Cold Start with Recommendation API RRS feed

  • Question

  • Hi,

    I have collection of services and would like to recommend:

    * other services when displaying given service (item to item)

    * services for current user (user to item)

    My problem is "cold start" I'll quite likely have no usage data...(or very limited but let's assume I have none).

    How should I start recommending items using Azure Recommendation API?

    Would suggested / acceptable way would be to:

    1) Display random services and start getting usage this way.

    2) Don't recommend anything until I have enough usage data based on user's normally browsing the site. (not ideal). Perhaps use custom clustering of services to at least start with some shape of item-to-item recommendation (or rather similarity). 

    3) Enter fictional / test user data for some services e.g. get few users / testers to use website a bit and generate some usage that would kick-off recommendation engine. (perhaps later even re-upload usage data and remove test entries?)

    Which way would be recommended approach?

    2 additional questions:

    a) When / how do I find out when I have enough usage data? Assuming I have 1000 services, what is the amount of usage that I would have to have for recommendation engine to start? Is this something that can be calculated? Without usage data (or 1 entry in usage data sample application for recommendation API will not complete building the model..).

    b) documentation shows that I can add Catalog data with features and Usage data that is userId-ItemId pairs (with optional even info). Why user feature data cannot be added, wouldn't it help with recommendation?

    Thanks in advance for help :)


    Tuesday, September 29, 2015 6:35 AM


  • Hi Lucas!

    If you have user & item profile data, you could build a regular regression model using those as features to bootstrap your model. Then, as you get more usage data you can switch over to collaborative filtering.

    Hope this helps,


    Monday, October 5, 2015 5:02 PM