Polars for Data Science

Discord | Documentation | User Guide | Want to Contribute?

pip install polars-ds

PDS (polars_ds)

PDS is a modern data science package that

1. is fast and furious
2. is small and lean, with minimal dependencies
3. has an intuitive and concise API (if you know Polars already)
4. has dataframe friendly design
5. and covers a wide variety of data science topics, such as simple statistics, linear regression, string edit distances, tabular data transforms, feature extraction, traditional modelling pipelines, model evaluation metrics, etc., etc..

It stands on the shoulders of the great Polars dataframe. You can see examples. Here are some highlights!

Parallel ML Metrics Calculation

import polars as pl
import polars_ds as pds
# Parallel evaluation of multiple ML metrics on different segments of data
df.lazy().group_by("segments").agg( 
    # any other metrics you want in here
    pds.query_roc_auc("actual", "predicted").alias("roc_auc"),
    pds.query_log_loss("actual", "predicted").alias("log_loss"),
).collect()

shape: (2, 3)
┌──────────┬──────────┬──────────┐
│ segments ┆ roc_auc  ┆ log_loss │
│ ---      ┆ ---      ┆ ---      │
│ str      ┆ f64      ┆ f64      │
╞══════════╪══════════╪══════════╡
│ a        ┆ 0.497745 ┆ 1.006438 │
│ b        ┆ 0.498801 ┆ 0.997226 │
└──────────┴──────────┴──────────┘

Quick and simple modeling on the fly (non-persistent)

E.g. running a quick linear regression and see the predictions and residues:

df.select(pds.lin_reg(pl.col("x1"), pl.col("x2"), target=pl.col("y"), add_bias=False, return_pred=True))

shape: (10_000, 1)
┌───────────────────────┐
│ lr_pred               │
│ ---                   │
│ struct[2]             │
╞═══════════════════════╡
│ {-0.3121,0.392769}    │
│ {-0.459507,-0.048989} │
│ {-0.469473,-0.215709} │
│ {-0.243764,-0.707016} │
│ {-0.511278,-0.785299} │
│ …                     │
└───────────────────────┘

Generating polynomial features and display a statsmodels-like regression summary:

import polars_ds as pds
from polars_ds.pipeline.transforms import polynomial_features
# If you want the underlying computation to be done in f32, set pds.config.LIN_REG_EXPR_F64 = False
df.select(
    pds.lin_reg_report(
        *(
            ["x1", "x2", "x3"] +
            polynomial_features(["x1", "x2", "x3"], degree = 2, interaction_only=True)
        )
        , target = pl.col("target")
        , add_bias = False
    ).alias("result")
).unnest("result")

┌──────────┬───────────┬──────────┬───────────┬───────┬───────────┬──────────┬──────────┬──────────┐
│ features ┆ beta      ┆ std_err  ┆ t         ┆ p>|t| ┆ 0.025     ┆ 0.975    ┆ r2       ┆ adj_r2   │
│ ---      ┆ ---       ┆ ---      ┆ ---       ┆ ---   ┆ ---       ┆ ---      ┆ ---      ┆ ---      │
│ str      ┆ f64       ┆ f64      ┆ f64       ┆ f64   ┆ f64       ┆ f64      ┆ f64      ┆ f64      │
╞══════════╪═══════════╪══════════╪═══════════╪═══════╪═══════════╪══════════╪══════════╪══════════╡
│ x1       ┆ 0.26332   ┆ 0.000315 ┆ 835.68677 ┆ 0.0   ┆ 0.262703  ┆ 0.263938 ┆ 0.971087 ┆ 0.971085 │
│          ┆           ┆          ┆ 8         ┆       ┆           ┆          ┆          ┆          │
│ x2       ┆ 0.413824  ┆ 0.000311 ┆ 1331.9883 ┆ 0.0   ┆ 0.413216  ┆ 0.414433 ┆ 0.971087 ┆ 0.971085 │
│          ┆           ┆          ┆ 32        ┆       ┆           ┆          ┆          ┆          │
│ x3       ┆ 0.113688  ┆ 0.000315 ┆ 361.29924 ┆ 0.0   ┆ 0.113072  ┆ 0.114305 ┆ 0.971087 ┆ 0.971085 │
│ x1*x2    ┆ -0.097272 ┆ 0.000543 ┆ -179.0377 ┆ 0.0   ┆ -0.098337 ┆ -0.09620 ┆ 0.971087 ┆ 0.971085 │
│          ┆           ┆          ┆ 76        ┆       ┆           ┆ 7        ┆          ┆          │
│ x1*x3    ┆ -0.097266 ┆ 0.000542 ┆ -179.4486 ┆ 0.0   ┆ -0.098329 ┆ -0.09620 ┆ 0.971087 ┆ 0.971085 │
│          ┆           ┆          ┆ 32        ┆       ┆           ┆ 4        ┆          ┆          │
│ x2*x3    ┆ -0.097987 ┆ 0.000542 ┆ -180.7579 ┆ 0.0   ┆ -0.099049 ┆ -0.09692 ┆ 0.971087 ┆ 0.971085 │
│          ┆           ┆          ┆ 6         ┆       ┆           ┆ 4        ┆          ┆          │
└──────────┴───────────┴──────────┴───────────┴───────┴───────────┴──────────┴──────────┴──────────┘

