Feature-aware open set multi-model for trajectory classification in mobile edge devices
One example method includes training an event predictor, of a machine learning model, wherein the training includes training an event predictor of the machine learning model, and training open set trajectory classifier, of the machine learning model. After the event predictor is trained, the event predictor is operable to receive an unknown trajectory class and predict an event class for the unknown trajectory class. Further, after the open set trajectory classifier is trained, the open set trajectory classifier is operable to receive an unknown trajectory and classify the unknown trajectory with a predicted trajectory classification. Finally, during training, the output of the open-set trajectory classifier may be used as an input to the event predictor.
1 . A method, comprising:
training an event predictor, of a machine learning model, wherein the training comprises:
providing, to the event predictor, a predicted trajectory classification that is unknown to the event predictor component; and
providing, to the event predictor, a trajectory feature that is associated with the predicted trajectory classification,
wherein after the event predictor is trained with the predicted trajectory classification and the trajectory feature, the event predictor is operable to receive an unknown trajectory class and predict an event class for the unknown trajectory class, or predict that it is in fact an unknown trajectory class; and
training an open set trajectory classifier, of the machine learning model, wherein the training of the open set trajectory classifier comprises:
performing a closed set training process using a labeled dataset that maps a trajectory and its trajectory features to a trajectory class; and
performing an open set training process using the labeled dataset, and performing the open set training process using the labeled dataset comprises performing a first reconstruction of the trajectories in the labeled dataset for those trajectories in the labeled dataset that are in a correct class, and performing a second reconstruction of the trajectories in the labeled dataset for those trajectories in the labeled dataset that are in an incorrect class,
wherein after the open set trajectory classifier is trained, the open set trajectory classifier is operable to receive an unknown trajectory and classify the unknown trajectory with the predicted trajectory classification.
2 . The method as recited in claim 1 , wherein the predicted trajectory classification is provided by the open set trajectory classifier to the event predictor.
3 . The method as recited in claim 1 , wherein the unknown trajectory comprises information about a movement and/or position of a mobile edge device, in which the machine learning model is deployed, in a physical operating domain.
4 . The method as recited in claim 1 , wherein the open set training process for the open set trajectory classifier accounts for a class imbalance across trajectory classes that are known to the open set trajectory classifier.
5 . The method as recited in claim 1 , wherein the machine learning model is operable to detect abnormal events occurring during operation of a mobile edge device in which the machine learning model is deployed.
6 . The method as recited in claim 1 , wherein the event predictor comprises an auto-encoder that receives, as inputs, the predicted trajectory classification and the trajectory feature.
7 . The method as recited in claim 1 , wherein the open set trajectory classifier comprises an auto-encoder that receives, as input, trajectory features of the unknown trajectory.
8 . The method as recited in claim 1 , wherein training of the machine learning model is automatically triggered as a result of a monitoring process that tracks performance of the machine learning model.
9 . The method as recited in claim 1 , further comprising deploying the machine learning model, after the machine learning model is trained, to a group of mobile edge devices.
10 . A non-transitory storage medium having stored therein instructions that are executable by one or more hardware processors to perform operations comprising:
training an event predictor, of a machine learning model, wherein the training comprises:
providing, to the event predictor, a predicted trajectory classification that is unknown to the event predictor component; and
providing, to the event predictor, a trajectory feature that is associated with the predicted trajectory classification,
wherein after the event predictor is trained with the predicted trajectory classification and the trajectory feature, the event predictor is operable to receive an unknown trajectory class and predict an event class for the unknown trajectory class, or predict that it is in fact an unknown trajectory class; and
training an open set trajectory classifier, of the machine learning model, wherein the training of the open set trajectory classifier comprises:
performing a closed set training process using a labeled dataset that maps a trajectory and its trajectory features to a trajectory class; and
performing an open set training process using the labeled dataset, and performing the open set training process using the labeled dataset comprises performing a first reconstruction of the trajectories in the labeled dataset for those trajectories in the labeled dataset that are in a correct class, and performing a second reconstruction of the trajectories in the labeled dataset for those trajectories in the labeled dataset that are in an incorrect class,
wherein after the open set trajectory classifier is trained, the open set trajectory classifier is operable to receive an unknown trajectory and classify the unknown trajectory with the predicted trajectory classification.
11 . The non-transitory storage medium as recited in claim 10 , wherein the predicted trajectory classification is provided by the open set trajectory classifier to the event predictor.
12 . The non-transitory storage medium as recited in claim 10 , wherein the unknown trajectory comprises information about a movement and/or position of a mobile edge device, in which the machine learning model is deployed, in a physical operating domain.
13 . The non-transitory storage medium as recited in claim 10 , wherein the open set training process for the open set trajectory classifier accounts for a class imbalance across trajectory classes that are known to the open set trajectory classifier.
14 . The non-transitory storage medium as recited in claim 10 , wherein the machine learning model is operable to detect abnormal events occurring during operation of a mobile edge device in which the machine learning model is deployed.
15 . The non-transitory storage medium as recited in claim 10 , wherein the event predictor comprises an auto-encoder that receives, as inputs, the predicted trajectory classification and the trajectory feature.
16 . The non-transitory storage medium as recited in claim 10 , wherein the open set trajectory classifier comprises an auto-encoder that receives, as input, trajectory features of the unknown trajectory.
17 . The non-transitory storage medium as recited in claim 10 , wherein training of the machine learning model is automatically triggered as a result of a monitoring process that tracks performance of the machine learning model.
18 . The non-transitory storage medium as recited in claim 10 , wherein the operations further comprise deploying the machine learning model, after the machine learning model is trained, to a group of mobile edge devices.