Unsupervised outlier detection in time-series data
Systems and methods for detecting patterns in data from a time-series and for detecting outliers in network data in an unsupervised manner are provided. In one implementation, a method includes the steps of obtaining network data from a network to be monitored and creating a window from the obtained network data. The method also includes the step of detecting outliers of the obtained data with respect to the window using an unsupervised deep learning process (e.g., using a Generalized Adversarial Network (GAN) learning technique and/or a Bidirectional GAN (BiGAN) learning technique) for enabling the learning of a data distribution. The unsupervised process, for example, does not require manual intervention.
1 . A non-transitory computer-readable medium configured to store a program executable by a processing system, the program including instructions configured to cause the processing system to:
obtain raw, heterogeneous time-series data from multiple components of a communication network to be monitored at the processing system,
convert all of the raw, heterogeneous time-series data to a two-dimensional image format by constructing tensors from all of the raw, heterogeneous time-series data,
create a multi-dimensional matrix comprising a plurality of image windows formed from the raw, heterogeneous time-series data using the processing system,
detect or forecast at least one of a network threshold crossing, a network alarm, a network user quality-of-experience, or a network anomaly by detecting, utilizing a deep neural network, outliers of the raw, heterogeneous time-series data with respect to the multi-dimensional matrix using an unsupervised deep learning process on the two-dimensional image format of all of the raw, heterogeneous time-series data, the unsupervised deep learning process including training a first path utilizing a Generalized Adversarial Network (GAN) learning technique to obtain a discriminator from a trained GAN data and training a second path utilizing a Bidirectional GAN (BiGAN) learning technique to obtain an encoder from a trained BiGAN data and executed by the processing system to transfer knowledge of other experimented domains for imbalanced datasets anomaly detection by identifying portions of the raw, heterogeneous time-series data as normal by the absence of an indicator that the identified portions are abnormal such that a determined low probability that another portion of the raw, heterogeneous time-series data is within the normal portions of the raw, heterogeneous time-series data indicates an anomaly, wherein a first probability is determined from the discriminator, a second probability is determined from the encoder, and the detected outliers are determined based on a combination of the first probability and the second probability; and
localize the detected or forecasted at least one of the network threshold crossing, the network alarm, the network user quality-of-experience, or the network anomaly within the raw, heterogeneous time-series data by localizing the detected outliers in the obtained, raw, heterogeneous time-series data by detecting at least one image window of the plurality of image windows containing the detected outliers.
2 . The non-transitory computer-readable medium of claim 1 , wherein detection of outliers includes a determination if the network threshold has been crossed.
3 . The non-transitory computer-readable medium of claim 1 , wherein the raw, heterogeneous time-series data is obtained for a first time-step and for a second time-step, and wherein the plurality of image windows are each created through creation of a first image window from the raw, heterogeneous time-series data for the first time-step and creation of a second image window from the raw, heterogeneous time-series data for the second time-step.
4 . The non-transitory computer-readable medium of claim 3 , wherein the instructions further cause the processing system to
subtract a background from the second image window to remove changes in the raw, heterogeneous time-series data and to obtain a resulting image window, and
use the resulting image window for training and outlier detection.
5 . The non-transitory computer-readable medium of claim 1 , wherein the plurality of image windows are each created by creation of a first image window covering a first time period and creation of a second image window covering a second time period, the first time period covering a shorter time than the second time period, and wherein the step of transferring knowledge of other experimented domains for imbalanced datasets anomaly detection includes the steps of training a first model using the first image window and training a second model using the second image window.
6 . The non-transitory computer-readable medium of claim 5 , wherein the instructions further cause the processing system to
calculate a first outlier score using the first image window and the first model,
calculate a second outlier score using the second image window and the second model,
calculate a score difference between the first outlier score and the second outlier score, and
if the score difference exceeds a predetermined threshold, retrain the first and second models with subsequent data.
7 . The non-transitory computer-readable medium of claim 5 , wherein the instructions further cause the processing system to
create a first prediction using the first image window having the first time period,
create a second prediction using the second image window having the second time period, and
compare the first prediction with the second prediction to detect a changing distribution in the data.
