FairDiff: Mitigating the Self-Reinforcing Matthew Effect in Diffusion Recommender Models

Song-Li Wu; Xianquan Wang; Zhaocheng Du; Weinan Gan; Jingyi Wang

cs.AISep 29, 2026

FairDiff: Mitigating the Self-Reinforcing Matthew Effect in Diffusion Recommender Models

Authors: Song-Li Wu, Xianquan Wang, Zhaocheng Du, Weinan Gan, Jingyi Wang

Organizations: Tsinghua University · University of Science and Technology of China · Huawei Noah’s Ark Lab

Abstract

While the "Matthew Effect" and filter bubbles are widely recognized outcome-level biases in recommender systems, we reveal that Diffusion Recommender Models (DRMs) uniquely compound this issue through their generative dynamics. Rather than merely inheriting data imbalances, DRMs trigger a self-reinforcing amplification of popularity bias. We identify that this phenomenon is driven by two compounding mechanisms. First, while optimization loss is universally dominated by high-frequency items across recommenders, DRMs suffer from a unique structural prior mismatch during generation. Because the forward terminal distribution of long-tailed data deviates significantly from the standard Gaussian prior, reverse sampling trajectories inherently collapse toward high-density popular items, fundamentally suppressing niche item generation. To dismantle this self-reinforcing loop, we propose FairDiff, a plug-and-play fairness-aware diffusion framework. To overcome the popularity-dominated loss, we introduce Popularity Condition Guidance (PCG). Rather than altering the training objective, PCG acts as an inference-time distributional reweighting mechanism, mathematically reshaping the score-based gradient field to penalize high-popularity regions and guide trajectories toward niche semantics. Furthermore, we design a Semantic Calibration (SC) Module to bridge the prior mismatch, aligning the forward and reverse distributions via one-step optimal transport. Comprehensive evaluations demonstrate that FairDiff achieves state-of-the-art performance while effectively mitigating the self-reinforcing Matthew Effect, highlighting its value as a general framework for DRMs.

Figures & tables

Figure 1: Comparison of DRMs and FairDiff.

Dataset	Metric	SASRec	BERT4Rec	TiMiRec	TIGER	BASRec	HSTU	ACVAE	CSRec	DreamRec	+FairDiff	DiffuRec	+FairDiff	CDiff4Rec	+FairDiff
Beauty	H@5	3.27	2.13	1.90	3.53	5.37	5.25	2.47	5.48	5.61	6.55 $\star$ $\star$	5.73	6.94 $\star$ $\star$	5.83	7.29 $\star$ $\star$
	H@10	6.26	3.72	3.34	6.23	7.03	6.29	3.88	7.08	7.13	7.99 $\star$ $\star$	7.21	7.96 $\star$ $\star$	7.71	8.69 $\star$ $\star$
	H@20	8.98	5.79	5.17	9.63	10.32	10.08	6.12	7.53	10.62	11.82 $\star$ $\star$	10.53	11.76 $\star$ $\star$	10.72	11.90 $\star$ $\star$
	N@5	2.40	1.32	1.24	2.59	3.29	3.04	1.71	3.28	3.49	3.92 $\star$ $\star$	3.64	4.13 $\star$ $\star$	3.90	4.39 $\star$ $\star$
	N@10	3.23	1.83	1.71	3.26	4.21	3.97	2.07	4.18	4.19	4.78 $\star$ $\star$	4.44	4.89 $\star$ $\star$	4.65	5.21 $\star$ $\star$
	N@20	3.66	2.35	2.17	3.71	4.95	4.82	2.63	5.08	4.99	5.85 $\star$ $\star$	5.32	5.93 $\star$ $\star$	5.41	6.01 $\star$ $\star$

Table 1: Performance comparison across datasets (metrics: HR@K (H@K) and NDCG@K (N@K), K

\in{5,10,20}

). The t-test results show that our performance advantage over the previous SOTA method is statistically significant. The improvement is statistically significant with

p<10^{-2}

(

\star

p<10^{-2}

\star\star

p<10^{-4}

Model	Recall $@$ 10 ( $\uparrow$ )	NDCG $@$ 10( $\uparrow$ )	$\Delta$ Recall ( $\downarrow$ )	$\Delta$ NDCG ( $\downarrow$ )	APLT ( $\uparrow$ )	$\Delta$ Exp ( $\downarrow$ )
SASRec	9.62	10.94	0.33	1.57	9.47	86.29
BERT4Rec	10.14	11.67	0.32	1.54	10.75	87.36
TiMiRec	10.34	12.82	0.16	1.37	13.73	84.98
TIGER	10.46	12.94	0.14	1.34	10.50	88.78
BASRec	6.45	12.76	0.35	1.66	9.49	90.54
HSTU	10.35	12.04	0.48	1.43	9.63	87.02

