1. Bibliographic Information

1.1. Title

Negative Feedback Really Matters: Signed Dual-Channel Graph Contrastive Learning Framework for Recommendation

1.2. Authors

Leqi Zheng, Chaokun Wang (Corresponding Author), Zixin Song, Cheng Wu, Shannan Yan, Ziyang Liu (Tsinghua University); Jiajun Zhang (University of Science and Technology of China).

1.3. Journal/Conference

This paper is associated with NeurIPS (Neural Information Processing Systems). Given the citations and references, it represents a state-of-the-art contribution to the 2025 research cycle in graph-based recommendation systems. NeurIPS is a top-tier (A*) conference in machine learning and artificial intelligence, known for high rigor and significant impact.

1.4. Publication Year

2025 (Based on the repository and metadata provided).

1.5. Abstract

Traditional recommender systems primarily focus on positive feedback (e.g., clicks, likes), leaving the abundant negative feedback (e.g., dislikes, low ratings) underutilized. Existing methods that attempt to use negative feedback suffer from poor model compatibility, ineffective information exchange, and high computational costs. This paper proposes the Signed Dual-Channel Graph Contrastive Learning (SDCGCL) framework. It features a dual-channel embedding for separate positive and negative graph processing, a cross-channel distribution calibration mechanism to ensure consistency, and an adaptive prediction strategy. Additionally, the authors introduce the DualFuse model and a two-stage optimization strategy to balance accuracy and efficiency. Experiments across four datasets show substantial improvements over 22 state-of-the-art baselines.

1.6. Original Source Link

The source code and data are released at: https://github.com/LQgdwind/nips25-sdcgcl. The paper is a research submission/publication for NIPS 2025.

---

2. Executive Summary

2.1. Background & Motivation

In the realm of digital platforms, recommender systems are vital for user satisfaction. Most current systems treat interactions as "unsigned" graphs—only recording that an interaction happened. However, real-world data is "signed": a user might watch a movie and hate it (negative feedback) or love it (positive feedback).

The Core Problem: Negative feedback is often ignored because it is harder to model. Unlike positive feedback, which suggests similarity, negative feedback suggests a "repulsion" or a boundary of interest. Previous researchers faced three hurdles:

1. Model Compatibility: Specialized "signed" models can't use the latest advances in general Graph Contrastive Learning (GCL).
2. Ineffective Exchange: Negative signals are often treated as "extra" rather than fundamental, leading to a shallow understanding of user preferences.
3. Efficiency: Processing all negative interactions (which can be as numerous as positive ones) makes training very slow.

Innovation: The paper introduces a model-agnostic framework (meaning it can be "plugged into" other existing models) that treats positive and negative signals as two separate channels that "talk" to each other to refine user representations.

2.2. Main Contributions / Findings

1. SDCGCL Framework: A model-agnostic architecture that separates positive and negative interactions into two channels, allowing existing GNNs to leverage signed data without structural changes.

2. Cross-Channel Calibration: A mechanism that prevents the positive and negative embedding spaces from drifting too far apart, ensuring the model remains stable.

3. DualFuse Model: A specific implementation that uses a "cross-channel fusion" where positive signals perturb negative ones (and vice versa) to create better contrastive views.

4. Two-Stage Optimization: A strategy that starts with the full graph and then switches to a Popularity-Guided Random Walk Sampling strategy, maintaining high accuracy while cutting training time by nearly 30%.

5. Performance: The proposed method outperformed the strongest baselines (like XSimGCL and NFARec) by up to 28% on datasets like Yelp.

   ---

3. Prerequisite Knowledge & Related Work

3.1. Foundational Concepts

3.1.1. Graph Neural Networks (GNNs)

A Graph Neural Network is a type of AI that learns from data structured as a graph (nodes connected by edges). In recommendation, Users and Items are nodes, and Interactions are edges. GNNs use Message Passing, where a node updates its "understanding" (embedding) by collecting information from its neighbors.

3.1.2. Graph Contrastive Learning (GCL)

Contrastive Learning is a technique where the model learns by comparing things. It takes an original graph, creates two slightly "distorted" versions (views), and tries to make the representations of the same node in both views as similar as possible. This helps the model learn robust features even when data is noisy or sparse.

3.1.3. Bayesian Personalized Ranking (BPR)

BPR is a common loss function (a way to measure error) in recommendation. It assumes that an observed interaction (like a user buying an item) should have a higher predicted score than a non-observed interaction.

3.1.4. Signed Graphs

In a standard graph, an edge either exists or it doesn't. In a Signed Graph, edges have weights: +1 for positive feedback (like) and -1 for negative feedback (dislike).

3.2. Previous Works

• Unsigned RS: Models like LightGCN simplify the GNN structure for recommendation but ignore the "meaning" (positive vs. negative) of the edges.
• Contrastive Models: SGL and XSimGCL use data augmentation to improve GCL but only on positive interactions.
• Signed RS: Models like SiReN and SIGformer use negative feedback but often rely on Balance Theory (e.g., "the enemy of my enemy is my friend"), which the authors argue is too restrictive for modern recommendation.

