Aggregate features for machine learning

A method comprising:

• reading a new record from an input source;
• accessing a plurality of aggregate definitions, each aggregate definition of the plurality of aggregate definitions specifying:
• a key,
• an output store indicating a location for aggregate feature records generated from the aggregate definition,
• a feature set indicating one or more information fields from the input source,
• a half-life value set indicating one or more time periods, and
• a set of aggregate operations to apply to a cross of the feature set and half-life value set to generate a set of aggregate metrics;
• determining, based on the accessing, a respective input key value for each of the plurality of aggregate definitions, wherein the respective input key value is determined based on the key for the aggregate definition and a corresponding value from the new record;
• identifying first aggregate feature records in output stores associated with the plurality of aggregate definitions, each first aggregate feature record having a value for the key that matches the respective input key value for the aggregate definition and wherein each first aggregate feature record includes one or more aggregate metrics, the one or more aggregate metrics representing a cross of the feature set, the half-life value set, and the set of aggregate operations specified in the aggregate definition;
• for each of the first aggregate feature records, updating the one or more aggregate metrics using the new record and a half-life decay applied to the one or more aggregate metrics;
• and
• using the first aggregate feature records as input to a neural network.

A method comprising:

• receiving a query parameter;
• determining, from a plurality of aggregate definitions, responsive aggregate definitions that are responsive to the query parameter, each aggregate definition of the plurality of aggregate definitions specifying:
• a key,
• an output store indicating a location for aggregate feature records generated from the aggregate definition,
• a feature set indicating one or more information fields from an input source,
• a half-life value set indicating one or more time periods, and
• a set of aggregate operations identifying one or more aggregate operators to apply to a cross of the feature set and half-life value set;
• obtaining responsive aggregate feature records that are responsive to the query parameter from the output stores specified in the responsive aggregate definitions;
• applying a half-life decay to each responsive aggregate feature record, the half-life decay for a particular aggregate feature record being based on a timestamp associated with the particular aggregate feature record and a timestamp for the query;
• and
• providing the responsive aggregate feature records as input to a neural network,
• wherein at least one aggregate definition includes two time periods in the half-life value set and each time period is applied to each responsive aggregate feature record as part of applying the half-life decay.