Notice: this is a preliminary description. Details will follow soon.

Given a XING user, the goal of the RecSys challenge is to develop a recommender system that predicts those job postings that a user will click on.

Background: http://2016.recsyschallenge.com/

A more detailed description will follow soon.

• impressions - which items were shown to which user at which point in time?
• interactions of the users - clicks on job postings
  • ID of the user (anonymized) who clicked
  • ID of the item (anonymized)
  • timestamp (Unix timestamp, rounded to 5 min)
• details about users (including details about social network size, work experience):
  • ID of the user (anonymized)
  • jobrole ID
  • career level ID (e.g. beginner, experienced, manager)
  • industry ID (IDs represent industries such as "Internet", "Automotive", "Finance")
  • discipline ID (IDs represent disciplines such as "Consulting", "HR", "Engineering & Technical")
  • number of contacts (ranges such as less than 10, 10-50)
  • estimated years of experience (ranges such as 2-5 years)
  • educational background ID (classes such as: no university degree, Bsc/Msc, doctor)
• details about job postings such as:
  • ID of the job posting (anonymized)
  • jobrole ID (IDs represent job roles such as "Java Software Developer", "Account Manager", etc.)
  • tag IDs (also from full-text description) (IDs represent, for example, skills such as "Java", "Data mining", "Leadership", etc. but may also represent jobroles or other concepts that are mentioned in the text)
  • career level ID (e.g. beginner, experienced, manager)
  • industry ID (IDs represent industries such as "Internet", "Automotive", "Finance")
  • discipline ID (IDs represent disciplines such as "Consulting", "HR", "Engineering & Technical")
  • location ID (city-level; IDs will follow a postal-code logic so that distances can be estimated)

The data that is provided is a sample of XING's dataset. XING users will be splitted into two groups A and B. The interactions and anonymized demographics about those users will be released. The dataset will moreover be enriched with a small percentage of fake users. Anonymization of profiles will be done by mapping jobroles, etc. into a numeric space and by using ranges for fields such as the number of contacts or years of experience.

Current plan: One large dataset that covers more than 20 Million interactions (possibly from a time period of around 6 months), more than 1 Million anonymized users and more than 500k job postings will be released for analysis and training.