Projects

This project presents a Hybrid Minkowski–Log–Cosh (MLC) loss function designed to improve the robustness of LSTM models on noisy, outlier-contaminated time-series data.

By combining a Minkowski-style exponent with the smoothness of the Log-Cosh function, MLC addresses the limitations of traditional losses like MSE and MAE, offering better stability and outlier resistance.

Highlights

• Improved stability and accuracy under noise
• Matched standard loss functions in computational cost
• Outperformed MSE/MAE in long-range forecasting

This research led to a first-author publication in IEEE Access (Q1).

📄 Read the paper
💻 View code on GitHub

Citation (APA)

Simsoba, K.-A., Oscar, N., & Mageto, T. (2025). A Hybrid Minkowski-Log-Cosh Loss Function for Robust LSTM-Based Time Series Forecasting. IEEE Access, 13, 187307–187319. https://doi.org/10.1109/ACCESS.2025.3626795

This project explores scalable recommendation systems based on matrix factorization, with a focus on Alternating Least Squares (ALS) for large-scale collaborative filtering.

The work progresses from baseline matrix factorization models to content-aware and hybrid recommender systems, addressing key challenges such as data sparsity, scalability, and efficient optimization in large user–item interaction matrices. Experiments are conducted on the MovieLens 32M dataset, enabling rigorous evaluation of convergence behavior, recommendation quality, and computational performance.

Ongoing work focuses on extending the ALS framework with auxiliary information to improve recommendation accuracy while maintaining scalability, with the goal of preparing a journal or arXiv manuscript.

• Project report / materials:
  Google Drive – Project Documentation

This project investigates generative and multimodal AI models for breast cancer diagnosis and staging, integrating histopathology images with structured clinical records.

The work addresses limitations of unimodal approaches by exploring multimodal learning frameworks that jointly model visual and clinical data to improve diagnostic accuracy and clinical relevance. Current efforts focus on model prototyping, modality fusion strategies, and evaluation protocols, with emphasis on robustness and interpretability.

This project is an ongoing research project within my second Master’s in Artificial Intelligence for Science, with the objective of producing a peer-reviewed publication in the medical AI domain.
Status: Ongoing.

This project examines the impact of online learning on university students’ academic performance during the COVID-19 pandemic using a statistical and data-driven approach.

Primary data were collected from students at the Universite de Kara, capturing academic outcomes, access to digital devices, internet quality, and learning conditions during periods of remote instruction. The analysis employed descriptive and inferential statistical methods to identify relationships between learning environments and student performance.

The findings provide empirical insights into the challenges of remote education in resource-constrained settings and inform recommendations for improving equity and resilience in higher education systems.

This work constitutes a research project within my second Master’s training, building on my background in applied statistics and serving as a foundation for subsequent research in educational and social data analysis.
Status: Ongoing.