Article Preview
Top1. Introduction
Non-Linear Machine Learning based regression models are one of the most challenging models in terms of verifying the correctness of predictions. Linear models, on the other hand, needs a lot of adjustment while modelling data which are mostly non-linear and violates most of the assumptions like normality. The generalized Additive model is being studied and improved with the best smoothing parameter function adapted to the model. The model is being applied to bioinformatics to study on Oral Squamous Cell Carcinoma (OSCC), which causes the highest mortality rate compared to other cancer types, especially for men. The generalized additive models are found to be useful, especially to model the complex non-linear impacts of some or all of the predictor variables. This paper introduces a Laplacian likelihood generalized additive model to tackle the non-linearity in oral squamous cancer between the predictor and dependent variables.
Machine Learning (ML) models could be non-linear but linear in terms of the undefined parameters in the variables. ML can also involve non-linear models that cannot be rendered linear in parameters after a transformation function being applied. It comprises of models which are linear having non-linear parameters. A big issue with fitting transformed variables is that after a transition, errors in the transformed variable cannot be assumed to be following a normal distribution. In predictive analytics, the error estimation is intended to be used for data which are distinct from the specific distribution related to the particular application. Without a robust mathematical justification, it is not desirable to feed transformed variables with non-linear predictive models. It is much more recommended to use a generic non-linear model approach than to transform the data and fit a specific model.
Generalized additive models (GAM) expand the standard linear models by linking with the expected value of Y to the input variables possible for non-linear modelling. It implies that an alternate distribution can be connected in addition to the normal distribution for the essential random variants. The Gaussian representations could be used for various statistical applications, but it is to be noted that certain types of problems are not suitable. GAM eases the standard parametric presumption and helps to discover complexity in the association among the dependent and the independent variable that might otherwise be neglected. Most non-parametric approaches, such as thin-plate smoothing spline and local regression method do not work well if a great deal of stand-alone variables is present in the model (Baquero et al., 2018).