📝 Abstract:
Rank fusion is a critical mechanism for synthesizing outputs from multiple retrieval models, yet traditional heuristic and linear supervised methods struggle to generalize across diverse data environments. This thesis introduces EmbMerge, a novel, lightweight transformer-based framework that maps discrete ranking features—such as items, scores, models, and positional ranks—into continuous embedding spaces. To demonstrate its efficacy, EmbMerge is evaluated across two fundamentally distinct domains: implicit feedback Cross-Domain Recommendation (CDR) and dense semantic Information Retrieval (IR). Our empirical evaluation reveals a critical architectural dilemma. In the CDR pipeline, results on datasets like Digix and Amazon prove the network relies heavily on explicit Item Embeddings to map latent topological relationships. Furthermore, because the evaluation relies on a strict leave-one-out methodology, pointwise Binary Cross-Entropy (BCE) loss is mathematically indispensable to stabilize gradients under single-positive constraints. Conversely, for the dense IR domain, we propose a residual prediction architecture trained via an InfoNCE contrastive loss, which refines semantic priors by predicting localized deltas. Ablation studies on the zero-shot BEIR benchmark demonstrate that while Item Embeddings maximize in-domain retrieval, they severely impede domain transfer by memorizing the training corpus. Consequently, achieving robust zero-shot generalization necessitates omitting explicit item representations, forcing the network to rely purely on universal ranking heuristics to prevent dataset memorization. Ultimately, EmbMerge successfully outperforms traditional rank fusion baselines in both domains. However, this research establishes that true domain-independent fusion is not a rigid, universal algorithm; rather, it requires the dynamic calibration of network inputs and objective functions based on the topological mapping requirements and label distributions of the underlying base models.