The construction of a knowledge organization system must be guided by a profound understanding of the inherent knowledge associations within the data to model. With a specific focus on the information interaction characteristics of investors, this study propose that the knowledge in the stock market ecosystem can be structured through five core association patterns: categorical, statistical, spatiotemporal, event-centric, and behavioral associations.

1. Categorical Associations: As the most fundamental and “explicit” pattern, this describes the taxonomic and hierarchical structures among conceptual entities in the stock market. It embodies the “multi-perspective” and “multi-layered” logic within big data. In the SKBG, this pattern is primarily realized through the financial ontology construction at the content layer, which defines classes and their subClassOf relationships, establishing a structured, multi-dimensional schema for understanding investors and their interactions.

2. Statistical Associations: Within the Data-Information-Knowledge-Wisdom (DIKW) hierarchy (McDowell, 2021), raw social posts and price ticks constitute “data” and processed metrics like sentiment scores or correlation coefficients are “information”. The SKBG framework advances this transformation by systematically contextualizing information within a rich semantic and relational fabric. However, when the correlation coefficient is interpreted within the context of financial theory or market behavior (e.g., interpreted as linkage due to the attention of common investors) and can be used to support prediction or decision-making, it is transformed into “knowledge”. Thus, this “implicit” pattern reveals correlations between knowledge units through computational analysis, highlighting the “computable” nature of knowledge associations. Examples include stock price co-movement and investor similarity. The SKBG supports this pattern through its constraint layer, where graph algorithms and Cypher queries are used to calculate correlation strength, structural patterns, and temporal changes, thereby discovering new knowledge that is not directly observable.

3. Spatiotemporal Associations: This pattern captures the “evolvable” characteristic of knowledge, describing how the states and relationships of entities change over time and across spaces (geographical or institutional). The SKBG’s novel social dynamics layer is specifically designed to manage this. It handles temporal relationships to describe evolutionary patterns and spatial influences on information migration, generating “new knowledge” about market dynamics.

4. Event-Centric Associations: As fundamental components of finance, events can trigger a series of market reactions and propagate through investor sentiment. This pattern is a domain-specific semantic association mode. The SKBG captures it by integrating Event as a core class in the ontology and modeling its connections to involved entities. The analysis of how events drive information diffusion is then enabled by combining the social network’s propagation paths with the ontological event semantics.

5. Behavioral Associations: This pattern encompasses all information interaction behaviors like commenting/following and emotional interactions among investors, which are driven by information needs. Behavior both relies on prior knowledge and generates new knowledge, exhibiting dynamic and spiraling evolutionary features. The SKBG integrates this directly by fusing the extracted social network into the content layer. These behavioral links are treated as first-class knowledge associations, allowing the study of how interactive behaviors are intertwined with static knowledge in a continuous process of knowledge accumulation and value discovery.

The five association patterns outlined above provide the theoretical typology for structuring knowledge in the stock market ecosystem. The categorical associations are embedded in our ontological framework, providing a static schema for organizing entities. Statistical associations are the core analytical lens, through which we apply network analysis and computational algorithms to uncover implicit correlations from the behavioral data, while spatiotemporal associations are considered in our longitudinal data collection. The event-centric and behavioral associations are global, derived patterns that enable the leap from “big data” to “big knowledge”, revealing the operational mechanics of the stock market complex system from the perspectives of information itself and the investors. Event-centric associations provide a semantic contextual layer, where we interpret observed statistical and behavioral anomalies within the context of market events. Behavioral associations serve as the primary data foundation, directly represented by the investor-investor and investor-stock attention/following networks we construct.

The design of SKBG framework is fundamentally guided by three interlocking principles derived from the core tenets of KO, each corresponding to a critical dimension of the five knowledge association patterns.

First, the principle of representational completeness demands that the framework extend beyond the taxonomic and hierarchical relations that have traditionally dominated KO systems. This expansion is necessary to incorporate the full spectrum of knowledge associations present in socio-technical systems, including not only categorical associations but also statistical correlations, spatiotemporal evolutions, event-centric linkages, and behavioral interactions. By ensuring that the model captures both the static semantics of what is known and the dynamic processes of how knowledge evolves and circulates, this principle directly operationalizes the multi-modal nature of knowledge in domains such as financial markets.

