まとめページへのリンク

Research Track Oral Presentations: RT11 Clustering and Visualization

Deep Landscape Forecasting for Real-time Bidding Advertising

Kan Ren (Shanghai Jiao Tong University); Jiarui Qin (Shanghai Jiao Tong University); Lei Zheng (Shanghai Jiao Tong University); Zhengyu Yang (Shanghai Jiao Tong University); Weinan Zhang (Shanghai Jiao Tong University); Yong Yu (Shanghai Jiao Tong University)

Bid landscape forecasting

Simple statistic -based (KDD16, ECML16)
Function form assumption

Survival analysis methodology

Survival analysis: time-to-event data analysis

Related work

Deep learning with survival analysis
Deep landscape forecasting
Deep learning solution with solution
Inference
Losing probability by multiply conditional prob

Model architecture.
Feature vector x and varying price feature vector
Loss function:

uncensored data
Based on C.D.E

Visualized analysis

Conclusion

A novel deep learning method for market price modeling
And distribution learning , in real-time bidding advertising
Fine-grained forecasting for individual auction

QA

features

User from web brousa
Adviser itself

Applied 2nd price auction

Background

Real time bidding

Predicting Path Failure In Time-Evolving Graphs

Jia Li (The Chinese University of Hong Kong); Zhichao Han (The Chinese University of Hong Kong); Hong Cheng (The Chinese University of Hong Kong); Jiao Su (The Chinese University of Hong Kong); Pengyun Wang (Noah’s Ark Lab, HuawTechnologies); Jianfeng Zhang (Noah’s Ark Lab, Huawei Technologies); Lujia Pan (Noah’s Ark Lab, Huawei Technologies)

Focus: Path classification in a time 0rvoluving graph, which predicts the status od a path in the near future

Define:

Time-evolving graph
Path availability
Path classification

Three properties

Node correlation
Graph structure dynamics
Temporal dependency

Model: two-layer LRGCN, obtain the hidden representation and Self-attentive
Static graph modeling
Relational GCN (R-GCN) by Kipf er al

Two-hop simplified
Adjacent graph snapshots modeling
Before diving into sequence of graph snapshot, we first focus on two adjacent time step t-1 and t
Time- evolving graph unit

Two challenges

Size invariance
Node importance
SAPE: self-attentive path embedding method
(Data: predicting path failure)

The benefit of graph evolution modeling
Training efficiency

conclusion
Path classification in time-evolution

Pairwise Comparisons with Flexible Time-Dynamics

Lucas Maystre Spotify lucasm@spotify.com Victor Kristof EPFL victor.kristof@epfl.ch Matthias Grossglauser EPFL matthias.grossglauser@epfl.ch

background

How can we quantify the skill of France team?
relate work: Latent skill model
Actual setting: Data come with a timestamp

We might want to ask
How strong was France in 1972? In 2018?

This work: kickscore: skill becomes a (latent) stochastic process
Covariance function defines time dynamics
Model is conditionally parametric
Covariance function
(Model inference: )
Inference algorithm
For each observation: using EP [minka, 2001] or CVI[khan et al ., 2017]
For each item: using SSM reformulation

Conclusion:

Pairwise comparison model with Flexi bee time-dynamics
Linear-time inference algorithm

Modeling Extreme Events in Time Series Prediction

Daizong Ding, Mi Zhang∗ Fudan University, China Xudong Pan, Min Yang Fudan University, China Xiangnan He University of Science and Technology of China

background:

time series prediction, RNN

relate-work:

extreme events in time-series data

Estimated distribution of labels y
Extreme event problem in DNN
Recalling extreme event in history
Attention mechanism
Extreme value theory
Extreme value loss
Conclusion:
Optimization:
Extreme events in time-series data
Why DNN is innately weak in predicting extreme events
(A framework )

Adversarial Substructured Representation Learning for Mobile User Profiling

Pengyang Wang∗ Missouri Univ. of Sci. and Tech. MO, USA pwqt3@mst.edu Yanjie Fu∗† Missouri Univ. of Sci. and Tech. MO, USA fuyan@mst.edu Hui Xiong Rutgers University NJ, USA hxiong@rutgers.edu Xiaolin Li Nanjing University Nanjing, China lixl@nju.edu.cn

background:

Mobile user profiling, similarity graph of users, transportation, OD pairs
Adaptive interfaces by inferring trip
Human activities are spatially , temporally, and socially structural
Spatial-temporal transition patterns
(Given a user and corresponding user activity)
Preserving entire-structure
Preserving substructures
Approximating substructure detector
(Summary: generator auto encoder linked with an approximated )
Study of Node and circle substructures

Research problem

Method users as activity graphs
Formulate modeling specific interests as preserving substructures of user activity graphs
Propose an adversarial substructure learning model to integrate substructure into representations

relate-work:
Conclusion: