Skip to main content

FaissImputer

Warning: FaissImputer 0.1.x has a known neighbor-mapping bug that can produce incorrect imputations, and it is incompatible with scikit-learn 1.8+. Use version 0.2.0 or newer.

PyPI Version License

A scikit-learn-compatible missing-value imputer with Faiss-backed neighbor search.

Current release: 0.3.14. See Releases for version history.

Comparison with KNNImputer

FaissImputer supports scikit-learn pipelines, but its defaults and options differ from KNNImputer. The comparison below describes FaissImputer 0.3.14.

Behavior FaissImputer KNNImputer
Donors Training rows observed in every non-empty feature by default; donor_policy="available" permits partially observed donors selected per missing feature. Donors selected per missing feature; other donor features may be missing.
Neighbors n_neighbors=3; complete mode requires at least that many complete donors. Available mode permits fewer eligible donors. n_neighbors=5; fewer usable neighbors are allowed.
Aggregation Mean (default) or median via strategy, with uniform weights by default. Distance and callable weights require strategy="mean" and either metric="l2" or metric="nan_euclidean". Mean with uniform (default), distance, or callable weights; no median strategy.
Numeric precision Converts inputs and produces imputed values as float32. Supports floating inputs including float64, without forcing conversion to float32.
All-missing training columns Dropped by default; keep_empty_features=True retains them with zero values under either donor policy. Dropped by default; keep_empty_features=True retains them with zero values.
Missing-value marker missing_values=np.nan by default; also accepts a finite numeric marker, matched before float32 conversion. Configurable missing_values, default np.nan.
Missing indicators add_indicator=True appends 0/1 columns for features missing during fit; disabled by default. add_indicator=True appends indicators for features missing during fit.

For a closer comparison, use donor_policy="available", strategy="mean", and metric="nan_euclidean", and match n_neighbors, weights, and missing_values. Available mode requires index_factory="Flat" and accepts "l2" or its "nan_euclidean" alias.

These settings do not guarantee identical donor choices or imputed values: float32 conversion, distance calculations, and ties can affect results. See the real-data comparison for measured per-cell differences, and the scikit-learn imputation guide for KNNImputer behavior.

Performance at a glance

FaissImputer can accelerate nearest-neighbor imputation, especially when queries reuse the same missing-feature patterns.

Which donor policy should I use?

A donor is a training row used to supply a missing value.

  • complete (default): Uses training rows observed in every non-empty feature as donors. Rows missing any of those features are excluded from neighbor search. Columns entirely missing during fit do not disqualify rows. Choose this when you have enough complete training rows.
  • available: Also allows partially observed training rows as donors. Donors are selected separately for each missing feature: they must contain that feature's value and share at least one originally observed feature with the query row. No fully observed training row is required.

The benchmarks below use fully observed training data for complete and partially missing training data for available. They compare each policy against KNNImputer, not the two policies against each other.

Released 0.3.10 benchmarks

Both runs compared PyPI FaissImputer 0.3.10 with scikit-learn 1.9.0's KNNImputer on synthetic float32 data, using one thread, 300 queries, 20 features, five neighbors, and uniform-weight mean aggregation.

Each query had four missing features. Complete training data was fully observed; available training data had approximately 10% MCAR missingness. Inputs were identical within each policy comparison.

Times measure the first transform after fitting, excluding fit time. Medians cover three seeds and three fresh runs per seed. Speedups are KNNImputer/FaissImputer ratios of unrounded median times.

20,000-row comparison — AMD runner

Donor policy Query missingness KNNImputer FaissImputer Speedup
complete One shared pattern 422.4 ms 22.0 ms 19.24×
complete Random patterns 409.8 ms 107.2 ms 3.82×
available One shared pattern 391.4 ms 166.2 ms 2.35×
available Random patterns 385.6 ms 167.5 ms 2.30×

Raw results and environment · Workflow run

Training-size comparison — Intel runner

Values above 1× mean FaissImputer is faster.

