Artificial intelligence system for training a classifier
An artificial intelligence system for training a classifier has a database of training data and a modeling system for building a classification model based on the training data. The database has a binary classification for each entity and binary tokens indicating whether or not one or more indicators about the entity are true. The classification model is based on a tempered indication of the tokens. The tempered indication is a ratio of a weighted sum of the tokens for each entity divided by a tempering factor for each of the entities. The tempering factor is a function of the unweighted sum of the tokens for each entity. Thus, the tempering factor will reduce the tempered indication when large numbers of low weight tokens are present so that the model does not over predict the probability of an entity being in the classification.
1 . An artificial intelligence system for training an automated classification system, the artificial intelligence system comprising:
a computer implemented training database comprising a plurality of training entity records, each of the records comprising:
an adjudicated binary class of a training entity j, wherein the binary class is determined by an adjudicator using data independent of tokens used to train the classifier having a binary value of 1 when the class is true and a binary value of 0 when the class is false; and
a plurality of class indicator tokens i for the training entity j, each of the tokens being a binary indication of the presence or absence of an indicator i in a publication by the training entity j, the binary indication having a binary value of 1 when the indication is true and a binary value of 0 when the indication is false; and
a computer implemented modeling engine comprising:
an input device, an output device, a permanent memory, a microprocessor; and
computer readable instructions stored on the permanent memory, which when executed by the microprocessor cause the microprocessor to:
read in, by the input device, the plurality of training entity records from the training database;
process, in parallel using multiple processing threads of the microprocessor to reduce computational processing time compared to sequential, wherein the multiple processing threads execute with interdependent calculations:
a first thread calculating weighted sums of the tokens i;
a second thread calculating unweighted sums of the tokens i; and
a third thread calculating tempering factors as functions of the unweighted sums calculated by the second thread;
calculate, by the microprocessor, for each training entity j, a tempered indication of the class, wherein the tempered indication is automatically set to zero when all tokens for the training entity have binary values of zero, the tempered indication being based on:
the ratio of a weighted sum of the tokens i associated with the training entity j and a tempering factor, the tempering factor being a function of an unweighted sum of the tokens i associated with the training entity j;
one or more indicator weights i associated with each of the tokens i in the weighted sum; and
one or more tempering parameters associated with either the weighted sum or the tempering factor;
generate a normalized asymptotic transformation of the tempered indication, wherein the transformation has a long tail relative to a logistic transformation and has lower probability values than the logistic transformation at higher values of the tempered indication;
calculate, by the microprocessor, for each training entity j, a probability of the class having a binary value of 1 using the normalized asymptotic transformation;
calculate, by the microprocessor, for each training entity j, a residual, the residual being the difference between the binary value of the class and the probability of the class having a binary value of 1;
calculate, by the microprocessor, an error function of the residuals for all of the entities;
calculate, by the microprocessor, iteratively adjusted values for the indicator weights i and the one or more tempering parameters using gradient descent that minimize the error function; and
output, by the output device, the indicator weights i and the one or more tempering parameters in a computer readable form to the automated classification system such that the indicator weights i and one or more tempering parameters can be used by the automated classification system to read in token data for a prospective entity h and use a model based on the indicator weights i and tempering parameters for determining a probability of the prospective entity h being in the class; and
wherein the class of each training entity j is associated with an event date, and the publication for each training entity j has a publication date after the training entity j's event date.
2 . The artificial intelligence system of claim 1 wherein the tempered indication is set to a value of 0 when all of the tokens for a training entity j have a binary value of 0.
3 . The artificial intelligence system of claim 1 wherein the tempering factor is set to a value of 1 when the unweighted sum of the tokens for a training entity j has a value of 1.
4 . The artificial intelligence system of claim 3 wherein:
a) the tempering parameters comprise a tempering weight; and
b) the tempering factor increases by the tempering weight when the unweighted sum of the tokens for a training entity j increases by 1.
5 . The artificial intelligence system of claim 4 wherein the tempering weight has a value of 0.1 or less.
6 . The artificial intelligence system of claim 1 wherein:
a) the tempering parameters comprise a ballast factor; and
b) the ballast factor is added to the weighted sum.
7 . The artificial intelligence system of claim 1 wherein the indicators comprise words, phrases or word stems.
8 . The artificial intelligence system of claim 1 wherein the error function is a sum of squares of the residuals.
9 . The artificial intelligence system of claim 1 wherein:
a) the prospective entity h is a claimant and an associated insurance claim;
b) the adjudicated binary class is a determination of whether or not the insurance claim is fraudulent; and
c) the publication is a social media publication by the claimant.