Collect the right data

The idea

Good project data is relevant, enough, varied, recent and collected with consent.

More is better?

Meera says: The more data we collect, the better our AI tool will be. What do you think?

• She is right: more records make a better tool
• Only if the data fits the problem
• She is wrong: fewer records make a better tool

Which data would you collect?

Gurpreet's class is building a tool that predicts whether the school bus will arrive late. For each data idea, tap a group. Read the reason, then press Next card.

Five checks

The data that helped had a job to do. It matched the bus problem, covered rain and road works, and came from drivers who agreed.

The rest failed a check: too old, too little, a route missing, or private and taken without asking. Easy-to-find data failed too, because it did not fit.

Good project data is relevant, enough, varied, recent, and collected with consent.

People say Garbage In, Garbage Out: poor data gives a poor tool.

Notes

Good project data is relevant, enough, varied, recent, and collected with consent.

Before collecting, ask five questions: does it fit, is there enough, does it include everyone, is it recent, and did people agree? Collect only what the project needs.

Check yourself

A tool that reads road signs is taught only with photos taken in bright daylight. What is the likely result at night?

Ravi's team predicts vegetable prices at the local market, using records from twelve years ago only. What should they do?

To learn how fast the bus goes, a team could track every student's phone. What should they do?

• It will do just as well, since signs look the same. Signs look the same to us, but a dark photo looks very different to the tool. It was taught only from bright ones.
• It may do worse, because it saw no dark photos — correct. A kind of photo missing from the data is a kind the tool may handle badly. Adding night photos would help.
• It will report a fault and stop. A tool does not warn you by itself. It still gives an answer, which may be wrong, so people must test it on night photos.
• Keep using them, because real records stay true. The records are real, but markets change over the years. Very old records may mislead the tool about today's prices.
• Use fewer records so the tool runs quicker. Speed is not the problem. The problem is that the data does not describe today.
• Add recent prices, since markets change over time — correct. Recent records show today's market. Very old records may mislead.
• Use the bus's own GPS, since student locations are not needed — correct. The project only needs the bus. Data about people that a project does not need should not be collected at all.
• Track the phones, but only after every student and parent agrees. Agreement matters, but it is not enough here. The project does not need students' locations, so asking for them is still too much.
• Track the phones quietly, since it helps the project. A good aim does not excuse secret collection. People have a right to know, and to say no.