- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper studies item cold-start recommendation through a leave-one-ID-out induction objective. The authors propose Softmax-Hull Embeddings (SHE), which construct an item representation as a content-similarity-weighted combination of neighboring item embeddings. The theoretical analysis bounds the empirical induction objective near a warm-model optimum using reconstruction error and a Lipschitz assumption relating content and behavioral representations. For efficient training, the method evaluates the recommendation loss on a single mixed embedding table that interpolates between each item’s own embedding and its neighborhood reconstruction. Experiments with MF and SASRec on Amazon and Onion report improved warm-target performance and competitive cold-target results.
Critical strengths
S1. The embedding construction is simple, interpretable, and readily integrated into existing recommenders. S2. The leave-one-ID-out formulation provides a useful framework for examining representation substitution.
Critical weaknesses
W1. The theoretical claims exceed what the analysis establishes. Proposition 2 proves that SHE minimizes a kernel-weighted least-squares objective over neighboring embeddings, but this does not establish optimality for leave-one-ID-out reconstruction or cold-item generalization. Moreover, Proposition 5 justifies the single-pass surrogate only locally around a warm optimum, whereas the actual training reinitializes the backbone and jointly optimizes its parameters. If the mixing matrix is invertible and the embeddings are unconstrained, the surrogate can even be viewed as a reparameterization of the original recommendation objective. The authors should clarify the scope of the optimality claim and empirically validate the surrogate against the original induction objective.
W2. The experimental evidence does not sufficiently establish the incremental contribution. SHE is closely related to kernel regression, proxy-based item representations (Seol et al., WSDM 2024), and training-time neighbor substitution (SwapRec, 2026). Comparisons with these approaches and a pretrained backbone using inference-only neighbor interpolation would better isolate the proposed training mechanism. Furthermore, Section 3.1 states that cold items use , whereas Section 4.3 attributes declining cold-target performance at larger to their untrained self-embeddings. This inconsistency requires clarification. Finally, differences in test-history handling between SHE and ID-based baselines may partly explain the warm-target gains, motivating a controlled comparison with identical input histories.
Questions for authors
See weakness.
Reviewer whia
Thank you for the comments on the paper. The revised paper provides pointers in purple, with a topic name and the weakness, e.g., "(scope: whia W1)".
Summary of results. In the revised paper, SASRec + SHE is the most accurate method over the full test set on both datasets (Table 2, "All"), and with the new content fine-tuning (Remark 1) it is also the best cold-target method on Amazon.
W1, Prop. 2. (tag: "scope") We narrowed the claims (abstract; text after Prop. 2 in Sec. 2.2): Prop. 2 characterizes SHE as the kernel-regression estimate for a given neighborhood and kernel. We no longer claim optimality for leave-one-out reconstruction or for cold-item generalization, and we assess the latter empirically.
W1, Prop. 5. (tag: "scope") We agree that Prop. 5 (Sec. 2.4) holds around a local minimum. Outside this neighborhood the evidence is empirical: with both E and ω randomly initialized (Sec. 4.2, "Initialization of E"; Table 3), SASRec + SHE still outperforms its backbones.
W1, reparameterization. (tag: "reparameterization") We commented on this in Remark 2 (Sec. 2.4; App. A). The surrogate is the warm loss reparameterized as F=WqE, but cold items are represented through E, not F. Their embeddings differ from post-hoc interpolation over a warm optimum by a term linear in the hull residual (I−Wˉ)F∗, which vanishes only if Q=0.
While the two parameterizations are equivalent on warm items (and hence for the training loss), they are not equivalent on cold items, as shown by the new inference-only ablation (Sec. 4.2; Table 4).
As requested, we also validated the single-pass objective against the leave-one-out condition of Eq. 1 on the trained models (new App. C.10; Table 11). When an item is withheld and rebuilt from its content neighbors, the pre-trained SASRec cannot recommend it (HR@10 .0010 on Amazon), whereas SASRec + SHE recommends it almost as accurately as from its own embedding (.0717 vs. .0745). The additional training loss caused by rebuilding an item is more than ten times lower than for the backbone (.0067 vs. .0900).
W2, baselines. (tags: "inference-only", "swap baselines") The table above also addresses W2 on inference-only interpolation. We further added a comparison with SwapRec, its inference-only counterpart (nearest-neighbor swap), and the proxy-based item representations of Seol et al. (2024), all using SASRec as backbone (new App. C.11; Table 12). Over all test sequences, SASRec + SHE is more accurate than all three on both datasets (on Onion, within the confidence intervals of the two swap methods).
W2, q=0. (tag: "q=0 at inference") Thank you for spotting this; the statement was wrong. Cold items always use q=0 while the q in Fig. 1 is the training self-share. We hence attribute the accuracy decline to a train–inference mismatch in q-values, and reframed the discussion accordingly.
W2, histories. (tag: "identical histories") Previous work (SwapRec, 2026) showed that sequential models underperform when cold-item inputs are represented by the randomly initialized embeddings as compared to "dropping" them (i.e., the setup in which we evaluate them). Nevertheless, we added in App. C.9 (Table 10) an evaluation on the test sequences whose history contains no cold item, under which all models receive identical inputs. On Amazon (19,370 such sequences), SHE improves over its backbone by 46% (SASRec) and 136% (MF) in HR@10 over all targets. On Onion, only 203 sequences do not contain a cold item as input, and although our approach does not reach the best accuracy, on this small sample the differences are not statistically significant.