It used to be said that low latency networks weren’t needed for the data centers that ran big e-commerce or social media sites, as most people are willing to wait an extra few microseconds for the latest update on their friends and the network itself wasn’t a gating factor in performance.


Product recommendation technology, which powers the “you might also like” messages on e-commerce sites, however is changing that.  New “recommender” systems require increased computing performance to better factor more data into their recommendations.  And they need to do this all in the time it takes a webpage to load.


An article in the October 2012 issue of IEEE Spectrum by professors, and recommender system pioneers, Joseph A. Konstan and John Riedl chronicles the evolution of the technology and its dramatic impact on e-commerce sales.


The most popular recommender systems use either user-user or item-item algorithms – that is they compare your purchases, likes, clicks and page views with other people (user-user) or they compare the items you like with other items to see what buyers of those items also purchased (item-item). 


The two main problems with these approaches are that the algorithms are rigid and that tastes and preferences change, both of which lead to bad recommendations.


Dimensionality reduction is a new way to make both algorithms much more accurate.  This method builds a massive matrix of people and their preferences. Then it assigns attributes or dimensions to these items to reduce the number of elements in the matrix.


Let’s take food for example.  A person’s matrix might show that they rated filet mignon, braised short ribs, Portobello mushrooms and edamame with sea salt very highly.  At the same time, they give low ratings to both fried chicken wings and cold tofu rolls. The dimensionality reduction then seeks to determine that person’s taste preferences: 


“But how do you find those taste dimensions? Not by asking a chef. Instead, these systems use a mathematical technique called singular value decomposition to compute the dimensions. The technique involves factoring the original giant matrix into two “taste matrices”—one that includes all the users and the 100 taste dimensions and another that includes all the foods and the 100 taste dimensions—plus a third matrix that, when multiplied by either of the other two, re-creates the original matrix.”


So in our example, the recommender might conclude that you like beef, salty things and grilled dishes, but that you dislike chicken, fried foods and vegetables.


But the number of calculations grows dramatically as the matrices grow in size.  A matrix of 250 million customers and 10 million products takes 1 billion times as long to factor as a matrix of 250,000 customers and 10,000 products.  And the process needs to be repeated frequently as the accuracy of the recommendations decreases as new ratings are received.


This can spawn a lot of east-west data center traffic, which is needed to complete these large matrix calculations. Because users don’t spend much time on a given web page, the data center network latency is critical to providing recommendations in a timely manner (time is money).


Intel® Ethernet Switch Family FM6000 ICs are perfect for these types of data centers because of their pioneering low layer 3 cut-through switching latency of less than 400 ns.  So, the next time you get a great book recommendation, there might just be an Intel switch helping to power that suggestion.