Posted by Admin: System Admin
Due to the great advances in mobility techniques, an increasing number of point-of-interest (POI)-related services have emerged, which could help users to navigate or predict POIs that may be interesting. Obviously, predicting POIs is a challenging task, mainly because of the complicated sequential transition regularities, and the heterogeneity and sparsity of the collected trajectory data. Most prior studies on successive POI recommendation mainly focused on modeling the correlation among POIs based on users’ check-in data. However, given a user’s checkin sequence, generally, the relationship between two consecutive POIs is usually both time and distance subtle. In this article, we propose a novel POI recommendation system to capture and learn the complicated sequential transitions by incorporating time and distance irregularity. In addition, we propose a feasible way to dynamically weight the decay values into the model learning process. The learned awareness weights offer an easy-to-interpret way to translate how much each context is emphasized in the prediction process. The performance evaluations are conducted on real mobility datasets to demonstrate the effectiveness and practicability of the POI recommendations. The experimental results show that the proposed methods significantly outperform the state-of-the-art models in all metrics. Machine learning is an important component of the growing field of data science. Through the use of statistical methods, different type of algorithms is trained to make classifications or predictions, and to uncover key insights in this project. These insights subsequently drive decision making within applications and businesses, ideally impacting key growth metrics. Machine learning algorithms build a model based on this project data, known as training data, in order to make predictions or decisions without being explicitly programmed to do so. Machine learning algorithms are used in a wide variety of datasets, where it is difficult or unfeasible to develop conventional algorithms to perform the needed tasks.
Cheng et al. [9] considered two main properties in the check-in sequence: 1) a personalized Markov chain and 2) region localization. They then proposed a novel matrix factorization method for POI recommendation that incorporates users’ movement constraint and the personalized Markov chain. He et al. [18] modeled successive check-in behaviors by proposing a third rank tensor that blends a personalized Markov chain with a latent pattern. LORE [45] utilizes sequential patterns to predict the probability of a user visiting a location by exploiting a dynamic location transition graph via an additive Markov chain. Note that sequential influence with geographical influence and social influence could also be fused into the recommendation framework. The authors also propose a gravity model that integrates the spatiotemporal, social, and popularity influences by estimating a power of law distribution. Matrix factorization [30], [38], [42] is a method to discover sequential mobility signals from the trajectory and then aims to forecast mobility based on these popular patterns. The basic idea is to factor the user-item matrix into two latent matrices which represent the user and item characteristics. Xiong et al. [39] produced latent vectors of users, items, and time bins via factorization for POI recommendation. Cheng et al. [8] further employed geographic influence and social influence in a matrix factorization method for performance improvement. Cheng et al. [9] solved the problem of successive POI by the factoring personalized Markov chain (FPMC) and included the geographic influence for predicting the POIs that the user may be interested in [32]. Several latent probabilistic generative models [14], [23], [25], [31], [35], [36], [40], [41] could effectively mimic users’ decision-making process by utilizing the multiplication of latent variables. Fu et al. [14] developed a generative probabilistic model for restaurant recommendation using multiple information fusion called GC-BCoNMF. This article exploits the latent features of each aspect in multirating, such as cuisine, environment, and cost to predict ratings. The WWO [23] probabilistic graphical model integrates the user interests and the evolving sequential preference with temporal interval assessment. Liu et al. [25] proposed a general geographical probabilistic factor model, Geo-PFM, to capture the geographical influences. Since the probability of a user visiting a POI is inversely proportional to the geographic distance between them, Geo-PFM could identify POIs in the same region that shares similar features. Qian et al. [31] proposed a spatiotemporal context-aware and translation-based recommendation framework (STA) to model the third-order relationship users, POIs, and spatiotemporal contexts for large-scale POI recommendation. TRM [35] utilizes the latent topic-region variable to effectively fuse the sequential influence and cyclic pattern with personal interest to the latent space. Wang et al. [36] introduced a novel probabilistic-generative model, SPORE, to capture the user decision-making process for choosing spatial items based on the latent variable topic region. Yin et al. [40] proposed a spatial-aware user preference modeling method based on latent factors to include the influence of spatial dynamics of personal preferences. The authors leverage the spatial-aware crowd’s preferences by exploiting the geographic autocorrelation of personal preferences with geographically hierarchical additive representations. Yin et al. [41] extended TRM to concurrently determine the semantic, temporal, and spatial patterns of check-in activities. The authors also developed a novel clustering-based branch and bound algorithm (CBB) to trim the POI search space. Disadvantages ? An existing system,DeNavi model doesn't contrast significantly with the previous literature in several aspects. Most prior works exploit a latent probabilistic generative model which utilizes the latent variables to calculate the prediction results. In order to cover many aspects, more information is necessary for the latent variables, such as user, POI, region, topic region, general public preference, and rating, to name a few. ? An existing methodology doesn’t implement Prediction Module of DeNavi method.
The system proposes a novel POI recommendation system, deep navigator (DeNavi), to handle time and distance irregular intervals in longitudinal check-in trajectories. We incorporate the irregularities of time and distance between the consecutive elements in a sequence into the memory unit to boost the performance improvement of the standard recurrent networks. Intuitively, the memory cell should decay in a way that the greater the elapsed time and distance, the smaller the effect of the previous memory on the current output. Based on this assumption, the elapsed time and distance are formulated into a proper weight exploitation. Also, the prediction is not only based on the users’ short-term interests but also on their long-term interests. There are three learning models in DeNavi: 1) DeNavi-LSTM considers the time and distance awareness based on LSTM and utilizes a subspace decomposition of the memory cell and hidden state by a decay function to discount the historical information according to interval differences; 2) DeNavi-GRU, likewise, also considers the time and distance awareness and utilizes lightweight learning for the hidden state by decay function; and 3) DeNavi-Alpha dynamically includes the time and distance awareness based on LSTM with SVM and Logistic Regression and utilizes the dynamic exponential weight moving average (EWMA) to adjust the importance of each context differently. Advantages 1) A novel learning model, DeNavi, is proposed to incorporate time and distance irregularities from check-in sequences. The time and distance differences are considered and included in the learning process for model training. 2) By carrying a decomposition of the memory cell out into short- and long-term effects, in the proposed model, the short-term counterpart is modified by the discounted weight and is then combined with the long-term counterpart before entering the sequential learning model. 3) We propose a feasible way to dynamically weight the decay values by integrating the EWMA into the model learning process. The learned awareness weights offer a convenient way to translate how much each context is emphasized in the prediction process. 4) We perform extensive experiments on two real-life mobility datasets. Our results demonstrate that DeNavi outperforms state-of-the-art mobility prediction models. It exhibits an outstanding generalization ability and is robust across trajectory datasets of different natures.