• by Steven Lee , May 31, 2007

In all the run-up about market exchanges being introduced outside of the norm, one element has been missing from the conversation. Understandably, this element is about as esoteric as anything in our business. In fact, it's downright geeky.

What I'm referring to is the role that Machine Learning will play in the new Web 2.0 world.

Let's examine the relationship between market exchanges and machine learning from both the vantage point of the buyer, who is mostly concerned with the simplicity and transparency of the exchange, and also from the vantage point of the seller, who understands the need to participate in the exchange.

But, most importantly, let's examine the notion from the perspective of the user -- the consumer being targeted. When we get to that perspective, you'll see that Machine Learning has entirely different, far more automated and sophisticated applications -- especially in video.

The Buyer's Perspective

One of the reasons DoubleClick and Right Media both generated a lot of buzz with their exchanges before ultimately being acquired is that providing such exchanges enables buyers to treat impressions-based inventory like a bid engine in the Search segment.

That's where many industry observers think most media is heading, and I'm in full agreement. If I'm a marketer in pursuit of the best value online, the learning that a fully automated exchange can provide is unsurpassed.

There's a reason Google has become the behemoth it's become, and much of that has to do with the auction model that they have all but perfected around text ads. For marketers, it couldn't be simpler.

The Seller's Perspective

But -- what about the publisher? Of course, text ads that appear on Google as search responses reside in a walled garden. Nobody at Google will complain about being outbid, since all the money goes to Google anyway. But, in an AdSense world, this is not always the case. Publishers do sometimes complain about how the vagaries of AdSense performance, as well they do about other market exchanges they do business with.

The Web is filled with thousands of sites published by thousands of companies and individuals. So, naturally, there will be winners and losers. If you're a publisher doing business with one of these market exchange networks -- or all of them and others -- you know that on a month-to-month basis, your revenue is not as predictable as you'd like it to be.

What you have to do then, Mr. or Ms. Publisher, is make your site smarter so as to better attract those dollars. Pay attention to the next stanza to learn how -- because as is so often the case -- it's all about the user, stupid!

The User

Publishers in all media have always acted as the trusted intermediaries between the buyer and the consumer, or user. With Machine Learning, the new intermediaries are the companies running the exchanges that enable the buying of the media, and these new intermediaries should provide more than just the market exchange that enables the transaction: They also should provide the intelligence that enables precise targeting of individual consumers.

That, after all, is what the future of interactive is about. What Machine Learning can provide is an analysis in real time that informs every transaction along the value chain. The most important one is the one that shows the user the precisely selected ad unit at the precisely most opportune time.

How can well-designed Machine Learning accomplish this? We've all seen how Google has begun leveraging machine learning to build what is essentially a behavioral targeting algorithm for search results. For our purposes, let's stick to video.

How Machine Learning Works

To get information about a video, it doesn't take advanced technology to extract information from the textual meta-data. Extracting information from the audio stream and visual cues requires more advanced technology, but such technology exists. For any given video, the level of strength of each of these "signals" will be different. It's important to use machine learning to discover the useful signals and filter the noise.

Relying on any individual signal by itself is dangerous. Only through a Machine Learning algorithm can a noisy stream of signals be converted into meaningful, contextual matching; providing the kind of targeting referenced above along with the brand protection that advertisers have been waiting for.

This can be done using algorithms to train statistical semantic models which inform the Machine Learning, accumulating data that builds predictive profiles that must constantly be kept up-to-date. For example, lonelygirl15 is a term that did not exist 12 months ago but is regarded today as "the first video blog."

An effective Machine Learning model needs to know this is a vlog about a teenager. As media accumulates, so does the data that informs this media. Only Machine Learning can accumulate it all while separating the actionable data from the fluff.

Similarly, Jet Blue was associated with flawless customer service until recently. So, the Machine Learning model needs to capture the latest information that airport delays is a bad association to Jet Blue.

While the Machine Learning activity is being carried out, data is gathered from sensors affixed to every object or site or portal involved. The data is fed into a reasoning engine -- a machine learning algorithm that analyzes the data, compares it to a large set of activity models, and infers which model is the best match. On a minute by minute, second by second basis, the best tools gather and act on more data than a human does in a lifetime.

Finally, machine learning is clearly key to ad optimization. For those who don't understand how optimization works, you don't want to serve an ad to promote extreme sports participation to grandparents in Florida. Machine learning is a critical component to learning these relationships. At the same time, it should provide protection from the brands who are wary of seeing their ads in the wrong place at the wrong time.

Brand Protector functionality should appropriately leverage Machine Learning algorithms which are constantly trained and updated to determine if a video clip is controversial or inappropriate. If the Machine Learning software is up to it, the learning experienced by the software through this process should continually refine the machine algorithms.

In other words, as media accumulates, so should the actionable data gathered by the software. Only then can the targeting and delivery be as precise as the buying dictates. Machine learning on the front end will completely upend the way that video or any other media is transacted by adding more actionable information in real time than anything we've ever seen. Anything else with a Market Exchange on the back end will merely create a new way to buy, and is not terribly exciting.