[Paper Review] Two Birds with One Stone: Unified Model Learning for Both Recall and Ranking in News Recommendation

Chuhan Wu, Fangzhao Wu+2 more|arXiv (Cornell University)|Apr 15, 2021

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes UniRec, a unified model that jointly learns user embeddings for both news recall and ranking in personalized news recommendation. By using the ranking user embedding as an attention query to select and combine basis user embeddings, UniRec achieves state-of-the-art performance in both tasks while reducing computational overhead through model unification.

ABSTRACT

Recall and ranking are two critical steps in personalized news recommendation. Most existing news recommender systems conduct personalized news recall and ranking separately with different models. However, maintaining multiple models leads to high computational cost and poses great challenge to meeting the online latency requirement of news recommender systems. In order to handle this problem, in this paper we propose UniRec, a unified method for recall and ranking in news recommendation. In our method, we first infer user embedding for ranking from the historical news click behaviors of a user using a user encoder model. Then we derive the user embedding for recall from the obtained user embedding for ranking by using it as the attention query to select a set of basis user embeddings which encode different general user interests and synthesize them into a user embedding for recall. The extensive experiments on benchmark dataset demonstrate that our method can improve both efficiency and effectiveness for recall and ranking in news recommendation.

Motivation & Objective

• To address the high computational and memory cost of maintaining separate models for news recall and ranking in large-scale systems.
• To unify the learning of user embeddings for both recall and ranking without sacrificing performance.
• To leverage user interest representations from ranking to enhance recall quality while filtering noisy interests.
• To meet strict online latency requirements in real-time news recommendation systems.

Proposed method

• A user encoder model processes historical clicked news to generate a user embedding for ranking using multi-head self-attention and position embeddings.
• The ranking user embedding is used as an attention query to select top-k basis user embeddings from a memory of M general user interest representations.
• The selected basis embeddings are dynamically weighted and combined to form a personalized user embedding for recall.
• Only the top-P basis embeddings (with highest attention weights) are used in inference to filter out noisy or irrelevant user interests.
• The news encoder uses BERT-style architecture to embed news content, while the user encoder models click behavior sequences with self-attention and positional encoding.
• The unified framework shares the news encoder and ranking user encoder across both recall and ranking, minimizing model redundancy.

Experimental results

Research questions

• RQ1 Can a single unified model effectively learn user embeddings for both news recall and ranking without performance degradation?
• RQ2 How can user interest representations from ranking be leveraged to improve recall quality?
• RQ3 What is the optimal number of basis user embeddings and how many should be selected for recall to balance performance and efficiency?
• RQ4 Can attention-based selection of basis embeddings filter out noisy user interests and improve recall diversity and accuracy?

Key findings

• UniRec outperforms strong baselines like NAML and NPA in ranking performance, demonstrating the effectiveness of multi-head self-attention and position embeddings.
• UniRec (top), which uses only the top 5 basis embeddings, achieves better recall performance than UniRec(all), which uses all basis embeddings, indicating selective attention improves noise filtering.
• The optimal number of basis user embeddings M is 20, as performance degrades when M is too small or too large due to insufficient coverage or overfitting.
• The optimal number of selected basis embeddings P for recall is 5, as performance declines when P is too small (incomplete coverage) or too large (inclusion of noisy embeddings).
• Extensive experiments on a real-world benchmark dataset confirm that UniRec achieves state-of-the-art results in both recall and ranking with a unified model.
• Case studies show that UniRec generates diverse and accurate recall results that align with inferred user interests across multiple categories.