Similarity encoding for learning with dirty categorical variables
This addresses a practical issue in machine learning for non-curated data, offering an incremental improvement over existing encoding methods for dirty categories.
The paper tackles the problem of learning from dirty categorical variables with high cardinality and redundancy by proposing similarity encoding, a generalization of one-hot encoding that uses similarities across categories. Results on seven real-world datasets show significant prediction gains compared to methods like one-hot encoding and bag of character n-grams.
For statistical learning, categorical variables in a table are usually considered as discrete entities and encoded separately to feature vectors, e.g., with one-hot encoding. "Dirty" non-curated data gives rise to categorical variables with a very high cardinality but redundancy: several categories reflect the same entity. In databases, this issue is typically solved with a deduplication step. We show that a simple approach that exposes the redundancy to the learning algorithm brings significant gains. We study a generalization of one-hot encoding, similarity encoding, that builds feature vectors from similarities across categories. We perform a thorough empirical validation on non-curated tables, a problem seldom studied in machine learning. Results on seven real-world datasets show that similarity encoding brings significant gains in prediction in comparison with known encoding methods for categories or strings, notably one-hot encoding and bag of character n-grams. We draw practical recommendations for encoding dirty categories: 3-gram similarity appears to be a good choice to capture morphological resemblance. For very high-cardinality, dimensionality reduction significantly reduces the computational cost with little loss in performance: random projections or choosing a subset of prototype categories still outperforms classic encoding approaches.