How to Leverage Machine Learning via Predictive APIs

How to implement Predictive APIs as part of your business strategy

There’s no doubt that whether predictive APIs are machine learning or, probably more accurately, its forebear, there’s no doubt that there is a value in their intelligence.

The above debate may have left you with a heady feeling but, when it comes down to it, predictive APIs lead to some simply sound business strategy applications. Here are just a few ways it can be applied to predict customer behaviour and behaviour anomalies.

  1. Churn: Subscription-based business models use churn analysis by learning what customers do, their profiles and whether they leave your service or not. You can then predict who might leave next and push extra resources like support their way.
  2. Up-sell: To detect who would most likely be interested in another product you’re selling and which product to push.
  3. Pricing optimization: What’s the best pricing based on characteristics of a product, often including competitive data.
  4. Sales optimization: Relating to the one above, how many sales of this product for this price, so you can play with the price and see how it affects sales.
  5. Fraud detection: By developing predictive patterns of fraud-like behavior.
  6. Credit scoring: To create a solid decision-making tool that takes the breadth of data that gets poured into this decision, while maintaining a clear causation.

As you can see, machine learning’s application can range from security to sales and marketing to even transportation and chemistry. It’s best to look at the data you have, what information you would like to get out of it, and to see if you can create repeatable data samples and examples.

What kind of predictive apps can you build via predictive APIs?

Of course, for any benefit, there’s an app for that. Here are a few of countless examples of applications built with predictive APIs enabling their machine learning.

  • Inbox by Google: It learns your email activity patterns in order to help prioritize what your eyes see, giving higher inbox placement for what you deem important.
  • Google now: Predicts where you’re going next.
  • Mention: Twitter and Facebook sentiment analysis of your brand, as overall positive, negative or neutral.
  • Snips: Makes crowd predictions on passenger flow across public transit, based on context like events and weather.
  • Zillow: Home value estimator.
  • Google Translate: language detection before translation is made.
  • Shadows of Mordor: In this video game, your rivals learn your moves and then start using them against you. Here, if you outsmart the machine, you get extra points.

From shopping to pricing to translating to almost impossible games, predictive APIs can improve the user experience while also improving your ability to sell something to that user.

How can you build a machine learning model via predictive APIs?

If you want to go the route of the folks quoted in this article, you can actually study for a PhD in the world of machine learning. The objective here is to get machine learning software into the hands of domain experts, and, eventually, to make ML as accessible as website development now is. This accessibility is done via predictive APIs.

The two phases of predictive APIs

  1. Train a model
    model = create_model (dataset)
  2. Predict with a model
    predicted_output = create_prediction (model, new_input)
Predictive APIs example code

How to distinguish between kinds of predictive APIs

Different needs require different kinds of predictive APIs. Dorard presented the most common ones in order from most complex, study-needed to the most abstract, low level to high level.

  1. Machine Learning Algorithm API: An API that exposes a machine-learning algorithm, which requires you to understand which algorithms to pick in which circumstances. It dives into a bit of ML and allows you to develop models of production.
    1. ErsatzLabs: Deep learning.
    2. BigML: Exposes decision tree algorithms allowing you to play with the perimeters.
    3. Random forests, which work in the opposite way, collecting a multiple of decision trees, and then using that to predict the individual ones.
    4. Microsoft Azure: Exposes ML algorithms.
  2. Automated Predictive API: Much easier to use, not needing to specify about algorithms or parameters. Each of these works with classification and regression.
    1. BigML
    2. Google Prediction API
    3. Wolfram Cloud
    4. DirectedEdge: Recommendations for the cloud
    5. Server-based recommendations to work offline
  3. Text Classification API: The input of your ML problem will be bits of text, with the outputs going to be classes.
    1. uClassify
    2. MonkeyLearn
    4. MetaMind
  4. Vertical Predictive API: An API that focuses on one particular problem and it enriches your data with data from other sources like clients’ social networks.
    1. Framed Data: churn
    2. Sift Science: fraud detection
    3. Infer: lead scoring
    4. personal assistant
  5. Fix-model Predictive API: No training needed, the model is already there and fixed.
    1. For text-specific parameters: Datumbox, AlchemyAPI, Semantria — like within Mention or for gender detection
    2. For “Siri-like” for predefined queries: Maluuba
    3. For vision: AlechemyAPI — like those that can detect for porn or object detection

How should you manage your predictive API research?

Dorard offers the seven questions he asks to decide whether an API will be good before starting to use it.

Can I get into using it right away or do I have to talk to a sales rep? This really comes down to the question of how fast you need it.

How much does it cost? Is there a Freemium option to test out and build a proof of concept on? Or do you want to pay more to get something more customized? If there isn’t pricing on the site, then you can assume you have to go through a sales rep and then it’ll cost you more time and money. He recommends particularly checking out how the pricing structure handles the amount of data and if it works with your financial plan.

Will you need to tune any parameters to get the predictive API to work for you? “Time that is spent doing that is time not spent getting value from predictions or working on the data that you’re sending to the APIs, cleaning the data, enriching the data.”

Is it cloud or server-based? Are you going to have to worry about scaling the ML system or does it just run in the cloud?

What kind of data do you need to train? If you don’t have to provide any training data, you’re using a fixed model. Otherwise, you have to spend time preparing that data. There are two kinds of predictive APIs. Generic predictive APIs let you specify your own data which is then used to make your own predictions. Specialized predictive APIs allow users to get answers to a specific question for an already existing model. The latter is significantly quicker to get started but this may not suit your needs.

What kind of machine learning problem are you looking to solve? Is it a classification, regression or recommendation model.

Who owns the model? Is the API provided in a white-box form to be downloaded on your laptop or phone? Do they simply tell you which ML algorithm was used, enabling you to replicate the results you get, but you can’t guarantee exact replication? Or, like Google Prediction APIs, is it a black-box model? Do what works best for your objectives.

Predictive APIs are for play first!

When it comes down to it, machine learning can be an overwhelming idea to take on. Instead of focusing on the how, we recommend getting right into it, going for predictive APIs that you can get into with a free trial to see if it fits you. Once you get into it, machine learning is a big word to do awesome things with your big data.

Tell us about your predictive API experience below!

images courtesy of Why And How To Leverage Predictive APIs In Any Application

Be sure to read the next Predictions article: Daily API RoundUp: Predikt, Pyrus, PatrolServer, ZipBooks


Comments (0)