13 min
Rules or examples? How a filter learns
The game's chat is full of "free skins" scams. Write a rule to stop them, then find out why the filters that really work learn from examples instead.
Ticket 01 · The guess
The filter that learned
Words to know
Sources and every number (2)
Sources
- Ridding Gmail of 100 million more spam messages with TensorFlow (Neil Kumaran, February 6, 2019) · Google Workspace blog, 2019
- A Plan for Spam (August 2002) · Paul Graham, 2002
Every number
- 99.9%
google.com ·2019 · Share of spam, phishing and malware Gmail says its protections block (Google, February 2019); "more than". - 100 million messages a day
google.com ·2019 · Extra spam Gmail blocked each day after adding new machine-learning protections built with TensorFlow, about, February 2019. - 200 messages
(example number: Simplified example: messages the moderators labeled.) · Simplified example: messages the moderators labeled. - 40 messages
(example number: Simplified example: labeled messages with a link from an account less than a week old.) · Simplified example: labeled messages with a link from an account less than a week old. - 34 messages
(example number: Simplified example: of those, the ones labeled scam.) · Simplified example: of those, the ones labeled scam. - 85%
(example number: Worked out: 34 messages ÷ 40 messages.) · Worked out: 34 messages ÷ 40 messages. - 160 messages
(example number: Worked out: 200 messages − 40 messages.) · Worked out: 200 messages − 40 messages. - 4 messages
(example number: Simplified example: scams among the other labeled messages.) · Simplified example: scams among the other labeled messages. - 2.5%
(example number: Worked out: 4 messages ÷ 160 messages.) · Worked out: 4 messages ÷ 160 messages. - 17%
(example number: Worked out, the wrong way: 34 messages ÷ 200 messages.) · Worked out, the wrong way: 34 messages ÷ 200 messages.