Transversal Algorithm Parameter Selection via Stochastic Region Contraction

The primary goal of this proposed project is to develop a general-purpose technique for the problem of algorithm parameter selection (APS) that achieves the best performance of the algorithm and demonstrates its effectiveness within graph analysis, for the problem of community detection, or clustering.  Cell Matrix Corporation (CMC) will develop a means to greatly simplify the tweaking typically required to obtain high fidelity clustering results for a given data set, and for a set of community detection algorithms.  The technique presented here, referred to as T-APS, will derive the necessary values to tune the selected clustering algorithm without knowing the state or operation of the underlying algorithm.  The technique proposed features (a) rapid convergence to a governing parameter set that is beneficial for overall responsiveness, (b)  T-APS has state, and will learn from past successes and failures, and (c) T-APS benefits through Bayesian learning to extrapolate results for unlabeled data sets.  In this Phase I effort, CMC will provide a limited prototype of T-APS that is operational for four different community detection algorithms. The solution generalizes to a wide range of potential applications of network data to analytic objectives and is based on a core set of basic assessments of the datasets that are easily obtained.