3.3. Technological Evolution

The field moved from Matrix Factorization (simple math tables) to Deep Learning (NCF), then to Graph-based models (NGCF, LightGCN), and recently to Self-Supervised/Contrastive models. This paper represents the next step: bringing Signed Data into the Contrastive Learning era.

---

4. Methodology

4.1. Principles

The core idea of SDCGCL is to stop trying to force positive and negative feedback into a single math operation. Instead, the framework creates two parallel "universes" (channels). One channel learns what the user likes, and the other learns what the user dislikes. These channels are then calibrated to ensure they share a common understanding of the user.

4.2. Core Methodology In-depth (Layer by Layer)

4.2.1. Dual-Channel Graph Embedding

The framework starts by splitting the interaction graph into $\mathcal{G}^+$ (positive) and $\mathcal{G}^-$ (negative). It then propagates embeddings through $L$ layers. For each channel (indicated by + or -), the layer-wise update is:

$ \mathbf { e } _ { u , l } ^ { + } = f \left( p ( \hat { \mathbf { A } } ^ { + } ) , \mathbf { e } _ { i , l - 1 } ^ { + } \right) , \quad \mathbf { e } _ { u , l } ^ { - } = f \left( p ( \hat { \mathbf { A } } ^ { - } ) , \mathbf { e } _ { i , l - 1 } ^ { - } \right) $

Symbol Explanation:

• $\mathbf{e}_{u,l}^+ / \mathbf{e}_{u,l}^-$: The user embedding at layer $l$ for the positive/negative channel.

• $f(\cdot)$: The message propagation function (e.g., LightGCN).

• $p(\cdot)$: A dropout function (randomly removing edges to prevent overfitting).

• $\hat{\mathbf{A}}^+ / \hat{\mathbf{A}}^-$: The normalized adjacency matrices (the "map" of connections) for both channels.

  After $L$ layers, the final embeddings $\mathbf{E}_u^+$ and $\mathbf{E}_u^-$ are obtained via an aggregation function $\mathrm{AGG}(\cdot)$, typically a mean pool of all layers:

$ \mathbf { E } _ { u } ^ { + } = \mathrm { A G G } \left( { \mathbf { e } _ { u , l } ^ { + } : l \leq L } \right) , \quad \mathbf { E } _ { u } ^ { - } = \mathrm { A G G } \left( { \mathbf { e } _ { u , l } ^ { - } : l \leq L } \right) $

4.2.2. Cross-Channel Distribution Calibration

To prevent the channels from becoming contradictory, the authors use two types of contrastive learning.

1. Intra-Channel: Ensures the positive channel is robust to its own noise.

2. Inter-Channel (Distribution Alignment): This is the key innovation. It aligns the "shape" of the data distribution in both channels using a difference measure $g^*$.

   $ \operatorname* { m i n } { \sum _ { u \in \mathcal { U } } g ^ { * } ( \sum _ { ( u , i ) \in \mathcal { G } ^ { + } } \frac { \mathbf { E } _ { u } ^ { + } \circ \mathbf { E } _ { i } ^ { + \intercal } } { | | \mathcal { N } _ { u } ^ { + } | | } , \sum _ { ( u , j ) \in \mathcal { G } ^ { - } } \frac { \mathbf { E } _ { u } ^ { - } \circ \mathbf { E } _ { j } ^ { - \intercal } } { | | \mathcal { N } _ { u } ^ { - } | | } ) } $

Symbol Explanation:

• $\circ$: The Hadamard product (element-wise multiplication).
• $\mathcal{N}_u^+ / \mathcal{N}_u^-$: The set of positive/negative neighbors for user $u$.
• $g^*$: A distance metric (the paper specifically uses Jensen-Shannon Divergence in the implementation).

4.2.3. Adaptive Prediction Strategy

For the final recommendation, the model predicts a score $\hat{y}_{u,i}$ by subtracting a weighted version of the negative preference from the positive preference:

$ \hat { y } _ { u , i } = ( 1 + k ) \mathbf { E } _ { u } ^ { + } \circ \mathbf { E } _ { i } ^ { + \intercal } - k \mathbf { E } _ { u } ^ { - } \circ \mathbf { E } _ { i } ^ { - \intercal } $

Symbol Explanation:

• $k$: A hyperparameter (usually between 0 and 1) that controls how much the "dislike" signal should penalize the final score.

  The following figure (Figure 2 from the original paper) provides an overview of this process:

  该图像是SDCGCL框架的示意图，展示了三个主要组成部分：双通道图嵌入、跨通道分布校准和自适应预测策略。负图采样池（左下角）用于实现高效的训练优化。

4.3. DualFuse Model

While SDCGCL is a general framework, DualFuse is the specific model the authors designed to make the most of it. Its unique feature is the Cross-Channel Graph Fusion.

