Scholarlink

<p>Collaborative filtering remains a cornerstone of modern recommender systems, yet individual models exhibit complementary strengths and systematic prediction biases that limit their standalone performance. We propose <b>AEEMU</b> (Adaptive Ensemble with Enhanced Meta-Learning and Uncertainty-aware filtering), a three-stage framework that integrates classical signal processing techniques with deep meta-learning to improve ensemble recommendation accuracy. In Stage&#xa0;1, four heterogeneous base recommenders—Matrix Factorization (MF), Neural Collaborative Filtering (NCF), Self-Attentive Sequential Recommendation (SASRec), and Light Graph Convolution Network (LightGCN)—independently generate predictions and learned embeddings. In Stage 2, embedding-level filters (wavelet, spectral, bilateral) denoise base model representations before a dual-pathway meta-network with MLP and attention branches produces instance-adaptive ensemble weights guided by a multi-objective loss function combining prediction accuracy, weight alignment, entropy regularization, and ranking constraints. In Stage 3, weight-smoothing filters (Kalman, adaptive) and prediction-refinement filters (EMA, median, Savitzky-Golay, particle, confidence, consensus) post-process the ensemble output to produce the final prediction. We evaluate <b>AEEMU</b> on four benchmark datasets (MovieLens&#xa0;100K, Amazon Digital Music, Book-Crossing, and Jester) using 5-fold cross-validation. The meta-network ensemble achieves competitive RMSE against recent graph-based methods (MBRCC, DualGCN) and substantial improvements over the ranking-oriented SimpleX baseline (45–60%, reflecting a ranking-to-rating evaluation mismatch). A comprehensive ablation study covering 23 filter configurations reveals that signal processing filters yield a +1.1% NDCG@5 ranking gain on ML-100K with marginal RMSE trade-off, while on sparse datasets the meta-network alone captures the primary ensemble benefit and filters introduce 6–14% RMSE perturbation, indicating that filter effectiveness is modulated by data density. All filters use fixed hyperparameters across datasets, demonstrating robustness without per-dataset tuning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AEEMU: Adaptive Ensemble with Enhanced Meta-Learning and signal processing for recommendation systems

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Antonio José Martínez Casares

摘要

Collaborative filtering remains a cornerstone of modern recommender systems, yet individual models exhibit complementary strengths and systematic prediction biases that limit their standalone performance. We propose AEEMU (Adaptive Ensemble with Enhanced Meta-Learning and Uncertainty-aware filtering), a three-stage framework that integrates classical signal processing techniques with deep meta-learning to improve ensemble recommendation accuracy. In Stage 1, four heterogeneous base recommenders—Matrix Factorization (MF), Neural Collaborative Filtering (NCF), Self-Attentive Sequential Recommendation (SASRec), and Light Graph Convolution Network (LightGCN)—independently generate predictions and learned embeddings. In Stage 2, embedding-level filters (wavelet, spectral, bilateral) denoise base model representations before a dual-pathway meta-network with MLP and attention branches produces instance-adaptive ensemble weights guided by a multi-objective loss function combining prediction accuracy, weight alignment, entropy regularization, and ranking constraints. In Stage 3, weight-smoothing filters (Kalman, adaptive) and prediction-refinement filters (EMA, median, Savitzky-Golay, particle, confidence, consensus) post-process the ensemble output to produce the final prediction. We evaluate AEEMU on four benchmark datasets (MovieLens 100K, Amazon Digital Music, Book-Crossing, and Jester) using 5-fold cross-validation. The meta-network ensemble achieves competitive RMSE against recent graph-based methods (MBRCC, DualGCN) and substantial improvements over the ranking-oriented SimpleX baseline (45–60%, reflecting a ranking-to-rating evaluation mismatch). A comprehensive ablation study covering 23 filter configurations reveals that signal processing filters yield a +1.1% NDCG@5 ranking gain on ML-100K with marginal RMSE trade-off, while on sparse datasets the meta-network alone captures the primary ensemble benefit and filters introduce 6–14% RMSE perturbation, indicating that filter effectiveness is modulated by data density. All filters use fixed hyperparameters across datasets, demonstrating robustness without per-dataset tuning.