Modern recommender systems trained on domain-specific data often struggle to generalize across multiple domains. Cross-domain sequential recommendation has emerged as a promising research direction to address this challenge; however, existing approaches face fundamental limitations, such as reliance on overlapping users or items across domains, or unrealistic assumptions that ignore privacy constraints. In this work, we propose a new framework, MergeRec, based on model merging under a new and realistic problem setting termed data-isolated cross-domain sequential recommendation, where raw user interaction data cannot be shared across domains. MergeRec consists of three key components: (1) merging initialization, (2) pseudo-user data construction, and (3) collaborative merging optimization. First, we initialize a merged model using training-free merging techniques. Next, we construct pseudo-user data by treating each item as a virtual sequence in each domain, enabling the synthesis of meaningful training samples without relying on real user interactions. Finally, we optimize domain-specific merging weights through a joint objective that combines a recommendation loss, which encourages the merged model to identify relevant items, and a distillation loss, which transfers collaborative filtering signals from the fine-tuned source models. Extensive experiments demonstrate that MergeRec not only preserves the strengths of the original models but also significantly enhances generalizability to unseen domains. Compared to conventional model merging methods, MergeRec consistently achieves superior performance, with average improvements of up to 17.21% in Recall@10, highlighting the potential of model merging as a scalable and effective approach for building universal recommender systems. The source code is available at github.com/DIALLab-SKKU/MergeRec.

Sequential recommendation (SR) aims to predict users' subsequent interactions by modeling their sequential behaviors. Recent studies have explored frequency domain analysis, which effectively models periodic patterns in user sequences. However, existing frequency-domain SR models still face two major drawbacks: (i) limited frequency band coverage, often missing critical behavioral patterns in a specific frequency range, and (ii) lack of personalized frequency filtering, as they apply an identical filter for all users regardless of their distinct frequency characteristics. To address these challenges, we propose a novel frequency-domain model, Mixture of User-adaptive Frequency FIlteriNg (MUFFIN ), operating through two complementary modules. (i) The global filtering module (GFM) handles the entire frequency spectrum to capture comprehensive behavioral patterns. (ii) The local filtering module (LFM) selectively emphasizes important frequency bands without excluding information from other ranges. (iii) In both modules, the user-adaptive filter (UAF) is adopted to generate user-specific frequency filters tailored to individual unique characteristics. Finally, by aggregating both modules, MUFFIN captures diverse user behavioral patterns across the full frequency spectrum. Extensive experiments show that MUFFIN consistently outperforms state-of-the-art frequency-domain SR models over five benchmark datasets. The source code is available at https://github.com/ilwoong100/MUFFIN.

Large language models (LLMs) have been widely adopted to enrich the semantic representation of textual item information in recommender systems. However, existing linear autoencoders (LAEs) that incorporate textual information rely on sparse word co-occurrence patterns, limiting their ability to capture rich textual semantics. To address this, we propose L3AE, the first integration of LLMs into the LAE framework. L3AE effectively integrates the heterogeneous knowledge of textual semantics and user-item interactions through a two-phase optimization strategy. (i) L3AE first constructs a semantic item-to-item correlation matrix from LLM-derived item representations. (ii) It then learns an item-to-item weight matrix from collaborative signals while distilling semantic item correlations as regularization. Notably, each phase of L3AE is optimized through closed-form solutions, ensuring global optimality and computational efficiency. Extensive experiments demonstrate that L3AE consistently outperforms state-of-the-art LLM-enhanced models on three benchmark datasets, achieving gains of 27.6% in Recall@20 and 39.3% in NDCG@20. The source code is available at https://github.com/jaewan7599/L3AE_CIKM2025.

Heejin Kook, Junyoung Kim, Seongmin Park, Jongwuk Lee. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Session-based recommendation (SBR) aims to predict users' subsequent actions by modeling short-term interactions within sessions. Existing neural models primarily focus on capturing complex dependencies for sequential item transitions. As an alternative solution, linear item-item models mainly identify strong co-occurrence patterns across items and support faster inference speed. Although each paradigm has been actively studied in SBR, their fundamental differences in capturing item relationships and how to bridge these distinct modeling paradigms effectively remain unexplored. In this paper, we propose a novel SBR model, namely Linear Item-Item model with Neural Knowledge (LINK), which integrates both types of knowledge into a unified linear framework. Specifically, we design two specialized components of LINK: (i) Linear knowledge-enhanced Item-item Similarity model (LIS), which refines the item similarity correlation via self-distillation, and (ii) Neural knowledge-enhanced Item-item Transition model (NIT), which seamlessly incorporates complicated neural knowledge distilled from the off-the-shelf neural model. Extensive experiments demonstrate that LINK outperforms state-of-the-art linear SBR models across six real-world datasets, achieving improvements of up to 14.78% and 11.04% in Recall@20 and MRR@20 while showing up to 813x fewer inference FLOPs. Our code is available at https://github.com/jin530/LINK.

