DiffCold: A Diffusion-based Generative Model for Cold-Start Item Recommendation

Abstract: Cold-start item recommendation remains a persistent challenge in real-world systems due to the absence of interaction histories. While prior models attempt to bridge this gap using item content features, they universally suffer from the \textbf{seesaw dilemma}: enhancing performance for cold items inevitably degrades performance for warm items, and vice versa. We identify that this dilemma stems from a fundamental \textbf{distributional disparity}: warm item embeddings occupy a complex behavioral manifold" shaped by rich interaction signals, whereas cold item embeddings are constrained to asemantic manifold" derived solely from auxiliary content. Existing methods often force a rigid mapping between these inconsistent spaces, causing the model to sacrifice the precision of warm representations to accommodate cold ones. To address this, we propose \textbf{DiffCold}, a diffusion-based generative model that unifies warm and cold representations. Unlike GANs or VAEs, DiffCold leverages conditional diffusion to reconstruct warm item embeddings from content, preserving the underlying manifold structure without degradation. We further tailor this paradigm with two specific designs: a \textbf{Retrieval-enhanced Aggregator} that initializes generation using semantically similar warm items to bypass inefficient noise, and a \textbf{Simulation-based Representation Alignment} module that enforces distribution consistency between generated and real embeddings via contrastive learning. Experiments on three benchmarks confirm that DiffCold resolves the seesaw dilemma, consistently outperforming state-of-the-art methods across all metrics.

• The paper introduces DiffCold, a diffusion-based model that improves overall, cold-item, and warm-item recommendation performance simultaneously, including up to 23.61% higher Recall@20 on Citeulike with SimGCL.
• DiffCold retrieves similar warm items to initialize denoising and uses simulation-based contrastive alignment to bridge semantic and behavioral embedding spaces without relying on unavailable cold-item interactions.
• Experiments across three datasets and three recommendation backbones show that DiffCold mitigates the cold-start seesaw dilemma while maintaining competitive memory use and training efficiency.

DiffCold addresses cold-start item recommendation by reframing it as a conditional generative problem over item ID embeddings. The paper's central diagnosis is that prior approaches—dropout-based, generative, and alignment-based alike—suffer from a "seesaw dilemma": improving cold-item performance degrades warm-item performance. The authors attribute this to a distributional disparity between warm items, whose embeddings lie on an interaction-driven "behavioral manifold," and cold items, whose embeddings are confined to a content-driven "semantic manifold." Methods that force rigid mappings or alignments between these spaces necessarily sacrifice one side (2606.12245).

Method

DiffCold is built on a DDPM-style forward-reverse process applied to pre-trained warm item ID embeddings, with content features serving as conditioning signals in the reverse process. During training, Gaussian noise is injected into positive-item embeddings under a linear variance schedule with a controllable noise scale; a denoising network $f_\theta$ directly predicts the clean embedding $\hat{e}_0^i$, trained with a reconstruction loss plus a BPR loss on the denoised representation against sampled negatives.

Two components tailor this paradigm to cold-start settings:

• Retrieval-enhanced Aggregator: rather than starting denoising from pure Gaussian noise—which prior work (DiffRec, DDRM) shows destroys personalization information—the model retrieves the $k$ most similar warm items via cosine similarity of content features ($k=10$) and mean-aggregates their embeddings as the inference starting point.
• Simulation-based Representation Alignment: because cold items are unavailable during training, DiffCold simulates their generation path using warm items: it constructs simulated noisy inputs from the aggregated retrieval-based starting point, denoises them, and aligns the resulting representations with real warm embeddings via InfoNCE loss. This explicitly enforces distributional consistency between generated and real embedding spaces during training itself.

A multi-task objective combines reconstruction, cold-side BPR and simulation alignment, and warm-side BPR plus a user–item contrastive alignment term ("warm-aug") that counteracts distributional drift of warm embeddings caused by cold-oriented optimization. At inference, each cold item's aggregated semantic start is noised to step $T$ and iteratively denoised conditioned on its content features.

Empirical results

Experiments use three benchmarks (Movielens, Citeulike, Xing) with 20% of items held out as cold, evaluated under three backbones (MF, LightGCN, SimGCL) against dropout-based (DropoutNet, Heater), generative (MetaEmb, GAR, GoRec), and alignment-based (CLCRec, ALDI, PAD-CLRec) baselines using Recall@20 and NDCG@20.

The headline claim is that DiffCold is the only method improving overall, cold, and warm performance simultaneously across all datasets and backbones. Representative improvements over the best baseline include 22.55% overall Recall@20 on Citeulike (MF backbone), 23.61% on Citeulike (SimGCL), 15.20% on Movielens (MF), and 11.16% on Xing (SimGCL). Notably, gains extend to warm items (e.g., +9.77% NDCG@20 warm on Movielens/MF), which is precisely where baselines fail: GoRec and DropoutNet show severe warm degradation, while CLCRec and ALDI leave warm performance unchanged relative to the backbone.

Supporting analyses address the two design problems directly. For the starting-point problem, replacing the aggregator with Gaussian noise, projected content features, or KNN cluster centroids (100/500 centroids) consistently underperforms retrieval aggregation, confirming that pure-noise initialization impairs personalization and that clustering is hyperparameter-sensitive. For distribution consistency, DiffCold achieves the lowest intra-cold/intra-warm cosine similarity gap and near-optimal Wasserstein distance and MMD versus Heater, GAR, and ALDI, with t-SNE visualizations showing integrated user/warm/cold distributions where baselines exhibit clear separation.

Ablations confirm all five loss terms contribute: removing the Aggregator, Sim-Align, or BPR-Cold primarily hurts cold performance, while removing BPR-Warm or Warm-Aug primarily hurts warm performance—an asymmetry consistent with the seesaw framing. Efficiency is competitive: DiffCold uses modest memory (1199 MB vs. 1741 MB for CLCRec on Movielens) and per-epoch time comparable to Heater and GAR, converging in fewer epochs than GAR. Hyperparameter sensitivity is reported as mild within reasonable ranges, and a linear schedule with $T=20$ steps is optimal; longer schedules increase cost without improving accuracy, and cosine/binomial schedulers degrade performance substantially.

Limitations and open questions

Several constraints qualify the results. The framework requires pre-trained warm embeddings from a fixed backbone, so its ceiling is tied to backbone quality, and evaluation covers only MF, LightGCN, and SimGCL. The Retrieval-enhanced Aggregator assumes content-feature similarity is a reliable proxy for behavioral similarity—a strong assumption for sparse-content domains—and the simulation-based alignment trains only on warm items, leaving open whether simulated generation faithfully matches true cold-item generation when cold items arrive with out-of-distribution content. The method also targets item cold-start only; user cold-start is not addressed. Finally, the diffusion process operates on relatively small embedding dimensions with short schedules ($T=20$); scalability to industrial-scale catalogs and interaction with streaming item arrival remain unexamined.

Conclusion

DiffCold contributes a diffusion-based generative formulation of cold-start item recommendation that decouples warm-manifold preservation from cold-embedding generation, supported by retrieval-based initialization and simulation-based contrastive alignment. Its empirical claim—that simultaneous improvement across overall, cold, and warm metrics resolves the seesaw dilemma—is substantiated across three datasets, three backbones, and eight baselines, with quantitative distributional evidence.