A study of learning likely data structure properties using machine learning models

Data structure properties are important for many testing and analysis tasks. For example, model checkers use these properties to find program faults. These properties are often written manually which can be error prone and lead to false alarms. This paper presents the results of controlled experiments performed using existing machine learning (ML) models on various data structures. These data structures are dynamic and reside on the program heap. We use ten data structure subjects and ten ML models to evaluate the learnability of data structure properties. The study reveals five key findings. One, most of the ML models perform well in learning data structure properties, but some of the ML models such as quadratic discriminant analysis and Gaussian naive Bayes are not suitable for learning data structure properties. Two, most of the ML models have high performance even when trained on just 1% of data samples. Three, certain data structure properties such as binary heap and red black tree are more learnable than others. Four, there are no significant differences between the learnability of varied-size (i.e., up to a certain size) and fixed-size data structures. Five, there can be significant differences in performance based on the encoding used. These findings show that using machine learning models to learn data structure properties is very promising. We believe that these properties, once learned, can be used to provide a run-time check to see whether a program state at a particular point satisfies the learned property. Learned properties can also be employed in the future to automate static and dynamic analysis, which would enhance software testing and verification techniques.