The implementation of deep learning techniques necessitates the gathering of information on several input variables or features in order to accurately train and forecast models. Data collection on input features may be expensive, thus it's critical to cut costs by creating a budget-constrained model (BCM) and choosing a subset of characteristics. In this research, we offer a method that uses Deep Neural Networks (DNNs) to exclude features that are not as significant for big data analysis. After a DNN model is created, we pinpoint the weak neurons and links and exclude some input features to reduce the model's cost to a predetermined amount. The experimental findings demonstrate the viability of our method and support user choice of an appropriate BCM within a specified budget.