I am working on an information retrieval problem statement. I have trained a deep learning model using the pairwise approach. And NDCG and MAP are very good compared to the pointwise approach, but it is bringing extra computation. While inference I have to rank 10k documents at once, In pairwise I have to create all possible pairs and then I will be able to rank based on the result of the pairwise model or I can apply to merge sort. Using this approach we cant deploy our model in production because it will too long to rank the documents.
So I am looking for an optimized sorting algorithm for the pairwise ranking model.
Thanks in advance.