Due to market deregulation and globalisation, competitive environments in various sectors continuously evolve, leading to increased customer churn. Effectively anticipating and mitigating customer churn is vital for businesses to retain their customer base and sustain business growth. This research scrutinizes 214 published articles from 2015 to 2023, delving into customer churn prediction using machine learning methods. Distinctive in its scope, this work covers key stages of churn prediction models comprehensively, contrary to published reviews, which focus on some aspects of churn prediction, such as model development, feature engineering and model evaluation using traditional machine learning-based evaluation metrics. The review emphasises the incorporation of features such as demographic, usage-related, and behavioural characteristics and features capturing customer social interaction and communications graphs and customer feedback while focusing on popular sectors such as telecommunication, finance, and online gaming when producing newer datasets or developing a predictive model. Findings suggest that research on the profitability aspect of churn prediction models is under-researched and advocates using profit-based evaluation metrics to support decision-making, improve customer retention, and increase profitability. Finally, this research concludes with recommendations that advocate the use of ensembles and deep learning techniques, and as well as the adoption of explainable methods to drive further advancements.
1 January 2022 in Expert Systems with Applications
Smart grid is an emerging technology which is considered to be an ultimate solution to meet the increasing power demand challenges. Modern communication technologies have enabled the successful implementation of smart grid (SG), which aims at provision of demand side management mechanisms (DSM), such as demand response (DR). In this paper, we propose a hybrid technique named as teacher learning genetic optimization (TLGO) by combining genetic algorithm (GA) with teacher learning based optimization (TLBO) algorithm for residential load scheduling, assuming that electric prices are announced on a day-ahead basis. User discomfort is one of the key aspects which must be addressed along with cost minimization. The major focus of this work is to minimize consumer electricity bill at minimum user discomfort. Load scheduling is formulated as an optimization problem and an optimal schedule is achieved by solving the minimization problem. We also investigated the effect of power-flexible appliances on consumers’ bill. Furthermore, a relationship among power consumption, cost and user discomfort is also demonstrated by feasible region. Simulation results validate that our proposed technique performs better in terms of cost reduction and user discomfort minimization, and is able to obtain the desired trade-off between consumer electricity bill and user discomfort.
All over the world, deaf people use sign language as the only reliable source of communication with each other as well as with normal people. These communicating signs are made up of the shape of the hand and movement. In Pakistan, deaf people use Pakistan sign language (PSL) as a means of communication with people. In scientific literature, many studies have been done on PSL recognition and classification. Most of these work focused on colored-based hands while some others are sensors and Kinect-based approaches. These techniques are costly and also avoid user-friendliness. In this paper, a technique is proposed for the recognition of thirty-six static alphabets of PSL using bare hands. The dataset is obtained from the sign language videos. At a later step, four vision-based features are extracted i.e. local binary patterns, a histogram of oriented gradients, edge-oriented histogram, and speeded up robust features. The extracted features are individually classified using Multiple kernel learning (MKL) in support vector machine (SVM). We employed a one-to-all approach for the implementation of basic binary SVM into the multi-class SVM. A voting scheme is adopted for the final recognition of PSL. The performance of the proposed technique is measured in terms of accuracy, precision, recall, and F-score. The simulation results are promising as compared with existing approaches.
Since last few decades, class imbalance has been one of the most challenging problems in various fields, such as data mining and machine learning. The particular state of an imbalanced dataset, where each class associated with a given dataset is distributed unevenly. This happens when the positive class is much smaller than the negative class. In this case, most standard classification algorithms do not identify examples related to the positive class. A positive class usually refers to the key interest of the classification task. In order to solve this problem, several solutions were proposed such as sampling-based over-sampling and under-sampling, changes at the classifier level or the combination of two or more classifiers. However the main problem is that most solutions are biased towards negative class, computationally expensive, have storage issues or taking long training time. An alternative approach to this problem is the genetic algorithm (GA), which has shown the promising results. GA is an evolutionary learning algorithm that uses the principles of Darwinian evolution, it is a powerful global search algorithm. Moreover, the fitness function is a key parameter in GA. It determines how well a solution can solve the given problem. In this paper, we propose a solution which uses entropy and information gain as a fitness function in GA with an objective to improve the impurity and gives a more balanced result without changing the original dataset. Experiments conducted on different datasets demonstrate the effectiveness of the proposed solution in comparison with several other state-of-the-art algorithms in term of Accuracy (Acc), Geometric mean (GM), F-measure (FM), kappa, and Matthews correlation coefficient (MCC).