Nevertheless the overall prediction ac curac y on the

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: rediction ac curac y. Because AGV is c onstr ucted to be an interpolation bet ween LPA and a ring lattice, the AGV, LPA, and SMW mechan isms are equivalent in specific parameter regimes and c orrespondingly show a nonnegligible overlap. Nevertheless, the overall prediction ac curac y on the test sets still (Example subgraphs) Notes on procedure • Similar to techniques in social sciences (p∗, exponential random graph models). • Network “motifs”, Milo et al Science, 2002. But motifs only up to n = 3 or n = 4 nodes. • Note the term “clustering” here refers to machine learning technique to categorize data, not “clustering coefficient” (transitivity). Build classifier from the training data (Learning Algorithm) • Alternating Decision Tree (ADT), (Freund and Schapire, 1997). Which subgraphs best distinguish the models? Validating classifier • Slight overlap in models which are variations on one-another. Validating classifier • (a) DMC and RDG produce similar statistical distributions. • (b) Classifier can discriminate between the two models. After classifier built, use it to characterize individual network realizations (Walk the Drosophila data through the ADT) • A given network’s subgraph counts determine paths in the ADT (decision nodes are rectangles) • The ADT outputs a real-valued prediction score, which is the sum of all weights over all paths. • The final weight for a model is related to probability that particular network realization was generated by that model. • Model with the highest weight wins (best describes that particular network realization). • DMC wins for Giot Drosophila data! Comparison by subgraph counts • Green is best (same median occurrence as in real Drosophila data). • 0 means the subgraph is in data, but not in model. Introducing noise • Classifier robust. • Also robust to p = 0.5 and different subgraph counts, n = 7, 8. Comments • Model selection not validation. (Relative judgement) (i.e., which of these 7 models fits...
View Full Document

Ask a homework question - tutors are online