Table 2: Evaluation of the self-reinforcing Matthew effect. DRMs achieve strong utility (Recall@10 and NDCG@10), but exhibit substantial popularity amplification, reflected by large

\Delta

Recall,

\Delta

NDCG, and

\Delta

Exposure as well as limited long-tail coverage (APLT). By contrast, FairDiff significantly mitigates popularity bias while simultaneously improving utility, achieving a superior fairness–utility trade-off. All metrics are normalized to [0,1]; higher values indicate better performance (except for

\Delta

metrics, where lower is better).

Metric	Dim	Pop	SASRec	BERT4Rec	TiMiRec	TIGER	BASRec	HSTU	ACVAE	CSRec	DiffuRec	+FairDiff	DreamRec	+FairDiff	CDiff4Rec	+FairDiff
H@20	Item	Cold	2.85	2.12	3.67	3.92	5.12	5.18	2.56	5.32	2.51	5.59	2.74	5.93	2.87	6.45
		Mid	7.12	4.37	7.45	8.12	7.98	4.12	7.55	7.41	7.87	10.07	8.02	10.68	10.54	12.73
		Hot	9.42	10.23	9.05	12.75	12.08	13.90	12.32	12.94	16.09	17.35	17.21	18.58	18.08	19.14
	User	Cold	5.72	7.32	8.05	9.27	9.12	7.15	9.44	9.48	6.14	11.75	7.14	14.78	7.56	16.28
		Mid	8.96	8.05	9.37	10.48	10.58	10.32	8.06	10.43	8.83	11.24	9.36	14.62	9.85	14.83
		Hot	9.88	9.54	11.05	11.38	11.63	12.28	9.82	12.41	12.36	13.12	13.27	14.07	14.13	15.28

Table 3: Performance across item popularity and user sequence length. Dim and Pop denotes dimension and popularity.

Variant	Recall $@$ 10	NDCG $@$ 10	$\Delta$ Recall	$\Delta$ NDCG	APLT	$\Delta$ Exp
CDiff4Rec	11.29	14.13	0.58	2.49	9.33	91.23
+ PC ( $w{=}1.0$ )	12.01	14.72	0.46	2.12	10.84	86.41
+ PCG ( $w{=}0.5$ )	12.94	15.53	0.22	1.63	13.41	80.36
+ PCG ( $w{=}0.0$ )	12.48	15.02	0.11	1.29	14.02	78.91
+ SC (OT INI)	12.36	15.01	0.34	1.88	12.67	84.75
+ SC + TS	13.02	15.84	0.26	1.54	14.61	79.83

Table 4: Ablation study of FairDiff. PC denotes Popularity Conditioning, TS denotes Truncated Sampling, and INI denotes initialization.

Figure 2: Visualization of item distributions in the forward–reverse process. Green denotes the item distribution, while red denotes the reconstruction distribution. The first column shows the forward initial and reverse final distributions; the second column shows the forward final and reverse initial distributions.

Dataset	# Sequence	# items	# Actions	Avg len	Sparsity
Beauty	22,363	12,101	198,502	8.53	99.93%
Toys	19,412	11,924	167,597	8.63	99.93%
ML-1M	6,040	3,416	999,611	165.50	95.16%
Steam	281,428	13,044	3,485,022	12.40	99.90%

Table 5: Statistics of sequential recommendation datasets. Avg len means the average length of sequences.

Dataset	Modalities	# Users	# Items	# Interactions	Sparsity
TikTok	V, A, T	9,319	6,710	59,541	99.90%
Baby	V, T	19,445	7,050	139,110	99.90%
Sports	V, T	35,598	18,357	256,308	99.96%

Table 6: Statistics of cross-domain multimodal recommendation datasets. V, A, and T denote visual, acoustic, and textual modalities, respectively.

Backbone	Datasets	Steps	Epochs	Training Time	Inference Time
DreamRec	Beauty	350	248	369 minutes	344 seconds
	Toys	323	365	472 minutes	218 seconds
	Steam	5430	215	1850 minutes	2245 seconds
	ML-1M	2649	301	1627 minutes	180 seconds
+FairDiff	Beauty	145	109	163 minutes	149 seconds
	Toys	131	153	195 minutes	92 seconds

Table 7: Efficiency Experiments.

