Abinash Pujahari

In machine learning (ML), class imbalance is a serious problem when datasets with a dominating majority class make it difficult to accurately classify minority samples. Traditional methods like the Synthetic Minority Oversampling Technique (SMOTE) may lead to producing the noise and overfitting. Consequently, model performance degrades. In order to address this problem, we develop an innovative technique Adaptive K-Nearest Neighbor-Based Oversampling (AKS), a novel approach that strategically generates synthetic minority samples using an adaptive neighborhood selection and non-linearity checks among the considered minority samples. AKS avoids to generate synthetic points in majority regions, which attempts to reduce over-fitting problem. Further, the proposed model is applied on ten well known imbalanced datasets to make them balanced. To test the utility of the balanced dataset, seven ML algorithms are applied on such datasets. To validate the usefulness of the proposed method, SMOTE, SMOTE-ENN and borderline-SMOTE are being utilized. Moreover, the models performance is validated through 5-fold cross-validation methodology. In addition, statistical analyses, including the Friedman test are conducted to rigorously evaluate and validate the robustness of the proposed method. The results shows that, the proposed technique performs at par or better than other data balancing techniques.

Recent graph neural network-based collaborative filtering models are capable of representing complex user-item interaction information, which is crucial when datasets exhibit high sparsity. Many heterogeneous GNN architectures have been introduced to capture higher-order user-item interaction information, such as neighborhood information among items and item-item interaction sequential information. However, recent GNN-based CF models overlook the relationship between item sequence information and the semantic similarity between items. Hence, in this study, we propose a GNN framework for collaborative filtering that fuses item-item interaction signals derived using item-item interaction sequential graph which in turn derived from the order of all user-item interactions, with semantic similarity, utilizing two graphs: the first is a user-item bipartite graph, and the second is used for propagating item-item semantic relationship information. This helps the recommender system model to effectively represent users’ preferences towards items by using item feature information. We conducted experiments on three benchmark datasets—ML-1M, Amazon Books, and Yelp — using standard ranking-based measures. The results show significant improvements in the recommendation quality over the related state-of-the-art baselines.

In many real-world datasets, the class distribution is often highly imbalanced, and minority class samples are located within the majority regions, leading to significant overlap. These challenges produce a model of bias and misclassifications, particularly for minority classes. To address these issues, we introduced a novel technique that integrates Affinity Propagation with Adaptive Synthetic Sampling (ADASYN) and Random Oversampling (ROS), collectively termed as APAROS. Using Affinity Propagation Clustering (APC), our proposed method categorizes the data set into Overlapping Clusters (OLC), Pure minority Clusters (PmC) and Pure Majority Clusters (PMC). Further, we applied existing oversampling methods; ADASYN in OLC and ROS in PmC, effectively balancing the dataset while minimizing the risk of miss-classification in overlapped regions. The proposed method was evaluated on ten benchmark imbalanced datasets and compared against existing resampling techniques such as ADASYN, ROS and Kmeans-SMOTE. Seven machine learning classifiers were employed to evaluate the models performance considering the metrics such as accuracy, precision, recall, f1_score, Matthews Correlation Coefficient (MCC) and Cohen’s Kappa coefficient. Experimental results consistently demonstrate that APAROS performs significantly better than traditional resampling techniques and produced superior classification performance, particularly in highly overlapped and imbalanced scenarios.