With the growth of e-commerce platforms, the number of fraud reviews has rapidly increased, making fraud detection for platform data critically important. To detect fraud reviews in platform data, recent approaches have leveraged graph neural networks that model users as nodes in a heterogeneous multi-relational graph structure. This approach represents platform users as nodes in a graph, and the relationships among users who share commonalities in their reviews or products as edges, in order to identify fraudulent reviewers. However, existing graph-based fraud detection methods may suffer from the unstable training of the classifier networks that process embedding vectors. In this paper, we identify that this issue arises from the excessive deviation and scale expansion caused by some of the aggregated values from adjacent nodes in conventional node embeddings, and we propose a scale regularization method to mitigate this. To verify the effectiveness of the proposed method, we conduct validation on the Amazon-Fraud dataset, which is a multi-relational graph dataset constructed from review data of Amazon E-Commerce. The experimental results show that the proposed scale regularization achieves superior performance compared to previous verified graph fraud detection models.

Time series anomaly detection is a technology that finds outliers in observed data over time, and is a significant research field associated with many applications or platforms. In this paper, we propose a method called Encoded-IF, a new and improved version of the isolation forest (IF) machine learning algorithm using Auto-Encoder to detect time series anomalies. Encoded-IF is a non-parametric and unsupervised method for detecting anomalies in time series data. It efficiently detects anomalies in each dimension of high-dimensional data and quickly learns nonlinear correlations between dimensions. Experiments were conducted on four publicly available time series data sets to evaluate the performance of the proposed method. The results improvement in performance of up to 4%.

As the volume of content that users search for in the internet environment rapidly increases, the demand for research on recommendation systems that provide personalized recommendations is also growing. Research on recommendation systems to date has mainly focused on GCN(graph convolutional network)-based models. However, the GCN-based models are less favorable for reflecting changing trends due to slower learning speeds. By the way, the GSP(graph signal processing)-based approaches, which rely on signal filtering techniques instead of convolutional operation, have been proposed recently. Therefore, in this paper, we analysis the effect of user ratings in GSP-based recommendation systems. The experimental results showed that the performance of recommendation systems are better when reflecting user rating scores in the GSP-based models.

This paper presents a novel image-level anomaly detection control method for defect inspection, which comprises feature- extraction from high-dimensional data and feature-learning to analyze the observed patterns. we focus on the feature-learning step, especially in Deep SVDD. To determine anomalies from images of object suspected to be defective, Deep SVDD learns the decision boundaries for the normal feature distribution in the latent space. When a new image is received by the defect inspection system, Deep SVDD maps its features to the latent space and detects anomalies based on whether the observed features fall within the decision boundaries for normal features. The decision boundary of Deep SVDD is determined based on the center-point, which is the center position of the decision boundary. The center-point is randomly fixed at the beginning of training of Deep SVDD. However, this limits the possibility of finding the optimal decision boundary. Therefore, we propose a center-point-moving method to optimize the center-point location and improve the performance of Deep SVDD. To confirm the suitability of the proposed method, we conducted experiments on MVTec-AD, a defect inspection dataset for several products in the manufacturing industry. Based on the results, the proposed method improved the performance of Deep SVDD in 11 of the 15 class domains in the MVTec-AD dataset.

In the manufacturing supply chain of smart devices, defect inspection for core components, such as lens actuator that constitute the final product, is essential for ensuring productivity and reliability. Especially, the ball-based lens actuator—which controls the movement of the lens through balls mounted in the camera module—is one of the essential devices that require inspection of structural features such as the position, number, and arrangement state of the balls. In the industrial component inspection domain, general-purpose anomaly detection model based on deep learning methods have primarily been used as a means of detecting visual defects; however, they are not suitable for inspecting anomalies in structural features, such as those found in lens actuator. For this reason, this study proposes an anomaly detection process based on structural features. In this study, the proposed method is designed to reduce false positives in defect inspection by explicitly detecting structural anomalies through pattern recognition and computer vision techniques. In this paper, the proposed method was applied to various types of experiments using a dataset of images collected from an actual actuator assembly process. The proposed method showed that it demonstrated superior performance, as well as robustness in invariant-feature recognition compared to general-purpose anomaly detection models. (13.75% higher precision than the conventional inspection system, and 1.48% to 35.84% higher recall than the conventional general-purpose anomaly detection models).