Is Google obsessed with algorithms?

[PraedSt]

Many of our scientists would probably disagree, but from the experience of my field (economics/finance), there is a limit to the use of mathematics and physics in predicting human behaviour. It may be a fuzzy limit, and it may only exist due the state of our incomplete knowledge, but right now- it's there.

Also, irrespective of limits, the application of common sense, however you care to define it, often leads to a 'better' answer, or the same answer, faster and cheaper.

I came across this today:
Google searches for staffing answers.

Concerned a brain drain could hurt its long-term ability to compete, Google Inc. is tackling the problem with its typical tool: an algorithm.
...
The Internet search giant recently began crunching data from employee reviews and promotion and pay histories in a mathematical formula Google says can identify which of its 20,000 employees are most likely to quit.
...
Google's algorithm helps the company "get inside people's heads even before they know they might leave," said Laszlo Bock, who runs human resources for the company.

Note that this is an HR initiative, and then note that this is what simply asking people gets you:

Current and former Googlers said the company is losing talent because some employees feel they can't make the same impact as the company matures. Several said Google provides little formal career planning, and some found the company's human-resources programs too impersonal.

[Argos]

Many of our scientists would probably disagree, but from the experience of my field (economics/finance), there is a limit to the use of mathematics and physics in predicting human behaviour.

Dunno... OK, that would be true for single individuals and small groups. But It seems that the behaviour of large groups can be approached very nicely by maths.

[nauthiz]

Google's entire empire is built on the phenomenon that statistical techniques that have approximately zero predictive power when applied to small data sets can often prove to be surprisingly effective when applied to huge data sets.

To take something that's a bit closer to home for me, computational linguistics, the current gold-standard algorithm for guessing the parts of speech for unknown words in a sentence seems kind of silly at first - it doesn't even involve the program knowing anything about grammatical rules. It also tends to be about as accurate as humans - the humans in this case tending to be graduate students in linguistics programs.

Granted, I doubt this would be anywhere near 100% accurate since there are probably a lot of pertinent variables that Google wouldn't be able to quantify. As long as it can manage to beat random chance by a decent margin, though, it would probably be beneficial to the company.

[PraedSt]

Dunno... OK, that would be true for single individuals and small groups. But It seems that the behaviour of large groups can be approached very nicely by maths.

Yes! But sort of...exhibit A, the credit crisis.

But your point seems to agree with my point. You'll note that Google is trying to work out which specific employees would be likely to leave, not an aggregate number. The latter is easy, a simple average would do.

[PraedSt]

To take something that's a bit closer to home for me, computational linguistics, the current gold-standard algorithm for guessing the parts of speech for unknown words in a sentence seems kind of silly at first - it doesn't even involve the program knowing anything about grammatical rules. It also tends to be about as accurate as humans - the humans in this case tending to be graduate students in linguistics programs.

I'll read up on this, sounds interesting. I'll note your last sentence though. I think in Google's case it would be cheaper and faster to rely on human judgment. Bosses should know about their subordinates, it's part of their job. Or it should be. This exercise is just HR gone mad. Dilbert!

As long as it can manage to beat random chance by a decent margin, though, it would probably be beneficial to the company.

True, but again Google has to compare it to the alternative, not random chance. Google also has to take into account costs, and the effect this will have on its own answers. (See OP- impersonal HR)

[PraedSt]

NOTE: The HR bashing is due to the fact this is an HR project. It could have been any department really. I think Google is wasting time and money, not specifically their HR dept.

[nauthiz]

I'm actually not sure about the idea that Google can best rely on human judgement. I think the only really firm evidence that approach has in its favor is that it happens to be the way we've always done things.

However, one big potential downside to that approach is that HR folks and supervisors are also workaday schlubs who don't come with a guarantee of motivation and competence any more than any other employee does.

That, and this new system they're coming up with could identify some key indicators that the humans just haven't been picking up on. That's the sort of thing that seems to happen pretty often when we start applying careful analysis of empirical data to tasks that were previously handled by trained monkeys who were left to figure out their own ad-hoc grab bag techniques all by themselves.

(ETA: Of course, we won't know until they've tried it. It's certainly possible that this won't bear fruit, but for the most part I'm not a fan of dismissing research and experimentation as a waste just because there's a chance it won't produce a financial benefit.)

