Machine Learning

To provide machine learning and data science experts with a more robust dataset for model training, the well-known Palmer Penguins dataset has been expanded from its original 344 rows to 100,000 rows. This substantial increase was achieved using an adversarial random forest technique, effectively generating additional synthetic data while maintaining key patterns and features. The method achieved an impressive accuracy of 88%, ensuring the expanded dataset remains realistic and suitable for classification tasks.

This dataset offers both Channel State Information (CSI) and Beamforming Feedback Information (BFI) data for human activity classification, featuring 20 distinct activities performed by three subjects across three environments. Collected in both line-of-sight (LoS) and non-line-of-sight (NLoS) scenarios, this dataset enables researchers to explore the complementary roles of CSI and BFI in activity recognition and environmental characterization.

As various modalities of genomic data are accumulating, methods to integrate across multi-omics datasets are becoming important. Error-correcting output codes (ECOC) is an ensemble learning strategy for solving a multiclass problem thru a decoding process that aggregates the predictions of multiple classifiers. Thus, it lends itself naturally to aggregating predictions across multiple views as well. We applied the ECOC to multi-view learning to see if this strategy can enhance classifier performance as compared to traditional techniques.

Hyperthermia is a thermal treatment for cancer which kills cancer cells by applying heat. Among different heating approaches, magnetic hyperthermia uses heat generation of magnetic nanoparticles to achieve local treatment. Magnetic nanoparticles induce heat when stimulated by an alternating magnetic field, and the heating efficiency depends on multiple parameters including magnetic properties of magnetic nanoparticles and magnetic field used to stimulate the nanoparticles.