Abstract: Techniques described herein involve optimizing resource utilization in machine learning models. Embodiments include receiving a plurality of data sets and altering the plurality of data sets to produce an altered plurality of data sets, comprising: calculating similarity scores between data sets in the plurality of data sets; if a particular similarity score exceeds a threshold value, removing, from the plurality of data sets, a corresponding data set associated with the particular similarity score; determining anomalous data in a given data set of the plurality of data sets based on an anomaly detection technique; removing data not matching the anomalous data from the given data set; and reducing a size of one or more data sets in the plurality of data sets based on a type of data in each data set. Embodiments further include performing an action using a machine learning model based on the altered plurality of data sets.