TLDR: DiffusionGS is a novel generative search model that leverages diffusion processes and user queries to accurately extract immediate user interests from their historical behaviors. Developed by researchers at Kuaishou, it introduces a User-aware Denoising Layer (UDL) to integrate user-specific profiles, dynamically adjusting attention for personalized ranking. The model has demonstrated superior performance over existing deep learning and generative ranking methods in extensive offline experiments and achieved significant improvements in key metrics during online A/B tests on the Kuaishou E-commerce platform, proving its effectiveness in real-world applications.
Personalized search ranking systems are vital for modern platforms like e-commerce sites and short-video apps, as they directly impact how users engage and how much revenue is generated. While current methods are good at understanding a user’s general interests from their past actions, they often don’t fully connect a user’s immediate search query with their historical behaviors.
A new approach called DiffusionGS, developed by researchers at Kuaishou Technology and City University of Hong Kong, aims to solve this problem. This novel and scalable method, powered by generative models, uses user queries as explicit ‘intent anchors’. These anchors help to pinpoint a user’s immediate interests from a vast and often noisy history of past actions. The core idea is to treat interest extraction as a conditional denoising task, where the user’s query guides a special diffusion process. This process creates a strong, user-intent-aware representation from their sequence of past behaviors.
DiffusionGS also introduces a ‘User-aware Denoising Layer’ (UDL). This layer incorporates specific user profiles into how the system pays attention to a user’s past actions, optimizing the attention distribution. By reframing queries as initial intent signals and using diffusion-based denoising, DiffusionGS offers a powerful way to capture how user interests change dynamically. Extensive tests, both offline and online, have shown that DiffusionGS performs better than existing state-of-the-art methods.
How DiffusionGS Works
The model takes a user’s historical behavior sequence as input, which includes both positive and negative engagements. Each behavior is represented by an item ID and its associated details like title, category, and price. To capture how user interests evolve over time, each behavior token is enhanced with a timestamp and a positional embedding.
The ‘diffusion process’ is inspired by models that reconstruct structured data from random noise. DiffusionGS uses a Denoising Diffusion Probabilistic Model (DDPM) to separate static item features from dynamic user interests. It gradually adds noise to a clean input sequence, then iteratively denoises it, guided by the user’s query. This process helps to recover clean, dynamic interest representations, improving the prediction of future behaviors. The goal here is to estimate dynamic user interest distributions, not necessarily to generate perfect item tokens, so moderate noise levels are sufficient.
For ‘condition-guided generation’, DiffusionGS uses ‘additive conditioning’ to inject query information. This means the query’s encoding is added directly into the model’s intermediate feature maps. This method was chosen because it has a more direct and pervasive influence on the model’s components, leading to better performance compared to simply concatenating queries or using cross-attention.
The ‘User-aware Denoising Layer’ (UDL) is a key innovation. Unlike traditional attention layers that apply uniform weights, UDL introduces a gating mechanism. This mechanism uses user features (like age and gender) and contextual information (like timestamp and search scenario) to dynamically adjust how much attention the model pays to each item in a user’s history. This allows for truly personalized ranking, where different users searching for the same item might prioritize different features based on their profiles.
Training and Results
The model’s training objective combines two goals: ensuring the generated representations accurately reflect the original input (using a KL divergence loss) and predicting user engagement (click-through rate or conversion rate) for target items (using a binary cross-entropy loss). This combined approach balances generative quality with predictive performance.
DiffusionGS was evaluated on a large-scale, real-world dataset from the Kuaishou E-commerce platform, which includes authentic search queries, rich user behavior histories, and detailed side information. It consistently outperformed several deep learning and generative ranking models in predicting both click-through rates (CTR) and conversion rates (CVR).
Internal analyses confirmed that the dynamic user interest captured by DiffusionGS is indeed guided by query signals. The study also validated that additive conditioning is the most effective way to inject query information, and that the User-aware Denoising Layer significantly enhances personalized ranking. Furthermore, the research showed that moderate noise levels in the diffusion process are optimal, and that the model’s performance improves as its capacity (number of UDL layers) increases, demonstrating a ‘scaling law’.
Also Read:
- Improving Product Search in Multi-Turn Conversations with Dynamic Reranking
- Spacetime-GR: A Generative Model for Context-Aware Online POI Recommendations
Online Deployment and Future Vision
The DiffusionGS model has been successfully deployed in the ranking stage of the Kuaishou E-Commerce Search Platform. For online deployment, where speed is crucial, the model uses a single denoising step instead of the full 50 steps used in offline experiments. This simplification results in a negligible performance drop (about 0.01% decrease in AUC) while significantly boosting inference efficiency.
An online A/B test conducted over 14 days, involving billions of user requests, showed significant improvements. DiffusionGS led to increases in total gross merchandise value (#GMV), total orders (#Total Orders), online CTR, and CVR compared to the base model. This demonstrates its practical value and ability to provide a more engaging shopping experience for users.
The researchers envision extending DiffusionGS towards a fully generative paradigm in the future, where the system directly models the user-to-item generation process. This could potentially eliminate the need for multi-stage pipelines, leading to a more cohesive and adaptive search and recommendation framework. For more details, you can read the full research paper here.
