Development of NMF methods considering consumers, products, and time for understanding relationship between purchasing behavior and psychology

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K13822
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 07090:Commerce-related
• Research Institution: Kyoto University
• Principal Investigator: Abe Hiroyasu 京都大学, 医学研究科, 助教 (40807963)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 消費者購買行動 / 消費者心理 / 非負値行列因子分解 / ベイズモデル / 零過剰モデル / 負の二項分布モデル / 基底数選択 / ウェブサイト閲覧履歴

Outline of Final Research Achievements

A novel new nonnegative matrix factorization (NMF) was developed. Using the user-item matrix data and user’s characteristics information, it can estimates the following two simultaneously: behavioral pattern expressed by items, and the characteristics of the users who did not take action against an item which the users should have done so according to the estimated behavioral pattern. Thanks to variational inference of both estimations, model order selection can be done without any tuning parameters. The interpretation procedure was demonstrated and the results are compared with existing methods through the application to the access log data of a book shopping website. In addition, focusing on the hierarchical structure of item labels, the other new NMF method was developed and its performance was evaluated. The method extracts cooccurrence relationships within each upper labels (homogeneous basis) and co-occurrence relationships between upper labels (heterogeneous basis).

Free Research Field

統計学

Academic Significance and Societal Importance of the Research Achievements

行列内の0要素を推定行動パターンと関係あるもの(related zero)とないもの(unrelated zero)に分け、後者の確率が高い対象や属性を商品(群)ごとに特定できるデータ解析法を開発した。また異なるラベルをもつ商品(群)の共起関係を抽出する方法を示した。これらの手法を様々な行動履歴データに適用することで、人々の行動に対して深い考察を与えることができる。また、マーケティングの視点では、単純な共起関係に基づかない、より繊細で熟考された施策提案を提供できる。なお開発手法はいずれも変分推論のおかげで調整パラメータが実質ない解析法となっており、分析者にとって使いやすいものとなっている。