Post on 30-Jul-2020
transcript
Bayesian Poisson Tensor Factorization (BPTF)
for feature selection from GDELT
Chennan Zhou, Feiyan Liu, Jianbo Gao
March 22 , 2017
➢ GDELT collects nearly all events about human activity in the world since 1979.
➢ How can we analyze the events we are concerned?
➢ Will BPTF proposed by David M. Blei and Hanna Wallach help us extract the
desired events?
Purpose
➢ Answer: extract the relevant events from. However, this is difficult since events
in GDELT do not contain sufficient information as which one is precisely which,
and each day hundreds of thousands events happen.
➢ Discussions
Outline
➢ Introduction: Tensor Factorization
➢ The decomposition of the events from GDELT by BPTF
𝑇 =
𝑖=1
𝑘
𝜋𝑖𝑢𝑖 ⊗𝑢𝑖 ⊗+𝜀𝑅.
➢Tensor analogue of matrix eigen-decomposition.
What is tensor (CP) factorization?
➢Goal: Given T with noise, 𝜀𝑅, recover factors 𝑢𝑖.
How BPFT works?
➢ Input
• 4 - way tensor, whose structure is as:
<Sender, Receiver, ActionType, Date>
• Elements (𝑦𝑖𝑗𝑎𝑡) of a tensor are the number of events that sender 𝑖 acts
on receiver 𝑗 with action-type 𝑎 on date 𝑡.
➢ Output
• k components, each of which consists of four factors:
Senders, Receivers, Action-type, Time-steps.
Real Event: In July 2016, four consecutive violent attacks took place in Germany in a week.
Violent Attacks In Germany
Violent Attacks In Germany
Decomposition compare with the original time series
Real Event: From June to July 2016, three successive terrorist attacks took place in Turkey.
Successive Explosion In Turkey
Successive Explosion In Turkey
Decomposition compare with the original time series
Real Event: On the evening of July 14, 2016, a truck was deliberately driven into crowds in Nice,
France, resulting in the deaths of 86 people and injuring 434.
Terrorist Attack In France
Terrorist Attack In France
Decomposition compare with the original time series
July 12, 2016 the South China Sea Arbitration
Examples that some of the components couldn’t be related to certain event
Examples that some of the components couldn’t be related to certain event
➢ When to use BPTF:
Discussions
• When we don’t know what happened in a period of time, it can extract some
events from database and obtain the evolution of events;
• It can provide guidance for the selection of features when we do not know
what fields to select from database.
➢ can not yet:
• Extract precisely event we are concerned;
• Predict the evolution of major events in the future;
• Describe complex network among countries during the evolution of a particular
event.
THANKS