to support decision making • Output given to people with limited analyDcs, data and modeling literacy • Hence, it is imperaDve to have interpretable machine learning methods so that predicDons are beLer adopted/trusted by decision makers.
in worst case in comparison to SVM • Renewed interest in field aRer Ruckert & Kramer(2008) – first to introduce rule-‐learning based on objecDve • Works of Gilbert(2008) highlight parallels between group tesDng and sparse signal recovery
supervised classificaDon through rule learning based on Boolean compressed sensing. The major contribuDons of the papers are: -‐ • Show that problem of learning sparse conjuncDve/ disjuncDve clause rules from training samples can be formulated as a group tesDng problem • Reduce the NP hard problem using relaxaDon to resemble the basic pursuit algorithm for sparse signal recovery • Establish condiDons under which the relaxaDon recovers exactly.
pool i • Aij is 0 or 1 based on whether subject j is part of ith group or not • To Find: -‐ wi is the true value of subject i which should be recovered
training samples {Xi, yi} where Xi belongs to X are the features and yi = 0 or 1 • We would like to learn a funcDon to map Xi to yi • Every classifier is made of set of clauses, each clause contains a set of boolean terms
processing technique for effecDvely measuring and reconstrucDng signals • There are clear similariDes, however we restrict boolean algebra instead of real algebra • We apply similar techniques for formulaDon of problem and use LP relaxaDon from compressed sensing
is K-‐ disjunct if the union of any K columns does not contain any other column of A • If there exists a w* with K non-‐zero entries and matrix K-‐disjunct, then the LP recovers it exactly
C k) K subsets, (1-‐e) fracDon of them saDsfy the property that union does not contain any other column • If matrix A is (e,K) disjunct, then LP recover the correct soluDon with probablity 1-‐e
Approach • This is a common technique • First step is to learn an AND rule (here using ideas from boolean sensing) • Once we know one AND rule, remove all training samples which are idenDfied by the rule • Now repeat the learning on remaining rules • This leads to DNF
the approach used in the paper on IRIS dataset. • IRIS dataset is a set of 150 tuples. There are four features : sepal width, sepal length, petal length, petal width. Each tuple is indicaDve of an iris flower. There are three types of flowers :-‐ setosa, versicolor, virginica
Varshney, Exact Rule Learning via Boolean Compressed Sensing. • Malioutov, D. and Malyutov, M. Boolean compressed sensing: LP relaxaDon for group tesDng. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process., pp. 3305–3308, Kyoto, Japan, Mar. 2012.