Training rows complete / fixed complete / random available / fixed available / random
1,000 6.75× 0.71× — slower 1.24× 1.19×
5,000 9.64× 1.46× 1.34× 1.34×
20,000 11.00× 1.63× 1.32× 1.33×
100,000 15.56× 2.30× 1.73× 1.64×

Raw results and environment · Workflow run

In both runs, complete outputs matched KNNImputer exactly; available outputs differed by at most 4.77e-7 on the tested inputs. Repeated outputs were unchanged.

Memory advantages varied with size. In the scaling run, available's median whole-worker peak RSS was 6–21% higher at 5,000 and 20,000 rows, but 15–27% lower at 100,000 rows, compared with KNNImputer.

These runs used different CPUs. Differences between their results do not measure a change between software versions. Speed and memory advantages are not guaranteed on other workloads.

Installation

FaissImputer requires Python 3.10 or newer.

python -m pip install --upgrade "faiss-imputer>=0.3.14"

Usage

Complete donors (default)

When at least one training column is non-empty, training data must contain enough rows observed in all non-empty columns to supply n_neighbors donors.

import numpy as np

from faiss_imputer import FaissImputer

X_train = np.array(
    [
        [1.0, 10.0, 100.0],
        [2.0, 20.0, 200.0],
        [3.0, 30.0, 300.0],
    ],
    dtype=np.float32,
)
X_missing = np.array(
    [
        [1.5, np.nan, 150.0],
        [2.5, 25.0, np.nan],
    ],
    dtype=np.float32,
)

imputer = FaissImputer(n_neighbors=2)
X_imputed = imputer.fit(X_train).transform(X_missing)
print(X_imputed)

Expected output:

[[  1.5  15.  150. ]
 [  2.5  25.  250. ]]

Partially observed donors

Set donor_policy="available" to choose donors separately for each missing feature. No fully observed training row is required.

import numpy as np

from faiss_imputer import FaissImputer

# No training row is completely observed.
X_train = np.array(
    [
        [0.0, 10.0, np.nan],
        [2.0, 30.0, np.nan],
        [1.0, np.nan, 20.0],
        [3.0, np.nan, 40.0],
    ],
    dtype=np.float32,
)
X_missing = np.array(
    [[0.1, np.nan, np.nan], [np.nan, np.nan, np.nan]],
    dtype=np.float32,
)

imputer = FaissImputer(n_neighbors=1, donor_policy="available")
X_imputed = imputer.fit(X_train).transform(X_missing)
print(X_imputed)

Expected output:

[[ 0.1 10.  20. ]
 [ 1.5 20.  30. ]]

By default, both policies return a new NumPy float32 array. The training and query inputs are not modified.

Feature names and pandas output

The following example requires FaissImputer 0.3.3 or newer. It also requires pandas, which is optional: python -m pip install pandas.

import pandas as pd
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler

from faiss_imputer import FaissImputer

X = pd.DataFrame({
    "age": [20.0, None, 40.0],
    "income": [100.0, 190.0, 300.0],
})

pipe = make_pipeline(
    FaissImputer(n_neighbors=1),
    StandardScaler(),
).set_output(transform="pandas")

result = pipe.fit_transform(X)

print(pipe.get_feature_names_out())  # ['age' 'income']
print(result.columns.tolist())      # ['age', 'income']

Use pipe.set_output(transform="default") to return NumPy arrays instead. For unnamed array inputs, feature names are generated as x0, x1, and so on.

Parameters

  • n_neighbors (default: 3): Positive integer specifying the maximum number of donors used for each missing feature. When non-empty training columns exist, the complete-donor policy requires at least this many rows observed in all of them.
  • metric (default: "l2"): Supports "l2", "nan_euclidean", and "ip". "nan_euclidean" is an alias for the existing NaN-aware L2 imputation path. Raw inner product is not cosine similarity. The available-donor policy accepts "l2" and "nan_euclidean". Callable metrics are not supported.
  • strategy (default: "mean"): Supports "mean" and "median" for aggregating donor values and calculating fallback column statistics.
  • index_factory (default: "Flat"): Faiss index description for the complete-donor policy. The available-donor policy accepts only "Flat" and uses the distance backend described below.
  • donor_policy (default: "complete"): Use training rows observed in all non-empty features with "complete", or allow partially observed training rows with "available".
  • weights (default: "uniform"): "uniform" or None preserves the existing unweighted aggregation. "distance" uses inverse Euclidean distance; a callable supplies custom weights. Distance and callable weights require strategy="mean" and either metric="l2" or metric="nan_euclidean" under either donor policy.
  • add_indicator (default: False): Append missingness indicators to the imputed output. Indicator columns are selected during fit() and remain fixed until refitting.
  • keep_empty_features (default: False): Drop columns that were entirely missing during fit(). Set to True to retain those columns with zero values. The selection remains fixed until refitting.
  • missing_values (default: np.nan): The missing-value marker used during both fit() and transform(). Accepts NaN or a finite Python/NumPy integer or floating-point scalar, such as -1 or 9999. Booleans, strings, None, and infinity are not supported as markers.

