ConvRec: Convolutional Sequential Recommender

Updated 5 July 2026

ConvRec is a convolution-based model that aggregates user interactions into a compact sequence representation via hierarchical down-sampling.
It replaces global self-attention with cascaded 1D convolutions, achieving linear time and memory complexity while leveraging localized patterns.
Experimental results across multiple domains show ConvRec outperforms attention-based models with noticeable gains in HR@10 and efficiency.

Searching arXiv for papers on "ConvRec" and related usages to ground the article. ConvRec is a convolution-based architecture for attribute-aware sequential recommendation introduced in "Rethinking Convolutional Networks for Attribute-Aware Sequential Recommendation" (Elsayed et al., 6 May 2026). In that setting, a model predicts the next item a user will interact with from a chronologically ordered interaction history augmented with item attributes and interaction context. The defining claim of ConvRec is that it replaces global self-attention with a hierarchy of 1D convolutions that progressively down-sample the sequence until a single compact sequence representation remains, while maintaining linear computational and memory complexity. In adjacent literature, the string "ConvRec" also appears as shorthand for conversational recommendation or as part of the name "ConvRec-R1," an LLM-based conversational recommender trained with reinforcement learning (Zhu et al., 23 Oct 2025). The term therefore has both a specific architectural meaning and a broader field-level ambiguity.

1. Definition and scope

In its specific technical sense, ConvRec denotes an alternative to attention-based attribute-aware sequential recommenders. The underlying problem assumes a set of users $\mathcal{U}$ and items $\mathcal{I}$ , where each user $u$ has a time-ordered interaction sequence

$S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$

and each interacted item carries side-information $a_k^u$ and interaction context $c_k^u$ . The prediction target is the next item $i_{|S^u|+1}^u$ (Elsayed et al., 6 May 2026).

The motivation is explicitly framed against self-attention-based models such as SASRec, BERT4Rec, and TiSASRec, which compute pairwise weights over all $L$ positions at cost $O(L^2 \cdot d)$ in time and memory, where $d$ is the embedding dimension. For long histories, this quadratic scaling is described as prohibitive in GPU memory and wall-clock time; the paper further argues that attention may under-utilize locality structure because neighboring interactions are often most predictive (Elsayed et al., 6 May 2026).

A separate conversational-recommendation line uses related nomenclature differently. "ConvRec-R1" is defined as a two-stage framework for end-to-end training of LLM-based conversational recommender systems, while other works use "ConvRec" as shorthand for the conversational recommendation problem setting rather than for a convolutional sequential model (Zhu et al., 23 Oct 2025). This suggests that the name is context-sensitive and should be interpreted by venue and task.

2. Architectural design

ConvRec consists of two main components: an item encoding module and a Convolutional Down-Scaling component (CDS). The item encoding module integrates attributes, context, and item identity into a shared representation. Raw attribute features $\mathcal{I}$ 0 and context features $\mathcal{I}$ 1 are projected as

$\mathcal{I}$ 2

then fused by

$\mathcal{I}$ 3

Item identity is represented by a lookup-table embedding

$\mathcal{I}$ 4

which is combined with the fused side-information as

$\mathcal{I}$ 5

(Elsayed et al., 6 May 2026).

Temporal interval information is injected by computing

$\mathcal{I}$ 6

followed by

$\mathcal{I}$ 7

and

$\mathcal{I}$ 8

This produces the sequence representation consumed by the convolutional hierarchy (Elsayed et al., 6 May 2026).

The CDS maps $\mathcal{I}$ 9 to a single vector $u$ 0 using $u$ 1 successive ConvBlocks. For block $u$ 2, with $u$ 3,

$u$ 4

then

$u$ 5

and

$u$ 6

where $u$ 7 are learnable scalars initialized to $u$ 8, and $u$ 9. After $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 0 blocks, the remaining sequence is collapsed as

$S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 1

The paper characterizes this design as hierarchical and down-scaled, with neighboring items aggregated gradually until a comprehensive sequence representation is formed (Elsayed et al., 6 May 2026).

3. Sequence modeling, scoring, and optimization

The forward pass summarized in Algorithm 1 first computes $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 2 through item/context encoding and interval fusion, initializes $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 3 and $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 4, applies the ConvBlocks for $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 5, collapses the final output into $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 6, then scores a target item via the interaction between the target embedding and the final sequence vector (Elsayed et al., 6 May 2026).

