Hacker News Reader

Comment by Red_Leaves_Flyy | Hacker News Reader

>I don't know how to make these different use cases clear without having pages and pages of explanation that nobody will read.

Expose the logic. Put a "?" Icon button on the add that lists every data point used in the decision making process, who collected the data, when, and how.

E.g:

1.1.2020 @ Starbucks on main st in 90210 you bought a medium coffee, a gift card, and a cookie.

1.3.2020 @ Barnes and noble you bought a book and m&MS.

So when the customer wonders why they're being shown advertisements for cat litter they'll be able to see how the ml algorithm thought they'd be interested.

If the logic can't be explained in real time to the advertised too then it surely can't be justified to the actual companies buying said advertisement.

View on HN

But that's not how sophisticated machine learning algorithms work. Without knowing specifics, it gets us in the right ballpark to imagine they ingest all that kind of information and build an opaque representation of the user/customer, which allows various similarities to be computed between users and ads, regarding how likely they are to be receptive or how likely they will click. There's no short list of events you did. It can even contain stuff like how quickly you move your mouse cursor, whether you use the scroll wheel or the scroll bar, or who knows what else.

The ML algorithms are the ultimate "source: dude trust me"

If all those metrics are being used to sell advertisements against me then I have every right to know each metric, all the details behind every single piece of data collected, from where when and whom, as well as the right to know to know how every data point and metric and is used to sell every specific advertisement to me. That means everything I just enumerated and more must be retained every time a query is run against my data or auction is ran to advertise to me.

What's that about a short list? Sorry, I may not have been clear, if you're using 1 data point or 1x10e999 I have the right to know every single one. I have the right to know the inferences drawn about me and how advertisers can target me. If companies aren't retiring data they are reckless and auditable records for every customer should help to rectify that imbalance. Likewise, if a model cannot be explained, proven mathematically, and reproduced then it's bad computer science. I'm sorry if this seems offensive but I'm sick of the whole move fast and break things when innocent people keep getting hurt or companies fleeced of their advertising dollars.

Either this stuff can be explained in reverse (so consumers know what privacy they've sacrificed, and advertisers know they're not being fleeced) or it can't and there's sketchy or bad computer science at work.

Even if that were possible it wouldn’t really explain why you got this ad vs some other one, so the explanation would fall flat. That is just “how” you got it not “why” you got it.

You're missing my point.

Many ads are sold in real time auctions. Every metric exposed or used in that process should be accounted for and auditable by the advertised too. Let the user, or journalists, or researchers figure out the why from the corpus.

Reply on news.ycombinator.com