collate

https://img.shields.io/pypi/v/collate.svg https://travis-ci.org/dssg/collate.svg?branch=master Documentation Status Updates Code Coverage

Aggregated feature generation made easy.

Overview

Collate allows you to easily specify and execute statements like “find the number of restaurants in a given zip code that have had food safety violations within the past year.” The real power is that it allows you to vary both the spatial and temporal windows, choosing not just zip code and one year, but a range over multiple partitions and times. Specifying features is also easier and more efficient than writing raw sql. Collate will automatically generate and execute all the required SQL scripts to aggregate the data across many groups in an efficient manner. We mainly use the results as features in machine learning models.

Inputs

Take for example food inspections data from the City of Chicago. The table looks like this:

inspection_id license_no zip inspection_date results violations ...
1966765 80273 60636 2016-10-18 No Entry   ...
1966314 2092894 60640 2016-10-11 Pass …CORRECTED… ...
1966286 2215628 60661 2016-10-11 Pass w/ C… …HAZARDOUS… ...
1966220 2424039 60620 2016-10-07 Pass   ...

There are two spatial levels in the data: the specific restaurant (by its license number) and the zip code. And there is a date.

An example of an aggregate feature is the number of failed inspections. In raw SQL this could be calculated, for each restaurant, as so:

SELECT license_no, sum((results = 'Fail')::int) as failed_sum
FROM food_inspections GROUP BY license_no;

In collate, this aggregated column would be defined as:

Aggregate({"failed": "(results = 'Fail')::int"}, "sum", {'coltype':'aggregate', 'all': {'type': 'mean'}})

Note that the SQL query is split into two parts: the first argument to Aggregate is the computation to be performed and gives it a name (as a dictionary key), and the second argument is the reduction function to perform. The third argument provides a set of rules for how to handle imputation of null values in the resulting fields.

Splitting the SQL like this makes it easy to generate lots of composable features as the outer product of these two lists. For example, you may also be interested in the proportion of inspections that resulted in a failure in addition to the total number. This is easy to specify with the average value of the failed computation:

Aggregate({"failed": "(results = 'Fail')::int"}, ["sum","avg"], {'coltype':'aggregate', 'all': {'type': 'mean'}})

Aggregations in collate easily aggregate this single feature across different spatiotemporal groups, e.g.:

Aggregate({"failed": "(results = 'Fail')::int"}, ["sum","avg"], {'coltype':'aggregate', 'all': {'type': 'mean'}})
st = SpacetimeAggregation([fail],
                               from_obj='food_inspections',
                           groups=['license_no','zip'],
                           intervals={"license_no":["2 year", "3 year"], "zip": ["1 year"]},
                           dates=["2016-01-01", "2015-01-01"],
                           date_column="inspection_date",
                           state_table='all_restaurants',
                           state_group='license_no',
                           schema='test_collate')

The SpacetimeAggregation object encapsulates the FROM section of the query (in this case it’s simply the inspections table), as well as the GROUP BY columns. Not only will this create information about the individual restaurants (grouping by license_no), it also creates “neighborhood” columns that add information about the region in which the restaurant is operating (by grouping by zip). The state_table specified here should contain the comprehensive set of state_group entities and dates for which output should be generated for them, regardless if they exist in the from_obj.

Even more powerful is the sophisticated date range partitioning that the SpacetimeAggregation object provides. It will create multiple queries in order to create the summary statistics over the past 1, 2, or 3 years, looking back from either Jan 1, 2015 or Jan 1 2016. Executing this set of queries with:

st.execute(engine.connect()) # with a SQLAlchemy engine object

will create four new tables in the test_collate schema. The table food_inspections_license_no will contain four feature columns for each license that describe the total number and proportion of failures over the past two or three years, with a date column that states whether it was looking before 2016 or 2015. Similarly, a food_inspections_zip table will have two feature columns for every zip code in the database, looking at the total and average number of failures in that neighborhood over the year prior to the date in the date column. The food_inspections_aggregation table joins these results together to make it easier to look at both neighborhood and restaurant-level effects for any given restaurant. Finally, the food_inspections_aggregation_imputed table fills in null values using the imputation rules specified in the Aggregate constructor.

Imputation Rules

Imputation rules should be specified in the form of a dictionary:

{
    'coltype': 'aggregate',
    'all': {'type': 'mean'},
    'max': {'type': 'constant', 'value': 137}
}

The coltype key of this dictionary must be one of aggregate, categorical, or array_categorical and informs how the imputation rules are applied.

The other keys of the dictionary are the reduction functions used by the aggregate (such as sum, count, avg, etc.) or all as a catch-all. Function-specific rules will take precedence over the catch-all rule. The values associated with these keys are each a dictionary with a required type key specifying the rule type and other rule-specific keys.

Currently available imputation rules:
  • mean: The average value of the feature (for SpacetimeAggregation the mean is taken within-date).
  • constant: Fill with a constant value from a required value parameter.
  • zero: Fill with zero.
  • zero_noflag: Fill with zero without generating an “imputed” flag. This option should be used only for cases where null values are explicitly known to be zero such as absence of an entity from an events table indicating that no such event has occurred.
  • null_category: Only available for categorical features. Just flag null values with the null category column.
  • binary_mode: Only available for aggregate column types. Takes the modal value for a binary feature.
  • error: Raise an exception if any null values are encountered for this feature.

Outputs

The main output of a collate aggregation is a database table with all of the aggregated features joined to a list of entities.

TODO: sample rows from the above aggregation.

Usage Examples

Multiple quantities

TODO

Multiple functions

TODO

Tuple quantity

TODO

Date substitution

TODO

Categorical counts

TODO

Naming of features

TODO

More complicated from_obj

TODO

Technical details