CD-CGCN: Community Debiasing Framework

• CD-CGCN is a universal framework for mitigating filter bubbles in recommender systems by exploiting community structures via community detection and adversarial debiasing.
• It integrates a Conditional Discriminator with a community-reweighted GCN and community-enhanced negative sampling to extract community attributes and rebalance recommendations.
• Experimental evaluations on datasets like Gowalla and Amazon-Book show significant improvements in ranking metrics and reduced intra-community bias compared to standard recommenders.

CD-CGCN is a universal framework for mitigating the filter bubble effect in recommender systems from a community detection perspective. Introduced in "Mitigating Filter Bubble from the Perspective of Community Detection: A Universal Framework" (Tang et al., 15 Aug 2025), it treats the filter bubble as a community-structured exposure bias on the user–item bipartite graph: users and items form communities, and low diversity manifests as repeated recommendation of intra-community items. The framework is model-agnostic and operates as a plug-in atop standard recommenders such as MF, NGCF, LightGCN, and DGCF. Its core design combines a Conditional Discriminator with a Community-reweighted Graph Convolutional Network, adversarial learning based on community labels, Community-enhanced Negative Sampling, a Gradient Reversal Layer, and a user-adaptive inference strategy (Tang et al., 15 Aug 2025).

1. Community-structured formulation of the filter bubble

In CD-CGCN, the filter bubble is explicitly grounded in community structure rather than treated only as a generic lack of diversity. Let users and items be sets $\mathcal{U}$ and $\mathcal{I}$ with $|\mathcal{U}| = m$, $|\mathcal{I}| = n$, and interaction matrix $R \in \mathbb{R}^{m \times n}$, where $R_{u,i} = 1$ if user $u$ interacted with item $i$, else $0$. The interaction data define a bipartite graph $\mathcal{G} = (\mathcal{U} \cup \mathcal{I}, R)$, and community labels are obtained by applying Louvain on the training graph only, producing $\mathcal{I}$0 (Tang et al., 15 Aug 2025).

This formulation ties recommendation bias to latent graph structure. On Gowalla, 78% of users’ historical interactions are intra-community; in the top-20 lists from standard recommenders, $\mathcal{I}$1 of recommended items are intra-community. The framework therefore interprets filter bubbles as a systematic overexposure to same-community items. This suggests that community detection is not merely an auxiliary analytic tool but a mechanism for identifying the structural bias inherited by recommendation models (Tang et al., 15 Aug 2025).

The paper further diagnoses the behavior of common recommenders during optimization. Tracking ILFBI@20 during training shows that LightGCN and NGCF hold ILFBI $\mathcal{I}$2 from early epochs, while MF and DGCF gradually trend toward similar levels. Grouping users by “ILFBI-init.”, defined as the intra-community proportion in training data, reveals that the ILFBI increment is larger for users whose histories already favor intra-community interactions. The reported interpretation is that base recommenders amplify the bubble via a feedback loop (Tang et al., 15 Aug 2025).

2. Architectural composition

CD-CGCN stands for Conditional Discriminator + Community-reweighted Graph Convolutional Network. It takes embeddings and scores from a base recommender, extracts community attributes, adversarially debiases them with respect to community labels, applies community-aware negative sampling, and fuses base and debiased predictions at inference (Tang et al., 15 Aug 2025).

Component	Role	Mechanism
CGCN	Extract community attributes	Community-reweighted propagation
CD	Predict community labels	Two-layer MLP with global embeddings
CNS	Increase intra-community repulsion	Negative sampling with probability $\mathcal{I}$3
GRL	Stabilize adversarial training	Gradient sign reversal with factor $\mathcal{I}$4
UIS	Personalize inference	Score fusion via $\mathcal{I}$5

The Community-reweighted GCN is parameter-free. Let $\mathcal{I}$6 and $\mathcal{I}$7 denote base embeddings, and initialize $\mathcal{I}$8 and $\mathcal{I}$9. The compatibility of neighbor item $|\mathcal{U}| = m$0 for user $|\mathcal{U}| = m$1 is defined as

