acceptodds
Under review as a conference paper at ICLR 2027

Semantic-Guided Structural information Borrowing for Item Cold-Start Recommendation on Graphs

Abstract

In graph-based recommender systems, items are ranked by embeddings learned from the structural information of the user-item interaction graph. A cold (new) item has no interaction history and is isolated in this graph, so it lacks such structural information and cannot be ranked effectively like warm items. Existing methods address this issue in two ways. The first is injecting content features, which provide a cold item with a semantic embedding, but not the structural embedding required for ranking. The second is complementing the structural information through simulated or reconstructed edges. Since these edges are generated instead of real, the complemented structural information may be noisy. To overcome these limitations, we propose an end-to-end Semantic-Guided Structural information Borrowing framework (SSB), which allows a cold item to borrow the structural information from its similar warm items to represent itself, thus the cold item obtains a structural embedding for accurate recommendation. Specifically, we use a fine-tuning-free LLM to help a cold item to accurately locate the similar warm items, ensuring that the borrowed structural information is indeed relevant. Then, we propose a soft-structure borrowing mechanism composed of a soft attention mechanism and hierarchical structure aggregation. This mechanism enables a cold item to flexibly determine how much information to borrow from each located similar warm item, thereby making a cold item obtain a continuous combination of warm item structural embeddings rather than embeddings learned from discrete edges. Moreover, we optimize a pseudo-cold mask method to address the lack of supervision signals caused by missing structural information of cold items, letting gradients flow back to the learnable parameters during training. Finally, we conduct extensive experiments on three Amazon datasets: Sports, Toys, and Beauty. The results show that SSB outperforms SOTA cold-start baselines on Sports and Toys, and is comparable to CGRC on Beauty.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Cold-start Recommendation+Graph Neural Networks+Large Language Models+Representation Learning+Self-supervised Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_gNqMOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses item cold-start in graph-based recommender systems. The proposed method, SSB, follows a locate-and-borrow scheme: for each cold item it retrieves the top-KK warm items that are nearest in the dense feature space of an external, fine-tuning-free LLM item encoder, and lets the cold item borrow the structural information of these warm semantic neighbors. In effect, this injects one-way soft edges from the cold item to its warm neighbors, each weighted by a learned soft attention. The model is built on LightGCN. The final representation of a cold item is a convex combination of the projected LLM content embedding and, for each LightGCN propagation layer, the attention-weighted aggregate of the neighbors' layer-wise structural embeddings. Combining the pseudo-cold mask method, the soft readout of structure, and the learned layer weights, the method outperforms existing cold-start baselines on Sports and Toys and is comparable to CGRC on Beauty.

Critical strengths

S1. The hierarchical structure aggregation is an original component. The default model (Eq. 6) reads a neighbor's representation Ej(l)E_j^{(l)} separately for each propagation layer l≥1l \ge 1, keeps the semantic prior as layer zero, and learns the layer weights β\beta. Decoupling how much to borrow from each neighbor (wjw_j) from which order of structural signal to read (βl\beta_l) is a well-motivated design, and I am not aware of prior work that borrows layer-wise representations in this way.

S2. Effect sizes are reported with variance. SSB and CGRC are run over five seeds with mean ± std (Table 2). The improvement over CGRC on cold items is convincing on Sports (Recall@20 8.28 ± 0.11 vs. 6.81 ± 0.23, +21.6%) and Toys (15.54 ± 0.16 vs. 14.78 ± 0.10), and the paper explicitly reports that SSB is comparable to CGRC on Beauty overall and below CGRC on the rarest bucket.

Critical weaknesses

W1. The formulation is not compared, analytically or empirically, with closely related existing methods, and the novelty is limited once the formulation is made explicit.

Eq. (5) (MLP fusion with a scalar gate) is presented as part of the method but is not used in the default model; the default model is Eq. (6), a normalized convex combination of the projected semantic embedding and attention-weighted neighbor embeddings. Since in LightGCN

Ej(1)=∑u∈N(j)A~ju Eu(0),E_j^{(1)} = \sum_{u \in N(j)} \tilde A_{ju}\, E_u^{(0)},

borrowing ∑jwjEj(l)\sum_j w_j E_j^{(l)} is exactly what LightGCN would compute for the cold item if it were attached, one-way, to the users of its KK neighbors through soft edges of weight wjA~juw_j \tilde A_{ju}. The method is therefore a representation-level, one-way edge augmentation of the user–item graph driven by a content kNN, with learned edge weights.

