Machine learning systems and methods for attributed sequences

Patent metadata

Field	Value
Publication number	US-12086718-B2
Application number	US-201816057025-A
Country	US
Kind code	B2
Filing date	Aug 7, 2018
Priority date	Aug 7, 2018
Publication date	Sep 10, 2024
Grant date	Sep 10, 2024

Abstract

Machine learning systems and methods for embedding attributed sequence data. The attributed sequence data includes an attribute data part having a fixed number of attribute data elements and a sequence data part having a variable number of sequence data elements. An attribute network module includes a feedforward neural network configured to convert the attribute data part to an encoded attribute vector having a first number of attribute features. A sequence network module includes a recurrent neural network configured to convert the sequence data part to an encoded sequence vector having a second number of sequence features. In use, the machine learning system learns and outputs a fixed-length feature representation of input attributed sequence data which encodes dependencies between different attribute data elements, dependencies between different sequence data elements, and dependencies between attribute data elements and sequence data elements within the attributed sequence data.

First claim

What is claimed is:

1. A machine learning system for embedding attributed sequence data comprising an attribute data part having a fixed number of attribute data elements and a sequence data part having a variable number of sequence data elements into a fixed-length feature representation for a fraud detection system, wherein the machine learning system comprises a multilayer feedforward neural network having an attribute data input layer and an attribute vector output layer which comprises a first predetermined number of units, operatively coupled to a long short-term memory (LSTM) network which comprises a second predetermined number of hidden units which is equal to the first predetermined number of units, wherein an output of the attribute vector output layer is operatively coupled to an input of an attribute vector input layer of the LSTM network, and wherein the attribute vector input layer of the LSTM network comprises a hidden state of the LSTM network at a first evaluation step, the machine learning system comprising:
a computing device; and
a computer-readable storage medium comprising a set of instructions that upon execution by the computing device cause the machine learning system to:
obtain a dataset comprising a plurality of attributed sequences based on user behavior associated with user actions; and
for each attributed sequence in the dataset,
train the multilayer feedforward neural network using the attribute data part of the attributed sequence via back-propagation with respect to a first objective function, and
train the LSTM network using the sequence data part of the attributed sequence via back-propagation with respect to a second objective function,

wherein training of the multilayer feedforward neural network is coupled with training the LSTM network such that, in use, the machine learning system is configured to:
identify common behaviors based on clusters in points in feature space;
determine a fixed-length feature representation of input attributed sequence data based on an analysis of the user behavior associated with the user actions within the fraud detection system that includes the identified common behaviors, wherein the fixed-length feature representation of input attributed sequence data comprises the hidden state of the LSTM network at a final evaluation step;
identify potential fraudulent behaviors based on isolated points within the fixed-length feature representation; and
output the fixed-length feature representation which encodes: i) dependencies between different attribute data elements in the attribute data part, ii) dependencies between different sequence data elements in the sequence data part, and iii) dependencies between attribute data elements and sequence data elements within the attributed sequence data.

2. The machine learning system of claim 1 wherein the LSTM network comprises a recurrent neural network.

3. The machine learning system of claim 1 wherein the set of instructions that upon execution by the computing device further cause the machine learning system to:
learn a nonlinear function of a concatenation which encodes dependencies between attribute data elements and sequence data elements within the attributed sequence data.

4. A training method of a machine learning system for embedding attributed sequence data comprising an attribute data part having a fixed number of attribute data elements and a sequence data part having a variable number of sequence data elements into a fixed-length feature representation for a fraud detection system, wherein:
the machine learning system comprises a multilayer feedforward neural network having an attribute data input layer and an attribute vector output layer which comprises a first predetermined number of units, operatively coupled to a long short-term memory (LSTM) network which comprises a second predetermined number of hidden units,
an output of the attribute vector output layer is operatively coupled to an input of an attribute vector input layer of the LSTM network,
the attribute vector input layer of the LSTM network comprises a hidden state of the LSTM network at a first evaluation step, and
the first predetermined number of units of the attribute vector output layer is equal to the second predetermined number of hidden units,
the training method comprising:
obtaining a dataset comprising a plurality of attributed sequences based on user behavior associated with user actions; and
for each attributed sequence in the dataset,
training the multilayer feedforward neural network using the attribute data part of the attributed sequence via back-propagation with respect to a first objective function, and
training the LSTM network using the sequence data part of the attributed sequence via back-propagation with respect to a second objective function,

wherein training of the multilayer feedforward neural network is coupled with training the LSTM network such that, in use, the machine learning system is configured to:
identify common behaviors based on clusters in points in feature space;
determine a fixed-length feature representation of input attributed sequence data based on an analysis of the user behavior associated with the user actions within the fraud detection system that includes the identified common behaviors, wherein the fixed-length feature representation of input attributed sequence data comprises the hidden state of the LSTM network at a final evaluation step;
identify potential fraudulent behaviors based on isolated points within the fixed-length feature representation; and
output a fixed-length feature representation of input attributed sequence data which encodes: i) dependencies between different attribute data elements in the attribute data part, ii) dependencies between different sequence data elements in the sequence data part, and iii) dependencies between attribute data elements and sequence data elements within the attributed sequence data.

5. The training method of claim 4 wherein the LSTM network comprises a recurrent neural network.

6. The training method of claim 4 wherein the multilayer feedforward neural network comprises:
an encoder having an encoder input layer which comprises the attribute data input layer and an encoder output layer which comprises the attribute vector output layer; and
a decoder having a decoder input layer coupled to the encoder output layer, and a decoder output layer which comprises a reconstructed estimate of an input to the encoder input layer,
and wherein:
the first objective function comprises a distance measure between the input to the encoder input layer and the reconstructed estimate, and
training the multilayer feedforward neural network comprises:

iteratively performing steps of forward- and back-propagation with the attribute data part of the attributed sequence as input to the encoder input layer until the distance measure satisfies a first convergence target.

7. The training method of claim 6 wherein the second objective function comprises a likelihood measure of incorrect prediction of a next sequence item at each one of a plurality of training time steps of the LSTM network, and training the LSTM network comprises:
iteratively repeating the plurality of training time steps until the likelihood measure satisfies a second convergence target, each iteration comprising:
at a first training time step, copying the output of the attribute vector output layer to the hidden state of the LSTM network; and
at a final training time step, computing the likelihood measure.

8. The training method of claim 7 wherein the distance measure comprises a mean-squared-error loss function and the likelihood measure comprises a categorical cross-entropy loss funct

Assignees

• Amadeus Sas

Inventors

• Zhuang Zhongfang
• Arora Aditya
• Zouaoui Jihane
• Kong Xiangnan
• Rundensteiner Elke

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/0455

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

• G06N3/09

  Supervised learning · CPC title

• G06N3/045

  Combinations of networks · CPC title