Analysing the Needs of Homeless People Using Feature Selection and Mining Association Rules

TL;DR

The paper addresses the challenge of understanding homeless populations in Europe by integrating a mobile data-collection app with AI tools for feature selection and association-rule mining. It employs a diverse set of methods, including filter/wrapper/embedded feature selectors and two association-rule strategies (Apriori/FP-growth and a PonyGE2 grammar-based approach), evaluated on the INE homeless dataset with $CCR$ as a primary metric. Key findings show that a compact feature subset (e.g., 6 features) can yield high predictive accuracy, and the generated rules reveal meaningful patterns linking health status to socio-demographic factors like chronic illness, disability, nationality, and age. The work provides practical, scalable support for NGOs by enabling efficient data collection and interpretable data-driven insights for targeted interventions.

Abstract

Homelessness is a social and health problem with great repercussions in Europe. Many non-governmental organisations help homeless people by collecting and analysing large amounts of information about them. However, these tasks are not always easy to perform, and hinder other of the organisations duties. The SINTECH project was created to tackle this issue proposing two different tools: a mobile application to quickly and easily collect data; and a software based on artificial intelligence which obtains interesting information from the collected data. The first one has been distributed to some Spanish organisations which are using it to conduct surveys of homeless people. The second tool implements different feature selection and association rules mining methods. These artificial intelligence techniques have allowed us to identify the most relevant features and some interesting association rules from previously collected homeless data.

Figures (4)

• Figure 1: Block diagram of the main SINTECH tools.
• Figure 2: Menus and functionalities of the mobile application.
• Figure 3: Variation of the CCR when decreasing the number of features.
• Figure 4: Most important variables to predict the health state.