Machine learning algorithms deployed for evolving streaming environments must handle the non-stationary data distributions, commonly referred to as concept drift. The presence of concept drift poses a major challenge for many real-world applications because it can severely degrade their predictive performance, hindering their ability to support robust decision-making. Consequently, the timely and efficient detection of drift events is critical for sustaining high accuracy over time. This study examines theoretically the concept drift characteristics and numerous drift detection algorithms across several categories. Furthermore, we evaluate their performance on both synthetic and real-world datasets exhibiting diverse streaming scenarios and drift characteristics, such as abrupt and gradual changes. This study aims to enhance understanding of the complex notion of concept drift characteristics and behavior of drift detectors, along with their applicability to diverse contexts.

To address the trade-off between computational efficiency and adherence to Kolmogorov-Arnold Network (KAN) principles, we propose TruKAN, a new architecture based on the KAN structure and learnable activation functions. TruKAN replaces the B-spline basis in KAN with a family of truncated power functions derived from k-order spline theory. This change maintains the KAN's expressiveness while enhancing accuracy and training time. Each TruKAN layer combines a truncated power term with a polynomial term and employs either shared or individual knots. TruKAN exhibits greater interpretability than other KAN variants due to its simplified basis functions and knot configurations. By prioritizing interpretable basis functions, TruKAN aims to balance approximation efficacy with transparency. We develop the TruKAN model and integrate it into an advanced EfficientNet-V2-based framework, which is then evaluated on computer vision benchmark datasets. To ensure a fair comparison, we develop various models: MLP-, KAN-, SineKAN and TruKAN-based EfficientNet frameworks and assess their training time and accuracy across small and deep architectures. The training phase uses hybrid optimization to improve convergence stability. Additionally, we investigate layer normalization techniques for all the models and assess the impact of shared versus individual knots in TruKAN. Overall, TruKAN outperforms other KAN models in terms of accuracy, computational efficiency and memory usage on the complex vision task, demonstrating advantages beyond the limited settings explored in prior KAN studies.

Studies on adopting Kolmogorov-Arnold Networks (KANs) for Continuous Reinforcement Learning (CRL) remain limited, even though traditional neural networks often struggle in CRL tasks. Our paper integrates a new KAN variant, termed TruKAN, into the Robust Policy Optimization (RPO) framework. TruKAN is distinguished by its knot placement, knot broadcasting and layer normalization. Knot locations can be either fixed or learned during training. Knot broadcasting determines whether knot sequences are shared across all output channels or maintained as separate, per-channel sets. We propose the TruKAN-RPO framework by integrating TruKAN into the actor and critic components of RPO. We evaluate the performance of two TruKAN-RPO models, TruKAN-RPO-Fixed and TruKAN-RPO-Learnable, and compare them with KAN-RPO and SineKAN-RPO. We train and test these different models using the "MuJoCo-MJX Humanoid" environment under different settings.

This paper explores integrating SineKAN networks into the Soft Actor-Critic (SAC) algorithm for Continuous Reinforcement Learning (CRL) to address the size and speed limitations of Kolmogorov-Arnold Networks (KANs). We modify SineKAN by adding residual connections to each hidden layer and then introduce two other variants. The latter retain the modified SineKAN’s architecture but differ in parameter update strategies and initialization. Experiments across several CRL environments show that SineKAN-based SAC models achieve strong performance and stability under certain conditions. Notably, one variant reduces the number of learnable parameters to less than one-quarter of the original SineKAN. We also analyze deeper architectures’ performance to assess scalability for complex CRL problems. Our findings provide insights into adopting SineKAN in continuous settings and guide future work on building more robust RL algorithms.

This study explores the integration of the Kolmogorov-Arnold Network (KAN) model into continuous reinforcement learning (CRL). We incorporate KAN and its improved variant, MultKAN, into the on-policy PPO algorithm, resulting in a new approach called KAN-PPO. Using the Mujoco Half-cheetah environment, we conduct extensive experiments to compare the performance and generalization of KAN-PPO models against MLP-PPO structures. Additionally, we examine the effects of dropout regularization through the DropKAN method in dense network settings. Our findings demonstrate the effectiveness of KAN and MultKAN models in CRL, highlighting their potential for complex learning tasks.

This paper is the first to explore the integration of Kolmogorov-Arnold networks (KANs) into off-policy methods for continuous reinforcement learning (CRL) tasks. We introduce KAN-SAC, a method that integrates the KAN model and its variants, namely MultKAN and SineKAN, with the Soft Actor-Critic (SAC) algorithm. The integration is based on the embedding of the KAN architecture in both actor and critic networks. Using the Mujoco Half-Cheetah environment as a case study, we evaluate the performance of these KAN-based SAC algorithms against traditional MLP-based SAC. Our results show that KAN models have great potential, even outperforming MLP models in certain scenarios. However, further refinement of these methods is needed before they can be used as a robust alternative in complex CRL applications.

Data distributions often change over time in non-stationary environments, rendering models trained on previous data invalid in deployed environments. To prevent a decline in predictive performance, machine learning models must monitor the incoming data stream to capture changes in data distributions, known as Concept Drift (CD). This study extends a probabilistic real-drift detection method by adopting a new strategy to detect the CD and adapt the underlying model to learn new concepts accordingly. Detecting such drift is critical, allowing for timely model updates based on new data, thereby mitigating potential prediction errors. Additionally, we refine the detector’s thresholds to better distinguish between drifting and non-drifting samples. Using established stream generators, we create diverse synthetic datasets to compare the performance of our enhanced method with state-of-the-art drift detection techniques under both abrupt and gradual drift scenarios. This broad range of drift scenarios allows for thorough assessment. The proposed approach demonstrates high performance across different base learners and varied data streams and often outperforms other methods.

This research primarily focuses on tackling the challenge of forecasting univariate time series. Existing methods for training forecasting models range from traditional statistical models to domain-specific algorithms, and more recently deep neural network models typically rely on raw time series observations, with alternative representations like higher-dimensional embedding used mainly for auxiliary analyses such as Topological Data Analysis (TDA). In contrast to conventional time series analysis methods, this study explores the impact of higher-dimensional embedding as a primary data representation. Leveraging this higher-dimensional embedding, we introduce a geometrical realization model that captures crucial data points of the embedding representation. Subsequently, we propose a deep neural network model inspired by N-BEATS, incorporating a TDA model, an attention model, and a convolutional neural network (CNN) model in parallel as sub-modules alongside the geometrical realization model. To assess the efficacy of the proposed model, we conduct evaluations on diverse time series datasets spanning various domains, including electricity load demands and M4 competition datasets. Furthermore, we conduct an ablation study to analyze the specific contributions of each sub-module towards the final predictions.

Data clustering has many applications in medical sciences, banking, and data mining. K-means is the most popular data clustering algorithm due to its efficiency and simplicity of implementation. However, K-means has some limitations, which may affect its effectiveness, such as all the features having the same degree of importance. To address these limitations and improve K-means accuracy, we adopt the Biogeography-Based Optimization (BBO) algorithm to select the most relevant features of datasets. Our primary idea is to reduce the intra-cluster distance while increasing the distance between clusters.

Concept drift detection in noisy data streams is challenging yet essential. This paper introduces NPRDD, a new concept drift detection algorithm that is robust to noise and accurately identifies Real drifts. NPRDD operates on a moving window of recent data, utilizing predicted class probabilities and cross-entropy-based surprise measures to weigh real drift candidates. In line with the Bayesian definition of Real concept drift, NPRDD considers a sample as a drift candidate when the classifier makes an error but is highly confident in its judgment. We evaluate NPRDD on synthetic datasets by varying the noise levels and comparing its performance with other well-established methods. Our results show that NPRDD outperforms other methods regarding ROC-AUC and Accuracy metrics.

Concept drift, which indicates data-distribution changes in streaming scenarios, can significantly reduce predictive performance. Existing concept drift detection methods often struggle with the trade-off between fast detection and low false alarm rates. This paper presents a novel concept drift detection algorithm, called SPNCD*, based on probabilistic methods, particularly Sum-Product Networks, that addresses this challenge by offering high detection accuracy and low mean lag time. Based on three benchmark datasets, the proposed method is evaluated against state-of-the-art algorithms, such as DDM, ADWIN, KSWIN, and HDDM_A. Our experiments demonstrate that SPNCD* outperforms the existing algorithms in terms of true positive rate, recall, precision, and mean lag time while improving the performance of the base classifier. The SPNCD* algorithm provides a reliable solution for detecting concept drift in real-time streaming data, enabling practitioners to maintain their machine learning models’ performance in dynamic environments.

Data clustering has many applications in machine learning, data mining and image processing. K-means is the most popular clustering algorithm due to its efficiency and simplicity of implementation. However, K-means has limitations, such as large feature spaces, which may affect its effectiveness. To improve K-means accuracy, we adopt the Biogeography-Based Optimization (BBO) evolutionary technique to select the most relevant features of datasets. We conducted several experiments to compare our approach with other methods, such as PCA and Particle Swarm Optimization (PSO). The results demonstrate the effectiveness of BBO for feature selection.

