3.3.1. Comment collection and preprocessing.

Dangdang.com, characterized by high product sales volume, extensive reviews, and deep specialization in the book market, served as the data source. The Octopus Collector tool was used to scrape review pages, capturing book titles, review text, and timestamps. The Jieba segmentation package and LTP NLP toolkit were employed for data preprocessing. Initially, invalid texts (duplicates, very short comments, containing only punctuation/emojis) were removed. Sentences were then segmented into clauses and stop words were filtered out. Finally, an effective review text corpus was constructed.

3.3.2. Causal event extraction.

Rule and template-based methods were adopted to identify commonly occurring causal relationships between events. Building upon the causal trigger word and template-based framework for causal event pair extraction by S. Zhao and Sorgente [30], and addressing systematic syntactic and expressive differences between Chinese and English [31], this study integrated the classification system for Chinese causal complex sentences proposed by Chu et al. [32] (Center-Conjoining, End-Dependent, Bi-Ended) to precisely locate cause and effect semantic segments within sentences. Accordingly, ten categories of Chinese causal rule-matching templates were designed (details in Table 1). Given the unstructured nature of Dangdang reviews often leading to sentence component omissions, where traditional Subject-Verb-Object (SVO) recognition performs poorly, the method by Shan et al. [33] was referenced. Events were represented lightweightly using combinations of core noun phrases and verb phrases. Regular expressions were designed for each template type to enable automated causality identification. For example, the template【{Cause}, < Conj...>{Effect}】matches conjunctions like【所以】(suǒyǐ, “so”), corresponding to the Center-Conjoining type. Using this template, a causal event pair like【theory combined with practice】【so】【good book】could be matched. Here, “theory combined with practice” is identified as {Cause}, the cause event, and “good book” as {Effect}, the result event.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Examples of causal syntactic rules.

https://doi.org/10.1371/journal.pone.0341504.t001

3.3.3. Top2Vec theme clustering.

Given the large scale and significant distribution sparsity of the causal event pairs obtained during extraction, and considering that ELGs primarily consist of nodes and directed edges where heterogeneous event pairs exhibit semantic co-occurrence at theme level, theme clustering on the extracted event set was necessary. Merging semantically homologous events into theme clusters not only reduces data dimensionality complexity to optimize graph structure clarity but also achieves semantic abstraction through event generalization at the cluster level. For complex causal scenarios like “one cause, multiple effects” or “one effect, multiple causes”, this clustering process supports feature aggregation of similar event samples, providing a structured analytical foundation for causal mechanism induction and event evolution trend prediction.

In the ELG construction context focused on structured expression of event causal logic, introducing Top2Vec for theme clustering leverages its advantages in deep semantic mining and automatic theme discovery. By mapping event texts into a unified semantic vector space, Top2Vec effectively captures latent semantic associations between documents and automates the identification of hidden thematic structures within large-scale unlabeled data, thereby providing robust support for event classification, causal reasoning, and knowledge organization. Furthermore, this approach enhances information extraction efficiency and graph construction quality, making the resulting ELG more interpretable and practically valuable. Therefore, this study utilized the Top2Vec model for cluster analysis.

3.3.4. Graph interpretation.

Combining the clustered theme demands with cause and result events, cause events were set as source nodes and result events as target nodes, with causal relationships between them as directed edges. The Gephi tool was then used to generate a visualized directed graph. Finally, based on this graph, the purchasing behavior of online book buyers was analyzed.

4.4.1. Analysis of core nodes.

Observations of the graph reveal “Intellectual Resonance” and “Purchase Guidance” as core hubs both visually and structurally. Combining node layout, connection strength, and semantic labels, the core nodes and their associative logic are analyzed as follows.”Purchase Guidance” appears as a large node at the center of the graph. It exhibits strong hub properties within a local sub-network through high-density connections to sub-node clusters like “bought this book,” “purchase books,” “bought this book because...” (Fig 4). Analyzing semantics and connection logic yields two key conclusions:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Nodes Associated with “Purchase Guidance”.

https://doi.org/10.1371/journal.pone.0341504.g004

Decision-Driven Multi-Path: Sub-nodes form three behavioral branches around “Purchase Guidance”:Social Trust: Nodes like “bought because big brother recommended,” “imitated purchase” demonstrate the influence of friend/family recommendations and group trends (“circle following”) on decisions. Social circle “trust endorsement” reduces user selection risk, becoming a non-commercial decision trigger.Such behaviors validate that “trust-based social proof” embedded in social relationships significantly mitigates users’ book selection risks [34]. Crucially, while algorithm-driven recommendations dominate consumption contexts, this mechanism’s non-transactional nature fosters stronger affective bonds with users.