This construction coincides with that of GraphDA (Fan et al., SIGIR 2023). GraphDA augments the adjacency matrix with an item–item block WIIW_{II} whose non-zeros are the top-KK neighbors under embedding similarity (their Eq. 5, binary and symmetric) and re-trains a LightGCN on the augmented graph. Compared term by term, SSB differs from GraphDA in: the similarity source (frozen text embeddings instead of pre-trained collaborative embeddings, which is what makes the construction applicable to cold items), one-way instead of symmetric edges, softmax attention instead of degree-normalized binary weights, layer-wise reading with learned β\beta instead of uniform layer averaging, and end-to-end instead of two-stage training. These differences are substantive, but the paper neither cites GraphDA nor isolates any of them experimentally, so it cannot be determined which of them, if any, accounts for the reported gains. Content-kNN item–item graphs in multimodal GNN recommenders (e.g., LATTICE, Zhang et al., MM 2021) and virtual-neighbor approaches for isolated nodes (e.g., Cold Brew, Zheng et al., ICLR 2022) may also be relevant; I have not examined them in detail and mention them only as pointers.

W2. The ablations are not controlled, and natural baselines are missing.

  1. wo_hierarchical (Fig. 2) changes three factors simultaneously — layer-wise reading vs. reading the mixed structural embedding EjE_j, inclusion of Ej(0)E_j^{(0)} in the borrowed value, and linear convex combination vs. MLP with a gate — so its −13.9 / −10.2 / −4.9% cannot be attributed to the layer weights.
  2. There is no ablation of the soft attention: neither uniform nor cosine-softmax weights over the same top-KK set. With K=3K = 3 (Beauty) and K=5K = 5 (Toys) the learned weights have little room to differ from fixed weights, so this ablation is inexpensive and informative.
  3. There is no fixed-weight kNN fold-in baseline — a frozen LightGCN with the cold item set to the similarity-weighted mean of its top-KK neighbors' embeddings (layer-wise or mixed). This is the parameter-free counterpart of the proposed method and the natural reference point for every learned component. Likewise there is no kNN edge-augmentation baseline (GraphDA-style) built from the same LLM similarity.
  4. All ablations use a single seed (§4.4), and the yy-axes in Fig. 2 are truncated.

Without items 2–3, the gains in Table 2 cannot be attributed to any of the proposed components.

Questions for authors
  1. Could you report (a) a fixed-weight baseline in which a frozen LightGCN is used and the cold item is set to the cosine-weighted (or uniform) mean of its top-KK warm neighbors' embeddings, and (b) SSB with the learned attention replaced by a softmax over the LLM cosine similarities on the same top-KK set? These would show how much of the gain in Table 2 is due to the learned components.
  2. The borrowing step can be read as a one-way soft edge augmentation of the user–item graph driven by a content kkNN, which is the construction of GraphDA (Fan et al., SIGIR 2023) with a different similarity source. Could you discuss how SSB relates to this line of work and, if possible, compare against a GraphDA-style item–item augmentation built from the same LLM similarity?
  3. Please report the learned layer weights βl\beta_l for all layers on each dataset (currently only β0\beta_0 is given for Beauty and Sports), ideally together with the norms ∥c(l)∥\|c^{(l)}\| so that the effective contribution of each layer can be assessed.
  4. To confirm my reading of §3.5: in the default model (Eq. 6), the MLP and gate of Eq. (5) are not used, and the neighbors' layer-zero embeddings Ej(0)E_j^{(0)} enter only through the attention keys, not through the borrowed values. Is this correct?
1 comment

Response by Authors

We thank the reviewer for the GraphDA comparison and the scrutiny of our ablations. We have narrowed the novelty claim and added the missing controls.

