Source-linked AI summary
Spam Review Detection with Graph Convolutional Networks
Ao Li, Zhou Qin, Runshi Liu, Yiqun Yang, Dong Li
TL;DR
Spam advertisements on Xianyu undermine user experience and create a setting where spammers exploit varied patterns to mislead buyers. The paper proposes GAS, a scalable GCN-based method combining local and global comment context, and reports superiority to a baseline plus effective million-scale deployment.
Problem
Spam reviews on Xianyu influence purchasing decisions, undermine user experience, and create a setting where spammers publish malicious information using varied patterns.
Method
GAS is a highly scalable graph-convolutional anti-spam method that combines local publisher-item context with global comment-feature context using heterogeneous and homogeneous graph representations.
Results
GAS significantly outperforms the baseline model, further improves results with global context, and is deployed with distributed processing for Xianyu’s large-scale data.
Takeaways & Limitations
The reported offline and online results support GAS as an effective approach for scalable spam-advertisement detection at Xianyu.
Abstract
from arXiv · showhide
Customers make a lot of reviews on online shopping websites every day, e.g., Amazon and Taobao. Reviews affect the buying decisions of customers, meanwhile, attract lots of spammers aiming at misleading buyers. Xianyu, the largest second-hand goods app in China, suffering from spam reviews. The anti-spam system of Xianyu faces two major challenges: scalability of the data and adversarial actions taken by spammers. In this paper, we present our technical solutions to address these challenges. We propose a large-scale anti-spam method based on graph convolutional networks (GCN) for detecting spam advertisements at Xianyu, named GCN-based Anti-Spam (GAS) model. In this model, a heterogeneous graph and a homogeneous graph are integrated to capture the local context and global context of a comment. Offline experiments show that the proposed method is superior to our baseline model in which the information of reviews, features of users and items being reviewed are utilized. Furthermore, we deploy our system to process million-scale data daily at Xianyu. The online performance also demonstrates the effectiveness of the proposed method.
1 INTRODUCTION
Xianyu’s spam-advertisement detection problem combines massive scale with adversarially varied comments. The proposed GAS method uses heterogeneous and global graph context to improve detection and supports million-scale daily processing.
- Problem: Spam advertisements constitute the majority of Xianyu’s spam comments, motivating the paper’s focus on their detection.The paper distinguishes spam advertisements from vulgar comments and focuses on the larger category.
- Challenges: Xianyu’s anti-spam system must handle over 1 billion goods published by over 10 million users while resisting adversarial actions.The two stated challenges are scalability and performance decay caused by spammers’ countermeasures.
- Approach: The model combines local context from publishers and items with global context from the feature distribution of all comments.Global context is added beyond the heterogeneous graph’s local context, further improving the result.
- Approach: GAS is a highly scalable anti-spam method based on graph convolutional networks for Xianyu.The method uses a heterogeneous graph with edge attributes and can generalize to meta-path-based heterogeneous GCNs.
- Deployment: The deployed system processes million-scale comments daily at Xianyu and remarkably identifies more spam comments while satisfying efficiency requirements.Offline experiments and online evaluation support the reported effectiveness against adversarial actions.
2 RELATED WORK
Prior work progresses from engineered review, reviewer, and product features to graph-based and graph-convolutional spam detection. This paper applies GCNs to spam review detection and combines heterogeneous and homogeneous graph contexts for Xianyu.
- Earlier spam detection methods primarily use engineered features from review content, reviewer behavior, and products.
- Graph-based methods model reviewer, store, review, or user-similarity relationships to detect spam or estimate trustworthiness.
- GCNs aggregate local-neighbor features across multiple hops, while GraphSAGE supports inductive embeddings and GAT weights neighbors with attention.
- Most graph methods focus on homogeneous graphs, whereas heterogeneous methods address multiple node or link types with attention-based aggregation.
- To the authors’ knowledge, this paper is the first to apply a GCN-based method to spam review detection.
- GAS extends GCNs to Xianyu’s attributed bipartite graph and incorporates a homogeneous Comment Graph to capture local and global context.
E · AGGl
The comment-edge aggregation collects information from the edge and its connected user and item nodes, while the overall GAS architecture also adds homogeneous-graph comment embeddings.
- For a user or item, aggregation collects neighboring node information together with attributes of connected comment edges.
- The heterogeneous GCN produces user, item, and comment embeddings, while the homogeneous Comment Graph produces an additional comment embedding.
- The displayed aggregation expression specifies how neighborhood information is calculated for item nodes.
I · AGGl
The heterogeneous GCN uses type-specific parameters and aggregation functions, attention-based neighbor weighting, and scalable time-aware mini-batch sampling.
- User and item nodes maintain different trainable parameters and aggregation functions within the heterogeneous graph convolution.
- Attention maps a feature vector and candidate feature vectors to a weighted sum, with weights computed by scaled dot-product attention.
- The heterogeneous method generalizes to meta-path-based graph convolution for graphs with varied node and edge types.
- Mini-batch training samples a fixed number of neighbors because whole-batch training is impractical at Xianyu Graph’s scale.
- When candidates exceed the sample limit, the method selects the temporally closest comments; when fewer exist, it pads with placeholders.
- Time-based sampling favors related comments, while padding preserves sparse neighborhood distributions without resampling.
