Fraud-Aware Selective CoT Distillation

• The paper introduces a novel fraud-aware selective CoT distillation mechanism that autonomously generates and selects multi-hop reasoning chains to reveal fraud cues in text-attributed graphs.
• The methodology combines semantic and structural cues through an integrated LLM-GNN pipeline, achieving up to +8.8% improvement in AUPRC over leading baseline models.
• The approach employs asymmetric co-training and caching strategies to reduce LLM calls and scale batch sizes efficiently, enabling significant computational speedups.

Fraud-aware selective Chain-of-Thought (CoT) distillation is a targeted mechanism developed to enhance graph-based fraud detection in text-attributed graphs (TAGs) by autonomously generating and selecting diverse, semantically-relevant reasoning paths. Integrated within the FraudCoT framework, it addresses the limitations of predefined prompting and decoupled LLM-GNN training pipelines by aligning semantic and structural cues from textual and relational data. This process involves autonomous LLM generation of multi-hop reasoning chains, their selective distillation based on fraud-relevance and diversity, and their integration into node representations for downstream graph neural network (GNN) classification—enabling state-of-the-art detection performance and scalable end-to-end optimization (Li et al., 30 Jan 2026).

1. Problem Formulation and Notation

Fraud-aware selective CoT distillation operates on a text-attributed graph (TAG) $G = (V, E, X)$, where $V$ is the set of nodes (e.g., reviews, transactions, users), $E \subseteq V \times V$ represents binary relations, and $X = \{x_v\,|\,v \in V\}$ comprises raw textual attributes. Each node $v$ is assigned a binary fraud label $y_v \in \{0, 1\}$, and $A \in \{0, 1\}^{|V|\times|V|}$ denotes the adjacency matrix, with $N(v)$ defining the set of neighbors of $v$. The objective is to learn a scoring function $f_\theta(v)$ approximating $V$0 for accurate node-level fraud detection (Li et al., 30 Jan 2026).

2. Autonomous Chain-of-Thought Generation

The distillation process begins with teacher LLM-based generation of graph-aware reasoning chains for selected nodes. For each distillation node $V$1, a composite prompt $V$2 is constructed by concatenating its raw text $V$3 and textual descriptions of its 1-hop or multi-hop neighbors $V$4. The teacher LLM $V$5 receives $V$6 and samples $V$7 chain-of-thoughts:

$V$8

Each $V$9 is a free-form path that references both local semantic features and relational (neighbor) signals, aiming to uncover multi-hop fraud cues embedded in the graph structure (Li et al., 30 Jan 2026).

3. Selective Distillation: Relevance Scoring and Diversity Enforcement

A fraud-aware CoT scoring function $E \subseteq V \times V$0 is proposed to select high-value chains:

$E \subseteq V \times V$1

where $E \subseteq V \times V$2 is cosine similarity and $E \subseteq V \times V$3 is a lightweight encoder (e.g., BERT-tiny), balancing semantic relevance and structural alignment. Scores are normalized to $E \subseteq V \times V$4 across samples.

To ensure that distilled CoTs are both fraud-relevant and diverse, the procedure selects:

• $E \subseteq V \times V$5: the top-$E \subseteq V \times V$6 CoTs with scores $E \subseteq V \times V$7,
• $E \subseteq V \times V$8: the bottom-$E \subseteq V \times V$9 CoTs with scores $X = \{x_v\,|\,v \in V\}$0, retaining only those with pairwise Levenshtein edit distance $X = \{x_v\,|\,v \in V\}$1 to enforce diversity.

The student model $X = \{x_v\,|\,v \in V\}$2 is distilled by minimizing:

$X = \{x_v\,|\,v \in V\}$3

where $X = \{x_v\,|\,v \in V\}$4 and $X = \{x_v\,|\,v \in V\}$5 are sets of positive and negative chains, and $X = \{x_v\,|\,v \in V\}$6 is an unlikelihood loss coefficient. LoRA (Low-Rank Adaptation) adapters with rank $X = \{x_v\,|\,v \in V\}$7–$X = \{x_v\,|\,v \in V\}$8 are employed for efficient parameterization during student finetuning. Empirically, $X = \{x_v\,|\,v \in V\}$9 nodes, $v$0 samples, $v$1–3, and $v$2 are optimal (Li et al., 30 Jan 2026).

