Formula for measuring unfairness

Fri 3 February 2017

Geoffrey De Smet

OptaPlanner lead

Load balancing is a common constraint for many OptaPlanner use cases. Especially when scheduling employees, the workload needs to be spread out fairly. Easier said than done. Let’s take a look at this challenging problem.

For example, let’s assign 15 undesirable tasks to 5 employees. Each task takes a day, but they have different skill requirements.

What is fair?

In a perfectly fair schedule, each employee will get 3 tasks, because the mean (= average) is 15 / 5 = 3. Or explained more formally, for the math geeks:

Mean (= average): \(\overline{x} = \frac{\sum_{i=1}^{n} x_i}{n}\)

So this schedule is perfectly fair:

	Ann	Beth	Carl	Dan	Ed	Quality	Mean
	Number of tasks per employee					Quality	Measurement of fairness
Schedule A	3	3	3	3	3	Fair	3

Problem solved? Unfortunately not, because there are other constraints. For example, there are 7 tasks that require a skill which only Ann and Beth posses. One of them will have to do at least 4 tasks.

Our schedule will not be perfectly fair, but how can we make it as fair as possible?

What is fairer?

Let’s look at two opposing definitions of fairness:

A schedule is fair if most users think it is fair to them.
A schedule is fair if the employee with most tasks has as little tasks as possible.

Obviously, since we want to treat all employees equally, the second definition is correct. Besides, if we make almost everyone happy by letting Ann do all the work, she would probably quit on us. That wouldn’t help.

	Ann	Beth	Carl	Dan	Ed	Quality
	Number of tasks per employee					Quality
Schedule	15	0	0	0	0	Unfair. Ann quits her job.

Sorting solutions by fairness

Let’s look at a few different solutions of the same dataset. Each one has 15 undesirable tasks:

	Ann	Beth	Carl	Dan	Ed	Quality
	Number of tasks per employee					Quality
Schedule A	3	3	3	3	3	Fairest
Schedule B	4	4	3	2	2
Schedule C	5	3	3	2	2
Schedule D	5	5	2	2	1
Schedule E	6	3	3	2	1
Schedule F	6	5	2	1	1
Schedule G	11	1	1	1	1	Unfairest

These are sorted from fair to unfair. Several of these are probably infeasible, due to hard constraints such as skill requirements.

Ann has most tasks each time. How do we compare 2 schedules in which Ann has the same number of tasks?

	Ann	Beth	Carl	Dan	Ed	Quality
	Number of tasks per employee					Quality
Schedule C	5	3	3	2	2	More fair
Schedule D	5	5	2	2	1	Less fair

In that case, we look at the second unfairest treated employee: Beth in this case. And we minimize her tasks.

Now that we’ve defined fairness, how do we implement it?

Measuring unfairness

Ideally, we want to calculate a number for each schedule to measures the fairness of that solution. The lower, the fairer. How do we calculate that number? Let’s look at some formula’s:

Deviation from the mean

Because a perfectly fair schedule has all employees working the average number of tasks, what if we simply sum the difference with the mean per employee?

Absolute deviation from the mean: \(f(n) = \sum_{i=1}^{n} |x_i - \overline{x}|\)
Average deviation from the mean: \(f(n) = \frac{\sum_{i=1}^{n} |x_i - \overline{x}|}{n}\)

	Ann	Beth	Carl	Dan	Ed	Absolute deviation	Average deviation
	Number of tasks per employee					Measurement of fairness
Schedule A	3	3	3	3	3	0	0.00
Schedule B	4	4	3	2	2	4 - Bad	0.80 - Bad
Schedule C	5	3	3	2	2	4	0.80
Schedule D	5	5	2	2	1	8 - Very bad	1.60 - Very bad
Schedule E	6	3	3	2	1	6	1.20
Schedule F	6	5	2	1	1	10	2.00
Schedule G	11	1	1	1	1	16	3.20

These measurements are terrible. Both claim that schedule B and C are equally fair. They aren’t. Both claim that schedule D is worse than E. It’s not: just ask Ann.

Variance and standard deviation

In statistics, variance and standard deviation are used to penalize outliers more. That sounds like exactly what we need:

Variance: \(f(n) = \frac{\sum_{i=1}^{n} (x_i - \overline{x})^2}{n}\)
Standard deviation: \(f(n) = \sqrt{\frac{\sum_{i=1}^{n} (x_i - \overline{x})^2}{n}}\)
Squared deviation from the mean (= variance * n): \(f(n) = \sum_{i=1}^{n} (x_i - \overline{x})^2\)
Root of squared deviation from the mean: \(f(n) = \sqrt{\sum_{i=1}^{n} (x_i - \overline{x})^2}\)

	Ann	Beth	Carl	Dan	Ed	Variance	Standard deviation	Squared deviation	Root squared deviation
	Number of tasks per employee					Measurement of fairness
Schedule A	3	3	3	3	3	0.00	0.00	0	0.00
Schedule B	4	4	3	2	2	0.80	0.89	4	2.00
Schedule C	5	3	3	2	2	1.20	1.10	6	2.45
Schedule D	5	5	2	2	1	2.80 - Bad	1.67 - Bad	14 - Bad	3.74 - Bad
Schedule E	6	3	3	2	1	2.80	1.67	14	3.74
Schedule F	6	5	2	1	1	4.40	2.10	22	4.69
Schedule G	11	1	1	1	1	16.00	4.00	80	8.94

These measurements are good, but still not ideal. They claim that schedule D and E are equally fair. They aren’t.

Maximum

What if we simply take the maximum of each row?

