Abstract: COVID-19 has had a profound impact on the global economy and financial activities, driving a surge in consumer demand for online shopping and a sharp increase in business volume for e-commerce platforms. Achieving accurate product recommendations based on consumer demand has become a hot topic of research. Online product reviews are crucial for consumers to assess product quality and merchant credibility, directly influencing purchasing decisions. Therefore, leveraging online review information for intelligent product recommendation is essential. Compared to user ratings, online text reviews can more precisely express various emotional product preferences, often coming in uncertain emotional descriptions such as ‘very like it’, ‘like it’, ‘dislike it’, etc. But transforming the data into fuzzy sets presents complexity and challenges.
Existing research lacks consideration for cases where a single criterion contains multiple emotional preferences, leading to information loss during comment processing and transformation. In real-world decision-making scenarios, customers often exhibit bounded rationality, necessitating further research on product recommendation for loss-averse customers based on rich emotional preference online text reviews. Presently, limited research considers customer loss aversion psychology, primarily utilizing the prospect theory and TODIM method to characterize bounded rational behavior. However, traditional TODIM methods encounter application paradoxes and complexities, resulting in biased purchase decisions. Therefore, a new product recommendation model is proposed based on online text reviews by integrating belief structures, uncertain probabilistic linguistic term sets (UPLTSs), and the generalized TODIM method to address multiple emotional preferences and characterize customer loss aversion psychology effectively.
Initially, a 5-level linguistic term is employed to convert user reviews, considering the nature of online reviews. Belief structures are introduced to fully leverage their powerful uncertain information processing capabilities in transforming multiple emotional preferences in text reviews. Next, ‘benefit type’ and ‘cost type’ criteria in product reviews are normalized before calculating criterion weights based on the similarity between evidence bodies. The higher the similarity, the greater the criterion weight assignment. Subsequently, belief structures, UPLTSs, and the generalized TODIM method are integrated to devise a novel product recommendation model suitable for loss-averse customers with multiple emotional preferences. To efficiently address uncertain probability linguistic issues, on one hand, the reliability function and plausibility function of belief structures are used to transform evaluations, resulting in an uncertain probability linguistic evaluation matrix. On the other hand, to mitigate the paradoxes and complexities associated with the traditional TODIM method in application, the generalized TODIM method is extended to an uncertain probability linguistic environment, effectively capturing customers’ loss aversion psychology as observed in real decision-making scenarios. This streamlines the computational processes and broadens the application scope.
To demonstrate the effectiveness of the recommendation model, a case study on mobile phone recommendations is conducted. Four phones are selected based on customer requirements, recommending the highest-rated phone. To validate the robustness and effectiveness of the proposed model, a sensitivity analysis is conducted using loss aversion coefficients, and a comparison is made with existing researches. The results show that changes in the degree of risk aversion do not affect the product recommendation order, indicating insensitivity of the recommendation results to changes in the level of risk aversion, thereby demonstrating the robustness of the proposed recommendation model. Through the comparative analysis, it is evident that in a product recommendation model based on text reviews, handling and transforming multiple emotional preferences, and considering customers’ bounded rational behavior are extremely necessary, resulting in recommendation outcomes that better align with human innate thinking habits, reflecting customers’ needs more authentically, and holding significant theoretical and practical value for actual product recommendation.
This study focuses on personalized product recommendation for specific needs. Future research will consider customer groups without specific needs and apply the proposed recommendation model to other fields. Through this extension, deeper insights will be provided for website optimization management strategies and a new research perspective will be offered for personalized recommendation problems in different domains.

Key words: online text review, belief structure, uncertain probabilistic linguistic term sets, the generalized TODIM method, product recommendation

摘要： COVID-19对全球经济和金融活动产生了深远影响,推动了消费者在线购物需求激增和电商平台业务量飙升。在线评论信息作为消费者辨别产品质量和商家信誉的关键途径,对购买决策有直接影响。与用户等级评分相比,在线文本评论更能精准地表达产品情感偏好,本文由此提出基于在线文本评论考虑顾客损失厌恶心理的产品推荐模型。首先,引入信度结构,充分发挥其强大的不确定信息处理能力,转化文本评论信息中的情感偏好,并基于证据距离公式测算相似性,获取属性权重。其次,为高效处理不确定概率语言问题,依据信度结构的可信度函数和似真度函数转化评价信息,获取不确定概率语言评价矩阵。进而,将广义TODIM方法拓展到不确定概率语言环境中,构造有效刻画顾客损失厌恶行为特征的产品推荐方法。最后,通过实例应用、灵敏度分析和对比分析,进一步验证了所提推荐模型的实践性和有效性。

关键词: 在线文本评论, 信度结构, 不确定概率语言术语集, 广义TODIM方法, 产品推荐