none
Movie Recommendation Experiment User Input recommendation RRS feed

  • Question

  • Curernt Movie Recommendation is basically recommend items for all users in the dataset what i want to do is to recommend items to user that comes from Web Input , one way to do that is take the input and use "Add Rows" to dataset and do select only input userId to get only that users ID recommendations .But its not trainable due to input , is there any other way to get recommendation for user input without adding rows to my dataset?
    Thursday, July 2, 2015 9:57 AM

Answers

  • Hi! I think I understand your issue now - are you saying that to score/predict new users you need to add them to the dataset used in the training graph? Are you perhaps interested in Converting your experiment to a scoring graph ready to be published?

    The basic idea of the above link is to right-click the output port of Train Matchbox Recommender and select Save As Trained Model. Then, just create a new graph (or copy of this one) with the training branch deleted and replaced with your recently saved model.

    This can then be published as a web service to accept new input!

    Let us know if this doesn't make sense and we'd be happy to help.

    Regards,

    AK

    Sunday, July 5, 2015 7:11 PM
    Moderator

All replies

  • You can have one dataset containing user-item-rating triples which is used for training, and a totally different dataset containing only user ids which is used for scoring. What's the problem with that?

    -Y-

    Thursday, July 2, 2015 12:49 PM
  • Let me show you with my model , my model is something like this , i have to add rows to user data to get recommendations for all and select userID data that i provide with the input . I am asking is there any other way cause the model is not training the red circle part.Sorry for my lack of knowledge i want suggestion from you . 

     






    Thursday, July 2, 2015 2:11 PM
  • Why is the model not training the red circle part?! By the way, it'll help if you had more clearly separated sentences and subsentences, because I find it difficult to parse what you say.

    -Y-

    Sunday, July 5, 2015 4:16 PM
  • Hi! I think I understand your issue now - are you saying that to score/predict new users you need to add them to the dataset used in the training graph? Are you perhaps interested in Converting your experiment to a scoring graph ready to be published?

    The basic idea of the above link is to right-click the output port of Train Matchbox Recommender and select Save As Trained Model. Then, just create a new graph (or copy of this one) with the training branch deleted and replaced with your recently saved model.

    This can then be published as a web service to accept new input!

    Let us know if this doesn't make sense and we'd be happy to help.

    Regards,

    AK

    Sunday, July 5, 2015 7:11 PM
    Moderator
  • Hi, AK, 

    I followed the way you said, and set Score Matchbox Recommendation as Item Recommendation and "From All Items"(production mode).

    But I found that my every input user's recommend items are all the same.

    Only I changed the setting from production mode into "From Rated Items(from model evaluation)" (evaluation mode), then the output of every user are different.

    But the "Score Matchbox Recommendation" documentation said, it need to choose production mode for web service.

    I can't figure out the reason. Can you give me some advise?

    Thanks in advance.

    Friday, July 31, 2015 10:49 AM