Other available simple models (non-persistent):

• Normal Linear Regression (pds.lin_reg)
• Lasso, Ridge, Elastic Net (pds.lin_reg, use l1_reg, l2_reg arguments)
• Rolling linear regression with skipping (pds.rolling_lin_reg)
• Recursive linear regression (pds.recursive_lin_reg)
• Non-negative linear regression (pds.lin_reg, set positive = True)
• Statsmodel-like linear regression table (pds.lin_reg_report)
• f32 support (pds.Config.LIN_REG_EXPR_F64 = False)
• binary logistic regression with L1, L2 parameters (pds.logistic_reg, doesn't work with F32 yet.)

Distances

Various string distances:

df.select( # Column "word", compared to string in pl.lit(). It also supports column vs column comparison
    pds.str_leven("word", pl.lit("asasasa"), return_sim=True).alias("Levenshtein"),
    pds.str_osa("word", pl.lit("apples"), return_sim=True).alias("Optimal String Alignment"),
    pds.str_jw("word", pl.lit("apples")).alias("Jaro-Winkler"),
)

Array, list distances:

df = pl.DataFrame({
    "x": [[1,2,3], [4,5,6]]
    , "y": [[0.5, 0.2, 0.3], [4.0, 5.0, 6.1]]
})

df.select(
    x = pl.col('x').cast(pl.Array(inner=pl.Float64, shape=3))
    , y = pl.col('y').cast(pl.Array(inner=pl.Float64, shape=3))
).lazy().select(
    pds.arr_sql2_dist('x', 'y')
).collect()

shape: (2, 1)
┌───────┐
│ x     │
│ ---   │
│ f64   │
╞═══════╡
│ 10.78 │
│ 0.01  │
└───────┘

Replace arr_sql2_dist with list_sql2_dist. Note: sql2 stands for squared l2 distance, which is the same as squared euclidean distance.

In-dataframe statistical tests

df.group_by("market_id").agg(
    pds.ttest_ind("var1", "var2", equal_var=False).alias("t-test"),
    pds.chi2("category_1", "category_2").alias("chi2-test"),
    pds.f_test("var1", group = "category_1").alias("f-test")
)

shape: (3, 4)
┌───────────┬──────────────────────┬──────────────────────┬─────────────────────┐
│ market_id ┆ t-test               ┆ chi2-test            ┆ f-test              │
│ ---       ┆ ---                  ┆ ---                  ┆ ---                 │
│ i64       ┆ struct[2]            ┆ struct[2]            ┆ struct[2]           │
╞═══════════╪══════════════════════╪══════════════════════╪═════════════════════╡
│ 0         ┆ {2.072749,0.038272}  ┆ {33.487634,0.588673} ┆ {0.312367,0.869842} │
│ 1         ┆ {0.469946,0.638424}  ┆ {42.672477,0.206119} ┆ {2.148937,0.072536} │
│ 2         ┆ {-1.175325,0.239949} ┆ {28.55723,0.806758}  ┆ {0.506678,0.730849} │
└───────────┴──────────────────────┴──────────────────────┴─────────────────────┘

