Can category information and temporal order improve detection of complementary products?
This explores whether two extra signals — what category a product belongs to, and the order people buy or view things in — help a system spot complements (products that go together, like a phone and a case) rather than substitutes.
This explores whether category metadata and the sequence of user actions sharpen a system's ability to find complementary products — and the corpus answers from the side, since it has more on the 'category' half than the 'temporal' half. Start with what 'complementary' even means as a signal. The clearest doorway is research showing that 'frequently-bought-together' and 'co-viewed' product networks are genuinely different objects: they attract different audiences and even shape product ratings differently Do different recommender types shape opinion convergence differently?. That matters here because complementarity is exactly the frequently-bought-together relationship, and the lesson is that co-purchase and co-view are not interchangeable proxies for it — which signal you mine changes what you detect.
On the category question, the corpus is encouraging. The recurring finding is that pure interaction data (who bought what) misses relationships that attribute or category data exposes. Knowledge-graph attention networks fold item attributes into the same graph as user-item clicks, and the payoff is precisely the 'high-order' connections — A relates to B through a shared attribute even when nobody bought both — that plain collaborative filtering can't see Can graphs unify collaborative filtering and side information?. Graph autoencoders make the same case from the cold-start angle: blending rating history with side information lets the system reason about brand-new items it has no co-purchase history for Can autoencoders solve the cold-start problem in recommendations?. Category is one of the most reliable side-information channels, so both point to yes — it helps, especially for items too new or too rare to have co-purchase signal yet.
There's a subtler payoff in how you encode that category. One line of work shows that item identifiers built from ID plus title plus attributes beat pure-ID or pure-text identifiers, because the attribute facet carries semantic meaning the bare ID lacks Can item identifiers balance uniqueness and semantic meaning?. The implication for complement detection: a model that knows an item's category as part of its representation can generalize 'cameras pair with tripods' to a tripod it has never seen, rather than memorizing specific co-purchase pairs.
The temporal-order half is where the corpus is thinner, and it's worth saying so plainly rather than padding. None of these notes directly model purchase sequence for complementarity. The nearest adjacent idea is that a user isn't one fixed taste vector but several personas, weighted dynamically depending on the candidate item Can attention mechanisms reveal which user taste explains each recommendation? — a hint that *which* complement is relevant shifts with context, which is a cousin of temporal context, but not a direct test of buy-order.
So the honest synthesis: category information has strong corpus support for improving complement detection, mostly by rescuing items that lack co-purchase history and by letting models generalize across attribute-similar products; temporal order is plausible but largely unaddressed in this collection. The most useful thing you might not have expected — that the 'complementary' relationship itself splits into co-purchase versus co-view variants with different behavior Do different recommender types shape opinion convergence differently? — means the first design decision isn't which features to add, but which definition of 'goes together' you're actually trying to detect.
Sources 5 notes
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.
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.
TransRec shows that combining numeric IDs, titles, and attributes into structured identifiers solves three problems simultaneously: distinctiveness from IDs, semantics from text, and generation grounding from structural constraints. Neither pure IDs nor pure text alone achieves all three.
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.