Skip to main content

Pandio's machine learning library.

Project description

Pandio Logo

PandioML - Machine Learning

This repository contains the PandioML Python library to develop and deploy machine learning for streaming data.

At a high level, PandioML provides three things:

  1. Pipelines

    Traditional pipelines based on scikit-learn pipelines.

  2. Datasets

    Connect and stream any source of data into a pipeline.

  3. Automation Of Pipelines & Datasets

    Use pandiocli to automate locally or automate on the platform powered by Apache Pulsar and Presto/Trino.

Quick Links - - Getting Started - Quick Start - PyPi PandioML - PyPi PandioCLI

About PandioML

Incremental Learning

As data is streamed in, stream learning models are created incrementally and updated continously. They are extremely powerful in a real-time environment, but can also be a viable replacement for traditional machine learning.

Adaptive Learning

Data distribution changes have far less impact from concept drift on a stream learning model due to the adaptive nature of incremental learning.

Resource Efficient

Due to the incremental learning, considerably less processing time and memory are needed to train a model.

Easy To Use

PandioML is based on scikit-learn, making it easy to pickup.

Open Source

Based on powerful open source technology itself, PandioML is released under SSPL License.

AI Orchestration

Not just a Python library, PandioCLI makes it easy to package and deploy PandioML pipelines and models to the Pandio platform.

Use Cases

Supervised Learning

When working with labeled data. Depending on the target type can be either classification (discrete values) or regression (continuous values)

Single & Multiple Output

Single-output methods predict a single target-label (binary or multi-class) for classification or a single target-value for regression. Multi-output methods simultaneously predict multiple variables given an input.

Concept Drift Detection

Changes in data distribution can harm learning. Drift detection methods are designed to rise an alarm in the presence of drift and are used alongside learning methods to improve their robustness against this phenomenon in evolving data streams.

Unsupervised Learning

When working with unlabeled data. For example, anomaly detection where the goal is the identification of rare events or samples which differ significantly from the majority of the data.

Building Pipelines

Scikit-learn pipelines on steroids! Familiar syntax with more helper methods to make building and debugging pipelines easier.

Deploying Models

After writing (or re-using) your pipeline and model, deploy to Pandio's AI Orchestration platform with a single command. Your model is now available for production use at any scale!

Our Philosophy

PandioML exists to provide a full featured experience from accessing data, developing a pipeline, and deploying to production as quickly as possible.

Our focus is on ease-of-use, quick iterations, and single command deploy to production. The library is meant to automate everything except the actual data science, to optimize pipeline development. For the actual data science, the focus is on accessibility. 40+ algorithms, dozens of metrics and stats, and a powerful pipeline framework save significant time.

The last core idea to PandioML is to keep it as low level as possible. All too often libraries try to abstract things away to make things simple, but in our experience it only complicates things. It might make a few things easier, or negate the need to write a little more code, but when debugging or auditing is needed, it becomes a real pain to dive into what is happening.

This library was meant to be for everyone, from zero experience to experts in the field.


Pandio is an AI Orchestration platform that helps companies accelerate their AI initiatives. Connect data, build pipelines, train models, and deploy models on your terms at unmatched speed and scale.


Python 3.5 - 3.8

PIP > 20.0.0

Docker (or Docker Desktop)

How It Works

Architecture Diagram

PandioML is built on the AI Orchestration platform. For the purposes of PandioML, this means that two things support the deployment of a PandioML project.

  1. Distributed Messaging

    This is the concept of a stream, queue, or pubsub providing input topic(s) and output topic(s). A PandioML project sits inbetween the input topic(s) and the output topic(s).

    This means to receive a prediction, a message must be placed on an input topic. The result is then placed on an output topic, which are defined in the project file.

  2. Serverless Functions

    This is a compute framework that runs the PandioML project in a Kubernetes environment.

When you build machine learning with PandioML, you automatically get a fully scalable production environment with each deployment of a project.

As you work locally, these components are simulated to allow quick iteration before deploying the project to production with a single command.

This is one of the most powerful features of PandioML. Your local development on training, iterating, evaluating, etc. is the same pipelines that you'll deploy as a microservice to You don't have to worry about packaging it as a microservice, using the PandioML and PandioCLI tools does all of that for you.


