RankNet using neural net technology
January 13th, 2006
Despite a lack of public comments on their relevancy ranking system based on “neural net” technology (called RankNet), MSN has been making strides to further improve the relevance of their search results ? something Google has yet to do. To understand how RankNet will affect you, you first have to understand what neural networks are.
Neural networks
Neural networks are based on the parallel architecture of animal brains. They consist of simple processing elements, a high degree of interconnectivity and adaptive interaction between their elements. The network adapts by changing the weight by an amount proportional to the difference between the desired output and the actual output. This creates a network that can “learn” at an increasing rate as it processes data.
RankNet
RankNet uses neural networks to evaluate a set of documents and determine not which are the most relevant to the search query but rather the factors that actually determine their relevance to the search query. A neural network in and of itself is fairly simple and straight forward, the difficult part is teaching the network how to properly apply weights to various features of a web page. To put this in perspective ? MSN utilizes 569 features to evaluate a document and each feature can be weighted differently depending on the circumstances.
As MSN’s neural network “learns” how to properly rank web pages on a consistent basis, they will be able to offer more relevant and timely search results than Google or Yahoo and begin taking a larger portion of the search engine market share. RankNet should help to decrease the amount of spam web sites and splogs. Legitimate business owners will benefit from increased traffic if they have an effective search engine optimization/internet marketing strategy in place and of course, users will benefit from more relevant search results.
