💻 computer science

Intent-Guided Diffusion with Adaptive Intent Fusion for User Cold-Start Recommendation

This paper proposes the Intent-Guided Diffusion Model (IGDM), a novel framework that enhances user cold-start recommendation by extracting transferable latent intent prototypes from historical data, adaptively fusing them with sparse user behaviors, and guiding a diffusion-based denoising process to refine preference estimation, thereby outperforming existing baselines on multiple real-world datasets.

Original authors: Lingxuan Li, Luofei Jia, Dafei Lin, Chongmin Wang, Yongxin Shi

Published 2026-06-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lingxuan Li, Luofei Jia, Dafei Lin, Chongmin Wang, Yongxin Shi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you just moved to a new city and joined a local book club. You haven't read many books there yet, and you haven't told the club leader much about your taste. The leader needs to recommend a book for you, but they have almost no information to go on. This is the "Cold-Start Problem" in the world of digital recommendations: how do you guess what a new user likes when they have barely left any digital footprints?

Most current systems try to guess by looking at what you did click (even if it's just one or two things) or by asking you to fill out a profile. But if you haven't clicked much, or if you don't want to share your profile, these systems often fail. They might guess wildly because they are working with too little data.

This paper introduces a new solution called IGDM (Intent-Guided Diffusion Model). Here is how it works, broken down into simple steps using analogies:

1. The "Library of Hidden Themes" (Intent Extraction)

Before trying to guess what you specifically like, the system looks at thousands of other users who have been around longer. It notices patterns.

The Analogy: Imagine the system is a librarian who has noticed that people who buy hiking boots, read survival guides, and watch nature documentaries often share a hidden "love for the outdoors." Even if they never said "I love hiking," their actions reveal this Intent.
What the paper does: It groups users into clusters based on these hidden patterns and creates "Intent Prototypes." Think of these as archetypes or "personality types" of users (e.g., "The Traveler," "The Foodie," "The Tech Geek").

2. The "Smart Matchmaker" (Adaptive Intent Fusion)

Now, the system looks at you, the new user. You only clicked on one thing: a picture of a mountain.

The Problem: One click isn't enough to know if you are a "Hiker" or just someone who likes "Scenic Views."
The Analogy: Instead of guessing based solely on that one click, the system acts like a smart matchmaker. It says, "This new user clicked on a mountain. That looks a lot like the 'Hiker' archetype we found earlier. Let's blend your tiny bit of data with the 'Hiker' profile."
What the paper does: It takes your limited data and mixes it with the "Intent Prototypes" that match you best. It doesn't just copy the prototype; it weighs them. If you look 70% like a Hiker and 30% like a Tech Geek, it creates a hybrid representation of you that is much richer than your single click alone.

3. The "Denoising Filter" (Intent-Guided Diffusion)

Even with the hybrid profile, your data is still "noisy" and incomplete because you are a new user. It's like trying to hear a song through a static-filled radio.

The Analogy: Imagine you have a blurry, low-resolution photo of a face. A standard system might just guess the features. The IGDM system uses a Diffusion Model. Think of this as a high-tech photo restoration tool.
- The Process: The system intentionally adds "static" (noise) to your profile to see what happens, then it uses the "Intent Match" (the Hiker profile) as a guide to slowly remove that static.
- The Result: It iteratively cleans up your profile, removing the random errors and filling in the gaps based on what a "Hiker" usually likes. It's like sculpting a statue: you start with a rough block of stone (noisy data) and chip away the excess until the true shape (your real preference) emerges, guided by the blueprint (the intent).

Why This is Better (The Results)

The authors tested this on three real-world datasets (movies, music, and local business reviews).

The Claim: When users have very little data (the "severe cold-start" scenario), IGDM consistently outperforms older methods.
The Takeaway: By using the "wisdom of the crowd" (the Intent Prototypes) to guide the cleaning process (Diffusion), the system can make much better guesses about new users than systems that try to guess based only on the tiny amount of data the new user provided.

In Summary

The paper argues that to help a new user, you shouldn't just look at their tiny footprint. Instead, you should:

Identify what kind of user they likely are by comparing them to established groups.
Blend their small data with those group patterns.
Refine that blend using a mathematical process that strips away uncertainty, leaving a clear picture of what they probably want.

The paper concludes that this method is particularly powerful when data is scarce, making it a strong tool for welcoming new users to recommendation platforms.

Technical Summary: Intent-Guided Diffusion with Adaptive Intent Fusion for User Cold-Start Recommendation

1. Problem Definition

The paper addresses the user cold-start recommendation problem, a fundamental challenge in recommender systems where new users have extremely limited historical interaction data (e.g., fewer than 20 interactions). In such scenarios, behavioral signals are sparse and incomplete, making it difficult to accurately model user preferences.

Existing approaches face specific limitations:

Auxiliary Information Methods: Rely on user profiles or item attributes, which are often unavailable or restricted by privacy constraints.
Meta-Learning Methods: Learn transferable knowledge across tasks but often become unstable under extremely sparse conditions where the limited support interactions are insufficient for reliable adaptation.
Intent-Aware Methods: Typically designed for sequential recommendation with sufficient history, failing to leverage transferable intent patterns for users with almost no history.
Diffusion-Based Methods: While effective at modeling complex distributions, current approaches often ignore latent intent information, limiting their ability to recover reliable preferences from noisy, sparse cold-start data.

The core challenge identified is how to exploit limited personalized behaviors while simultaneously introducing transferable intent-level knowledge to compensate for sparse observations and progressively recover preference-consistent user representations.