W1. Relation to existing formulations

We agree that content-kNN item-item augmentation is not a new primitive. At the first structural layer, the borrowing read can be expressed as a one-way content-kNN edge augmentation. GraphDA obtains neighbors from pretrained collaborative embeddings, adds a symmetric top-KK item-item block, and retrains LightGCN. LATTICE builds modality-specific item graphs from multimodal features. Cold Brew performs latent neighborhood discovery for isolated nodes in a teacher-student framework.

SSB targets the zero-interaction setting. Its neighbors come from frozen item text, the warm graph is not modified, and the cold item performs a conditional one-way read with separate neighbor and layer weights. We do not attribute every gain to these differences. The novelty is the formulation and operationalization of text-only locate-and-borrow, not the invention of content-kNN augmentation.

W2. Controlled ablations

The original wo_hierarchical variant is a compound intervention. It changes layer-wise reading to a mixed structural embedding, changes whether the neighbors' layer-zero state enters the borrowed value, and replaces the linear fusion in Equation (6) with the MLP and gate in Equation (5). Its drop cannot be attributed solely to hierarchy. We now label this component ablation as a single-seed diagnostic.

Under the same encoder, neighbors, warm graph, pseudo-cold mask, fusion form, and training budget, the Beauty results over five seeds are:

ReadoutRecall@20NDCG@20
Learned attention, two layers15.38 ± 0.507.98 ± 0.26
Uniform average16.25 ± 0.228.66 ± 0.08
Fixed LLM-cosine16.23 ± 0.418.66 ± 0.13
Learned attention, one layer15.52 ± 0.558.01 ± 0.27
Structure-only uniform14.18 ± 0.207.46 ± 0.15

Uniform and LLM-cosine weighting outperform learned attention. The one-layer and two-layer learned variants are nearly tied. Learned attention and hierarchy are therefore not universal sources of gain.

Q1. Fixed-weight and learned-readout comparisons

For the frozen-LightGCN diagnostic, a cold item is the uniform or LLM-cosine mean of its top-KK warm neighbors without learned borrowing or pseudo-cold training. On Beauty with seed 42, these variants obtain 11.59/6.32 and 11.59/6.52 Recall@20/NDCG@20. This diagnostic uses one seed.

Within SSB training, fixed LLM-cosine obtains 16.23/8.66 over five seeds, compared with 15.38/7.98 for learned attention. The gain is therefore not attributable to learned attention.

Q2. GraphDA-style comparison

A true cold item has no collaborative embedding, so we adapt GraphDA with the same frozen LLM similarities, retain its item-item augmentation and retraining procedure, and select the item-item weight on validation data. The selected value is 0.5 on all datasets.

DatasetGraphDA R@20GraphDA N@20SSB R@20SSB N@20
Beauty16.27 ± 0.368.46 ± 0.1815.37 ± 0.507.98 ± 0.26
Sports8.11 ± 0.104.09 ± 0.058.28 ± 0.113.99 ± 0.07
Toys16.13 ± 0.138.99 ± 0.0915.54 ± 0.168.42 ± 0.06

GraphDA-style is stronger on Beauty and Toys. On Sports, SSB has higher Recall@20 while GraphDA-style has higher NDCG@20. We report this dataset-dependent tradeoff and do not claim uniform superiority.

Q3. Learned layer weights

All structural states are L2L_2-normalized, so each layer norm is 1.0 and the effective contribution equals βl\beta_l.

Datasetβ0\beta_0 semanticβ1\beta_1 structuralβ2\beta_2 structuralNorm
Beauty0.3470 ± 0.00820.3811 ± 0.01780.2718 ± 0.00971.0
Sports0.1227 ± 0.00450.7485 ± 0.03440.1288 ± 0.03031.0
Toys0.1574 ± 0.00990.5975 ± 0.03770.2451 ± 0.02781.0

Layer 1 has the largest effective weight on all datasets. This replaces the earlier incorrect statement that Beauty assigns its largest weight to the semantic layer.

