How do aspect-aware retrieval and surrogate models compare as explainability approaches?
This explores two different routes to making recommender systems explainable — pulling in external evidence (aspect-aware retrieval) versus building a stand-in model that mimics the recommender's own reasoning (surrogate models) — and what each one buys you.
This explores two genuinely different philosophies of explanation, and the corpus has one clean example of each. Aspect-aware retrieval, as in ERRA Can retrieval enhancement fix explainable recommendations for sparse users?, treats explanation as an evidence problem: when a user's history is too thin to justify a recommendation, go fetch relevant reviews from the outside and personalize which aspects of the product to highlight. The explanation is assembled from real external signal. The surrogate approach, as in RecExplainer Can LLMs explain recommenders by mimicking their internal states?, treats explanation as a mind-reading problem: train an LLM to imitate the recommender — its outputs (behavior alignment), its internal embeddings (intention alignment), or both — so it can narrate why the black-box model decided what it did.
The sharpest way to see the contrast is what each is faithful to. The surrogate model is faithful to the recommender — its whole goal is to reproduce the target system's internal states, so its explanations describe the actual decision. Retrieval enhancement is faithful to the world — it grounds explanations in genuine review text and user-relevant aspects, but those reasons are plausible justifications, not necessarily the literal computation the recommender ran. So one tells you what the model is doing; the other tells you what's true about the item and user. RecExplainer's hybrid mode is interesting precisely because it tries to have it both ways: faithful to the model's internals while staying human-readable.
They also fail in opposite places. Retrieval inherits every weakness of retrieval itself — and the corpus is blunt that those weaknesses are structural, not tuning problems Where do retrieval systems fail and why?. Most relevant here: embeddings measure semantic association, not task relevance Do vector embeddings actually measure task relevance?, so an aspect-aware retriever can surface reviews that sound related to the recommendation without actually explaining it. The surrogate model sidesteps the retrieval-relevance trap entirely, but takes on a harder burden: it has to faithfully mirror a model it doesn't have privileged access to, and if the imitation drifts, the explanation is confidently wrong.
There's a bridge worth noticing. Rationale-driven evidence selection Can rationale-driven selection beat similarity re-ranking for evidence? shows that letting an LLM reason about which evidence matters beats picking by similarity by a wide margin — which is essentially importing surrogate-style reasoning into the retrieval step. That hints the real answer isn't retrieval versus surrogate but retrieval guided by a model that understands relevance, blunting the embedding-association problem that otherwise undercuts aspect-aware methods.
The takeaway a curious reader might not expect: "explainability" in recommenders isn't one goal. Retrieval-based explanation optimizes for explanations users find convincing and well-grounded; surrogate-based explanation optimizes for explanations that are actually true to the system. Those can diverge, and which one you want depends on whether you're trying to persuade the user or audit the model.
Sources 5 notes
ERRA combines model-agnostic review retrieval with personalized aspect selection to address data sparsity that embedded methods cannot solve. Retrieval augmentation provides richer signal when user history is sparse, while aspect personalization ensures explanations match user context rather than generic defaults.
RecExplainer trains LLMs via three alignment methods: behavior (mimicking outputs), intention (incorporating neural embeddings), and hybrid (combining both). The hybrid approach produces explanations that are simultaneously faithful to the target model and intelligible to users by balancing internal-state inspection with human-readable reasoning.
RAG systems fail at three structural levels: adaptive triggering (fixed intervals waste context), semantic-task mismatch (embeddings measure association, not relevance), and mathematical limits (embedding dimension constrains representable document sets). These require fundamentally different retrieval approaches, not tuning.
Embeddings encode co-occurrence patterns, making semantically close but role-distinct concepts highly similar. This works in simple demos but fails in production where underspecified queries have many wrong-but-associated candidates.
METEORA uses LLM-generated rationales with flagging instructions to select evidence, achieving 33% better accuracy with 50% fewer chunks than similarity re-ranking across legal, financial, and academic domains. The method also improves adversarial robustness substantially.