Using traits of web macro scripts to predict reuse

Digging for Diamonds: Identifying Valuable Web Automation Programs in Repositories

“…Prior work showed that our model was empirically just as accurate as more complex models on CoScripter data [17]. Moreover, the accuracy of our model was comparable to that of other research that applies machine learning to software engineering problems.…”

“…Motivated by the limitations in existing repositories, our automated machine learning approach categorizes end-user web automation programs based on likelihood of high value, rather than content [17]. Evaluating the effectiveness of this approach requires several steps: characterizing the traits of programs to be used for predictions, identifying target thresholds for several indications of value, and finally testing the model.…”

“…Finally, the algorithm collects and returns the generated constraints. (Our prototype of the training algorithm previously included a step to discard traits where this difference in proportions was below a certain level, but experiments on the prototype showed that this extra step did not affect model accuracy much on average [17], so we now omit this step. )…”

“…A typical repository might contain thousands of these programs, but as few as 10% of these might be of value to the community at large [17]. The other 90% include programs that are unreliable, as well as those that are so highly personalized to particular users and are so hard to read or modify that it is inconvenient for others to customize the programs to meet their own needs.…”

“…In prior work, we prototyped our model for use on one particular repository of web automation programs, with four specific measures of the programs' value [17]. We evaluated the prototype model for that repository and found its accuracy to be comparable to that of more complex machine learning models, including those used in other areas of software engineering.…”