4. Integration of Distilled CoTs into GNN Representation

After student finetuning, a distilled CoT is generated for each node $v$3 as $v$4 and concatenated with the raw text to yield $v$5. This extended textual attribute is processed by an LLM-style encoder $v$6 to obtain initial node embeddings:

$v$7

These enriched embeddings encode both direct and multi-hop semantic-structural fraud cues, providing a more informative input for the downstream GNN (Li et al., 30 Jan 2026).

5. Asymmetric LLM–GNN Co-Training and Efficiency

A key component is the asymmetric co-training pipeline. In each batch $v$8, only target nodes $v$9 are freshly encoded:

$y_v \in \{0, 1\}$0

while neighbor nodes $y_v \in \{0, 1\}$1 utilize cached embeddings $y_v \in \{0, 1\}$2 computed once at initialization. Message passing employs $y_v \in \{0, 1\}$3 heterogeneous GraphSAGE layers:

$y_v \in \{0, 1\}$4

The total loss is a weighted sum of binary cross-entropy node classification and distillation loss:

$y_v \in \{0, 1\}$5

Typically, $y_v \in \{0, 1\}$6. Only target nodes propagate gradients through $y_v \in \{0, 1\}$7, greatly reducing backpropagation cost.

In terms of efficiency, the asymmetric strategy decreases per-epoch LLM calls from $y_v \in \{0, 1\}$8 (naive) to $y_v \in \{0, 1\}$9, with a one-time $A \in \{0, 1\}^{|V|\times|V|}$0 cache build. Using the DigitalMusic dataset, this yields epoch time reduction from $A \in \{0, 1\}^{|V|\times|V|}$1 s (naive) to $A \in \{0, 1\}^{|V|\times|V|}$2 s (asymmetric), corresponding to a $A \in \{0, 1\}^{|V|\times|V|}$3 speedup, and permits batch size scaling from $A \in \{0, 1\}^{|V|\times|V|}$4 to $A \in \{0, 1\}^{|V|\times|V|}$5 (Li et al., 30 Jan 2026).

6. Empirical Evaluation and Baselines

FraudCoT, incorporating fraud-aware selective CoT distillation, is experimentally validated on InstantVideo (37,126 nodes, 9.9M edges), DigitalMusic (64,706 nodes, 7.7M edges), and PromotionAbuse (371K nodes, 1.3M edges). Comparisons are drawn against:

• GNNs: GraphSAGE, HGT, ConsisGAD, PMP, GAAP
• Pure LLMs: Qwen-8B, InstructGLM
• Graph-enhanced LLMs: LLaGA, GraphGPT, HiGPT
• LLM-enhanced GNNs: TAPE, FLAG

Evaluation metrics include Macro-F1, AUROC, and AUPRC (using scikit-learn implementations). FraudCoT achieves up to $A \in \{0, 1\}^{|V|\times|V|}$6 improvement in AUPRC over the best baseline (PromotionAbuse), with consistent performance gains across all public and industrial dataset splits. Ablation studies confirm that both the selective distillation and asymmetric co-training components are essential for optimal results (Li et al., 30 Jan 2026).

Dataset	Nodes	Edges	Best AUPRC Gain
InstantVideo	37,126	9.9M	up to +8.8%
DigitalMusic	64,706	7.7M	up to +8.8%
PromotionAbuse	371,000	1.3M	up to +8.8%

7. Implementation Strategies and Practical Considerations

Implementation guidelines include:

• Use LoRA adapters with rank $A \in \{0, 1\}^{|V|\times|V|}$7–$A \in \{0, 1\}^{|V|\times|V|}$8.
• Set unlikelihood loss $A \in \{0, 1\}^{|V|\times|V|}$9.
• Sample $N(v)$0 neighbors, with hop number $N(v)$1.
• Cache node embeddings on CPU or host memory to minimize GPU LLM calls.
• Apply early stopping based on validation AUROC; batch sizes $N(v)$2–$N(v)$3 provide optimal throughput.
• Precompute all text encodings for CoT scoring to accelerate distillation selection.

A plausible implication is that this design allows practitioners to scale LLM-enhanced GNNs to industrial TAGs with both superior predictive performance and orders-of-magnitude throughput improvements, without requiring expensive end-to-end LLM inference on all neighbors during training (Li et al., 30 Jan 2026).