Performance Comparison of Data Mining Techniques for Rain Prediction Models in Indonesia

Rain prediction is a crucial topic that continues to gain interest across the globe. Rain has a massive impact on various aspects of human life such as in agriculture, health, transportation, etc, and also some natural disasters. Various impacts of rain on human life prompts us to build a model to understand and predict rain to provide early warning for various use cases in various fields. Previous research on rain modeling using Data Mining (DM) techniques had suffered from low accuracy caused by the limited availability of the training data and their meteorological attributes. This research aims to address those issues by building the rain model using a richer and more abundant rain data in Indonesia. Four DM techniques are used and compared in this research i.e. the C4.5/J48, Random Forest (RF), Naïve Bayes (NB), and Multilayer Perceptron (MLP). The experimental results showed that the MLP and J48 algorithm can provide the best accuracy (up to 78,4%), which is better than previous research. Other key findings in this research include: (a) the selection of DM techniques has little effect on the model accuracy; (b) a larger training dataset generally improves model accuracy and a larger test dataset is necessary to get a representative realworld test accuracy, and (c) the two most influential attributes in rain modeling are the relative humidity and the minimum temperature, and we suggest to include cloud condensation nuclei in the next research to complete the model.