CIKM CUP 2016 Track 1: Cross-Device Linking

The goal of the competition is to find browsing logs which belong to the same user.

More details at http://cikmcup.org and https://competitions.codalab.org/competitions/11171

My solution

• Convert user ids into integers so they occupy less RAM
• Split the train data into 2 folds based on connected components
• Use Elastic Search and More-Like-This queries to find top pair candidates
• Split logs into sessions (using 30 minute intervals) and compute the user "profile" (log features):
  • Number of sessions
  • Clicks within session
  • Duration of breaks between sessions
  • Starts and ends of sessions
  • Title-based, Domain-based and Url-based similarities within sessions
• For candidates retrieved with Elastic Search, compute the following features:
  • Absolute difference between the profile features
  • Cosine between domains, full urls and titles
• Train an xgboost model for predicting if a candidate pair corresponds to the same user or not

Files:

• 1_prepare_data.py: preprocesses the data
• 2_data_to_elastic.py: puts the log data to elastic search
• 3_candidates_elastic.py: uses elastic search for retrieving top 70 candidates for each user
• 4_session_vectorizers.py: "trains" count vectorizers for urls, domains and titles for user sessions
• 5_user_profiles.py: extracts profile information from each user log
• 6_pair_features.py: computes features for each candidate pair
• 7_model.py: trains the xgb model and creates the submission file

Presentation

• This solution was presented at Berlin Machine Learning meetup. See the slides here.