Influence Maximization For Location Promotion In Geo-social Networks

In recent years,the rapid development in Internet-of-Things(Io T)technology and the rise of online social media have facilitated the prosperity of location-based social networks(LBSNs).The location-based influence maximization problem,which aims to find top influential seed users from an LBSN such that a target location can be promoted to attract the most individuals,has drawn increasing attention.It finds application in many significant Io T scenarios such as viral marketing,advertisement deployment,etc.However,most existing works only applied a single location to distinguish users and ignored the valuable information embedded in rich geo-spatial data like user check-in records.Besides,LBSNs in the real world are often large-scale;how to efficiently solve the influence maximization problem in large-scale networks is also a considerable challenge.Based on the above problems,this thesis research the following two aspects:First,this thesis proposes a representation learning model,All2 Vec,for extracting user preferences.All2 Vec is an improved model based on the POI2 Vec model.Since the POI2 Vec only considers the geographical relationship between the user and the current check-in location,the user preference captured by the model cannot reflect the real preference of users due to its randomness.Instead,the All2 Vec model models both the user and the check-in location jointly,considering the interactions between the user and each check-in location.All2 Vec considers both the sequential relationship between the check-in locations and the influence of geographical space.Finally,All2 Vec computes the user preference weight by integrating distance and user behavior preference.Experimental results show that the user preference obtained by the All2 Vec model is significantly more accurate than the POI2 Vec model in user behavior prediction.Second,this thesis combines representation learning with influence maximization and defines the Location Promotion Influence Maximization(LPIM)problem.Through theoretical analysis of the problem,it is proved that the problem has theoretical properties such as submodularity and monotonicity,and an algorithm named IM2 Vec,which is based on Reverse Influence Sampling(RIS)and can return a(1-1 ? 0)-)-approximate solution,is proposed.In addition,this thesis determines the asymptotic lower bound of OPT through the user preference weights,which dramatically reduces the sampling size and further improves the algorithm efficiency compared with the original algorithm.The simulation results show that IM2 Vec is significantly better than other comparison algorithms in terms of effectiveness and efficiency.For solution quality evaluation,IM2 Vec has achieved the influence spread values in all experiments.In terms of efficiency,the running time of the IM2 Vec algorithm is much less than other algorithms,and with the increase of the number of seeds in the seed set,the efficiency advantage of IM2 Vec is more prominent.In summary,the main contributions of this thesis are outlined as follows: for the LPIM problem,this thesis proposes a general framework that combines the problem with the latest representation learning.The framework can better capture user behavior preference through joint modeling of user behavior,location sequence,and distance.In addition,by improving the RIS algorithm,the sampling size is further reduced,and the algorithm’s efficiency is improved.The work of this thesis improves the practicability of the traditional influence maximization method and has good theoretical and application values.