Second, to maintain logical coherence across this expanded representational scope, the principle of semantic consistency governs the integration of formally defined ontological structures with the fluid, often informal data generated through social interaction. This is achieved by grounding all social entities and relations within a unified financial-social ontology, whether a transient reply in a forum or a long-term investment holding. Through this ontological anchoring, along with a layer of constraint-based reasoning that enforces logical rules and temporal validity, the framework bridges the persistent divide between the stable semantics of knowledge graphs and the dynamic, often unstructured nature of social networks. This ensures that the resulting knowledge structure remains machine-interpretable and logically sound, even as it accommodates the real-time flow of social data.

Finally, the principle of associative reasoning ensures that the system functions not merely as a passive repository but as an active engine for knowledge discovery. This principle enables multi-modal inference pathways that traverse and interconnect the different association types. For instance, a reasoning process might begin with a behavioral association, cross-reference it with categorical associations, situate it within a spatiotemporal context, and link it to an event-centric trigger, ultimately materializing a new statistical association. By supporting this kind of synthetic, cross-pattern reasoning, the SKBG framework embodies the ultimate objective of advanced knowledge organization: to transform a complex, interconnected dataset into an intelligible and actionable epistemic structure that reveals patterns, supports explanation, and generates new insight.

These three principles (representational completeness, semantic consistency, and associative reasoning) form a cohesive theoretical foundation that translates the abstract taxonomy of knowledge associations into a robust, functional architecture. They ensure that the SKBG operates as a genuine knowledge organization system, capable of representing, querying, and reasoning about the intricate socio-epistemic phenomena that define modern information ecosystems.

The content layer is the process of redefining the structured knowledge from the ontology and the unstructured relationships from the social network within a unified property graph model. This process is the technical realization of the foundational categorical associations and the direct integration of behavioral associations. Its primary function is to resolve the structural and semantic heterogeneity between the formal ontology and the informal social network. This is achieved through a process of semantic lifting, where social entities are mapped to ontological classes and social interactions are either directly represented as edges or reified into node-centric behavioral constructs. The output of this layer is a coherent property graph where all entities have well-defined types and relationships, ready for further reasoning and analysis.

The constraint layer imbues the SKBG with machine-interpretable intelligence. It operates by defining and executing a set of rules over the fused graph, and operationalizes reasoning over statistical associations and enforces the consistency of categorical associations. These rules serve consistency preservation and knowledge discovery.

We implement Cypher-based algorithms to check for inconsistencies in the categorical associations and introduce new rules to discover and validate statistical associations and other implicit patterns, the methodological details will be described in Chapter 4.

The social dynamics layer ensures the SKBG is not a static snapshot but a living representation of the market. This novel layer is specifically designed to capture the “evolvable” nature of knowledge, directly addressing spatiotemporal associations and the dynamics of event-centric associations. It introduces the concepts of Time and Space as first-class entities, allowing every behavior and event to be temporally anchored. This enables historical querying and trend analysis. Furthermore, this layer hosts dynamic computation processes, such as recalculating the influenceScore of an Investor node based on the evolving network structure. This functionality is crucial for modeling the propagation of information during specific events, effectively linking behavioral associations with event-centric and spatiotemporal associations.

In summary, the SKBG framework presented in this chapter provides a theoretically grounded and architecturally sound blueprint for constructing a next-generation knowledge organization system. The Content Layer defines what is known, the Constraint Layer defines what is logically consistent and derivable, and the Social Dynamics Layer defines how knowledge evolves.This layered approach ensures that the SKBG is not merely a concatenation of two graphs, but a logically self-consistent and semantically rich knowledge representation suitable for complex analytical tasks.

We collected data on A-shared stocks associated with the China Securities Index (CSI) 300 Index from the China Stock Market & Accounting Research (CSMAR) database, including information such as industry classification, concepts, shareholding ratios, and basic trading data for the year 2023. The CSI 300 index is a standard benchmark representing large-cap, liquid stocks, making it a pertinent universe for studying market-wide phenomena. The Eastmoney forum is the largest online investment community in China, providing a rich and ecologically valid source of retail investor interactions. Consequently, the findings are most generalizable to the social dynamics surrounding large-cap stocks within China’s dominant retail investor ecosystem.

