The demo illustrates the MIQUBO method by finding an optimal feature set for
predicting survival of Titanic passengers. It uses records provided in file
formatted_titanic.csv
, which is a feature-engineered version of a public
database of passenger information recorded by the ship's crew (in addition to a
column showing survival for each passenger, it contains information on gender,
title, class, port of embarkation, etc). Its output is a ranking of subsets of
features that have high MI with the variable of interest (survival) and low
redundancy.
Note: This example solves a binary quadratic model (BQM) directly on a D-Wave quantum computer's quantum processing unit (QPU). The Feature Selection for CQM example solves this same problem using a constrained quadratic model on a Leap™ quantum-classical hybrid solver.
python titanic.py
Statistical and machine-learning models use a set of input variables (features) to predict output variables of interest. Feature selection, which can be part of the model design process, simplifies the model and reduces dimensionality by selecting, from a given set of potential features, a subset of highly informative ones. One statistical criterion that can guide this selection is mutual information (MI).
Ideally, to select the k
most relevant features, you might maximize I(Xs;Y)
,
the MI between a set of k
features, Xs
, and the variable of interest, Y
.
This is a hard calculation because the number of states is exponential with k
.
The Mutual Information QUBO (MIQUBO) method of feature selection
formulates a quadratic unconstrained binary optimization (QUBO) based on an
approximation for I(Xs; Y)
, which is submitted to the D-Wave quantum computer
for solution.
There are different methods of approximating the hard calculation of optimally
selecting k
of n
features to maximize MI. The approach followed here
assumes conditional independence of features and limits conditional MI
calculations to permutations of three features. The optimal set of features,
S
, is then approximated by:
The left-hand component, I(Xi;Y)
, represents MI between the variable of
interest and a particular feature; maximizing selects features that best predict
the variable of interest. The right-hand component, I(Xj;Y |Xi)
, represents
conditional MI between the variable of interest and a feature given the prior
selection of another feature; maximizing selects features that complement
information about the variable of interest rather than provide redundant
information.
This approximation is still a hard calculation. MIQUBO is a method for formulating it for solution on the D-Wave quantum computer based on the 2014 paper, Effective Global Approaches for Mutual Information Based Feature Selection, by Nguyen, Chan, Romano, and Bailey published in the Proceedings of the 20th ACM SIGKDD international conference on knowledge discovery and data mining.
X. V. Nguyen, J. Chan, S. Romano, and J. Bailey, "Effective global approaches for mutual information based feature selection", https://dl.acm.org/citation.cfm?id=2623611
Released under the Apache License 2.0. See LICENSE file.