2

In our case, we have users' click stream, items' attributes (like category, tags and so on), favorites about item, and collections for items. How can we combine these data as Myrrix's input data?

Sean Owen
  • 66,182
  • 23
  • 141
  • 173
piaolingxue
  • 75
  • 6

1 Answers1

3

Basically you are trying to model interactions between users and items. The way you would model the different interactions is by assigning a strength indicator to each. For instance, you could argue that a click has a strength of 2, a favorite a strength of 5 and perhaps a purchase a strength of 15 (I'm just saying numbers out of the top of my head).

Example of input data:

user1,item1,2 => he view the item
user1,item1,5 => he made the item a favorite
user1,item1,15 => he purchased the item

Now, internally, Myrrix will add all of these values, to indicate quite a strong preference for the item, hence you would keep all the interactions (and not just the strongest one)

The meta-data you might have on the users or the item can also be introduced to Myrrix as "tags" to better inform the model. So you could say that an user is "female" or an item is "jeans". You can have multiple tags per user or item and each tag can be assigned a weight as well.

Julian Ortega
  • 947
  • 4
  • 11
  • 1
    I agree. I can add a few more guidelines: http://myrrix.com/tuning-quality/ For example, weights can be chosen by relative frequency in the data. Something that is 20x rarer can have 20x higher weight. Also, see the new "tag" API as a convenient way of adding category/tag information: http://myrrix.com/rest-api/ – Sean Owen Apr 12 '13 at 13:07
  • Got it,if the input data has same UserID and ItemID, the strength are incremented rather than replace!and I have a question again,If i restart Myrrix with the same inputDir, whether the same csv data in that inputDir will be read again or increment the old value or just ignore? – piaolingxue Apr 15 '13 at 01:25