Item-Centric Recommendations

• Item-centric recommendations are approaches that model items directly using co-occurrence, relation-aware embeddings, and interaction states to improve retrieval and explanation.
• They employ techniques like factorization machines, skip-gram based embeddings, and cold-start exploration to capture nuanced item relationships and contextual dynamics.
• This paradigm enhances catalog coverage and performance metrics (e.g., CTR, MAP, satisfaction) in both sequential and page-level recommendation settings.

Searching arXiv for the cited item-centric recommendation papers to ground the article in current arXiv records. Item-centric recommendations are recommender-system methods that treat items, item–item relations, item–interaction states, or item-derived modules as primary modeling units for retrieval, ranking, exploration, or explanation. In the literature, the term spans several related formulations: item–item similarity models over implicit-feedback sequences, interaction-conditioned item embeddings, page-level selection of item recommendation modules on an item details page, cold-start exploration that seeks the best users for a new item, and retrieval-augmented large-language-model pipelines that retrieve items rather than users (Zhao et al., 2018, Demir et al., 2017, Lo et al., 2021, Wang et al., 12 Jul 2025, Kim et al., 19 Nov 2025). Across these variants, the central premise is that catalog structure, item relations, and item-specific uncertainty can be modeled directly, often yielding scalable retrieval, broad catalog coverage, and interpretable recommendation pathways.

1. Conceptual scope and problem formulations

Early and contemporary item-centric methods share a common abstraction: they replace or supplement a user–item matrix with structures in which items are related to other items through co-occurrence, relations, or shared content. In implicit-feedback recommendation, a standard formulation learns a scoring function $r(u,i)$ over users and items, often as a dot product $p_u^\top q_i$, and optimizes pointwise, pairwise, or sampled-softmax objectives over positive-only logs (Rendle, 2021). Item-centric systems specialize this general setting by making item–item transitions, item neighborhoods, or item-conditioned candidate generation the dominant inductive bias.

One influential formulation uses local co-occurrence in interaction sequences. SoundCloud’s item–item music recommender constructs a co-occurrence matrix by sliding a window of size $W=10$ over positive user–track interaction sequences and then fits a Factorization-Machine model over a sparse indicator vector that encodes the track, the context track, and optional side features such as the creator (Demir et al., 2017). A related but more explicitly relational formulation is MoHR, which defines a set of explicit item–item relation types $\mathcal{R}$ together with an extra latent relation $r_0$, learns shared item embeddings, user embeddings, and relation embeddings, and scores next-item recommendations by combining a long-term term with a mixture of relation-specific experts (Kang et al., 2018). In both cases, the recommendation object is not an isolated item but an item situated inside a learned item neighborhood or relation space.

A stricter inversion of the usual paradigm appears in item-centric cold-start exploration. Instead of solving “for each user, which item maximizes predicted satisfaction,” the item-centric objective seeks, for each item, the users with the highest predicted $p(S_+|u,i)$ relative to the item’s own reliability (Wang et al., 12 Jul 2025). This redefinition is notable because it distinguishes item-centric recommendation from merely item-based similarity: here the optimization target itself is item-first.

2. Item representations, interaction states, and relation-aware embeddings

A defining development in item-centric recommendation is the move from plain item identifiers to richer item-centered tokens. Etsy’s interaction-based model forms a vocabulary $P=\mathcal{L}\times\mathcal{I}$, where each token $p=(\ell,i)$ denotes an item together with an interaction type such as view, favorite, add-to-cart, or purchase (Zhao et al., 2018). User sessions are sequences of such tokens, and a symmetric skip-gram window of size $m=5$ is used to predict neighboring item–interaction tokens. Each token has an input embedding $u_p\in\mathbb{R}^d$ and an output embedding $p_u^\top q_i$0, with $p_u^\top q_i$1, and the conditional probability of a context token is modeled by a softmax over all tokens in $p_u^\top q_i$2. Negative sampling replaces the full softmax, and the negative construction is customized to the partial order purchase $p_u^\top q_i$3 add-to-cart $p_u^\top q_i$4 favorite $p_u^\top q_i$5 view, so that higher-priority unobserved interactions on the same item can serve as negatives (Zhao et al., 2018). This design makes interaction type part of the semantic unit rather than a downstream feature.

