\section{Method}

This section explains the three main components of our approach: automatically extracting patterns from student submissions, learning classification rules for correct and incorrect programs, and using those rules to generate hints.

\subsection{Extracting patterns}
\label{sec:extracting-patterns}

We construct patterns by connecting pairs of leaf nodes in a program’s AST. For this paper we always select a pair of nodes from the same clause: either two nodes referring to the same variable (like the examples in Fig.~\ref{fig:sister}), or a value (such as the empty list \code{[]} or \code{0}) and another variable or value in the same \textsf{compound} or \textsf{binop} (like the blue dotted pattern in Fig.~\ref{fig:sum}). For example, in the clause\footnote{The second occurrence of variables \code{A}, \code{B} and \code{C} is marked with ’ for disambiguation.}

\begin{Verbatim}
a(A,\textsf{ }B):-
  b(A\textsf{'},\textsf{ }C),
  B\textsf{'} is C\textsf{'}\textsf{ }+\textsf{ }18.
\end{Verbatim}

\noindent
we select the following pairs of nodes: \{\code{A},\,\code{A\textsf{'}}\}, \{\code{B},\,\code{B\textsf{'}}\}, \{\code{C},\,\code{C\textsf{'}}\}, \{\code{B\textsf{'}},\,\code{18}\} and \{\code{C\textsf{'}},\,\code{18}\}.

For each selected pair of leaf nodes $(a,b)$ we build a pattern by walking the AST in depth-first order, and recording nodes that lie on the paths to $a$ and $b$. We omit \textsf{and} nodes, as explained in the previous section. We also include certain nodes that do not lie on a path to any selected leaf. Specifically, we include the functor or operator of all \textsf{compound}, \textsf{binop} or \textsf{unop} nodes containing $a$ or $b$.

Patterns constructed in this way form the set of features for rule learning. To keep this set at a reasonable size, we only use patterns that have appeared in programs submitted at least five times.

\subsection{Learning rules}

%\begin{figure}[t]
%\noindent \textsc{Procedure $learn\_rules(D, sig, acc)$}
%\\
%\\
%\noindent Given: data \emph{D}, threshold of significance \emph{sig}, threshold of accuracy \emph{acc}, and a 
%procedure for learning rules $learn\_target\_rules(target, D, D_{acc}, sig, acc)$ that learns a set of rules for 
%class \emph{target} from data \emph{D}, where: a) each attribute-value pair in the condition part 
%needs to be significant using the likelihood-ratio test with significance level $p<sig$,
%b) classification accuracy of rules on data $D_{acc}$ should be at least $acc$, and c) only positive values of 
%attributes are mentioned in conditions. 
%\begin{enumerate}
%  \item Let \emph{I-rules} = $learn\_target\_rules(incorrect, D, D, sig, acc)$
%  \item Let $D_c$ be data $D$ without programs covered by \emph{I-rules}
%  \item Let \emph{C-rules} = $learn\_target\_rules(correct, D, D_c, sig, acc)$
%  \item Return \emph{I-rules} and \emph{C-rules}
% \end{enumerate}
% \caption{An outline of the algorithm for learning rules. The method \emph{learn\_rules}, which induces rules  for a specific class, is a variant of the CN2 algorithm~\cite{cn21991} implemented within the Orange data-mining suite~\cite{demsar2013orange}. In all our experiments, \emph{sig} was set to 0.05 and \emph{acc} was set to 0.9.}
% \label{figure:algorithm}
%\end{figure}

We represent submitted programs in the feature space of AST patterns described above. One pattern corresponds to one boolean feature with value set to $true$ when pattern is present and $false$ when pattern is absent. We classify programs as correct or incorrect based on predefined unit tests for each problem, and use these labels for machine learning.

Since programs can be validated with appropriate unit tests, our goal is not classifying new submissions, but to discover patterns related to program correctness. This machine-learning approach is called \emph{descriptive induction} -- the automatic discovery of patterns describing regularities in data. We use rule learning for this task because conditions of rules can be easily translated to hints.

Before explaining the algorithm, let us discuss the reasons why a program can be incorrect. Our teaching experience indicates that bugs in student programs can often be described by 1) some incorrect or \emph{buggy} pattern, which needs to be removed, or 2) some missing relation (pattern) that should be included before the program can be correct. We shall now explain how both types of errors can be identified with rules.

