Stop Thinking, Just Do!

BASICS OF PREDICTIONIO

PredictionIO acts as a server that collects data and serve prediction results through REST APIs/SDKs. Conceptually, the 3 main building blocks are: App, Engine and Algorithm.

App

App in PredictionIO Server is like a database or collection in a database server. It usually corresponds to the application you are building. Relevant data, such as user behavior, is collected by an app. An app contains one or more prediction engines. App data is shared among these engines.

Engine

An engine must belong to a prediction type (or engine type), such as Item Recommendation or Item Similarity. Each Engines process data and construct predictive model independently. Therefore, every engine serves its own set of prediction results. In an app, for example, you may create two engines: one for recommending news to users and another one for suggesting new friends to users. An algorithm must be deployed in each engine.

Algorithm

A number of built-in algorithms are available for use in each type of engine. An algorithm, and the setting of its parameters, determines how predictive model is constructed. In another word, prediction accuracy or performance can be improved by tuning a suitable algorithm. PredictionIO comes with tools and metrics for algorithm evaluation.

DATA COLLECTION

To make prediction, PredictionIO app collects mainly 3 types of data: User Data, Item Data and Behavioral Data.

User Data

Each user record corresponds to a unique user or customer of your application. The only required data attribute is a user ID string, which normally matches with the one in your database. You may also provide any extra data attributes, such as age, gender, location and income.

Item Data

An item record corresponds to an object of your application. An object can be anything, e.g., a book, a restaurant, a document, a product, a video or any other content. Each item record requires 2 data attributes: an Item ID string and an array of item types. You may also provide any extra data attributes to each item record. Similar to an user ID, an item ID normally matches with the ID of the corresponding object in your database. An item type is a string that categorizes an item.

Item Type

There may be one or more type of items in your application. They can be categorized with item type strings. Each item can also belong to more than one item type. Let say you are building a software marketplace service that categorizes software as Freeware or Shareware. Each software is further categorized as Developer Tool, Database Server and System Utility etc. In this case, the item type array of an item record may look like this: [‘Software’, Freeware’, ‘Developer Tool’].

Behavioral Data

User-to-item actions and user-to-user actions are collected as behavioral data. They are used for constructing predictive models. A behavior record looks like this: User A likes Item X, where like is a user-to-item action type.

Built-in Action Type

PredictionIO comes with some user-to-item action types:

like
A user explicitly likes an item.

dislike
A user explicitly dislikes an item.

rate
A user gives rating to an item. It is a 1-5 scale score. 1 is the worst and 5 is the best. (so 3 is neutral) Please convert the score manually if your application follows a different rating scale.

view
A user browses an item. It may be regarded as a soft implicit preference.

conversion
A user performs an action that implies the strongest preference, e.g. when a user purchases a product, when a user download the content or when a user signs up a service.

You may also add your own custom action type.

PREDICTIVE MODELING

Predictive modeling is a process by which a model is built (or trained) to predict future or unknown outcome. The accuracy and the performance of the model is determined by the algorithm you select as well as the parameter settings of this algorithm.

Every engine manages a predictive model independently in PredictionIO. In another word, there is one deployed algorithm running in each engine.

Note Depends on the implementation of the engine type, the system can train the model with the collected data in batch mode or in real-time.

Algorithm

Algorithm is sometimes called ‘learning algorithm’ in Machine Learning. It determines how your system learns to predict from data. Each type of engine in PredictionIO comes with different built-in algorithms available for use. An algorithm is deployed with default parameters when you create a new engine. You can deploy another one to fit your needs.

Algorithm Parameter

Some algorithms require you to specify parameter values. (They are sometimes referred to as hyperparameters in academic disciplines.) Parameters adjust how an algorithm learns. For instance, a regularization parameter tries to ensure that the model does not overfit its data. Some algorithms provide an automatic tuning feature to help you find better parameter settings. This feature requires a lot of computational resources though.

Choose an Algorithm

Algorithms can rely on very different assumptions and theories. There is no one-size-fit-all solution that is suitable for every prediction problem. To improve prediction accuracy for your specific case, you may need to evaluate various combination of algorithms and parameter settings. To learn more about it, read Accuracy Optimization.

SYSTEM DESIGN OF PREDICTIONIO

PredictionIO is mainly built with Scala. Scala runs on the JVM, so Java and Scala stacks can be freely mixed for totally seamless integration. PredictionIO Server consists of a few components:

• Admin Server
• IO Server
• Scheduler
• Data Store
• Data Processing Stack

Admin Server

PredictionIO’s Admin Server component provides a web interface for developers to manage applications, engines and algorithms. It is built on top of Play Framework.

IO Server

IO Server offers scalable REST API services to communicate with your web or mobile app. It is responsible for handling data input and prediction output. It is built on top of Play Framework.

Scheduler

A scalable scheduler that can be used to manage schedules for executing tens, hundreds, or even tens-of-thousands of jobs. Quartz is the default scheduler.

Data Store

Data store manages the collected data, the predictive model and the cached prediction results. MongoDB is the default data store.

Data Processing Stack

Built on top of solid data frameworks and technology, such as Hadoop, Cascading, Scalding and Mahout, PredictionIO can handle a huge amount of data efficiently. A variety of machine learning algorithms are available for you to implement with just a few clicks.

References

[1] PredictionIO Guide, PredictionIO Concepts, 2013.