Future work: Rule Graph Analysis

[SankBad]

This is inspired from the paper Association rule learning. By interpreting the rules as a graph, we can use tools from graph theory to extract information about the configuration space that can improve the quality of the learned model. Later, we can inspect properties of this rule graph to sort reported errors by those most likely to be valid.

Procedure

1. We define the rule graph as a directed hypergraph H = (V, E), with vertices V = {keywords} and labeled, weighted edges E = {(Vs ,Vt ,l,w)}. The set of edges is constructed from the learned rules, using the source and target keyword sets as sources and targets, respectively and the confidence as weights.
2. We then define a measure of degree, D(v), for each vertex v as the sum of in-degree and out-degree. The rule graph is then used to calculate the degree of the keywords, which indicate the estimated importance of each keyword.

We can now use this measure to rank of our genrated errors. The more rules of high confidence are extracted for a keyword by the learner, the higher the D(v) of the corresponding vertex in the rule graph. In our final analysis, we use this classification of rules to order the reported errors by estimated importance.

Acceptance criteria: Calculated order of errors based on Rule graph Analysis