In some embodiments, a number of inputs for a machine learning model may be dynamically scaled based on a record update. In some embodiments, an account record may be obtained based on an obtained record update, and a first model may determine a category value based on the account record. A graph that includes a node mapping to the account record may be obtained. In some embodiments, a threshold may be selected based on the category value, and the graph may be traversed via a set of paths to select nodes of the graph, such that the paths are based on the threshold. Record values associated with the set of nodes may be provided to a second model to determine whether the record update satisfies a set of criteria. If so, the record update may be stored in an aggregated set of record updates.