Next-item scoring is defined as

$S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 7

where, for a candidate item $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 8, the candidate embedding is denoted $S^u = [i_1^u, i_2^u, \ldots, i_{|S^u|}^u],$ 9. Training uses binary cross-entropy with one positive item and $a_k^u$ 0 negatives:

$a_k^u$ 1

These formulations place ConvRec squarely in the standard implicit-feedback next-item prediction regime, but with an encoder that is explicitly attribute-aware and interval-aware (Elsayed et al., 6 May 2026).

The paper’s interpretation of the hierarchy is that local dependencies are handled first and then merged progressively. A plausible implication is that the model operationalizes a locality bias through convolution rather than through explicit pairwise attention weights. The residual pathway using both $a_k^u$ 2 and pooled progressive residuals indicates that the architecture is designed to retain global information from the original sequence while compressing length layer by layer (Elsayed et al., 6 May 2026).

4. Complexity and efficiency profile

ConvRec’s principal systems claim is linear complexity. Each convolutional block $a_k^u$ 3 costs $a_k^u$ 4. Because the stride is chosen such that $a_k^u$ 5, the paper states that the output length satisfies $a_k^u$ 6. Summing across blocks gives

$a_k^u$ 7

which, with typical $a_k^u$ 8, is stated to sum to $a_k^u$ 9. Peak memory is reported as $c_k^u$ 0 because only intermediate representations of shrinking length need to be stored, rather than an $c_k^u$ 1 attention map (Elsayed et al., 6 May 2026).

The empirical efficiency comparison is reported on the Games domain with batch size 128 on an A40 GPU. There, memory scales as $c_k^u$ 2 for ConvRec versus $c_k^u$ 3 for ProxyRCA, and wall-time per batch at $c_k^u$ 4 is given as $c_k^u$ 5 for ConvRec versus $c_k^u$ 6 for ProxyRCA (Elsayed et al., 6 May 2026).

The broader significance of these claims is methodological. The paper explicitly positions ConvRec against the dominant assumption that long-range sequential modeling in recommendation should be mediated by self-attention. Its argument is not merely that convolution is cheaper, but that a carefully structured convolutional hierarchy can also be more effective for extracting sequential patterns. This suggests a reevaluation of the standard attention-centric design space for recommender encoders (Elsayed et al., 6 May 2026).

5. Experimental results

The experimental study uses four Amazon domains: Beauty, Games, Fashion, and Men. The reported dataset statistics are as follows.

Dataset	#Users	#Items
Beauty	52 204	57 289
Games	31 013	23 715
Fashion	45 184	166 270
Men	34 244	110 636

A second set of reported statistics gives the interaction volume, average sequence length, maximum sequence length, and number of attributes: Beauty has 394 908 interactions, AvgSeq 7, MaxSeq 293, and #Attrs 6 507; Games has 287 107 interactions, AvgSeq 9, MaxSeq 860, and #Attrs 506; Fashion has 358 003 interactions, AvgSeq 8, MaxSeq 192, and #Attrs 2 048; Men has 254 870 interactions, AvgSeq 7, MaxSeq 308, and #Attrs 2 048 (Elsayed et al., 6 May 2026).

The baselines are divided into sequence-only methods—SASRec, BERT4Rec, TiSASRec, CosRec—and attribute-aware methods—S $c_k^u$ 7-Rec, SASRec++, DIF-SR, CARCA, CARCA-LIP, ProxyRCA. The evaluation protocol is leave-one-out, with the last item per user used as test and the second-last as validation, alongside 100 negative samples per test user. The reported metrics are HR@10 and NDCG@10 (Elsayed et al., 6 May 2026).

