What preference signals beyond reviews can improve recommendation steering?
This explores what signals other than star ratings and written reviews can sharpen the steering of recommendations — and the corpus turns out to read 'preference signal' much more broadly than the rating matrix.
This explores what signals beyond reviews and ratings can sharpen recommendation steering, and the collection's most interesting move is to widen what counts as a 'preference signal' in the first place. The narrowest answer — and the most surprising — comes from watching how humans actually recommend to each other. An analysis of a thousand real recommendation dialogues found that the conversations that landed weren't the ones full of preference questions; they were full of personal opinion sharing, encouragement, expressions of similarity, and appeals to credibility Do recommendation strategies beyond preference questions work better?. In other words, the steering signal is partly social and rhetorical, not just a tally of what you liked.
A second cluster says the richest untapped signal is the *side information* sitting next to the rating — item attributes, metadata, and the structure connecting them. Knowledge-graph attention networks fold item attributes and user-item interactions into a single graph so that attribute-similarity and user-similarity steer recommendations together, surfacing high-order connections plain collaborative filtering never sees Can graphs unify collaborative filtering and side information?. Graph autoencoders use that same side information to recommend for brand-new users and items where no rating history exists at all Can autoencoders solve the cold-start problem in recommendations?. And when reviews *are* present but sparse, aspect-aware retrieval pulls in relevant fragments from other users' reviews and selects which aspects matter to *this* person — turning thin signal into steerable, personalized explanation Can retrieval enhancement fix explainable recommendations for sparse users?.
The most counterintuitive signal is your social graph — but not in the way you'd guess. Methods that assume friends share your taste underperform a model that deliberately exploits friends with *different* preferences, because the network's real value is influencing your anomalous, off-pattern choices rather than confirming your usual ones Can friends with different tastes improve recommendations?. Steering, here, comes from diversity in the signal, not similarity.
There's also a signal hiding *inside* the user. Rather than collapsing a person into one preference vector, attention-weighted persona models represent each user as several tastes and let the candidate item decide which persona to listen to — so a recommendation can be traced to the specific facet of you it satisfies, and diversity falls out for free Can attention mechanisms reveal which user taste explains each recommendation?. Pushed further, the text-to-text framing treats every scrap of metadata, interaction, and description as natural language, letting one model fuse heterogeneous signals and even steer toward items and domains it never saw in training Can one text encoder unify all recommendation tasks?.
The quiet warning across all of this: steering signals also steer *outcomes*. The type of product network you recommend through — bought-together versus co-viewed — changes whether opinions converge or diverge, because each draws a different audience Do different recommender types shape opinion convergence differently?. The richer your signals, the more your recommender behaves as persuasion infrastructure rather than a neutral mirror How do recommendation feeds shape what people see and believe?. Worth knowing before you go signal-hunting: more steering power is also more responsibility for what it converges people toward.
Sources 9 notes
Analysis of 1,001 human recommendation dialogues shows successful recommendations correlate with personal opinion sharing, encouragement, similarity signals, and credibility appeals—not just preference questions. Opinion and experience sharing appear in 30% and 27% of recommendation sentences respectively.
KGAT merges user-item interaction graphs with item knowledge graphs into a Collaborative Knowledge Graph, using attention-based propagation to capture both user-similarity and attribute-similarity signals simultaneously—including high-order connections that standard supervised learning methods miss.
GHRS uses graph features and deep autoencoders to integrate rating history with side information, enabling predictions for new users and items by discovering non-linear relationships that linear hybrid methods miss.
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.
Social Poisson Factorization uses friends' diverse tastes to recommend items outside users' usual preferences, outperforming methods that pull friends' representations together. Networks add value through influence on anomalous choices, not taste similarity.
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.
P5 converts user-item interactions and metadata into natural language and trains a single encoder-decoder across five recommendation task families, matching task-specific models while achieving zero-shot transfer to new items and domains. Unification trades efficiency for composability.
Research shows that frequently-bought-together and co-viewed recommendation networks produce different opinion convergence patterns. The mechanism: each recommender type attracts different audience segments with different prior expectations, shaping both who sees products together and how they rate them.
Research shows recommendation systems operate as political actors: feed weights influence producer behavior, network topology drives opinion convergence, and automation enables targeted persuasion at population scale. These effects compound through rating contamination and selection biases.