Convolutional neural network (cnn)-based suggestions for anomaly input

Patent metadata

Field	Value
Publication number	US-2019108439-A1
Application number	US-201715726268-A
Country	US
Kind code	A1
Filing date	Oct 5, 2017
Priority date	Oct 5, 2017
Publication date	Apr 11, 2019
Grant date	—

Abstract

The technology disclosed determines one or more field values in a set of field values for a particular field in a fielded dataset that are similar to an input value using six similarity measures. A factor vector is generated per similarity measure and combined to form an input matrix. A convolutional neural network processes the input matrix to generate evaluation vectors. A fully-connected network evaluates the evaluation vectors to generate suggestion scalars for similarity to a particular input value. Thresholding is applied to suggestions scalars to determine one or more suggestion candidates for the particular input value.

First claim

What is claimed is:

1 . A neural network-based suggestion method, including:
determining that one or more field values in a set of field values are similar to an input value for a particular field in a fielded dataset, including
comparing a particular input value to unique field values for the particular field by
applying a plurality of similarity measures and
generating a factor vector that has one scalar for each of the unique field values;

evaluating the factor vector using convolution filters in a convolutional neural network (CNN) to generate evaluation vectors for similarity to the unique field values; further evaluating the evaluation vectors using a fully-connected (FC) neural network to produce suggestion scalars for similarity to the particular input value; and using the suggestion scalars to determine one or more suggestion candidates for the particular input value.

2 . The method of claim 1 , wherein the plurality of similarity measures include semantic similarity, syntactic similarity, soundex similarity, character-by-character format similarity, field length similarity, and dataset frequency similarity.

3 . The method of claim 1 , further including constructing an input to the CNN by column-wise arranging one or more factor vectors in an input matrix.

4 . The method of claim 3 , wherein the convolution filters apply row-wise on the input matrix.

5 . The method of claim 1 , further including:
automatically constructing positive and negative examples for inclusion in a training dataset by:
for a given linguistic similarity measure,
determining a first set of similar field values from a vocabulary and determining a second set of dissimilar field values from the vocabulary; and
randomly selecting some field values from the first and second sets as positive and negative examples respectively;

iterating the determining and the selecting for a plurality of similarity measures; and storing the randomly selected field values for the plurality of similarity measures as the training dataset.

6 . The method of claim 5 , further including training the CNN and the FC neural network using the positive and negative examples in the training dataset.

7 . The method of claim 1 , further including:
using at least one cost function to evaluate performance of the CNN and the FC neural network during training.

8 . The method of claim 1 , wherein the CNN is a one-layer CNN.

9 . The method of claim 1 , wherein the CNN is a two-layer CNN.

10 . The method of claim 1 , further including:
determining which field values for a particular field in the fielded dataset are anomalous, including
comparing a particular unique field value to other unique field values for the particular field by
applying a plurality of similarity measures and
generating a factor vector that has one scalar for each of the unique field values;

evaluating the factor vector using the convolution filters in the CNN to generate evaluation vectors;
further evaluating the evaluation vectors using the FC neural network to produce an anomaly scalar for the particular unique field value; and thresholding the anomaly scalar to determine whether the particular unique field value is anomalous.

11 . A system including one or more processors coupled to memory, the memory loaded with computer instructions to provide neural network-based suggestions, the instructions, when executed on the processors, implement actions comprising:
determining that one or more field values in a set of field values are similar to an input value for a particular field in a fielded dataset, including
comparing a particular input value to unique field values for the particular field by
applying a plurality of similarity measures and
generating a factor vector that has one scalar for each of the unique field values;

evaluating the factor vector using convolution filters in a convolutional neural network (CNN) to generate evaluation vectors for similarity to the unique field values; further evaluating the evaluation vectors using a fully-connected (FC) neural network to produce suggestion scalars for similarity to the particular input value; and using the suggestion scalars to determine one or more suggestion candidates for the particular input value.

12 . The system of claim 11 , wherein the plurality of similarity measures include semantic similarity, syntactic similarity, soundex similarity, character-by-character format similarity, field length similarity, and dataset frequency similarity.

13 . The system of claim 11 , further implementing actions comprising constructing an input to the CNN by column-wise arranging one or more factor vectors in an input matrix.

14 . The system of claim 13 , wherein the convolution filters apply row-wise on the input matrix.

15 . The system of claim 11 , further implementing actions comprising:
automatically constructing positive and negative examples for inclusion in a training dataset by:
for a given linguistic similarity measure,
determining a first set of similar field values from a vocabulary and
determining a second set of dissimilar field values from the vocabulary; and
randomly selecting some field values from the first and second sets as positive and negative examples respectively;

iterating the determining and the selecting for a plurality of similarity measures; and
storing the randomly selected field values for the plurality of similarity measures as the training dataset.

16 . The system of claim 15 , further implementing actions comprising training the CNN and the FC neural network using the positive and negative examples in the training dataset.

17 . The system of claim 11 , further implementing actions comprising:
using at least one cost function to evaluate performance of the CNN and the FC neural network during training.

18 . The system of claim 11 , wherein the CNN is a one-layer CNN.

19 . The system of claim 11 , further implementing actions comprising:
determining which field values for a particular field in the fielded dataset are anomalous, including
comparing a particular unique field value to other unique field values for the particular field by
applying a plurality of similarity measures and
generating a factor vector that has one scalar for each of the unique field values;

evaluating the factor vector using the convolution filters in the CNN to generate evaluation vectors; further evaluating the evaluation vectors using the FC neural network to produce an anomaly scalar for the particular unique field value; and thresholding the anomaly scalar to determine whether the particular unique field value is anomalous.

20 . A non-transitory computer readable storage medium impressed with computer program instructions to provide neural network-based suggestions, the instructions, when executed on a processor, implement a method comprising:
determining that one or more field values in a set of field values are similar to an input value for a particular field in a fielded dataset, including
comparing a particular input value to unique field values for the particular field by
applying a plurality of similarity measures and
generating a factor vector that has one scalar for each of the unique field values;

evaluating the factor vector using convolution filters in a convolutional neural network (CNN) to generate evaluation vectors for similarity to the unique field values; further evaluating the evaluation vectors using a fu

Assignees

• Salesforce Com Inc

Inventors

• Lu Chang
• Zhang Lingtao

Classifications

• G06N3/08 Primary

  Learning methods · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06Q30/020121 Primary

  Sales lead analysis · CPC title