The Fifth Elephant
The Fifth Elephant

The Fifth Elephant 2026 Winter Edition

Call for Data Problems and Proposals

All submissions

AMIT RAI

@amit12018

[Add a tentative/catchy title for talk/showcase/problem discussion]

Submitted Oct 11, 2026

🎤 I am submitting: To speak I have a submission for: Showcase - a cool idea I have built/working on - 20 mins

Submission type
Showcase — new ideas — 20 mins
Title
Beyond Cosine: Learning How Bi-Encoders Measure Relevance
Describe your session in two paragraphs
Bi-encoders represent queries and items as vectors, but ranking usually relies on a fixed similarity function such as cosine. What if we learned the scoring geometry from relevance judgments instead? This session explores a lightweight, relevance-guided scorer that predicts positive dimension-wise weights at the midpoint of each query–item embedding pair. Using e-commerce search as a case study, we compare fixed cosine scoring with learned local weighting under frozen, jointly trained, and staged encoder configurations. The geometric motivation is simple: different query–item interactions may benefit from different “rulers,” even within the same embedding space.
Early experiments show promising ranking improvements, particularly with frozen representations, but attributing those gains to local geometry requires stronger controls. I’ll discuss the method, preliminary findings, and the experimental questions that remain: does local directional weighting outperform global reweighting or pair-dependent scaling, how should checkpoint selection and evaluation be designed, and where can a pair-dependent scorer fit into an efficient retrieval pipeline? This is a work-in-progress research showcase, inviting critique and collaboration rather than presenting a settled claim that non-Euclidean geometry is necessary for retrieval.
Takeaways

  1. A concrete way to extend bi-encoder ranking beyond fixed cosine similarity without using a cross-encoder.
  2. How to distinguish gains from local weighting, additional scoring capacity, and training choices through controlled evaluation.
    Which audiences will benefit most?
  • Researchers and practitioners working on information retrieval, embeddings, and metric learning.
  • Search and recommendation engineers interested in lightweight learned ranking functions.
  • Researchers interested in geometric representation learning and rigorous evaluation.
    Bio
    Organization: Target
    Their current research investigates relevance ranking with bi-encoders, learned distance functions, and the relationship between representation learning and scoring geometry.

We want to determine whether midpoint-conditioned directional weighting adds value beyond simpler global or scalar weighting. We also welcome help with independent evaluation datasets, initialization provenance, and practical integration into candidate-retrieval-and-reranking systems.
Matchmaking tags

  • I can help with: ideas / technique / critique
  • I need help with: dataset / collaborator / critique
  • I’d like to meet: researchers / practitioners / dataset owners / tool builders
    Topic/domain tags
    Information retrieval · Representation learning · Evaluation
All submissions

Comments

{{ gettext('Login to leave a comment') }}

{{ gettext('Post a comment…') }}
{{ gettext('New comment') }}
{{ formTitle }}

{{ errorMsg }}

{{ gettext('No comments posted yet') }}

Hosted by

The Fifth Elephant
The Fifth Elephant
Jumpstart better data engineering and AI futures