« Huge steaming piles of data (Project Cocktail Part 2) | Main | Just. One. Word. »


Alex Schultz

Re. the alphabetical bias although in the rest of the data it is also strong, in the data above I wouldn't say it is mindblowing. The reason I say that is relationships such as "Slippery Dick Slippery Nipple " are to be expected and similarly with the vast majority of the other cocktails on the list above with the same few starting letters.

I will certainly agree with you that I need to look at CTR once I add the recommendations (but I do need a starting point with which to produce the recommendation list) so probability of click will be one key factor (as pointed out in the article above) but I think probability of liking will be a key factor too (i.e. if I view cocktail a and then visit cocktail b what is the average rating b is given in that scenario).

I definitely want to do something with a chain of cocktails (i.e. if I saw a and then b what is the most likely cocktail next) but baby steps first baby steps :).


...this looks like a huge alphabetical bias.

I would recommand developping a method that, on each page, would pick up and list a few "Suggestions", let's say five. The first three would be those that have the best clickthrough rate from the current cocktail page, the two others would be chosen at random (in order to be able to account for new cocktails or new trends in user interests).

This way you do not have anything to manage once it's functional. You just need to log in how many times you displayed each suggestion on any given cocktail page, and how many times it's been clicked on.

The next stage would then be to build session profiles, where you do not just take into account the cocktail currently viewed, but also those browsed through during the same session. These could make up a "pool" of cocktails, from which - based on the individual best matches - you could compile the best suggestions.

Would that make sense?

The comments to this entry are closed.

My Photo
Blog powered by Typepad

Become a Fan