2. Methodology: Intent-Guided Diffusion Model (IGDM)

The authors propose IGDM, a framework integrating intent modeling with diffusion-based representation refinement. The model consists of three primary components:

A. Intent Extraction Module

To capture transferable preference patterns, the model extracts latent intent prototypes from historical interaction sequences of existing (non-cold-start) users.

User interaction sequences are encoded into dense representations.
K-means clustering is applied to partition users into $K$ clusters based on sequence similarity.
Each cluster represents a latent intent type, and its intent prototype ( $I_j$ ) is defined as the mean representation of users within that cluster. These prototypes serve as transferable intent priors.

B. Adaptive Intent Fusion Module

This module dynamically integrates the limited behavioral features of a cold-start user with the extracted intent prototypes.

The similarity between the cold-start user's representation ( $S_c$ ) and each intent prototype ( $I_j$ ) is computed.
Adaptive Weights: A softmax function generates normalized weights ( $\omega_{cj}$ ) based on these similarities, controlled by a temperature parameter $\tau$ .
Fused Representation: The final fused representation ( $z_c$ ) is a weighted combination of the user's own features and the intent prototypes:
$z_c = \alpha S_c + (1 - \alpha) \sum_{j=1}^{K} \omega_{cj} I_j$
Here, $\alpha$ balances personalized features and intent priors. This mechanism ensures that cold-start users receive a stable, intent-aware representation even with sparse data.

C. Intent-Guided Diffusion Module

The fused representation ( $z_c$ ) acts as a conditional signal to guide a diffusion process that refines the user representation.

Forward Process: Gaussian noise is progressively added to the initial user representation ( $S_{c0}$ ) over $T$ steps, creating a noisy trajectory.
Reverse (Denoising) Process: A neural network ( $\epsilon_\theta$ $ϵ_{θ}$ ) predicts the noise injected at each step. Crucially, this process is intent-guided:
- The model uses Classifier-Free Guidance to strengthen the condition. It computes a weighted difference between the noise predicted with the intent condition ( $z_c$ ) and the noise predicted with a null intent vector ( $z_\emptyset$ ).
- The guidance strength is controlled by a hyperparameter $w$ .
Refinement: Through iterative denoising, the model recovers a refined user representation that aligns better with latent preferences, effectively filtering out noise caused by data sparsity.

Optimization

The model is trained using a joint objective function combining:

Diffusion Loss ( $\mathcal{L}_{diff}$ ): Minimizes the mean squared error between predicted and ground-truth noise.
Recommendation Loss ( $\mathcal{L}_{rec}$ ): Uses Bayesian Personalized Ranking (BPR) loss to optimize the final matching score between the refined user representation and items.

3. Key Contributions

The paper outlines three main contributions:

A Novel Framework: Proposes an Intent-Guided Diffusion framework that synergizes transferable intent modeling with diffusion-based representation refinement to improve preference learning under sparse conditions.
Adaptive Fusion Mechanism: Designs a mechanism that dynamically weights latent intent prototypes based on their relevance to the cold-start user, mitigating the instability of representation learning caused by limited observations.
Empirical Validation: Conducts extensive experiments on three real-world datasets (MovieLens-1M, Last.FM, Yelp), demonstrating consistent superiority over representative baselines, particularly in highly sparse interaction settings.

4. Experimental Results

The authors evaluated IGDM against three categories of baselines: traditional cold-start methods (e.g., GC-MC, DropoutNet), meta-learning methods (e.g., MeLU, MetaCS), and sequential recommendation methods (e.g., SelfGNN).

Overall Performance: IGDM consistently outperformed all baselines in terms of Recall@k, NDCG@k, and MAP@k across all three datasets.
Sparse Setting Advantage: The performance gains were most pronounced under severe cold-start conditions (e.g., users with 1–5 interactions). For instance, on the MovieLens-1M dataset, IGDM showed significant improvements over the best baseline (ColdNas) in Level 1 cold-start scenarios.
Ablation Studies: Removing either the Adaptive Intent Fusion module or the Intent-Guided mechanism resulted in clear performance degradation, confirming that both components are essential for stabilizing the model and recovering preference signals.
Hyperparameter Sensitivity: The model achieved optimal performance with a diffusion step count ( $T$ ) of 50 and an intent guidance weight ( $w$ ) of 0.1. Too few steps failed to explore the latent space, while too many diluted the original signal.
Representation Quality: t-SNE visualizations indicated that IGDM produces user representations that are more structurally compact and discriminative compared to baselines like ColdNas, suggesting better capture of latent similarities while preserving individual differences.

5. Significance and Claims

The paper claims that IGDM offers a robust solution for user cold-start recommendation by addressing the dual challenges of data sparsity and unstable adaptation.

Significance: The work demonstrates that combining transferable intent priors (learned from historical data) with iterative denoising (via diffusion) allows the system to progressively refine noisy, sparse user representations into reliable preference estimates.
Modesty and Limitations: The authors acknowledge that the current framework is specific to user cold-start and has not yet been extended to item cold-start or dual cold-start settings. They also note that the transferability of intent prototypes may be limited if new users exhibit preference patterns significantly different from historical groups. Furthermore, the diffusion-based refinement introduces additional computational overhead, which may impact scalability in large-scale scenarios.

In conclusion, the paper posits that leveraging latent intent structures to guide diffusion processes is a promising direction for overcoming the limitations of existing cold-start recommendation methods, particularly in environments where behavioral data is scarce.