Abstract
In the era of information overload, recommender systems are essential
tools for guiding user interactions and enhancing decision-making processes.
However, the increasing complexity and scale of data present challenges in
personalization, accuracy, and evaluation of recommendations. This thesis
addresses these challenges by introducing advanced methods and techniques
to enhance the performance and utility of recommender systems. The research
investigates both metric-based and non-metric approaches to tackle optimization
inefficiencies, embedding robustness, and evaluation gaps. The work
introduces a contextual grid triplet network that models user-item-context
relationships through a novel triplet input structure, effectively integrating
contextual information into recommendations. Leveraging pairwise learning—
a method for ranking positive items over negative ones based on their
relevance—the network captures intra-grid patterns (within individual user
or item grids) and inter-grid patterns (across user-item-context grids) to
improve contextual relevance and accuracy. To optimize learning, a hinge
distribution loss function is proposed, addressing the fixed-margin limitation
of traditional triplet hinge loss by dynamically adjusting margins based on
evolving positive and negative distance distributions. This ensures better handling
of hard triplets (positive items close to negatives in embedding space),
improving convergence and learning efficiency. For metric-based methods,
the thesis tackles the impostor problem—negative items that resemble positives
and disrupt embedding space structure—through the Weighted Batch
Approximate-Collaborative Metric Learning (WBACML) framework. This
approach dynamically evolves a target distance (a threshold derived from
negative item distributions) to enhance the separation of positive and negative
items. Non-metric methods like Weighted Approximate-Rank Pairwise
(WARP) are addressed by integrating inter-item associations and incorporating
the triangularity property (ensuring consistent distance relationships
across embeddings) through a Weighted Batch Approximate-Rank Pairwise
Loss (WBARP), improving similarity propagation and overall modeling accuracy.
Evaluation metrics are another key focus of this work. Traditional
metrics, such as Mean Average Precision (MAP) and Normalized Discounted
Cumulative Gain (NDCG), emphasize ranking accuracy but neglect rating
migration patterns—scenarios where relative rankings remain correct, but
predicted scores deviate significantly from true values, affecting user perception
and trust. To address this, the Standard Discounted Cumulative Squared
Error (SDCSE) metric is introduced, bridging predictive accuracy and ranking
utility by penalizing large deviations in predicted scores for highly ranked
items, thus ensuring alignment with user expectations. Through extensive
comparative and ablation studies, this thesis demonstrates the effectiveness
of its proposed methods across diverse datasets, significantly advancing the
robustness, accuracy, and evaluation of recommender systems. By tackling
context-aware modeling, optimization challenges, and evaluation limitations,
this work contributes to the development of more precise, transparent, and
user-aligned recommendations.