Named after the Norse God of Foresight, our Crime Prediction model takes in longitude and latitude time series data and gives a propensity at a certain time for these crimes.
Available now (Q2 2019)
The prediction model is accessed through a RESTful JSON API. Just enter the longitude and latitude of the crime, the crime type (1 to 11) and the inference will come back with a probability of getting a custodial sentence for that crime.
Predictive models that use lots of past historic data, to train complex algorithms such as neural networks or deep learning algorithms, can be what is called "over fitted". This essentially means that when unseen data is presented to the predictive model that is outside of the cohort or population of data that the model is trained on, the accuracy of the model rapidly diminishes. A neural network that could offer in excess of 0.85 Gini coefficient on one population could offer a 0.25 or lower Gini with a different population.
By using an unsupervised machine learning algorithm, that is still trained on looking for the anomalies associated with a certain population, but not subject to the bias, therefore more ethical, and overfitting, a more realistic 0.65 or above Gini can be achieved.
Our Crime Prediction platform streams data from the publicly available police data base and uses time series data mapped to longitude and latitude.
Also because all the algorithms have been written from scratch using GLM we can expose the weightings for each prediction, giving us the ability to show how the prediction is created, what we call the magic donut - or explainable AI - XAI.