in Thoughts

Search Engine Design meets Economics

Also a 2-year-old post for a search engine which is build on an other foundation. Nice idea.

Google is working on penalizing over-optimized pages and still it probably won’t affect the results for long which isn’t too surprising. If there’s a method to exploit certain circumstances people will probably do it.

The goal here is to let the users themselves inform the search engine as to what content is good — hence the plus-ones, social search and all of that stuff. But all of this is still indirect evidence. Unless you could plug a computer inside the head of a person and watch their every thought, the only real data you have for input is server logs, click-throughs, and all kinds of other things that computers do, not people.

I just don’t see this being solved any time soon. But I do see it getting so complex and unwieldy that it continues to frustrate searchers and content producers alike. —DanielBMarkham

In economics there’s this ultra interesting thing called a market mechanism. Basically is a system that leads to some (desired) results. Could we apply the logic of building economical mechanisms, with the help of mechanism design, to build a search engine which is self-regulating?

What the fuck are you talking about?

Let’s make it super easy. You have an apple and what to sell it at the highest price (this is your desired outcome). How should you do sell your apple? Making a fix price and waiting for a buyer? Do a second price auction? Create an algorithm for pricing your apple?

Let’s see how these scenarios play out.

Fix price

You set the price for your apple to $5. There are three cases. The maximal value in your market is either less than $5, i.e. nobody buys your apple and it will rot. It is exactly $5, i.e. you will get your $5. Or it is higher than $5, i.e. you still only get your $5.

Algorithm

You look at past prices of apples. You try to find factors which influence the price: season, size, color, freshness. You build your model and price your apple adequately. The problem is that your apple may be an outlier, e.g. really big or has a new color. On average, you get what you can expect on average. This model is a bit better than the first one because you will sell your apple with a higher probability. Still you can underestimate the market value of your apple.

Auction

You take your apple to an action and let people bid secretly. After everyone submitted their bid the highest bidder will pay the second highest price. In this case, you can detect the highest market value (which is interesting for the future). Furthermore, you will receive, if there are enough bidders, the highest price for your apple.

There’s tons of literature on this topic. Go read Experimental Auctions if you are interested. You can understand it even if you didn’t know much about micro economics / game theory.

How can we take this idea in the realm of search engines?

This was my initial question. We can design market mechanisms that work quite well to reach a desired price. The interesting thing about market mechanisms is that they work well even if they are gamed (to a certain degree). My idea than was. What if we build a system where we expect people to game it instead of hoping that the users will play by the rules?

Analogy

In the moment, it looks like Google is using some kind of the algorithm approach. They say that they don’t interfere directly from data but look for factors. One example is the amount of links (super simplified). This was the basic page rank algorithm. Many links = good site. Of course, people who want to profit gamed the system. They started to build link farms, because many links = good site, and pushed their site.

How could this look like?

I don’t have a real idea. We have to define some desired outcome. For example, for a question, we want the right answer. Now it become gritty. What is the right answer?

Let’s look at the other side of the table. What do web site providers want? Views (generally). So, the web site provider has the incentive to give the right answer if the user can penalize or reward the web site provider.

In the simplest sense, a “was this website helpful?” would be sufficient. However, the web site provider has an incentive to game his helpfulness. We have two options. a) try to fix the gaming of helpfulness or b) fixing the reward/penalty mechanism. I tend to lean to b) because a) could end in a rabbit hole. I have no good idea to fix this in the moment but if someone could fix this problem, he would swim in money shortly afterwards.

Write a Comment

Comment