ORP – The Open Recommendation Platform
The Open Recommendation Platform (ORP) is a distributed system of entities designed to deliver recommendations. It consists of recommendation providers and recommendation consumers that interact and communicate over a standardized protocol.
ORP was developed by plista and is the underlying protocol of the NewsREEL challenge that is hosted by the DAI lab. You can find all the specifications of ORP in our github.
When you register a recommender with ORP, we will push data, i.e. page impressions, clicks on item, information about news article changes, as a json message to your algorithm which then has a narrow time window to form a proper response and send it back. You can see the technical specification of how the data is structured and how requests and responses look like here:
The News Recommendation Evaluation Lab Challenge (NewsREEL) is a yearly challenge by the Distributed Artificial Laboratory (DAI lab) of the TU Berlin and plista to give researchers, professionals and individuals an unique opportunity in creating and testing recommendation systems.
For the challenge, we will redirect a certain amount of traffic from our system to you – meaning you get direct, unfiltered access to a steady stream of recommendation requests for you to work with. The goal is, of course, to create a recommendation system capable of providing useful news recommendations.This years 2017 NewsREEL challenge will start soon. If you want to find out more about it, check out the NewsREEL home page and if you want to participate in the challenge, look here.
The NewsREEL challenge will present you with a lot of real-world problems such as a continuous stream of data that has to be processed in near real time, changes in the set of items that are recommendable, noise, etc, but at the same time, this is the beauty of the challenge. There are two categories of the challenge that you have to master: NewsREEL Live and NewsREEL Replay.
NewsReel Live is a challenge that provides you with a real-world stream of recommendation requests from our servers. The data is spread across all participants and each participating algorithm has a time limit of 100ms to process a request and generate a recommendation.
When a user clicks on the recommendation your algorithm submitted, it will no longer be counted as an impression, but a click. The goal is for your recommender to have as high a CTR (click through rate) as possible.
The NewsREEL Replay challenge is similar to NewsREEL live, but here, the Idomaar system of the DAI lab is used to generate an offline stream. Even though it uses offline data, you still have the same time constraints as in the live case. In this task, the goal is the prediction accuracy which is defined as the number of successful recommendations divided by the total number of recommendations.
ORP stands for Open Recommendation Platform. ORP enables users to test and track their algorithm's statistics, including: impressions, clicks, ctr and any errors caught.
Yes. As ORP utilizes Twitter Bootstrap, the site and its components resize no matter what screen size you use. Although, the zoom functionality is imparied on small resolutions. A notebook or desktop is recommended for an optimal experience.
The Dashboard displays a general overview of your account. The area graphs display the overall statistics for each algorithm associated with your account. Both enabled and disabled algorithms are shown to help you gauge your results.
The Statistics page displays a single modular graph and table to convey the data accumulated over time. The graph and table assist in visualizing the growth and decay of the selected algorithms across local publishers.
The Debug page displays all errors caught or generated throughout the process of displaying your ads.
The account page allows editing of vital account information, such as: + Email + Password
The graphs with zooming and data manipulation functionality appear on the Statistics and Debug pages. These line graphs display a default period of 3 months prior from your current visiting date for all account related algorithms on all publishers. These dates and the data sources for the graph can be modified in the form directly below the corresponding graph. Above each graph there are buttons containing both a title and an averaged value. Upon clicking one of these buttons the correct dataset will be displayed in the graph. Scrolling over the line on the upper graph will display important data points and their associated date. The lower graph is utilized as a controller. By clicking and while still holding down the left mouse button, dragging the mouse, a new intermediate date selection will be rendered above.