Hi, I'm building a fashion startup where we make users swipe on personalized clothing recommendations. I have 2 questions: one on how the recommender works with replaced dataset and the other regarding best practices on training my data.
The first issue I have is as items go out of stock, I need to remove them from my dataset. However, I can't just remove one item -- I'm required to replace the whole dataset. My concern is that every time I do this, I think personalize begins treating all of them as cold items because the level of recommendations very visibly tanks. In an industry where I need to keep updating my dataset what should I do? I'm using aws-user-personalization which is supposed to offer the best support with cold items but it's still inaccurate.
The second is about training data. The reason why I created this interface was to solve a problem with fashion websites: brands don't know if a user dislikes an item because scrolling past something could mean they didn't see it or they didn't like it. On my interface, users swipe left to dislike and right to like so I collect data on explicit preferences. However, I keep reading that you should only train your recommender on positive data so thus far I've only trained it on right swipes. However, does this mean all the data on left swipes is being disregarded? Is there a better way to approach this?
Thank you so much! Would really appreciate any assistance on either question. Do let me know if I can provide any more details about solutions I'm using.
AWS OFFICIALUpdated 8 months ago