This research adhered to strict ethical guidelines and legal standards. All data from the Eastmoney forum were collected from publicly accessible web pages. To protect user privacy, all raw user IDs and IP addresses were immediately hashed into irreversible anonymized identifiers upon collection. No attempt was made to link these online identities to real-world personal information. All analyses presented are based on aggregated network patterns or anonymized individual behaviors, aligning with the provisions for scientific research under China’s Personal Information Protection Law (PIPL).

After filtering out stocks that were suspended or had missing data during trading, a total of 288 stocks remained for analysis. Subsequently, the paper crawled investor comment text data, user attention (focused stocks and other investors), and other information interaction data from the stock forums of the 288 stocks on the Eastmoney.com Stock Forum website (https://guba.eastmoney.com/). We capture details of 2,584,057 online investor IDs and their corresponding pseudonyms, stock forum seniority, IP addresses, and more. Further insights include attention data for 1,441,579 investors towards specific stocks, as well as details on the followers of 1,627,370 investors.

The stock market public opinion text contains a large number of short investor comment texts and a small amount of news public opinion information. Among them, due to the poor language standardization of investor comment texts, low information content, and obvious value sparsity, it is difficult to directly extract knowledge from these texts. For these short texts, the study adopts the method of identifying the topics of investor comments as the “key knowledge” of comment information, that is, investors posting a comment can be converted into investors paying attention to a topic; for news events, element extraction is performed, and the entity relationships and theme events obtained are used as the “complementary knowledge” of the social knowledge graph.

We take the “Kweichow Moutai” stock bar as an example to identify the hot topics that investors pay attention to every month to describe the characteristics of investor comment texts. The study uses the BERTopic topic model to identify hot topics (Grootendorst, 2022). BERTopic performs topic modeling based on BERT word vectors. It uses Transformer and c-TF-IDF to create dense clusters and achieves better results than the LDA model in short texts. In the process of training the model, we use open-source pre-trained Chinese word vectors in the financial field to further enhance the effect of BERTopic, and obtains the theme of each month in 2022 through topic merging and screening. The study limits the number of topics to less than 20 and obtains the topic distribution for 12 months after parameter tuning. After eliminating topics that have no practical significance or only express investor emotions, they are summarized and summarized to form the hot topics of investor comments shown in Table 3.

Event element extraction is the core task of event extraction, which aims to extract corresponding event arguments, roles, etc. from the text according to the event representation framework. This paper regards event element extraction as a sequence labeling task, that is, given a sequence to be labeled $X=\{x_1,x_2,\mathrm{\dots },x_N\}$, a label $t_j$ is predicted for each $x_i$, $i\in [1,N]$, $t_j$$\in$T=[$t_1$, $t_2$, …, $t_M$] is a set of all possible labels. The study proposes the idea of jointly extracting event arguments, roles, and trigger words based on sequence labeling, that is, by designing a labeling strategy, while extracting event elements such as entities, time, and place in financial news, it can also obtain two types of roles of the event agent and the patient, and obtain the key type of each event by labeling the core trigger words. In model selection, the study combines the financial field pre-training model FinBERT with dynamic word representation capabilities, Bi-LSTM that can add trigger word information encoding, and CRF structure with global feature constraint effect, and introduces the Attention mechanism. The experimental results are shown in Table 4.

With prepared data, we proceeded to instantiate the ontology and populate the graph, following the layered mapping methodology. The static ontology was first brought to life by creating nodes for all resolved entities, while the core of the fusion occurred with the dynamic ontology instantiation and behavioral reification. Here, each social media post was not modeled as a simple edge but was reified into a rich node. This node was then linked to its author via [:PERFORMS] and to the discussed listed companies via [:HAS_TOPIC] relationships, which were created in Section 5.1.2. This design elegantly captured behavioral associations and linked them to financial entities. Furthermore, corporate events like earnings announcements were instantiated as nodes and connected via [:PROVIDES_CONTEXT_FOR] to all relevant PostingBehavior nodes within a defined timeframe, explicitly forging event-centric associations. Finally, to anchor the graph in spatiotemporal associations, all Behavior and Event nodes were explicitly linked to Time nodes through [:OCCURS_AT] relationships.

Following the instantiation of the dynamic ontological framework, a critical step of refining and calculating dynamic associations was undertaken to imbue the graph with nuanced, computable relationship semantics. This involved three key advancements beyond simple relationship creation.

