Multi-Relationship Features based Residual Network for Fake Review Detection

The proliferation of e-commerce platforms has significantly affected user experience due to the advent of fake reviews. To circumvent platform detection, fake reviews have become increasingly challenging to identify. Contemporary approaches for detecting fraudulent reviews utilising Graph Neural Networks (GNNs) frequently experience network degradation and gradient vanishing problems during extensive training. Furthermore, it examines the impact of camouflage outcomes on the skewed tagging of reviews, which undermines the robustness of GNN detection models. Multi-relationship review detection on residual networks (MRDRN) is presented to combine multi-relational review elements to identify fraudulent reviews. Initially, the approach employs a residual network for comprehensive feature extraction to alleviate network degradation. It proposes an innovative neighbour-mixing sampling strategy, which conducts low-order and high-order neighbour-mixing sampling predicated on feature similarity among reviewers. This mitigates the issue of uneven review labelling and facilitates the acquisition of more nuanced review characteristics. Secondly, a multi-relational review feature fusion technique is developed, which mitigates the effects of review camouflage by incorporating the network architecture of reviews inside a relation and the comprehensive fusing of multi-relational review features. Experiments performed on three real-world datasets indicate that MRDRN exhibits superior detection capabilities and enhanced robustness relative to baseline approaches.