從高維度消費紀錄挖掘隱藏偏好

Abstract

理解消費者行為在許多領域中都被認為是重要的信息，尤其是在\r\n市場營銷中。但是，複雜的行為以及高維度、動態數據使得從中提取\r\n有意義的洞察變得困難。為了解決這些問題，我們結合了非負矩陣分\r\n解 (NMF) 和遞迴神經網絡 (RNN)，提出深度動態神經網路 (Dynamic\r\nDeep NMF)，來捕捉動態模式。NMF 的分解幫助我們總結消費主題\r\n和用戶對主題的興趣。而 RNN 的遞迴特性則幫助我們捕捉消費者\r\n的動態模式。我們設計了一個模擬實驗，產生模擬數據以測試 NMF\r\n和 Dynamic Deep NMF 的性能。最後，我們使用一個實證數據來展示\r\nDynamic Deep NMF 會找到什麼隱藏主題，以及它如何捕捉動態用戶\r\n行為。

To understand users’ consumption behavior is found critical in many fields,\r\nespecially marketing. But the complex behavior embedded in high-dimensional,\r\ndynamic transaction data make it hard to extract meaningful insights. To\r\ntackle such problems, we combine the non-negative matrix factorization(NMF)\r\nand recurrent neural network (RNN) to develop a Dynamic Deep NMF in order to capture dynamic patterns and elicit hidden preferences. The decomposition of NMF helps us to summarize the consumption topics and users’ interests among the topics. And the recurrent properties of RNN helps us to\r\ncapture the dynamic pattern of users’ interests. We also develop a simulation experiment to generate synthetic data to test the performances of NMF and Dynamic Deep NMF. Finally, we use an empirical dataset to demonstrate\r\nwhat hidden topics the Dynamic Deep NMF could find and how the method captures dynamic user behaviors.