Sequential prediction learning framework and algorithm

by Python awesome

DTCN

This is the implementation of our paper “Sequential Prediction of Social Media Popularity with Deep Temporal Context Networks“.

Dataset

To successfully test performance, we created TPIC Dataset, a temporal popularity image collection dataset.

Overview

Our DTCN contains three main components, from embedding, learning to predicting. With a joint embedding network, we obtain a unified deep representation of multi-modal user-post data in a common embedding space. Then, based on the embedded data sequence over time, temporal context learning attempts to recurrently learn two adaptive temporal contexts for sequential popularity. Finally, a novel temporal attention is designed to predict new popularity (the popularity of a new user-post pair) with temporal coherence across multiple time-scales.

Environment

The code is pure python. Keras is chosen to be the deep learning library here. Environment is configured by Anaconda. The environment file is saved as “environment.yml”.

  • Ubuntu 16.04
  • Python 2.7
  • Cuda 10.0
  • cudnn 7.6.5

Setup

conda env create -f environment.yml

Prequisition

  • Clone the repository to your local machine
  • Acquire relevant dataset
  • Extract the image feature with ResNet (2048 dims)
  • Run script by seeing example.

Usage

DATA_HOME=test_data/TRIM_DATA
KERAS_BACKEND=theano \
THEANO_FLAGS='mode=FAST_RUN,device=cuda0,nvcc.fastmath=True,optimizer=fast_run' \
python main.py \
-feature_path $DATA_HOME/USER_20W_SORTED_BY_TIME.txt \
-meta_path $DATA_HOME/ResNet_20W_2048_SORTED_BY_TIME.txt \
-label_path $DATA_HOME/LABEL_20W_SORTED_BY_TIME.txt \
-algorithm SHARED_DTCN \
-nb_epoch 1000 \
-start_cross_validation 2 \
-total_cross_validation 3 \
-identifier_path $DATA_HOME/USERID_20W_SORTED_BY_TIME.txt \
-timestamps_path $DATA_HOME/TIMESTAMP_20W_SORTED_BY_TIME.txt \
-visual_mlp_enabled y \
-timestep 10 \
-time_align y \
-time_dis_con continue \
-time_context_length 18 \
-time_unit_metric hour \
-discrete_time_start_offset 2 \
-discrete_time_unit 4 \
-train_set_partial 9 \
-merge_mode concat \
-dual_time_align n \
-time_weight_mode time_flag \
-dual_lstm n

Citation

@inproceedings{Wu2017DTCN,
  title={Sequential Prediction of Social Media Popularity with Deep Temporal Context Networks},
  author={Wu, Bo and Cheng, Wen-Huang and Zhang, Yongdong and Qiushi, Huang and Jintao, Li and Mei, Tao},
  booktitle={IJCAI},
  year={2017},
  location = {Melbourne, Australia}}

Please concat us (social.media.prediction@gmail.com) if you have further questions or cooporations

GitHub

View Github