$|\mathcal{U}| = m$2

An analogous definition gives $|\mathcal{U}| = m$3 for item $|\mathcal{U}| = m$4 with respect to user $|\mathcal{U}| = m$5. The propagation rule for layer $|\mathcal{U}| = m$6 is

$|\mathcal{U}| = m$7

$|\mathcal{U}| = m$8

Community embeddings are then aggregated as

$|\mathcal{U}| = m$9

Because the $|\mathcal{I}| = n$0 weights amplify intra-community edges, CGCN makes same-community embeddings cluster more strongly, thereby explicitly extracting the community attributes that the adversarial component is designed to counteract (Tang et al., 15 Aug 2025).

The Conditional Discriminator is a two-layer MLP $|\mathcal{I}| = n$1 with parameters $|\mathcal{I}| = n$2. It receives the concatenation of a community embedding and a shared global embedding indicating entity type:

$|\mathcal{I}| = n$3

With one-hot community labels $|\mathcal{I}| = n$4 and $|\mathcal{I}| = n$5, it is trained using cross-entropy,

$|\mathcal{I}| = n$6

The intended effect is to make community labels predictable from extracted attributes while the encoder, through adversarial training, is pushed to reduce community-identifiable signals in its embeddings (Tang et al., 15 Aug 2025).

3. Optimization and adversarial debiasing

The recommendation objective uses BPR over triplets $|\mathcal{I}| = n$7:

$|\mathcal{I}| = n$8

where $|\mathcal{I}| = n$9 and $R \in \mathbb{R}^{m \times n}$0 denotes base-model parameters. The discriminator contributes an adversarial loss

$R \in \mathbb{R}^{m \times n}$1

The combined min–max objective is

$R \in \mathbb{R}^{m \times n}$2

with

$R \in \mathbb{R}^{m \times n}$3

Minimizing with respect to $R \in \mathbb{R}^{m \times n}$4 improves recommendation quality while making community labels harder to predict; maximizing with respect to $R \in \mathbb{R}^{m \times n}$5 strengthens community classification. The appendix proves that this alternating min–max procedure is equivalent to joint minimization with a Gradient Reversal Layer defined by $R \in \mathbb{R}^{m \times n}$6 and gradient $R \in \mathbb{R}^{m \times n}$7 (Tang et al., 15 Aug 2025).

Under the GRL formulation, the objective becomes

$R \in \mathbb{R}^{m \times n}$8

and the training reduces to joint minimization,

$R \in \mathbb{R}^{m \times n}$9

This design is presented as a stabilization device for end-to-end training (Tang et al., 15 Aug 2025).

Community-enhanced Negative Sampling further shifts the geometry of the learned representation. Given a triplet $R_{u,i} = 1$0 with positive item $R_{u,i} = 1$1, the negative item $R_{u,i} = 1$2 is sampled from non-interacted items; with probability $R_{u,i} = 1$3, $R_{u,i} = 1$4 is sampled from $R_{u,i} = 1$5’s same community, and with probability $R_{u,i} = 1$6, it is sampled uniformly from all non-interacted items. The stated effect is to increase repulsion among intra-community embeddings and raise the chance of recommending inter-community items (Tang et al., 15 Aug 2025).

4. User-adaptive inference and evaluation protocol

CD-CGCN measures filter-bubble severity with two community-level metrics in addition to standard ranking metrics such as Precision@k, Recall@k, and NDCG@k. The Intra-List Filter Bubble Index at cutoff $R_{u,i} = 1$7 is

$R_{u,i} = 1$8

which quantifies the proportion of recommended items that share the user’s community. The Community Gini Index is

$R_{u,i} = 1$9

where $u$0 is the number of communities, $u$1, and $u$2 are counts per community in ascending order (Tang et al., 15 Aug 2025).

Inference is tailored to each user’s filter-bubble state. For user $u$3, let $u$4 denote the intra-community proportion in $u$5’s training interactions, and let $u$6 denote the average over users. CD-CGCN defines

$u$7

and fuses scores as

$u$8

where $u$9 and $i$0 is the base model’s score. Users with larger $i$1 therefore receive stronger CD-CGCN influence at inference, promoting inter-community recommendations where the bubble is more severe (Tang et al., 15 Aug 2025).