In SoundCloud’s ITEM and ITEMc models, the representation is instead a sparse feature vector over track ID, context-track ID, and side-feature IDs, with a Factorization-Machine prediction

$p_u^\top q_i$6

where $p_u^\top q_i$7 is the latent dimension (Demir et al., 2017). When only track and context-track indicators are active, the interaction term reduces to an item–item dot product; adding creator features introduces additional cross-terms. This formulation preserves the item–item viewpoint while allowing side information to participate in the same latent geometry.

MoHR generalizes item-centric modeling to heterogeneous item relationships. It learns a single shared item embedding $p_u^\top q_i$8 for each item, a user embedding $p_u^\top q_i$9, and a relation embedding $W=10$0 for each explicit or latent relation. Its gating function

$W=10$1

selects which relation type is most plausible for the current user and previous item, while each expert scores a candidate next item by a translational form $W=10$2 (Kang et al., 2018). The resulting score decomposes next-item prediction into relation selection and relation-specific item transition.

Relation-aware item representations are further extended by KGIF, which initializes a collaborative knowledge graph with TransD, then explicitly fuses entity and relation embeddings through dynamic projection, Hadamard products, a two-layer MLP, and attentive propagation (Jeon et al., 7 Jan 2025). In KGIF, item attributes and item–item relations are treated on equal footing, and final user–item prediction is an inner product over concatenated multi-layer fused embeddings. This suggests that item-centric recommendation need not be limited to pairwise co-occurrence; it can also be framed as explicit information fusion over a heterogeneous graph.

3. Side information, cold start, and catalog coverage

A major rationale for item-centric methods is their ability to handle sparse or newly introduced inventory. In the SoundCloud study, ITEMc augments the item–item model with creator as side information and improves average Mean Percentile Rank from 0.0449 for ITEM to 0.0422, while the implicit matrix-factorization baseline IMPL obtains 0.0901; the creator feature yields the biggest lift in the long tail (Demir et al., 2017). The mechanism is straightforward: side information enters the same latent interaction function as the track and context-track fields.

A different strategy is to transfer collaborative knowledge into content weights. The wrapper feature-weighting framework of “Deriving item features relevance from collaborative domain knowledge” learns a nonnegative weight vector $W=10$3 so that a weighted content similarity $W=10$4 approximates a target collaborative similarity matrix $W=10$5 (Dacrema et al., 2018). In the reported experiments, the learned wrapper weights close approximately 80–90% of the gap between pure content-based filtering and collaborative filtering, improve raw tf–idf by +15–25% relative MAP and +10–20% relative Precision@10 across all datasets, and improve cold-start item recommendations by up-weighting the content features most predictive of collaborative similarity (Dacrema et al., 2018). This is item-centric because recommendation at inference time depends only on item features and item–item weighted similarities.

The eBay two-tower system addresses marketplace cold start by computing item embeddings solely from static content attributes: title text, aspects, and category ID (Wang et al., 2021). User embeddings are computed from multi-modal on-site behaviors, including item views and search-query “pseudo-items,” and the model is trained with sampled softmax over approximately 10 million unique impressions. Offline, the recurrent-plus-search variant improves Recall@20 from 0.12 to 0.13 relative to a Recently-Viewed-Items baseline, and online it increases the surface rate by approximately 6% at 90% coverage of listing pages (Wang et al., 2021). The item tower is thus cold-start resilient because it does not require prior item interaction statistics.

Cold-start exploration is treated even more directly in the item-centric control of Wang et al. Each new item has an intrinsic satisfaction rate $W=10$6, updated with impressions and satisfied impressions, and an item is shown to a user only if

$W=10$7

with $W=10$8 in the experiments (Wang et al., 12 Jul 2025). Online evaluation reports +50% and +40% improvements on two internal Satisfaction Metric signals, a 20% reduction in exploration impressions, and a +10% increase in recommendable corpus size (Wang et al., 12 Jul 2025). Here item-centricity is operationalized as uncertainty-aware allocation of users to items.