Q4. Default model and layer-zero values

The reviewer's reading is correct. The default model uses Equation (6), while Equation (5) and its MLP plus scalar gate are only a single-layer fallback. Neighbor EjE_j is used as the attention key, neighbor Ej(0)E_j^{(0)} does not enter the borrowed value, values start at l=1l=1, and c(0)c^{(0)} is the cold item's semantic prior.

Manuscript changes

We added GraphDA, LATTICE, and Cold Brew to Related Work, narrowed the novelty claim, added the GraphDA-style and fixed-readout comparisons, repaired the hierarchy interpretation, reported all five-seed layer weights, and clarified Equations (5) and (6). The original component ablation and frozen-LightGCN diagnostic remain single-seed.

Reviewer_aXmMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the item cold-start problem in graph-based recommender systems by leveraging content features and augmenting structural information for cold items. Specifically, the proposed approach first uses an LLM without fine-tuning to identify similar warm items based on the content features of cold items. It then constructs structural information for cold items through a soft-structure borrowing mechanism. To provide training signal, a pseudo-cold mask treats a fraction of warm items as cold during training. The authors evaluate the proposed method against various baselines using Recall@K and NDCG@K. The results demonstrate improvements over existing methods on some of the evaluated datasets.

Critical strengths
  1. The authors formulate item cold-start as a locate-and-borrow problem, providing a clear framework for reconstructing the missing structural information of cold items. The approach does not require LLM fine-tuning, which simplifies deployment.

  2. The experiments evaluate cold and warm items separately across different models, offering useful insights into how these methods perform on each item group.

  3. The paper is clearly written and easy to follow, with an accessible presentation of both the motivation and the proposed approach.

Critical weaknesses
  1. The proposed method uses an attention-weighted mechanism to transfer information from the content space to the LightGCN space, but the paper lacks comparisons with simpler alternatives or an in-depth analysis of their limitations. Consequently, it remains unclear whether the gains come from the attention, gating, and hierarchical components or primarily from the semantic information provided by the LLM.

  2. SSB does not consistently outperform the baselines across datasets, particularly on Beauty. If warm items have richer structural information, one would expect the borrowed structure to be of higher quality. However, the paper does not adequately explain why this appears to benefit CGRC more than SSB or clarify the mechanism underlying this difference.

  3. The paper makes claims such as semantically similar items tending to connect to the same users in sparse datasets, but provides no quantitative evidence to support them. Such claims could be directly examined by measuring the user overlap between a cold item’s top-K semantic neighbors and the cold item itself. Without this analysis, the proposed explanation remains an unsupported assertion.

  4. Figure 1 is unclear and does not effectively illustrate the model architecture or the flow of data information between its components, making the proposed framework difficult to understand.

Questions for authors
  1. How does SSB compare with a simpler baseline that averages the top-K neighbors’ LightGCN embeddings instead of using attention-weighted aggregation?

  2. Could the authors provide a more in-depth justification for the proposed design choices? For example, the paper states that structural information may contain noise, but how does this noise affect recommendation performance, and to what extent does the proposed method mitigate its effects compared with simpler alternatives?

  3. Could the authors investigate and address the performance limitations on Beauty? If the proposed method consistently outperforms the baselines across the reported metrics on this dataset, I would consider raising my score.

1 comment

Response by Authors

We thank the reviewer for identifying the attribution problem, the unsupported shared-user statement, and the unclear architecture figure. We have revised the claims and added controlled analyses.

W1. Sources of the gain

We agree that the original paper placed too much emphasis on learned attention and gating. We retrained all variants with the same frozen LLM encoder, semantic neighbors, warm graph, pseudo-cold mask, optimization objective, and training budget. Table 1 reports five-seed Beauty results.

Table 1. Neighbor-readout and controlled hierarchy results on Beauty.

MethodRecall@20NDCG@20
Learned attention, two layers15.38 ± 0.507.98 ± 0.26
Uniform average16.25 ± 0.228.66 ± 0.08
Fixed LLM-cosine16.23 ± 0.418.66 ± 0.13
Learned attention, one layer15.52 ± 0.558.01 ± 0.27
Structure-only, uniform14.18 ± 0.207.46 ± 0.15

Uniform weighting improves Recall@20 and NDCG@20 over learned attention by 0.87 and 0.68 percentage points, with paired tests at p≈0.003p \approx 0.003 and p≈0.002p \approx 0.002. Fixed LLM-cosine is also better, with p≈0.008p \approx 0.008 and p≈0.006p \approx 0.006. The structure-only result remains substantially lower, which isolates semantic-guided access to warm structure from learned attention.

