Skip to content

Timelined Information Retrieval

I was thinking about how I search through my email this morning and worked out that sometimes I know more about *when* an email happened than what it said or who it was from. This is a rare thing, but generalizing, I quickly worked out that this would be a great addition to any/all search interface(s) if done well.

I want to be able to specify where in time I think my known item search should look. I think it could be done fairly simply with a well-designed normal distribution curve.

I want to see a timeline (aka a landscape-oriented rectangle) with a distribution curve that I can drag around. I would be centering the curve on *when* I wanted to focus my search.

The search itself would still do fulltext and weight like before, but now, would scale that prior weighting by how well it fit under my specified curve.

I have not done any due-diligence in looking through the information retrieval literature, but I have not seen this interface before and it seems like it would be very helpful for certain types of known-item, time-based queries.

Things that are not within my “window of interest” would be punished with a reduced relevance score in my overall search results. Things that matched my curve, in time, would receive a boost. Otherwise, the search behaves as it always has. This would simply be an addition parameter that gives more power to the searcher who knows *when* they’re looking for.

Has anyone seen anything like this before? Where?

Tags: - - - - -

View blog reactions

{ 2 } Comments

  1. Jane | November 7, 2007 at 1:18 pm | Permalink

    It seems like another extension of your relevance game. Or are you thinking of an event algorithm of some sort, Terrell?

  2. Terrell Russell | November 7, 2007 at 2:30 pm | Permalink

    Jane,

    I’m mostly thinking that this would be helpful – and looking for anywhere it’s already been put into practice.

    This doesn’t necessarily have anything to do with reputation – except tangentially when you may know that some event happened in the past (related to your current reputation decision-making) and you’re looking for it in a stack of ‘all prior events’. If you can find it more easily, more quickly, then it could affect your decision-making – because it’s relevant during that decision-making window of time. As a technique for IR, having a ‘draggable probability density function’ seems like a worthy path of inquiry as a researcher.