Three methods are generally viable for correcting errors in occupancy data:
- Correlation inference
- Reasonableness and manual intervention
- Accuracy correction for precise data
This method uses multiple proxies to attempt to make the same measurement, and then applying a specific rule or heuristic to determine the safest inference that can be made.
For example, a Reservation Management System and IPDD may both be used to measure Physical Occupancy of specific reservable workstations. If the Reservation Management System reports a reservation for a specific person, and that person’s laptop is also reported by IPDD as being docked during the same period as the reservation, then this complete correlation gives very high confidence in the inference that this person Physically Occupied this workstation during the period.
On the other hand, if the IPDD reported use of the reserved workstation until an hour after the reservation ended, and in the absence of any user initiated “extend reservation” event, then a rule may be applied that states that the continued IPDD presence events for this houe indicates continued Physical Occupancy, notwithstanding the reservation terminating.
Similarly, if both Reservation Management and IPDD reported presence at the same times, but at different workstations, it is likely that a rule would infer the actual Physical Occupancy Space from the IPDD data rather than the reservation data.
Note, however, that these examples are also only inferences with varying degrees of confidence – it isn’t possible to know from the data alone what the person was actually doing, as neither proxy tells us that.
Reasonableness and manual intervention
Once reasonableness test have been introduced, they are most commonly used to highlight suspect data and initiate a process of manual intervention to try to understand why the data failed the reasonableness test.
This might involve establishing the wider context of the anomaly. For example, a spike in building Physical Occupancy might be caused by an all-hands meeting, and therefore the data is correct, just unusual. See also the Unintended consequences story.
It is good practice, when a reasonableness test fails, and the cause is established, to review and if possible refine the reasonableness test to make it stronger. For example, if all-hands meetings were published on a system somewhere, this system could be automatically checked when occupancy peaks were detected to see if there was an all-hands meeting on that day, and the tolerances for what constitutes and “unreasonable” peak adjusted accordingly.
Accuracy correction for precise data
This third method of error correction is included to deal with proxies that have high precision (i.e. they always produce the same outputs for a given semantically equivalent input) but low accuracy.
To help picture such a system, imagine a crossbow with a scope and every time it is fired with the sight perfectly central on the bullseye, the arrow hits the target in the exact same spot 20 cm above the bullseye. This crossbow is inaccurate but very precise. Clearly, in this case, the scope would be adjusted so that it was both accurate and precise.
However, in the world of occupancy, the proxies that provide the data are often “adjusted” for some other primary purpose – secure access control, network access etc. – and so there is limited opportunity to re-calibrate them.
Instead, the outputs can be adjusted. In the crossbow example, this is a bit like moving the target after the arrow is fired. As long as the proxy is precise, a compensating adjustment can be used to adjust the output data to make it more accurate.
How accurate the results become depends on the reliability of the compensating adjustment. For example, a high precision ACS might only provide information about people entering the building who have been issued with a security badge of their own. If the Physical Occupancy of the building is required, then this cohort is only part of the total – there will also be visitors who either don’t have a badge at all (e.g. from another organization) or whose badge does not grant them access to this building (for example colleagues from another location). It may be possible to establish via a study that these additional non-badge holding occupants consistently number between 4% and 6% of the badge-holding occupants on any one day. With this information, the ACS data could simply be inflated by 5% and the new figure would now be bothe precise and accurate (assuming that the ACS was precise in the first place, which is a big assumption!)