ZoRRO: A Zero-Weight Personalized Recommender System for Scalable News Recommendation

Technical Summary: ZoRRO – A Zero-Weight Personalized Recommender System for Scalable News Recommendation

Problem Statement
News recommendation platforms face unique challenges distinct from other domains, primarily driven by the constant influx of new articles and their extremely short lifespans. This dynamic environment exacerbates the cold-start problem for both new items and users, necessitating systems that can adapt rapidly to fresh content and shifting user interests. While deep learning-based recommender systems (DLRS) have advanced significantly, they often demand substantial computational resources, specialized expertise, and continuous maintenance. In practice, the complexity of these models often limits their adoption, and their reported performance gains are not always justified in production environments where latency, scalability, and retrieval efficiency are critical. There is a growing need for practical, scalable approaches that balance effectiveness with operational efficiency without relying on heavy computational demands.

Methodology: ZoRRO
The authors propose ZoRRO (Zero-Weight Personalized Recommender System), a framework designed to be training-free and zero-weight, specifically optimized for the dynamic nature of news. ZoRRO operates by combining article recency with representation-based similarity, eliminating the need for model retraining.

The core relevance score $r(c)$ for a candidate article $c$ is computed based on two components:

1. Intrinsic Relevance ($\phi$): This captures an article's inherent importance, modeled using an exponential decay function ($e^{-\lambda t}$) that assigns higher weights to recent articles. This applies to both the candidate article and the articles in the user's history.
2. Relational Relevance ($\tau$): This measures the similarity between a candidate article and the user's historical articles. The framework utilizes cosine similarity on two types of representations:
   • Document Embeddings ($\mathbf{e}$): Generated by language models (e.g., BERT, RoBERTa, or contrastive learning models) to capture textual semantics.
   • Category Vectors ($\mathbf{v}$): One-hot encoded vectors indicating the article's category.

The final relational relevance is the sum of content-based and category-based similarities. The overall score aggregates these factors, allowing the system to seamlessly integrate new content and adapt to user preferences instantly. The framework is flexible regarding the specific similarity function or representation method used, though the authors found contrastive-learning embeddings and the combination of semantic and categorical features yielded the best results.

Key Contributions

• Framework Design: Introduction of ZoRRO, a simple, efficient, zero-weight, and training-free recommendation framework that combines recency with representation-based similarity.
• Performance vs. Efficiency: Demonstration through large-scale offline experiments that a well-tuned lightweight method can outperform or match strong neural baselines (such as NRMS, LSTUR, and NPA) while achieving inference speeds more than 600× faster.
• Online Validation: Validation of the approach in a live online A/B test, revealing that strong offline ranking performance does not automatically guarantee the best online Click-Through Rate (CTR).
• Beyond-Accuracy Analysis: Provision of both offline and online analyses showing that models with similar CTR or ranking performance can induce meaningfully different recommendation distributions regarding topic, sentiment, and diversity, impacting editorial alignment and user exposure.

Experimental Results

• Offline Evaluation: Using the Ekstra Bladet News Recommendation Dataset (EB-NeRD) with over 37 million impression logs, ZoRRO achieved the highest performance across standard ranking metrics (AUC, MRR, nDCG) on the hidden test set, outperforming the strongest neural baseline, NRMS (AUC 63.80 vs. 62.46).
• Inference Speed: ZoRRO demonstrated a throughput of approximately 1,954 requests per second with an average latency of 0.51 ms, compared to roughly 2–3 requests per second and latencies exceeding 300 ms for neural baselines.
• Online A/B Test: In a six-day live test on the Ekstra Bladet website, NRMS achieved the highest CTR (4.33%), closely followed by ZoRRO (4.19%), with both significantly outperforming a "Most Popular" baseline (2.96%).
• Distributional Differences: Despite similar CTRs, the models exhibited distinct behaviors. ZoRRO produced more concentrated recommendations with a stronger focus on news content (47.2% vs. 21.3% for NRMS) and lower diversity. NRMS exposed users to a broader range of topics, particularly sports. Furthermore, ZoRRO's top-5 recommendations had a narrower time span (approx. 9h 50m) compared to NRMS (20h 40m), reflecting its stronger reliance on recency.

Significance and Claims
The paper positions ZoRRO as a practical and efficient solution for large-scale news recommendation, highlighting that complex deep learning models are not always necessary to achieve state-of-the-art performance. The authors emphasize that the gap between offline ranking metrics and online user engagement (CTR) is significant; a model that ranks well offline may not perform best online, and vice versa.

Crucially, the study underscores the importance of evaluating recommender systems using metrics beyond accuracy alone. The authors argue that models with similar accuracy can lead to vastly different topical and sentiment distributions, which has direct implications for the reliability, editorial alignment, and deployment decisions of news platforms. ZoRRO serves as a strong baseline that balances effectiveness with the operational constraints of real-time news environments.