With distance weighting, if any selected donor has distance zero, only the selected zero-distance donors contribute to that missing feature.

Callables receive a two-dimensional array of NaN-aware Euclidean distances and must return real weights of the same shape. They should operate independently on each row, without relying on batch size or neighbor order. Unavailable neighbors have NaN distances and are excluded regardless of the returned weight. Returned NaN weights contribute zero.

Invalid shapes, non-real weights, infinite weights for usable donors, zero weight sums, or results outside the finite float32 range raise ValueError. Fitted fallback column statistics remain unweighted.

Important behavior

Shared behavior

  • Inputs must be two-dimensional numeric array-like data. missing_values identifies missing entries, with NaN as the default. Observed values are converted to float32; infinity and observed values outside the finite float32 range are not accepted.
  • transform() requires the same number of features as fit().
  • An entirely missing query row uses column means or medians learned during fit() for non-empty training columns.
  • A failed fit(), including a failed refit, clears the fitted state.

A numeric marker is matched by exact equality before float32 conversion. Distinct observed values remain observed even if they round to the same float32 value as the marker. Use the same marker when preparing training and query data; with a numeric marker, actual NaN entries are rejected. A marker outside the float32 range is allowed because matching entries are removed before conversion. Change the marker by refitting the estimator.

For example, FaissImputer(n_neighbors=1, missing_values=-1) fitted on [[0, 10], [2, 20]] transforms [[2, -1]] to [[2, 20]]. Both missing indicators and empty-column handling use the same missingness decisions as imputation.

With add_indicator=True, appended values are 1 for originally missing query entries and 0 otherwise. Indicator features are selected from all training rows before donor filtering, in original column order. A feature first missing only at transform time does not receive a new indicator column.

Indicator names use missingindicator_<feature_name>, such as missingindicator_age or missingindicator_x0. Both get_feature_names_out() and pandas output include the added columns.

Columns entirely missing during fit() are excluded from donor selection and neighbor search. With keep_empty_features=False, they are omitted from the imputed output. With True, they remain in their original positions and always contain zero, even if a later query supplies an observed value in that column. transform() still requires the full original input schema.

If every training column is empty, fitting succeeds without a neighbor index. Transform returns no imputed columns by default, or all-zero columns with keep_empty_features=True; no donor-count minimum applies in this case.

With add_indicator=True, indicators still refer to the original input columns and query missingness, including dropped or zero-filled columns. Output names and pandas columns follow the retained imputed columns first, then the indicators in their original feature order. NaN-aware distance normalization continues to use the original input feature count, including for callable weights.

Complete-donor policy

  • Donors must be observed in all non-empty training columns. When any such columns exist, at least one complete donor is required.
  • When non-empty training columns exist, n_neighbors cannot exceed the number of complete donors.
  • Neighbor search uses only the originally observed columns of each query row.
  • The default index_factory="Flat" performs exact neighbor search.

Available-donor policy

  • Supports metric="l2" or metric="nan_euclidean", with index_factory="Flat" required. Mean supports all weight options; median requires uniform weighting.
  • A donor must observe the feature being imputed and share at least one originally observed feature with the query row.
  • Donors are ranked by squared L2 distance over shared observed features, scaled by the total feature count divided by the shared feature count.
  • Each missing feature uses up to n_neighbors eligible donors. Fewer eligible donors are allowed.
  • If no eligible donor exists for a feature, its fitted column mean or median is used.
  • Rows missing all non-empty training features are ignored. Entirely missing training columns follow keep_empty_features.

