Mining Temporal Rules from Program Execution Traces

Software evolution incurs difficulties in program comprehension and software verification, and hence increases the cost of software maintenance. In this study, we propose a novel technique, to mine from program execution traces a sound and complete set of statistically significant temporal rules of arbitrary lengths. The extracted temporal rules reveal invariants that the program observes, and will consequently guide developers to understand the program behaviors, and facilitate all downstream applications like verifications. Different from previous studies that are restricted to mining two-event rules (e.g., 〈lock〉 → 〈unlock〉), our algorithm discovers rules of arbitrary lengths. Furthermore, in order to facilitate downstream applications, we represent the mined rules as temporal logic, so that existing model checkers or other formal analysis toolkits can readily consume our mining results. We performed case studies on JBoss Application Server (JBoss AS) and a buggy Concurrent Versions System (CVS) application, and the result clearly demonstrates the usefulness of our technique in recovering underlying program designs and detecting bugs.