Spatio-temporal Convolutional Residual Network for Commercial Activeness Prediction


Published in Multimedia Tools and Applications (SCI Journal)


AUTHORS: Dongjin Yu, Xinfeng Wang, Ping Liang, Xiaoxiao Sun

If there is any comment, please mail to yudj@hdu.edu.cn.

Experimental results

Source code:
data set:

Table 1. Public Yelp Dataset Statistics

City Year #Review #Check-in
Las Vegas 2013 148,371    962,717
2014 217,579 1,028,731
2015 284,881 1,042,245
2016 324,064    945,062
2017 359,982    839,641
Totonto 2013   43,773    113,289
2014   57,664    116,554
2015   77,859    137,310
2016   85,561    136,932
2017   88,338    130,724

Table 2. Hyper-parameter Setting

Key Value
optimizer Adam
lr 0.001
beta_1     0.9
beta_2     0.9
epsilon 1e-08
decay 1e-06

Table 3. Comparison of Different Methods

Method MSE MAPE
HA   53.79   22.98
ARIMA   61.76   21.92
XGBoost   28.68   22.15
Our Method   26.25   16.20
Our Method without periodic deminsion   27.50   17.47
Our Method without residual network 567.70 100.00

Fig 1. Comparison of DSTRN with different parameters

Fig 2. Visualization resluts of real and predicted activeness of Las Vegas and Toronto in 2018

Las Vegas

Las Vegas



Toronto

Toronto