Despite their simplicity, linear autoencoder (LAE)-based models have shown comparable or even better performance with faster inference speed than neural recommender models. However, LAEs face two critical challenges: (i) popularity bias, which tends to recommend popular items, and (ii) neighborhood bias, which overly focuses on capturing local item correlations. To address these issues, this paper first analyzes the effects of two existing normalization methods for LAEs, i.e., random-walk and symmetric normalization. Our theoretical analysis reveals that normalization highly affects the degree of popularity and neighborhood biases among items. Inspired by this analysis, we propose a versatile normalization solution, called Data-Adaptive Normalization (DAN), which flexibly controls the popularity and neighborhood biases by adjusting item- and user-side normalization to align with unique dataset characteristics. Owing to its model-agnostic property, DAN can be easily applied to various LAE-based models. Experimental results show that DAN-equipped LAEs consistently improve existing LAE-based models across six benchmark datasets, with significant gains of up to 128.57% and 12.36% for long-tail items and unbiased evaluations, respectively. Refer to our code in https://github.com/psm1206/DAN.

In sequential recommendation (SR), neural models have been actively explored due to their remarkable performance, but they suffer from inefficiency inherent to their complexity. Linear SR models exhibit high efficiency and achieve competitive or superior accuracy compared to neural models. However, they solely deal with the sequential order of items (i.e., sequential information) and overlook the actual timestamp (i.e., temporal information). It is limited to effectively capturing various user preference drifts over time. To address this issue, we propose a novel linear SR model, named TemporAl LinEar item-item model (TALE), incorporating temporal information while preserving training/inference efficiency. It consists of three key components. (i) Single-target augmentation concentrates on a single target item, enabling us to learn the temporal correlation for the target item. (ii) Time interval-aware weighting utilizes the actual timestamp to discern the item correlation depending on time intervals. (iii) Trend-aware normalization reflects the dynamic shift of item popularity over time. Our empirical studies show that TALE outperforms ten competing SR models by up to 18.71% gains across five benchmark datasets. It also exhibits remarkable effectiveness for evaluating long-tail items by up to 30.45% gains. The source code is available at https://github.com/psm1206/TALE.

Collaborative filtering (CF) is a pivotal technique in modern recommender systems. The learning process of CF models typically consists of three components: interaction encoder, loss function, and negative sampling. Although many existing studies have proposed various CF models to design sophisticated interaction encoders, recent work shows that simply reformulating the loss functions can achieve significant performance gains. This paper delves into analyzing the relationship among existing loss functions. Our mathematical analysis reveals that the previous loss functions can be interpreted as alignment and uniformity functions: (i) the alignment matches user and item representations, and (ii) the uniformity disperses user and item distributions. Inspired by this analysis, we propose a novel loss function that improves the design of alignment and uniformity considering the unique patterns of datasets called Margin-aware Alignment and Weighted Uniformity (MAWU). The key novelty of MAWU is two-fold: (i) margin-aware alignment (MA) mitigates user/item-specific popularity biases, and (ii) weighted uniformity (WU) adjusts the significance between user and item uniformities to reflect the inherent characteristics of datasets. Extensive experimental results show that MF and LightGCN equipped with MAWU are comparable or superior to state-of-the-art CF models with various loss functions on three public datasets.

Because implicit user feedback for the collaborative filtering (CF) models is biased toward popular items, CF models tend to yield recommendation lists with popularity bias. Previous studies have utilized inverse propensity weighting (IPW) or causal inference to mitigate this problem. However, they solely employ pointwise or pairwise loss functions and neglect to adopt a contrastive loss function for learning meaningful user and item representations. In this paper, we propose Unbiased ConTrastive Representation Learning (uCTRL), optimizing alignment and uniformity functions derived from the InfoNCE loss function for CF models. Specifically, we formulate an unbiased alignment function used in uCTRL. We also devise a novel IPW estimation method that removes the bias of both users and items. Despite its simplicity, uCTRL equipped with existing CF models consistently outperforms state-of-the-art unbiased recommender models, up to 12.22% for Recall@20 and 16.33% for NDCG@20 gains, on four benchmark datasets.

Implicit feedback has been widely used to build commercial recommender systems. Because observed feedback represents users' click logs, there is a semantic gap between true relevance and observed feedback. More importantly, observed feedback is usually biased towards popular items, thereby overestimating the actual relevance of popular items. Although existing studies have developed unbiased learning methods using inverse propensity weighting (IPW) or causal reasoning, they solely focus on eliminating the popularity bias of items. In this paper, we propose a novel unbiased recommender learning model, namely BIlateral SElf-unbiased Recommender (BISER), to eliminate the exposure bias of items caused by recommender models. Specifically, BISER consists of two key components: (i) self-inverse propensity weighting (SIPW) to gradually mitigate the bias of items without incurring high computational costs; and (ii) bilateral unbiased learning (BU) to bridge the gap between two complementary models in model predictions, i.e., user- and item-based autoencoders, alleviating the high variance of SIPW. Extensive experiments show that BISER consistently outperforms state-of-the-art unbiased recommender models over several datasets, including Coat, Yahoo! R3, MovieLens, and CiteULike.

Unbiased recommender learning has been actively studied to alleviate the inherent bias of implicit datasets under the missing-not-at-random assumption. Existing studies solely address the bias of positive feedback but do not account for the bias of missing feedback, which heavily affects their sub-optimal performance gains. This paper proposes a dual recommender learning framework that simultaneously eliminates the bias of clicked and unclicked data. Specifically, the proposed loss function adopts two propensity weighting to effectively estimate the true positive and negative preferences from clicked and unclicked data. We also prove that the proposed loss function converges to the ideal loss function for both clicked and unclicked data. Because of the model-agnostic property, it can be applied to any existing unbiased learning models. Experimental results show that the proposed method outperforms state-of-the-art unbiased models up to 5.54-24.56% for [email protected] on three datasets.