The reported training loop is: run Louvain on the training graph; sample triplets using CNS; compute base embeddings; perform $i$2 layers of CGCN propagation; obtain discriminator outputs from concatenated embeddings with global embeddings; compute $i$3, $i$4, and $i$5; then jointly update $i$6 and $i$7 with Adam. At inference, compute base and CD-CGCN scores for all candidate items, compute $i$8, fuse scores, rank, and return top-$i$9 items (Tang et al., 15 Aug 2025).

To evaluate inter-community preference modeling more directly, the paper constructs a community-debiased test set. The motivation is that original test sets are heavily biased toward intra-community interactions; the example given is Gowalla’s test set, where 80.9% of items are intra-community. The debiasing procedure limits each user’s test set to at most one item per community. After debiasing, the reported scales are Gowalla: $0$0, Amazon-Book: $0$1, and ML-2k: $0$2 (Tang et al., 15 Aug 2025).

5. Experimental findings

The evaluation uses three real-world datasets with a 70/10/20 train/validation/test split and community detection on training data only: Gowalla with 29,858 users, 40,981 items, 1,027,370 interactions, and 18 communities; Amazon-Book with 52,643 users, 91,599 items, 2,984,108 interactions, and 9 communities; and ML-2k with 2,095 users, 5,254 items, 367,761 interactions, and 6 communities. Base models are MF, NGCF, LightGCN, and DGCF. Optimization uses Adam, embedding size 64, global embedding size 16, $0$3 regularization $0$4 for base embeddings and $0$5 for CD, and $0$6 (Tang et al., 15 Aug 2025).

Across models and datasets, CD-CGCN is reported to improve both ranking quality and filter-bubble metrics. Representative gains include the following (Tang et al., 15 Aug 2025):

Dataset	Base	Representative reported gains
Gowalla	NGCF	P@20 +10.9%, R@20 +9.2%, NDCG@20 +10.2%, ILFBI@20 −5.1%, CGI@20 −0.7%
Gowalla	DGCF	P@20 +11.1%, R@20 +8.9%, NDCG@20 +9.7%, ILFBI@20 −2.7%
Amazon-Book	MF	P@20 +13.1%, R@20 +14.9%, NDCG@20 +15.5%, ILFBI@20 −11.5%, CGI@20 −3.4%
Amazon-Book	NGCF	P@20 +14.3%, R@20 +15.5%, NDCG@20 +15.8%, ILFBI@20 −7.7%
Amazon-Book	DGCF	P@20 +17.4%, R@20 +20.4%, NDCG@20 +20.0%
ML-2k	MF	P@100 +5.7%, NDCG@100 +8.3%, ILFBI@100 −12.9%

The framework is also reported to achieve better accuracy–bubble trade-offs than MMR, Fairness, and IPS. Training trajectories show that ILFBI@20 remains lower across epochs for CD-CGCN than for baselines, and t-SNE visualizations on Amazon-Book show cross-community blending of embeddings after debiasing, in contrast to sharp same-community clusters before debiasing (Tang et al., 15 Aug 2025).

The debiased test sets produce a stronger effect size: the accuracy improvements of CD-CGCN are significantly larger on debiased test sets than on original ones. The paper interprets this as evidence that the framework better captures inter-community preferences, rather than merely exploiting the original test distribution’s intra-community skew (Tang et al., 15 Aug 2025).

Ablation on LightGCN over Amazon-Book reports that removing CGCN, CD, or CNS degrades both accuracy and bubble metrics, and removing the User-adaptive Inference Strategy reduces accuracy substantially. One example states that CD-CGCN improves P@20 from 0.0172 to 0.0184 and reduces ILFBI@20 from 0.8082 to 0.7398; removing UIS drops P@20 to 0.0168 despite a strong ILFBI reduction. This suggests that the adaptive fusion coefficient $0$7 is central to managing the accuracy–debiasing trade-off (Tang et al., 15 Aug 2025).