  1. Getting Started
  2. Quick Start
  3. Pipelines
  4. Datasets
  5. Metrics
  6. Stats
  7. Schemas and Schema Registry
  8. Full Form Fraud Example
  9. Full Movie Ratings Example
  10. Full Credit Card Fraud Pipeline Example
  11. Full Credit Card Fraud Dataset Example
  12. How To Load A Model
  13. Hyperparameter Tuning
  14. Full Lead Scoring Pipeline Example
  15. Full Lead Scoring Dataset Example



pip install pandioml

To run pipelines locally, Docker is required.

Docker removes the need to install complex data science environments on your computer.

Get it free here:



The pandioml.model.* module handles all of the available algorithms and models.

Module Description
GaussianNB Gaussian Naive Bayes
MultinomialNB Naive Bayes classifier for multinomial models.
ComplementNB Naive Bayes classifier for multinomial models.
BernoulliNB Bernoulli Naive Bayes.
K-means Incremental k-means.
ADWINBaggingClassifier ADWIN Bagging classifier.
BaggingClassifier Online bootstrap aggregation for classification.
BaggingRegressor Online bootstrap aggregation for regression.
AdaBoostClassifier Boosting for classification
SRPClassifier Streaming Random Patches ensemble classifier.
EpsilonGreedyRegressor Epsilon-greedy bandit algorithm for regression.
UCBRegressor Upper Confidence Bound bandit for regression.
EWARegressor Exponentially Weighted Average regressor.
SuccessiveHalvingClassifier Successive halving algorithm for classification.
SuccessiveHalvingRegressor Successive halving algorithm for regression.
StackingClassifier Stacking for binary classification.
FFMClassifier Field-aware Factorization Machine for binary classification.
FFMRegressor Field-aware Factorization Machine for regression.
FMClassifier Factorization Machine for binary classification.
FMRegressor Factorization Machine for regression.
FwFMClassifier Field-weighted Factorization Machine for binary classification.
FwFMRegressor Field-weighted Factorization Machine for regression.
HOFMClassifier Higher-Order Factorization Machine for binary classification.
HOFMRegressor Higher-Order Factorization Machine for regression.
HardSamplingClassifier Hard sampling classifier.
HardSamplingRegressor Hard sampling regressor.
RandomOverSampler Random over-sampling.
RandomSampler Random sampling by mixing under-sampling and over-sampling.
RandomUnderSampler Random under-sampling.
HoeffdingTreeClassifier Hoeffding Tree or Very Fast Decision Tree classifier.
HoeffdingAdaptiveTreeClassifier Hoeffding Adaptive Tree classifier.
ExtremelyFastDecisionTreeClassifier Extremely Fast Decision Tree classifier.
LabelCombinationHoeffdingTreeClassifier Label Combination Hoeffding Tree for multi-label classification.
HoeffdingTreeRegressor Hoeffding Tree regressor.
HoeffdingAdaptiveTreeRegressor Hoeffding Adaptive Tree regressor.
iSOUPTreeRegressor Incremental Structured Output Prediction Tree (iSOUP-Tree) for multi-target regression.
StackedSingleTargetHoeffdingTreeRegressor Stacked Single-target Hoeffding Tree regressor.
KNNClassifier k-Nearest Neighbors classifier.
KNNADWINClassifier K-Nearest Neighbors classifier with ADWIN change detector.
SAMKNNClassifier Self Adjusting Memory coupled with the kNN classifier.
KNNRegressor k-Nearest Neighbors regressor.
AccuracyWeightedEnsembleClassifier Accuracy Weighted Ensemble classifier
AdaptiveRandomForestClassifier Adaptive Random Forest classifier.
AdaptiveRandomForestRegressor Adaptive Random Forest regressor.
AdditiveExpertEnsembleClassifier Additive Expert ensemble classifier.
BatchIncrementalClassifier Batch Incremental ensemble classifier.
ClassifierChain Classifier Chains for multi-label learning.
ProbabilisticClassifierChain Probabilistic Classifier Chains for multi-label learning.
MonteCarloClassifierChain Monte Carlo Sampling Classifier Chains for multi-label learning.
DynamicWeightedMajorityClassifier Dynamic Weighted Majority ensemble classifier.
LearnPPNSEClassifier Learn++.NSE ensemble classifier.
LearnPPClassifier Learn++ ensemble classifier.
LeveragingBaggingClassifier Leveraging Bagging ensemble classifier.
MultiOutputLearner Multi-Output Learner for multi-target classification or regression.
OnlineAdaC2Classifier Online AdaC2 ensemble classifier.
OnlineBoostingClassifier Online Boosting ensemble classifier.
OnlineCSB2Classifier Online CSB2 ensemble classifier.
OnlineRUSBoostClassifier Online RUSBoost ensemble classifier.
OnlineSMOTEBaggingClassifier Online SMOTEBagging ensemble classifier.
OnlineUnderOverBaggingClassifier Online Under-Over-Bagging ensemble classifier.
OzaBaggingClassifier Oza Bagging ensemble classifier.
OzaBaggingADWINClassifier Oza Bagging ensemble classifier with ADWIN change detector.
RegressorChain Regressor Chains for multi-output learning.
StreamingRandomPatchesClassifier Streaming Random Patches ensemble classifier.
ADWIN Adaptive Windowing method for concept drift detection.
DDM Drift Detection Method.
EDDM Early Drift Detection Method.
HDDM_A Drift Detection Method based on Hoeffding’s bounds with moving average-test.
HDDM_W Drift Detection Method based on Hoeffding’s bounds with moving weighted average-test.
KSWIN Kolmogorov-Smirnov Windowing method for concept drift detection.
PageHinkley Page-Hinkley method for concept drift detection.
EvaluateHoldout The holdout evaluation method or periodic holdout evaluation method.
EvaluatePrequential The prequential evaluation method or interleaved test-then-train method.
EvaluatePrequentialDelayed The prequential evaluation delayed method.
VeryFastDecisionRulesClassifier Very Fast Decision Rules classifier.
RobustSoftLearningVectorQuantization Robust Soft Learning Vector Quantization for Streaming and Non-Streaming Data.
HalfSpaceTrees Implementation of the Streaming Half–Space–Trees (HS–Trees)


