基於同儕的深度學習：超商會員將會購買什麼？

Peer-Based Deep Learning: What Retail Customers Will Buy?

游達(Yu Ta)

指導教授 ： 莊皓鈞 ； 周彥君

國立政治大學/商學院/資訊管理學系/碩士(2019年)

https://doi.org/10.6814/NCCU201900418

摘要

本研究針對商品間隱含相依關係的線下零售業，提出一種基於同儕的機器學習模型―商品迭代式的深度學習模型，作為推薦系統的演算法。此法可以針對不同的目標商品改變輸出及輸入的資料，每一次迭代都可以篩選出對於目標商品最活躍的客群，著重學習目標商品與其他商品的交互購買模式，且迭代過程中也可以對某些資料較少的商品進行過取樣，以增加較少被購買商品的訓練次數，此外，本研究僅需訓練一個模型便能針對各個商品進行推薦。本研究以臺灣分布最廣的零售業―超商為例，實際測試此模型捕捉會員購物喜好的表現，結果顯示模型未見過的特定會員購買過的特定商品，此模型購買量預測可比「歷史預測法」平均改善 29%、有一半的商品降低 44%以上，相對於線性迴歸模型、或是在推薦系統中被廣泛使用的演算法―協同過濾，此深度學習模型不論是「本季買過的所有商品」、或是「上季未買，但本季新買的商品」中，皆獲得「各特定會員商品組之誤差分布」、「各商品誤差分布」及「平均誤差」此三項指標之最佳結果。

關鍵字

推薦系統 ； 深度學習 ； 超商

並列摘要

For the offline retail industry with implicit dependence between items, this study proposes a peer-based machine learning model - item iterative deep learning - as an recommendation algorithm. This model can change the output and input data for different target items. Each iteration can filter out the most active customer groups for the target item, focusing on the interactions between the target item and other items. Besides, the iterative process can over-sample some items with less data to increase the number of training samples for less purchased items. In addition, this method only needs to train one model for all target items, instead of training many models for different target items. This study empirically tests predictive performance of the proposed item iterative deep learning using data from Taiwan's most widely-distributed retailing sector, convenience stores. The results show that for specific items purchased by specific members that have not been seen by the model, comparing with the historical predicting method, the proposed model has improved by an average of 29% and by more than 44% for half of those items. Compared to the linear regression model or the collaborative filtering that is widely-adopted in recommender systems, the proposed model has obtained the best result of the two situations below: “all items bought this season” and "not bought in the previous season, but newly purchased in this season", in the mean and distribution of different prediction error metrics.

並列關鍵字

recommendation system ； deep learning ； retail industry

參考文獻

張志斌（2018年9月 5 日）。有關辦理「106年連鎖式便利商店經營概況調查」案。公平交易委員會新聞稿。取自https://www.ftc.gov.tw/internet/main/doc/docDetail.aspx?uid=126&docid=15589

Aggarwal, C. C.(2016). Recommender systems . Cham: Springer International Publishing.

Bansal, T., Belanger, D., & McCallum, A. (2016, September). Ask the gru: Multi-task learning for deep text recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems (pp. 107-114). ACM.

Basu, C., Hirsh, H., & Cohen, W.(1998). Recommendation as classification: Using social and content-based information in recommendation. In Proceedings of the 1998 National Conference on Artificial Intelligence, pp. 714–720

Breese, J. S., Heckerman, D., & Kadie, C.(1998). Empirical analysis of predictive algorithms for collaborative filtering. In Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence, pp. 43-52.