The available-donor policy uses batched NaN-aware distances and Faiss neighbor selection, with float64 safeguards for detected numerical risks. This backend is also used when all training donors are complete.

Batching trades memory for throughput. Its internal batch-sizing budget is not a total process-memory limit: fitted donor data, distance calculations, and temporary arrays also consume memory. Larger batches can improve speed but increase peak memory usage. FaissImputer does not set the number of threads; configure thread limits in the application when needed.

Donor-side arrays are prepared during fit() and reused across query batches. This reduces repeated work but increases retained memory after fitting; lower peak memory is not guaranteed.

Benchmarks

Performance depends on donor policy, data size, and missingness patterns. Neither faster execution nor lower memory use than KNNImputer is guaranteed. Similar average errors do not imply identical imputed values; ties and numerical precision can affect donor selection.

  • Complete-aggregation recovery pilot: compares v0.3.8 source with the implementation prepared for 0.3.10 on one runner, across three seeds and five query/feature shapes. Complete-policy first transform was 16.69-60.96% faster; available-policy changes ranged from -3.59% to +3.84%. These are source-checkout measurements with ordinary-scale inputs, not published-wheel measurements or a KNNImputer comparison.
  • Available-donor batching and threads: compares 16/64/128 MiB batch budgets and 1/2/4 threads, including repeated 500,000-row measurements and a single-run million-row pilot. Documents both speed gains and memory tradeoffs.
  • Real-data pilot: compares SimpleImputer, KNNImputer, and both Faiss donor policies under MCAR and selected MAR missingness. This small dataset is not a scalability test.
  • Complete-donor patterns: measures the impact of training size and query missingness patterns.
  • Partial-donor development snapshot: historical results for measured commit 0a3cc077, not the final 0.3.0 implementation.
  • Historical 0.2.0 benchmark: results for the earlier complete-donor-only implementation.

Reports provide measurement conditions, results, and links to reproduction code and raw data.

Example notebook

See Imputing Missing Values with Faiss Imputer for a complete-donor example.

Contributing

Contributions are welcome! Please open an issue or create a pull request. Further work is tracked in the roadmap.

Author

License

This project is licensed under the MIT License.

Third-Party Licenses

FaissImputer depends on Meta's Faiss, which is distributed under the MIT License.

FaissImputer is not affiliated with or endorsed by Meta or the Faiss maintainers.

Metadata

Release files for faiss-imputer 0.3.14

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for faiss-imputer 0.3.14
File Size Uploaded Provenance Actions
faiss_imputer-0.3.14.tar.gz 51.6 kB With provenance Details for faiss_imputer-0.3.14.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for faiss-imputer 0.3.14
File Interpreter ABI Platform Actions
faiss_imputer-0.3.14-py3-none-any.whl Python 3 none any Details for faiss_imputer-0.3.14-py3-none-any.whl

Total release size: 70.5 kB

Release files / faiss_imputer-0.3.14.tar.gz

Download URL faiss_imputer-0.3.14.tar.gz
Size 51.6 kB
Tags Source
SHA-256 checksum
How to use checksums
71c6f417f22acdf2e8f8e7080c28f25e924d24a75dbd6d4fcda6b773ef801e64
BLAKE2b-256 checksum
How to use checksums
374f6c5c9e017b99885a827f1ead7fb95850876fdab23a5c6a58dfa5f5beed6b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 13, 2026.

Transparency log

Release files / faiss_imputer-0.3.14-py3-none-any.whl

Download URL faiss_imputer-0.3.14-py3-none-any.whl
Size 19.0 kB
Tags Python 3
SHA-256 checksum
How to use checksums
9b83a37c61829f68837086428356b4be3e5ebe8135284c4715503f800880ce68
BLAKE2b-256 checksum
How to use checksums
332b8ea5e62ae3160ef89bae088171c4eaf7d9be34549d1c2530c013e66599a4
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 13, 2026.

Transparency log

Security for faiss-imputer 0.3.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for faiss_imputer-0.3.14.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 13, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.3.14 This release

2 release files Full provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page