Variable neighbourhood search for parameter tuning in support vector machines

Variable neighbourhood search for parameter tuning in support vector machines
Document type
Carrizosa, Emilio; Martín-Barragán, Belén; Romero Morales, Dolores
Said Business School, Oxford University
Date of publication
11 September 2012
Saïd Business School Working Paper Series 2012-20
Trends: economic, social and technology trends affecting business
Business and management
Material type

This item is only available to registered users

Register now or Log in

Your use of this content is subject to the terms and conditions of this portal

As in most Data Mining procedures, how to tune the parameters of a Support Vector Machine (SVM) is a critical, though not suffciently explored, issue. The default approach is a grid search in the parameter space, which becomes prohibitively time-consuming even when just a few parameters are to be tuned. For this reason, for models involving a higher number of parameters, different metaheuristics have been recently proposed as an alternative. In this paper we customize a continuous Variable Neighborhood Search to tune the parameters of the SVM. Our framework is general enough to allow one to address, with the very same method, several popular SVM parameter models encountered in the literature. The structure of the optimization problem at hand is successfully exploited. This is done by expressing the problem of parameter tuning as a collection of nested problems which can be solved sequentially. As algorithmic requirements we only need a routine which, for given parameters, nds the SVM classi er. Hence, as soon as an SVM library or any routine for solving linearly constrained convex quadratic optimization problems is available, our approach is applicable. The experimental results show the usefulness of our tuning method for different SVM parameter models analyzed in the literature: we can address tuning problems with dimensions for which grid search is infeasible, and, at the same time, our rather general approach yields comparable results, in terms of classi cation accuracy, against ad-hoc benchmark tuning methods designed for specific SVM parameter models.

Related to Trends: economic, social and technology trends affecting business

Follow the money: exploring the link between UK growth and workers' pay packets

Downloadable briefing note looking at the relationship between productivity and pay growth

From precarity to empowerment: women and the future of work

Downloadable paper on ensuring a fair future of work for women

Economic insecurity: the case for a 21st century safety net

Downloadable paper on how economic insecurity is experienced by workers

Agents of change: making innovation agencies as innovative as those they support

Downloadable paper on how government innovation agencies can become better

More items related to this subject

Related to Said Business School, Oxford University

Brand tribalism: a netnographic exploration of virtual communities

The article comes to the conclusion that brands can benefit from having a presence in Facebook and other Social Network Sites (SNS) if the objectives and the nature of the brand suit the environment.

The role of small and medium enterprise retailing in Britain

This article reviews the issues surrounding the operation and the contribution of the SME retail sector to the local community and the whole UK economy. Extant literature suggests that one of the greatest

Backs to the wall: how public relations carries discomfort for organisations

Public relations plays a sensitive role in dealing with the discomfort felt within organisations about the ethical implications of their operations. Organisational discomfort seems to be on the increase,

A monetary system of equations with inflationary expectations for the USA

The paper is an empirical investigation that places Livingston’s expectations of the Consumer Price Index (CPI) with the rate of inflation centre stage in a monetary system of equations with real money

More items related to this publisher