There are significant differences in the types of dynamic interactions between institutional investors and individual investors. The interaction between the former is mainly reflected in the shareholding (position) relationship, while the interaction between individual investors (this article refers to online investors) is mainly reflected in the information exchange process. social relationships generated. In addition, according to the length of investor interaction, interaction types can be divided into three types: long-term steady-state interaction, short-term interaction and instantaneous interaction. For long-term steady-state interaction relationships, they can be regarded as a static relationship, so only the triplet $(s,p,o)$ is needed to describe the interaction between entities; for short-term interaction relationships, they can be regarded as temporal association relationships generated within a certain period of time, so the time $t_{start}$ and $t_{end}$ are needed to represent the start and end moments of the relationship respectively, that is, the five-tuple $(s,p,o,t_{start},t_{end})$ is needed to characterize this dynamic interaction; and the instantaneous interaction relationship is a relationship that only occurs at a certain moment, so it only needs to be represented by the timestamp T representing a certain moment, so the four-tuple $(s,p,o,T)$ can be used to describe the instantaneous interaction relationship between two entities.

The emotional interaction relationship of investors is mainly reflected through the information they publish. If investors have a “co-attention” behavior for a certain stock, and the emotional scores of the comments published by the two investors in the same period are similar, it is considered that the two have a high degree of emotional consistency; if the emotional scores of the comments published by two investors who pay attention to a certain stock are quite different, it means that there is emotional difference between the two, otherwise it is considered that there is no emotional interaction relationship between the investors. The study calculated the emotional scores of all comments in the forum by calling Baidu API (the score range is [0,1]), and calculated the emotional interaction score (Sentimental Interaction Score, SIS) between investors, as shown in the calculation Eqn. 1. This score was materialized as a property on the [:SIMILAR_TO] edge between investors or used to create new [:SENTIMENT_ALLY]/[:SENTIMENT_OPPOSITE] relationships, thereby making implicit sentiment dynamics explicit and queryable.

(1)$$SI{S}_{AB\cdot T}=|Sentimen{t}_{A\cdot T}-Sentimen{t}_{B\cdot T}|$$

The symbolic representation of multiple semantics is the process of uniformly representing the multi-modal knowledge associations between knowledge units with investors as the core as vector tuples. For example, there are both “follow” and “comment” relationships between investor A and stock A, and the comments have emotional attributes, so the result after symbolic representation is shown in Fig. 4. Symbolic representation simplifies the modeling complexity, and while ensuring the computability of large social knowledge graphs, it can also provide better interpretability for graph-based query and analysis tasks.

Fig. 4.

Example of multiple semantic symbolic representation results.

According to the scenario and service requirements, different weights are assigned to each element of the symbolically represented multi-semantic associated knowledge tuple. The resulting set of values ​​that can reflect the degree of knowledge association between each knowledge unit is the Knowledge Association Degree (KAD), and its value range is [0,1]. At time T, KAD of knowledge unit a to b shows in Eqn. 2, where ${\sum }_1^n\eta =100\%$, and $r_i(1\le i\le n)$ represents the symbolic representation result of the multi-semantic relationship between node a and node b (excluding timestamp representation).

(2)$$KA{D}_{a\to b|T}={\eta }_1{r}_1+{\eta }_2{r}_2+\mathrm{\dots }+{\eta }_n{r}_n$$

KAD indicates the similarity between knowledge units. Since the weight is defined as a hyperparameter, it can be flexibly calculated according to the actual application scenario. The core purpose of this study is to propose a unified data intelligence modeling framework. To ensure universality, the study was treated with equal weights.

The construction process culminated in the final graph generation and validation. The fully populated SKBG in Neo4j resulted in a substantial and richly connected knowledge base, comprising approximately 1.5 million nodes and 1.8 million relationships, including key types such as Investor, Stock, PostingBehavior, HOLDS_SHARE_OF, FOLLOWS, and HAS_TOPIC (shows in Fig. 5). To ensure the graph’s logical integrity, the consistency reasoning algorithms from the methodological framework were executed. Through the comprehensive and iterative process, a semantically consistent SKBG was generated, providing a powerful and verifiable substrate for the multi-faceted application scenarios that follow.

Fig. 5.

