Efficient Pointwise-Pairwise Learning-to-Rank for News Recommendation

TL;DR

This work addresses efficient learning-to-rank for news recommendation by integrating pointwise relevance with pairwise preferences in a single PLM framework called GLIMPSE. It employs multi-task fine-tuning to train a model for both Rel and Pref as text-generation tasks and introduces an RTL-based inference that refines an initial pointwise ranking with at most $m$ passes on the top-$k$ items, yielding $O(K)$ complexity per pass. A Markov-chain theoretical analysis establishes conditions under which RTL refinements improve expected ranking metrics, linking pairwise accuracy to gains in $\text{MRR}$ and related measures. Empirically, GLIMPSE outperforms state-of-the-art baselines on MIND and Adressa datasets, with notable improvements in AUC and DCG metrics while maintaining practical runtimes, demonstrating the approach's practical impact for scalable, high-quality news recommendation.

Abstract

News recommendation is a challenging task that involves personalization based on the interaction history and preferences of each user. Recent works have leveraged the power of pretrained language models (PLMs) to directly rank news items by using inference approaches that predominately fall into three categories: pointwise, pairwise, and listwise learning-to-rank. While pointwise methods offer linear inference complexity, they fail to capture crucial comparative information between items that is more effective for ranking tasks. Conversely, pairwise and listwise approaches excel at incorporating these comparisons but suffer from practical limitations: pairwise approaches are either computationally expensive or lack theoretical guarantees, and listwise methods often perform poorly in practice. In this paper, we propose a novel framework for PLM-based news recommendation that integrates both pointwise relevance prediction and pairwise comparisons in a scalable manner. We present a rigorous theoretical analysis of our framework, establishing conditions under which our approach guarantees improved performance. Extensive experiments show that our approach outperforms the state-of-the-art methods on the MIND and Adressa news recommendation datasets.

Figures (2)

• Figure 1: An illustration of the proposed framework. GLIMPSE consists of a multi-task training approach where the PLM is fine-tuned by considering both the relevance prediction and the pairwise preference tasks. During inference, the relevance predictions are used to obtain an initial pointwise ranking, which is subsequently improved by performing one or more right-to-left (RTL) passes using pairwise comparisons.
• Figure 2: Comparison of the distribution shift after RTL passes. The figures show how often the first positive label is at each position. The figure on the left shows the distribution obtained from a pointwise model using 5% of the data (see main text), the figure in the middle is based on the same weak pointwise model, but uses an RTL pass on the top-5 items. It can be seen that RTL passes progressively move the first positive label to the front of the ranked list.

Theorems & Definitions (2)

• Theorem 3.1
• Corollary 3.1.1