Mining Specification Parameters for Multi-Class Classification

We present a method for mining parameters of temporal specifications

for signal classification. Given a parametric formula and a set of labeled traces,

we find one parameter valuation for each class and use it to instantiate the 

specification template. The resulting formula characterizes the signals in a class by dis-

criminating them from signals of other classes. We propose a two-step approach:

first, for each class, we approximate its validity domain, which is the region of

the valuations that render the formula satisfied. Second, we select from each 

validity domain the valuation that maximizes the distance from the validity domain

of other classes. We provide a statistical guarantee that the selected parameter

valuation is at a bounded distance from being optimal. Finally, we validate our

approach on three case studies from different application domains.