The Cutting Plane Procedure


  1. The Cutting Plane Procedure is a non-parametric method that finds the cutting plane and its normal vector given the Xi's and the corresponding vector of choices. The method

    1. Maximizes Correct Classification of the Choices

    2. Is iterative and Always Converges

    3. Is Robust and Almost Always Finds the True Cutting Plane and Normal Vector

  2. The Cutting Plane Procedure Consists of the Following Steps:

    1. First, given a reasonable guess for the normal vector, the cutting plane is moved through the space along the normal vector and placed at the point that maximizes the correct classification of the choices. Conditioned on the normal vector, the conditional global maximum in classification is always found.

      Intuitively, given b1 and b2 such that b12 + b22 = 1, this search stage finds the optimal b0.

    2. Second, the current cutting plane is used to construct a set of points that in turn is used to change the orientation of the Cutting Plane (and Normal Vector) in the Space. This is accomplished as follows:

      1. The Correctly Classified Points are projected onto the surface of the current cutting plane. The Incorrectly Classified points are left at their current positions in the space. Let Y represent the matrix of points constructed from the (in)correctly classified Xi's.

      2. Place Y at its centroid by subtracting off the column means from the corresponding entries in each column. Denote this matrix by Y*. Note that the columns of Y* sum to zero.

      3. Find the least squares line through Y*. Note that this is equivalent to rotating the cutting line through the space towards the errors.

        The least squares line is given by the Eckart-Young Theorem (1936). Namely, let the singular value decomposition of Y* be:

        Y* = ULV'

        where, in terms of the example discussed above, U is a 12 by 2 matrix, L is a 2 by 2 diagonal matrix of singular values, and V is an 2 by 2 matrix such that

        U'U = I2 and V'V = VV' = I2

        with the singular values arranged in decreasing sequence on the diagonal of L.

        To find the least squares line simply set the smallest singular value in L equal to zero and then remultiply; namely

        B = ULsV'

        where Ls is equivalent to L except that the smallest singular value is set equal to zero. B is the least squares line through the points in Y* and the normal vector to this line is the corresponding singular vector in V -- in this case the second column of V. This is the new normal vector.

        Intuitively, in this search stage b0 is set to zero and new estimates are found for b1 and b2 that maximize correct classification.

    3. Third, Go to (a)