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

Technical Summary: DiffCold

Problem Statement
Cold-start item recommendation remains a persistent challenge in real-world systems due to the absence of historical interaction data for new items. Existing approaches, ranging from dropout-based methods to generative and alignment-based models, universally suffer from a "seesaw dilemma." Strategies designed to improve predictions for cold items often degrade the performance for warm (existing) items, and vice versa. The authors identify the root cause of this dilemma as a fundamental distributional disparity: warm item embeddings occupy a complex "behavioral manifold" shaped by rich interaction signals, whereas cold item embeddings are constrained to a "semantic manifold" derived solely from auxiliary content features. Previous methods fail because they attempt to force a rigid mapping between these inconsistent spaces, sacrificing the precision of warm representations to accommodate cold ones.

Methodology: DiffCold
To resolve this, the authors propose DiffCold, a diffusion-based generative framework that unifies warm and cold representations. Unlike direct regression or adversarial approaches (GANs/VAEs), DiffCold leverages the conditional diffusion process to reconstruct warm item embeddings from content features, preserving the underlying manifold structure without degradation. The framework consists of three core components:

1. Forward-Reverse Process:

   • Training: The model learns to reconstruct warm item ID embeddings ($e^0_i$) from noisy versions ($e^t_i$) using content features ($c_i$) as conditional guidance. The forward process injects Gaussian noise, and the reverse process learns a denoising function $f_\theta$ to predict the original embedding.
   • Inference: For cold items, the model generates embeddings by initiating from a starting distribution and iteratively refining them through the learned denoising process.

2. Retrieval-enhanced Aggregator (Solving the Starting-Point Problem):

   • Directly using pure Gaussian noise as the starting point for inference is shown to damage personalized information.
   • DiffCold introduces a mechanism where, for a new cold item, the system retrieves the $k$ most similar warm items based on the cosine similarity of their content features.
   • The ID embeddings of these retrieved warm items are aggregated (via mean pooling) to form a robust initial representation ($\bar{e}^0_i$) for the denoising process, effectively harnessing semantic information to bypass inefficient noise.

3. Simulation-based Representation Alignment (Solving the Distribution-Consistency Problem):

   • To ensure the generated cold embeddings align with the true distribution of warm embeddings, the authors propose a simulation strategy during training.
   • Since cold items are unavailable during training, the model uses warm items to simulate the cold generation process. It generates a simulated starting point for a warm item using the Retrieval-enhanced Aggregator and then applies the forward-reverse diffusion process.
   • A contrastive learning loss (InfoNCE) is applied to align these simulated representations with the actual warm item representations. This enforces distributional consistency between the generated cold embeddings and the real interaction space.

Training Objective
The model is optimized using a multi-task loss function ($L_{overall}$) that balances:

• Reconstruction Loss ($L_{recon}$): Ensures the diffusion model can accurately reconstruct original embeddings.
• Cold Recommendation Loss ($L_{bpr-cold}$): Optimizes the generated cold embeddings for ranking tasks using Bayesian Personalized Ranking (BPR).
• Simulation Alignment Loss ($L_{sim-align}$): Aligns simulated representations with real warm representations via contrastive learning.
• Warm Recommendation Loss ($L_{bpr-warm}$) & Warm Augmentation ($L_{warm-aug}$): Ensures the model maintains high performance on warm items and aligns user-warm item distributions.

Key Contributions

1. Novel Perspective: The paper attributes the seesaw dilemma to the rigid coupling of semantic and behavioral manifolds and introduces the first diffusion-based framework specifically tailored to resolve this disparity.
2. Retrieval-enhanced Aggregator: A design that leverages semantic neighbors to provide a robust initialization for the denoising process, solving the starting-point problem.
3. Simulation-based Representation Alignment: A module that simulates the cold generation process using accessible warm items during training and aligns representations via contrastive learning to ensure distributional consistency.
4. SOTA Performance: Extensive experiments demonstrate that DiffCold resolves the seesaw dilemma, consistently outperforming state-of-the-art methods across all metrics (Overall, Cold, and Warm) on three datasets.

Experimental Results
The authors evaluated DiffCold on three benchmark datasets: Movielens, Citeulike, and Xing, using three backbone models (MF, LightGCN, SimGCL).

• Performance: DiffCold consistently achieved the highest Recall@20 and NDCG@20 scores across all categories (Overall, Cold, Warm). For instance, on the Movielens dataset with an MF backbone, DiffCold improved the Overall Recall@20 by 15.20% over the best baseline.
• Seesaw Resolution: Unlike baselines (e.g., Heater, GAR, ALDI) which showed trade-offs where improving cold performance hurt warm performance, DiffCold achieved simultaneous improvements.
• Distribution Analysis: Quantitative analysis using Wasserstein Distance and Maximum Mean Discrepancy (MMD) confirmed that DiffCold achieves the lowest divergence between cold and warm item distributions, validating the effectiveness of the alignment module.
• Efficiency: While DiffCold requires slightly more memory than some baselines, it maintains competitive training times per epoch and converges faster than complex alignment-based methods like CLCRec.

Significance
The paper claims that DiffCold represents a fundamental shift in addressing cold-start recommendation by moving away from rigid mapping or adversarial approaches. By utilizing the progressive destruction and reconstruction capabilities of diffusion models, DiffCold successfully decouples the preservation of warm item fidelity from the generation of cold items. The work validates that resolving the distributional disparity between semantic and behavioral manifolds is key to overcoming the seesaw dilemma, offering a robust solution that improves recommendation quality for both new and existing items without compromise.