Customer revisit prediction using in-store sensor data = 매장 내 센서 데이터를 활용한 고객 재방문 예측

Cited 0 time in webofscience Cited 0 time in scopus
  • Hit : 82
  • Download : 0
With the advancement of sensor technology, offline data collection has become possible, and many retail analytics companies are beginning to offer solutions that provide data collection and analysis. Thereby, store managers can grasp the status of their stores, thus trying to satisfy the customers' experience. Many of these efforts are carried out in order to secure regular customers for continuous store management and profit generation. To get closer to customers, companies strive to understand customers' interest and profile. Furthermore, they make an effort to predict customers' potential lifetime values, purchasing patterns, revisits, and stickiness. Among these objectives, customer revisit is a feasible and valuable metric to study since it can be recognized by only using customer foot-traffic data. This is very important to note since purchase data and user profiles are considered as proprietary information and difficult to obtain outside the company, but customer mobility becomes relatively easy to obtain through location monitoring technology once we get the customers' permission through their mobile device. By knowing customers' visitation pattern, store managers can indirectly gauge the expected revenue. Targeted marketing can also be available by knowing customers' revisit intention. By offering discount coupons, merchants can encourage customers to accidentally revisit a store nearby. Also, they can offer a sister brand with finer products to provide new shopping experiences to loyal customers. In this way, they can increase the revenue as well as satisfy their customers. My thesis focuses on these closely related questions---revisit prediction---to capture the potential regular customers of the store. To achieve the goal, we formally design predictive analytics and develop two frameworks using mobility data captured from in-store sensors. In the first part, we introduce a traditional machine learning model with carefully designed handcrafted features. We design extensive handcrafted features using semantic areas of the stores, and we investigate the predictive powers of feature groups and semantic levels of areas. We confirm the effectiveness of considering customer mobility by showing the performance improvement of 4.7-24.3%. Furthermore, we provide an in-depth analysis regarding the effect of the data collection period as well as missing customers. Throughout this chapter, we look forward to sharing a series of processes to solve the predictive analytics problem by finding the right features. In the second part, we introduce a survival analysis model powered by a deep architecture. We propose this model to challenge more realistic prediction settings having partial observations with the imbalanced distribution. Unlike the framework in the first part, our new SurvRev model can predict the event rate of the next 365 days for each visit. We are able to handle partial observations by survival analysis, and the underlying deep learning architecture effectively learns the hidden representation of customers and their visits. By optimizing a custom loss function, our SurvRev model can be tuned for various prediction purposes. Throughout this chapter, we introduce our various efforts to refine the model and verify its superiority over other revisit prediction models. We successfully apply our models to mobility datasets collected from seven flagship stores in downtown Seoul, including more than 5.7 million visits over 2.5 years. For fertilizing research, we also release a benchmark dataset of customer indoor movement patterns. We hope that our research and datasets can be used for offspring studies that require understandings of customers' shopping patterns.
Lee, Jae-Gilresearcher이재길researcher
한국과학기술원 :지식서비스공학대학원,
Issue Date

학위논문(박사) - 한국과학기술원 : 지식서비스공학대학원, 2019.8,[v, 98 p. :]


Revisit prediction▼ahuman behavior prediction▼apredictive analytics▼aretail analytics▼ain-store motion pattern analysis▼auser modeling▼asensor data▼amobility data▼afeature engineering▼adata mining▼amachine learning▼adeep learning▼asurvival analysis▼alongtitudinal data; 재방문 예측▼a고객 행동 예측▼a예측 분석▼a소매 분석▼a매장 내에서 수집된 발자취 분석▼a유저 모델링▼a특성 추출▼a데이터 마이닝▼a기계 학습▼a딥 러닝▼a생존 분석▼a종적 데이터

Appears in Collection
Files in This Item
There are no files associated with this item.


  • mendeley


rss_1.0 rss_2.0 atom_1.0