Financial SKBG visualization results stored in Neo4j.

We first focus on a micro-level analysis: generating nuanced, multi-dimensional portraits of investors. Traditional models often rely on a single data dimension, such as network structure or trading history, leading to a fragmented view. Our approach, powered by the SKBG, overcomes this by synthesizing categorical, behavioral, and statistical associations into a unified investor profile, enabling the identification of subtle yet influential investor archetypes that would otherwise remain hidden.

We programmatically query the SKBG to compile a comprehensive feature vector for each investor. From the categorical associations defined in the static ontology, we extract fundamental attributes like the investor’s type (RetailInvestor or InstitutionalInvestor) and their portfolio composition, calculated from holdsShareOf edges (e.g., number of holdings, concentration in specific sectors). From the behavioral and spatiotemporal associations instantiated in the dynamic ontology, we compute temporal dynamics including their posting_frequency, avg_sentiment_score across all their PostingBehavior nodes, and their activity_volatility (the standard deviation of their daily post count). From the statistical associations materialized through graph algorithms, we derive powerful network metrics such as a global influence_score computed using the PageRank algorithm, and community_centrality measure to quantify their role as a bridge between different investor groups. With the rich feature vectors for all investors assembled, we proceed to investor typology construction through unsupervised clustering. We employ the K-means algorithm on the standardized feature matrix to uncover natural groupings within the investor population without preconceived biases. The optimal number of clusters is determined using the silhouette score. This data-driven method reveals distinct archetypes based on the synthesized knowledge within the SKBG. This multi-modal feature extraction is formalized through a dedicated computational procedure (see Algorithm B.1 in Appendix Table 11) that systematically queries the fused graph.

The analysis results shown in Table 5 reveals several distinct investor typologies that validate the SKBG’s analytical value. Notably, we identify a “Cross-Modal Influencer” archetype, investors who exhibit categorical characteristics of retail participants but demonstrate behavioral and statistical patterns typically associated with institutional actors. These individuals maintain highly concentrated portfolios (low portfolio_diversity), exhibit strategic timing in their social interactions (high activity_consistency), and wield network influence (influence_centrality) disproportionate to their formal classification. The discovery of this archetype exemplifies the unique insight generated through multi-associational analysis, as it emerges precisely from the tension between different knowledge association patterns within SKBG. By identifying “Cross-Modal Influencers” through their characteristic fusion of retail categorization with institutional-grade behavioral patterns, financial regulators can prioritize monitoring efforts on actors who may disproportionately impact market sentiment. Similarly, investment services can leverage these nuanced portraits to develop targeted engagement strategies that account for both the formal classification and actual behavioral patterns of investor segments. This scenario thus validates the SKBG not merely as a data repository but as a sophisticated analytical framework that generates actionable intelligence by revealing the complex interrelationships between different dimensions of financial knowledge.

We establish KMeans clustering on investor features without SKBG-based features as the baseline and quantify SKBG’s marginal gain, the cluster results shows in Fig. 6. Table 6 further represents the experimental results with a rigorous baseline comparison. SKBG improves silhouette score by 2.52% and prediction accuracy by 0.52%; Cluster separation remains stable with consistent classification (ARI (Adjusted Rand Index) = 0.5831), demonstrating SKBG’s incremental value for investor portrait generation. These results confirm that SKBG contributes meaningful marginal value to investor clustering and user portrait generation without compromising cluster separation, validating the effectiveness of integrating social knowledge graph features.

Fig. 6.

Investor clustering results comparison. SKBG, Social Knowledge Big Graph.

From KO perspective, this scenario demonstrates how SKBG synthesizes knowledge associaction of categorical classification (investor type), behavioral traces (posting patterns), and statistical networks (influence scores) into a unified investor portrait. This moves beyond traditional siloed representations, offering a more holistic and semantically rich model for organizing investor knowledge.

This scenario moves beyond simple network diffusion analysis by leveraging the integrated knowledge associations to reconstruct not just how false information spreads, but why certain pathways are exploited and who benefits from the propagation dynamics. The approach transforms traditional rumor tracking into a sophisticated investigation of market manipulation patterns.

