Techniques for improving the accuracy of automated predictions

One or more non-transitory computer readable media comprising instructions which, when executed by one or more hardware processors, cause performance of operations comprising:

• generating initial predictions for a population of individual prediction targets using a plurality of input parameters;
• for each individual prediction target in the population of individual prediction targets, obtaining actual performance data;
• identifying erroneous predictions by comparing initial predictions with actual performance data;
• generating, by one or more computing devices, a feature representation for each individual prediction target by applying a trained neural network to the input parameters;
• calculating pairwise distances between individual prediction targets using the feature representations;
• forming one or more clusters of individual prediction targets based on the pairwise distances, each cluster comprising:
• a core set of underperforming entities that are individual prediction targets associated with erroneous predictions in the initial predictions, wherein the pairwise distances between the underperforming entities in the core set are shorter than the average distances between the underperforming entities and other prediction targets;
• and
• a boundary set of performing entities that are individual prediction targets associated with accurate predictions in the initial predictions, and positioned in the periphery of the core set;
• identifying at least one input parameter that differentiates the core set from the boundary set within each cluster;
• determining a critical value range for each identified input parameter, wherein values for the core set fall within the critical range, and values for the boundary set fall outside the critical range;
• modifying, by the one or more computing devices, a decision-making threshold or a model parameter of the automated prediction mechanism based on the identified input parameters and corresponding critical value ranges.

A system comprising:

• at least one device including a hardware processor;
• the system being configured to perform operations comprising:
• generating initial predictions for a population of individual prediction targets using a plurality of input parameters;
• for each individual prediction target in the population of individual prediction targets, obtaining actual performance data;
• identifying erroneous predictions by comparing initial predictions with actual performance data;
• generating, by one or more computing devices, a feature representation for each individual prediction target by applying a trained neural network to the input parameters;
• calculating pairwise distances between individual prediction targets using the feature representations;
• forming one or more clusters of individual prediction targets based on the pairwise distances, each cluster comprising:
• a core set of underperforming entities that are individual prediction targets associated with erroneous predictions in the initial predictions, wherein the pairwise distances between the underperforming entities in the core set are shorter than the average distances between the underperforming entities and other prediction targets;
• and
• a boundary set of performing entities that are individual prediction targets associated with accurate predictions in the initial predictions, and positioned in the periphery of the core set;
• identifying at least one input parameter that differentiates the core set from the boundary set within each cluster;
• determining a critical value range for each identified input parameter, wherein values for the core set fall within the critical range, and values for the boundary set fall outside the critical range;
• modifying, by the one or more computing devices, a decision-making threshold or a model parameter of the automated prediction mechanism based on the identified input parameters and corresponding critical value ranges.