EXPLORING TRANSFER LEARNING IN URBAN COMPUTING PERSPECTIVE
Traditional machine learning algorithms are mainly characterized by the training a model with the basic assumption that the train data and the test data are drown from the same distribution. When we are facing a distribution shift between the train data and the test data the performance of the predictive model degrade considerably. In some scenarios it become expensive the collect valuable training data that match the feature space and predicted data distribution. Transfer leaning, a new leaning paradigm come to rescue by have the ability to improve the predictive power of a learner from a particular domain A using the wealth of transferable knowledge from a related or similar domain B. There has been a multiple machine learning application that have seen enormous success with transfer learning such as sentiment analysis, multi-language classification, machine translation, human activity classification.
In this thesis we study the use of transfer learning to solve urban computing issue in cities where data are scarce. We focus on particular urban computing application which is based on predictions. To solve the data scarcity we had to tackle multiple challenging including: transferring from multiple views or sources by learning the their intrinsic dependency; building strong learning model that can predict service of interest one target cities using service from one or multiple source cities by overcoming the divergence in term of distribution between; Building model that can intelligently leant both spatial and temporal dependency between the target and sources; tackling the missing and corrupted data issue; finding a mechanism which allows to find the right source to transfer from; building a mechanism that can detect useful information and penalize information that can mislead the transfer leaning process. Our main contribution alongside this thesis can be describe as follows:
1. A Multi-view diction learning algorithm is proposed based on low rank tensor which is used to learn the intrinsic relationship existing between views.
2. A novel unsupervised multi-view transfer learning based on low rank constrained called IMUTransfer is proposed whose main idea is to find a domain invariant subspace where view from the source domain is coupled with those from the target domain to combat noisy, corrupted and missing observation.
3. An unsupervised multi-source transfer learning is also proposed based on deep auto-encoder coupled with multi-view dictionary learning called MVT-DAE.
4. We develop a mechanism that can select the best subset of source in which we can conduct transfer from in a multi-source transfer setting.
5. Multi-source deep spatial temporal transfer learning based on a ConvLSTM structure called DeepMuTrans-ST. This approach main to conduct transfer learning with spatial temporal dependency which is pertinent in many urban computing applications.
 Abdoullahi Diasse, Zhiyong Li, Multi view deep unsupervised transfer learning via joint auto encoder coupled with dictionary transfer. Intelligent Data Analysis, vol. 23, no. 3, pp. 555-571, 2019(SCI)
 Abdoullahi Diasse, Zhiyong Li, Big Cities transfer learning: An unsupervised multi-view cross-domain classification with misses，ACM ICMLC 2018 Proceedings of the 2018 10th International Conference on Machine Learning and Computing Pages 312-321 (EI)
 Abdoullahi Diasse, Zhiyong Li, Multi source spatial temporal deep transfer learning for crowd flow prediction: Under review at Knowledge based system Abdoullahi Diasse, Zhiyong Li,(Under review at KBS , SCI)
版权所有©湖南大学2017 湖南大学党委宣传部 地址：湖南省长沙市岳麓区麓山南路麓山门 邮编：410082 Email：firstname.lastname@example.org 域名备案信息：[www.hnu.edu.cn,www.hnu.cn/湘ICP备05000239号] [hnu.cn 湘教QS3-200503-000481 hnu.edu.cn 湘教QS4-201312-010059]