In this paper, we present Modified ADASYN (M-ADASYN), an oversampling technique designed to address class imbalance in fine-grained fraud classification tasks. Unlike traditional oversampling methods that often generate noisy synthetic samples near decision boundaries, M-ADASYN incorporates Reciprocal Nearest Neighbors (RNN) to focus on densely populated minority regions, minimizing overlap with majority class data. This approach improves the quality of synthetic samples, reduces false positives, and enhances model generalization. Extensive experimental evaluation demonstrates that M-ADASYN outperforms methods such as SMOTE, ADASYN, and GAN-based approaches, by an average of 3 percentage points improvement in classification accuracy. Furthermore, M-ADASYN is shown to reduce training time while maintaining model performance across multiple classification models, including Random Forest and LightGBM. The proposed method’s integration of RNN ensures more efficient handling of extreme class imbalance, leading to improved accuracy and computational efficiency.

In this work, we propose Dynamic Kernel Tuning (DKT), which dynamically selects a subset of kernels per attribute for the pedestrian attribute recognition problem. By tuning kernels for each attribute class in real time, DKT optimizes both training and inference times. Experimental evaluation of DKT on the Market-1501S dataset achieves a reduction in training time from 234 seconds to 32.5 seconds and a reduction in inference time from 35 ms to 2.3 ms, without impacting accuracy. Additionally, we benchmark model quantization methods and establish theoretical results to foster future research in dynamic kernel tuning.

Missing data imputation of transportation occupancy data faces a big challenge, specifically in situations where the entire trip of a vehicle is missing due to no sensors. Traditional imputation algorithms such as k-nearest neighbor (KNN), Multi-recurrent Neural Network (MRNN), Bidirectional recurrent imputation for time series (BRITS), and Transformer-based algorithms like self-attention-based imputation for time series (SAITS) and time-aware-self-attention-based imputation for time series (TA-SAITS) usually have poor adaptiveness to dataset variations since their performances depend highly on dataset characteristics. Even typical ensemble techniques such as bagging, boosting, and stacking do not update dynamically regarding various patterns in the same dataset. To overcome those challenges, this work proposes an Adaptive Ensemble Imputation based on Deep Reinforcement Learning, AEI-DRL. AEI-DRL uses a sliding window approach to scroll over sub-parts of the data and learns the optimal weights assignment of multiple imputation methods. This enables AEI-DRL to dynamically adjust weightage towards alternative algorithms or their combinations for different data segments. Experimental results show improvements in imputation results using AEI-DRL. It achieved a 5.3% and 2.6% reduction in mean square error (MSE) on the transportation occupancy dataset compared with the second-best method. On a non-transport dataset, AEI-DRL was able to record a reduction of MSE 16%, hence proving its applicability across time-series datasets with all sorts of diverse missing data patterns.

Simple Online and Real-time Tracking with a Deep Association Metric (DeepSORT) has been widely used due to its simplicity and strong empirical performance on Multiple Object Tracking (MOT). However, in real-world applications involving low frame rate (LFR) videos, DeepSORT requires practitioners to tune its hyperparameters to handle abrupt changes in motion and reduce erroneous tracks. Additionally, it is currently unknown how sensitive the DeepSORT's performance is to changes in these hyperparameters for LFR videos. Thus we conduct a sensitivity analysis of DeepSORT to explore the effect of these hyperparameters on the overall performance in LFR videos. Our main aim is to understand the impact of hyperparameter and identify crucial choices for DeepSORT in LFR-MOT. We finally present practical recommendations based on our extensive empirical study for those interested in getting the most out DeepSORT for LFR-MOT in real-world settings.