Abstract: In one embodiment, a device in a network maintains a machine learning-based recursive model that models a time series of observations regarding a monitored entity in the network. The device applies sparse dictionary learning to the recursive model, to find a decomposition of a particular state vector of the recursive model. The decomposition of the particular state vector comprises a plurality of basis vectors. The device determines a mapping between at least one of the plurality of basis vectors for the particular state vector and one or more human-readable interpretations of the basis vectors. The device provides a label for the particular state vector to a user interface. The label is based on the mapping between the at least one of the plurality of basis vectors for the particular state vector and the one or more human-readable interpretations of the basis vectors.

Abstract: In one embodiment, a device in a network receives a notification of a particular anomaly detected by a distributed learning agent in the network that executes a machine learning-based anomaly detector to analyze traffic in the network. The device computes one or more distance scores between the particular anomaly and one or more previously detected anomalies. The device also computes one or more relevance scores for the one or more previously detected anomalies. The device determines a reporting score for the particular anomaly based on the one or more distance scores and on the one or more relevance scores. The device reports the particular anomaly to a user interface based on the determined reporting score.

Abstract: In one embodiment, a device in a network receives an indication of a connection between an endpoint node in the network and a conferencing service. The device retrieves network data associated with the indicated connection between the endpoint node and the conferencing service. The device uses a machine learning model to predict an experience metric for the endpoint node based on the network data associated with the indicated connection between the endpoint node and the conferencing service. The device causes the endpoint node to use a different connection to the conferencing service based on the predicted experience metric.

Abstract: In one embodiment, a device in a network receives traffic records indicative of network traffic between different sets of host address pairs. The device identifies one or more address grouping constraints for the sets of host address pairs. The device determines address groups for the host addresses in the sets of host address pairs based on the one or more address grouping constraints. The device provides an indication of the address groups to an anomaly detector.

Abstract: In one embodiment, a networking device at an edge of a network generates a first set of feature vectors using information regarding one or more characteristics of host devices in the network. The networking device forms the host devices into device clusters dynamically based on the first set of feature vectors. The networking device generates a second set of feature vectors using information regarding traffic associated with the device clusters. The networking device models interactions between the device clusters using a plurality of anomaly detection models that are based on the second set of feature vectors.

Abstract: In one embodiment, a device in a network identifies a plurality of applications from observed traffic in the network. The device forms two or more application clusters from the plurality of applications. Each of the application clusters includes one or more of the applications, and wherein a particular application in the plurality of applications is included in each of the application clusters. The device generates anomaly detection models for each of the application clusters. The device tests the anomaly detection models, to determine a measure of efficacy for each of the models with respect to traffic associated with the particular application. The device selects a particular anomaly detection model to analyze the traffic associated with the particular application based on the measures of efficacy for each of the models.

Abstract: In one embodiment, a device in a network identifies a new interaction between two or more nodes in the network. The device forms a feature vector using contextual information associated with the new interaction between the two or more nodes. The device causes generation of an anomaly detection model for new node interactions using the feature vector. The device uses the anomaly detection model to determine whether a particular node interaction in the network is anomalous.

Abstract: In one embodiment, a device in a network detects an anomaly in the network by analyzing a set of sample data regarding one or more conditions of the network using a behavioral analytics model. The device receives feedback regarding the detected anomaly. The device determines that the anomaly was a true positive based on the received feedback. The device excludes the set of sample data from a training set for the behavioral analytics model, in response to determining that the anomaly was a true positive.

Abstract: In one embodiment, a networking device in a network causes formation of device clusters of devices in the network. The devices in a particular cluster exhibit similar characteristics. The networking device receives feedback from a device identity service regarding the device clusters. The feedback is based in part on the device identity service probing the devices. The networking device adjusts the device clusters based on the feedback from the device identity service. The networking device performs anomaly detection in the network using the adjusted device clusters.

Abstract: In one embodiment, a device in a network maintains a plurality of anomaly detection models for different sets of aggregated traffic data regarding traffic in the network. The device determines a measure of confidence in a particular one of the anomaly detection models that evaluates a particular set of aggregated traffic data. The device dynamically replaces the particular anomaly detection model with a second anomaly detection model configured to evaluate the particular set of aggregated traffic data and has a different model capacity than that of the particular anomaly detection model. The device provides an anomaly event notification to a supervisory controller based on a combined output of the second anomaly detection model and of one or more of the anomaly detection models in the plurality of anomaly detection models.

Abstract: In one embodiment, a device in a network receives an indication that a network anomaly detected by an anomaly detector of a first node in the network is associated with scanning activity in the network. The device receives labeled traffic data associated with the detected anomaly that identifies whether the traffic data is associated with legitimate or illegitimate scanning activity. The device trains a machine learning-based classifier using the labeled traffic data to distinguish between legitimate and illegitimate scanning activity in the network. The device deploys the trained classifier to the first node, to distinguish between legitimate and illegitimate scanning activity in the network.

Abstract: In one embodiment, a device in a network determines cluster assignments that assign traffic data regarding traffic in the network to activity level clusters based on one or more measures of traffic activity in the traffic data. The device uses the cluster assignments to predict seasonal activity for a particular subset of the traffic in the network. The device determines an activity level for new traffic data regarding the particular subset of traffic in the network. The device detects a network anomaly by comparing the activity level for the new traffic data to the predicted seasonal activity.

Abstract: In one embodiment, a device in a network loads an anomaly detection model for warm-start. The device filters input data for the model during a warm-start grace period after warm-start of the anomaly detection model. The model is not updated during the warm-start grace period based on the filtering. The device determines an end to the warm-start grace period. The device updates the anomaly detection model using unfiltered input data for the anomaly detection model after the determined end to the warm-start grace period.