Search code examples
recommendation-engineazure-machine-learning-service

probability on azure recommendations api


I am using the azure recommendation api on http://recommendations.azurewebsites.net/. I prepared the catalog to be like <Item Id>, <Item Name>, <Item Category>, <Features list> and the usage file : <userId>, <ItemId>. Now when I test the recommender, I always get a probability of 0.5 for all items, so I had to presume something is not right. In order to know what's the problem I added two items to the catalog one with same features as an other item but with different name and id, and an other item with different id and one different feature. I still get the 0.5 probability and now i'm sure something is not right but I still can figure out what the problem.

here is a screenshot of what I get when I add the item to the cart

Is there any possibility to use the azure ml matchbox recommender with features and without ratings?


Solution

  • Tayehi,

    Nice to meet you. I am the program manager in charge of the recommendations API. 2 things:

    1. If you get a 0.5 probability you are most likely getting "default recommendations". This usually means that you do not have enough training data or that there are not enough co-occurrences for the item you are testing in the data. To describe the extreme case, imagine an item A that only gets purchased with an item B only one or two times -- it would be hard to say with confidence (statistical significance) that someone that likes item A is also likely to like item B.

    2. It looks like you are still using the old recommendations API. I would like to encourage you to use our newer version (the Recommendations API cognitive service). Please take a look at https://azure.microsoft.com/en-us/documentation/articles/cognitive-services-migration-from-dm/to help you in this process.

    Thanks! Luis Cabrera Cortana Intelligence Applications.