1. The inputs q and u are needed because the similarity of l to l_antecedent is defined in the context of the query and the user. For example, two listings may be somewhat similar to each other for users bookings for 2 guests, but may not be similar for users bookings for 5 guests.

2. Yes, all features of the listings are input as they can be potentially useful to infer similarity.

3. That's right. The exact iterative algorithm can be found in the paper as Algorithm 2 on page 6:

https://arxiv.org/pdf/2210.07774