{Mukul Prasad Lei Liu Wei-Peng Chen Mehdi Bahrami}

Abstract

Data pre-processing is one of the key steps in creating machine learning pipelines for tabular data. One of the common data pre-processing operations implemented in AutoML systems is to encode categorical features as numerical features. Typically, this is implemented using a simple alphabetical sort on the categorical values, using functions such as OrdinalEncoder, LabelEncoder in Scikit-Learn and H2O. However, often there exist semantic ordinal relationships among the categorical values, such as: quality level (i.e., [’very good’ > ’good’ > ’normal’> ’poor’]), or month (i.e., [’Jan’< ’Feb’ < ’Mar’]). Such semantic relationships are not exploited by previous AutoML approaches. In this paper, we introduce BERT-Sort, a novel approach to semantically encode ordinal categorical values via zero-shot Masked Language Models (MLM) and apply it to AutoML for tabular data. We created a new benchmark of 42 features from 10 public data sets for sorting categorical ordinal values for the first time, where BERT-Sort significantly improves semantic encoding of ordinal values in comparison to the existing approaches with 27% improvement. We perform a comprehensive evaluation of BERT-Sort on different public MLMs, such as RoBERTa, XLM and DistilBERT. We alsocompare the performance of raw data sets against encoded data sets through BERT-Sort in different AutoML platforms including AutoGluon, FLAML, H2O, and MLJAR to evaluate the proposed approach in an end-to-end scenario.