Interactive adaptation of machine learning models for time series data

Abstract

A computer-implemented method can comprise establishing programmatic connections to a digitally stored first database comprising over one million records, each of the records comprising time series data organized as an event with a timestamp and an event data, the first database being part of a HADOOP cluster that is programmatically coupled to a HIVE data warehouse manager and a PRESTO query engine, reading a configuration file that specifies one or more tables in the first database and for each particular table among the one or more tables, forming and submitting a plurality of PRESTO queries to the first database, each of the PRESTO queries specifying one or more data aggregation operations, and in response thereto, receiving a set of aggregated records of the first database, training a machine learning model using a portion of the aggregated records as a training dataset, determining a plurality of outlier values that are represented in the set of aggregated records, one or more change points that are represented in the set of aggregated records, a plurality of seasonality patterns that are represented in the set of aggregated records, receiving, from a second computer, input specifying a plurality of hyper-parameters for the machine learning model, evaluating the aggregated records of the first database using the machine learning model to output forecasted event data as a second time series, generating presentation instructions which when rendered using a computer display device cause displaying one or more graphical visualizations in a graphical user interface of the computer display device, the graphical visualizations comprising a time series analysis, a graphical timeline of forecasted event data, and a performance summary of the machine learning model.