Contact
Lehrstuhl für Datenbanksysteme und Data Mining
Oettingenstraße 67
80538 München
Germany
Room:
F 104
Phone:
+49-89-2180-9319
Email:
aljoud@dbs.ifi.lmu.de
Website:
https://www.linkedin.com/in/mamdouh-aljoud-70386b112/
Office hours:
By appointment
I am interested in Data-centric AI and the applications of AutoML in unsupervised settings such as Clustering. Currently I am investigating the use of Meta-learning in deep clustering algorithms.
I'm looking for Bachelor and Master theses student. If you find my interests appealing, please reach out to me via Email and include your CV and transcript.
Available Theses:
- Extrapolating Subset-Based Neural Architecture Search for Efficient Deep Clustering Optimization (BA)
This thesis explores the feasibility of extrapolating Neural Architecture Search (NAS) optimizations from a small subset to an entire dataset within the context of deep clustering algorithms. Deep clustering, which integrates deep learning with clustering objectives, can benefit significantly from NAS by identifying optimal model architectures tailored to the dataset. This research investigates whether the NAS-derived architecture optimized on a smaller subset retains its effectiveness when scaled to the entire dataset, potentially offering a method to accelerate NAS processes. Through systematic experimentation, we aim to assess the transferability of the optimized architecture, its impact on clustering performance, and any limitations associated with subset-based optimizations. The findings will contribute to more efficient NAS methodologies in deep clustering, enhancing computational efficiency while maintaining performance standards.
- Evaluating Internal Clustering Metrics for Effective Architecture Recommendation in Deep Clustering (BA)
This thesis benchmarks internal clustering metrics to evaluate their effectiveness in recommending architectures that align with the best-performing choices according to an external metric, specifically unsupervised clustering accuracy (UCA). This study systematically compares a range of internal metrics, assessing their alignment with UCA to determine which internal metrics most reliably recommend the highest-performing architectures. By examining the consistency between internal metric recommendations and actual clustering performance, this research aims to identify metrics that serve as accurate proxies for architecture selection in deep clustering. The findings provide valuable insights into metric selection for optimizing deep clustering methods, offering a pathway toward more efficient and effective architecture recommendation in unsupervised learning.
- Neural Architecture Search for Autoencoder-Based Deep Clustering on Tabular Data (MA)
This thesis explores the application of Neural Architecture Search (NAS) to identify optimal autoencoder architectures tailored for deep clustering algorithms, specifically when applied to tabular data. The research begins with a selection of tabular datasets and examines which statistical features best characterize these datasets in ways relevant to clustering performance. Using these insights, a search space for suitable autoencoder architectures is defined. Finally, a recommendation system is developed to suggest the most promising architectures based on dataset characteristics