Prediction of Public Satisfaction Levels with Public Services Using Random Forest and Decision Tree

Abstract

Evaluation of public satisfaction in government agencies generally stops at calculating an
index, so there is no initial picture of groups of service users who tend to be less satisfied.
This study builds a classification model of service user satisfaction levels at the Medan Center
for Health Quarantine using Random Forest and Decision Tree, then compares the
performance of the two. The data came from 1,045 respondents to the 2025 Service Quality
Perception Survey and Anti-Corruption Perception Survey with 13 assessment items on a
scale of 1 to 6. Labels were formed from respondent indices into four categories, and the data
were divided into 80% for training and 20% for testing. On the test data, Random Forest
achieved an accuracy of 93.30% and a weighted F1-score of 92.70%, while Decision Tree
achieved 81.82% and 81.79%. Five-fold cross-validation with 3 repetitions showed the same
pattern (F1-score 90.36% versus 82.72%, Wilcoxon test p = 0.0001). However, in the Poor
class, which contained only 6 test respondents, Random Forest had a recall of 0.33 and
Decision Tree 0.83. The infrastructure and online service system items were the most
influential items in the three feature importance analyses. The labels were calculated from the
same items used as features, so the results should be interpreted as the models' ability to learn
the rules for grouping the index.