ItemRAG extends the cold-start theme into LLM-based recommendation. It constructs an item–item retrieval database from co-purchase histories and text embeddings, augments direct co-purchase neighborhoods with neighborhoods of semantically similar items, and uses co-purchase frequencies as sampling weights (Kim et al., 19 Nov 2025). On Amazon domains, it improves Hit-Ratio@1 over a zero-shot LLM by up to 43% and remains robust in cold-start settings, with only about a 1–2% drop from standard to cold-start recommendation (Kim et al., 19 Nov 2025).

4. Sequential, session-based, and page-level item-centric systems

Many item-centric methods are inherently sequential because item relevance depends on recent item context. Etsy’s interaction embeddings are trained on session windows but explicitly model only first-order co-occurrence within the skip-gram neighborhood; the authors note that there is no explicit sequential dynamics beyond window $W=10$9 (Zhao et al., 2018). This limitation motivates deeper session models.

GRAINRec represents each session as a directed item graph in which, for every pair of positions $\mathcal{R}$0 with $\mathcal{R}$1, a directed edge $\mathcal{R}$2 is added (Rath et al., 2024). It alternates GRU-based GNN layers, intended to capture short-range dependencies, with attention layers, intended to capture long-range dependencies, and then scores candidate items by inner product with a local–global session representation. To meet service-level constraints, it precomputes a nearest-neighbor matrix of size $\mathcal{R}$3 and scores only the union of neighbors of items in the session. The reported system achieves an average improvement of 1.5% across offline metrics and, in a two-week A/B test, +10% click-through rate and +9% attributable demand (Rath et al., 2024).

A different sequential formulation targets multiply repeated interactions. The sequence-aware retail recommender of Equihua et al. treats each user’s purchase history as a sequence of item tokens, embeds item IDs, processes them through two stacked LSTM layers, and predicts per-item purchase probabilities with a five-layer MLP (Equihua et al., 2023). Offline, it outperforms collaborative filtering and matrix factorization on several top-rank metrics, and a live A/B test reports a 5% uplift in total sales and over 50% increase in per-customer expenditure (Equihua et al., 2023). The item-centric aspect lies in modeling repeated item consumption directly in the sequence rather than collapsing history into static user factors.

Item-centricity can also operate at the page or module level. On an item details page, the “hero” item defines the page context, and the system must select and order recommendation modules or carousels beneath it (Lo et al., 2021). The TRNN architecture models page-level module ordering as a sequential decision process, with a two-stage LSTM that first encodes page and family context and then scores remaining candidate modules. The production system uses inverse propensity scoring, multi-task losses for click-through, purchase-through, and user-intent signals, ONNX serving, beam search, and a constraint that no two consecutive modules come from the same family. Online A/B testing reports +2.48% CTR and +7.34% PTR over a static configuration (Lo et al., 2021). This widens the notion of item-centric recommendation from item-to-item retrieval to page composition around a focal item.

5. Retrieval, serving architectures, and industrial-scale efficiency

Industrial item-centric recommenders are typically retrieval-first systems. Etsy explicitly uses the inner product $\mathcal{R}$4 as an unnormalized affinity between item–interaction tokens and indexes output embeddings with HNSW in Faiss because the token space has approximately 200 million distinct tokens (Zhao et al., 2018). The reported HNSW parameters are $\mathcal{R}$5, $\mathcal{R}$6, and $\mathcal{R}$7, and serving forms a user-specific candidate set of size approximately 200 before passing it to a downstream re-ranker (Zhao et al., 2018). Offline, view/cart embeddings cover approximately 70% of active items and more than 80% of traffic, whereas co-purchase covers only 9% of items (Zhao et al., 2018). The coverage result is central: item-centric embeddings can expose substantially more of the catalog than sparse co-purchase tables.

