💻 computer science

A Contrastive Multi-Decomposition Approach with Social-Guided Signals for Robust Recommender Systems

This paper proposes COMPASS, a robust recommender system that enhances performance under noisy and sparse data by integrating multi-resolution low-rank decomposition views with social-guided contrastive learning and adaptive optimization strategies.

Original authors: Mohammad Aliabadi, Alireza Abdollahpouri, Parham Moradi

Published 2026-08-12

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Aliabadi, Alireza Abdollahpouri, Parham Moradi

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 are walking into a massive, chaotic library where millions of people are shouting out what books they liked, but the room is so noisy that you can barely hear them. Some people are whispering, others are screaming, and a few are just making random noises that sound like they like a book but actually don't. This is the world of recommender systems, the digital engines behind Netflix, Spotify, and Amazon that try to guess what you want next. To make these guesses, computers use something called Graph Neural Networks (GNNs). Think of a GNN as a super-smart librarian who doesn't just look at one person's list, but looks at the entire web of connections between every reader and every book. However, this librarian has a problem: when the library is too quiet (sparse data) or too noisy (bad feedback), the librarian gets confused and starts recommending books you'll hate.

The big question scientists are asking is: How do we teach this librarian to ignore the noise and find the real patterns, even when the data is messy? The answer usually involves contrastive learning, which is like asking the librarian to compare two slightly different versions of the same story to see what stays the same (the truth) and what changes (the noise). But most current methods try to fix the noise by randomly shaking the library shelves or adding fake books, which can accidentally throw away the good stuff along with the bad.

This paper introduces a new, clever librarian named COMPASS (Contrastive Multi-Decomposition Approach with Social-Guided Signals). Instead of randomly shaking the shelves, COMPASS uses a special pair of glasses that let it see the library in layers. It looks at the "big picture" of what everyone likes (global patterns) and then zooms in to see the tiny, specific details of what just a few people like (local details). It also listens to what your friends say about books, but only as a gentle suggestion, not a strict rule. The researchers tested this on real data from music, movies, and books, and found that COMPASS is much better at ignoring the noise and giving great recommendations, even when the data is very messy or when users haven't read many books yet.

The Problem: A Noisy Library

Imagine you are trying to figure out what your friend likes to eat. You ask them, "Do you like pizza?" and they say "Yes." But what if they were just being polite, or maybe they were hungry and said yes to everything? In the world of online recommendations, this happens all the time. Users click on things they don't actually like, or they just browse without buying. This is called noisy implicit feedback.

Current computer programs that try to guess what you like (recommender systems) often treat every click as a perfect truth. They build a giant map (a graph) connecting users to items. But if the map has too many wrong connections, the computer gets lost. It's like trying to navigate a city where half the street signs point to the wrong place. Most existing methods try to fix this by randomly deleting some connections or adding fake ones to see what happens. But this is a bit like trying to fix a broken map by throwing darts at it and hoping you hit the right spot. It often removes the good clues along with the bad ones.

The Solution: COMPASS and Its Special Glasses

The authors of this paper, Mohammad Aliabadi, Alireza Abdollahpouri, and Parham Moradi, created a new system called COMPASS. Instead of randomly guessing which connections are real, COMPASS uses a mathematical trick called Singular Value Decomposition (SVD).

Think of SVD as a way to take a giant, messy photo of the library and break it down into layers of clarity.

The Low-Rank Layer (The Big Picture): This layer shows the most obvious patterns. It's like seeing the library from a drone: you can clearly see the main aisles and the most popular sections. This is great for users who haven't read many books yet (sparse users) because it gives them a safe, general recommendation based on what everyone likes.
The High-Rank Layer (The Fine Details): This layer zooms in to show the tiny, specific details. It's like walking down the aisle and noticing that a specific person loves a very obscure genre. This is great for users who have read a lot of books (active users) because it captures their unique, specific tastes.

Most systems only look at one of these layers. COMPASS, however, puts on a pair of Gumbel-Softmax glasses. These glasses automatically decide how much to look at the "big picture" versus the "fine details" for each person. If you are a new user, the glasses focus on the big picture. If you are a super-fan, the glasses zoom in on the details. This happens automatically, without the computer needing to be told what to do.

