# US 2021/0334673 A1 · Online unsupervised anomaly detection

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0334673 A1
- Document ID: US20210334673A1
- Status: Published application
- Kind: A1
- Application number: 16/856,071
- Filed: 2020-04-23
- Published: 2021-10-28
- Earliest claimed: 2020-04-23
- Assignees: Actimize Ltd.
- Applicants: Actimize LTD.
- Inventors: Danny BUTVINIK
- Web page: https://patentgrep.com/patent/US20210334673A1

## CPC

- G06N 5/04: Inference or reasoning models
- G06N 20/00: Machine learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A computerized-method for real-time detection of anomalous data, by processing high-speed streaming data. In a computerized-system receiving a data-stream comprised of unlabeled data points, and operating an Anomalous Data Detection (ADD) module. The ADD module receives at least one of: (i) k number of data point neighbors for each data point; (ii) X number of data points in a predetermined period of time; (iii) d number of dimensions of each data point, threshold; and (iv) n number of data points that said ADD module is operating on, in a predefined time unit. Then, the ADD module prepares a dataset having n data points from the received X data points; and then identifies one or more data points, from the received data stream, as outliers to send an alert with details related to the identified outliers, thus, dynamically evaluating local outliers in the received data stream.

## Claims

### Claim 1

A computerized-method for real-time detection of anomalous data, by processing high-speed streaming data, said computerized-method comprising:
in a computerized-system comprising a processor and a memory, receiving by the processor, a data stream comprised of unlabeled data points,
operating by the processor an Anomalous Data Detection (ADD) module, said ADD module is configured to:
(i) receive: k, X, d, threshold, and n,
wherein k is a number of data point neighbors for each data point,
wherein X is a number of data points in a predetermined period of time,
wherein d is a number of dimensions of each data point,
wherein n is a number of data points that said ADD module is operating on, in
a predefined time unit;
(ii) prepare a dataset having n data points from the received X data points; and
(iii) identify one or more data points, from the received data stream, as outliers to send an alert with details related to the identified outliers thus, dynamically evaluating local outliers in the received data stream.

### Claim 2 (depends on 1)

The computerized-method of claim 1, wherein the preparation of the dataset is comprising:
(i) fetching X data points from a data storage device, according to at least one preconfigured criterion;
(ii) retrieving random n data points from the retrieved X data points to yield a dataset;
(iii) for each data point in the dataset:
a. applying at least one classification algorithm to yield a set of results from each applied classification algorithm and to determine k data points neighbors;
b. marking the data points in the set of results as related to the dataset;
c. calculating a local density, wherein said local density is corresponding to a calculated heterogenous nearest neighbors Local Outlier Factor (hLOF) area of the data point;
d. associating an outlier counter and zeroing said outlier counter; and
e. marking said data point as a potential-outlier by increasing by 1 each associated outlier counter that its calculated local density is higher than 1.

### Claim 3 (depends on 2)

The computerized-method of claim 2, wherein the identifying of one or more data points as outliers is performed by:
a. reading a first data point and a second data point from the received data stream;
b. applying the at least one classification algorithm on the first data point and on each data point that is related to the prepared dataset, to yield a first group of results and to delete a data point that is not in the first group of results, from the dataset;
c. checking the associated outlier counter of the deleted data point to mark the deleted data point as an outlier when the value of the associated outlier counter is greater than the received threshold;
d. applying the at least one classification algorithm on the first data point and on each data point in the first group of results to yield a second group of results;
e. for each data point in the second group of results:
i. applying the at least one classification algorithm on the second data point and on each data point in the second group of results, to yield a third group of results for each data point and to delete at least one data point that is not in the third group of results, from the second group of results;
ii. calculating a local density, wherein said local density is corresponding to a calculated heterogenous nearest neighbors Local Outlier Factor (hLOF) area of the first group of results;
f. checking the associated outlier counter of the at least one deleted data point that is not in the third group of results to mark the at least one deleted data point as an outlier when the value of the associated outlier counter is greater than the received threshold;
g. applying the at least one classification algorithm on the second data point and on each data point in the third group of results, to yield a fourth group of results;
h. for each data point in the fourth group of results:
i. applying at least one classification algorithm to yield a fifth group of results from each applied classification algorithm and to determine k data points neighbors;
ii. calculating a local density, wherein said local density is corresponding to a calculated heterogenous nearest neighbors Local Outlier Factor (hLOF) area of the data point;
iii. marking said data point as a potential-outlier by increasing by 1 each associated outlier counter that its calculated local density is higher than 1; and
i. marking the fifth group of results as a dataset; and
j. repeating step (a) through step (j).

### Claim 4 (depends on 3)

The computerized-method of claim 3, wherein the first group of results, the second group of results, the third group of results, the fourth group of results and the fifth group of results include determined k data points neighbors.

### Claim 5 (depends on 1)

The computerized-method of claim 1, wherein the alert that is sent is presented to a user via a display unit.

### Claim 6 (depends on 1)

The computerized-method of claim 1, wherein the alert that is sent is forwarded to a preconfigured application.

### Claim 7 (depends on 1)

The computerized-method of claim 1, wherein the data stream is a financial data stream and the detected anomalous activity is fraud.

### Claim 8 (depends on 1)

The computerized-method of claim 1, wherein the data stream is selected from a group of: environmental sensors, social network analysis, computer network traffic, web searches or other.

### Claim 9 (depends on 2)

The computerized-method of claim 2, wherein the at least one classification algorithm is selected from:
a. k Nearest Neighbors (kNN) algorithm;
b. Inverse of k Nearest Neighbors (IkNN) algorithm;
c. Joint of k Nearest Neighbors (JkNN) algorithm;
or any combination thereof.

### Claim 10 (depends on 2)

The computerized-method of claim 2, wherein the local density is calculated by applying a distance-based method that is calculating an average distance between the read data point and k nearest data point neighbors.

### Claim 11 (depends on 10)

The computerized-method of claim 10, wherein the k nearest data point neighbors are calculated according to distance metrics.

### Claim 12 (depends on 11)

The computerized-method of claim 11, wherein the distance metrics are Euclidean distance metric, Manhattan distance metric or other distance metrics.

### Claim 13 (depends on 2)

The computerized-method of claim 2, wherein the calculated heterogenous nearest neighbors Local Outlier Factor (hLOF) area is a set of data points which is a combination of the set of results of each operated classification algorithm.

### Claim 14 (depends on 1)

The computerized-method of claim 1, wherein the received threshold is determined according to the number of standard deviations of a data point from the mean.

### Claim 15 (depends on 1)

The computerized-method of claim 1, wherein the number of dimensions of a data point is a number of attributes therein.

### Claim 16 (depends on 1)

The computerized-method of claim 1, wherein the data point neighbors are determined as such, according to preconfigured distance metrics.