Equation (5) is now explicitly identified as the single-layer MLP-plus-gate fallback and is not used in the default hierarchical model in Equation (6). Learned attention is presented as one optional neighbor readout. The revised contribution is semantic-guided structural borrowing with configurable neighbor and layer readouts, not a claim that learned attention is universally superior.

W2. Performance on Beauty

The original learned-readout model does not dominate CGRC on Beauty. Over five seeds, CGRC obtains 15.67 ± 0.31 Recall@20 and 7.55 ± 0.27 NDCG@20, while SSB obtains 15.37 ± 0.50 and 7.98 ± 0.26. CGRC also has stronger warm performance. We now state this limitation directly.

This difference is consistent with the two designs. CGRC reconstructs item-specific edges and exploits the dense Beauty graph directly. SSB combines several warm structural states, giving higher cold NDCG but potentially losing item-specific recall. We no longer infer that richer warm structure makes semantic borrowing uniformly better.

Uniform structural borrowing reaches 16.25 Recall@20 and 8.66 NDCG@20 on Beauty, exceeding CGRC on both cold metrics. We do not extend this claim to all warm metrics.

W3. Semantic-neighbor user alignment

We have removed the statement that semantically similar items generally connect to the same users. Table 2 reports exact and soft alignment against degree-matched random neighbors.

Table 2. Exact and soft semantic-user alignment.

DatasetExact overlap, semantic/randomUser-centroid cosine, semantic/random
Beauty2.61% / 0.096%0.124 / 0.014
Sports0.44% / 0.017%0.152 / 0.002
Toys3.22% / 0.025%0.207 / 0.006

The supported statement is that semantic neighbors provide a relatively enriched but absolutely sparse signal about the cold item's audience. They do not determine an identical user set.

W4. Figure 1

We have expanded the caption and Section 3.2 with a step-by-step description of text encoding, semantic retrieval, layer-wise structural states, configurable borrowing, and scoring. The training-only pseudo-cold path is now distinguished from inference. No synthetic edge is used at inference.

Q1. Comparison with top-KK averaging

The comparison is given in the W1 table. Uniform averaging improves both cold metrics over learned attention, so semantic-guided structural borrowing remains effective without learned attention.

Q2. Structural noise

At 25% corruption on Beauty, NDCG@20 changes from 7.97 to 6.60 with learned attention, from 8.72 to 7.96 with uniform weighting, and from 7.57 to 6.12 with structure-only aggregation. At 100% corruption all variants degrade severely. These results do not support a general claim that learned attention suppresses noisy neighbors. We remove that claim and describe fixed weighting as a lower-variance readout.

Q3. Beauty limitation

The learned-readout limitation is now explicit. Uniform structural borrowing still exceeds CGRC on both cold metrics over five seeds. The borrowing mechanism remains effective, while the original learned-readout instantiation is not uniformly best. We do not claim superiority across every reported metric.

Manuscript changes

  • Contributions center on semantic locating and structural borrowing.
  • Learned attention is optional, and Equation (5) is identified as a fallback.
  • The Beauty limitation and sparse semantic-user alignment are stated explicitly.
  • The Figure 1 data flow and corruption analysis have been clarified.

Comments on the paper

1 comment

Response by Authors

We thank the reviewer for the two soundness concerns on pseudo-cold graph visibility and the attribution of learned attention. We agree and have revised the claims and analyses.

W1 and Q1. Pseudo-cold graph visibility

We agree that retaining pseudo-cold edges is not equivalent to using a zero-degree item. Training excludes the target from its own neighbor retrieval and uses no cold validation or test edge. Retained training edges can nevertheless propagate its embedding through users and multi-hop neighbors, so the original exact-task equivalence claim was too strong.

