Enterprise Organization of National Program of Excellence in Software

The Data Intelligence and Learning (DIAL, Professor: Jong-Wook Lee) lab has had three papers accepted for publication at SIGIR 2025, the world's most prestigious information retrieval conference, and will present them in Padua, Italy in July.

[List of Papers]

1. Why is Normalization Necessary for Linear Recommenders? (SIGIR'25)

2. Linear Item-Item Models with Neural Knowledge for Session-based Recommendation (SIGIR'25)

3. DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation (SIGIR'25)

Study 1: Seongmin Park, Mincheol Yoon, Hye-young Kim, Jongwuk Lee, “Why is Normalization Necessary for Linear Recommenders?”, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2025

This study focuses on the fact that linear autoencoder (LAE)-based recommendation systems have comparable recommendation performance and fast inference speed to neural network-based models despite their simple structure. However, LAE faces two structural limitations: popularity bias, which over-recommends popular items, and neighborhood bias, which relies excessively on local correlations between items. To address these issues, in this paper, we propose a novel normalization method, Data-Adaptive Normalization (DAN), that can be applied to the LAE model.

DAN is a normalization technique designed to flexibly control two biases depending on the characteristics of the data, and consists of two core components: (i) item-adaptive normalization and (ii) user-adaptive normalization. First, item-adaptive normalization controls the influence of popular items through the regularization parameter α and provides a denoising effect to LAE. This allows LAE to significantly improve the recommendation performance for unpopular items (tail items) by moving away from the performance centered on popular items (i.e., popularity bias) that the existing LAE mainly recommends. Second, user-adaptive normalization controls the neighborhood bias using the parameter β. This technique suppresses high-frequency components and preserves important low-frequency components, thereby helping to better reflect the overall global pattern rather than local correlations.

The effectiveness of DAN is experimentally verified on six representative recommendation datasets (ML-20M, Netflix, MSD, Gowalla, Yelp2018, Amazon-book). LAE models (LAE_DAN, EASE_DAN, RLAE_DAN) applied with DAN showed consistent performance improvement over the existing LAE model on all datasets, and recorded performance improvements of up to 128.57% and 12.36% in tail items and unbiased evaluation, respectively. DAN also showed superior performance compared to the state-of-the-art collaborative filtering models.

In addition, Figure 1 (Case study) shows the recommendation results for a specific user according to the regularization method, and the following observations were made:

(1) LAE without regularization (W/O) recommends only five highly popular action movies even though the user watched three romantic movies. On the other hand, the three regularization methods (RW, Sym, DAN) effectively reflect user preferences by recommending “Step Up 2” as the top item related to “Step Up 1” viewed by the user.

(2) DAN provides the most balanced recommendation that maintains user preferences while appropriately mitigating popularity bias. RW regularization still has strong popularity bias with 4 out of 5 items being popular. Sym regularization overly mitigates popularity bias with 4 out of 5 items being unpopular. DAN recommends the items most relevant to user preferences while balancing popular and unpopular items.

Figure 1: Interaction history of user #91935 on the ML-20M dataset and the top-5 recommendation list of the four regularization methods. The red border is the head (top 20%) item, and the blue border is the tail (bottom 80%) item.

Furthermore, this study analyzes how the effect of the regularization parameters (α, β) varies depending on the Gini index and homophily characteristics of the dataset, and also presents guidelines for setting parameters suitable for each dataset. Through this, it is shown that the proposed DAN technique can be established as a general and practical solution that can precisely control bias depending on the data characteristics. For more information about this paper, please refer to the following address. https://dial.skku.edu/blog/2025_dan

Study 2: Minjin Choi, Sunkyung Lee, Seongmin Park, Jongwuk Lee, “Linear Item-Item Models with Neural Knowledge for Session-based Recommendation”, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2025

This study focuses on the problem of session-based recommendation (SBR), which predicts the next action based on the user’s current session interaction. The SBR field is largely divided into two paradigms. One is a neural network-based model that is strong in capturing complex sequential transition patterns, and the other is a linear item-item model that effectively learns co-occurrence patterns between items and provides fast inference speed. However, each paradigm is specialized in capturing different types of item relationships, and an effective integrated method to simultaneously achieve high accuracy of neural network models and efficiency of linear models is still lacking.

