| The rapid development of electronic science and technology enables to record details of human behaviors. With the availability of large-scale data, such as those from e-commerce to smart-phone communications, makes it possible to probe into and quantify the dynamics of human behaviours. Patterns of human behaviours have attracted increasing academic interests, since the quantitative understanding of human behavior is helpful to uncover the origins of many socioeconomic phenomena. Additionally, the study on human behaviours, which has significant applications in science and engineering, commerce, as well as defense, in terms of specific tasks such as recommendation and human-behavior prediction etc. For example, since the discovery of non-Poissonian statistics of human behaviors such as human interaction activities and mobility trajectories, more and more scientists have been paying attention to the role of these patterns in epidemic spreading. In this thesis, which contains(i) the empirical analysis of different types of human behaviours, from online to offline, from individual level to collect level.(ii) We introduce some protype models to mimic the observed pattern of human behaviours.(iii) To study the effects of the non-Poissonian statistics of human behaviours,we investigate how the scale-free characteristic of human contact activities influences epidemic spreading.(iv) In the end, we study how we utilize the human behaviours data to recommendation and human-behavior prediction.Firstly, we analyze many real-life huge datasets such as short message communication, movie-watching in Netflix and MovieLens, the transaction in Ebay, the bookmarkcollecting in Delicious, the posting in FreindFeed and Twitter, human interest trails and other online and offline of human trajectories, etc. Empirical analysis reveals some common statistical features of human behavior:(i) The total number of user′s actions, the user′s activity, and the interevent time all follow heavy-tailed distributions.(ii) There exists a significantly negative correlation between the user′s activity and the width of the interevent time distribution.(iii) The collective behaviors of rating media follow a process embodying self-similarity and long-range correlation. At individual level, we find that the dynamic behaviors of a few users have extremely small scaling exponents associating with long-range anticorrelations.(iv) We uncover fat-tailed(possibly power-law)distributions associated with the three basic quantities, such as the length of continuous interest, the return time of visiting certain interest, and the interest ranking and transition.(v) We found that there exists a power-law scaling between the frequency of visits to an academic forum and the number of corresponding visitors, with the exponent being about0.75.(vi) The expansion process of academic forums follows a power-law scaling with the exponent = 0.54, which is determined by the analysis of exploration of new academic forums.(vii) The statistical distributions of real time interval and the number of visits taken to revisit same academic forum both follow a power law, indicating the existence of memory effect in academic forum activities.(viii) We find a superlinear scaling relation between the mean frequency of visit, ? ?, and its fluctuation : ～ ? ?with ≈ 1.2, and so on.Secondly, on the basis of these empirical results, we develop some models to fit them, such as:(i) To distinguish the effects of endogenous mechanisms like the highestpriority-first protocol and exogenous factors like the varying global activity versus time,we propose a new timing method by using a relative clock and employ a model, in which agents act either in a constant rate or with a power-law inter-event time distribution, and the global activity either keeps unchanged or varies periodically versus time. Our analysis shows that the heavy tails caused by the heterogeneity of global activity can be eliminated by setting the relative clock, yet the heterogeneity due to real individual behaviors still exists.(ii) We develop a biased random-walk model, incorporating preferential return to previously visited interests, inertial effect, and exploration of new interests, to account for the observed fat-tailed distributions in interest dynamics.(iii) A dynamic model,incorporating the ingredients of exploration and preferential return with memory effect,is proposed to account for the observed scaling laws in human academic activities.(iv) To mimic the striking finding in human movements, we develop a model with two essential ingredients: preferential return and exploration, and show that these are necessary for generating the scaling extracted from real data.Thirdly, we focus on epidemic spreading in the hierarchical geographical networks with non-Poissonian statistics of human behaviors. On the one hand, we investigate how the scale-free characteristic of human contact activities influences epidemic spreading.We find that, compared with homogeneous contact pattern and null model, correlations between time delay and network hierarchy can remarkably slow down epidemic spreading, and result in a upward cascading multi-modal phenomenon. More importantly,high-layer seeds arouse large variabilities, while low-layer seeds result in several comparable peaks of variabilities, which makes the prediction of epidemic spreading hard.On the other hand, we employ a metapopulation modeling framework to investigate the susceptible-infected epidemic process evolving on hierarchical networks in which agents randomly walk along the edges and establish contacts in network nodes. Our results reveal that a shifted power-law-like negative relationship between the peak timing of epidemics 0and population density ??, and a logarithmic positive relationship between0and the network size , can both be explained by the gradual enlargement of fluctuations in the spreading process. Additionally, we provide a quantitative discussion of the efficiency of a border screening procedure in delaying epidemic outbreaks on hierarchical networks, yielding a rather limited feasibility of this mitigation strategy but also its non-trivial dependence on population density, infector detectability, and the diversity of the susceptible region. Our results suggest that the interplay between the human spatial dynamics, network topology, and demographic factors can have important consequences for the global spreading and control of infectious diseases.In the end, we deploy the human behaviours on recommendation and human-behavior prediction. Collaborative Filtering() algorithms are widely used in a lot of recommender systems, however, the computational complexity of is high which thus hinders their use in large scale systems. We implement user-based algorithm on a cloud computing platform, namely , to solve the scalability problem of . To predict human next actions(or trajectories), we comprehensively analyzed the features of individual’s mobility, which suggests that the human mobility patterns are heterogeneous and weakly weekly irregular. Then we presented various predictors for the mobile data and apply these predictors to make trajectory prediction, suggesting better prediction accuracy. Furthermore, by implementing a blending strategy, the best prediction accuracy is obtained because extensively considering the temporal-spatial features of human mobility. |
PDF Full Text Request