Interpolating isotonic regression for binary classification of spam, explicit material or malware using interpolation based on a Delaunay triangulation
View Patent ↗Described is a machine learning system for binary classifications. The system improves interpolation schemes used in isotonic regressions by providing a continuous function that also is monotonic. The system outputs a probability estimating function on a signal space that is both monotonic and varies continuously with the input signals. More specifically, described is an interpolation function that is continuous and piecewise linear on Delaunay simplices. Accordingly, the resulting probability estimation function may more accurately match actual probabilities especially when training data is sparse.
1. A computer-implemented method for interpolating isotonic regression in a binary classifier system, comprising:
receiving a set of training examples for content items, each training example including a positive or negative label for a classification and a respective value for each of D attributes of the content items, wherein the values are provided as D-dimensional coordinates;
determining an interpolation function based on a Delaunay triangulation of coordinates in the received set of training examples, wherein the determined function satisfies monotonicity constraints and includes maximizing a likelihood of an observation subject to the monotonicity constraints;
receiving a first observation for a first content item, the first observation including D-dimensional coordinates; and
determining a probability that the first observation is positive for the classification based on the determined interpolation function, wherein the classification is one of spam, explicit material, and malware.
2. The method of claim 1 , wherein the attributes include at least one of a text attribute and a non-text attribute.
3. The method of claim 1 , wherein the coordinates are values between 0 and 1.
4. The method of claim 1 , wherein the monotonicity constraints are based on a matrix of a coordinate representation of a linear map from barycentric coordinates to projective feature space coordinates.
5. The method of claim 1 , wherein the maximized likelihood of the observation subject to the monotonicity constraints is based on the interpolation function linearly interpolating over simplices.
6. A system for interpolating isotonic regression in a binary classifier system, comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
receiving a set of training examples for content items, each training example including a positive or negative label for a classification and a respective value for each of D attributes of the content items, wherein the values are provided as D-dimensional coordinates;
determining an interpolation function based on a Delaunay triangulation of coordinates in the received set of training examples, wherein the determined function satisfies monotonicity constraints and includes maximizing a likelihood of an observation subject to the monotonicity constraints;
receiving the observation for a content item, the observation including D-dimensional coordinates; and
determining a probability that the observation is positive for the classification based on the determined interpolation function, wherein the classification is one of spam, explicit material, and malware.
7. The system of claim 6 , wherein the attributes include at least one of a text attribute and a non-text attribute.
8. The system of claim 6 , wherein the coordinates are values between 0 and 1.
9. The system of claim 6 , wherein the monotonicity constraints are based on a matrix of a coordinate representation of a linear map from barycentric coordinates to projective feature space coordinates.
10. The system of claim 6 , wherein the maximized likelihood of the observation subject to the monotonicity constraints is based on the interpolation function linearly interpolating over simplices.