Maximum: \(f(n) = \underset{0 < i \leq n}\max x_i\)

	Ann	Beth	Carl	Dan	Ed	Maximum
	Number of tasks per employee					Measurement of fairness
Schedule A	3	3	3	3	3	3
Schedule B	4	4	3	2	2	4
Schedule C	5	3	3	2	2	5 - Bad
Schedule D	5	5	2	2	1	5
Schedule E	6	3	3	2	1	6 - Bad
Schedule F	6	5	2	1	1	6
Schedule G	11	1	1	1	1	11

That’s worse than variance: it only looks at one employee. Furthermore, it completely discards fairness between the remaining employees. That might be ok if there’s one employee, but not if there are thousands.

List of maximums

What if we don’t use any formula but just store the list of numbers sorted by decreasing size?

	Ann	Beth	Carl	Dan	Ed	List of maximums
	Number of tasks per employee					Measurement of fairness
Schedule A	3	3	3	3	3	[3,3,3,3,3]
Schedule B	4	4	3	2	2	[4,4,3,2,2]
Schedule C	5	3	3	2	2	[5,3,3,2,2]
Schedule D	5	5	2	2	1	[5,5,2,2,1]
Schedule E	6	3	3	2	1	[6,3,3,2,1]
Schedule F	6	5	2	1	1	[6,5,2,1,1]
Schedule G	11	1	1	1	1	[11,1,1,1,1]

That will compare perfectly. In OptaPlanner it can be implemented by adding 5 score levels for this dataset.

However, besides obvious memory consumption issues when scaling to thousands of employees, this isn’t compatible with other soft constraints…

No constraint is an island

Fairness is typically a soft constraint. But there are other soft constraints that we’ll need to optimize for too, so we’ll need to weight them against each other.

An example

For example, presume there’s a soft constraint on priority violations, that’s 10 times as important as a fairness violation. Let’s also add a schedule with 1500 tasks, to see how it scales out:

	Number of tasks	Priority violations	Fairness violations	Soft score	Quality
Schedule F	15	1	?	?	Best
Schedule C	15	2	?	?
Schedule D	15	2	?	?	Worst
Schedule X	1500	100	?	?	Different dataset

Number of tasks

Priority violations

Fairness violations

Soft score

Quality

Schedule F

Best

Schedule C

Schedule D

Worst

Schedule X

1500

100

Different dataset

To calculate the soft score, we sum the fairness violations with 5 times the priority violations and make that negative.

Let’s start the process of elimination…

Represented by a single number

List of maximums isn’t represented as single number (because it uses multiple score levels), so it’s difficult to mix in priority violations:

Number of tasks

Priority violations

Measurement of fairness

Soft score

List of maximums

Schedule F

[6,5,2,1,1]

ERROR?

Schedule C

[5,3,3,2,2]

ERROR?

Schedule D

[5,5,2,2,1]

ERROR?

Schedule X

1500

100

[8,8,7,7,7,7,…]

ERROR?

Grow with the number of violations

If we scale out to 1500 employees, we notice that maximum gets dwarfed by the priority violations:

	Number of tasks	Priority violations	Maximum	Soft score
	Number of tasks	Priority violations	Measurement of fairness	Soft score
Schedule F	15	1	6	16
Schedule C	15	2	5	25
Schedule D	15	2	5	25
Schedule X	1500	100	8	1008 - Dwarfed

Similarly, average deviation from the mean, variance and standard deviation get dwarfed on bigger datasets too:

	Number of tasks	Priority violations	Average deviation	Variance	Standard deviation
	Number of tasks	Priority violations	Measurement of fairness
Schedule F	15	1	2.00	4.40	2.10
Schedule C	15	2	0.80	1.20	1.10
Schedule D	15	2	1.60	2.80	1.67
Schedule X	1500	100	1.50 - Dwarfed	2.50 - Dwarfed	1.58 - Dwarfed

As the number of fairness violations grow, so should the fairness measurement.

Do not grow exponentially with the number of violations

On the other hand, as the dataset grows, the fairness violations shouldn’t dwarf the other violations either. Squared deviation does that:

	Number of tasks	Priority violations	Squared deviation
	Number of tasks	Priority violations	Measurement of fairness
Schedule F	15	1	80
Schedule C	15	2	6
Schedule D	15	2	14
Schedule X	1500	100 - Dwarfed	10201

Conclusion

That just leaves:

absolute deviation from the mean which compares terribly for fairness
root squared deviation which isn’t perfect but works well enough

So the recommended approach is:

Root of squared deviation from the mean: \(f(n) = \sqrt{\sum_{i=1}^{n} (x_i - \overline{x})^2}\)

	Number of tasks	Priority violations	Root squared deviation	Soft score
	Number of tasks	Priority violations	Measurement of fairness	Soft score
Schedule F	15	1	4.69	14.69
Schedule C	15	2	2.45	22.45
Schedule D	15	2	3.74	23.74
Schedule X	1500	100	101.00	1101.00

Additional notes

Part-time employees

To deal with unequal employees, for example if some employees work half as many hours as other employees, multiply their number of tasks by the inverse of their FTE (full time equivalent) before feeding into this formula.

Other reasons to treat employees unequally (such as disabilities or talent retention) can be handled in a similar fashion or with separate constraints, depending on the requirement.

Update

There is a formula that seems to compare perfectly, see this mathexchange answer, but it might suffer from numerical instability and overflow.

Permalink

Formula for measuring unfairness

Geoffrey De Smet

What is fair?

What is fairer?

Sorting solutions by fairness

Measuring unfairness

Deviation from the mean

Variance and standard deviation

Maximum

List of maximums

No constraint is an island

An example

Represented by a single number

Grow with the number of violations

Do not grow exponentially with the number of violations

Conclusion

Additional notes

Part-time employees

Update

Comments