Learning Spatial-Temporal User Behaviors with Symbolic Graphs

  • 鄧 善云

Student thesis: Doctoral Thesis


Recently interesting user behaviors have been found in different fields Therefore many research focuses on understanding user behaviors needs and motivations through observation techniques data analysis and machine learning methodologies By learning different user behaviors the prediction of user behaviors can lead to applications for indispensable services such as trajectory patterns mining social links prediction and personalized recommender systems Due to the high penetration rate of cost-effective mobile devices it becomes possible to continuously collect user behavior records with temporal and spatial information As such many spatial-temporal learning methodologies have been proposed to figure out the hidden information of user behaviors with the spatial or temporal sense In my dissertation we attempt to learn spatial-temporal user behaviors which are left unexplored thus far in different scenarios such as indoor spaces social networks and E-learning systems More specifically we investigated the following three kinds of user behaviors spatial-temporal stop-by behaviors in indoor spaces temporal worship behaviors in celebrity-dived networks and temporal rating behaviors in E-learning systems Learning Spatial-Temporal Stop-by Behaviors with Uncertain Symbolic Graphs in Indoor Spaces: In this work we explore a new mining paradigm called Indoor Stop-by Patterns (ISP) to discover user stop-by behavior in mall-like indoor environments The discovery of ISPs enables new marketing collaborations such as a joint coupon promotion among stores in indoor spaces (e g shopping malls) Moreover it can also help in eliminating the overcrowding situation To pursue better practicability we consider the cost-effective wireless sensor-based environment and conduct the analysis of indoor stop-by behaviors on real data The proposed Probabilistic Top-k Indoor Stop-by Patterns Discovery (PTkISP) framework incorporates the probabilistic model to identify top-k ISPs over uncertain data collected from sensing logs Moreover we develop an uncertain symbolic model and devise an Index 1-itemset (IIS) algorithm to enhance the accuracy and efficiency Our experimental studies on one synthetic data and one real data show that the proposed PTkISP framework can efficiently discover high-quality ISPs and can provide insightful observations for marketing collaborations Learning Temporal Worship Behaviors with Heterogeneous Symbolic Graph in Celebrity-Dived Networks: We in this work explore a new link prediction paradigm called ‘worship’ prediction to discover worship links between users and celebrities on celebrity-dived networks The prediction of ‘worship’ links enables valuable social services such as viral marketing popularity estimation and celebrity recommendation However as the concern of business security and personal privacy only public-accessible statistical social properties instead of the detailed information of users can be utilized to predict the ‘worship’ labels To address these issues a novel learning framework is devised including a heterogeneous factor graph with new discovered statistical properties and a Gaussian estimation based learning algorithm with active learning Our experimental studies on real data including Instagram Twitter and DBLP show that the proposed learning framework can overcome the problem of missing labels and efficiently discover worship links Learning Temporal Rating Behaviors with Dependent Symbolic Graphs in E-Learning Systems: In this work we address an important issue on the exploration of user rating behaviors from an interactive question-answering process in E-learning systems A novel interactive learning system called CagMab is devised to interactively recommend questions with a round-by-round strategy which contributes to applications such as a conversational bot for self-evaluation The flow enables users to discover their weakness and further helps them to progress Even though formulating the problem with the multi-armed bandit framework provides a solution it often leads to suboptimal results for interactive unknowns recommendation as it simply relies on the contextual features of answered questions To address this issue we develop a novel interactive learning framework by borrowing strengths from the dependency of concept-aware graphs for learning user ratings Our experimental studies on real data show that the proposed framework can effectively predict user ratings in an interactive fashion for the recommendation in E-learning systems In summary we in my dissertation focus on learning spatial-temporal user behaviors with symbolic graphs in indoor spaces social networks and E-learning systems The experimental results of learning the proposed user behaviors show that incorporating symbolic graphs with machine learning algorithms can significantly improve the accuracy performances
Date of Award2019
Original languageEnglish
SupervisorKun-Ta Chuang (Supervisor)

Cite this