Comparison of residual-driven CF-BiLSTM ensemble boosting models with benchmarks for recommender systems

Abstract

Collaborative filtering methods and deep sequential recommender systems have each achieved strong empirical results but address different aspects of the recommendation problem: collaborative filtering captures structured latent user–item interaction patterns, while deep text encoders leverage the semantic content of user-generated reviews. Combining the two modalities in a principled and computationally efficient manner remains an open challenge.

This report builds on the residual multimodal boosting framework of Mboumi (2026),which embeds a temporal collaborative filtering component and a Bidirectional Long Short-Term Memory (BiLSTM) text encoder inside a multi-class LogitBoost outer loop. Rather than fusing modalities through feature concatenation, the framework treats review text as a residual correction signal: at each boosting iteration the BiLSTM targets the systematic prediction errors left unexplained by the collaborative filtering component, with fusion performed additively on the log-odds scale. The primary contribution of this work is to investigate the role of word embedding initialisation within this framework, by replacing the domain-specific SA-Tweedie embeddings used in Mboumi (2026) with general-purpose pre-trained GloVe embeddings (Pennington et al., 2014), yielding the CF-GloVe-BiLSTM-Boost model.

Experiments are conducted on two Amazon review datasets — the 2014 Amazon Beauty corpus (28,798 reviews; 1,340 users; 733 items) and the 2023 Amazon Movies & TV corpus (546,978 reviews; 11,462 users; 20,508 items) — and the proposed model is evaluated against four classes of baseline: seven temporal collaborative filtering variants due to Koren and Bell (2011b) as implemented by Steiner (2017), a gradient-boosted CF baseline without the text encoder (CF-Boost, useclf2=False), four state-of-the-art deep learning recommenders (BERT4Rec-style, NCF, SASRec, TransformerRec), and the original CF-Tweedie-BiLSTM-Boost of Mboumi (2026). Performance is measured on mean squared error (MSE), NDCG@5, and Recall@5.

CF-GloVe-BiLSTM-Boost substantially outperforms all baseline groups on MSE and NDCG@5 across both datasets. Compared to the best temporal CF baseline, MSE is reduced by 18.8% on the 2014 data and 45.4% on the 2023 data. Compared to the best deep learning baseline, MSE is reduced by 30.3% and 56.2% respectively. Significance tests confirm that all MSE and NDCG@5 improvements are statistically significant (p ≤0.010) with large effect sizes (|d|≥1.77 in every case). The sole exception is Recall@5 on the 2023 dataset, where NCF achieves a significantly higher value (0.288 vs. a mean of 0.233; p = 0.035; d = 1.83), suggesting that attention-based sequential models capture broader item coverage at the cost of rating accuracy.

Comparing the two BiLSTM-Boost variants reveals a nuanced embedding–domain interaction: on the 2014 beauty corpus, where reviews are short and stylistically homogeneous, CF-Tweedie-BiLSTM-Boost achieves significantly higher NDCG@5 ([Delta] = +0.003; p = 0.003; d = 1.82), while MSE is equivalent; on the larger and linguistically richer 2023 corpus, CF-GloVe-BiLSTM-Boost achieves a substantially lower MSE ([Delta] = +0.032; p < 0.001; d = 4.20), while ranking metrics are equivalent. GloVe embeddings thus provide stronger initialisation for rating prediction in diverse review corpora but do not consistently improve and may slightly hinder — top-k ranking on short, homogeneous reviews. Hyperparameter sensitivity analysis over 2,100 configurations demonstrates that CF-GloVe-BiLSTM-Boost generalises reliably across the full grid of boosting iteration counts and learning rates, with tight dev–test alignment throughout, in contrast to the erratic behaviour of the ablated CF-Boost model.