On the 100-negative evaluation, ConvRec is reported to outperform all baselines on HR@10 and NDCG@10. The paper provides the following HR@10 comparison against the best baseline, which in all four cases is ProxyRCA:

Domain	Best-baseline (HR@10)	ConvRec (HR@10)
Beauty	0.609	0.629 ± 0.001
Games	0.805	0.819 ± 0.001
Fashion	0.661	0.688 ± 0.003
Men	0.613	0.618 ± 0.001

The corresponding relative gains are reported as $c_k^u$ 8 on Beauty, $c_k^u$ 9 on Games, $i_{|S^u|+1}^u$ 0 on Fashion, and $i_{|S^u|+1}^u$ 1 on Men. On full-item evaluation for Beauty and Games, ConvRec is also reported to lead by $i_{|S^u|+1}^u$ 2– $i_{|S^u|+1}^u$ 3 relative gains (Elsayed et al., 6 May 2026).

6. Ablations, inductive bias, and interpretation

The ablation study on Beauty HR@10 attributes measurable contributions to several components. Removing intervals yields $i_{|S^u|+1}^u$ 4 versus the full model’s $i_{|S^u|+1}^u$ 5; removing residuals yields $i_{|S^u|+1}^u$ 6 versus $i_{|S^u|+1}^u$ 7; replacing the hierarchy with a single convolution over the entire sequence yields $i_{|S^u|+1}^u$ 8; and using average pooling only yields $i_{|S^u|+1}^u$ 9 (Elsayed et al., 6 May 2026).

Kernel and stride selection is treated as a key architectural variable. On Beauty, the best reported setting is $L$ 0, producing HR@10 $L$ 1, and the paper states that $L$ 2 proved superior to overlap. On Games, sequence-length-group analysis shows that ConvRec outperforms ProxyRCA for all $L$ 3 in $L$ 4, especially on shorter sequences (Elsayed et al., 6 May 2026).

The paper’s explanation for why convolution is effective on long sequences has three elements: locality bias, hierarchical down-sampling, and linear scaling. Convolutions focus on neighbors first and then progressively merge information, which the authors state mirrors natural user behavior patterns; hierarchical down-sampling captures both short-term transitions and long-range context in fewer operations; and linear scaling allows training on longer histories without quadratic blowup, preserving more of the user’s past (Elsayed et al., 6 May 2026).

A common misconception in this area is that attention is intrinsically superior for sequence modeling. The ConvRec results do not support that as a universal claim. Instead, they indicate that in attribute-aware next-item prediction, a hierarchical convolutional encoder can surpass state-of-the-art attention-based recommenders under the reported evaluation protocols. A second misconception is that convolution necessarily sacrifices global context; the CDS is explicitly constructed to reach a single comprehensive representation through gradual aggregation and residual pathways (Elsayed et al., 6 May 2026).

7. Terminological ambiguity and neighboring uses

The term "ConvRec" is not semantically unique across recommendation research. In sequential recommendation, it refers to the convolutional model just described (Elsayed et al., 6 May 2026). In conversational recommendation, closely related naming appears in at least three distinct ways.

First, "ConvRec-R1" designates a two-stage LLM-based conversational recommender trained by supervised fine-tuning followed by reinforcement learning with Rank-GRPO. Its workflow is: user dialogue $L$ 5 Stage 1 SFT $L$ 6 Stage 2 RL $L$ 7, after which the model generates a ranked list directly in natural language (Zhu et al., 23 Oct 2025).

Second, "ConvRec" is used generically as shorthand for conversational recommendation in "Retrieval-Augmented Conversational Recommendation with Prompt-based Semi-Structured Natural Language State Tracking," which introduces RA-Rec, a retrieval-augmented, LLM-driven dialogue state tracking system. That system uses prompt-based intent classification, semi-structured JSON state tracking, dense review retrieval with TAS-B and FAISS, and LLM-based response generation, but does not define "ConvRec" as a standalone model name (Kemper et al., 2024).

Third, conversational recommendation is also the application context for cold-start preference elicitation in PEBOL, which formulates natural-language preference elicitation as Bayesian optimization over latent utilities and uses NLI-based updates with LLM-guided query generation (Austin et al., 2024). Dataset work such as U-NEED likewise situates itself in E-commerce conversational recommendation and related tasks including dialogue understanding, user needs elicitation, recommendation, generation, and evaluation (Liu et al., 2023).

This broader usage pattern implies that "ConvRec" has become overloaded. In current arXiv usage, it may denote a specific convolutional sequential recommender, a named conversational recommender variant such as ConvRec-R1, or the conversational recommendation area in abbreviated form. For technical precision, disambiguation by full paper title or task domain is therefore essential.