Key Technologies For Point-of-interest Recommendation On Location-based Social Networks

The increasing popularity of mobile devices and the rapid development of wireless communication technologies have accelerated the development and application of location-based social networks(LBSNs),such as Foursquare and Yelp.LBSNs connect cyberspace with the physical world,bringing the human into a digital age.In LBSNs,users can learn localization-related knowledge from the shared information from others,and find shops,restaurants and related services that are interesting to them at any time and place;users can also share their experiences via smart devices by posting their check-ins.The availability of large-scale interaction data in LBSNs provides an excellent opportunity for business and academia to gain insights into user mobility behavior and preference.The recommendation service provided by LBSNs can greatly enhance the user experience.Recently,a recommender system using machine learning and data mining technologies have rapidly developed in terms of both theory and application.However,due to the lack of deep mining and understanding about multimedia content information and user behavior under complex contexts,the quality and performance of the recommender system need to be improved regarding cold-start recommendation under the data sparsity environment,mining of sequence dependency relations under complex context,and activity recommendation.To solve these problems,we propose a number of new methods for mining user preference and develop several recommendation models by successfully mining multimodal content infornation,sequence dependency under complex contexts,and user activity preferences.The main work and contributions of this paper are as follows:(1)We propose a Point-of-Interest(POI)recommendation model by integrating multimodal content information.To resolve the cold-start issue and improve the robustness of POI recommendations under the data sparsity conditions,we develop a deep multimodal rank learning(DMRL)model to improve both the accuracy and robustness of POI recommendations.The model is a generative probabilistic model that combines multimodal content information of POIs with a Bayesian personalized ranking(BPR)learning framework.Firstly,we establish a time-dependent and geographical constrained user preference model to exploit the temporal dynamics and spatial dependence of user behaviour.Then a deep multimodal network,which links implicit feedback to semantic data,is constructed and integrated into BPR.This allows us to extract content and information of semantic representations that have different modalities in a task-oriented and supervised way.Finally,we utilize a ranking-based adaptive sampling strategy to accelerate convergence and improve model accuracy for model optimization.We then conducted experiments using two large-scale datasets obtained from Foursquare and Yelp;our DMRL model significantly outperforms other state-of-the-art models in terms of accuracy and MRR.For Accuracy@10,the cold-start POI recommendations are improved by 19.74%for the Foursquare dataset and 53.89%for the Yelp dataset.By parameterizing and varying the proportion of sparse data,the experiments showed that our DMRL model also gives better and more robust recommendations.(2)We propose a POI recommendation model by mining context awareness sequence dependency.After analyzing user's check-in sequence dependency and the interaction between different types of input contexts,we then proposed a context awareness deep neural network model,where we integrate sequence context,input contexts and user preferences into a cohesive framework.Firstly,we model sequence context and interaction of different types of input contexts by extending the recurrent neural network to capture the user's dynamic preference from check-in records.After that,we design a feedforward neural network to capture user's general preferences from check-in data and incorporate that into MCI-DNN.Finally,the user's dynamic preference and general preference are used to predict the check-in probability for target POIs.To deal with different types of input contexts in the form of multi-field categorical,we adopt embedding representation technology to automatically learn dense feature representations of input contexts,and the interaction of location and different kinds of input contexts were learned in RNN unit.Experiments on Foursquare and Gowalla datasets respectively show that the proposed model outperforms the current state-of-the-art approaches by about 57.12%and 76.4%regarding F1-score@5.(3)We propose a POI recommendation model by mining geographical-temporal awareness sequence dependency relationship.To perform mining and enhance the understanding of the global sequence dependency of user mobility as well as capture the subtle POI-POI connections,and distinguish the relevant check-ins from the irrelevant,we further propose a geographically-temporally awareness hierarchical attention network(GT-HAN).GT-HAN consists of two attention networks.The first one is a geographical-temporal attention network that models the global temporal dependency of a check-in sequence and the geographical relations between POIs.The second one is a co-attention network that is used to capture the changing user preferences(dynamic user preferences),giving the general preference of the target user and the context at the current state by adaptively assigning different weights to each check-in behavior.Our tests on Foursquare and Gowalla datasets demonstrated the significant improvement given by GT-HAN compared to the state-of-the-art approaches.For Accuracy@10,the proposed model outperforms the second-best model by more than 46.3%for the Foursquare dataset and 37.8%for the Gowalla dataset.(4)We propose an activity recommendation based on user's check-in data.To solve the problem of data sparsity for activity recommendation in LBSNs,we describe a collaborative tensor-topic factorization(CTTF)model by mining tips posted by users and semantic content information about locations.CTTF model incorporates user interest topics and activity topics into a tensor factorization framework to improve activity recommendations for users.We represent user activity feedback with a third-order tensor and penalize false preferences inferred from check-ins using term frequency-inverse document frequency(TF-IDF).To solve the problem of data sparsity,a biterm topic model(BTM)that performs better than latent Dirichlet allocation(LDA)on the short text was used to extract user interest topics from tips and to extract activity topics from semantic location data.The interest and activity topics are then incorporated into the tensor factorization framework to infer the latent relations between users,activities and time.Experiments on the real-world datasets show that the CTTF model significantly outperforms current state-of-the-art approaches.