Predicting the risk of suffering chronic social exclusion with machine learning

Abstract

The fight against social exclusion is at the heart of the Europe 2020 strategy: 120 million people are at risk of suffering this condition in the EU. Risk prediction models are widely used in insurance companies and health services. However, the use of these models to allow an early detection of social exclusion by social workers is not a common practice. This paper describes a data analysis of over 16K cases with over 60 predictors from the Spanish region of Castilla y León. The use of machine learning paradigms such as logistic regression and random forest makes possible a high precision in predicting chronic social exclusion. The paper is complemented with a responsive web available online that allows social workers to calculate the risk of a social exclusion case to become chronic through a smartphone.

Towards Social Care Prediction Services Aided by Multi-agent Systems

“…Although the architecture has not been implemented, many of the ideas and proposals have been put into practice for the generation of an online social exclusion prediction service in the Spanish region of Castilla y Le´on [17]. Future work includes: implementing the architecture in a multi-agent plat form; extending the decision support system for the labeling service; considering different machine learning problems as multi-instance learning and multi-label learning; and, a better exploitation of the ontologies and semantic resources to include forms of advanced learning such as case-based reasoning, transfer learn ing, and graph mining.…”

“…According to the number of variables and their possible values, the Layer 2 agents may report for their study: a hierarchical clustering dendrogram to compare the similarity between the representatives (so that if one receives a label, the most similar cases could, but not necessarily, receive similar labels); multi-dimensional visualizations (such as Chernoff faces or Star diagrams); plots of the two or three most important variables obtained after a dimensionality reduction process if they are reasonably representative of all dimensions considered, etcetera. In a prediction service of the chronicity of social exclusion built with the ideas presented here [17], 63 variables were considered, hence labeling a few cases was very complex because of the number of dimensions to be studied.…”

Discovering Effective Connectivity in Neural Circuits: Analysis Based on Machine Learning Methodology

“…There is no universal rule to define the percentages of cases assigned to training and validation. Here we follow previous work and Pareto’s Principle ( Pareto, 1896 ; Dunford et al, 2014 ; Serrano et al, 2017 ) and use 80% of positives for training and 20% for validation. A larger percentage in validation may lose too much information needed for learning, but a smaller percentage may give an improper confidence because of the variability associated to a small amount of cases.…”