The* model contains all of the datasets and generators available.

Module Description Schema Labeled
FormSubmissionGenerator Uses the Faker Python package to generate an infinite amount of form submissions. schema No
WebHostingDataset Contains 12,496,728 server resource metric events recorded over a 3 month period of time. schema No
PersonProfileDataset Generates an infinite stream of user Profiles using the Faker Python library. schema No
CreditCardFraud A dataset of 1,296,675 credit card transactions with a percentage labeled as fraud. schema Yes
AgrawalGeneratorDataset A generator for data regarding home loan applications with the ability to balance and add noise. schema Yes
SineGeneratorDataset A generator for data regarding sine values with the ability to balance and add noise. schema Yes
LEDGeneratorDataset A generator for data regarding a digit displayed on a seven-segment LED display with the ability to add noise. schema Yes
PhishingDataset 1250 entries of webpages that are classified as phishing or not. schema Yes
MovieRatingDataset 100,000 movie ratings from different types of individuals. schema Yes
RestaurantVisitorsDataset This dataset contains 252,108 records over a 16 week period to 829 Japanese Restaurants. schema Yes

Additional custom datasets can be created using the pandiocli dataset generate tool to use your own data.


The pandioml.metrics.* contains helper methods to calculate metrics relating to the pipelines.

Module Description
Accuracy Accuracy score, which is the percentage of exact matches.
BalancedAccuracy Balanced accuracy.
BinaryMetric Mother class for all binary classification metrics.
ClassificationMetric Mother class for all classification metrics.
ClassificationReport A report for monitoring a classifier.
CohenKappa Cohen's Kappa score.
CrossEntropy Multiclass generalization of the logarithmic loss.
ExactMatch Exact match score.
ExamplePrecision Example-based precision score for multilabel classification.
ExampleRecall Example-based recall score for multilabel classification.
ExampleF1 Example-based F1 score for multilabel classification.
ExampleFBeta Example-based F-Beta score.
F1 Binary F1 score.
FBeta Binary F-Beta score.
GeometricMean Geometric mean score.
Hamming Hamming score.
HammingLoss Hamming loss score.
Jaccard Jaccard index for binary multi-outputs.
KappaM Kappa-M score.
KappaT Kappa-T score.
LogLoss Binary logarithmic loss.
MAE Mean absolute error.
MacroF1 Macro-average F1 score.
MacroFBeta Macro-average F-Beta score.
MacroPrecision Macro-average precision score.
MacroRecall Macro-average recall score.
MCC Matthews correlation coefficient.
Metric Mother class for all metrics.
Metrics A container class for handling multiple metrics at once.
MicroF1 Micro-average F1 score.
MicroFBeta Micro-average F-Beta score.
MicroPrecision Micro-average precision score.
MicroRecall Micro-average recall score.
MultiClassMetric Mother class for all multi-class classification metrics.
MultiFBeta Multi-class F-Beta score with different betas per class.
MultiOutputClassificationMetric Mother class for all multi-output classification metrics.
MultiOutputRegressionMetric Mother class for all multi-output regression metrics.
MSE Mean squared error.
Precision Binary precision score.
Recall Binary recall score.
RegressionMetric Mother class for all regression metrics.
RegressionMultiOutput Wrapper for multi-output regression.
RMSE Root mean squared error.
RMSLE Root mean squared logarithmic error.
ROCAUC Receiving Operating Characteristic Area Under the Curve.
Rolling Wrapper for computing metrics over a window.
R2 Coefficient of determination ($R^2$) score
SMAPE Symmetric mean absolute percentage error.
TimeRolling Wrapper for computing metrics over a period of time.
WeightedF1 Weighted-average F1 score.
WeightedFBeta Weighted-average F-Beta score.
WeightedPrecision Weighted-average precision score.
