Recommender Systems

Can retrieval enhancement fix explainable recommendations for sparse users?

When users have few historical interactions, embedded recommendation models struggle to generate personalized explanations. Can augmenting sparse histories with retrieved relevant reviews—selected by aspect—overcome this fundamental data limitation?

Note · 2026-05-03 · sourced from Recommenders Personalized
How do recommendation feeds shape what people see and believe? What breaks when specialized AI models reach real users?

Three frameworks exist for explainable recommendation: post-hoc methods that generate explanations for a black-box model, embedded methods that learn explanations as part of model representation, and multi-task methods that share latent representations between recommendation and explanation. Post-hoc explanations have limited diversity. Embedded methods rely on historical interactions to learn representations, so they fail when the user or item has insufficient data. Multi-task methods often use Transformers but encounter limited review content, leading to performance decline in generation tasks.

ERRA's contribution is two enhancements added to multi-task learning. Retrieval enhancement is model-agnostic: identify and select the most relevant reviews from a corpus (the dataset's training set, used as a retrieval corpus rather than Wikipedia). This addresses data sparsity by augmenting the user's limited historical reviews with relevant external ones. Aspect enhancement is personalized: extract aspects from each user and item review using extraction tools, then select the most relevant aspects per user-item pair. The same user might care about environment when choosing a hotel and price when choosing a phone — aspect selection adapts to the recommendation context.

The combined effect: retrieval enhancement gives the model more information to work with when user history is sparse, and aspect enhancement makes the explanations user-specific rather than generic. The general pattern: when sparsity hits an embedding-based model, retrieval augmentation is a more effective remedy than larger embeddings — and personalizing the augmentation (aspect-aware retrieval) further improves quality.


Source: Recommenders Personalized

Related concepts in this collection

Concept map
13 direct connections · 73 in 2-hop network ·medium cluster

Click a node to walk · click center to open · click Open full network for a force-directed map

your link semantically near linked from elsewhere
Original note title

explainable recommendation needs aspect-aware retrieval enhancement — sparse user history defeats embedded-only methods