Efficient Content-based Recommendation Model Training via Noise-aware Coreset Selection

TL;DR

This work tackles the data-efficiency and noise challenges in content-based recommender systems by introducing Noise-aware Coreset Selection (NaCS). NaCS integrates gradient-guided submodular coreset construction with progressive label self-correction and Monte Carlo dropout denoising to build compact, high-quality training subsets that are robust to noisy implicit feedback. The method achieves substantial efficiency gains, recovering up to $93-95\%$ of full-dataset performance with as little as $1\%$ of the data across large-scale datasets, and demonstrates transferability to text-based recommendation with LLM-inspired backbones. Overall, NaCS offers a practical, scalable pathway to maintain accuracy while dramatically reducing training costs in CRSs, including settings involving LLMs.

Abstract

Content-based recommendation systems (CRSs) utilize content features to predict user-item interactions, serving as essential tools for helping users navigate information-rich web services. However, ensuring the effectiveness of CRSs requires large-scale and even continuous model training to accommodate diverse user preferences, resulting in significant computational costs and resource demands. A promising approach to this challenge is coreset selection, which identifies a small but representative subset of data samples that preserves model quality while reducing training overhead. Yet, the selected coreset is vulnerable to the pervasive noise in user-item interactions, particularly when it is minimally sized. To this end, we propose Noise-aware Coreset Selection (NaCS), a specialized framework for CRSs. NaCS constructs coresets through submodular optimization based on training gradients, while simultaneously correcting noisy labels using a progressively trained model. Meanwhile, we refine the selected coreset by filtering out low-confidence samples through uncertainty quantification, thereby avoid training with unreliable interactions. Through extensive experiments, we show that NaCS produces higher-quality coresets for CRSs while achieving better efficiency than existing coreset selection techniques. Notably, NaCS recovers 93-95\% of full-dataset training performance using merely 1\% of the training data. The source code is available at \href{https://github.com/chenxing1999/nacs}{https://github.com/chenxing1999/nacs}.

Figures (8)

• Figure 1: The overview of NaCS. First, we progressively construct the coreset while refining the labels. Next, we apply Monte Carlo Dropout to estimate uncertainty and remove highly uncertain samples, resulting in the final clean coreset.
• Figure 2: Example process of CRSs.
• Figure 3: Log loss of trained DeepFM model on 1% coreset. "Sample ID" is the index in the sorted array based on log loss.
• Figure 4: Runtime relative to the full dataset of DCNv2 backbone with various budgets. "Select" and "Denoise" represent the runtime of our two steps. "Original" represents the training time on the original dataset.
• Figure 5: Performance and runtime trade-off between our method and complete dataset training. The red line ("Performance") represents the AUC ratio between the model trained on our result coresets and the original dataset. "m" stands for minutes.