include the dataset MovieLens, Netflix Prize and Million Songs in the evaluation. We prepare them into implicit feedback compatible dataset. To download our pre-processed data, please use this link and decompress it into data folder.

How to Use

In the common datasets, we provide a one-stop shop for all the methods including Ours, IPW, EASE and Tail Optimization and all datasets including Movielens20, Netflix Prize and MSD. The main body is written in train_body.py.

To use it, simply run train_body.py. In main function, you could specify the concrete parameters to reproduce the result.

The parameters required to reproduce our results.

Dataset	Params
MovieLens	ml20m_ease_lr2.00e+00_alr1.00e+00_wd8.00e-06_nstd1.50e+00_awd0.00e+00_ast2bmlp2_alossrecall_dist_onesFalse_lam1.00e+00_l2_itemarl_metric
Netflix Prize	nflx_ease_lr1.00e+00_alr1.00e+00_wd5.00e-05_nstd1.00e+00_awd0.00e+00_ast2mlp50_alossrecall_dist_onesFalse_lam1.00e+00_l2_itemarl_metric
Million Song	msd_ease_lr4.00e+01_alr1.00e-01_wd1.00e-05_nstd5.00e-01_awd1.00e-05_ast2mlp2_alossrecall_onesTrue_lam1.00e+00_l2_itemarl_metric

You can fill these parameters in train_body.py or jupyter notebook and invoke train_body function to train.

Dependency

The program will auto run on GPU if the device is avaiable and on CPU otherwise. It requires pytorch. We modified the python package recsys_metrics https://github.com/zuoxingdong/recsys_metrics for evaluation.

Citation

If you find this project useful, please cite us

@misc{xu2023posit,
      title={POSIT: Promotion of Semantic Item Tail via Adversarial Learning}, 
      author={Qiuling Xu and Pannaga Shivaswamy and Xiangyu Zhang},
      year={2023},
      eprint={2308.03366},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}

Result

Dataset	Method	Coverage@k Top 100	Coverage@k Top 50	Coverage@k Top 20	NDCG@K Top 100	Recall@K Top 100	Recall@K Top 50	Recall@K Top 20	Item Recall@k Top 100	Item Recall@k Top 50	Item Recall@k Top 20
Movie Lens	POSIT	1793	1052	530	0.4214	0.6369	0.5244	0.3928	0.0465	0.0294	0.0146
Movie Lens	IPW	1531	972	519	0.4209	0.6364	0.5226	0.3926	0.0433	0.0270	0.0131
Movie Lens	CVar	1488	940	502	0.4200	0.6353	0.5210	0.3911	0.0426	0.0262	0.0127
Movie Lens	Rerank	1718	948	490	0.4190	0.6300	0.5199	0.3906	0.0419	0.0257	0.0123
Movie Lens	EASE	1461	922	489	0.4199	0.6356	0.5209	0.3906	0.0419	0.0257	0.0123
Movie Lens	Causal	1084	417	259	0.3695	0.5599	0.4560	0.3411	0.0143	0.0083	0.0041
Movie Lens	Post	1586	922	489	0.3927	0.5525	0.4448	0.4118	0.0419	0.0257	0.0123
Movie Lens	MP	100	50	20	0.1905	0.3300	0.2351	0.1617	0.0067	0.0035	0.0016
Standard Deviation		±10	±5	±2	±0.0009	±0.0009	±0.0009	±0.0009	±0.0004	±0.0004	±0.0004
Netflix Prize	POSIT	1968	1256	694	0.3953	0.5565	0.4472	0.3633	0.0958	0.0631	0.0342
Netflix Prize	IPW	1802	1179	670	0.3942	0.5560	0.4464	0.3625	0.0833	0.0526	0.0269
Netflix Prize	CVar	1795	1177	667	0.3933	0.5537	0.4449	0.3617	0.0842	0.0532	0.0271
Netflix Prize	Rerank	1950	1190	655	0.3919	0.5478	0.4435	0.3612	0.0802	0.0502	0.0252
Netflix Prize	EASE	1762	1153	654	0.3930	0.5541	0.4448	0.3613	0.0804	0.0502	0.0252
Netflix Prize	Causal	1394	962	564	0.3551	0.5034	0.4024	0.3233	0.0251	0.0468	0.0249
Netflix Prize	Post	1759	1155	654	0.3927	0.5525	0.4448	0.3612	0.0777	0.0468	0.0249
Netflix Prize	MP	100	50	20	0.1587	0.2743	0.1749	0.1161	0.0083	0.0044	0.0019
Standard Deviation		±6	±4	±2	±0.0009	±0.0009	±0.0009	±0.0009	±0.0004	±0.0004	±0.0004
Million Song	POSIT	6569	3692	1625	0.3915	0.5116	0.4290	0.3341	0.2479	0.2030	0.1433
Million Song	IPW	6466	3673	1626	0.3909	0.5099	0.4290	0.3340	0.2464	0.2018	0.1419
Million Song	CVar	6337	3556	1567	0.3839	0.5070	0.4240	0.3268	0.2401	0.1917	0.1314
Million Song	Rerank	6368	3624	1609	0.3896	0.5062	0.4278	0.3339	0.2444	0.2007	0.1412
Million Song	EASE	6304	3623	1609	0.3898	0.5084	0.4279	0.3339	0.2445	0.2007	0.1412
Million Song	Causal	-	-	-	-	-	-	-	-	-	-
Million Song	Post	6342	3632	1611	0.3899	0.5072	0.4279	0.3340	0.2453	0.2013	0.1416
Million Song	MP	100	50	20	0.0582	0.0986	0.0680	0.0427	0.0027	0.0018	0.0010
Standard Deviation		±9	±5	±2	±0.0009	±0.0009	±0.0009	±0.0009	±0.0004	±0.0004	±0.0004