Why did conversational recommenders drop both item and user similarity signals?
This explores why conversational recommender systems (CRS) — the ones that chat with you to suggest things — stopped using the two classic signals that powered older recommenders: 'people like you liked X' (user similarity) and 'items like the ones you liked' (item similarity).
This explores why conversational recommenders abandoned the two collaborative-filtering signals — user-to-user and item-to-item similarity — that made traditional recommenders work. The short answer from the corpus: it wasn't a deliberate design choice so much as a side effect of how CRS framed the problem. By treating the live chat as the whole story, these systems only infer your taste from the current dialogue session, throwing away the historical behavior and look-alike-user structure that similarity signals are built from Can conversational recommenders recover lost preference signals from history?. A single conversation is simply too thin a slice of a person to reconstruct 'users like you' or 'items like these.'
The fix being proposed is to stop choosing between dialogue and history and instead feed the recommender three preference channels at once — the current session, your past conversations, and similar users — all conditioned on what you seem to want right now Can conversational recommenders recover lost preference signals from history?. That reframing matters because the rest of the corpus shows just how much signal a chat-only view leaves on the table. When CRS models the *order* items get mentioned in a conversation, accuracy improves over treating mentions as an unordered bag — sequence itself is preference information that flat models discard Does conversation order matter for recommending items in dialogue?.
There's also a subtler reason the field got away with dropping these signals for so long: the benchmarks were rewarding shortcuts. More than 15% of the 'correct' items in a standard CRS dataset were already named earlier in the same conversation, so a trivial baseline that just parrots back mentioned items beats most trained models Do conversational recommender benchmarks actually measure recommendation skill?. If the scoreboard rewards copying from the current chat, there's little pressure to rebuild the harder user- and item-similarity machinery.
What's interesting is that the broader recommendation literature never lost faith in those similarity ideas — it just expressed them differently. Some work represents each user as *multiple* personas weighted per candidate item, which is collaborative structure made interpretable Can attention mechanisms reveal which user taste explains each recommendation?. Other work patches the sparse-data problem CRS suffers from by retrieving outside text — sentiment-matched reviews or aspect-aware passages — to enrich a thin conversation Can review sentiment alignment fix sparse CRS dialogue? Can retrieval enhancement fix explainable recommendations for sparse users?. And LLM-based recommenders are now re-importing large-corpus retrieval strategies that are essentially industrial-strength similarity search How should LLM-based recommenders retrieve from massive item corpora?.
So the deeper takeaway isn't that similarity signals were wrong — it's that conversational systems traded them away for a tempting simplification (the chat is enough), and the current research wave is busy buying them back through history channels, sequence modeling, and retrieval augmentation. The thing you didn't know you wanted to know: the 'conversational' framing that made these systems feel smart is the same framing that made them forget who you are between conversations.
Sources 7 notes
Current CRS systems only use the active dialogue session to infer preferences, losing item-CF and user-CF signals proven valuable in traditional recommenders. Integrating current session, historical dialogues, and look-alike users—conditioned on current intent—recovers essential user representation structure.
TSCR models items and entities in the order they appear in CRS dialogue, using transformers to learn dependencies between sequential mentions. This recovers information that bag-of-mentions approaches discard, improving recommendation accuracy on standard benchmarks.
Over 15% of ground-truth items in INSPIRED are items already mentioned earlier in conversation. A naive baseline that copies mentioned items outperforms most trained models, showing the metric rewards shortcut learning rather than real recommendation ability.
AMP-CF represents each user as multiple latent personas weighted dynamically by candidate item. This makes recommendations both diverse and interpretable—each suggestion traces to the specific persona preference it satisfies—without requiring post-hoc reranking.
RevCore demonstrates that retrieving user reviews with polarity matching the user's stance—then integrating them into dialogue history and generation—produces more informative and aligned recommendations. Sentiment-coordinated filtering prevents contradictory context that random review retrieval would introduce.
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.
RecLLM identifies four retrieval patterns—dual-encoder, direct LLM search, concept-based, and search-API lookup—each optimized for different corpus sizes, latency budgets, and training constraints. Hybrid approaches mixing multiple strategies likely work best for real systems.