Homelessness is a rising issue in Australia, which has recently expanded beyond the stereotypical demographics, to full time workers who’ve faced speed-bumps in life, and chose the streets. We aim to address this new form of homelessness, by enabling early intervention that prevents its snowballing effect, to prematurely stop homelessness.

We used a FNN, the simplest form of a neural-net, as a proof of concept for our idea. Here, we conducted supervised learning optimised by gradient descent, which involves iteratively tuning the parameters of the neural-net, by having the model attempt to match an unseen data (X) to its label (Y), compute the difference between the model’s response and the expected response by a loss function, and finally adjust the model’s parameters in a direction opposite to the derivative of its loss via back-propagation. Data used to train this model was synthetically generated by a separate hand-adjusted weighted logistic model.

The model plateaued at ~61% accuracy on synthetic data, likely due to rounding continuous homelessness risk values into binary (1/0) in the synthetic data generation, which caused information loss. However, web interface interactions reveals correct mapping of attributes to their respective effects (e.g., DV+ → Chance+), indicating success.