Abstract:
The processing of judicial data by artificial intelligence systems or machine learning methods is
expected to raise the transparency of the operation of justice by increasing transparency,
effectiveness and efficiency, in particular, the predictability of legal application and the
uniformity of case law. There are a number of areas in which AI could have a significant impact
on the legal system. This study attempts to develop predictive model for court decision of Jimma
Zone High Court by using machine learning approach. The proposed predictive model was
trained on a dataset which includes every major criminal cases happened from 2010-2014 E.C.
in Jimma Zone, with detailed information about the criminal, type of the crime and the jury
decision. In the study, machine learning process cycle was employed for developing the
proposed predictive model. A regression predictive model was constructed by using various
machine learning algorithms to predict court decision. Among the various machine learning
algorithms applied for the predictive model are Linear Regression, Huber Regression, Random
Sample Consensus Regression or RANSAC, Theil Sen regression, and Extreme Gradient
Boosting. These algorithms were evaluated on the dataset by using k-fold cross-validation
testing procedure, where k=5. Accordingly, the proposed machine learning models showed
different results on the given dataset by using MAE evaluation metric reveals that Extreme
Gradient Boosting regression algorithm appears to be the best-performing, scores MAE of about
4.080. On the other hand, Theil Sen performed the worst, MAE of about 17.146. Linear, Huber,
and RANSAC have also shown they do not have skill on the given dataset. They score MAE of
about 14.899, 13.195, and 16.020, respectively. Then, Extreme Gradient Boosting was used as a
final model and made predictions on sample rows of data. Finally, the model was deployed using
Gradio GUI library which helps to create user interfaces and share with a link to colleagues or
stakeholders. As a future work, investigation needs to consider tuning hyperparameters, and
calculating optimized values for these parameters has to be considered