Datasets	Metrics	T-DiffRec	TI-DiffRec	GRU4Rec	+PDRec	+FairDiff	SASRec	+PDRec	+FairDiff	CL4SRec	+PDRec	+FairDiff
Toy	N@1	0.1033	0.1058	0.0878	0.0899	0.1137	0.1095	0.1247	0.1426	0.1125	0.1254	0.1575
	N@5	0.1564	0.1618	0.1515	0.1617	0.1749	0.1779	0.2023	0.2218	0.1802	0.2041	0.2471
	N@10	0.1758	0.1823	0.1755	0.1879	0.2114	0.2020	0.2286	0.2371	0.2046	0.2305	0.2415
	HR@5	0.2055	0.2151	0.2128	0.2300	0.2451	0.2423	0.2752	0.2913	0.2438	0.2776	0.3048
	HR@10	0.2657	0.2787	0.2874	0.3112	0.3473	0.3169	0.3568	0.3764	0.3195	0.3595	0.3864
	AUC	0.5911	0.5968	0.5670	0.5909	0.6127	0.5771	0.6060	0.6239	0.5805	0.6068	0.6398

Table 8: Temporal Modeling Results between backbones and FairDiff on four datasets.

Methods	Movie $\rightarrow$ Music						Book $\rightarrow$ Movie						Book $\rightarrow$ Music
Methods	20%		50%		80%		20%		50%		80%		20%		50%		80%
	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE	MAE	RMSE
TGT	4.4803	5.1580	4.4989	5.1736	4.5020	5.1891	4.1831	4.7536	4.2288	4.7920	4.2123	4.8149	4.4873	5.1672	4.5073	5.1727	4.5024	5.2308
CMF	1.5209	2.0158	1.6893	2.2271	2.4186	3.0936	1.3632	1.7918	1.5813	2.0886	2.1577	2.6777	1.8284	2.3829	2.1282	2.7275	3.0130	3.6948
DCDCSR	1.4918	1.9210	1.8144	2.3439	2.7194	3.3065	1.3971	1.7346	1.6731	2.0551	2.3618	2.7702	1.8411	2.2955	2.1736	2.6771	3.1405	3.5842
SSCDR	1.3017	1.6579	1.3762	1.7477	1.5046	1.9229	1.2390	1.6526	1.2137	1.5602	1.3172	1.7024	1.5414	1.9283	1.4739	1.8441	1.6414	2.1403

Table 9: Performance comparison of CDR to cold-start users in three CDR scenarios.

Dataset	Metric	Dimension	Popularity	SASRec	DIN	BERT4Rec	ComiRec	TiMiRec	TIGER	BASRec	HSTU	ACVAE	CSRec	DiffuRec	+FairDiff	DreamRec	+FairDiff	CDiff4Rec	+FairDiff
Steam	H@20	Item	Cold	0.32	0.41	0.25	0.21	0.38	0.45	0.52	0.50	0.36	0.55	0.61	0.82	0.78	0.95	0.85	1.02
			Mid	4.15	4.82	3.97	3.25	4.68	4.95	5.12	4.98	4.72	5.36	5.21	5.45	5.02	5.26	5.38	5.62
			Hot	15.28	16.95	14.87	13.65	16.25	15.82	16.53	16.10	15.76	17.23	17.85	19.12	19.08	20.23	18.95	21.05
		User	Cold	6.25	7.18	6.02	5.38	6.95	7.52	7.88	7.65	7.02	8.15	7.95	16.82	17.49	18.56	8.52	19.10
			Mid	12.35	13.18	11.97	10.85	12.88	13.52	13.85	13.62	12.95	14.23	14.58	16.95	17.36	17.52	15.10	18.02
			Hot	14.82	15.95	14.56	13.98	15.78	16.25	16.68	16.35	15.89	17.52	17.95	18.88	16.27	17.35	17.23	19.05

Table 10: Performance across item popularity and user sequence length on sequential recommendation datasets.