Making Polars More Convenient

import polars_ds as pds
df = pl.DataFrame({
    "group": ['A', 'A', 'B', 'B', 'A']
    , "a": [1, 2, 3, 4, 5]
    , "b": [4, 1, 99, 12, 33]
})
df.group_by("group").agg(
    *pds.E(['a', 'b'], ["min", "max", "n_unique", "len"])
)

shape: (2, 8)
┌───────┬───────┬───────┬───────┬───────┬────────────┬────────────┬─────────┐
│ group ┆ a_min ┆ b_min ┆ a_max ┆ b_max ┆ a_n_unique ┆ b_n_unique ┆ __len__ │
│ ---   ┆ ---   ┆ ---   ┆ ---   ┆ ---   ┆ ---        ┆ ---        ┆ ---     │
│ str   ┆ i64   ┆ i64   ┆ i64   ┆ i64   ┆ u32        ┆ u32        ┆ u32     │
╞═══════╪═══════╪═══════╪═══════╪═══════╪════════════╪════════════╪═════════╡
│ A     ┆ 1     ┆ 1     ┆ 5     ┆ 33    ┆ 3          ┆ 3          ┆ 3       │
│ B     ┆ 3     ┆ 12    ┆ 4     ┆ 99    ┆ 2          ┆ 2          ┆ 2       │
└───────┴───────┴───────┴───────┴───────┴────────────┴────────────┴─────────┘

Streamable Tabular Machine Learning Data Transformation Pipeline

See SKLEARN_COMPATIBILITY for more details.

import polars as pl
import polars.selectors as cs
from polars_ds.pipeline import Pipeline, Blueprint

bp = (
    Blueprint(df, name = "example", target = "approved", lowercase=True) # You can optionally 
    .filter(pl.col("city_category").is_not_null())
    .linear_impute(features = ["var1", "existing_emi"], target = "loan_period") 
    .impute(["existing_emi"], method = "median")
    .append_expr( # generate some features
        pl.col("existing_emi").log1p().alias("existing_emi_log1p"),
        pl.col("loan_amount").log1p().alias("loan_amount_log1p"),
        pl.col("loan_amount").clip(lower_bound = 0, upper_bound = 1000).alias("loan_amount_clipped"),
        pl.col("loan_amount").sqrt().alias("loan_amount_sqrt"),
        pl.col("loan_amount").shift(-1).alias("loan_amount_lead_1") # shift(-1) is a lead transform
    )
    .scale( # target is numerical, but will be excluded automatically because bp is initialzied with a target
        cs.numeric().exclude(["var1", "existing_emi_log1p"]), method = "standard"
    ) # Scale the columns up to this point. The columns below won't be scaled
    .append_expr(
        # Add missing flags
        pl.col("employer_category1").is_null().cast(pl.UInt8).alias("employer_category1_is_missing")
    )
    .one_hot_encode("gender", drop_first=True)
    .woe_encode("city_category") # No need to specify target because we initialized bp with a target
    .target_encode("employer_category1", min_samples_leaf = 20, smoothing = 10.0) # same as above
)

print(bp)

pipe:Pipeline = bp.materialize()
# Check out the result in our example notebooks! (examples/pipeline.ipynb)
df_transformed = pipe.transform(df)
df_transformed.head()

Since Polars >=1.34 supports collect_batches(), you can also use this to perform batched machine learning

for df_batch in pipe.transform(df, return_lazy=True).collect_batches():
    X_batch, y_batch = your_function_to_turn_df_batch_into_model_inputs(df_batch)
    ml_model.update(X_batch, y_batch)

See pipeline examples for more details and caveats.

Nearest Neighbors Related Queries

Get all neighbors within radius r, call them best friends, and count the number. Due to limitations, this currently doesn't preserve the index, and is not fast when k or dimension of data is large.

df.select(
    pl.col("id"),
    pds.query_radius_ptwise(
        pl.col("var1"), pl.col("var2"), pl.col("var3"), # Columns used as the coordinates in 3d space
        index = pl.col("id"),
        r = 0.1, 
        dist = "sql2", # squared l2
        parallel = True
    ).alias("best friends"),
).with_columns( # -1 to remove the point itself
    (pl.col("best friends").list.len() - 1).alias("best friends count")
).head()

shape: (5, 3)
┌─────┬───────────────────┬────────────────────┐
│ id  ┆ best friends      ┆ best friends count │
│ --- ┆ ---               ┆ ---                │
│ u32 ┆ list[u32]         ┆ u32                │
╞═════╪═══════════════════╪════════════════════╡
│ 0   ┆ [0, 811, … 1435]  ┆ 152                │
│ 1   ┆ [1, 953, … 1723]  ┆ 159                │
│ 2   ┆ [2, 355, … 835]   ┆ 243                │
│ 3   ┆ [3, 102, … 1129]  ┆ 110                │
│ 4   ┆ [4, 1280, … 1543] ┆ 226                │
└─────┴───────────────────┴────────────────────┘

