blbarker / spark-tk-1

spark-tk

spark-tk is a library which enhances the Spark experience by providing a rich, easy-to-use API for Python and Scala. It adds new machine learning capabilities and other operations, like working with DICOM images for example.

Overview

Spark-tk simplifies applying machine learning to big data for superior knowledge, discovery and predictive modeling across a wide variety of use cases and solutions. Its APIs span feature engineering, graph construction, and various types of machine learning. The APIs are geared at an abstraction level familiar to data scientists (similar to Python pandas, scikit-learn) and removes the complexity of cluster computing and parallel processing. The library works alongside Spark and makes it easier to program. The lower-level Spark APIs are also seamlessly exposed through the library. Applications written with Spark-tk will have access the best of both worlds for the given situation. All functionality operates at full scale according to the Spark configuration.

Frame Interface

Spark-tk uses a Frame object for its scalable data frame representation, which is familiar and intuitive to data researchers compared to low level HDFS file and Spark RDD/DataFrame/DataSet formats. The library provides an API to manipulate the data frames for feature engineering and exploration, such as joins and aggregations. User-defined transformations and filters can be written and applied to large data sets using distributed processing.

Graph Analytics

Spark-tk uses a Graph object for its scalable graph representation, based on a Frame holding vertices and another Frame holding edges. Graph representations are broadly useful.

• Use Case: linking disparate data with arbitrary edge types and then analyzing the connections for powerful predictive signals that can otherwise be missed with entity-based methods.

Working with graph representations can often be more intuitive and computationally efficient for data sets where the connections between data observations are more numerous and more important than the data points alone. Spark-tk brings together the capabilities to create and analyze graphs, including engineering features and applying graph-based algorithms. Since the graphs are built using frames, Frame operations may be seamlessly applied to graphs.

• Use Case: applying a clustering algorithm to a vertex list with features developed using graph analytics.

Spark-tk supports importing and exporting graphs to the OrientDB's scalable graph database. Graph databases allow users to run real-time queries on their graph data.

Machine Learning

The toolkit provides algorithms for supervised, unsupervised, and semi-supervised machine learning using both entity and graphical machine learning tools. Examples include time-series analysis, recommender systems using collaborative filtering, topic modeling using Latent Dirichlet Allocation, clustering using K-means, and classification using logistic regression. Available graph algorithms such as label propagation and loopy belief propagation exploit the connections in the graph structure and provide powerful new methods of labeling or classifying graph data. Most of the Machine Learning is exposed through the Models API. The Models API provides a simplified interface for data scientists to create, train, and test the performance of their models. The trained models can then be used for predictions, classifications and recommendations. Data scientists can also persist models by using the model save and load methods.

Image Processing

Spark-tk includes support for ingesting and processing DICOM images in a distributed environment. DICOM is the international standard for medical images and related information (ISO 12052). Sparktk provides queries, filters, and analytics on collections of these images.

Documentation

API Reference pages for Python and Scala are located here.

Example:

Create a TkContext

>>> from sparktk import TkContext

>>> tc = TkContext()

Upload some tabular data

>>> frame1 = tc.frame.create(data=[[2, 3],
...                                [1, 4],
...                                [7, 1],
...                                [1, 1],
...                                [9, 2],
...                                [2, 4],
...                                [0, 4],
...                                [6, 3],
...                                [5, 6]],
...                          schema=[("a", int), ("b", int)])

Do a linear transform

>>> frame1.add_columns(lambda row: row.a * 2 + row.b, schema=("c", int))

>>> frame1.inspect()
[#]  a  b  c
=============
[0]  2  3   7
[1]  1  4   6
[2]  7  1  15
[3]  1  1   3
[4]  9  2  20
[5]  2  4   8
[6]  0  4   4
[7]  6  3  15
[8]  5  6  16

Train a K-Means model

>>> km = tc.models.clustering.kmeans.train(frame1, "c", k=3, seed=5)

>>> km.centroids
[[5.6000000000000005], [15.333333333333332], [20.0]]

Add cluster predictions to the frame

>>> pf = km.predict(frame1)

>>> pf.inspect()
[#]  a  b  c   cluster
======================
[0]  2  3   7        0
[1]  1  4   6        0
[2]  7  1  15        1
[3]  1  1   3        0
[4]  9  2  20        2
[5]  2  4   8        0
[6]  0  4   4        0
[7]  6  3  15        1
[8]  5  6  16        1

Upload some new data and predict

>>> frame2 = tc.frame.create([[3], [8], [16], [1], [13], [18]])

>>> pf2 = km.predict(frame2, 'C0')

>>> pf2.inspect()
[#]  C0  cluster
================
[0]   3        0
[1]   8        0
[2]  16        1
[3]   1        0
[4]  13        1
[5]  18        2