Mining user satisfaction decision stages from online reviews is helpful for understanding user preferences and conducting user-centered product improvements. Therefore, this study develops a two-stage nonlinear user satisfaction decision model (USDM). First, we use word2vec technology and lexicon-based sentiment analysis to mine the sentiment polarity of each product attribute in the reviews. Then, we develop KANO mapping rules using utility functions to classify consumer preferences based on attribute importance. Based on this, a two-stage nonlinear USDM is developed to describe post-purchase evaluation behavior. In the first non-compensatory stage, consumers determine their initial satisfaction level based on the performance of basic attributes. If the performance of these attributes is poor, it is almost impossible for users to be satisfied. In the compensatory stage, the performance of the remaining attributes collectively affects final satisfaction through participation in user utility calculation. With the use of reviews from JD.com, we develop a genetic algorithm to determine feasible solutions for the USDM and verify its validity and robustness. The USDM is proven to be effective in predicting user satisfaction compared to other classic models and machine learning algorithms. This study provides a universal pattern for user satisfaction decisions and extends the study on preference analysis.

中文翻译：

---

基于在线评论挖掘的两阶段非线性用户满意度决策模型：考虑非补偿和补偿阶段

从在线评论中挖掘用户满意度决策阶段有助于了解用户偏好并进行以用户为中心的产品改进。因此，本研究开发了一个两阶段非线性用户满意度决策模型（USDM）。首先，我们使用word2vec技术和基于词典的情感分析来挖掘评论中每个产品属性的情感极性。然后，我们使用效用函数开发 KANO 映射规则，根据属性重要性对消费者偏好进行分类。在此基础上，开发了两阶段非线性 USDM 来描述购买后评估行为。在第一个非补偿阶段，消费者根据基本属性的表现来确定自己的初始满意度。如果这些属性的表现很差，用户几乎不可能满意。在补偿阶段，剩余属性的表现通过参与用户效用计算来共同影响最终满意度。利用京东的评论，我们开发了一种遗传算法来确定 USDM 的可行解决方案，并验证其有效性和鲁棒性。与其他经典模型和机器学习算法相比，USDM 被证明可以有效预测用户满意度。这项研究为用户满意度决策提供了一个通用模式，并扩展了偏好分析的研究。