Compatibility

Under some mild assumptions, (e.g. columns implement to_numpy()), PDS works with other eager dataframes. For example, with Pandas:

from polars_ds.compat import compat as pds2

df_pd["linear_regression_result"] = pds2.lin_reg(
    df_pd["x1"], df_pd["x2"], df_pd["x3"],
    target = df_pd["y"],
    return_pred = True
)
df_pd

The magic here is the compat module and the fact that most eager dataframes implement the array protocal.

Other

Other common numerical functions such as: pds.convolve, pds.query_r2, pds.principal_components, etc. See our docs for more information.

Getting Started

import polars_ds as pds

To make full use of the Diagnosis module, do

pip install "polars_ds[plot]"

How Fast is it?

Feel free to take a look at our benchmark notebook!

Generally speaking, the more expressions you want to evaluate simultaneously, the faster Polars + PDS will be than Pandas + (SciPy / Sklearn / NumPy). The more CPU cores you have on your machine, the bigger the time difference will be in favor of Polars + PDS.

Road Map

1. All yours. Just let me know, and I will evaluate and get back to you!

Minimum Polars Support + Streaming Compatibility

The core package depends on Python Polars >=1.4.0, with the exclusions listed in pyproject.toml. The source compatibility matrix pins Polars 1.44.2 and 2.0.0, and runs the changed compatibility contracts on the minimum version, 1.4.0. Python 3.9 uses Polars 1.x; Polars 2 requires Python >=3.10.

Compatibility regressions exercise lazy auto, streaming and in-memory execution, including nulls, single-row batches, multiple chunks and grouped expressions. Whole-column plugins may materialize their input inside a streaming query; this does not make them bounded-memory algorithms. Fixed-seed random expressions use the complete input or group so batch boundaries do not restart the sequence. Polars large-index builds are not supported.

The Pipeline explode transform retains empty-list and null input rows by default on both versions, including previously serialized pipelines. Statistical binning preserves the linear quantiles and interval labels of Polars 1 via the still-available qcut API; bin_quantiles uses different breakpoints and is not a numerical replacement.

Build From Source

The guide here is not specific to LTS CPU, and can be used generally.

The best advice for LTS CPU is that you should compile the package yourself. First clone the repo and make sure Rust is installed on the system. Create a python virtual environment and install maturin in it. Next set the RUSTFLAG environment variable. The official polars-lts-cpu features are the following:

RUSTFLAGS=-C target-feature=+sse3,+ssse3,+sse4.1,+sse4.2,+popcnt,+cmpxchg16b

If you simply want to compile from source, you may set target cpu to native, which autodetects CPU features.

RUSTFLAGS=-C target-cpu=native

If you are compiling for LTS CPU, then in pyproject.toml, update the polars dependency to polars-lts-cpu:

polars >= 1.4.0 # polars-lts-cpu >= 1.4.0

Lastly, run

maturin develop --release

If you want to test the build locally, you may run

# pip install -r requirements-test.txt
pytest tests/test_*

If you see this error in pytest, it means setuptools is not installed and you may ignore it. It is just a legacy python builtin package.

tests/test_many.py::test_xi_corr - ModuleNotFoundError: No module named 'pkg_resources'

You can then publish it to your private PYPI server, or just use it locally.

Credits

1. Some statistics functions are taken from Statrs (MIT) and internalized. See here
2. Linear algebra routines are powered mostly by faer
3. String similarity metrics are soooo fast because of RapidFuzz

Other Projects

1. Caching for Polars
2. Polars FWF Reader

AI Usage Disclosure

Since this project is mostly maintained by a single person, and a single person cannot be fluent with all topics in scientific programming, some of the code in this package is guided by AI and reviewed by me. I will not accept vibe-coded functions. For AI-generated PRs, see contributing.md. Once again, please understand that I have limited attention and if you submit an AI PR with thousands of lines of code, I may not keep up and decide not to include a large portion of the code due to maintenance reasons.