Algorithm repository

This is the repository of algorithms for the MIP.

Algorithms, written in their native language (R, Matlab, Python, Java...) are encapsulated in a Docker container that provides them with the runtime environment necessary to execute this function.

The environment variables provided to the Docker container are used as parameters to the function or algorithm to execute.

Currently, we expect the Docker containers to be autonomous:

they should connect to a database and retrieve the dataset to process
they should process the data, taking into account the parameters given as environment variables to the Docker container
they should store the results into the results database.

The format of the results should be easily shared.

For algorithms providing statistical analysis or machine learning, we require the results to be in PFA format in its YAML or JSON form.
For algorithms providing visualisations, we support different formats, including Highcharts, Vis.js, PNG and SVG.
For algorithms providing tabular data, we expect a JSON output in this format: Tabular Data Resource

List of algorithms

hbpmip/python-anova: Anova algorithm

This is a Python implementation of Anova.

hbpmip/python-correlation-heatmap: Correlation heatmap

Calculate correlation heatmap, only works for real variables. Run it on single node or in a distributed mode. First, intermediate mode calculates covariance matrix from a single node, then aggregate mode is used after intermediate to combine statistics from multiple jobs and produce the final graph.

hbpmip/python-distributed-pca: PCA - principal components analysis

Calculate PCA, only works for real variables. Run it on single node or in a distributed mode. First, intermediate mode calculates covariance matrix from a single node, then aggregate mode is used after intermediate to combine statistics from multiple jobs and produce the final graph.

Code is shared with hbpmip/python-correlation-heatmap

hbpmip/python-distributed-kmeans: K-means

Implementation of distributed k-means clustering (https://github.com/MRN-Code/dkmeans) in Python. It uses Single-Shot Decentralized LLoyd (https://github.com/MRN-Code/dkmeans#single-shot-decentralized-lloyd).

Intermediate mode calculates clusters on a single node, while aggregate mode is merging the clusters according to least merging error (e.g. smallest distance between centroids).

hbpmip/python-histograms: Histograms

Calculates histogram of nominal or real variable grouped by nominal variables in independent variables. Histogram edges are taken from minValue and maxValue property of dependent variable. If not available, then these values are calculated dynamically from dependent values (this won't work in distributed mode though).

hbpmip/python-jsi-hedwig: Hedwig method

Hedwig method for semantic subgroup discovery. (https://github.com/anzev/hedwig).

hbpmip/python-jsi-hinmine: HINMINE

The HINMINE algorithm for network-based propositionalization is an algorithm for data analysis based on network analysis methods.

The input for the algorithm is a data set containing instances with real-valued features. The purpose of the algorithm is to construct a new set of features for further analysis by other data mining algorithms. The algorithm outputs a data set with features, generated for each data instance in the input data set. The features represent how close a given instance is to the other instances in the data set. The closeness of instances is measured using the PageRank algorithm, calculated on a network constructed from instance similarities.

hbpmip/python-knn: k-nearest neighbors

Implementation of k-nearest neighbors algorithm (https://en.wikipedia.org/wiki/K-nearest_neighbors_algorithm) in Python.

Run it on single node or in a distributed mode.

hbpmip/python-linear-regression: Linear and logistic regression

Python implementation of multivariate linear regression. It supports both continuous and categorical as independent variables. Run it on single node or in a distributed mode. Python implementation of logistic regressions on one class versus the others. Only single-node mode is supported

hbpmip/python-sgd-regression: SGD family of regressions

This is a Python implementation of scikit-learn estimators (http://scikit-learn.org/stable/modules/scaling_strategies.html) using Stochastic Gradient Descent and the partial_fit method for distributed learning.

Implemented methods:

linear_model - calls SGDRegressor or SGDClassifier
neural_network - calls MLPRegressor or MLPClassifier
naive_bayes - calls MixedNB (mix of GaussianNB and MultinomialNB), only works for classification tasks
gradient_boosting - calls GradientBoostingRegressor or GradientBoostingClassifier, does not support distributed training.

hbpmip/python-summary-statistics: Summary statistics

It calculates various summary statistics for entire dataset and also for all subgroups created by combining all possible values of nominal covariates. Run it on single node or in a distributed mode.

hbpmip/python-tsne: t-SNE

The python-tsne is a wrapper for the the A-tSNE algorithm developed by N. Pezzotti. The underlying algorithm is an improvement on the Barnes-Hut tSNE (http://lvdmaaten.github.io/publications/papers/JMLR_2014.pdf) using an approximated k-nearest neighbor calculation.

hbpmip/java-jsi-clus-fire: k-nearest neighbors

hbpmip/java-jsi-clus-fr: k-nearest neighbors

hbpmip/java-jsi-clus-pct: k-nearest neighbors

hbpmip/java-jsi-clus-pct-ts: k-nearest neighbors

hbpmip/java-jsi-clus-rm: k-nearest neighbors

hbpmip/java-jsi-streams-modeltree: k-nearest neighbors

hbpmip/java-jsi-streams-regressiontree: k-nearest neighbors

hbpmip/python-longitudinal: Longitudinal

hbpmip/r-3c: 3C

hbpmip/r-ggparci: ggParci

hbpmip/r-heatmaply: Heatmaply

hbpmip/java-rapidminer-knn: 🌑 k-NN k-NN

k-NN implemented with RapidMiner. Deprecated, replaced by hbpmip/python-knn

java-rapidminer-naivebayes: 🌑 Naive Bayes Naive Bayes

Naive Bayes implemented with RapidMiner. Deprecated, replaced by hbpmip/python-naivebayes

hbpmip/r-linear-regression: 🌑 Linear regression Linear regression

Linear regression implemented in R, with support for federated results. Deprecated, replaced by hbpmip/python-linear-regression

Algorithm capabilities

Algorithm	Description	Predictive	Federated results	In production	Used for	Runtime engine
hbpmip/python-anova	Anova		✔️ 🔜	✔️	Regression	Woken
hbpmip/python-correlation-heatmap	Correlation heatmap		❌	✔️	Visualisation	Woken
hbpmip/python-distributed-pca	PCA		✔️	✔️	Visualisation	Woken
hbpmip/python-distributed-kmeans	K-means		✔️	✔️	Clustering	Woken
hbpmip/python-histograms	Histograms		✔️	✔️	Visualisation	Woken
hbpmip/python-jsi-hedwig	Hedwig		❌	✔️		Woken
hbpmip/python-jsi-hinmine	HINMINE		❌	✔️		Woken
hbpmip/python-knn	k-NN	✔️	✔️	✔️	Clustering	Woken
hbpmip/python-linear-regression	Linear regression	✔️	✔️	✔️	Regression	Woken
hbpmip/python-linear-regression	Logistic regression	✔️	❌	✔️	Regression, Classification	Woken
hbpmip/python-sgd-regression	SGD Linear model	✔️	✔️	✔️	Classification	Woken
hbpmip/python-sgd-regression	SGD Neural Network	✔️	❌	✔️	Classification	Woken
hbpmip/python-sgd-regression	SGD Naive Bayes	✔️	❌	✔️	Classification	Woken
hbpmip/python-sgd-regression	SGD Gradient Boosting	✔️	❌	✔️	Classification	Woken
hbpmip/python-summary-statistics	Summary statistics		✔️	✔️	Data exploration	Woken
hbpmip/python-tsne	t-SNE		❌	✔️	Visualisation	Woken

Acknowledgements

This work has been funded by the European Union Seventh Framework Program (FP7/20072013) under grant agreement no. 604102 (HBP)

This work is part of SP8 of the Human Brain Project (SGA1).

kkech/algorithm-repository