The Social Whisper

COMPASS also listens to your friends, but it does it carefully. In many systems, if your friend likes something, the computer forces you to like it too. But what if your friend is just being weird? COMPASS treats social connections as a soft whisper. It says, "Hey, your friend likes this, so maybe you will too," but it doesn't force the issue. If the computer sees that your friend's taste doesn't match your actual behavior, it ignores the whisper. This helps the system stay robust even if some social connections are fake or weak.

The Balancing Act

Training a computer to do all these things at once is tricky. The computer has to learn to recommend items (the main job) while also learning to ignore noise (the extra job). If it focuses too much on ignoring noise, it might forget how to recommend. If it focuses too much on recommending, it might get confused by the noise.

The authors used a gradient-balanced weighting strategy. Imagine a seesaw where the weights are constantly adjusting themselves. The computer watches how hard it is to learn each part and automatically shifts the weight to keep the seesaw balanced. This ensures the computer learns both skills perfectly without one overpowering the other.

What They Found

The team tested COMPASS on three real-world datasets: Yelp (businesses), LastFM (music), and Douban (books). They compared it against ten other top-tier systems.

Better Accuracy: COMPASS consistently beat the other systems. On the Douban dataset, it improved the accuracy of recommendations by up to 5.76% (measured by NDCG@20) compared to the next best system. That might sound small, but in the world of recommendation engines, it's a huge win.
Handling Noise: The researchers deliberately added "noise" to the data, making 20% of the interactions fake. Even with this much corruption, COMPASS kept performing well, while other systems crashed. It suggests that COMPASS is very good at ignoring the fake signals.
Helping New Users: The system was especially good at helping users with very few interactions (sparse users). By focusing on the "big picture" for these users, it could still make good guesses when there wasn't much data to go on.
Speed: Despite doing all this extra math, the system didn't get much slower. The extra calculations (the SVD layers) were done once before the training started, so during the actual learning, it was just as fast as the other systems.

The Takeaway

The paper suggests that the secret to a great recommender isn't just having more data or adding more noise to the mix. It's about looking at the data from different angles. By breaking the user-item connections into layers of "big picture" and "fine detail," and letting the system decide which angle to focus on for each person, we can build systems that are smarter, more robust, and better at handling the messy reality of human behavior.

The authors note that while this works well, there is still room to grow. Future work could involve making the "social whispers" even smarter by figuring out which friends are trustworthy, or using the same multi-layer idea for social networks themselves. But for now, COMPASS shows that a little bit of structural decomposition goes a long way in making our digital recommendations feel a bit more human.

Technical Summary: COMPASS for Robust Recommender Systems

Problem Statement

Graph Neural Networks (GNNs) have become a dominant paradigm for collaborative filtering by modeling high-order user-item relationships. However, their performance often degrades in real-world scenarios characterized by noisy implicit feedback and sparse interactions. Existing approaches typically represent the user-item graph as a single structure or rely on stochastic graph augmentations (e.g., random edge perturbations). These methods face two primary limitations:

Single Granularity: They often process structural information at a single level, failing to distinguish between stable global collaborative patterns and unreliable local connections.
Inflexible Social Integration: Social information is frequently coupled with fixed propagation architectures rather than being utilized as soft semantic guidance, and existing contrastive methods often rely on manually designed views or stochastic perturbations that may remove informative connections along with noisy ones.

Methodology: COMPASS

The paper proposes COMPASS (Contrastive Multi-Decomposition Approach with Social-Guided Signals), a framework designed to capture user preferences across multiple structural resolutions without relying on stochastic graph perturbations. The architecture consists of four core components:

1. Multi-Resolution View Generation

Instead of stochastic augmentations, COMPASS constructs deterministic complementary views using rank-truncated Singular Value Decomposition (SVD) of the normalized interaction matrix.

Low-Rank Components: Capture dominant, global collaborative patterns (broad community-level information).
High-Rank Components: Preserve finer interaction details and localized preferences.
Mechanism: Multiple views are generated for a set of ranks $Q$ , creating a hierarchy of representations based on classical low-rank approximation theory.

2. Data-Driven View Fusion

A Gumbel-Softmax gating mechanism is introduced to learn how to combine these multi-resolution views adaptively.

Activity-Aware Partitioning: Users are divided into "sparse" and "active" groups based on their interaction counts relative to the median.
Adaptive Weighting: The model learns separate rank distributions ( $\pi_S$ for sparse users, $\pi_A$ for active users). Sparse users are guided to rely more on low-rank (global) components, while active users leverage higher-rank (fine-grained) components.
Efficiency: This fusion introduces only $3|Q|$ additional scalar parameters.

3. Social-Guided Contrastive Regularization

Social relations are incorporated not as a rigid propagation path but as a soft contrastive signal.

A social view is constructed using a symmetric social adjacency matrix with edge dropout.
A contrastive objective (InfoNCE) aligns the collaborative embeddings with the social embeddings. This encourages consistency between social homophily and collaborative preferences without imposing hard constraints, allowing the model to downweight noisy social links if necessary.

4. Adaptive Gradient-Balanced Optimization

To address the imbalance in gradient magnitudes between the recommendation objective (BPR) and the contrastive objectives (InfoNCE), the paper employs a dual-optimizer strategy.

Bounded Learnable Weights: The weights for the BPR and InfoNCE losses are learned via sigmoid-mapped parameters with bounded intervals.
Dual-Optimizer: A primary optimizer updates model parameters $\Theta$ at every batch, while a secondary optimizer updates the loss weights at a lower frequency (every $T$ epochs) to ensure stability and prevent rapid fluctuations.

Key Contributions

The authors summarize their contributions as follows:

Multi-resolution Graph Decomposition: A framework based on hierarchical SVD truncation that represents interactions at different approximation levels, capturing both dominant patterns and residual structures.
Data-Driven View Fusion: A Gumbel-Softmax gating mechanism that adaptively balances global and local information based on user activity, requiring minimal additional parameters.
Adaptive Optimization: A gradient-balanced learning strategy with bounded adaptive loss weights to stabilize the joint optimization of recommendation and contrastive tasks.
Robustness Evaluation: Extensive experiments demonstrating stable performance under varying sparsity levels and injected interaction noise.

Experimental Results

The framework was evaluated on three real-world datasets: Yelp, LastFM, and Douban.

Performance: COMPASS consistently outperformed strong baselines, including LightGCN, SGL, SimGCL, and recent social recommendation models (e.g., RGCML, RecDiff).
- On the Douban dataset, COMPASS achieved the highest gains, with improvements of 6.32% in Recall@20 and 5.76% in NDCG@20 over the best baseline (RGCML).
- Improvements were also observed on LastFM and Yelp, with NDCG@20 gains of 3.08% and 2.75%, respectively.
Ablation Studies: Removing the multi-resolution branch caused the most significant performance degradation (e.g., ~10.4% drop in Recall@20 on Douban), confirming the importance of multi-resolution modeling. The social view and adaptive optimization provided complementary, though slightly smaller, benefits.
Robustness:
- Sparsity: COMPASS showed the largest performance margins over baselines for users with low activity (few interactions), indicating its effectiveness in cold-start or sparse scenarios.
- Noise Injection: When up to 20% of interactions were corrupted, COMPASS maintained stable performance with a more gradual degradation rate compared to baselines.
Efficiency: The one-time SVD pre-computation cost was negligible (under 1 second for the largest dataset). During inference, only the collaborative branch is retained, resulting in complexity identical to LightGCN ( $O(|E_c|d)$ ).

Significance and Claims

The paper claims that COMPASS offers a robust alternative to stochastic augmentation methods by leveraging deterministic, multi-resolution spectral decomposition. Its significance lies in:

Handling Sparsity and Noise: By distinguishing between global patterns and local details, the model can better filter out unreliable signals in sparse data.
Adaptive Representation: The ability to dynamically adjust the reliance on different structural resolutions based on user activity addresses the heterogeneity of user data histories.
Soft Social Guidance: Treating social relations as a soft regularizer rather than a hard propagation mechanism allows the model to remain robust against noisy or weak social links.
Stable Optimization: The adaptive gradient-balancing strategy ensures that the auxiliary contrastive objectives do not overwhelm the primary recommendation task during training.

The authors conclude that the combination of multi-resolution decomposition, adaptive fusion, and socially guided contrastive learning provides a unified framework that improves recommendation accuracy and robustness without introducing substantial computational overhead.