陳履謙

Lyu-CianChen

碩士論文（2017）

基於適地性服務之地點與行動分析方法研究：以零售點推薦為例

On Method of Location and Mobility Analytics with Location-based Service: A Case Study of Retail Store Recommendation

關鍵字 Keywords

城市探勘、時間與空間探勘、適地性服務、零售點推薦、機器學習、特徵萃取

Urban mining、Spatial and temporal data mining、Location-based Service、Retail store recommendation

摘要

近年來，適地性服務(Location-based Service)搭上行動網路普及的風潮，於企業而言已成為一行銷的重要利器。於零售業者，零售店面位置之選擇與適當的行銷手法，能大幅度的增加產品曝光度及獲利。但零售店面位置選擇之議題，在分析資料之蒐集上著實困難，故本研究利用適地性社群網路資料，透過分析使用者足跡與興趣點之關聯性，發展一個零售點推薦之方法。

透過目標區域關聯性分析與人群移動模式萃取，本研究建立零售點推薦之特徵值(Features)。本研究設計興趣點間之產業密度(Density)、區域類別(Category)與區域/產業分群(Clustering)，以及區域飽和度的計算，以核密度估計(Kernel Density Estimation, KDE)、k-平均演算法(k-Means)等方法，計算區域關聯性對零售點選擇之影響。另外，本研究建立人群移動模式擷取方法，以循序模式探勘(Sequential Pattern Mining)方法選取最常行動之路徑，以量化人群行動對候選零售點之影響力。本研究分別對興趣點分析與行動路徑產生多個特徵值，並採以排名品質指標(Nomalized Discounted Cumulative Gain, NDCG)衡量多個特徵值之有效性。最後，本研究使用Foursquare與Facebook的資料集進行零售點推薦之實驗，以CART決策樹(Classification and Regression Tree)演算法進行零售點推薦。在本研究所萃取的特徵值中部分表現相當優異，且建立了以適地性服務進行零售點推薦的方法，並實際以台南地區的店面進行推薦。

Abstract

In recent years, with the popularization of mobile network, the location-based service (LBS) has made great strides, becoming an efficient marketing instrument for enterprises. For the retail business, good selections of store and appropriate marketing techniques are critical to increasing the profit. However, it is difficult to select the retail store because there are numerous considerations and the analysis was short of metadata in the past. Therefore, this study uses LBS, and provides a recommendation method for retail store selection by analyzing the relationship between the user track and point-of-interest (POI).

This study uses regional relevance and human mobility analysis to establish the feature values of retail store recommendation. This study proposes (1) architecture of the data model available for retail store recommendation by influential layers of LBS, and System-based solution for recommendation of retail stores; (2) Industry density, area categories and region/industry clustering methods of POIs. Uses KDE and KMeans to calculate the effect of regional functionality on the retail store selection; (3) Provide a mobility (sequence) feature extraction method to measure mobility track. Finally, give an evaluation and implementation for the result.