24. ホークス型モデルの効果
Evolution with Hawkes interaction feature averaging processes
24
User’s and item’s features can evolve and be
influenced by the characteristics of their interactions.
…
We choose the exponential kernel ?? ? = ???(???)
to reduce the influence of each past event. In other
words, only the most recent interaction events will
have bigger influences.
36. 予測と評価
Predictions and metrics
36
a. アイテム?レコメンデーション
? ある時刻tにおけるユーザの生存確率Sui(t)を算出
? Sui(t)を昇順に並べてアイテムをレコメンド
? 評価はMean Average Rank(MAR)
b. 時刻予測
? 次イベントの確率密度f(t)を算出
? 評価はMean Absolute Error(MAE)
37. 比較
competitors
37
? TimeSVD++: the classic matrix factoraization
? FIP: a static low rank latent factor model
? STIC: a semi-hidden markov model (時刻予測のみ)
? Poisson Tensor: Poisson Regression
? Dynamic Poisson: the Poisson factorization with a
Kalman filter(アイテム?レコメンドのみ)
? Low Rank Hawkes: a Hawkes Process based model
epock
based
models
43. 実世界データでの結果
Results - Yelp
43
LOW RANK HAWKES might be good at differentiating the rank of
intensity better than COEVOLVE . However, it may not be able to learn
the actual value of the intensity function accurately. Hence our method
has the order of magnitude improvement in the time prediction task.