To isolate this effect, we remove every edge incident to a pseudo-cold target during the training epoch and restore the complete warm graph before validation and testing. Table 1 reports the five-seed Beauty control.

Table 1. Beauty graph-visibility control.

Training graphCold R@20Cold N@20Warm R@20Warm N@20
Retain pseudo edges15.38 ± 0.507.98 ± 0.268.87 ± 0.063.75 ± 0.04
Excise pseudo edges16.77 ± 0.298.84 ± 0.088.65 ± 0.143.56 ± 0.03

Edge excision improves cold Recall@20 by 1.39 points and cold NDCG@20 by 0.86 points, with paired p≈0.0015p \approx 0.0015 and p≈0.0017p \approx 0.0017. It also improves rare Recall/NDCG@20 from 13.44/6.63 to 16.25/7.82 but reduces warm performance. Thus the cold result does not rely on indirect information from retained edges. Retained edges mainly preserve a stronger warm representation. The revised manuscript no longer equates pseudo-cold training with a deployment-time isolated graph.

W2 and Q2. Readout and fusion

We agree that wo_hierarchical is a compound intervention. It changes layer-wise reading to the mixed structural embedding, changes whether neighbor layer zero enters the borrowed value, and switches Equation (6)'s linear combination to Equation (5)'s MLP with a scalar gate. Its full drop cannot be attributed to multi-layer aggregation alone. Equation (5) is now described as a single-layer fallback.

We compare neighbor readouts while fixing the semantic encoder, neighbors, warm graph, pseudo-cold mask, layer combination, and training budget. Table 2 reports five-seed Beauty results.

Table 2. Beauty readout and controlled hierarchy results.

MethodRecall@20NDCG@20
Learned attention, two layers15.38 ± 0.507.98 ± 0.26
Uniform average16.25 ± 0.228.66 ± 0.08
Fixed LLM-cosine16.23 ± 0.418.66 ± 0.13
Learned attention, one layer15.52 ± 0.558.01 ± 0.27
Structure-only, uniform14.18 ± 0.207.46 ± 0.15

Uniform weighting improves Recall@20 and NDCG@20 over learned attention by 0.87 and 0.68 points, with p≈0.003p \approx 0.003 and p≈0.002p \approx 0.002. Fixed LLM-cosine is also better, with p≈0.008p \approx 0.008 and p≈0.006p \approx 0.006. The one-layer and two-layer learned variants are nearly tied. Learned attention is therefore presented as an optional, task-dependent readout. The stable contribution is semantic locating plus access to warm structural states, as indicated by the lower structure-only result.

Q3. Joint warm-cold ranking

We add a joint-ranking diagnostic in which all warm and cold items form one candidate pool for warm users with test interactions. Train and validation history is excluded, and ground truth is the union of warm and cold test interactions. Table 3 reports the principal results.

Table 3. Joint ranking at K=20K=20.

DatasetMethodOverall RCold RCold NCold share
BeautySSB7.939.585.2725.0%
BeautyCGRC5.585.332.4417.8%
BeautyGraphDA8.526.302.7811.8%
SportsSSB2.915.502.6437.4%
SportsCGRC3.032.231.0617.9%
SportsGraphDA3.852.311.099.6%
ToysSSB6.049.615.4125.4%
ToysCGRC4.014.872.5316.2%
ToysGraphDA5.894.952.379.6%

Cold exposure and relevance improve substantially on all datasets, and SSB gives the best overall ranking on Toys. On Beauty and especially Sports, overall performance remains below stronger graph-augmentation baselines. On Sports, warm NDCG@20 is 1.75 versus 2.71 for CGRC and 2.89 for LightGCN. We therefore report the cold benefit without claiming universal superiority. CGRC uses seed 42 on Beauty and seed 123 on Sports and Toys, while SSB and GraphDA use seed 42. We label this comparison diagnostic and do not use it for seed-level significance.

Q4. Rare-item candidate pool

The ≤2\leq 2 analysis changes only the ground-truth subset. All cold items remain candidates. Every true cold item has zero training interactions. The ≤2\leq 2 label refers to withheld validation or test interactions, not to training degree.

Manuscript changes

We incorporate these changes in the revision.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.