Machine learning platform and feature generation for efficient data processing
A system enables agile model development to speed up innovation by data scientists. Model training and deployment are coordinated and standardized to reduce redundancy. Data is obtained for feature generation and reformatted and de-sensitized for storage. The features are stored in locations available to all models and training modules of a system so data does not need to be adjusted for new models. To generate a machine learning model, the system establishes a cohort for evaluation by the model. A model template and features for use by the model are identified. The selected template and features are used for experimentation and evaluation. Model training artifacts, such as model weights are subsequently recorded in a model store and the model scripts and settings can then be registered in a centralized database where it can be accessed for execution.
1 . A computer-implemented method of feature generation comprising:
creating a first model template and storing the first model template in a template store, wherein the first model template is configured to use a first feature set comprised of a first training feature set and a first prediction feature set;
identifying a set of users associated with an organization;
obtaining, for each user, a set of data values associated with the user's interactions with the organization;
normalizing the set of data values using domain-specific knowledge;
determining a score date, wherein the score date is determined with respect to an occurrence of an event for each of the users;
dividing the sets of data values into a set of user data values that occurs before the score date and a set of user data values that occurs on or after the score date;
generating training feature sets for training models using the normalized set of user data values that occur before the score date;
generating prediction feature sets for making predictions using the trained models using the normalized set of user data values that occur on or after the score date;
storing the generated training and prediction feature sets in a centralized feature store accessible by the organization so that the organization does not have to re-format the training and prediction feature sets each time a new model is trained or an existing model is adjusted, wherein the generated training and prediction features sets are comprised of the first training feature set and the first prediction feature set, respectively, for use with the first model template;
selecting the first model template from the template store;
selecting the first feature set associated with the first model template from the centralized feature store;
training, by at least one processor, a first machine learning model using the first model template and the first training feature set by training weights of the first machine learning model using labeled feature data from before the score date;
testing the trained first machine learning model using the first prediction feature set;
storing the trained first machine learning model in a model store; and
storing parameters and configuration data of the trained first machine learning model in a centralized location accessible by the organization.
2 . The computer-implemented method of claim 1 , wherein the organization is related to clinical care.
3 . The computer-implemented method of claim 1 , further comprising determining, for each user, a user tenure value that represents an amount of time that the user has been actively associated with the organization and normalizing the set of data values based on the user tenure value for each user.
4 . The computer-implemented method of claim 1 , wherein obtaining, for each user, the set of data values associated with the user's interactions with the organization comprises:
detecting sensitive data values in source data; and
processing the source data to exclude the sensitive data values.
5 . The computer-implemented method of claim 4 , wherein processing the source data includes at least one of pseudonymization or de-sensitization.
6 . The computer-implemented method of claim 1 , wherein obtaining, for each user, a set of data values associated with the user's interactions with the organization comprises:
receiving raw data in a plurality of formats;
reformatting the raw data for storage in a data store; and
annotating the reformatted data with metadata.
7 . The computer-implemented method of claim 1 , wherein the first feature set is stored with an associated timeline for depreciation or deletion.