ValueBasedPrioritization

Article

Academic article: https://github.com/freeradical13/ValueBasedPrioritization/raw/master/value_based_prioritization.pdf

Installation

https://pypi.org/project/vbp/

pip3 install vbp

vbp

Value Based Prioritization (vbp) uses value theory to quantitatively prioritize potential actions to accomplish a goal.

The vbp.run module may be used from the command line to perform different VBP actions such as listing actions (list), counting actions (count), predicting values (predict), running Modeled VBP (modeled_value_based_prioritization), and more. For usage, run:

python3 -m vbp.run

Any non-terminal output goes to the vbpoutput sub-folder.

Here is a simple example counting the number of groupings of underlying causes of death for the United States for the default data type:

python3 -m vbp.run count UCODUnitedStates

Alternatively, if installed through pip, a vbp script wrapper may be used:

vbp count UCODUnitedStates

The academic article above includes footnotes with details on how to run vbp to produce the output of each step.

This package provides abstract classes and utility methods to run VBP, mostly focused on Modeled VBP which uses time series data to predict future values and prioritize actions based on the relative predicted values.

The vbp.DataSource class is the base abstract class for VBP.

The vbp.TimeSeriesDataSource abstract class inherits from vbp.DataSource and may be used for Modeled VBP. The vbp.ExampleDataSource class demonstrates a simple data source based on vbp.TimeSeriesDataSource.

Built-in Modeled VBPs include Underlying Cause of Death models for the United States (vbp.ucod.united_states.UCODUnitedStates) and the World (vbp.ucod.world.UCODWorld). These data sources both inherit from vbp.ucod.icd.ICDDataSource which inherits from vbp.TimeSeriesDataSource.

Running

The model type is specified with --ets, --ols, and/or --prophet. These are not mutually exclusive; if combined during modeled_value_based_prioritization, an average is taken of the results. The default is --ets.

By default, action names are obfuscated to reduce bias during model building and testing. Specify --do-not-obfuscate to show actual names.

Some data sources have different data types (e.g. mutually exclusive groupings of data). Add the -a argument before the data source name to run for all data types. Add the --data-type X argument after the data source name to specify a specific data type.

In general, a list of actions may be specified to run for just that list; otherwise, without such a list, all actions are processed.

Examples:

python3 -m vbp.run modeled_value_based_prioritization UCODUnitedStates

python3 -m vbp.run modeled_value_based_prioritization UCODUnitedStates --do-not-obfuscate "Heart disease" Cancer

Exponential Smoothing

Using exponential smoothing:

python3 -m vbp.run modeled_value_based_prioritization ${DATA_SOURCE} --ets

Specify --ets-no-multiplicative-models to only use additive models.

Specify --ets-no-additive-models to only use multiplicative models.

Linear Regression

Using linear regression.

python3 -m vbp.run modeled_value_based_prioritization ${DATA_SOURCE} --ols

Specify --ols-max-degrees X to model higher degrees.

Prophet

Using Facebook Prophet.

python3 -m vbp.run modeled_value_based_prioritization ${DATA_SOURCE} --prophet

United States

As of 2019-03-01, the unzipped U.S. mortality data consumes ~36GB of disk. It will be downloaded and unzipped automatically when a function is used that needs it.

Long-term, comparable, leading causes of death

Generate data/ucod/united_states/comparable_data_since_1959.xlsx for all long-term, comparable, leading causes of death in https://www.cdc.gov/nchs/data/dvs/lead1900_98.pdf:

python3 -m vbp.run prepare_data UCODUnitedStates

Rows 1900:1957 and the sheet Comparability Ratios in data/ucod/united_states/comparable_ucod_estimates.xlsx were manually input from https://www.cdc.gov/nchs/data/dvs/lead1900_98.pdf

Open comparable_data_since_1959.xlsx and copy rows 1959:Present.

Open comparable_ucod_estimates.xlsx and paste on top starting at 1959.

Process comparable_ucod_estimates.xlsx with its Comparability Ratios sheet to generate comparable_ucod_estimates_ratios_applied.xlsx:

python3 -m vbp.run prepare_data UCODUnitedStates --comparable-ratios

Final output:

python3 -m vbp.run modeled_value_based_prioritization UCODUnitedStates --data-type UCOD_LONGTERM_COMPARABLE_LEADING

World

As of 2019-03-01, the unzipped World mortality data consumes ~320MB of disk. It will be downloaded and unzipped automatically when a function is used that needs it.

When testing, writing data spreadsheets takes a lot of time and may be avoided with --do-not-write-spreadsheets.

Creating a new Data Source

Review vbp/example.py for a simple example. The basic process is:

1. Create a sub-class of vbp.DataSource in somename.py
2. Implement all @abc.abstractmethod methods and override any other superclass methods as needed.
3. Import somename.py at the top of vbp/run.py

Development

Prerequisites:

pip3 install numpy pandas matplotlib statsmodels scipy fbprophet

Updating PyPI package:

# Edit version in setup.py and __init__.py
python3 setup.py bdist bdist_wheel
python3 -m twine upload --skip-existing dist/*
# https://pypi.org/project/vbp/