[PraedSt]

However, one big potential downside to that approach is that HR folks and supervisors are also workaday schlubs who don't come with a guarantee of motivation and competence any more than any other employee does.

computational linguistics, the current gold-standard algorithm for guessing the parts of speech for unknown words in a sentence

This is machine translation? Any chance of the name of this algo? So I can look it up...

[nauthiz]

There are a few different variations; the one I'm thinking of is the Viterbi algorithm which is based on Markov chains. Here's a Wikipedia article on the topic that gives a fairly good overview.

[Gillianren]

I think in Google's case it would be cheaper and faster to rely on human judgment.

In the long term? I doubt it. "Better" is an arguable point--the HR person in my last job was dreadful--but cheaper and faster? Humans cost money, and one human can only see so many employees at a time.

[NEOWatcher]

However, one big potential downside to that approach is that HR folks and supervisors are also workaday schlubs who don't come with a guarantee of motivation and competence any more than any other employee does...

Yes; but they are workaday schlubs who are supposed to specialize in identifying these kinds of issues.

That, and this new system they're coming up with could identify some key indicators that the humans just haven't been picking up on. That's the sort of thing that seems to happen pretty often when we start applying careful analysis of empirical data to tasks that were previously handled by trained monkeys who were left to figure out their own ad-hoc grab bag techniques all by themselves.

I would agree with you if it weren't for the fact that they have already identified the problem. If this were done in conjunction and not to the impedence of other efforts, I would be all in favor.

This (seemingly) is only identifying who might have a problem with the problem. I would think that with such a large staff, that the exit interviews would be much more valuable. Those are the people who actually do have a problem with the problem.

It kind of reminds me of those restaurant surveys...
How was your food, waiter, experience, etc on a scale of 1 to 10. That's great for marketing, but does nothing to identify that the bread always tastes moldy.

[Fazor]

I'll get back to you later; I need to go devise an algorithm to see how I feel about this subject.

[Fazor]

It kind of reminds me of those restaurant surveys...
How was your food, waiter, experience, etc on a scale of 1 to 10. That's great for marketing, but does nothing to identify that the bread always tastes moldy.

My company has been on this "Customer survey bender" for the last few years. I mean, they hit this stuff hard. They haven't started using it against our commissions (yet) but they do use it to dictate certain advantages we may or may not get.

The it-woud-be-funny-if-it-wouldnt-hurt-us thing is, they send us graphs and charts each month as the update the data. These graphs chart our agency versus our district verses the state average versus the company average.

A monkey can look at them and realize that they're not giving accurate results (they just don't jive like they should, and obvious 40-50% discrepancies, and a suspicious way the company's average stays the exact same from month to month, with literally no variation.)

Yet, the higher-ups think these are a god send and rock-hard proof of how good or bad an agency is.

(I'll note that our marks are typically extremely high, so I'm not complaining because I feel we're getting cheated.)

[NEOWatcher]

...They haven't started using it against our commissions (yet) but they do use it to dictate certain advantages we may or may not get...

Done correctly, I do see the advantage of compiling the number for performance based issues. But; that still has nothing to do with what it is you need to do better.

In your case (IIRC you are in insurance)...

...(I'll note that our marks are typically extremely high, so I'm not complaining because I feel we're getting cheated.)

It's a perfect example of "done correctly". I'm sure the agencies that are falling behind are probably areas that are high risk. So now you're comparing people who deal with "customers" against people who deal with "victims". Of course the outcome changes.

[Fazor]

It's a perfect example of "done correctly". I'm sure the agencies that are falling behind are probably areas that are high risk. So now you're comparing people who deal with "customers" against people who deal with "victims". Of course the outcome changes.

I write almost exclusively high risk.

Here's where the problems are:

For instance, one question asked "Satisfaction with Agency" (Mind you, these are phone surveys so no underlines). We get high scores here.

Another question; "How likely are you to recommend [this company]?" Here, the scores are much, much lower. Yet the company uses their rose-colored glasses, and deduces that people love the agency, but are unhappy with the agency and less likely to recommend it to friends.

Of course, if pressed to answer why there's such a discrepancy in scores, I might say that people like their service, but aren't particularly thrilled with the company (rates, rules, coverage, whatever). But I don't have my MBA, so obviously I don't understand surveys or how to ask proper questions. :-P

[kleindoofy]

... a mathematical formula Google says can identify which of its 20,000 employees are most likely to quit. ...