Market Signal: Nodes like “came to read because it’s a bestseller” reflect the guiding role of commercial labels like bestseller lists and platform recommendations. This indicates users rely on external market signals to “simplify” book selection, reducing information search costs.This contrasts with the “spontaneous impulse purchases” characteristic identified by Wang Yanyan et al. [35]. in live-streaming commerce research. Book consumption’s market-signal dependency tends to induce content homogenization, necessitating vigilance against algorithmic suppression of content diversity.

Behavioral Characteristic: Nodes like “buy many books,” “buy physical books,” “searched and then bought books” dissect variations in purchasing behavior forms, reflecting personalized habits like “bulk buying,” “preference for physical format,” and “active searching.”This phenomenon highlights how preferences for physical books epitomize the affective and archival value of print media, in contrast to the prevailing convenience-driven paradigm of digital consumption [36]. Such predilections align with the “scarcity of experiential consumption” theory in cultural economics [37].

Cross-Node Value Coupling: “Purchase Guidance” connects upwards to nodes like “Intellectual Resonance,” “Teacher Recommendation & Shared Benefit,” “Must-Read Children’s Books,” “Children’s Fun Land,” “Nostalgic Reasons,” “Discount & Premium Enjoyment,” “Author Influence,” and “Mirror of History.” It points directly downwards to behavior nodes like “buy books,” “purchase this book,” forming a complete chain: Preceding Influences (Content/Price) → Decision Filtering (Purchase Guidance) → Action Implementation (Buy Books). This structure proves that purchase decisions involve a collaborative filtering process of multi-dimensional signals (content value, price stimulus, social trust, scenario demand), rather than being a linearly driven single-factor behavior.However, traditional planned behavior theory underrepresents the complexity of social-market dynamics in the digital era [38], particularly overlooking cost-conscious segments’ heightened sensitivity to perceived price-piracy associations.

“Fruit of Reflection” appears as a core node, acting as an aggregation hub for post-reading value output and behavioral extension. With high-density connections to over 30 semantic sub-nodes like “can’t put it down,” “provokes thought,” “creates resonance,” and to core nodes like “Book Reading,” “Purchase Guidance,” and “Aesthetic Binding Preference” (Fig 5), it constructs a network structure centered on “Intellectual Resonance,” radiating to reading experience, purchase decisions, and cultural ritualistic feelings. This layout visually demonstrates that “Intellectual Resonance” is the energy transfer station where reading behavior extends into subsequent value chains. Emotional touches, cognitive reconstruction, and social interactions triggered by reading are all translated from experiential perception into behavioral drivers through “Intellectual Resonance.” Sub-nodes around it can be grouped into four dimensions:Emotional Resonance: Nodes like “can’t put it down,” “like,” “deeply resonate” anchor the value of books as emotional carriers, revealing user empathy with stories and emotional tones as a core source of reading stickiness.Distinct from research on aesthetic-driven visual stimuli dominant in cover design studies [39], our event logic graph demonstrates that content-triggered empathy constitutes the primary driver of reading engagement persistence. This finding resonates with the core affective priming mechanisms posited by cultural consumption theory.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Nodes associated with “intellectual resonance”.

https://doi.org/10.1371/journal.pone.0341504.g005

Cognitive Reconstruction: Nodes like “provokes thought,” “induces deep thought,” “see the big from the small” reveal the impact of reading on restructuring users’ cognitive frameworks, proving deep reading is essentially a knowledge reproduction process.Content Evaluation: Nodes like “good content,” “easy to understand,” “grew fond of Xiong Peiyun” both retroactively validate pre-purchase decisions (e.g., high-quality content drives deep resonance) and provide emotional anchors for repurchase loyalty (e.g., preference for authors/series).Social Propagation: The connection between “colleague recommendation” and “Purchase Guidance” embodies users’ desire to recommend based on reading gains, turning books into social currency, driving the closed loop “Read → Resonate → Socially Share → Re-purchase. Furthermore, “Intellectual Resonance” weaves multi-dimensional behavioral loops through its connections. Upwards linkage with “Book Reading” shows edges like “easy to understand,” “triggered self-reflection” reveal reading quality directly influences resonance depth. Connection rightwards to “Purchase Guidance” forms an implicit path; for instance, recommendation desires sparked by resonance can transform into motives for purchase decisions within social circles. Downwards linkage to “Aesthetic Binding Preference” reflects the collection logic in nodes like “good content”; users might be driven by profound resonance to seek cultural ritualistic feelings (e.g., demand for beautiful bindings or collector’s editions increases after deep engagement with quality content).

