Personalization Through Tracking Triplets of Users, Queries, and Web Pages

What value might the tracking of links to pages that searchers click upon within search engine results pages have for personalized search?

Imagine a search engine keeping track of each user (u), as they perform a query (q) on the search engine, and seeing which pages (p) they click upon, and collecting those selections in what they call “triplets” of data, represented like this – (u,q,p).

Then consider that the search engine might map and compare those triplets of information against each other to see what kinds of relationships and associations exist between people making the same searches, faced with similar results. That information could then be used to personalize results shown to individuals.

You’ve probably seen something similar at Amazon Books, where they tell you that “people who purchased this book, also bought these other books.” Instead of just looking at two points of data – the first book, and the second book, this method would look at three – who the person searching is, what they are searching for, and what other people who may share similar interests conducting similar searches may have chosen in the past.

That’s the topic of a new patent application from Microsoft. The patent filing shares a number of authors with a Microsoft research paper that appears to be closely related, CubeSVD: A Novel Approach to Personalized Web Search. Both the patent and the paper are very math heavy, but the concepts behind them are worth pursuing.

Augmenting user, query, and document triplets using singular value decomposition
Invented by Hua-Jun Zeng, Jian-Tao Sun, Wei-Ying Ma, Zheng Chen, and Benyu Zhang
Assigned to Microsoft
US Patent Application 20070055646
Published March 8, 2007
Filed September 8, 2005

Abstract

A system for augmenting click-through data with latent information present in the click-through data for use in generating search results that are better tailored to the information needs of a user submitting a query is provided.

The augmentation system creates a three-dimensional matrix with the dimensions of users, queries, and documents. The augmentation system then performs a three-order singular value decomposition of the three-dimensional matrix to generate a three-dimensional core singular value matrix and a left singular matrix for each dimension.

The augmentation system finally multiplies the three-dimensional core singular value matrix by the left singular matrices to generate an augmented three-dimensional matrix that explicitly contains the information that was latent in the un-augmented three-dimensional matrix.

Personalization without profiles

The patent filing isn’t all numbers and symbols. It does contain some text that tries to explain the processes involved in plain English. It talks about a couple of well known ranking methods, including Google’s PageRank, and notes that one of the failings of those methods is that they don’t pay attention to the person who is submitting a query. Here’s what they say:

For example, a zoologist who submits the query “jaguar” would get the same results as a car enthusiast who submits the same query. In such a case, the zoologist may be interested in web pages related to animals, whereas the car enthusiast may be interested in web pages related to automobiles.

It also discusses the difficulties of some present day personalization methods which try to create user profiles to base which results they should show to individuals. Those methods usually attempt to customize results for each user by basing results shown to them upon personal profiles that are created manually or automatically, and then adapting search results based upon those personal profiles.

A difficulty in using personal profiles is that if they are created manually, people tend to not want to share their personal information with a search engine – so requiring people to fill out an extensive amount of information concerning their interests is unlikely to result in many people using such a service.

Creating a profile automatically in the background, based upon previous searches could require a large amount of user history data being collected. Regardless of whether a manual or automatic approach to creating a profile is followed, we’re told in the patent that there’s a question as to whether “complex user behavior can be accurately modeled by a personal profile.”

Conclusion

Imagine instead a system that pays attention while you are searching, and compares your search selections during a session to others who made similar queries, and chose similar results. It may then start reranking results that you see based upon the searches that you perform and the selections that you make.

If you would like to learn more about some of the math described in the paper and the patent application, a good starting point may be Dr. E. Garcia’s tutorials on Singular Value Decomposition (SVD) and Latent Semantic Indexing (LSI). In addition to being a great source of information on those topics, the tutorials also do a wonderful job of debunking marketing myths around latent semantic indexing, and are highly recommended.

Share

10 thoughts on “Personalization Through Tracking Triplets of Users, Queries, and Web Pages”

  1. Pingback: Link Grabbag
  2. This sounds very interesting for shopping sites (like Amazon), but I’m not sure, what’s the point of tracking the user variable (u) for search engines?

  3. Hi Robert,

    By tracking a variable based upon users, a search engine is capable of understanding more about the context of a query. For example, someone searching for [football] from the UK probably wants to see search results for teams like Manchester United. Someone in the United States searching for [football] probably doesn’t. Or look at the example I provided in the post:

    For example, a zoologist who submits the query “jaguar” would get the same results as a car enthusiast who submits the same query. In such a case, the zoologist may be interested in web pages related to animals, whereas the car enthusiast may be interested in web pages related to automobiles.

    Those are good reasons for tracking the user variable.

Comments are closed.