魏海涛, 李柯, 赫晓慧, 田智慧. 融入空间关系的矩阵分解POI推荐模型[J]. 武汉大学学报 ( 信息科学版), 2021, 46(5): 681-690. DOI: 10.13203/j.whugis20200355
引用本文: 魏海涛, 李柯, 赫晓慧, 田智慧. 融入空间关系的矩阵分解POI推荐模型[J]. 武汉大学学报 ( 信息科学版), 2021, 46(5): 681-690. DOI: 10.13203/j.whugis20200355
WEI Haitao, LI Ke, HE Xiaohui, TIAN Zhihui. Integrating Spatial Relationship into a Matrix Factorization Model for POI Recommendation[J]. Geomatics and Information Science of Wuhan University, 2021, 46(5): 681-690. DOI: 10.13203/j.whugis20200355
Citation: WEI Haitao, LI Ke, HE Xiaohui, TIAN Zhihui. Integrating Spatial Relationship into a Matrix Factorization Model for POI Recommendation[J]. Geomatics and Information Science of Wuhan University, 2021, 46(5): 681-690. DOI: 10.13203/j.whugis20200355

融入空间关系的矩阵分解POI推荐模型

Integrating Spatial Relationship into a Matrix Factorization Model for POI Recommendation

  • 摘要: 兴趣点(point of interest, POI)推荐是在基于位置的社交网络中流行起来的个性化服务。针对数据稀疏和隐性反馈的使用等问题,提出了一种关系型矩阵分解模型——合作竞争矩阵分解(cooperative competition matrix factorization, CC‑MF)。该模型根据用户与POI间的相互关系建模,融入空间关系,并将空间关系细分为空间距离关系和空间拓扑关系,挖掘POI之间、POI与用户之间的空间关系, 以缓解数据稀疏问题;同时使用加权最小二乘准则构建目标函数,缓解隐性反馈问题。在现实世界签到Foursquare数据集上进行实验,结果显示: (1)CC‑MF模型显著提高了推荐结果的准确性;(2)考虑空间拓扑关系的空间距离因素能够进一步提升推荐系统的性能。因此,CC‑MF模型具有良好的拓展性和解释性,且缓解了数据稀疏和隐性反馈使用问题。

     

    Abstract:
      Objectives  Point of interest (POI) recommendation is the prevalent personal service in location‑based social network(LBSN), and aims to provide personalized recommendation services by using the information carried by LBSN. The utilization of spatial relationship information as the side information supplies a chance to product better POI recommend. However, thousands of users and POIs in the LBSN make the user‑POI check‑in matrix very large and sparse.In addition, check‑in record data is typical implicit feedback data, which cannot directly reflect the user?s preference. To tackle the aforementioned challenges, we propose a relational matrix factorization model based on cooperative competition matrix factorization (CC‑MF).
      Methods  The CC‑MF model can simulate the relationship between users and POIs, and divides spatial relationships into spatial distance relationship and spatial topological relationship. In order to alleviate the problem of data sparsity, the model excavates the spatial relationships among POIs, POIs and users by integrating spatial relationships. Firstly, we use nonlinear function to establish the spatial distance relationship between users and POIs, which can connect the relationship between users and POIs. Then, k‑nearest neighbor (kNN) algorithm is used to calculate the geo‑neighbors of POI by considering the spatial distance factor of spatial topological relationship, which can further alleviate the sparsity of data. Finally, the spatial relationship is integrated into the matrix factorization model. Meanwhile, the weighted least square method is used as the objective function of the CC‑MF model to relieve the implicit feedback problem. Experiments are carried out on the real‑world check‑in Foursquare datasets. We test the recommendation performance of the proposed model and baseline methods, and analyze the crucial influence of different spatial relationships on POI recommendation. The precision and recall are used as evaluation metrics.
      Results  The results show that: (1) The CC‑MF model significantly improves the precision and recall of the recommendation results. (2) Considering the spatial distance factor of the spatial topological relationship can further improve the performance of the recommendation system.
      Conclusions  Therefore, CC‑MF model can make use of spatial relationship better and more comprehensive.The proposed CC‑MF model has better scalability and better interpretability, and can alleviate the problems of data sparsity and implicit feedback usage.

     

/

返回文章
返回