The general retrieval problem is formalized in the implicit-feedback overview by Rendle, where a dot-product model maps directly to approximate Maximum Inner Product Search. Brute-force retrieval costs $\mathcal{R}$8 per query, and sublinear methods such as HNSW and FAISS are presented as standard mechanisms for top-$\mathcal{R}$9 retrieval over millions of item vectors (Rendle, 2021). This connection explains why item-centric models frequently adopt shared vector spaces and inner-product scoring.

eBay’s production architecture illustrates a complementary scaling pattern. Item embeddings are computed in batch, clustered with $r_0$0 centroids, and at serving time the system first scores centroids against the user embedding, then retrieves within the top $r_0$1 clusters, allocating a total of approximately $r_0$2 items proportionally to cluster scores (Wang et al., 2021). Results are stored in a key-value store and served in under 5 ms (Wang et al., 2021). By contrast, GRAINRec retains online session dynamics and controls latency by restricting inference to the union of precomputed neighbors, achieving 95th-percentile latency under 50 ms on CPUs (Rath et al., 2024). The page-level TRNN system likewise reports P95 response time under 50 ms under peak QPS of hundreds per second (Lo et al., 2021). Together, these systems show that item-centric recommendation is as much an indexing and serving problem as a representation-learning problem.

6. Evaluation, interpretability, misconceptions, and research directions

Evaluation in item-centric recommendation is heterogeneous because the target object varies. Etsy measures hit rate over candidate sets, defining $r_0$3 as the average fraction of later-interacted items of target interaction type $r_0$4 captured by the retrieved set (Zhao et al., 2018). SoundCloud uses Mean Percentile Rank, where lower is better and random ranking yields approximately 0.5 (Demir et al., 2017). Session and sequence models report hit@10, mrr@10, ndcg@10, MAP@1, MAP@10, or business metrics such as attributable demand, total sales uplift, and expenditure (Rath et al., 2024, Equihua et al., 2023). Page-level systems add unbiased estimators such as Direct Method and Doubly Robust estimators for CTR and PTR (Lo et al., 2021). This diversity reflects a substantive fact: item-centric recommendation is not a single task but a family of tasks with different retrieval units and business surfaces.

Interpretability is frequently stronger than in purely latent user-centric systems. Etsy’s per-interaction buckets support explanations of the form “Because you X’d item A you may Y item B” (Zhao et al., 2018). MoHR can decompose the recommendation score into relation-specific contributions and identify whether “also-bought,” “also-viewed,” or another relation dominated the decision (Kang et al., 2018). KGIF provides explainable recommendations through attention-weighted path visualization in the collaborative knowledge graph (Jeon et al., 7 Jan 2025). ItemRAG makes item–item evidence explicit in prompt text by summarizing retrieved co-purchase neighborhoods (Kim et al., 19 Nov 2025).

A common misconception is that item-centric recommendation excludes personalization. The literature does not support that view. eBay learns a shared user–item space (Wang et al., 2021), MoHR conditions relation gating on both user and last item (Kang et al., 2018), TRNN conditions page-level module selection on user and hero-item context (Lo et al., 2021), and item-centric cold-start exploration still compares predicted user satisfaction against item-centric uncertainty thresholds (Wang et al., 12 Jul 2025). Item-centricity therefore concerns the primary organizational unit of modeling, not the absence of user dependence.

Limitations are equally recurrent. Sparse interaction types degrade interaction-specific embeddings at Etsy; the model lacks explicit sequential dynamics beyond the skip-gram window (Zhao et al., 2018). KGIF notes potential over-smoothing at greater propagation depth and increased computation for large-scale relations (Jeon et al., 7 Jan 2025). ItemRAG identifies summarization cost and static item summaries as scalability constraints (Kim et al., 19 Nov 2025). IAM shows that standard token-level attention can obscure item-level collaborative structure and proposes separate intra-item and inter-item attention layers to re-center the item as the fundamental unit in LLM recommendation (Zhang et al., 20 Mar 2026). Future directions proposed across the literature include user–interaction–item triples, joint multimodal embeddings, adversarial or otherwise improved negatives, temporal and sequential item attributes, personalized attention masks, hybrid user+item exploration, and deeper sequence models such as session transformers (Zhao et al., 2018, Jeon et al., 7 Jan 2025, Zhang et al., 20 Mar 2026).

In aggregate, item-centric recommendation has evolved from item–item co-occurrence models into a broader design space encompassing interaction-aware embeddings, relation mixtures, knowledge-graph fusion, sequence and session models, page composition, item-first exploration, and item-based RAG. The unifying principle is stable: recommendations improve when item structure is modeled explicitly rather than treated as a by-product of user-centric scoring.