Skripsi
OPTIMALISASI IDENTIFIKASI POLA KECELAKAAN LALU LINTAS MELALUI UNSUPERVISED FEATURE SELECTION DAN SHAP
The analysis of traffic accident patterns using unsupervised clustering algorithms on high-dimensional data is often hindered by the presence of uninformative variables (noise) and the limitations of conventional statistical methods in interpreting the meaning of the formed clusters. This research aims to optimize the identification and interpretation of accident patterns through the integration of the Unsupervised Feature Selection (UFS) and SHAP methods. The Sequential Filtering on Variance and Correlation approach was applied to reduce data dimensionality, followed by clustering with K-Means, and variable interpretation through a Random Forest surrogate model and SHAP values. The UFS process successfully filtered 98 encoded variables into the 15 most relevant features, utilizing the Société de l'assurance automobile du Québec (2016-2022) dataset comprising 886,795 incident records. This reduction was proven to optimize the clustering structure by minimizing distance distortion, validated by a consistent increase in internal evaluation metric scores compared to the use of raw data. The K-Means algorithm identified two highly contrasting risk patterns: low-speed urban intersection incidents involving multiple vehicles and high-speed single-vehicle incidents on rural roads. The application of SHAP analysis successfully dismantled the "black box" nature of the clustering algorithm by providing transparent explanations regarding the direction of variable influence, the magnitude of impact, and identifying the transition point on the speed limit that determines pattern classification. Overall, this integrated framework proved effective in addressing the problems of high dimensionality and shallow interpretation, thereby generating accident risk profile insights that are objective, structurally accurate, and comprehensively understandable.
No other version available