Combination Response Surface Modeling for Sparse Interaction Data

The patent applies preliminary action by using active learning to selectively identify the most informative experiments to perform first. The system prioritizes experiments that will provide maximum information gain for model training, rather than performing dense sampling across all possible conditions. This allows the model to achieve good predictive performance with a smaller, more strategically selected set of experiments, reducing both labor and cost while maintaining prediction accuracy.

The patent applies segmentation by decomposing the complex high-dimensional prediction problem into multiple manageable components. The system uses separate models or model components for different aspects of the prediction task, such as handling different types of item combinations or different response dimensions separately. This modular approach makes the overall system more tractable while maintaining the ability to handle high dimensionality through composition of simpler models.

The patent applies feedback through active learning loops where the machine learning model continuously evaluates its own predictive uncertainty and uses this information to guide subsequent experiment selection. The system identifies regions of chemical space or combination spaces where predictions are most uncertain and prioritizes experiments in those regions. This feedback-driven approach efficiently fills data gaps and reduces sparsity by focusing experimental resources on the most informative under-sampled regions, improving comprehensiveness without requiring exhaustive sampling.