6. Complexity, implementation, and scope conditions

The framework is designed to be plug-and-play. It only requires learned embeddings and scores from the base model plus community labels extracted from the interaction graph. Louvain is run once on training data and is described as scaling well to large graphs. CGCN is parameter-free message passing over the interaction graph for $0$8 layers, with per-layer complexity linear in the number of edges, $0$9; symmetric normalization uses precomputed sums. Memory overhead is described as minimal because there are no extra parameters beyond global embeddings and the Conditional Discriminator. The discriminator itself is a small MLP over concatenated embeddings of dimensions $\mathcal{G} = (\mathcal{U} \cup \mathcal{I}, R)$0, and GRL introduces negligible overhead (Tang et al., 15 Aug 2025).

The implementation is in PyTorch with Adam optimization and standard ranking metrics. The paper does not announce a public code release, but it states that methodological details are sufficient for reproduction. Practical guidance given in the paper includes initializing global embeddings randomly with small norms, starting with moderate $\mathcal{G} = (\mathcal{U} \cup \mathcal{I}, R)$1 such as 0.3–0.5 and $\mathcal{G} = (\mathcal{U} \cup \mathcal{I}, R)$2 such as 0.1–0.5, using early stopping on both accuracy and ILFBI/CGI, and expecting particularly strong gains with richer base encoders such as NGCF and DGCF, while still obtaining benefits with MF and LightGCN (Tang et al., 15 Aug 2025).

The main limitations are also explicitly stated. The method depends on community detection quality; misassigned labels can weaken debiasing. Current labels are static and computed once on training data, so dynamic communities and temporal drift are not handled. Community labels for very sparse users or items may be unreliable, which creates cold-start issues. Finally, excessive adversarial strength $\mathcal{G} = (\mathcal{U} \cup \mathcal{I}, R)$3 can hurt accuracy, even though the user-adaptive inference strategy partially mitigates this. Proposed future directions include dynamic or evolving communities, integration of side information or content-based signals, more personalized debiasing strength, and causal formulations (Tang et al., 15 Aug 2025).

7. Terminological ambiguity and related acronyms

The label “CD-CGCN” is potentially ambiguous in the recent literature. In recommender systems, it denotes the framework built from a Conditional Discriminator and a Community-reweighted Graph Convolutional Network (Tang et al., 15 Aug 2025). In contrast, "Self-Supervised Contrastive Graph Clustering Network via Structural Information Fusion" uses the acronym “CGCN,” not “CD-CGCN”; the paper explicitly states that the term “CD-CGCN” does not appear there, and that if encountered elsewhere it likely refers to the same self-supervised contrastive deep graph clustering model, which the paper names simply “CGCN” (Ji et al., 2024). That method addresses unsupervised graph clustering through an AE, a GAE/IGAE, structural information fusion, and contrastive alignment implemented as squared $\mathcal{G} = (\mathcal{U} \cup \mathcal{I}, R)$4 losses, rather than recommendation debiasing (Ji et al., 2024).

A further nearby acronym is “CDGCN,” short for Community Detection Graph Convolutional Network, introduced for overlap-aware speaker diarization (Wang et al., 2023). That method builds KNN subgraphs over speech segments, refines local edges with a GCN, applies Leiden for speaker-community partitioning, and adds Graph-OSD for overlap-aware labeling. Its target problem, inputs, losses, and evaluation regime are distinct from those of CD-CGCN in recommendation (Wang et al., 2023).

A common misconception is therefore to treat these names as interchangeable. They are not identical methods. The shared acronyms arise from overlapping uses of “community detection,” “graph convolutional network,” and “contrastive” or “conditional discriminator” terminology, but the underlying tasks differ: recommendation debiasing in (Tang et al., 15 Aug 2025), self-supervised graph clustering in (Ji et al., 2024), and speaker diarization in (Wang et al., 2023). The most precise usage reserves “CD-CGCN” for the recommender-system framework of (Tang et al., 15 Aug 2025).