The lower image was reconstructed using a prediction model generated by a random forest algorithm. The data provided to train the algorithm consisted of a random sample of only 3% of the pixels from the full logo (shown in the upper image).
Interested in learning about modern data science techniques and applications? We offer training in data science and the R programming language specifically tailored for actuaries.
Adopting a new model can seem like a big step. We are able to deploy machine learning models in the cloud for visualisation and communication purposes. We can implement final models both within client based systems, or as independent cloud based solutions.
In traditional GLM models interactions between predictors must be carefully isolated and inserted manually. In contrast, the non-parametric nature of many machine learning algorithms means that all interactions can be included by default. Beyond simply using these models for predictions, we have a suite of tools to analyse the interactions present in those models, which can be used as pointers to refine and improve traditional models.
Applying modern machine learning techniques to traditional personal lines pricing data, we were able to provide predictive models which were more accurate (when measured against an independent test set) than a standard GLM model. Such techniques enable clients to exploit their data to the fullest.