|M.Sc Student||Suchinsky David|
|Subject||Investigating Accident Severity and Type on Inter-Urban|
Roads in Israel Using Data Mininig
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Shlomo Bekhor|
|Full Thesis text|
The complex attributes of vehicle crashes make it difficult to predict, prevent and mitigate their severity. Many studies have focused on identifying the factors that impact crash severity, with scopes focusing on certain types of crashes, certain types of drivers, or certain types of roads. The results’ incompatibility makes it difficult to evaluate which infrastructure, environmental or driver characteristics affect a crash’s severity, and to what degree. In addition, it is difficult to transform the large databases of crash records, where each crash has many independent variables, into useful models by using traditional statistical techniques. Furthermore, there is a desire to include speed-related independent variables in models of crash severity, which historically have not been included.
This study uses data mining on Israel’s historic crash records combined with the free-flow speed on road segments to create crash severity and crash type prediction models for Israel’s inter-urban roads. Multi-layer perception artificial neural networks were chosen to create the prediction models because the technique is powerful, efficient, flexible and easy to use. The resulting networks are fed a choice of independent variables that define the situation in which a hypothetical crash occurred, and the network’s outputs are the probabilities of each severity level or crash type. A knowledge discovery in databases methodology was followed to identify the problem, obtain and preprocess the data, perform the data mining, evaluate and interpret the results and finally to use the discovered knowledge.
The results of interpreting the two models reveals that variables do not always have the same effect depending on the values of the other independent variables, and therefore it’s hard to generalize the impact any single variable will have. The power of neural networking is not in the ability to interpret the network and understand the complex relationships, but in the ability to model desired scenarios and quickly obtain results. As long as there is enough data, and variation among that data, to create an accurate neural network, it is a highly successful modeling technique.