Thursday, August 02, 2007

Self-optimizing ad systems

After talking about some of Google's early steps toward personalized advertising, Philipp Lenssen goes a little visionary, writing:
I wonder, with that massive amount of ads + searches Google has, if there’s some merit in allowing the software to figure it out for itself... evolutionary algorithms, self-learning style.

Search sessions are automatically grouped into general patterns, and then random ads are presented, and when an ad performs well, more ads from that ad segment will be displayed next time, and so on, causing a "survival of the fittest ad" environment.

Then when Google meets the press in 2012, they can tell the journalists, "We don’t have a clue anymore how our ads work, but click-throughs are higher than ever."
AdWords and AdSense already self-optimize depending on ad clickthrough rates, but Philipp is talking about something more here.

Rather than have ad systems target to specific keywords, let's make the ad systems target to micro-groups of intent, where intent is determined both by current actions and past behavior.

Rather than specify exactly who to target an ad to using keywords, the keywords merely would be hints to the ad engine, a starting place for a likely target audience. Ads submitted enter a great pool of experimentation where ads are shown to different audiences with different behavior, culled where they fail, reinforced where they succeed.

As much as I like this idea, ad systems appear to be headed in the opposite direction, with offerings from Microsoft and Yahoo touting the additional controls -- the knobs and levers -- they give to advertisers.

It is not surprising that advertisers want control over their ads, but, in the long-term, ad systems are most effective when they serve two audiences, consumers and advertisers. Consumers pay attention to useful and helpful ads. Advertisers want effective ads that consumers pay attention to and use.

Relevant ads are useful, helpful, and effective. While giving up control may be hard for advertisers, it is inevitable that they will have to do so. It is impossible for advertisers to manually tune their ads to millions of micro-audiences with subtle variations of intent. Only an automated solution can deliver that.

Eventually, we will see organic, self-organizing advertising systems. Advertisers merely will give their ads some guidance as they nudge them out the door. Then, they will sit back and watch as the ads find the audiences they seek.

Update: If you might enjoy a version of this for search relevance instead of ad relevance, you might also be interested in my earlier post, "The perils of tweaking Google by hand".

8 comments:

Robert said...

True. Customers consider personalized ads as helpful. As for my thesis project I've created a self optimizing ad system. It uses online services and customer profiles to learn from clients in-store and news / synopsis from imdb. Hopefully we'll see the online / offline world blending some day.

jeremy said...

Search sessions are automatically grouped into general patterns, and then random ads are presented, and when an ad performs well, more ads from that ad segment will be displayed next time

I suppose something like this is certainly within the realm of possibility. But as you were saying, Greg, about the advertising having to serve both the consumers and the advertisers.. I myself as a consumer would quickly lose patience with the ads I was seeing as the system "trained" itself. I would begin to associate the ads column with randomness, which is what it really would start out as. And if I then never clicked anything, because I stopped looking at it, because it was random junk, it would never be able to train itself.

Greg Linden said...

Hi, Jeremy. I would hope that ads so poorly targeted as to appear random would be very infrequent in the experiments -- on the order of 1% of the time -- but it is an interesting question to ask how much exploration would be necessary and how far that exploration would have to be away from known good targets to provide useful data to improve targeting.

jeremy said...

Sounds like the classic AI dilemma of exploration vs. exploitation.

Andrew Goodman said...

Greg, I enjoy your blog.

This post sounds plausible on the surface, but I wonder how much actual experience you have in running real life online ad campaigns, specifically AdWords campaigns?

Intent is easy enough to measure in hindsight, based on current methods of tracking ROI and other user behavior data from paid search campaigns.

If someone wanted to go beyond today's current forms of automation to serve their own ads in a more rule-based way, that would make sense. I'm not sure it should be up to the ad serving company to wrest control and creativity out of the advertisers' hands.

And your as-yet-unexamined concept of "intent" is actually muddy and has large enough holes to drive a truck through. There are a lot of pseudo-researchers making a lot of noise today about "purchase intent," but a tiny percentage of them offer any real insight into that process. Integrating the study of intent (and/or growing awareness) as it grows from a variety of sources - online, offline, paid, organic, wom, pr, etc. - is far more complex than the pat scenario of "self-training systems" allows.

Not every transaction even looks the same, needless to say. Sales cycles differ, some relationships must be initiated through a lengthy process of persuasion (content), some require an opt-in; businesses win or lose based on total order size, repeat business, etc.

So working to an idealized model of the "customer" and the "transaction," you can envision a lot more elegance to the process than is likely to emerge anytime in the next decade.

As Godin has taken to saying, targeting is a hunting/military analogy anyway.

My take is that a lot of the people who are suddenly intrigued with advertising are misinterpreting the whole field and looking at it from the perspective of one or two generations ago.

And more: we *already* know a lot about intent from current AdWords campaign data. I'm not sure this requires a futuristic self-training ad system.

Greg Linden said...

Good point, Andrew, I likely am oversimplifying things.

Even so, using the hunting analogy, if you start hunting targets individually rather than going after large groups -- one-to-one marketing -- it seems impossible to scale that using manual targeting.

The big question here is whether the performance of some self-optimizing ad system doing fine-grained targeting would be high enough to justify the loss in control over doing manual coarse-grained targeting.

That probably depends on how well the system ends up being able to determine the intent of small groups based on current and past behavior. A tricky problem, as you said, but one many hope to conquer.

My guess is that this kind of fine-grained targeting system eventually will outperform the current norm by a large enough margin that advertisers will feel they have little choice but to switch away from the coarse-grained frameworks where they had more control. But, that is speculation, so please take it as such.

Theo Van Rooy said...

Genetic style "exploration" algorithms have one main weakness, they often get locked onto local maxima and minima.

Furthermore, when the machine attempts to learn an ad's preferred user group, what kind of numbers must we see in terms of clicks to really establish a pattern: 50, 100, 1000? Most ads only receive clicks every 1000-2000 views. That means each ad would need to deliver to an incredibly broad spectrum of potential consumers with a very large number of views to "organically" find its way into the optimal user group (wich we can't really assume is not a suboptimal local maximum).

I think that the AI approach is the correct approach for this problem...but I don't disagree with Google's hand tweaking, nor Yahoo's or MS's. Sometimes those machines need a little kickstart in the right direction.

Kaila Colbin said...

Greg, what a great piece! I completely agree with you about the need for ads to serve the consumers as well as the advertisers.

You suggest a self-optimizing system based on intent, where intent is determined from past and present actions. I'd like to throw another possibility into the mix: a system that takes into account who the user is and what the user really cares about, rather than demographics or search history.

Personalization is one of many waves of the future, yet consumers are understandably wary of privacy transgressions. It is possible, though, to identify relevance without tracking user history; I blog for a company that has done just that.

I look forward to reading your future posts!