WeightedRecall Weighted-average recall score.
WrapperMetric Metric Wrapper
BallHall Ball-Hall index
BIC Bayesian Information Criterion (BIC).
CalinskiHarabasz Calinski-Harabasz index (CH).
Cohesion Mean distance from the points to their assigned cluster centroids. The smaller the better.
DaviesBouldin Davies-Bouldin index (DB).
GD43 Generalized Dunn's index 43 (GD43).
GD53 Generalized Dunn's index 53 (GD53).
Hartigan Hartigan Index (H - Index)
IIndex I-Index (I).
InternalMetric Mother class of all internal clustering metrics.
MSSTD Mean Squared Standard Deviation.
PS Partition Separation (PS).
CR2 R-Squared
RMSSTD Root Mean Squared Standard Deviation.
SD The SD validity index (SD).
Separation Average distance from a point to the points assigned to other clusters.
Silhouette Silhouette coefficient [^1], roughly speaking, is the ratio between cohesion and the average distances from the points to their second-closest centroid.
SSB Sum-of-Squares Between Clusters (SSB).
SSW Sum-of-Squares Within Clusters (SSW).
XieBeni Xie-Beni index (XB).
WB WB Index
Xu Xu Index


The pandioml.stats.* contains helper methods to calculate metrics relating to the pipelines.

Create Your Own Dataset or Generator

Use the class, inherit it, define the required methods, and use your own data inside of PandioML!

To make it available on Pandio's platform, pandiocli upload it using the PandioCLI, then pandiocli deploy it.

Build Streaming Pipelines

The pandioml.core.pipelines contains a pipeline framework to build traditional machine learning pipelines.

The PandioML framework for building pipelines is similar to scikit-learn's, but differs in that the pipeline is meant to process a single record, instead of a batch of records.

Click here to read our detailed Pipelines guide.


Reproducible pipelines are a critically important part of data science.

PandioML provides a tool to store any artifact associated with a pipeline.

3 artifacts are automatically stored for you:

  1. Dataset

  2. Pipeline

  3. Model

If you'd like to store additional things, such as metrics, charts, hypertuning parameters, or anything else, it can be done by importing the artifact module like so:

from pandioml.core.artifacts import artifact

Then, call the add method of the artifact as follows:

dict = artifact.add('config_params', {'foo': 'bar'})

The first parameter is a unique name for it, the second is the item to be stored. If it can be pickled, it can be stored as an artifact. In addition, there are times when it is preferred to defer the artifact, so a callable is also acceptable as the item to be stored. When called, a single argument is passed called storage_location that contains where the artifacts will be stored. This makes it easy to provide custom logic when storing artifacts.

Bonus, the method will return the artifact, so that you can easily add items and define things in a single line.

Additionally, a save method exists that can be manually called to store artifacts. The storage medium used to store these artifacts can also be extended. Currently, File and AWS S3 storage backends are supported. The default storage backend is File.

This method is called automatically when the function stops running. You may call it yourself, but not too often as it is an expensive operation depending on the size of the artifacts.

ARTIFACT_STORAGE inside of defines where the artifacts are stored. This can be overwritten through the call to save as well.

Schema Registry Support

Type safety is critically important to any data science initiative. If you can't rely on a float being a float, that can have dangerous consequences.

PandioML treats type safety as a first class citizen. Before data is streamed into your pipeline, you can require that it validates against a defined schema. If it fails validation, your pipeline will never see it.

