User1 purchases = infant car seat, infant stroller User2 purchases = infant car seat, infant stroller, infant crib mobile
The obvious recommendation for User1 is an infant crib mobile. From the purchase history the users look similar. Here similarity is in “taste”. User or item information that does not relate to taste may be misleading for recs. If you look at their profiles: User1: male, 55 years old, upper 75% income User2: female, 29 years old, lower 25% income User1 is actually a doting grandfather, User2 a doting mother. Their profiles are quite dissimilar though their taste is similar. The point being that those other pieces of data may not relate to user similarity *of taste*. Going through the cross-recommendation process applies cooccurrence analysis to the data that checks to see if the secondary data correlates in an important way with the action you know is important. For this reason it’s usually best to start out ignoring that information and using just <UID> <ITEMID> for the important action. Later you may find uses for the extra data, or may consider viewing or purchasing from a certain category as a secondary action and use cross-recommendations to improve things. On Dec 4, 2014, at 7:17 AM, Yash Patel <[email protected]> wrote: Cross Recommendors dont seem applicable because this dataset doesn't represent different actions by a user,it just contains transaction history.(ie.customer id,item id,shipping location,sales amount of that item,item category etc) Maybe location,sales per item(similarity might lead to knowledge of people who share same purchasing patterns) etc. On Wed, Dec 3, 2014 at 5:28 PM, Ted Dunning <[email protected]> wrote: > On Wed, Dec 3, 2014 at 6:22 AM, Yash Patel <[email protected]> > wrote: > >> I have multiple different columns such as category,shipping location,item >> price,online user, etc. >> >> How can i use all these different columns and improve recommendation >> quality(ie.calculate more precise similarity between users by use of >> location,item price) ? >> > > For some kinds of information, you can build cross recommenders off of that > other information. That incorporates this other information in an > item-based system. > > Simply hand coding a similarity usually doesn't work well. The problem is > that you don't really know which factors really represent actionable and > non-redundant user similarity. >
