Most outlier detection algorithms expect either purely categorical or purely numeric data, yet real-world datasets typically contain mixed types. When working with numeric detectors like Isolation Forest or Local Outlier Factor, categorical columns must be converted to numerical representations that preserve distance relationships between data points.

One-hot encoding dominates prediction tasks, but proves problematic for outlier detection because it creates sparse, high-dimensional representations that distort geometric relationships. Count encoding offers a compelling alternative by replacing categories with their occurrence frequencies, which better maintains meaningful distance calculations in the transformed space.

Target encoding fails in outlier detection contexts since these problems lack target variables—unlike churn prediction or sales forecasting where historical outcomes guide encoding decisions. The unsupervised nature of outlier detection eliminates entire families of encoding techniques that rely on labeled data.

Python libraries like Category Encoders provide numerous methods, though many require target columns unavailable in outlier detection workflows. Practitioners should prioritize one-hot and count encoding approaches that work without supervision, accepting that mixed-type data inherently complicates outlier identification compared to homogeneous datasets.