Grants-sorting tool opens up new insights for funders

Following five years of development and testing by data scientists and developers, CLASSIEfier is now integrated into Australia’s leading grants administration software, SmartyGrants, which channels $7 billion in grants each year.

The system is free for SmartyGrants users, with the data scientists behind the new system confident it will generate valuable insights.

The algorithmic system works by swiftly reading grant applications to automatically assign labels defined by CLASSIE, the social sector "dictionary”. CLASSIEfier takes its name from that dictionary.

Developed by SmartyGrants’ parent company, Our Community, the CLASSIE taxonomy classifies the social sector by organisation, subject, population, activity, and transaction type.

Already widely used, CLASSIE has been adopted by the Australian Charities and Not-for-Profit Commission’s online charity search portal.

CLASSIE has been available to SmartyGrants users for several years, but until now users were required to assign

CLASSIE labels manually when filling in, assessing or managing forms.

CLASSIEfier removes the need for human classification with its keyword-matching algorithm.

New tool gives more power to grantmakers

Our Community chief data scientist Dr Paola Oliva-Altamirano said grantmakers are able to use CLASSIEfier to classify all current and past records, or could classify specific programs or rounds.

Results are available instantly via labels applied to each classified application form, while collective results are shown on the grantmaker’s dashboard, she said.

SmartyGrants has also released a suite of Excel templates that grantmakers can upload into SmartyGrants to create aggregated views:

• Applications by subject
• Applications by beneficiary
• Total funding by subject
• Total funding by beneficiary
• Funding split by subject
• Funding split by beneficiary.

Dr Oliva-Altamirano said the system has been set up to give grantmakers extensive control, including deciding what sections of CLASSIE to use, and how granular results should be.

“We put the power in the hands of the grantmaker because only they know what view of the data they need,” Dr Oliva-Altamirano said.

“For example, you may be a general grantmaker who is interested in applying the entire subject classification to see how much money you’re putting towards the arts versus sports versus the environment.

“Or you might be an arts-only grantmaker, who only wants to apply the arts portion of the taxonomy, but you want a lot of granularity – you want to know what proportion of your budget went to ballet versus contemporary dance versus hip hop.

“Context is everything. Only you know what view will be most meaningful for the people who view your reports (and maybe that view will change on a daily basis, depending on the audience for the results). We’ve turned control over to you. You can reclassify your data however you like and as many times as you like.”

Users can also decide how many labels to apply.

“Some grantmakers will want to see only one label for subject and one for beneficiaries per application,” Dr Oliva-Altamirano said. “That allows you to produce a nice pie chart; however, forcing CLASSIEfier to make just one selection makes it really hard to get an accurate result. The algorithm can’t know what is the “right” result or the most important result for you.

“Take the example of a project to provide ballet classes to women in their 40s. For some people the gender of the beneficiaries will be the most important label; for others, the age group is most important. If you ask CLASSIEfier to return only one result it has to pick one.

“For this reason, we recommend you avoid applying single classifications. You may be able to produce a nice pie chart but it’s not telling you the full story.”

Making the most of the CLASSIEfier system

“In testing, we have found that level two is the sweet spot for the subject classification – that will give you a complete set of classifications that are not too specific. For beneficiaries, the sweet spot is level one because it’s a top-heavy classification so level one already gives you a lot of detail. But it will depend on your own needs. We suggest you try it out and change the filters to see the different results that come out so you can assess what are the right settings for your program.”

Dr Oliva-Altamirano said the data science team was aware of the ethical dimensions of using algorithms, and the tool’s design reflected that.

Initially, for example the data science team investigated using a self-improving machine-learning algorithm; however, ethical concerns arose when the machine threw up some results that the team found concerning.

“The problem with machine-learning algorithms is that they can learn from and then reinforce the biases that exist in the world,” says Dr Oliva-Altamirano. “We were seeing some results coming through that reinforced harmful racial stereotypes, but it was very difficult to control the bot.”

She said the team switched to using a keyword-matching algorithm in part because it allowed the data science team to correct for biases.

“We rely on our users to tell us when they see something unfair in the results so we can keep improving the algorithm for future users.”