Hacker News Reader

> In order to preserve the usefulness of those models and the hundreds of millions in R&D spent to develop them, you have to carefully understand exactly what products and algos will be affected by just deleting a single record.

My answer to that is 'who cares?'

It's not my problem, let me delete all my data.

In fact it's required by law in some places.

Exactly. The ability to delete user data should have been something engineered from the ground up. They are paying the penalty of technical debt and hopefully learning some tough lessons.

Taking the approach of a scientist, i.e. a data scientist, this is a hard problem to solve. Not only does Facebook have to implement the simple user.delete.all.data(), but think about the number of internal products that are machine learning based. In order to preserve the usefulness of those models and the hundreds of millions in R&D spent to develop them, you have to carefully understand exactly what products and algos will be affected by just deleting a single record.

This is not a science problem and it is not hard to solve: a scientist would delete the data and re-derive their models because they want the best data possible. This is how you "preserve the usefulness of those models," by ensuring that the data has integrity.

facebook has to figure out a way to understand exactly how thousands of records per user, being deleted, will affect it's models.

This is looking through the telescope backwards. You create your models with the data that you have, and if some data is removed then you regenerate.

You can bet your ass they groom their data after detecting stuff like new botnets and fake clicks, and individual user data does not have to be different.

That is an entire community and classification of people that facebook has to figure out how to reconcile the loss of data for, and fix its machine learning models

Again, they already know how to do this, and they do it all the time. There are no nuances to it outside of roadblocks installed by FB themselves.

It's 110% a business problem, and 110% Zuckerberg's choice to make it difficult, and to invent difficulties in bringing it about.

I don't think that's what this article is about.

I for one would be OK with a machine learning model that's been trained using my data. I wouldn't demand that it be "re-trained" just because I asked for my history to be deleted. I think most people, given an explanation of what a machine learning model is, would be fine with this.

The clear history thing is about personally identifiable information though. It's not having your information dangling around after you leave Facebook. It's having the option to delete your shadow profile.

Throughout all the hearings, FB execs have been very wishy-washy about how they handle your browsing history, particularly what they collect outside of FB properties.

People should continue to call them out on this until they get their act together.

The problem is that these lookalike models are constantly being re-trained to build an accurate profile that advertisers can then target. If you clear your history, the data that is used to train these models no longer exist. OC is saying FB needs time to adjust their models for a potential mass deletion of input data.

1. Anonymize the data in the machine learning training corpora.

2. That's it. Problem solved.

Question for extra credits:

These machine learning models you speak of, do they mostly do stuff that...

a) is in the interest of the users

b) goes against the interest of the users

Society does not have an obligation to tolerate negative externalities just to protect someone's business model. That is precisely what regulation is for.

Unfortunately, anonymizing the data is not what's being demanded. People want to completely delete their data, which presents a harder problem from a modelling perspective.

Also, worth noting that FB has 2 primary users: you and advertisers. The second class of users is very interested in having your data. Unfortunately, FB seems to favor them bc they are the user who's paying them.

> Unfortunately, anonymizing the data is not what's being demanded. People want to completely delete their data, which presents a harder problem from a modelling perspective.

99.9% of the people are not even aware of the distinction. This is an attempt to follow the letter of the law to create an excuse not to follow the spirit of the law. What people want is personal privacy and cognitive freedom. They don't care about implementation details.

> Also, worth noting that FB has 2 primary users: you and advertisers. The second class of users is very interested in having your data. Unfortunately, FB seems to favor them bc they are the user who's paying them.

You are right. Sorry for the typo. I did not mean "users", I meant the "used".

Supermarkets have two types of users: customers and other businesses that use supermarkets to distribute their products. It would be in the interest of those other businesses if, for example, food safety standards were more lax. But not in the interest of the customers. The food industry has a lot more money and power than the customers, so we solve this with regulation. It's kind of Civilization 101, really.

> Anonymize the data

In our world of Big Data, where it's easy to combine many disparate databases, effectively anonymizing data is also a very hard problem to solve.

> In order to preserve the usefulness of those models and the hundreds of millions in R&D spent to develop them, you have to carefully understand exactly what products and algos will be affected by just deleting a single record.

Indeed, that sounds like a hard problem.

So the question becomes "how much of the hundreds of millions in R&D are being allocated to solve the record-deletion problem?". I'd like to see spending reports, new research publications from Facebook (e. g. retroactive record-deletion, attempts to build models that are robust to deletion, attempts to build k-anonymity into models from the start).

> I know there are more nuances here, but I think there are some challenges Facebook has in implementing this, and business wise it's a drain on resources that will hurt them.

I can give Facebook the benefit of the doubt and suppose that they're trying to figure out how to transition smoothly without destroying their models. But I'd also like to see a hard timeline when they promise to roll out user.delete.all.data() anyways (business value be damned).

The real test of whether a company really "takes your privacy seriously" is whether they ever reach a point where privacy becomes a priority, i. e. whether there are decisions to do the right thing even when it could be bad for the bottom line (under the current business model).

It might be a difficult problem to solve, but there’s no evidence that Facebook is working on it or that they ever intended to work on it.

The easiest explanation is that Zuckerberg was lying. We could give him the benefit of the doubt but since he’s always lying it makes more sense to save time and assume he’s lying about this as well.

Based on your use of the word always to describe Zuckerberg as a liar, I infer that you are upset by FB, understandably. I would add that I am not a fan of FB and I am deactivated, but unless you work at FB, making comments on if they are working on it or not is just speculation. I would say a more probable speculation is >=1 employee is working on this tool.

> have to implement the simple user.delete.all.data()

That may not be as simple as you think. FB probably has thousands of data types and each one needs to define how a deletion should affect everything else. E.g. if I delete my comment, what happens to your reply? That one is obviously solved already since it's a core part of the product, but there are probably lots of pieces of data that were never designed to be deleted so will need effort to support.

Yep, totally agree. And, these lookalike models are the bread & butter that drive FB's advertising revenue, so it won't be taken lightly. My guess is that if this feature is ever launched, it'll be in a form that is necessarily not heavily adopted.

Facebook quiet literally figured out how to build a global social network connecting billions of people.

Given they solved that insanely hard problem pretty well, I’d suggest they could also solve this if they had sufficient motivation.

Oh no, not the poor neural networks!