The identification of synergistic drug combinations remains a significant challenge in oncology due to the large amount of existing drugs and complex interactions between these drugs. In this work, we propose an active learning framework applied to the NCI ALMANAC dataset to efficiently uncover promising drug pairs that conventional screening methods might overlook due to lack of time and resources to handle these nearly countless combinations. Building on established greedy sampling strategies—such as GSx, which selects samples based on maximal minimum distance in the input space, and GSy, which focuses on output diversity—we introduce modifications to potentially enhance sample selection diversity and predictive performance. First, we explore replacing the traditional greatest minimum distance criterion with a greatest average distance metric, hypothesizing that this adjustment captures the overall variability in the data differently than the traditional method. Second, we redefine the improved greedy sampling (iGS) approach by standardizing the distance metrics from both the input (GSx) and output (GSy) spaces using Z-score normalization (or alternative standardization methods) prior to their aggregation, rather than combining them multiplicatively. We conduct a comprehensive comparative analysis against traditional methods to evaluate improvements in model convergence, prediction accuracy, and the ability to identify rare but potent drug combinations. We also explore other active learning strategies as Query By Committee (QBC) and others. Our preliminary findings suggest that these tailored active learning techniques offer a promising pathway toward more efficient and insightful exploration of high-dimensional drug interaction landscapes.

Relative Risk (RR) is a highly interpretable parameter in epidemiology and biostatistics, based on both binary input and outcome. It is frequently used in vaccine development to measure the relative efficacy between two treatment groups.
Researchers are often tempted to use generalized linear models (GLMs) to estimate the logarithmic RR with respect to a set of baseline covariates. However, this approach has inherent flaws, as GLMs do not account for variation dependence in Relative Risk on its nuisance parameters. Richardson et al. have developed an unconstrained and variation-independent doubly robust nuisance model using the log Odds Product (OP).
To expand on this work, we will explore alternative nuisance models—both those developed by us and those from other researchers—and compare their computational robustness to that of the log Odds Product (OP).
Additionally, using the brm R package (which streamlines the methods proposed by Richardson et al.), we will analyze a dataset where Relative Risk serves as the target of inference and compare these results to those obtained using regression methods.

Music Emotion Recognition (MER) is a prominent area of research in engineering and data science. With the development of music feature extraction systems, the focus has been selecting relevant features and building predictive models based on them. This study aims to build a small structure that can extract music features, and compute the parameters used in classifying emotions. In this study, Marsyas is used to extract music features, and then LASSO regression model is applied to estimate the valence and arousal with the music features. The calculated valence and arousal are used to classify the music emotion based on Russell's Circumplex Model. This approach provides a view of the whole process of classifying music emotion, from extracting the basic features to calculating the parameters, to classifying the emotion.