top of page

Support Vector Machine Learning

Active learning in Relativity works with a classification index. Document reviewers indicate in a single choice field whether or not documents are responsive, and the system learns from these decisions using Support Vector Machine (SVM) learning. SVM predicts whether or not documents which have not been reviewed will be relevant based on the coders' decisions.

The documents are compared to coded documents using a spatial model with multiple dimensions. So a document to be evaluated will be measured on its distance from a coded document. A hyperplane divides responsive from non-responsive documents.

Documents are ranked on a scale from 100 to 0. A document that is clearly relevant will have a score closer to 100; a clearly non-responsive document's score will tend toward zero.

Support vector machine learning is not a concept unique to Relativity but is a model used in the general field of machine learning.

The hyperplane is a line drawn through the responsive and non-responsive sets. It's a linear classifier. The appropriate hyperplane will maximize the distance from the two nearest points from both classes. So in this example, H3 is the most appropriate hyperplane.

SVM algorithms can also be used to classify images and handwriting. The support vectors are those which are closest to the margin of the hyperplane. So a 'support vector' will be a document that is most difficult to code responsive or non-responsive. The support vector machine model is different from linear regression,which uses all points to determine the optimal hyperplane. SVM only uses the difficult points or documents to establish the boundary. Changing a single support vector will cause the decision boundary or hyperplane to move.

bottom of page