Against this backdrop, in this paper, we propose a novel SBR model, LINK (Linear Item-Item model with Neural Knowledge), which effectively integrates knowledge from linear models and neural network models. LINK aims to achieve both high accuracy and fast inference speed by combining two types of knowledge within a single unified linear framework. To this end, LINK includes two core components. (i) LIS (Linear knowledge-enhanced Item-item Similarity model) enhances the linear model’s ability to capture item similarity (co-occurrence) and learns high-dimensional relationships between sessions through self-distillation technique. (ii) NIT (Neural knowledge-enhanced Item-item Transition model) effectively injects neural network knowledge into linear models through a unique method that distills complex sequential transfer knowledge from pre-trained arbitrary neural network models and utilizes it as a regularization term when learning linear models. As shown in Figure 2, the effectiveness of the LINK model has been verified through extensive experiments using six real-world SBR datasets, including Diginetica, Retailrocket, and Yoochoose. The experimental results show that LINK consistently and significantly improves performance (up to 14.78% in Recall@20 and up to 11.04% in MRR@20) compared to existing state-of-the-art linear SBR models (such as SLIST and SWalk) on all datasets. This demonstrates that LINK successfully overcomes the limitations of linear models by incorporating neural network knowledge. In addition, LINK maintains the high inference efficiency (up to 813 times fewer FLOPs), which is a key advantage of linear models, while showing competitive or superior prediction accuracy compared to complex state-of-the-art neural network models. Further analysis shows that linear models are strong in the relationship between unpopular items, while neural network models are strong in the complex patterns of popular items, and LINK effectively combines these two strengths to achieve balanced performance. 

Figure 2: Comparison of Accuracy (Recall@20) and Inference Operations (FLOPs)

In conclusion, LINK presents a novel hybrid approach that provides a practical trade-off between accuracy and efficiency in the SBR field. In particular, the NIT component provides the flexibility to leverage knowledge from various models without being bound to a specific neural network architecture, making it a practical solution that can continuously improve performance as neural network models evolve in the future. For more information about this paper, please refer to the following address: https://dial.skku.edu/blog/2025_link

Study 3: Hye-young Kim, Minjin Choi, Sunkyung Lee, Ilwoong Baek, Jongwuk Lee, “DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation”, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), 2025

This study proposes a side-information integrated sequential recommendation (SISR) model that utilizes additional information such as category and brand in sequential recommendation, which predicts the next preferred item based on the user’s past consumption history. The proposed model, Dual Side-Information Filtering and Fusion (DIFF), removes noise in the user sequence and effectively fuses various attribute information to achieve more precise and expressive user preference modeling. DIFF includes the following three core techniques:

Figure 3: Frequency signals and fusion techniques of sequential recommendation systems that integrate additional information

(1) Frequency-based noise filtering: DIFF performs frequency domain transformation to remove signals that are not related to actual user preferences, such as accidental clicks or short-term interests. After converting the item ID and each attribute sequence to the frequency domain, it removes irregular or low-importance frequency components. This allows us to strengthen only the core signals that reflect actual user preferences, and enables more sophisticated noise removal by applying filtering to multiple sequences. (2) Dual multi-sequence fusion: DIFF utilizes intermediate fusion and early fusion methods, which have different strengths, together to effectively integrate the denoised sequences. We note that previous studies have tended to limit or exclude the use of early fusion methods due to concerns about information invasion, which has led to the overlooking of the ability to model correlations between various attributes. DIFF learns sophisticated user representations that encompass both IDs and attributes by integrating multidimensional attribute information through early fusion and supplementing ID-centric preference learning through intermediate fusion. Through the complementary combination of the two fusion methods, DIFF can effectively capture not only the overall structure of user tastes but also detailed attribute preferences. (3) Representation alignment to prevent information invasion: Item IDs and each attribute embedding are located in different representation spaces. Therefore, early fusion that combines them with a simple fusion function (e.g. summation, concatenation, gating) may cause information invasion, where specific information is overly emphasized or distorted. To prevent this, DIFF designs an alignment loss to make the vector spaces of item IDs and attribute embeddings close together, thereby maintaining appropriate differences while sharing meaning. DIFF has been validated on four representative public benchmark datasets (Yelp, Beauty, Toys, Sports), and has achieved superior performance in all indicators compared to existing state-of-the-art sequential recommendation models. In particular, it has demonstrated a new state-of-the-art performance by recording performance improvements of up to 14.1% and 12.5% on Recall@20 and NDCG@20, respectively. In addition, DIFF's robustness against noise is very remarkable. Considering noise in a realistic usage environment such as accidental clicks and temporary changes in interest in the user sequence, we conducted noise simulation experiments by randomly replacing items in the test sequence. As a result, DIFF showed the least performance degradation compared to other models even under low noise conditions of 5%, and maintained stable high performance even under high noise conditions of 25%. For more information about this paper, please refer to the following address:

https://dial.skku.edu/blog/2025_diff