Dataset	Metric	BPR	LightGCN	VBPR	MMGCN	DualGNN	SLMRec	BM3	MGCN	Freedom	DiffMM	+FairDiff	Imp.	DiffCL	+FairDiff	Imp.
TikTok	H@5	0.013	0.022	0.017	0.025	0.027	0.028	0.031	0.034	0.035	0.037	0.053	+43.24%	0.043	0.058	+34.9%
	H@10	0.022	0.031	0.028	0.038	0.043	0.044	0.049	0.054	0.056	0.068	0.099	+45.6%	0.075	0.108	+44.0%
	H@20	0.040	0.052	0.049	0.062	0.068	0.071	0.078	0.085	0.089	0.110	0.128	+16.4%	0.122	0.139	+13.9%
	N@5	0.007	0.014	0.010	0.016	0.017	0.017	0.019	0.021	0.022	0.025	0.041	+64.0%	0.026	0.048	+84.6%
	N@10	0.012	0.020	0.015	0.022	0.023	0.024	0.026	0.028	0.030	0.032	0.044	+37.5%	0.037	0.055	+48.6%
	N@20	0.019	0.028	0.022	0.031	0.033	0.034	0.037	0.040	0.042	0.045	0.052	+15.6%	0.049	0.067	+36.7%

Table 11: Multimodal Recommendation Results.

Setting	Algorithm	N@1	N@5	N@10	N@20	N@50	HR@5	HR@10	HR@20	HR@50	AUC
Game $\rightarrow$ Toy	T-DiffRec (M)	0.0981	0.1520	0.1727	0.1934	0.2375	0.2029	0.2673	0.3494	0.5780	0.5924
	TI-DiffRec (M)	0.1053	0.1598	0.1806	0.2008	0.2407	0.2111	0.2759	0.3562	0.5623	0.5932
	SASRec (M)	0.1267	0.2019	0.2261	0.2490	0.2785	0.2722	0.3472	0.4380	0.5873	0.5951
	PDRec(M)	0.1302	0.2093	0.2348	0.2574	0.2873	0.2826	0.3616	0.4515	0.6026	0.6106
	+PCG	0.1395	0.2137	0.2497	0.2674	0.2935	0.3048	0.3784	0.4672	0.6173	0.6392
	+SC	0.1382	0.2125	0.2489	0.2682	0.2931	0.3045	0.3777	0.4668	0.6160	0.6386

Table 12: Cross-domain Recommendation Results.

Explore similar work

Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders
May 16, 2026Jun Yin, Bangguo Zhu, Peng Huo +5Sequential Recommendation Recommendation Quality
School of Computer Science and Engineering, Central South University, Changsha, China · National Super Computing Center, Tianjin, China · Faculty of Data Science, City University of Macau, Macau, China +1
The Pitfall of Scaling Up: Uncovering and Mitigating Popularity Bias Amplification in Scaling Transformer-based Recommenders
Jun 20, 2026Weiqin Yang, Yue Pan, Chongming Gao +4Multimodal Recommendation Model Recommendation
Zhejiang University · State Key Laboratory of Blockchain and Data Security, Zhejiang University · College of Computer Science and Technology, Zhejiang University +2
Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation
Jul 21, 2026Yongsen Zheng, Ruilin Xu, Guohua Wang +2User Modeling Recommendation
Nanyang Technological University, Singapore · Digital Trust Centre Singapore · Sun Yat-sen University +3

May 16, 2026, cs.IRJ/K move · Enter open · S save

Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders

Jun Yin, Bangguo Zhu, Peng Huo, Ruochen Liu+4

School of Computer Science and Engineering, Central South University, Changsha, China · National Super Computing Center, Tianjin, China · Faculty of Data Science, City University of Macau, Macau, China+1

Sequential Recommendation Recommendation Quality Token-Level Uncertainty

Open paper AskPDF arXiv

Recently, Generative Recommenders (GRs), characterized by a unified end-to-end framework, have exhibited astonishing potential in transforming the recommendation paradigm. Despite their effectiveness, we recognize that GRs are still susceptible to the long-standing issue of popularity bias that has pervaded the recommendation community. Although a few studies have attempted to extend traditional debiasing methods to GRs, their effectiveness is marginal, and the fundamental reason why GRs suffer from popularity bias remains under-explored. To bridge this gap, this study focuses on two core aspects in GRs: the optimization of generative framework and the item tokenization based on semantic index. Based on theoretical analyses, we identify that the severe popularity bias emerges from the confluence of a token-level optimization flaw and the undifferentiated property of item tokenization. Accordingly, this study develops a novel generative recommender system, called Ghost, by designing the asymmetric unlikelihood optimization and the skeleton-founded tokenization. Extensive empirical evaluations across three datasets, alongside multiple SOTA baselines, reveal that Ghost substantially alleviates popularity bias and promotes fairer recommendations, while incurring slight degradation to the overall recommendation utility.

Abstract

Explore similar work

Echoes in Filter Bubble: Diagnosing and Curing Popularity Bias in Generative Recommenders

The Pitfall of Scaling Up: Uncovering and Mitigating Popularity Bias Amplification in Scaling Transformer-based Recommenders

Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation