💻 computer science

Debiasing Message Passing to Mitigate Popularity Bias in GNN-based Collaborative Filtering

This paper proposes DPAA, a novel framework that mitigates popularity bias in GNN-based collaborative filtering by integrating adaptive, embedding-aware interaction weights and layer-wise weighting directly into the message passing process to counteract the amplification of popular items and enhance long-tail item exposure.

Original authors: Md Aminul Islam, Ahmed Sayeed Faruk, Sourav Medya, Elena Zheleva

Published 2026-05-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Md Aminul Islam, Ahmed Sayeed Faruk, Sourav Medya, Elena Zheleva

Original paper licensed under CC BY 4.0 (http://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 through a massive, bustling marketplace (the internet) trying to find a gift. You ask a group of friends (an algorithm) for suggestions.

The Problem: The "Rich Get Richer" Crowd
In this marketplace, there are a few famous, flashy stalls selling the same popular items (like the latest iPhone or a viral video). Because everyone keeps pointing at these stalls, your friends keep shouting, "Go there! Go there!"

Even if you actually prefer a quiet, unique artisan stall in the back (a "long-tail" item), your friends ignore it. Why? Because the famous stalls have so many connections that the "whispers" of recommendation travel to them over and over again. The more popular an item is, the louder it gets, drowning out the unique, niche items that might actually be perfect for you. This is called Popularity Bias.

The Old Solutions: Trying to Fix It After the Fact
Previous attempts to fix this were like telling your friends, "Okay, stop shouting about the famous stalls," after they had already made their list. Or, they tried to manually lower the volume on the famous stalls using a static rule (like "if a stall has 100 signs, turn the volume down by 50%").

The paper argues these methods fail because the damage is already done. By the time the friends finish their conversation, the famous stalls have already dominated the conversation. The unique stalls never got a fair chance to be heard.

The New Solution: DPAA (The Smart Moderator)
The authors propose a new method called DPAA (Debiasing Popularity Amplification in Aggregation). Think of DPAA as a smart moderator who steps in while the friends are talking, not after.

Here is how DPAA works, using two main tricks:

1. The "Stable Ear" (Interaction-Level Weighting)

When your friends start talking, they might be a bit confused or noisy at the very beginning.

The Problem: If the algorithm tries to judge popularity based on what it thinks right now, it might get it wrong because it's still learning.
The Fix: DPAA uses a "Stable Ear." It starts by listening to a trusted, pre-trained expert (a model that has already seen the data) to decide who is popular. As the algorithm gets smarter and more confident during the training process, it slowly switches to listening to its own "growing" judgment.
The Result: This prevents the algorithm from making wild, noisy guesses early on. It ensures that when it decides to turn down the volume on a popular item, it's a calm, calculated decision, not a panic reaction.

2. The "Deep Dive" (Layer-Level Weighting)

Imagine your friends are passing a message down a line.

Layer 1: "I saw this popular item." (This is usually just the obvious, popular stuff).
Layer 2: "My friend's friend saw this niche item."
Layer 3: "My friend's friend's friend saw this amazing hidden gem."

In normal algorithms, the first layer (the popular stuff) gets the most attention, and the deeper layers (the hidden gems) get ignored.

The Fix: DPAA acts like a conductor who says, "Hey, the deeper the message goes, the more important it is!" It deliberately amplifies the voices from the deeper layers of the conversation.
The Result: This forces the algorithm to pay attention to the "hidden gems" found through long chains of connections, rather than just the items that are popular right next to you.

What Happens When They Test It?
The researchers tested this new "Smart Moderator" on real-world data (like online shopping and video watching) and also created fake data where they could control exactly how "biased" the popularity was.

The Outcome: DPAA consistently outperformed all other methods. It didn't just help the niche items; it actually improved recommendations for everyone.
The Magic: It broke the "Rich Get Richer" loop. It didn't just silence the popular items; it made sure the unique items got a fair hearing, resulting in a better, more diverse list of recommendations for the user.

In Summary
The paper claims that by fixing the bias while the algorithm is learning (during the message passing) rather than trying to fix it later, and by using a smart mix of "old wisdom" and "new learning," we can stop popular items from hijacking the recommendation system. This leads to a marketplace where you are more likely to find that perfect, unique gift you were actually looking for.

Technical Summary: Debiasing Popularity Amplification in Aggregation (DPAA)

Problem Statement

Graph Neural Network (GNN)-based Collaborative Filtering (CF) models have achieved state-of-the-art performance by propagating user-item signals over interaction graphs. However, these models are highly susceptible to popularity bias. Real-world interaction data typically follows a long-tail distribution where a small fraction of popular items accounts for the majority of interactions. In GNNs, message passing repeatedly propagates signals from high-degree popular items across local and higher-order neighborhoods. This process amplifies the influence of popular items while suppressing long-tail (niche) items, creating a "rich-get-richer" feedback loop.

Existing debiasing approaches—such as inverse propensity weighting (IPW), regularization, causal modeling, and post-processing—often fail in GNN settings because they do not directly counteract the bias propagated during the aggregation process. Recent in-aggregation weighting methods (e.g., APDA, NAVIP) rely on static heuristics or unstable embedding estimates during early training, which can distort representation learning by propagating noisy signals.

Methodology: DPAA

The authors propose Debiasing Popularity Amplification in Aggregation (DPAA), a framework that integrates adaptive, embedding-aware interaction weighting and layer-wise weighting directly into the GNN message passing process.

1. Interaction-Level Weighting (IIW)

To mitigate popularity bias at the edge level, DPAA assigns dynamic weights to user-item interactions based on their alignment in the embedding space.

Inverse Interaction Weight: The method defines an interaction weight as the complement of the cosine similarity between user and item embeddings. High similarity (often indicative of popularity-driven alignment) results in lower weights, while low similarity (niche items) receives higher weights.
Stabilized Estimation: To address the instability of embeddings in early training epochs, DPAA employs a time-dependent convex combination of weights derived from:
- A pre-trained model (providing stable signals early in training).
- The current evolving model (capturing adaptive signals as training progresses).
- A transition coefficient ( $\beta_t$ ) is dynamically adjusted based on the stability of the current embeddings (measured by average embedding change between epochs), ensuring a smooth shift from pre-trained reliance to current-model reliance.

2. Layer-Level Weighting (LW)

To address bias beyond individual interactions, DPAA introduces a layer-wise weighting mechanism that controls the contribution of different GNN layers.

Mechanism: Deeper layers, which aggregate signals from multi-hop neighbors, are assigned higher weights ( $\lambda_l = (l+1)^\eta$ ).
Rationale: While shallow layers often capture immediate interactions dominated by popular items, deeper layers incorporate signals from diverse, higher-order collaborative contexts. Amplifying these layers helps surface long-range interactions with underexposed, relevant items.

3. Unified Aggregation

The final weight for an interaction at layer $l$ combines both IIW and LW. The framework also utilizes initial residual connections to prevent over-smoothing, ensuring that node-specific information is preserved as the network deepens.

Key Contributions

The paper outlines four primary contributions:

Direct Bias Correction: DPAA is a unified framework that directly mitigates popularity amplification during message passing by combining stabilized, adaptive embedding-aware weights with layer-wise weights.
Theoretical Characterization: The authors provide theoretical proofs (Lemmas 4.1 and 4.2) demonstrating that:
- IIW reduces the relative message-passing contribution of popular items, with the reduction magnitude scaling with the severity of popularity bias and degree imbalance.
- LW structurally amplifies gradient influence toward deeper layers, shifting representation learning away from shallow, popularity-dominated interactions toward higher-order collaborative signals.
Semi-Synthetic Data Generation: The authors introduce a novel process for generating semi-synthetic click data using Zipfian probability sampling. This allows for the controlled variation of popularity bias severity ( $s$ ), enabling systematic evaluation of debiasing methods under different bias intensities—a capability lacking in prior studies.
Empirical Performance: Extensive experiments on real-world and semi-synthetic datasets demonstrate that DPAA consistently outperforms state-of-the-art methods across various bias severities and sparsity levels.

Experimental Results

The authors evaluated DPAA on three real-world datasets (Coat, KuaiRec, Yahoo! R3) and semi-synthetic data with controlled bias severity.

Overall Performance: DPAA achieved the best performance across all datasets and metrics (Recall@20, NDCG@20, HR@20). Notably, it showed significant gains on KuaiRec (up to 24.05% improvement in NDCG) and Coat (11.13% improvement in Recall), outperforming baselines like APDA, NAVIP, DAP, and causal inference methods.
Robustness to Bias Severity: In experiments varying the popularity bias severity parameter ( $s$ ) from 0 to 9, DPAA remained robust. While other methods degraded significantly as bias increased, DPAA maintained superior performance, particularly under severe bias conditions.
Popular vs. Niche Items: Unlike some debiasing methods that trade off popular item performance for niche items, DPAA improved performance for both groups simultaneously. This suggests the method reduces popularity-driven distortions without sacrificing preference-relevant signals.
Ablation Studies: Removing either the interaction-level weighting (IIW) or layer-wise weighting (LW) resulted in clear performance drops, confirming the necessity of both components. Furthermore, the study validated that combining pre-trained and current embeddings for weight estimation yields more stable and accurate results than relying on either source alone.

Significance and Claims

The paper claims that DPAA addresses a critical gap in GNN-based recommendation by intervening directly at the source of bias amplification: the aggregation process. By theoretically characterizing how interaction and layer weights reduce popularity dominance, the authors argue that their method offers a more principled and effective approach than post-hoc corrections or static re-weighting.

The introduction of a semi-synthetic data generation process is highlighted as a significant methodological contribution, allowing for rigorous, controlled evaluation of debiasing techniques that was previously difficult due to the inability to manipulate bias severity in real-world data. The authors position DPAA as a step toward more equitable recommendation systems that can effectively serve both popular and long-tail items, though they acknowledge limitations regarding other types of biases (e.g., selection or position bias) and the need for responsible deployment guardrails in practice.