Using Polars to Improve String Similarity Performance in Python
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
IntroductionString similarity is central to textual record linkage, and is often calculated with Levenshtein distance or Jaro-Winkler distance. In polars-strsim, we leverage the Polars DataFrame interface to surpass all existing Python libraries in computing these distances. Polars-strsim is especially useful for larger-than-memory datasets. Objectives and ApproachPolars is a library built to analyse and manipulate tabular data. It is written in Rust, but can be called from Python for ease of use. Polars-strsim implements the Levenshtein and Jaro-Winkler algorithms as a Polars extension. We compare performance against nine other Python libraries on three million pairs of first names. ResultsWhen computing Levenshtein/Jaro-Winkler distance for this benchmark, polars-strsim is respectively 5x/4x faster than the current fastest Python alternative and 47x/35x faster than the median. The Polars streaming engine can be seamlessly used to manipulate larger than memory datasets. One practical application of this in record linkage is to find all pairs of records in a dataset where the Levenshtein distance computed across one (or more) fields is below a certain threshold (e.g., during blocking). With polars-strsim, this calculation can be performed in Python with CPU parallelism even if the input and/or output doesn’t fit in memory. Conclusions/ImplicationsThis work introduces a powerful library for record linkage practitioners. Polars-strsim provides the convenience of Python, the performance of a strongly/statically typed compiled language, and the full power of the Polars streaming engine when computing string similarity, and can be updated to implement additional similarity algorithms.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.002 | 0.010 |
| Open science | 0.004 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it