Research & Projects

M.S Thesis

TRACE: Transparent Recognition of Affective Cues via Expression

Advisor: Dr. Felix Hamza-Lup · Georgia Southern University · Expected 2027

Microexpressions are brief, involuntary facial expressions lasting 40-200 milliseconds that emerge when individuals attempt to conceal emotions. Unlike conventional expressions, they reveal authentic affective states largely resistant to conscious control, making them a powerful signal for clinical diagnosis, security screening, behavioral analysis, and human-computer interaction.

TRACE develops a computational framework for automatic microexpression detection and classification. The approach focuses on precise mapping of Facial Action Units (FACS) to discrete emotional states, enabling models to capture causally meaningful facial dynamics rather than surface-level correlations.

The research aims to advance computing toward human-aware, transparent, and socially beneficial intelligent systems grounded in measurable, biologically meaningful facial behavior.

Research Projects

Topology-Aware Evolutionary Algorithm System to Solve Graph Coloring

Spring 2025 - Link to report

Developed a topology-aware evolutionary algorithm (EA) to address the NP-hard graph coloring problem, which involves assigning colors to graph vertices such that no adjacent vertices share the same color while minimizing the total number of colors used. The system leverages heuristic initializations (DSatur, RLF, greedy), adaptive mutation, and local search with tabu lists to navigate the complex search space.

Initially explored a two-pass approach with Graph Neural Network (GNN)-guided color probabilities, but experiments revealed the GNN was unnecessary - robust heuristic initializations and local search alone achieved equivalent performance. The streamlined single-pass EA was evaluated on DIMACS benchmark datasets, achieving optimal colorings for MYC (5/5) and most REG graphs (10/14), with near-optimal results on SGB queen graphs.

Cine Oracle
A Trope-based Movie Recommendation Tool

Fall 2024 - Link to presentation

Built a movie recommendation system that goes beyond conventional genre or rating-based methods by analyzing the storytelling tropes that make audiences love particular narratives. Inspired by the Open Syllabus project, CineOracle leverages the TV Tropes collaborative database - a catalog of recurring storytelling devices across film, TV, and books - to uncover why audiences connect with certain movies.

The system combines two complementary signal sources. Semantic embeddings are generated from trope descriptions using a sentence-transformer LLM (all-MiniLM-L6-v2, 384-dim), then reduced via a two-stage PCA → UMAP pipeline to 64 dimensions while preserving both global variance and local neighborhood structure. Structural embeddings are produced by constructing a heterogeneous graph of movies and tropes (nodes) connected by movie–trope associations (edges), then training Node2vec random walks to capture graph-level relationships in 64 dimensions.

Recommendations are generated by identifying important tropes for each movie via the graph embeddings, computing movie similarities based on shared trope patterns, and surfacing films with the most similar storytelling DNA.

Page updated

Google Sites