Identification of E-commerce Anomalies Using Bayesian Semi-Supervised Tensor Decomposition Approach

In e-commerce, customer reviews are the highest priority because they impact an organization's profit. Due to the growing acceptance of online review mechanisms, competitors and suppliers of services are more motivated than ever to trick buyers by faking service recommendations. Fake reviews are now very common on e-commerce sites because fraudulent reviews significantly negatively impact the public. Humans find it difficult to interpret anomalies that may not be available in individual variables. Dense blocks are strongly connected parameters in tensors that represent indicators of atypical or anomalous behaviour in commercial platforms. However, the speed, precision, and flexibility of the existing approaches for detecting such dense blocks are insufficient. The typical Natural Gradient Descent NGD approach can slowly converge and may suffer from numerical instability. Tensor-based approaches for anomalous identification in e-commerce platforms have been suggested to solve this issue in recent years. Hence, an E-commerce Anomaly Detection using an Enhanced Bayesian Semi-Supervised Tensor Decomposition (EAD-EBSTD) approach is proposed to increase the factorization's precision that represents the tensor's underlying structure using the available label data. Canonical Polyadic Decomposition and Non-negative Matrix Factorization (CPD- NMF) hybridization are used to discover the dense review blocks in the input matrices, thus increasing the precision and effectiveness of the decomposition process. The Hessian-based enhanced NGD approach adjusts the gradient descent algorithm's learning rate by using the Hessian matrix, which describes the curvature of the loss function. The performance metrics show good scores, recall, precision, and Area Under the Curve Receiver Operating Characteristics curve (AUC-ROC), a good convergence for detecting fake review anomalies.