- Algorithm 1 propagates edge, user, and item hidden states through the heterogeneous Xianyu Graph.
- Each comment edge updates from its prior state and the prior states of its linked user and item.
I · AGGl
The method combines text embeddings, heterogeneous user-item context, and a homogeneous Comment Graph to capture both local and global information for spam detection. Time-aware neighbor sampling and embedding smoothing support scalable training and improved separability, while increased propagation depth introduces noise.
- Scalable aggregation: Time-related sampling selects a fixed number of temporally closest neighboring comments for item and user aggregation in mini-batches.The sampling strategy replaces random neighbor selection and supports training on massive data.
- Comment representation: TextCNN converts comment text into embeddings that are trained end-to-end with the graph neural network.Word2vec-pretrained word embeddings are used as TextCNN inputs.
- Global context: The Comment Graph connects comments with similar contents, providing supplementary global context when local user-item context is insufficient.Approximate KNN Graph construction makes similarity-based graph generation practical instead of scanning all comment pairs.
- GAS model: GAS combines inductive GCN embeddings from the Comment Graph with heterogeneous GCN embeddings from the Xianyu Graph in an end-to-end model.The final comment representation concatenates the Comment Graph embedding with embeddings learned from the heterogeneous graph.
- Local context: The heterogeneous Xianyu Graph represents users and items as nodes and comments as edges, enabling local-context aggregation.For each comment edge, the associated user and item embeddings are learned from their neighboring nodes.
- Smoothing and depth: Embedding smoothing improves sample separability, whereas increasing propagation layers provides little performance benefit because the added noise cannot be ignored.Smoothing averages self and neighbor features; logistic-regression evaluation reports improved AUC and F1-score for smoothed embeddings.
4 EXPERIMENTS
The experiments evaluate GAS against feature-based and neural baselines offline, then test deployed models online and examine cases showing how local and global context improve spam detection.
- Offline evaluation: The offline evaluation compares GBDT, TextCNN+MLP, GAS-local variants, and GAS using AUC, F1-score, and recall@90% precision.GAS-local-1 and GAS-local-2 use one and two propagation layers on the Xianyu Graph, while GAS adds one Comment Graph layer.
- Offline evaluation: GAS-local-1, GAS-local-2, and GAS outperform GBDT and TextCNN+MLP, and GAS gains further improvement over GAS-local-2 by incorporating global context.The offline comparison is shown in Table 2 and Figure 7.
- Offline evaluation: 4.33 F1-score lift: GAS outperforms the deployed baseline system, while fixed 90% accuracy yields an extra recall of 16.16%.These are author-reported overall offline results against the deployed baseline system.
- Online evaluation: In daily production, GAS-local-1 and GAS consistently detect more spam comments than TextCNN+MLP, while GAS consistently exceeds GAS-local-1.The online comparison uses models deployed in Xianyu’s production environment, with detected spam manually checked at 90% precision.
- Case study: Case analyses associate local-context gains with similar advertisements linked by users or items, while global context helps capture deformed advertisements connected through the Comment Graph.The analyzed global-context cases include adversarial advertisements that are not connected through the same account or item.
5 SYSTEM IMPLEMENTATION AND DEPLOYMENT
The deployed GAS system uses a daily graph-building and distributed TensorFlow workflow to detect spam at Xianyu’s large operational scale.
- System workflow: Xianyu constructs a heterogeneous graph daily from recent-month comment logs stored on MaxCompute.The workflow begins when app comments are logged, after which the daily graph and KNN Graph are constructed.
- System workflow: A distributed TensorFlow implementation of GAS detects spam, removes detected spam from the app, and may penalize malicious accounts.The workflow is presented as the production anti-spam platform architecture.
- Distributed implementation: The parameter-server architecture distributes graph storage, data fetching, training, and prediction across 8 parameter servers and 8 workers.Each worker uses an Nvidia V100 GPU, 6 CPU cores, and 32GB memory; the passage introduces these resources as the distributed deployment configuration.
- Distributed implementation: Graph structure, vertex features, and edge features are stored as adjacency lists and feature data on parameter servers because the Xianyu Graph is too large for one machine.The distributed storage design also supports memory-efficient graph representation.
- Efficiency considerations: Even with caching, workers spend about 41% of time fetching information from parameter servers, while caching saves about 30% of training time.The network-connected parameter-server design avoids memory overflow but reduces efficiency relative to local memory access.
- Efficiency considerations: 2 hours: the distributed implementation reduces the offline GAS training time to two hours.This training-time result refers to the offline experiment described in Section 4.1.
6 CONCLUSION
The paper addresses Xianyu’s scalability and adversarial-action challenges with GAS, integrating local and global comment context. Offline and online evaluations support its effectiveness, while deployment details demonstrate operational use at Xianyu.
- Conclusion: Xianyu spam detection faces scalability and adversarial-action challenges addressed by an end-to-end GCN-based Anti-Spam algorithm.The conclusion identifies these as the two main challenges and names the proposed method GAS.
- Conclusion: GAS incorporates both local and global comment context, and offline and online evaluations demonstrate the method’s effectiveness at Xianyu.The conclusion also states that real-world cases further examine how the different contexts alleviate adversarial actions.
- Conclusion: The paper concludes by describing GAS implementation, deployment, and workflow at Xianyu.This connects the proposed algorithm to its production use within the platform.