tomwagg / difi

Did I Find It?

difi

Did I Find It?

About

difi is a simple package that takes pre-formatted linkage information from software such as MOPS, pytrax, or THOR and analyzes which objects have been found given a set of known labels (or truths). A key performance criteria is that difi needs to be fast by avoiding Python for loops and instead uses clever pandas.DataFrame manipulation.

Installation

The following installation paths are available:
Anaconda
PyPi
Docker
Source

Anaconda

difi can be downloaded directly from anaconda:
conda install -c moeyensj difi

Or, if preferred, installed into its own environment via:
conda create -n difi_py38 -c moeyensj difi python=3.8

PyPi

difi is also available from the Python package index:
pip install difi

Docker

A Docker container with the latest version of the code can be pulled using:
docker pull moeyensj/difi:latest

To run the container:
docker run -it moeyensj/difi:latest

The difi code is installed the /projects directory, and is by default also installed in the container's Python installation.

Source

Clone this repository using either ssh or https. Once cloned and downloaded, cd into the repository.

To install difi in its own conda enviroment please do the following:
conda create -n difi_py38 -c defaults -c conda-forge --file requirements.txt python=3.8

Or, to install difi in a pre-existing conda environment called difi_py38:
conda activate difi_py38
conda install -c defaults -c conda-forge --file requirements.txt

Or, to install pre-requisite software using pip:
pip install -r requirements.txt

Once pre-requisites have been installed using either one of the three options above, then:
python setup.py install

Or, if you would like to make an editable install then:
python setup.py develop

You should now be able to start Python and import difi.

Example

The example below can be found in greater detail in this Jupyter Notebook.

Assumed Inputs

difi is designed to analyze a set of linkages made by external software where some of the underlying true linkages are known. It needs just two DataFrames of data:

• 1. a DataFrame containing observations, with a column for observation ID and a column for the underlying truth (don't worry! -- difi can handle false positives and unknown truths as well)

• 2. a DataFrame describing the linkages that were found in the observations by the external software. This DataFrame needs just two columns, one with the linkage ID and the other with the observation IDs that form that linkage

What Can I Find?

In most cases the user can determine what known truths in their observations dataframe can be found by their respective linking algorithm. difi has two simple findability metrics:

The 'min_obs' metric: any object with this many or more observations is considered findable.

The 'nightly_linkages' metric: any object with this many or more observations is considered findable.

Which objects are findable?

What observations made each object findable?

A summary of what kinds of objects are findable might be useful.

Did I Find It?

Now lets see what the external linking software did find.

difi assumes there to be three different types of linkages:

• 'pure': all observations in a linkage belong to a unique truth
• 'partial': up to a certain percentage of non-unique thruths are allowed so long as one truth has at least the minimum required number of unique observations
• 'mixed': a linkage containing different observations belonging to different truths, we avoid using the word 'false' for these linkages as they may contain unknown truths depending on the use case. We leave interpretation up to the user.

Thanks to the power of pandas it can be super easy to isolate the different linkage types and analyze them separately. Selecting 'pure' linkages:

Selecting 'partial' linkages:

Selecting 'mixed' linkages:

Understanding the specifics behind each linkage is one thing, but how did the linking algorithm perform on an object by object basis.

Tutorial

A detailed tutorial on difi functionality can be found here.