We analyze the Computing Education Research (CER) literature to discover what theories, conceptual models and frameworks recent CER builds on. This gives rise to a broad understanding of the theoretical basis of CER that is useful for researchers working in that area, and has the potential to help CER develop its own identity as an independent field of study.Our analysis takes in seven years of publications (2005)(2006)(2007)(2008)(2009)(2010)(2011) 308 papers) in three venues that publish long research papers in computing education: the journals ACM Transactions of Computing Education (TOCE) and Computer Science Education (CSEd), and the conference International Computing Education Research Workshop (ICER). We looked at the theoretical background works that are used or extended in the papers, not just referred to when describing related work. These background works include theories, conceptual models and frameworks. For each background work we tried to identify the discipline from which it originates, to gain an understanding of how CER relates to its neighboring fields. We also identified theoretical works originating within CER itself, showing that the field is building on its own theoretical works.Our main findings are that there is a great richness of work on which recent CER papers build; there are no prevailing theoretical or technical works that are broadly applied across CER; about half the analyzed papers build on no previous theoretical work, but a considerable share of these are building their own theoretical constructions. We discuss the significance of these findings for the whole field and conclude with some recommendations.

Program comprehension (PC) is a research field that has been extensively studied from different points of view, including human program understanding and mental models, automated program understanding, etc. In this paper, we discuss algorithm recognition (AR) as a subfield of PC and explain their relationship. We present a method for automatic AR from Java source code. The method is based on static analysis of program code including various statistics of language constructs, software metrics, as well as analysis of roles of variables in the target program. In the first phase of the method, a number of different implementations of the supported algorithms are analyzed and stored in the knowledge base of the system as learning data, and in the second phase, previously unseen algorithms are recognized using this information. We have developed a prototype and successfully applied the method for recognition of sorting algorithms. This process is explained in the paper along with the experiment we have conducted to evaluate the performance of the method. Although the method, at its current state, is still sensitive to changes made to target algorithms, the encouraging results of the experiment demonstrate that it can be further developed to be used as a PC method in various applications, as an example, in automatic assessment tools to check the algorithms used by students, the functionality that is currently missing from these tools.

No abstract

We discuss algorithm recognition (AR) and present a method for recognizing algorithms automatically from Java source code. The method consists of two phases. In the first phase, the recognizable algorithms are converted into the vectors of characteristics, which are computed based on static analysis of program code, including various statistics of language constructs and analysis of Roles of Variables in the target program. In the second phase, the algorithms are classified based on these vectors using the C4.5 decision tree classifier. We demonstrate the performance of the method by applying it to sorting algorithms. Using leave-one-out cross-validation technique, we have conducted an experimental evaluation of the classification performance showing that the average classification accuracy is 98.1% (the data set consisted of five different types of sorting algorithms). The results show the applicability and usefulness of roles of variables in AR, and illustrate that the C4.5 algorithm is a suitable decision tree classifier for our purpose. The limitations of the method are also discussed.

Computing educators often rely on black-box analysis to assess students' work automatically and give feedback. This approach does not allow analyzing the quality of programs and checking if they implement the required algorithm. We introduce an instrument for recognizing and classifying algorithms (Aari ) in terms of white-box testing to identify authentic students' sorting algorithm implementations in a data structures and algorithms course. Aari uses machine learning techniques to classify new instances. The students were asked to submit a program to sort an array of integers in two rounds: at the beginning of the course before sorting algorithms were introduced, and after taking a lecture on sorting algorithms. We evaluated the performance of Aari with the implementations of each round separately. The results show that the sorting algorithms, which Aari has been trained to recognize, are recognized with an average accuracy of about 90%. When considering all the submitted sorting algorithm implementations (including the variations of the standard algorithms), Aari achieved an overall accuracy of 71% and 81% for the first and second round, respectively.In addition, we analyzed the students' implementations manually to gain a better understanding of the reasons of failure in the recognition process. This analysis revealed that students have many misconceptions related to sorting algorithms, which results in problematic implementations that are more inefficient compared with those of standard algorithms. We discuss these variations along with the application of the tool in an educational context, its limitations and some directions for future work.