8 . A system for detecting outliers of network data, the system comprising:
one or more processors; and
a memory in communication with the one or more processors, the memory configured to store instructions for detecting outliers of network data,
wherein the instructions, when executed, cause the one or more processors to
obtain raw, heterogeneous time-series data from multiple components of a communication network to be monitored,
convert all of the raw, heterogeneous time-series data to a two-dimensional image format by constructing tensors from all of the raw, heterogeneous time-series data,
create a multi-dimensional matrix comprising a plurality of image windows from the obtained, raw, heterogeneous time-series data,
detect or forecast at least one of a network threshold crossing, a network alarm, a network user quality-of-experience, or a network anomaly by detecting, utilizing a deep neural network, outliers of the obtained, raw, heterogeneous time-series data with respect to the multi-dimensional matrix using an unsupervised learning process on the two-dimensional image format of all of the raw, heterogeneous time-series data, the unsupervised deep learning process including training a first path utilizing a Generalized Adversarial Network (GAN) learning technique to obtain a discriminator from a trained GAN data and training a second path utilizing a Bidirectional GAN (BiGAN) learning technique to obtain an encoder from a trained BiGAN data, wherein the unsupervised learning process transfers knowledge of other experimented domains for imbalanced datasets anomaly detection by identifying portions of the raw, heterogeneous time-series data as normal by the absence of an indicator that the identified portions are abnormal such that a determined low probability that another portion of the raw, heterogeneous time-series data is within the normal portions of the raw, heterogeneous time-series data indicates an anomaly, wherein a first probability is determined from the discriminator, a second probability is determined from the encoder, and the detected outliers are determined based on a combination of the first probability and the second probability; and
localize the detected or forecasted at least one of the network threshold crossing, the network alarm, the network user quality-of-experience, or the network anomaly within the raw, heterogeneous time-series data by localizing the detected outliers in the obtained, raw, heterogeneous time-series data by detecting at least one image window of the plurality of image windows containing the detected outliers.
9 . The system of claim 8 , wherein the instructions are further configured to cause the one or more processors to
obtain the raw, heterogeneous time-series data for a first time-step,
obtain the raw, heterogeneous time-series data for a second time-step,
create a first image window from the raw, heterogeneous time-series data for the first time-step,
create a second image window from the raw, heterogeneous time-series data for the second time-step,
subtract a background from the second image window to remove changes in the raw, heterogeneous time-series data and to obtain a resulting image window, and
utilize the resulting image window for training and outlier detection.
10 . The system of claim 8 , wherein the instructions are further configured to cause the one or more processors to
create a first image window for a first time-period size,
create a second image window for a second time-period size,
train a first model using the first image window, and
train a second model using the second image window.
11 . The system of claim 10 , wherein the instructions are further configured to cause the one or more processors to
calculate a first outlier score using the first image window and the first model,
calculate a second outlier score using the second image window and the second model,
calculate a score difference between the first outlier score and the second outlier score, and
if the score difference exceeds a predetermined threshold, retrain the first and second models with subsequent network time-series data.
12 . The system of claim 10 , wherein the instructions are further configured to cause the one or more processors to
create a first prediction using the first image window having the first time-period size,
create a second prediction using the second image window having the second time-period size, and
compare the first prediction with the second prediction to detect changing network distribution data.
13 . A method comprising the steps of:
obtaining raw, heterogeneous time-series data from multiple components of a communication network to be monitored at a processing system;
converting all of the raw, heterogeneous time-series data to a two-dimensional image format by constructing tensors from all of the raw, heterogeneous time-series data;
creating a multi-dimensional matrix comprising a plurality of image windows formed from the raw, heterogeneous time-series data using the processing system;
detecting or forecasting at least one of a network threshold crossing, a network alarm, a network user quality-of-experience, or a network anomaly by detecting, utilizing a deep neural network, outliers of the raw, heterogeneous time-series data with respect to the multi-dimensional matrix using an unsupervised deep learning process on the two-dimensional image format of all of the raw, heterogeneous time-series data, the unsupervised deep learning process including training a first path utilizing a Generalized Adversarial Network (GAN) learning technique to obtain a discriminator from a trained GAN data and training a second path utilizing a Bidirectional GAN (BiGAN) learning technique to obtain an encoder from a trained BiGAN data and executed by the processing system to transfer knowledge of other experimented domains for imbalanced datasets anomaly detection by identifying portions of the raw, heterogeneous time-series data as normal by the absence of an indicator that the identified portions are abnormal such that a determined low probability that another portion of the raw, heterogeneous time-series data is within the normal portions of the raw, heterogeneous time-series data indicates an anomaly, wherein a first probability is determined from the discriminator, a second probability is determined from the encoder, and the detected outliers are determined based on a combination of the first probability and the second probability; and
localizing the detected or forecasted at least one of the network threshold crossing, the network alarm, the network user quality-of-experience, or the network anomaly within the raw, heterogeneous time-series data by localizing the detected outliers in the obtained, raw, heterogeneous time-series data by detecting at least one image window of the plurality of image windows containing the detected outliers.