will you “reconsume” the near past? fast prediction on ...lastfm, brightkite, gowalla data sets....
TRANSCRIPT
The Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI-15, Austin, USA).
Contact: [email protected], [email protected].
Reconsume items which have recently been consumed by a same user.
Alternated now and then with novel consumption behaviors.
Short-Term REConsumption (STREC) Behaviors
Jun Chen, Chaokun Wang, Jianmin Wang
Will You “Reconsume” the Near Past? Fast Prediction on Short-term Reconsumption Behaviors
School of Software, Tsinghua University, Beijing 100084, P.R. China
A switch problem to narrow down the problem domains.
Broad applications, e.g. intelligent marketing, recommendersystems, web revisitation, information re-finding.
Challenges: 1) high dynamics of short-term behaviors; 2)multiple influential factors; 3) No representative features.
Motivations
Feature Extraction
A. Item Popularity
Average normalized frequencies of items.
B. Item Reconsumption Ratio
Average normalized reconsumption probability of items.
C. User Reconsumption Ratio
Reconsumption probability of users.
D. Window Reconsumption Ratio
Fraction of reconsumptions in the current window.
ℎ𝐼𝑃 𝑥 =log(1 + 𝑓𝑟𝑒𝑞(𝑥))
max𝑦∈𝐗log(1 + 𝑓𝑟𝑒𝑞 𝑦 )ℎ𝐼𝑃 𝑊𝑘 =
1
|𝑊𝑘| 𝑥∈𝑊𝑘
ℎ𝐼𝑃 𝑥
ℎ𝐴𝐼𝑅𝑅 𝑥 = log(1 + 𝑢∈𝐔 𝑡∈𝑇𝑢 𝕝𝑡=𝑥∧𝑡∈𝐶𝑘
𝑢,𝑡
𝑢∈𝐔 𝑡∈𝑇𝑢 𝕝𝑡=𝑥)
ℎ𝐼𝑅𝑅 𝑥 =ℎ𝐴𝐼𝑅𝑅 𝑥
max𝑦∈𝐗ℎ𝐴𝐼𝑅𝑅 𝑦ℎ𝐼𝑅𝑅 𝑊𝑘 =
1
|𝑊𝑘| 𝑥∈𝑊𝑘
ℎ𝐼𝑅𝑅 𝑥
ℎ𝑈𝑅𝑅 𝑢 = 𝑡∈𝑇𝑢 𝕝𝑡∈𝑊𝑘
𝑢,𝑡
|𝑇𝑢|ℎ𝑊𝑅𝑅 𝑊𝑘 = 1 −
|𝐷𝑆(𝑊𝑘)|
𝑘
A
B
C D
argmin𝐰𝒬∗𝒬 𝐰 =
𝑢∈𝐔
𝑡∈𝑇𝑢
( 𝐰𝑇𝑑𝑖𝑎𝑔 𝐱𝑢,𝑡2𝐰− 𝕝
𝑡∈𝑊𝑘𝑢,𝑡)2
s.t.𝐰𝑇𝐰 = 𝟏
argmin𝐰𝒬∗𝒬 𝐰 =
𝑢∈𝐔
𝑡∈𝑇𝑢
( 𝐰𝑇𝑑𝑖𝑎𝑔 𝐱𝑢,𝑡2𝐰− 𝕝
𝑡∈𝑊𝑘𝑢,𝑡)2 + 𝜆𝐰𝑇𝐰
argmin𝐰ℒ∗ℒ 𝐰 =
𝑢∈𝐔
𝑡∈𝑇𝑢
(𝐰𝑇𝐱𝑢,𝑡 − 𝕝𝑡∈𝑊𝑘𝑢,𝑡)2
s.t. 𝒊𝐰𝒊 = 𝟏
argmin𝐰ℒ∗ℒ 𝐰 =
𝑢∈𝐔
𝑡∈𝑇𝑢
(𝐰𝑇𝐱𝑢,𝑡 − 𝕝𝑡∈𝑊𝑘𝑢,𝑡)2 + 𝜆
𝒊
𝐰𝒊
Fast Prediction Methods
1. Linear Method
Prℒ 𝑢, 𝑡 = 𝐰𝑇𝐱𝑢,𝑡
𝐱𝑢,𝑡 = {ℎ𝐼𝑃 𝑊𝑘 , ℎ𝐼𝑅𝑅 𝑊𝑘 , ℎ𝑈𝑅𝑅 𝑢 , ℎ𝑊𝑅𝑅 𝑊𝑘 }𝑇
2. Quadratic Method
Pr𝒬 𝑢, 𝑡 = 𝐰𝑇𝑑𝑖𝑎𝑔 𝐱𝑢,𝑡
2𝐰
Experiments
Collected a new App using data set, ManicTime.
LastFM, BrightKite, Gowalla data sets.
On average, 80% prediction accuracy on 4 real-world datasets.
SVM method is prone to be overfitted.
The linear and the quadratic methods do not have much difference inprediction accuracy.