A Flexible Monotone Neural Network for Data Mining

Incorporating the monotonicity prior into deep learning models that arc used for data mining in domains like finance, e-commerce marketing might be crucial at times. One class of existing methods uses monotonicity regularization to train monotonic neural networks, but not all of them are assured to provide monotonic predictors and others require additional validation procedures. By altering the network structure, a different class of approaches, including Deep Lattice Networks (DLN), can ensure monotonicity. The link between the inputs and outputs of the DLN, however, is step-wise and counterintuitive due to its complicated structure. In this study, a monotonous network called the Deep isotonic Embedding Network (DIEN) is developed. Different modules are used by DIEN to address monotonic and non-monotonic features individually, and the outputs of these modules are then linearly merged to get the prediction result. We demonstrate how DIEN can ensure that the learned model is monotonic in relation to particular features. In addition, compared to DLN, DIEN does not need to construct intricate structures like lattice. Results from experiments on both synthetic and real-world datasets demonstrate that DIEN outperforms existing methods.