4.4.2. Holistic analysis.

Analyzing the causal relationships among theme events in Dangdang’s online purchasing behavior allows parsing users’ internal logic and demand characteristics across four dimensions: reading motivation, purchase decision, feedback behavior, and group heterogeneity.

4.4.2.1 Multi-Dimensional reading motivation. Motivations manifest as three core orientations: intellectual enlightenment, historical insight, and emotional connection. Firstly, the need for intellectual stimulation is shown through the pursuit of mental resonance and cognitive expansion after reading, reflecting users’ tendency to view books as vehicles for enlightenment and value reconstruction, often aligned with preferences for philosophical or social science texts. Secondly, historical insight motivation focuses on the cognitive tradition of “learning from history,” where users explore practical reference points through historical texts, demonstrating the empowering value of history for real-world decisions. Thirdly, emotional connection motivation manifests as affective projection onto specific creators or cultural symbols, such as purchasing driven by identification with Su Dongpo’s literary style or Liu Cixin’s sci-fi narrative, essentially an extension of “fandom economy” into book consumption, reflecting the driving role of creator charisma on choices.

4.4.2.2 Multi-faceted purchase decision mechanism. Purchase decisions are jointly driven by three intertwined factors: product attributes, social trust, and price perception. Regarding product attributes, physical aspects like binding design and printing quality become core decision factors for some users, pointing to the aesthetic and collectible value of physical books. In the social trust dimension, recommendations from authorities or trusted individuals like teachers or colleagues act as key decision drivers, especially in student groups or parent-child education scenarios, highlighting the transmission logic of trust within social networks influencing information credibility and consumption choices. Price perception exhibits duality: promotional tactics like discounts and points stimulate both repeat and first-time purchases. However, the negative association between “low price - piracy/low quality” also reveals that price strategies require quality assurance as a prerequisite to avoid degrading user experience.Moreover, unlike algorithm-dependent programmatic advertising paradigms [40], social trust mechanisms highlight the irreplaceable role of interpersonal networks in decision-making processes, particularly within information-overloaded environments.

4.4.2.3 Dual nature of feedback behavior: emotion and supervision. User feedback behavior possesses dual attributes: emotional venting and quality supervision. Negative reviews driven by poor experiences constitute a quality supervision mechanism for platforms and products, their content directly pointing towards optimization directions for products and services. Consumption feedback driven by nostalgic feelings, like “revisiting classics” or “switching from e-book to physical copy,” implies a demand for emotional awakening of “reading memories,” reflecting the cultural ritualistic feeling and emotional continuity of reading—such feedback serves both as emotional anchors for individual reading histories and provides clues for platforms to mine the long-tail value of “classic IPs.”Finally, group heterogeneity was evident. Clustering results further revealed distinct subgroup characteristics: Children’s reader groups require balancing content fun and educational value, suggesting optimization through “child-friendly” narrative design and home-school collaborative recommendation mechanisms. History enthusiast groups value the real-world explanatory power of historical texts, necessitating enhanced thematic depth and cross-scenario interpretation (e.g., “History + Workplace,” “History + Life”). Price-sensitive groups require maintaining quality baselines during promotions to avoid the “low price - low quality” vicious cycle.

Overall, online book consumption is a closed-loop process of “Motivation Triggering → Decision Implementation → Feedback Iteration.” Strategies like precise content matching, activating social trust, balancing price and quality, and establishing negative review response mechanisms are required to synergistically achieve user need fulfillment and enhanced lifetime value.Crucially, this study’s interpretation of user motivations (e.g., “cognitive empowerment,” “affective bonding”) constitutes neither speculative inference nor subjective conjecture. Rather, it derives from functional categorization grounded in users’ recurring self-reported expressions within reviews (e.g., “provokes introspection,” “developing admiration for the author”), systematically mapped against the Stimulus-Organism-Response (S-O-R) framework and Theory of Planned Behavior [41]. While event logic graphs do not directly measure psychological states, the captured causal chains (e.g., “author’s unique perspectives → purchase decision”) serve as observable proxy variables for behavioral intent. This methodological approach aligns with explanatory frameworks central to digital ethnography and computational social science.