Instead of just dropping nodes to create contrastive "views," DualFuse uses information from the other channel as a perturbation. For example, to create a contrastive view for the positive channel ($\mathbf{Z}_u^+$), it adds a normalized version of the negative embedding ($\mathbf{e}_{u,l^*}^-$):

$ \mathbf { Z } _ { u } ^ { + } = \displaystyle \frac { 1 } { L + 1 } \sum _ { l = 0 } ^ { L } ( \mathbf { e } _ { u , l ^ { * } } ^ { + } + \frac { \mathbf { e } _ { u , l ^ { * } } ^ { - } } { | | \mathbf { e } _ { u , l ^ { * } } ^ { - } | | } ) $

This ensures that the contrastive learning process is "aware" of both likes and dislikes simultaneously.

4.4. Optimization Design: Two-Stage Strategy

Processing every negative interaction is expensive. The authors propose:

1. Stage 1 (Warm-up): Train on the full signed graph for a short time (usually just 1 epoch) to get a stable base.

2. Stage 2 (Sampling): Use Popularity-Guided Random Walk Sampling (Algorithm 1) to pick the most informative negative edges. This reduces the graph size while keeping the "important" dislikes.

   ---

5. Experimental Setup

5.1. Datasets

The authors used four public datasets. They binarized ratings: scores $\ge 4$ are positive, and $< 4$ are negative.

• ML-1M & ML-10M (MovieLens): Movie ratings. Very dense.
• Yelp: Business reviews. High percentage of negative feedback (~33.7%).
• Amazon: Book reviews. Very sparse, mostly positive feedback (~80.6%).

5.2. Evaluation Metrics

5.2.1. Recall@20

1. Conceptual Definition: Recall measures the proportion of items the user actually liked that the system successfully recommended in the top 20 list.
2. Mathematical Formula: $ \mathrm{Recall@K} = \frac{\sum_{u \in \mathcal{U}} |R_u \cap T_u|}{\sum_{u \in \mathcal{U}} |T_u|} $
3. Symbol Explanation: $R_u$ is the list of top-K recommended items; $T_u$ is the set of items the user actually interacted with in the test set.

5.2.2. NDCG@20 (Normalized Discounted Cumulative Gain)

1. Conceptual Definition: Unlike Recall, NDCG cares about the position. If a liked item is at #1, it's worth more than if it's at #20.
2. Mathematical Formula: $ \mathrm{NDCG@K} = \frac{\mathrm{DCG@K}}{\mathrm{IDCG@K}}, \text{ where } \mathrm{DCG@K} = \sum_{i=1}^K \frac{2^{rel_i} - 1}{\log_2(i+1)} $
3. Symbol Explanation: $rel_i$ is a 1 if the item at rank $i$ is a hit, 0 otherwise. IDCG is the "Ideal" DCG (a perfect ranking).

5.3. Baselines

The authors compared against 22 methods, including:

• Unsigned GCL: SGL, XSimGCL, LightGCL.

• Signed Models: SiReN, SIGformer, NFARec.

  ---

6. Results & Analysis

6.1. Core Results Analysis

The authors demonstrate that SDCGCL-DualFuse is the new state-of-the-art. Notably, it performs exceptionally well on Yelp, where negative feedback is more prevalent, showing a 28.01% improvement in Recall.

The following are the results from Table 1 of the original paper:

6.2. Ablation Studies / Parameter Analysis

The authors removed components to see what mattered most (Table 2 in the paper):

• w/o Fusion: Dropping the cross-channel fusion caused a massive performance drop (over 70% on MovieLens).
• w/o Rec: Dropping the supervised recommendation loss caused the embeddings to become random (visualized in Figure 4).

Sampling Rate Analysis: In Table 3, the authors found that a sampling rate ($\rho$) of 0.01 (using only 1% of negative edges) actually yielded better results than using the full graph (0.3282 vs 0.3226 Recall). This is because sampling acts as a filter, removing noise and focusing the model on the most significant negative signals.

---

7. Conclusion & Reflections

7.1. Conclusion Summary

The paper successfully proves that negative feedback really matters. By designing a model-agnostic framework (SDCGCL) that treats likes and dislikes as distinct but related channels, the authors solved the compatibility and information exchange issues that plagued previous signed recommendation models. Their DualFuse implementation and two-stage training strategy further provide a blueprint for high-performance, efficient recommendation in the real world.

7.2. Limitations & Future Work

• Industrial Scale: While tested on large datasets like ML-10M, deploying this in real-time systems with billions of edges might require further engineering.
• Cold-Start: The paper doesn't explicitly solve the problem for new users who have neither positive nor negative feedback yet.
• Dynamic Feedback: User tastes change. Future work could include a temporal element to see how "dislikes" evolve over time.

7.3. Personal Insights & Critique

This paper is highly rigorous. The decision to make the framework model-agnostic is its greatest strength; it allows researchers to upgrade their existing LightGCN or SGL models simply by adding the negative channel and the calibration loss.

One potential critique is the choice of the hyperparameter $k$ (the weight of negative feedback). The paper assumes $k$ is constant across all users, but in reality, some users are "haters" (express more negative feedback) while others are "positivity-biased." An individualized $k$ based on user history might have yielded even higher performance. Nonetheless, the two-stage sampling strategy is a brilliant solution to the "negative feedback explosion" problem, making this framework very practical for real-world applications.