Abstract:

Against the backdrop of an abundant supply of consumer goods and the continuous elevation of consumer demand levels， brands have been positioned as a core factor influencing purchasing decisions. However， in mainstream personalized recommendation systems， users' behavioral data regarding product attributes is primarily relied upon， with brands being simplified to mere identifiers or even completely ignored. As a result， the recommendation logic is reduced to a mere function filter， which is incapable of understanding the symbolic value and emotional connections that are built by brands in the consumer's mind. This limitation is manifested in a triple dilemma： the loss of commercial value， the emergence of user experience pain points， and the homogenization of platform competition. Therefore， how to move beyond product-attribute-level recommendations， delve into and quantify brand characteristics based on consumer mentalities， and thereby identify individuals' personalized brand preferences has been identified as a key research question for the transition from product matching to brand matching.To address this issue， the lipstick category is taken as an example in this study. By analyzing over 200，000 user comments on the Weibo platform， two core types of brand relationships are constructed. The first is Brand User Association， which is measured by using a multi-level attitude identification framework to determine bloggers' emotional tendencies towards brands， followed by the calculation of the overlap of users who hold the same attitude towards two brands. The second is Brand Perceived Similarity， which is based on a brand feature system encompassing 4 main dimensions， 11 sub-dimensions， and 83 style categories. Each brand is represented as a normalized feature vector， with similarity being mapped through vector distance. Based on this， 620 consumers are recruited to conduct genuine preference tests for 30 mainstream lipstick brands in a simulated e-commerce environment. The data are divided into 10 known brands and 20 target brands. For each “user-target brand” pair， a 34-dimensional feature vector is constructed， integrating users' preferences for known brands， the perceived similarity and user association between target brands and known brands， and statistical aggregation features. Finally， multiple machine learning algorithms are employed to train classification models in order to verify their predictive power.The research results indicate that a hybrid collaborative model integrating perceived similarity and user association can be effectively used to identify consumer brand preferences. The analysis reveals that the perceived similarity network reflects the mental competition landscape that is shaped by marketing narratives， while the user association network unveils the market competition landscape that is determined by price and channels. More importantly， significant heterogeneity has been found to exist in consumer preferences： At the group level， average perceived similarity follows a normal distribution， whereas average user association shows a right-skewed distribution， suggesting that consumers seek moderate similarity cognitively but tend to explore across communities behaviorally. At the individual level， as the number of preferred brands increases， both the similarity and association within their brand portfolios are observed to significantly decrease， revealing a strategic shift from homogenized identification to heterogeneous diversification. This study has theoretically expanded the intelligent recommendation perspective by incorporating brand mentalities； empirically revealed the personalized hybrid mechanism driving brand preferences； methodologically innovated a dynamic identification framework combining text mining and machine learning； and practically provided platforms with a viable path for differentiated marketing， transitioning from precise pushing to deep resonance.

Key words: brand preference, social networks, user-generated content, perceptual similarity, user relevance