Show/Hide Menu
Hide/Show Apps
Logout
Türkçe
Türkçe
Search
Search
Login
Login
OpenMETU
OpenMETU
About
About
Open Science Policy
Open Science Policy
Open Access Guideline
Open Access Guideline
Postgraduate Thesis Guideline
Postgraduate Thesis Guideline
Communities & Collections
Communities & Collections
Help
Help
Frequently Asked Questions
Frequently Asked Questions
Guides
Guides
Thesis submission
Thesis submission
MS without thesis term project submission
MS without thesis term project submission
Publication submission with DOI
Publication submission with DOI
Publication submission
Publication submission
Supporting Information
Supporting Information
General Information
General Information
Copyright, Embargo and License
Copyright, Embargo and License
Contact us
Contact us
Transformer-Based Multi-Model Rank Fusion for Cross-Domain Recommendation and Dense Information Retrieval
Download
mehmet_erdeniz_aydoğdu_tez.pdf
Date
2026-9-04
Author
Aydoğdu, Mehmet Erdeniz
Metadata
Show full item record
Item Usage Stats
79
views
0
downloads
Cite This
Rank fusion is an effective approach for combining outputs from multiple ranking models, though existing methods often have difficulty generalizing across different data environments. This thesis introduces EmbMerge, a transformer-based framework that frames rank fusion as a sequence modeling task by mapping discrete ranking features—items, scores, models, and positional ranks—into continuous embedding spaces. The framework is evaluated across two distinct domains: implicit feedback Cross-Domain Recommendation (CDR) and dense semantic Information Retrieval (IR). To handle differing domain constraints, EmbMerge introduces two scoring architectures: a Standard model for global list scaling, and a Direct Prediction model for localized fusion. Our empirical evaluation highlights key architectural trade-offs. In the CDR pipeline, experiments on datasets like Digix and Amazon show that the network relies heavily on explicit item embeddings and absolute scoring, optimized via pointwise Binary Cross-Entropy (BCE) loss. For the dense IR domain, however, the Direct Prediction architecture employs a residual "neural delta" approach trained via InfoNCE contrastive loss. Ablation studies on the zero-shot BEIR benchmark indicate that although item embeddings improve in-domain retrieval, they can hinder domain transfer due to corpus memorization. Thus, zero-shot generalization is more effectively achieved by excluding explicit item representations. Overall, EmbMerge demonstrates strong performance against traditional baselines. This research suggests that domain-independent fusion benefits significantly from dynamically adapting network inputs, scoring architectures, and objective functions to the specific structural requirements of the underlying base models.
Subject Keywords
Rank Fusion
,
Cross-Domain Recommendation
,
Information Retrieval
,
Learning-To-Rank
URI
https://hdl.handle.net/11511/121145
Collections
Graduate School of Natural and Applied Sciences, Thesis
Citation Formats
IEEE
ACM
APA
CHICAGO
MLA
BibTeX
M. E. Aydoğdu, “Transformer-Based Multi-Model Rank Fusion for Cross-Domain Recommendation and Dense Information Retrieval,” M.S. - Master of Science, Middle East Technical University, 2026.