Method and system for managing features for optimal model performance and application thereof
AssigneeVerizon Patent and Licensing Inc.
InventorVarun Unnikrishnannair
Life of this patent
Estimating the term…
Abstract
The present teaching relates to manage data features for machine learning a prediction model. Based on event sequence data, features are extracted and feature segments are created. Critical signals are generated based on specified targets to be predicted that representing the operation of a network. Target-based weights on respective feature segments are determined based on historical data representing past operation states of the network and are used to encode the feature segments. The encoded feature segments are used to generate target-based optimal training data, which is used to train models for predicting the specified targets based on real0time network data.
Claims20
A method, comprising:
- collecting event sequence data characterizing states of operation of a network;
- extracting features from the event sequence data;
- creating, based on the features, feature segments, each of which includes one or more of the features;
- receiving information specifying at least one target to be predicted based on data that is representative of corresponding at least one aspect of the operation of the network;
- generating at least one critical signal corresponding to the at least one target;
- determining, based on historical data representing past operational states of the network, a target-based weight indicative of relevance of each of the different feature segments with respect to each of the at least one critical signal;
- encoding the feature segments based on the target-based weights thereof with respect to the at least one critical signal;
- generating target-based optimal training data based on the encoded feature segments;
- and
- training, using the target-based optimal training data, one or more prediction models for predicting the at least one target based on real-time data from the network.
A machine-readable and non-transitory medium having information recorded thereon, wherein the information, when read by the machine, causes the machine to perform the s=following steps:
- collecting event sequence data characterizing states of operation of a network;
- extracting features from the event sequence data;
- creating, based on the features, feature segments, each of which includes one or more of the features;
- receiving information specifying at least one target to be predicted based on data that is representative of corresponding at least one aspect of the operation of the network;
- generating at least one critical signal corresponding to the at least one target;
- determining, based on historical data representing past operational states of the network, a target-based weight indicative of relevance of each of the different feature segments with respect to each of the at least one critical signal;
- encoding the feature segments based on the target-based weights thereof with respect to the at least one critical signal;
- generating target-based optimal training data based on the encoded feature segments;
- and
- training, using the target-based optimal training data, one or more prediction models for predicting the at least one target based on real-time data from the network.
A system, comprising:
- a feature extractor implemented by a processor and configured for
- collecting event sequence data characterizing states of operation of a network, and
- extracting features from the event sequence data;
- a feature segment generator implemented by a processor and configured for creating, based on the features, feature segments, each of which includes one or more of the features;
- a critical signal generator implemented by a processor and configured for
- receiving information specifying at least one target to be predicted based on data that is representative of corresponding at least one aspect of the operation of the network, and
- generating at least one critical signal corresponding to the at least one target;
- a feature weight determiner implemented by a processor and configured for determining, based on historical data representing past operational states of the network, a target-based weight indicative of relevance of each of the different feature segments with respect to each of the at least one critical signal;
- a feature segment weight encoder implemented by a processor and configured for encoding the feature segments based on the target-based weights thereof with respect to the at least one critical signal;
- an optimal training data generator implemented by a processor and configured for generating target-based optimal training data based on the encoded feature segments;
- and
- a machine learning engine implemented by a processor and configured for training, using the target-based optimal training data, one or more prediction models for predicting the at least one target based on real-time data from the network.
Dependencies are inferred from the claim text.