Examples with negative weights

[ClaireSuperlinked]

When playing with the netflix example I decided I wanted to find movies that matched the genre string "action sci-fi thriller" with a positive weight while biasing results away from genre "romance" using a negative weight.

To do this I defined another space using the same field. I had mixed results and noticed I often ended up with movies
that weren't "romance" nor "action sci-fi thriller" - the genre "family" was dominating the results. I tried to think why this was happening and my working theory is that the weight of my negative biasing was so great that it pushed the matching space away from "action sci-fi thriller" towards the opposite of "romance". Is "family" the opposite of "romance" in terms of genre similarity... maybe. Maybe opposite isn't the right word - but I'm thinking of something further away in the vector space.

This whole exercise raised another question: What is the best way to perform positive bias with one input and negative bias with another on the same field?

This seems like a real-world scenario for recommendation engines where users have both liked and disliked various items. Is there a better way than making a second space for the field? Taking this further, it seems sensible to input query strings as a list of tuples where each tuple comprises a value with a corresponding weight for that value e.g., [("Action", 4), ("Sci-fi Thriller", 8), ("Romance", -2) ].

[morkapronczay]

Great observation!
Can you please provide the code that produced this experience? I would love to dig a bit more deep into this. My initial feeling is that with 2 .similars referring to the same space with positive and negative weights I can comprehend these results. With separate spaces and separate .similars referring to them this is counterintuitive for me and would like to investigate further.

About your suggestion on tuple input: I genuinely like it, we will have to see how it fits into our product more broadly.

[morkapronczay]

Thanks @matthew-pest! Thank you for reproducing your example!
I will try to get back to you ASAP, although it might be only early next week, apologies

[morkapronczay]

Hi @matthew-pest! Sorry for taking too much time to answer! The report you filed indeed shows a bug in our system, it is verified. We are in the process of fixing it, will let you know when it is done!

[morkapronczay]

After prioritisation of issues this will be solved right at the beginning of the new year! See you in 2024!
Happy holidays :)

[morkapronczay]

Hey Hey! fixing turned out to be a little more complicated, but I can now say that we have got there to do it finally. Sorry for the long wait, should work as expected from version 5.4.1