Debiasing Sequential Recommendation with Time-aware Inverse Propensity Scoring

Sirui Huang; Jing Long; Qian Li; Guandong Xu; Qing Li

cs.IRMar 5, 2026

Debiasing Sequential Recommendation with Time-aware Inverse Propensity Scoring

Authors: Sirui Huang, Jing Long, Qian Li, Guandong Xu, Qing Li

Organizations: Hong Kong Polytechinic University Hong Kong SAR, China · University of Technology Sydney Sydney, Australia · Curtin University Perth, Australia · Education University of Hong Kong Hong Kong SAR, China

Abstract

Sequential Recommendation (SR) predicts users next interactions by modeling the temporal order of their historical behaviors. Existing approaches, including traditional sequential models and generative recommenders, achieve strong performance but primarily rely on explicit interactions such as clicks or purchases while overlooking item exposures. This ignorance introduces selection bias, where exposed but unclicked items are misinterpreted as disinterest, and exposure bias, where unexposed items are treated as irrelevant. Effectively addressing these biases requires distinguishing between items that were "not exposed" and those that were "not of interest", which cannot be reliably inferred from correlations in historical data. Counterfactual reasoning provides a natural solution by estimating user preferences under hypothetical exposure, and Inverse Propensity Scoring (IPS) is a common tool for such estimation. However, conventional IPS methods are static and fail to capture the sequential dependencies and temporal dynamics of user behavior. To overcome these limitations, we propose Time aware Inverse Propensity Scoring (TIPS). Unlike traditional static IPS, TIPS effectively accounts for sequential dependencies and temporal dynamics, thereby capturing user preferences more accurately. Extensive experiments show that TIPS consistently enhances recommendation performance as a plug-in for various sequential recommenders. Our code will be publicly available upon acceptance.

Figures & tables

Figure 1. (a) An example of a biased SR model that learns user preference solely from interactions. (b) Example of a system that considers exposure. Here, $\mathbf{C}_{t}$ denotes user interactions ( e.g., clicks) at time $t$ , $\mathbf{U}_{t}$ represents the user’s preference at time $t$ , and $\mathbf{E}_{t}$ is the recommendation list containing all items exposed to the user at time $t$ .

Figure 2. Structural Causal Model (SCM) of SR.

Figure 3. Overview of the Time-aware Inverse Propensity Scoring (TIPS) framework, designed as a plug-in for any SR model.

Dataset	#User	#Item	#Interaction	Time Period(mm/dd/yyyy)
ML-1M	5950	3532	574619	04/25/2000-02/28/2003 (34.2 months)
ML-10M	66028	10254	4980475	01/09/1995-01/05/2009 (168 months)
Music4All	14124	99596	5109577	12/30/2013-03/26/2019 (62.8 months)
GoodReads	17660	29972	409600	02/13/2001-11/02/2017 (200.3 months)

Table 1. The statistics of datasets.

Model	ML-1M				Music4All				GoodReads				ML-10M
Model	HR@5	HR@10	NDCG@5	NDCG@10	HR@5	HR@10	NDCG@5	NDCG@10	HR@5	HR@10	NDCG@5	NDCG@10	HR@5	HR@10	NDCG@5	NDCG@10
Traditional Sequential Models
SASRec	0.3351	0.4501	0.2249	0.2836	0.5581	0.6469	0.4119	0.4473	0.4814	0.5571	0.3622	0.3877	0.4755	0.6319	0.3164	0.3704
TiSASRec	0.3525	0.4819	0.2387	0.2968	0.5808	0.6924	0.4429	0.4772	0.4981	0.6017	0.3812	0.4113	0.4987	0.6721	0.3375	0.3924
GRU (Base)	0.3467	0.4731	0.2482	0.2895	0.5460	0.6253	0.3983	0.4336	0.4607	0.5372	0.3501	0.3732	0.4593	0.6027	0.3061	0.3538
+USR	0.3728	0.5123	0.2658	0.3119	0.5762	0.6501	0.4173	0.4581	0.4763	0.5567	0.3627	0.3862	0.4941	0.6502	0.3260	0.3807

Table 2. Comparison of recommendation performance between HyperG and two categories of sequential models.

