Exploiting ID-Text Complementarity via Ensembling for Sequential Recommendation

Liam Collins, Bhuvesh Kumar, Clark Mingxuan Ju, Tong Zhao, Donald Loveland, Leonardo Neves, Neil ShaharXiv 2025 · DOI · arXiv

Introduction

Sequential Recommendation (SR) is a core task in modern information retrieval, aiming to predict a user's next interaction based on their historical sequence of activities. Historically, these models have relied on Item IDs—unique tokens assigned to every product. However, the recent success of Large Language Models (LLMs) has shifted the research focus toward Modality Features, specifically text descriptions.

Currently, the field is divided into two primary camps: the Replacement Camp, which argues that text embeddings are sufficient to make IDs redundant, and the Fusion Camp, which attempts to combine both signals through complex alignment layers or multi-stage training. This paper challenges both perspectives by demonstrating that ID and text signals are fundamentally complementary and that complex fusion often "washes out" the unique benefits of each. Instead, the authors propose EnsRec, a simple ensembling method that outperforms state-of-the-art models by letting each signal "speak for itself."

Quantifying Complementarity

To move beyond anecdotal evidence, the authors first quantify the difference between what ID-based models and text-based models learn. They utilize two primary metrics:

  1. Jaccard Index ($\tilde{J}$): This measures the overlap of successful recommendations between two models. A low Jaccard Index indicates that the models succeed on different sets of users/items, suggesting high complementarity.
  2. "Genie" Metric: This represents a theoretical upper bound, calculating the performance achieved if an oracle could perfectly choose the better-performing model for every single user interaction.

Their analysis revealed a crucial insight: an ID-based model and a text-based model are significantly more complementary to each other than two ID-based models with different architectures. This suggests that the source of the information (ID vs. text) is more critical for performance than the specific neural architecture used.

The EnsRec Methodology

Rather than designing a complex neural network to mix features during training, the authors propose a decoupled approach called EnsRec.

Independent Training

The process begins by training two models in isolation:

  • ID-based Model: A standard sequential model (such as SASRec) trained using only unique Item IDs.
  • Text-based Model: A model trained using text embeddings from a frozen LLM (specifically SentenceT5-XXL) to represent items.

Score Ensembling

At inference time, the scores from both models are combined to produce the final recommendation list. In its simplest form, the ensemble sums the scores:

$$s_{Ens}(i) = s_{ID}(i) + s_{text}(i)$$

The authors also explored a generalized version, $EnsRec_{\alpha, \tau}$, which introduces a weight ($\alpha$) to balance the two signals and a temperature parameter ($\tau$) to control the sharpness of the score distributions.

Experimental Results

The authors evaluated EnsRec on four standard benchmarks: the Steam dataset and three categories from the Amazon Review 2018 dataset (Beauty, Toys, and Sports).

Despite its architectural simplicity, EnsRec consistently outperformed sophisticated, state-of-the-art baselines including LLM-ESR (which uses cross-attention) and UniSRec (which utilizes whitening layers). Key performance highlights include:

  • Toys Dataset: EnsRec achieved a 12.3% improvement in NDCG@10 and a 6.0% improvement in Recall@10 over the strongest baseline.
  • The "Genie" Gap: While EnsRec performed exceptionally well, a significant gap remains between its performance and the theoretical "Genie" maximum. This suggests that while simple ensembling is effective, there is still substantial room for improvement in how these signals are integrated.

Discussion: Memorization vs. Generalization

The success of EnsRec stems from how it handles the distinct strengths of each feature type. ID models are highly effective at memorization—learning specific collaborative patterns (e.g., "users who bought this specific item usually buy that specific item"). In contrast, text models excel at semantic generalization (e.g., "this user has a preference for sci-fi themes").

Complex fusion models often force these two distinct signals into the same mathematical space. The authors argue that this "alignment" can be harmful, as it forces the models to ignore the unique, non-overlapping signals that make each feature type useful. By keeping the models independent, EnsRec preserves these unique strengths.

Conclusion and Future Work

EnsRec demonstrates that in the quest for better recommendation systems, complexity is not always a prerequisite for performance. By acknowledging the inherent complementarity of IDs and text, researchers can achieve better results with simpler, more interpretable architectures.

Future research directions include developing user-wise gating mechanisms to dynamically weight the ID or text model for specific users and extending this complementarity analysis to other modalities such as images and video.

Summary updated

Abstract

Modern Sequential Recommendation (SR) models commonly utilize modality features to represent items, motivated in large part by recent advancements in language and vision modeling. To do so, several works completely replace ID embeddings with modality embeddings, claiming that modality embeddings render ID embeddings unnecessary because they can match or even exceed ID embedding performance. On the other hand, many works jointly utilize ID and modality features, but posit that complex fusion strategies, such as multi-stage training and/or intricate alignment architectures, are necessary for this joint utilization. However, underlying both these lines of work is a lack of understanding of the complementarity of ID and modality features. In this work, we address this gap by studying the complementarity of ID- and text-based SR models. We show that these models do learn complementary signals, meaning that either should provide performance gain when used properly alongside the other. Motivated by this, we propose a new SR method that preserves ID-text complementarity through independent model training, then harnesses it through a simple ensembling strategy. Despite this method's simplicity, we show it outperforms several competitive SR baselines, implying that both ID and text features are necessary to achieve state-of-the-art SR performance but complex fusion architectures are not.
BibTeX
@misc{collins2025exploiting,
  author = {Liam Collins and Bhuvesh Kumar and Clark Mingxuan Ju and Tong Zhao and Donald Loveland and Leonardo Neves and Neil Shah},
  title = {{Exploiting ID-Text Complementarity via Ensembling for Sequential Recommendation}},
  year = {2025},
  eprint = {2512.17820},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.17820},
}
Figure