Embed Progressive Implicit Preference in Unified Space for Deep Collaborative Filtering

TL;DR

GNOLR tackles the challenge of integrating multiple implicit feedback signals by modeling user engagement as a progressive ordinal process. It maps unstructured feedback into ordered labels and employs a nested, category-specific neural OLR within a twin-tower embedding framework to unify representations across feedback types. The approach yields a single, retrieval-friendly embedding space and a hierarchical optimization objective, improving both ranking accuracy and candidate retrieval efficiency, as demonstrated on nine real-world datasets. The work advances beyond traditional feedback-wise multi-task models by reducing task conflict and enabling one-stage ranking with robust performance under data imbalance. Theoretical and empirical results highlight GNOLR’s ability to capture engagement progression and leverage cross-task information for scalable recommendation.

Abstract

Embedding-based collaborative filtering, often coupled with nearest neighbor search, is widely deployed in large-scale recommender systems for personalized content selection. Modern systems leverage multiple implicit feedback signals (e.g., clicks, add to cart, purchases) to model user preferences comprehensively. However, prevailing approaches adopt a feedback-wise modeling paradigm, which (1) fails to capture the structured progression of user engagement entailed among different feedback and (2) embeds feedback-specific information into disjoint spaces, making representations incommensurable, increasing system complexity, and leading to suboptimal retrieval performance. A promising alternative is Ordinal Logistic Regression (OLR), which explicitly models discrete ordered relations. However, existing OLR-based recommendation models mainly focus on explicit feedback (e.g., movie ratings) and struggle with implicit, correlated feedback, where ordering is vague and non-linear. Moreover, standard OLR lacks flexibility in handling feedback-dependent covariates, resulting in suboptimal performance in real-world systems. To address these limitations, we propose Generalized Neural Ordinal Logistic Regression (GNOLR), which encodes multiple feature-feedback dependencies into a unified, structured embedding space and enforces feedback-specific dependency learning through a nested optimization framework. Thus, GNOLR enhances predictive accuracy, captures the progression of user engagement, and simplifies the retrieval process. We establish a theoretical comparison with existing paradigms, demonstrating how GNOLR avoids disjoint spaces while maintaining effectiveness. Extensive experiments on ten real-world datasets show that GNOLR significantly outperforms state-of-the-art methods in efficiency and adaptability.

Figures (9)

• Figure 1: An illustration comparing the prior paradigm with GNOLR. Prior methods embed entities into feedback-wise, incommensurable spaces, requiring separate ranking and fusion before displaying to users. In contrast, GNOLR unifies various feedback in a shared embedding space, aligning the spatial proximity with the user preference progression for seamless one-stage ranking and improved prediction.
• Figure 2: Comparison of three architectures. NSB (left) represents the predominant collaborative filtering framework for multiple implicit feedback signals, known as Naive Shared Bottom, which models each task independently. Neural OLR (middle) extends OLR to neural modeling using a shared encoder. GNOLR (right) further generalizes OLR with nested encoding and subtasks to enhance expressibility and unify the embedding of user engagement across tasks.
• Figure 3: The impact of $a_c$ and $\gamma$ on Sigmoid predictions. $a_c$ shifts the Sigmoid curve horizontally. $\gamma$ modifies the steepness of the Sigmoid curve, controlling which region of the input space receives greater focus during learning.
• Figure 4: Visualization of the angular distribution between user and item embeddings under single- and multi-task settings. We fix the user directions and plot item directions. NSB* uses sample re-weighting for better performance.
• Figure 5: Parameter sensitivity w.r.t. $a$ and $\gamma$ on KR-Pure.

Theorems & Definitions (1)

• definition 1