Backbone	Variant	ML-1M
Backbone	Variant	HR@10		NDCG@10
Attention	HyperG (Ours)	0.5558		0.3418
	HyperG $\neg\text{time}$	0.5550	$\downarrow$ 0.08%	0.3472	$\uparrow$ 0.54%
	HyperG $\neg\text{IPS}$	0.5356	$\downarrow$ 2.02%	0.3238	$\downarrow$ 1.80%
	HyperG $\neg\text{EP\&time}$	0.5311	$\downarrow$ 2.47%	0.2874	$\downarrow$ 5.44%
CVAE	HyperG (Ours)	0.4954		0.3366

Table 3. Ablation studies of HyperG on the ML-1M.

Figure 4. (a)(b) Results for different values of hyperparameters $\mu$ and $\gamma$ , and (c)(d) distribution of the difference between the average propensity score of positive items and that of negative items, using HyperG (Ours) and traditional IPS (Base).

Exposure Bias

See all 10 papers

Generative Archetype-Grounded Item Representations for Sequential Recommendation
Jun 9, 2026Yifan Li, Jiahong Liu, Xinni Zhang +5Sequential Recommendation Representation Learning
The Chinese University of Hong Kong Hong Kong SAR, China · McGill University Montreal, Canada · Tongji University Shanghai, China
Retrieve-then-Adapt: Retrieval-Augmented Test-Time Adaptation for Sequential Recommendation
Apr 7, 2026Xing Tang, Ziqiang Cui, Jingyang Bin +6Sequential Recommendation Test-Time Adaptation
School of Artificial Intelligence, Shenzhen Technology University, Shenzhen, China · City University of Hong Kong, Hong Kong, Hong Kong SAR · Shenzhen Technology University, Shenzhen, China +2
Rethinking Convolutional Networks for Attribute-Aware Sequential Recommendation
May 6, 2026Shereen Elsayed, Ngoc Son Le, Ahmed Rashed +1Sequential Recommendation Convolutional Neural Networks
Information Systems and Machine Learning Lab (ISMLL), University of Hildesheim, Germany · VWFS Data Analytics Research Center (VWFS DARC), University of Hildesheim, Germany · Volkswagen Financial Services AG, Germany

Jun 9, 2026J/K move · Enter open · S save

Generative Archetype-Grounded Item Representations for Sequential Recommendation

Yifan Li, Jiahong Liu, Xinni Zhang, Hao Chen+4

The Chinese University of Hong Kong Hong Kong SAR, China · McGill University Montreal, Canada · Tongji University Shanghai, China

Sequential Recommendation Representation Learning Large Language Model-Based Recommendation User Preference Modeling

Open paper AskPDF arXiv

Sequential recommendation aims to predict users' next interaction with items by analyzing their historical behavior. However, the limited quality of item representations remains a critical bottleneck. While pre-trained large language models (LLMs) can provide rich semantic representations, existing approaches only rely on static encoding of fixed attributes, overlooking the crucial role of target audiences in defining item identity. Moreover, the semantic space struggles to reflect actual user behavior, resulting in a significant gap between semantic representations and behavioral patterns. To address these limitations, we propose GenAIR, a general framework that empowers sequential recommendation with Generative Archetype-grounded Item Representations. Specifically, we first leverage an LLM to analyze item metadata and infer textual description of the Archetype, which represents the conceptual profile of the item's ideal target audience. We then extract the corresponding embeddings in a single forward pass. Further, to ground these generative archetypes in real-world behavior, we introduce a behavioral calibration objective, which explicitly incorporates behavioral signals from actual interactions. This objective adjusts the structure of the embedding space to reflect empirical patterns. GenAIR enables seamless integration with most existing models while maintaining high efficiency. Comprehensive experiments conducted on three real-world datasets demonstrate that GenAIR significantly improves the performance of various sequential recommendation models and consistently outperforms state-of-the-art baseline approaches. Implementation codes are available at https://github.com/AI-Santiago/GenAIR.

Abstract

Exposure Bias

Explore similar work

Generative Archetype-Grounded Item Representations for Sequential Recommendation

Retrieve-then-Adapt: Retrieval-Augmented Test-Time Adaptation for Sequential Recommendation

Rethinking Convolutional Networks for Attribute-Aware Sequential Recommendation