To discover buggy patterns, the algorithm first learns rules that describe incorrect programs (I-rules). We use a variant of the CN2 algorithm~\cite{cn21991} implemented within the Orange data-mining toolbox~\cite{demsar2013orange}. Since we use rules to generate hints, and since hints should not be presented to students unless they are likely to be correct, we impose additional constraints on the rule learner:
\begin{enumerate}
 \item The classification accuracy of each learned rule must exceed a certain threshold (in our experiments we used 90\% as threshold).
 \item Each conjunct in a condition must be significant with respect to the likelihood-ratio test (in our experiments significance threshold was set to $p=0.05$).
 \item A conjunct can only specify the presence of a pattern: we allow feature-value pairs with only $true$ as value. 
\end{enumerate}
\noindent The former two constraints are needed to induce good rules with significant patterns, while the latter constraint assures that rules mention only presence (and not absence) of patterns as reasons for a program to be incorrect. This is important, since these conditions ought to contain frequent patterns symptomatic of incorrect programs to be used by the hint generation procedure. 

With respect to the second type of error, we could try the same approach and learn rules using the above algorithms for the class of correct programs (C-rules). Having accurate rules for correct programs, the conditional part of these rules would define sufficient combinations of patterns that render a program correct. However, it turns out that it is difficult to learn accurate rules for correct programs, since these rules would have to contain all patterns that need to be in the program and would have to specify the absence of all incorrect patterns, yet a conjunct can only specify the presence of a pattern. If specifying absence of patterns was allowed in rules' condition, the learning problem would still be difficult, since there can be many incorrect patterns, resulting in many C-rules.

A possible way to solve this problem is to remove programs that are covered by rules for incorrect class. This way all known buggy patterns are removed from the data, therefore will not be included in C-rules. However, removing incorrect patterns also removes the need for relevant patterns. For example, if all incorrect programs were removed, the a single C-rule “$\mathsf{true} ⇒ \mathsf{correct}$” would suffice, which cannot be used to generate hints. We achieved the best results by learning from both data sets: we use the original data set (with all programs) to learn rules, while we use the data set without buggy patterns to test whether a rule achieves the required classification accuracy (90\%).

Even though our main interest is discovery of patterns, we can still use induced rules to classify new programs, for example to evaluate the quality of rules. The classification procedure has three steps: 1) if an I-rule covers the program, classify it as incorrect; 2) else if a C-rule covers the program, classify it as correct; 3) otherwise, if no rule covers the program, classify it as incorrect.

\subsection{Generating hints}

Once we have induced the rules for a given problem, we can use them to provide hints based on buggy or missing patterns. To generate a hint for an incorrect program, each rule is considered in turn. We consider two types of feedback: \emph{buggy} hints based on I-rules, and \emph{intent} hints based on C-rules.

First, all I-rules are checked to find any known incorrect patterns in the program. To find the most likely incorrect patterns, the rules are considered in the order of decreasing quality. If all patterns in the rule “$p_1 ∧ ⋯ ∧ p_k ⇒ \mathsf{incorrect}$” match, we highlight the relevant leaf nodes. As an aside, we found that most I-rules are based on a single pattern. For the incorrect \code{sum} program from the previous section, our method produces the following highlight:

\begin{Verbatim}
sum([],0).          % \textit{base case:} the empty list sums to zero
sum([H|T],\red{\underline{Sum}}):-    % \textit{recursive case:}
  sum(T,\red{\underline{Sum}}),       %  sum the tail and
  Sum is Sum + H.   %  add first element (\textit{bug:} reused variable)
\end{Verbatim}

If no I-rule matches the program, we use C-rules to determine the student’s intent. C-rules group patterns that together indicate a high likelihood that the program is correct. Each C-rule thus defines a particular “solution strategy” defined in terms of AST patterns. We reason that a hint alerting the student to a missing pattern could help them complete the program without revealing the whole solution.

When generating a hint from C-rules, we consider all \emph{partially matching} rules “$p_1 ∧ ⋯ ∧ p_k ⇒ \mathsf{correct}$”, where the student’s program matches some (but not all) patterns $p_i$. For each such rule we store the number of matching patterns, and the set of missing patterns. We then return the most common missing pattern among the rules with most matching patterns.

For example, if we find the following missing pattern for an incorrect program implementing the \code{sister} predicate:

\begin{Verbatim}[fontfamily=sf]
(clause (head (compound (functor "\code{sister}") (args var))) (binop var "\code{\textbackslash{}=}"))\textrm{,}
\end{Verbatim}

\noindent
we could display a message to the student saying “you are missing a comparison between \code{X} and some other value, with the form \code{X} \code{\textbackslash{}=} \code{?}”.


%%% Local Variables:
%%% mode: latex
%%% TeX-master: "aied2017"
%%% End: