Abstract: The key defects of existing user mobility discovery methods, such as PMM (periodic mobility model), W3 etc., are the lack of the capability to objectively and quantitatively reflect the effect of time factor on the users location change and geographical relationship, which makes it difficult to accurately predict the user's future check-in location. To solve this problem, this paper proposes a location transfer evolution algorithm based on vector auto-regressive (LTE), which takes in the large volume real user check-in location history as training datasets and learns the regularity of users location transfer over the change of time. The extensive experiments on real check-in datasets of Foursquare and Gowalla are conducted, which shows that compared with W3, LTE (the proposed algorithm)     improves the prediction accuracy by 4.43%—21.31%, and compared with PMM, LTE improves the prediction accuracy by 25.07%—38.50%.

Key words: location prediction, cluster marker transfer matrix, user mobility, vector autoregressive, cluster marker transfer vector sequence

摘要： 现存用户移动性规律发现方法，如PMM（periodic mobility model）、W3等的核心缺陷在于，不能将时间对于用户访问位置变化与地理位置之间关联制约的两种重要影响真实、量化地反映出来，因而无法准确预测用户未来的签到位置。针对该问题，提出基于向量自回归的位置转移演化算法（location transfer evolution algorithm based on vector autoregressive，LTE），基于用户签到位置的变化历史，学习得到用户位置转移随时间、空间变化的规律性，且基于此进行用户位置的准确预测。基于Foursquare和Gowalla真实签到数据集，进行大量、深入的实验分析，实验结果表明，相对于W3，LTE算法的准确率提升4.43%~21.31%，相对于PMM，LTE算法的准确率提升25.07%~38.50%。

关键词: 位置预测, 簇标记转移矩阵, 用户移动性, 向量自回归, 簇标记转移向量序列