A spam filter is a single neuron with a sigmoid on its output:
Here is a suspicious-words score and is a known-sender score, each between and . An email is flagged as spam when and let through otherwise.
The filter in production is P. A retraining run has produced two candidates to replace it, Q and R:
| filter | |||
|---|---|---|---|
| P (production) | |||
| Q | |||
| R |
On today's four emails, the three filters output these values of :
| P | Q | R | |||
|---|---|---|---|---|---|
| A | |||||
| B | |||||
| C | |||||
| D |
All three flag A and C as spam and let B and D through.
The team cares about one thing only. The replacement must make the same spam / not-spam call as P on every email it will ever see, whatever that email's two scores are.
Which of the candidates is guaranteed to do that?
Select all that apply.