0

With the introduction of the new aws-similar-items recipe I thought this would bring a big improvement to the "because you watched x" but for a lot of the items it seems to just show popular items, with items not even having any genres/descriptions/titles in common.

Is there anyway to make this weight more towards item metadata than historical data?

devnotdev
  • 85
  • 8

1 Answers1

1

You currently cannot control the weight of item metadata vs historical/interaction data with the Similar-Items recipe. However, the SIMS recipe can also be used for the "because you watched x" use case. SIMS only considers the interactions dataset so item metadata does not come into play. Furthermore, you can control for popularity with SIMS using the popularity_discount_factor algorithm hyperparameter.

James J
  • 621
  • 3
  • 6
  • Thanks but I seem to be running into use-cases where it basically doesn't use the item metadata at all, normally for low interaction items and it just returns popular. Even if i add a Genre filter it just seems to ignore it – devnotdev Jan 10 '22 at 16:00
  • @devnotdev, what does your items schema look like? Does your genre column have "categorical: true" and does your description column have "textual: true"? Without these attributes in your schema, string based columns are not included in training. Also, you can only have one "textual" column so if description is your textual column, title does not make sense to include (although you can concatenate them in your items dataset). – James J Jan 10 '22 at 22:46
  • the genre column does have categorical: true and the description "textual: true" and this is the only textual column we have. But as mentioned for items with more interactions, the similar returns decent results but with low interactions it doesn't seem to use the item metadata and is just returning popular. – devnotdev Jan 11 '22 at 09:59