Alibaba’s 11 IJCAI 2017 Papers: Cutting‑Edge AI Research Highlights

IJCAI 2017 Overview

IJCAI is one of the most prestigious AI conferences, covering machine learning, sustainable computing, image and speech recognition, video technology, and more. In 2017 it received 2,540 submissions and accepted 660 papers (26% acceptance rate).

Alibaba’s Accepted Papers

Alibaba had 11 papers accepted at IJCAI 2017, six from the Alibaba‑Zhejiang University Frontier Technology Joint Research Center, three from Ant Financial, and two from Tmall and Cainiao Logistics (Workshop on AI Applications in E‑Commerce).

Image Gradient‑based Joint Direct Visual Odometry for Stereo Camera

Authors: Zhu Jiank​e

Proposes a novel stereo visual odometry method using multi‑scale pyramid and gradient‑based image representation, improving convergence and robustness to illumination changes.

Link: https://www.ijcai.org/proceedings/2017/0636.pdf

DeepFacade: A Deep Learning Approach to Facade Parsing

Authors: Liu Hantang, Zhang Jialiang, Zhu Jiank​e, Xu Zhuhong

Introduces a symmetry‑constrained deep neural network for building façade parsing, achieving state‑of‑the‑art results on ECP and eTREIMS datasets.

Link: https://www.ijcai.org/proceedings/2017/0320.pdf

CFNN: Correlation Filter Neural Network for Visual Object Tracking

Authors: Li Yang, Xu Zhan, Zhu Jiank​e

Presents a correlation‑filter neural network that requires no pre‑training and leverages cyclic sampling for effective online tracking.

Link: https://www.ijcai.org/proceedings/2017/0309.pdf

Deep Optical Flow Estimation Via Multi‑Scale Correspondence Structure Learning

Authors: Zhao Shanshan, Li Xi, Ao Ma

Proposes the MSCSL framework that learns multi‑scale correspondence structures via a spatial‑ConvGRU, enabling end‑to‑end optical flow estimation.

Link: https://www.ijcai.org/proceedings/2017/0488.pdf

Group‑wise Deep Co‑saliency Detection

Authors: Wei Lina, Zhao Shanshan, Ao Ma, Li Xi, Wu Fei

Designs an end‑to‑end fully convolutional network that jointly learns group‑level and single‑image features for robust co‑saliency detection.

Link: https://www.ijcai.org/proceedings/2017/0424.pdf

Boosted Zero‑Shot Learning with Semantic Correlation Regularization

Authors: Pi Te, Li Xi, Zhang Zhongfei

Introduces a semantic‑correlation regularization (SCR) embedded in boosting to improve zero‑shot classification by aligning model predictions with semantic structures.

Link: https://www.ijcai.org/proceedings/2017/0362.pdf

Local Linear Factorization Machines

Authors: Chenghao Liu, Teng Zhang, Peilin Zhao, Jun Zhou, Jianling Sun

Proposes LLFM, which integrates local encoding with factorization machines via joint optimization of anchors, local codes, and FM parameters, achieving superior prediction accuracy.

Link: https://www.ijcai.org/proceedings/2017/0319.pdf

Learning User Dependencies for Recommendation

Authors: Yong Liu, Peilin Zhao, Xin Liu, Min Wu, Lixin Duan, Xiaoli Li

Presents Probabilistic Relational Matrix Factorization (PRMF) that automatically learns user‑user dependencies using a matrix‑variate normal model, improving recommendation performance on four real‑world datasets.

Link: http://static.ijcai.org/proceedings-2017/0331.pdf

Online Multitask Relative Similarity Learning

Authors: Shuji Hao, Peilin Zhao, Yong Liu, Steven C. H. Hoi, Chunyan Miao

Introduces OMTRSL, an online framework for multitask relative similarity learning with an active learning component, offering strong theoretical guarantees and practical efficiency.

Link: https://www.ijcai.org/proceedings/2017/0253.pdf

Workshop Papers (AI Applications in E‑Commerce)

Solving a New 3D Bin Packing Problem with Deep Reinforcement Learning Method

Authors: Hu Haoyuan, Zhang Xiaodong, Wang Longfei, Xu Yinghui

Applies a Pointer Network‑based deep reinforcement learning approach to 3D bin packing, achieving ~5% improvement over heuristic methods.

Life‑stage Inference in E‑Commerce: A Dynamic Merging Based Approach

Authors: Zhou Zhongsheng, Zhang Yidong, Shu Zhichao, Deng Yuming, Wang Xiaoqing

Proposes a dynamic fusion method to infer user life‑stage for personalized recommendations, maintaining a compact set of probability distributions for robust prediction.