This work presents a deep averaged reinforcement-learning approach to learn improvement heuristics for route planning. The proposed method is tested on the Traveling Salesman Problem (TSP). While learning improvement heuristics using machine learning models are prosperous, these methods suffer from low generalization and forgetfulness of the agents during the training process. We have applied the stochastic weight averaging method during the training phase to solve these issues, which smothers the training convergence and prevents the forgetting of optimized learned policies, and consequently provides better results. The agent can learn the optimized policy while holding a moving average of the previously learned policies during the training epochs. In order to assess the performance of our proposed approach, we conducted comparative experiments considering other known methods from the literature. The results demonstrate our proposed method’s superiority in training trends and optimization.

Our research addresses the actual behavior of the credit-card fraud detection environment where financial transactions containing sensitive data must not be amassed in a considerable amount to train robust classifiers. We introduce an adaptive learning approach that adjusts frequently and efficiently to new transaction chunks; each chunk is discarded after each training step. Our approach combines transfer learning and incremental feature learning. The former improves the feature relevancy for subsequent chunks, and the latter increases performance during training by dynamically determining the optimal network architecture for each new chunk. We show the effectiveness and efficiency of our approach experimentally on an actual fraud dataset.

Detecting fraud accurately in credit cards is critical as this financial sector incurs significant losses for cardholders. Nonetheless, most studies adopted standard machine learning and few incremental learning, which are inadequate for addressing credit card challenges, such as rapid data arrival, unlimited data, data sensitivity, and performance decline over time. For this purpose, we propose a chunk-based incremental feature learning approach that optimises the fraud model topology for each new chunk and keeps track of one chunk each time. The model consists of several connected sub-models, where a new sub-model is optimally created for each new chunk. To avoid the network growing indefinitely, we limit the number of sub-models. To this end, we retain the most relevant sub-models to the current chunk's data distribution and re-combine them to create the optimal model. We evaluate our approach using two credit card datasets: the first of medium scale contains 2-day payments in 2013, and the second of considerable scale possesses 6-month payments in 2019. We split these datasets into multiple chunks to learn and test incrementally. We compare our approach with static learning methods trained with different scenarios. Moreover, we vary the number of historical sub-models to check their impact on the predictive performance.

Our study proposes a dimensionality reduction approach to efficiently process a service monitoring application’s high-dimensional and unlabeled time-series dataset. The approach aims to improve data quality and lower the feature space optimally. Since the dataset is vast and the reduction approach requires colossal resources, we divide it into several weekly sub-datasets. Using clustering methods and metrics, we evaluate the approach’s efficacy on each sub-dataset thoroughly and show that the information loss was minimal after data transformation. Moreover, we assess each sub-dataset’s trustworthiness and similarity to verify whether the new data acquire the same cluster labels as the initial data. Since the experiments reveal a high data quality, the industrial partner can utilize the new data in their decision-making tasks.

This study aims to improve disease detection accuracy by incorporating a discrete version of the Whale Optimization Algorithm (WOA) into a supervised classification framework (KNN). We devise the discrete WOA by redefining the related components to operate on discrete spaces. More precisely, we redefine the notion of distance (between individuals in WOA), and propose a random exploration function to include more diversity. The latter includes the random move defined in the WOA algorithm, as well as two other random techniques based on the crossover and mutation operators. To assess the performance of our proposed method, we conducted experiments on two benchmark medical datasets. The results demonstrate the efficacy of the hybrid approach, WOA+KNN.

Network traces represent a critical piece of data for network security. Due to lack of expertise, companies are forced to outsource their network traces to third parties to perform analytics on the traces and provide security feedback and recommendations. However, these companies are reluctant to share their network traces, as they comprise sensitive information (e.g., IP addresses). Therefore, the network traces are anonymized to ensure the privacy of the data and preserve its utility. The latter guarantees that the essence of the data remains valid after anonymization, otherwise the analytics are useless. Existing solutions, such as CryptoPAN, preserves the data utility (by preserving the IP prefixes), but are vulnerable to semantic attacks.In this paper, we propose an anonymization solution, which is based on the Feistel, which is widely used in encryption systems, such as DES and Twofish. Our solution preserves both data privacy and its utility at the same time. We validate our solution using the Kddcup99 dataset and measure the data leakage (dual of privacy) provided by our solution. We evaluate the security of our solution using the avalanche property, which is widely used to measure the security of encryption systems. Moreover, the efficacy of our solution is evaluated against Injection attacks. Overall, the obtained results, avalanche property and resistance to Injection attacks, are appealing.