Algorithms for recommender systems.

TODO list:

• item cf (item based collaborative filtering)

  • different similarity measurements (cosine, L2-distance, Jacard, Pearson)
  • item popularity and freshness

• user cf

  • user-user similarity calculation
  • topic constrained user cf (for news recommendation)
  • expert cf (firstly detect domain experts, and rec items to users based on similar experts)

• factorization machines

  • svd++
  • libfm

• keyword/topic based methods

  • variable combinations of keywords, like N-gram
  • Bayesian inference of p(topic|user) and topic popularities

• random walk

  • transition matrix

• SNS based methods

  • trust cf (user-friend relationships instead of user-user similarities in user cf)

• LBS based methods

  • local hot reranking

• demographic based methods

  • rec by demographic segmentation

• context-aware methods

  • tensor factorization
  • sequential pattern based prediction
  • learning using context of user choise

• other methods

  • editor's choise
  • rec by top list
  • rec by commodity consumption cycle

• recommend items to items

  • click cooccurence based methods
  • keyword, topic based methods
  • personalized relevance model
  • recommend products by similar images

• recommend users to users

  • EdgeRank
  • Twitter's algorithm

• merging

  • blending (weights of algorithms, weights of items of algorithms, user feedback)
  • ensemble (LR, RBM, GBM, random forest)
  • switching (switch methods by context)
  • cascading (multi-level model)

• evaluation

  • offline metrics (MAP, nDCG, AUC, diversity...)
  • online metrics (CTR, percent conversion, user active degree, long tail item exploration)

• challenges

  • page optimization: by explanation, by method, clustering (k-means, AP clustering, HDP)
  • long-term interest vs. short-term action
  • exploitation vs. exploration: multi-armed bandits
  • accurate vs. diverse
  • freshness vs. stability
  • navigation vs. attention