The generative recommendation storyline (see Sections 1.1, 5.3, and 9.x) has shown us that models can "directly generate" item sequences, or even "think before recommending." But generative capability can also solve more practical engineering pain points — data sparsity, missing features, and result homogenization. These are exactly the problems industrial recommender systems have struggled with for years.
Diffusion models offer a unique set of tools for these problems, thanks to their generative paradigm of "gradually adding noise, then learning to remove it." They neither score items like discriminative models nor generate token by token like autoregressive models; instead, they "sculpt" the target in a continuous latent space through multi-step denoising. This part unfolds along two main threads: data augmentation (using diffusion to generate high-quality pseudo-interactions or cross-scenario samples) and feature augmentation and diversity optimization (using diffusion to fill in missing features and generate diverse slates).
What This Part Covers
| Section | Topic | The Big Idea |
|---|---|---|
| 10.1 | Diffusion model basics | Forward noising / reverse denoising, DDPM, latent diffusion, conditional generation and two guidance strategies; special designs for recommendation (noise scale, x₀-prediction) |
| 10.2 | Diffusion for data augmentation | DiffuASR generates "prequel" sequences to extend short-history users; Diff-MSR transfers knowledge across scenarios to ease cold start |
| 10.3 | Diffusion applications in recommendation | AsymDiffRec uses asymmetric diffusion to complete missing features; DMSG uses conditional diffusion to generate diverse slates |
What You'll Be Able to Do After This Part
- 🟢 Explain the two inverse Markov processes of forward diffusion and reverse denoising, and write the direct sampling formula for any t
- 🟢 Distinguish data-space diffusion from latent-space diffusion, and explain why recommendation prefers the latter
- 🟡 Describe the difference between ε-prediction and x₀-prediction, and why recommendation often uses the latter
- 🟡 Recount how DiffuASR / Diff-MSR / AsymDiffRec / DMSG each use diffusion to solve a specific pain point
- 🔴 Critically assess the applicability boundaries of diffusion models in recommendation (latency, supporting infrastructure) and future directions
- Complete the tiered practice problems in each section to consolidate the through-line from basics to deployment
Core Concepts
| Concept | Section | Relevance |
|---|---|---|
| Forward / reverse process | 10.1 | The core inverse mechanism pair of diffusion models |
| Latent diffusion | 10.1 | Recommendation favors LDM due to high dimensionality and sparsity |
| Conditional generation + guidance | 10.1 | Steer generation with user history or text |
| Sequence / cross-scenario augmentation | 10.2 | Eases data sparsity and cold start |
| Asymmetric diffusion / slate generation | 10.3 | Feature completion and diversity optimization |
Prerequisites
- Read Sections 1.1 (two paradigms) and 5.3 (the evolution of the generative paradigm, especially semantic IDs and end-to-end generation)
- Basic probability, variational autoencoders (VAE), and Transformer attention
This part leans toward engineering applications. The math is about "why it is designed this way" — no need to derive every line; just grasp the pain point each method targets.
Tips for This Part
- Always read with the pain point in mind. Each diffusion method maps to a concrete engineering problem (sparsity / cold start / missingness / homogenization).
- Keep the "spaces" straight. Data space vs latent space, forward space vs reverse space — pick the wrong space and the design goes wrong.
- Don't treat diffusion as a replacement for discriminative models. The methods here are tool-style generative capabilities (augmenting data / features / diversity), not an end-to-end replacement for ranking.
Let's dive in! 🚀