Bisecann: A Deep Learning Model from the Bisection Mechanism and Neural Network for Time Series Cases

Modeling in deep learning (DL) makes many people curious about a solvable case. This approach serves to discover regularities/patterns and relationships within the framework of large data sets with the lowest possible error rate. The development of this technique uses programming to learn from the past or part of the object, and then this information is used to identify data. The vanishing gradient is one of the weaknesses of DL; one way to overcome this is by adding a bias factor to the artificial neural network (ANN) model to shift the activation function to the right or left. The bias factor is usually constant; some modeling uses old data values. This paper aims to create a new deep learning model architecture by adopting the mechanism of one of the closed numerical methods, namely Bisection. The actual function is to replace the flexible bias value in the ANN and depends on the current and previous data input. The use of these two methods in this paper is then called BISECANN. This paper also presents accuracy observations to determine the level of model-learning evaluation in several cases. It compares several evaluation metrics from other DL methods in predicting the same case.