Predictive Scoring provides Feature Guess functionality.

Treasure Data suggests highly relevant features (predictors) derived from attributes:

The Feature Guess function:

  1. Sends a request to API server

  2. Samples 100,000 values from every attribute in every profile in a master segment

  3. Applies pre-defined set of rules for each attribute

Attributes detected as meaningless are automatically dropped, and pairs of a type (quantitative, categorical, or categorical array) and preprocessing rules (such as, how to fill in missing values, extract day of week from the timestamp) for potentially informative attributes are suggested in the input boxes.

The pre-defined set of guessing rules are based on:

  • attribute type (number, string, array)

  • attribute name

  • cardinality

  • regular expression

  • mean

  • standard deviation

  • percentile

of the sampled values.

  • No labels

1 Comment

  1. This page can be move just after create will be useful understand the flow when creating.

    There is exiting Bug[even in V4] Which is good to mention if we can . The bug is if we add any value in Categorical array Field which is not of type Array then PSM save will pass but train will fail. When we have any this kind of PSM in our entire PS that PS will end up into erroneous . That is if we try to train any existing PSM or try to Create any new PSM and Train – train will always fail Because the PS contains an PSM which is giving trouble. 

    We don't Have defect for this since its know behavior from v4 and its known to all stalk-holders so far .

    [This should go to train Doc]One more issue is good to mention - When the train fails we don't show the failed state as of now in UI , we show empty state to Train the model . This will be handled Post M1