Forget the Models – You Need a Machine Learning Pipeline


Machine Learning has taken the business world by storm and left a trail of buzzwords in its wake. Models, algorithms, deep learning – these are the terms that will turn heads at parties. But the reality is they’re only one part of the equation… and not even the hard part.
Most learning algorithms can only interpret clean, tidy sets of data. In the real world, data is messy and unstructured. What your business needs is a multi-step framework which collects raw data, transforms it into a machine-readable form, and makes intelligent predictions — an end-to-end Machine Learning pipeline.


Preprocessing → Cleaning → Feature Engineering → Model Selection → Prediction → Deployment


In this post, we break down the steps of the Machine Learning pipeline and explain why your business needs each one in order to deploy a scalable ML solution.


Step 1: Data Preprocessing

Data is the first ingredient in any machine learning recipe, and gathering and consolidating that is the first instruction. If your business is starting from scratch, this can be a huge undertaking. Raw data must be preprocessed in real-time, at huge scales, from disparate sources, and in various formats. Where ML is concerned, the more data the better. If you can get this right, you’ll have a funnel which pulls data into your pipeline from every corner of your business


Step 2: Data Cleaning

Next, your data flows to the Cleaning step. Unreliable data can confuse even the most sophisticated ML algorithms. Scrubbing away outliers, missing values, duplicates, class imbalances, and other errors ensures that your data tells a consistent story that an algorithm can learn.


Step 3: Feature Engineering

Once it’s been cleaned, your data is ready to be transformed through feature engineering. A feature is a way to describe each data point with the information you’ve collected. These will serve as inputs into the learning process, so the trick in this step lies in creating features which represent your data in the most predictive ways possible.

Feature engineering is typically the most challenging and critical step in the ML pipeline. It requires much more than math chops – choosing the most predictive from an infinite pool of potential features requires intuition about the problem at hand. Not only will your pipeline need to select the right ones, it must also crunch through huge amounts of data to build them.


Step 4: Model Selection

Once the data wrangling is over, the fun stuff begins. When you feed data into an ML algorithm, it learns the relationship between your features and your goal. This relationship is your Machine Learning model. Give it enough of the right features and your algorithm will learn a model which reliably turns data into predictions.

Mathematically speaking, training a model is the most complex step in the pipeline. But practically it’s usually the most straightforward. Today’s algorithms are mostly tried and true. Their methods are general purpose, agnostic to the problem you’re trying to solve. Just plug in your data and your goal — whether it’s predicting customer behavior or diagnosing disease — and the algorithm will do the rest.

Still, you’ll have to make some choices. There are hundreds of distinct Machine Learning algorithms (neural networks, logistic regression, decision trees, etc). Each one learns in different ways, and it’s difficult to know which will perform best on your dataset.

So how do you know which one to use? Stage a competition. Use labeled data to train a handful of models with a handful of algorithms. Evaluate how well each one performs on a fresh set of data, and a winner will emerge.


Step 5: Prediction

When you’ve got the best model, put it to work making predictions on new, unlabeled data. For example, your model might train on past examples of customers who have churned in order to learn which behaviors are indicative of churn in general. Once it’s seen enough historical data, you can use that model to predict each customer’s current likelihood of churning in the future.


Step 6: Deployment

You’ve built a step-by-step pipeline which starts with raw data and ends with intelligent predictions. Plugging this pipeline directly into your business bends it pipeline into a continuous loop. The system is always learning from fresh data, always generating fresh predictions, and your business is always reaping the benefits.

How do I get started?

The path toward implementing Machine Learning in your business is well-defined but full of roadblocks, none bigger than preparing your unstructured data. To integrating predictive intelligence across your business, you need an end-to-end pipeline which handles the entire ML life cycle from start to finish.

Most businesses know that Machine Learning is a must for staying ahead of the competition. But for many, the cost and complexity of building a pipeline is just too high. Organizations often spend multiple years and millions of dollars building a scalable pipeline. Others are too daunted to even get started.

Vidora is hard at work lowering these barriers so that any business can deploy ML in minutes. Cortex, our self-service Machine Learning platform, automates every step of the pipeline so that you can focus on applying ML to your organization’s most pressing problems. Marketers, product managers, and analysts shouldn’t need to moonlight as data scientists in order to make data-driven decisions. With Cortex in their toolbox, they don’t have to.