Type safety in PandioML prevents potential disastrous situations and makes pipeline development more efficient and productive.

To give an example, the following class defines a schema that the data must follow. Additionally, the data comes in as an object built right from the schema.

class Transaction(Record):
    trans_date_trans_time = String()
    cc_num = Integer()
    merchant = String()
    category = String()
    amt = Float()
    first = String()
    last = String()
    gender = String()
    street = String()
    city = String()
    state = String()
    zip = Integer()
    lat = Float()
    long = Float()
    city_pop = Integer()
    job = String()
    dob = String()
    trans_num = String()
    unix_time = Integer()
    merch_lat = Float()
    merch_long = Float()
    is_fraud = Integer()
    weekday = Integer()
    weekend = Integer()
    month = Integer()
    day = Integer()
    hour = Integer()

Using the above as an example, as data is streamed into your pipeline, it comes in as a Transaction object.

This allows you to quickly access properties like so: event.cc_num

Data science is so much easier with type safety offered by schemas!

The Schema Registry is provided by the platform. For most use cases, the inbound data is automatically handled for most implementations of PandioML.

The outbound data, or the prediction from PandioML, must be packaged up in a schema which is then consumed by the service that needs the prediction. It is just as important for this to have a schema as is the data sent into PandioML.

To see an example of returning an object with a schema, see this example: Restaurant Visits

This example shows a schema being created, then data being put into the object, and then the object being returned. This object is then sent to the output topic.

Debugging With PandioML Interactive Sessions

Have you ever asked yourself one of these questions?

  • What in the heck happened?
  • What does that error mean?
  • That value is not right... what part of my code did that?
  • My pipeline does not work, what is the problem?

If you have, you're not alone. If you haven't, feel free to skip this section!

To help find the problem and fix it quickly, PandioML has the ability to start an interactive session from any point in your code. This will halt execution and open up an interactive Python session that lets you inject code into your code execution in real time. No need to add a bunch of print statements to figure out what is happening!

To start an interactive session, import the interactive method from here:

from pandioml.core import interact

Then call it like so: interact(banner='Helpful Note Where This Was Triggered', local=locals())

When you run your code, it will halt where this method is called, letting you inject any Python code you would like into the execution.

With local, you can change scope of the variables, and access local=globals() if you'd like.

You can also add many of these calls all throughout your code. When you're ready to continue the execution of your code, simply type CTRL + D.

Happy debugging!



This example combines a NaiveBayes model, with the FormSubmissionGenerator, and a pipeline to demonstrate how to predict whether an email is from or


This example combines a LinearRegression model, with the RestaurantVisitorsDataset, and a pipeline to demonstrate how to predict how many visits a restaurant will receive.

This is the core of PandioML. Every project starting point is the file. All of the helper methods, magic sauce, embedded packages, cli tools, etc. exist to help build functions that run on the Pandio platform. In each of these functions is typically a model that makes accurate predictions. What data it uses, the algorithm, feature extractions, fitting, predicting, pipelining, and even labeling, is completely up to you. The function is your sandbox, where all the fun begins. Everything in PandioML is meant to help make creating value with machine learning easier.

Only two things are required to be defined in the function file generated for each project, the model you'd like to use, and the pipelines you'd like to execute against each individually streamed event.

Any number of classes or methods may be created to assist in the building of the function.

The pandioml.function.FunctionBase is the core component to building a model with PandioML. It provides helper methods and abstract methods that make it easy to build and deploy models.

The pandioml.function.Context provides local helper methods to simulate what is available when deployed to production to allow faster local iterative testing.

The pandioml.function.Storage provides access to a distributed key value storage service that is eventually consistent.

The pandioml.function.Logger provides local helper methods to simulate what is available when deployed to production to allow faster local iterative testing.


All contributions are welcome.

The best ways to get involved are as follows:

  1. Issues

    This is a great place to report any problems found with PandioML. Bugs, inconsistencies, missing documentation, or anything that acted as an obstacle to using PandioML.

  2. Discussions

    This is a great place for anything related to PandioML. Propose features, ask questions, highlight use cases, or anything else you can imagine.

If you would like to submit a pull request to this library, please read the contributor guidelines.


PandioML is licensed under the SSPL license.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

pandioml-1.0.13.tar.gz (62.9 kB view hashes)

Uploaded source

Built Distribution

pandioml-1.0.13-py3-none-any.whl (64.6 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page