If the complete set of possible feature values for a column is unknown, categorical_column_with_hash_bucket is the appropriate function to use. This function hashes each possible value in the feature column to an integer ID as they are encountered during training, eliminating the need to know all possible values upfront. For cases where all possible values are known and limited in number, categorical_column_with_vocabulary_list would be more suitable, as it assigns an auto-incremental ID starting from 0 to each key in the list. Reference: TensorFlow Documentation