Auto-encoder enhanced self-diagnostic components for model monitoring

Patent metadata

Field	Value
Publication number	US-11836746-B2
Application number	US-201414558700-A
Country	US
Kind code	B2
Filing date	Dec 2, 2014
Priority date	Dec 2, 2014
Publication date	Dec 5, 2023
Grant date	Dec 5, 2023

Abstract

A diagnostic system for model governance is presented. The diagnostic system includes an auto-encoder to monitor model suitability for both supervised and unsupervised models. When applied to unsupervised models, the diagnostic system can provide a reliable indication on model degradation and recommendation on model rebuild. When applied to supervised models, the diagnostic system can determine the most appropriate model for the client based on a reconstruction error of a trained auto-encoder for each associated model. An auto-encoder can determine outliers among subpopulations of consumers, as well as support model go-live inspections.

First claim

What is claimed is:

1. A non-transitory computer-readable medium containing instructions to configure one or more data processors to perform operations to enhance capabilities of a fraud detection computing system, the operations comprising:
receiving, by a first autoencoder, historical customer transaction profile data input, representing past spending patterns of one or more customers;
receiving by a second auto encoder, historical transaction data input, representing deviation of transactional behavior of the one or more customers from their peers;
comparing the historical customer transaction profile and the historical transaction data inputs with data in a stored model to determine a likelihood of a fraudulent transaction or a fraudulent behavior; and
sorting, based on a selection criterion, extracted original data sampling from the historical transaction data feeds into a plurality of partitions;
encoding data inputs to one or more latent variables in at least one hidden layer of a neural network including the first autoencoder and the second autoencoder, such that quantity of the one or more latent variables is less than the quantity of the data inputs being input to the at least one input layer of the neural network, the one or more latent variables defining one or more first data patterns being different from one or more second data patterns of the stored model,
the encoding of the data inputs comprises encoding at least an input x into a latent representation z, such that the input x can be reconstructed from z according to z=σ(W E x+b E ), where W E and b E are the encoding weights and bias, respectively, and σ is the logistic function; and

decoding the one or more latent variables to generate a reconstructed data set of the extracted original data sampling, the reconstructed data set comprising a quantity of data outputs to at least one output layer of the neural network,
the decoding of the one or more latent variables reconstructs the input x according to x=W D z+b D ;

calculating a reconstruction error for at least one partition of the plurality of partitions utilizing the one or more latent variables, the reconstruction error representing a deviation of the quantity of data outputs of the reconstructed data set from the quantity of data inputs of the extracted original data sampling for the at least one partition;
calculating a total reconstruction error for the neural network based on combining the reconstruction error calculated for the plurality of partitions; and
minimizing the reconstruction error, E R , with respect to encoding and decoding parameters, W E , b E , W D , b D on a training set by minimizing an associated loss function to correct the neural network misalignment based on a periodic check on the total reconstruction error on a selected sampled data set selected from a plurality of sample data sets, the selected sample data set having a lowest reconstruction error, wherein use of the stored model for detecting the fraudulent transaction or the fraudulent behavior by the fraud detection computing system is determined based on the one or more first data patterns and minimized reconstruction error, thereby improving scalability and fault-tolerance of the fraud detection computing system.

2. The non-transitory computer-readable medium in accordance with claim 1 , wherein the stored model is in a go-live state, and wherein the historical learning of the stored model is in a fixed state.

3. The non-transitory computer-readable medium in accordance with claim 1 , wherein the one or more customers includes a plurality of subpopulations, and wherein the operations of the neural network further comprise identifying an outlier of the reconstruction error associated with at least one of the plurality of subpopulations.

4. The non-transitory computer-readable medium in accordance with claim 3 , wherein diagnostic operations of the neural network further comprise selecting, from a plurality of models stored by the analytics module, a best model according to a lowest reconstruction error for the outlier associated with the at least one of the plurality of subpopulations.

5. The non-transitory computer-readable medium in accordance with claim 1 , wherein the stored model is an unsupervised model.

6. The non-transitory computer-readable medium in accordance with claim 1 , wherein the stored model is a supervised model, and wherein the historical learning of the stored model includes human input data.

7. A computer-implemented method for enhancing capabilities of a fraud detection computing system, the operations comprising:
receiving, by a first autoencoder, historical customer transaction profile data input, representing past spending patterns of one or more customers;
receiving by a second auto encoder, historical transaction data input, representing deviation of transactional behavior of the one or more customers from their peers;
comparing the historical customer transaction profile and the historical transaction data inputs with data in a stored model to determine a likelihood of a fraudulent transaction or a fraudulent behavior; and
sorting, based on a selection criterion, extracted original data sampling from the historical transaction data input into a plurality of partitions;
encoding data inputs to one or more latent variables in at least one hidden layer of a neural network including the first autoencoder and the second autoencoder, such that quantity of the one or more latent variables is less than the quantity of the data inputs being input to the at least one input layer of the neural network, the one or more latent variables defining one or more first data patterns being different from one or more second data patterns of a stored model,
decoding the one or more latent variables to generate a reconstructed data set of the extracted original data sampling, the reconstructed data set comprising a quantity of data outputs output to at least one output layer of the neural network;
calculating a reconstruction error for at least one partition of the plurality of partitions utilizing the one or more latent variables, the reconstruction error representing a deviation of the quantity of data outputs of the reconstructed data set from the quantity of data inputs of the extracted original data sampling for the at least one partition;
calculating a total reconstruction error for the neural network based on combining the reconstruction error calculated for the plurality of partitions; and
minimizing the reconstruction error, E R , with respect to encoding and decoding parameters, W E , b E , W D , b D on a training set by minimizing an associated loss function to correct the neural network misalignment based on a periodic check on the total reconstruction error on a selected sampled data set, selected from a plurality of sample data sets, the selected sample data set having a lowest reconstruction error, wherein use of the stored model for detecting the fraudulent transaction or the fraudulent behavior by the fraud detection computing system is determined based on the one or more first data patterns and minimized reconstruction error, thereby improving scalability and fault-tolerance of the fraud detection computing system.

8. The method in accordance with claim 7 , wherein the one or more customers includes a plurality of subpopulations, and wherein the method further comprises identifying, by the at least one data processor, an outlier of the reconstruction error associated with at least one of the plurality of subpopulations.

9. The method in accordance with claim 7 , wherein the encoding of the data input comprises encoding at least an input x into a latent representation z, such that the input x can b

Assignees

• Fair Isaac Corp

Inventors

• Zhang Jun
• Zoldi Scott Michael

Classifications

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06Q30/0201 Primary

  Market modelling; Market analysis; Collecting market data · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N5/045

  Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence · CPC title