How to make a wise machine learning platforms comparison

Data features

Organizations should first consider the data types and data management features the product or service offers and whether they are available for a customer's on-premises systems, in a private or public cloud, or in a hybrid IT environment. For instance, does the platform support big data initiatives and allow users to build machine learning models with data gathered from a variety of sources, including text, images, multimedia and location systems?

Data preparation features should also factor into a buying decision and machine learning platforms comparison, including data aggregation, sorting, filtering and integration. Does the platform handle data preparation faster by identifying common quality problems, for example?

Also important are the visualization and exploratory tools the product or service offers. Many data discovery tools use visual presentation tools that can aid users in finding patterns or specific items within a data set. Visualizations take the form of dashboards, reports, charts and tables.

Traditionally, data visualization was limited to standard charting, limited graphical representations, key performance indicators or other simple methods. But the rise of big data/analytics has also emphasized user-based data analysis and discovery, access to larger volumes of data, and the ability to create more advanced presentations.

Automation features

Automation is all the rage as organizations try to boost efficiency and speed time to market. That includes machine learning processes.

Some platforms for machine learning include the ability to automate multiple processes, such as creating, training, tuning and deploying models. The faster companies can complete these processes, the quicker they can get start using machine learning and see the business benefits.

For instance, some platforms have built-in monitoring tools that measure the model performance and automatically retrain the models to adjust to new data. Other machine learning platforms feature automatic model training and tuning within a user-specified time limit.

Another area with potential time and cost savings is IT resource provisioning. Some machine learning-related services automate all resource provisioning and monitoring so organizations can focus on model development and deployment without worrying about infrastructure.

Supported algorithm and modeling methods

Machine learning consists of many different algorithms and modeling methods, each with its own particular strengths and use cases. Any machine learning platforms comparison should include finding out which methods the platforms support.

At a high level, this breaks down into supervised vs. unsupervised learning methods. With supervised learning, data scientists and analysts who know what they want a machine to learn can expose it to a huge set of training data, examine the output and then alter the parameters if necessary until the appropriate results emerge. They can then have the machine predict the results of the new data to see what it has learned.

Organizations can use supervised learning for tasks like determining a potential customer's financial risk based on past financial performance. These models can also provide companies with insight into how customers will act in a given circumstance or customer preference based on past behavior.

With unsupervised learning, a machine explores a data set to identify inconspicuous patterns that link different variables, eventually grouping data into clusters based on statistical properties. An example of unsupervised learning is the clustering algorithm used to perform probabilistic record linking, which extracts connections between data elements to identify individuals and organizations.

Then there's deep learning, which is a way of performing unsupervised or reinforcement learning by mimicking some aspects of the way humans learn. This relies on neural networks to identify characteristics of data sets in great detail. Organizations can use deep learning for fraud detection and predictive maintenance.

Another machine learning platforms comparison consideration should be whether a platform offers the option of pre-trained models and build-your-own models. An organization might want to take the reins and train models using its own data or install pre-trained models that come with training data developed by the vendor. One of the benefits of pre-trained models is that users are able to score and classify new content right away rather than having to train the model first.

Integration, ease of use and management features

Among the other major factors to consider when looking at platforms for machine learning is how easily they integrate with other systems, as well as ease-of-use and management considerations.

On the integration side, consider the platform's API support. Companies will most likely be looking to integrate machine learning tools with a variety of other systems, so the ability to easily integrate is critical.

If there are any barriers to creating and training models, data workers might not be as open to using the platform and productivity might suffer.

In terms of ease of use and workflow, factors to consider include whether the platform guides users with features such as intuitive, graphical interfaces; lets multiple users concurrently analyze structured and unstructured data using visual interfaces; enables concurrent access to data in a multiuser environment; and generally supports users with a variety of skill sets.

Much of the focus with machine learning is on creating models. Does the platform allow developers, data scientists and other users to quickly and easily build, train and deploy models at any scale? If there are any barriers to creating and training models, data workers might not be as open to using the platform and productivity might suffer.

Container support is also important for a smooth workflow, and it is an important consideration in a machine learning platforms comparison. Can the platform work with open source containers, which some companies use to test their models before training or hosting them in production? Organizations might choose to run development and production workloads in containers to create shared and reusable environments.

Important management features include addressing data security and privacy, resource oversight, and governance. What provisions does the machine learning platform offer for data protection, user identification and access control? This is all the more vital in multiuser environments where various users share models and collaborate.

If you consider a cloud machine learning platform, does the provider offer an adequate security infrastructure to ensure the protection of client data?

For governance, something else to look into is whether the machine learning platform has a management console to enable centralized control of applications and data.

Open source options and resources

There are many open source resources that can serve as useful machine learning components. These include programming languages, notebooks, libraries, visualizations, data management platforms and frameworks. Here are brief descriptions of some of the leading options.

R
This is an open source programming language for statistical computing and graphics that is supported by the R Foundation for Statistical Computing. It is widely used by statisticians and data mining professionals to develop software for data analytics.

Python
An open source programming language maintained by the Python Software Foundation. It supports applications including web development, scientific and numeric computing, software development, and enterprise management.

Deeplearning4j
A deep learning programming library written for Java and Java virtual machines. This computing framework has broad support for deep learning algorithms.

H20
Produced by H2O.ai, H2O is an open source, distributed, in-memory machine learning platform with linear scalability. It supports the most widely used statistical and machine learning algorithms, including gradient-boosted machines, generalized linear models, deep learning and more.

D3.js (Data-Driven Documents)
This is a JavaScript library for manipulating documents based on data and particularly for producing dynamic, interactive data visualizations in web browsers.

Plotly
Plotly's JavaScript library is free, open source software that can be used to create sophisticated, interactive charts in JavaScript for finance, engineering and the sciences.

Jupyter
A web application that allows users to run live code and embed visualizations and explanatory text in one place. It was created by Project Jupyter, a nonprofit, open source project born out of the IPython Project as it evolved to support interactive data science and scientific computing across all programming languages.

Apache Zeppelin
A web-based notebook that enables data-driven, interactive data analytics and collaborative documents with SQL, Scala and other platforms. Zeppelin supports data ingestion, discovery, analytics and visualization, as well as collaboration.

Apache Spark
An open source, cluster computing framework maintained by the Apache Software Foundation. Spark provides an interface for programming entire clusters with implicit data parallelism and fault tolerance. It also powers libraries including SQL and MLlib for machine learning.

Hadoop
The Hadoop software library is a framework that supports distributed processing of large data sets across clusters of computers using simple programming models. It scales up from single servers to thousands of machines by design, with each offering local computation and storage.

TensorFlow
An open source software library for high-performance numerical computation. Its flexible architecture offers easy computation deployment across a variety of platforms, as well as for desktops, clusters of servers and mobile devices. The TensorFlow library comes with support for machine learning and deep learning.

Spark ML
An open source software package that provides a set of high-level APIs to help users create and tune practical machine learning pipelines. Spark ML standardizes APIs for machine learning algorithms to make it easier to combine multiple algorithms into a single pipeline or workflow.