3 ms·

One surprising thing I learned by working at these large companies (I worked at Google for 2 years): User data older than ~90 days is essentially worthless. Oth

One surprising thing I learned by working at these large companies (I worked at Google for 2 years): User data older than ~90 days is essentially worthless. Other companies that I am aware of essentially do not use such data for anything except people management (teams with better historical metrics can hire more, etc).

DaiPlusPlus 6y ago
> User data older than ~90 days is essentially worthlesl Worthless, commercially - yes. But priceless and utterly invaluable for historical interest and academic research! I hope Google keeps their Ngram, search-stats and even geolocation history for everyone (anonymized and aggregated, of course) - it'll be an invaluable tool for researchers of all kinds to be able to see, for example, how our movement patterns varied from ~2015 when we started collecting that en-mass to some point in the future. Ditto Google Street View archives - Street View is now almost 13 years old and I'm looking forward to being able to see how things were in everyday life back in 2008 compared to however things are when I'm old and grey).
vl 6y ago
This is of course is incorrect. I worked at Google as well, and I can tell you that for text-based ML services 90 days often is not enough.
- mgraczyk 6y ago
  What I mean is that the data for a particular user is worthless after 90 days (really more like 30). Aggregated data is useful for building models, but individual user data is not useful for targeting that particular user. For example, suppose you are ranking search results for Alice. It may be beneficial to include old data in your language model or indexing models. However, it would likely not be beneficial to include Alice-specific features that go back more than 30-90 days. Such features include counters for her actions (clicks, queries, scroll), embeddings of her activity, etc. On the other hand, site and query specific features would likely be useful for years.
Viliam1234 6y ago
> User data older than ~90 days is essentially worthless. For an average user, sure. But suppose the user suddenly becomes e.g. an important politician. Now having their history older than 90 days is interesting. Hard disks are cheap, you can keep everything and wait for the ticket that wins the lottery. If you are a company like Google of Facebook, it is almost guaranteed that you will have juicy material about most people who become important in a few decades. In a few decades, any politician who says "anti-trust" will immediately have their porn history or edgy things they wrote as a teenager leaked to public.
- zelly 6y ago
  When has this ever happened?
- mgraczyk 6y ago
  If anybody at Google even attempted to do this, that employee would be fired.
scoot_718 6y ago
I'm sure they think that. Frankly I'm sick of youtube thinking that watching one video suddenly means I watch nothing else.

DaiPlusPlus 6y ago

> User data older than ~90 days is essentially worthlesl Worthless, commercially - yes. But priceless and utterly invaluable for historical interest and academic research! I hope Google keeps their Ngram, search-stats and even geolocation history for everyone (anonymized and aggregated, of course) - it'll be an invaluable tool for researchers of all kinds to be able to see, for example, how our movement patterns varied from ~2015 when we started collecting that en-mass to some point in the future. Ditto Google Street View archives - Street View is now almost 13 years old and I'm looking forward to being able to see how things were in everyday life back in 2008 compared to however things are when I'm old and grey).

vl 6y ago

This is of course is incorrect. I worked at Google as well, and I can tell you that for text-based ML services 90 days often is not enough.

mgraczyk 6y ago
What I mean is that the data for a particular user is worthless after 90 days (really more like 30). Aggregated data is useful for building models, but individual user data is not useful for targeting that particular user. For example, suppose you are ranking search results for Alice. It may be beneficial to include old data in your language model or indexing models. However, it would likely not be beneficial to include Alice-specific features that go back more than 30-90 days. Such features include counters for her actions (clicks, queries, scroll), embeddings of her activity, etc. On the other hand, site and query specific features would likely be useful for years.

Viliam1234 6y ago

> User data older than ~90 days is essentially worthless. For an average user, sure. But suppose the user suddenly becomes e.g. an important politician. Now having their history older than 90 days is interesting. Hard disks are cheap, you can keep everything and wait for the ticket that wins the lottery. If you are a company like Google of Facebook, it is almost guaranteed that you will have juicy material about most people who become important in a few decades. In a few decades, any politician who says "anti-trust" will immediately have their porn history or edgy things they wrote as a teenager leaked to public.

zelly 6y ago
When has this ever happened?
mgraczyk 6y ago
If anybody at Google even attempted to do this, that employee would be fired.

scoot_718 6y ago

I'm sure they think that. Frankly I'm sick of youtube thinking that watching one video suddenly means I watch nothing else.