Meta-learning systems and/or methods for error detection in structured data

A meta-learning based error detection system, comprising:

• at least one processor and a memory, the at least one processor being configured to at least:
• access, from a first data store, historical datasets that each include historical data and historical data attributes, as well as a base classifier for each historical attribute for each of the historical datasets, each base classifier being pre-trained to classify between clean and dirty data, wherein historical clustering features are determined for each historical attribute in each historical dataset, and wherein clusters of a cluster model are generated based on determined historical clustering features for each historical attribute in each historical dataset;
• access, from a second data store, a dirty dataset that includes dirty data and a plurality of dirty attributes;
• automatically generate, for each one of the plurality of dirty attributes in the dirty dataset, corresponding dirty clustering features;
• assign, based on the corresponding dirty clustering features that are generated, each one of the plurality of dirty attributes of the dirty dataset to one of the clusters of the cluster model, wherein the base classifier for each historical attribute is associated with at least one of the clusters of the cluster model;
• select, from among a plurality base classifiers and for each assigned one of the plurality of dirty attributes, the base classifier(s) that is/are associated with the cluster(s) in the cluster model that a corresponding dirty attribute was assigned;
• automatically generate, for each corresponding dirty attribute of the plurality of dirty attributes in the dirty dataset, corresponding, dirty features that describe data for the corresponding dirty attribute of the dirty dataset;
• selectively modify the automatically generated extracted dirty features to conform to expected inputs of the selected base classifiers;
• generate a set of meta-features by transforming the set of dirty features into the set of meta-features by using the at least one selected base classifier;
• obtain labels for the generated meta-features;
• generating, based on the labeled meta-features, a new meta-classifier for use against the dirty dataset;
• and
• predict, based on the newly generated meta-classifier and for a plurality of data samples of the dirty dataset, about whether a state of a respective data sample of the dirty dataset is erroneous.