I wonder if the Google people have ever heard of the "self-fulfilling prophecy"?

Several said Google provides little formal career planning, and some found the company's human-resources programs too impersonal.

Well then, it's high time Google came up with an algorithm to deal with that.

[PraedSt]

There are a few different variations; the one I'm thinking of is the Viterbi algorithm which is based on Markov chains. Here's a Wikipedia article on the topic that gives a fairly good overview.

Thanks. Looks good.

[PraedSt]

one human can only see so many employees at a time.

Now this is valid, and something I didn't consider. Volume. Google is a large company. Also, after you've done the hard work, the system could run itself and you could monitor in real time.

You never know, it could even render a couple of HR spots redundant.

Like that's ever going to happen...

[Fazor]

Well then, it's high time Google came up with an algorithm to deal with that.

Now I'm picturing two or three stuffed shirts approaching a pale, nervous worker at his cubicle.

Shirt1: "Sad to see you go, employee 0567.
Shirt2: "Are you sure we can't convince you to stay, employee 0567?"
Shirt3: "We wish you luck in your future endeavors, employee 0567."

Employee: "Wha-what? Are you ... firing me?"

Shirt1: "Our algorithms indicate you are leaving. We wish you the best, employee 0567!"
Shirt2: "We were sorry to have recieved your two weeks notice sometime in the near future, employee 0567!"
Shirt3: "It's too bad our algorithm shows you will not reconsider, employee 0567."

Employee: "But...but, I don't want to leave."

Shirt1: "We beg to differ, employee 0567."
Shirt2: "Computer programs don't lie, employee 0567."
Shirt3: "Clean out your desk by noon, employee 0567."

[NEOWatcher]

Now I'm picturing two or three stuffed shirts approaching a pale, nervous worker at his cubicle.

Two Precogs will do, it's usually the 3rd red-ball that gets thrown out.

[kleindoofy]

Well, I do remember reading a few years ago about companies who used statistical predictions about employee loyality.

It turned out that the company's attitude towards employees who had bad prognosis' changed to the negative, resulting in a change in the employees' attitudes towards the company. It was a Catch 22 of cause and effect and the results were usually totally erroneous.

The people at Google should google that article.

[Fazor]

The people at Google should google that article.

They tried; unfortunately googling the term "Statistical analysis of employee loyalty" pulls up the article, but it's buried somewhere between the 357 blogs about cats, and 500,000+ porn sites that also return from those keywords.

[nauthiz]

Well, I do remember reading a few years ago about companies who used statistical predictions about employee loyality.

It turned out that the company's attitude towards employees who had bad prognosis' changed to the negative, resulting in a change in the employees' attitudes towards the company. It was a Catch 22 of cause and effect and the results were usually totally erroneous.

Ah, ack, that would suck. I was thinking this would be more useful for reducing turnover by helping The Powers that Be to focus their efforts at improving employee satisfaction.

[tdvance]

Dunno... OK, that would be true for single individuals and small groups. But It seems that the behaviour of large groups can be approached very nicely by maths.

I agree with that to a degree--because, after all, it's a general fact that if you measure some quantity with a numerical value, and do this on a lot of individual samples, and if there's reason to believe the individuals are similar (not the same, like molecules, but humans are all the same species, er, probably )--then there is a strong tendency for the measurements to fall into a normal distribution.

Thus, averaging over large groups mitigates a lot of individual differences.

You can't predict that X will quit tomorrow, but you might estimate, and not be too far off, that X% will quit within the next Y months, and it's conceivable one could predict that this group of Z employees has a particularly high percentage of the future-quitters, and thus give them raises, proving that the way to get ahead is to be a quitter or.... well, maybe not that exactly!!!! (of course, being able to quit because someone else will pay you more in a heartbeat is a good reason to give a raise in many cases--it's usually not the duds that are in that category).

[tdvance]

I said the first paragraph a little wrong--it's not the individual measurements that tend to be a normal distribution in general (though they often are for a lot of things, like IQ or height, for example), but if you take the average of the group, then take the average of group #2, then the average of group #3, and so on, those averages will tend to be normally distributed.

Incidentally, the reason that peoples' heights tend to be normally distributed is that it is sort of an average of all the growth patterns from infancy to adulthood.