Specifically, we employ a multi-stage filtering algorithm that identifies suspicious content based on deviation from established patterns: posts with sentiment scores exceeding two standard deviations from the historical norm for a given stock, coordinated posting behaviors from known influencer clusters identified in Scenario 1, and information cascades that exhibit abnormal acceleration compared to typical diffusion patterns. The core innovation in this approach lies in the cross-modal validation that connects information propagation with trading behavior and portfolio positions. For each identified misinformation cascade, we execute additional queries to detect potential coordination patterns: we analyze whether amplifiers in the propagation network hold concentrated positions in the target stock (exploiting categorical associations from portfolio data), whether the timing of their amplification activity correlates with unusual trading volume (leveraging spatiotemporal associations), and whether the sentiment trajectory aligns with price manipulation strategies (utilizing behavioral associations). This multi-dimensional analysis transforms simple retweet chains into evidence of potentially coordinated activity, details shown as Algorithm B.2 in Appendix Table 11.

Table 7 shows the results validating the superiority of the SKBG-enhanced framework over the baseline model across three core evaluation metrics. For misinformation detection, the SKBG-enhanced model achieves an F1-score of 0.8159, representing a 12.77% marginal gain compared to the baseline’s 0.7235. This improvement stems from the model’s ability to integrate SKBG features like influencer identity, portfolio concentration, and propagation network structure rather than relying solely on surface-level content attributes such as sentiment deviation and posting frequency. By anchoring suspicious content to the behavioral and relational context of market participants, the framework reduces false positives caused by isolated abnormal posts and enhances the precision of identifying coordinated misinformation campaigns. In addition, the SKBG-enhanced approach delivers substantial gains in metrics directly tied to market manipulation investigation capabilities. In propagation path accuracy (evaluated on Top 20 amplifiers), the model achieves a score of 0.6890, marking a 52.43% improvement over the baseline’s 0.4520. This advancement is critical because traditional diffusion models often misidentify high-frequency posters as key amplifiers, whereas the SKBG framework prioritizes nodes with high PageRank influence and concentrated portfolio positions. For manipulation correlation, measured by the Adjusted Rand Index (ARI), the SKBG-enhanced model reaches 0.5870, a striking 86.94% marginal gain compared to the baseline’s 0.3140. This result underscores the framework’s unique strength in linking information propagation cascades to underlying trading coordination. The model effectively quantifies the association between information operations and market manipulation behaviors by cross-validating amplification timing with unusual trading volume and sentiment trajectories with price manipulation patterns.

This scenario exemplifies how SKBG transforms misinformation analysis into a knowledge integrity verification task. By integrating behavioral (propagation acts), categorical (actor profiles), and event-centric (market context) associations within a single semantic model, the framework enables cross-modal consistency checking. This moves beyond structural network analysis to assess whether information flows align with the ontological attributes and temporal contexts of involved entities. The approach embodies the KO principle of associative reasoning, demonstrating that semantic integration across association types is essential for detecting coordinated epistemic manipulation in socio-technical systems.

The third scenario escalates the analytical scope to macroscopic market phenomena, demonstrating how the SKBG enables stock market price comovement effect analysis and risk early warning by uncovering the latent social mechanisms that drive correlated price movements beyond fundamental economic factors. While traditional financial models explain comovement through industry classification, geographic location, or fundamental similarities, our approach reveals how investor behavior and information diffusion patterns captured in the SKBG create artificial linkages between otherwise unrelated securities, providing a novel framework for systemic risk assessment.

The methodology employs a multi-layered association analysis that systematically investigates different dimensions of stock interconnectedness. We established a baseline through traditional price correlation analysis, calculating pairwise correlation coefficients between all stocks in our universe based on their daily returns. We compute social proximity based on investor interaction intensity, information proximity from the co-mention patterns in social discourse, and behavioral proximity from the sentiment spillover and attention synchronization between investor communities. This multi-dimensional proximity framework is implemented through follow procedure (see Algorithm B.3 in Appendix Table 11).

The analytical power emerges through temporal predictive validation that tests whether SKBG-derived proximity measures lead observable price correlations. We implement a Granger causality framework that examines whether changes in social and information proximity between stock pairs statistically precede the emergence of significant price comovement. In this context, we do not focus on the causal relationships between factors influencing risk contagion, but only on temporal sequence. Therefore, Granger causality tests between asset returns can achieve the research objective and serve as an effective alternative to describe the sequence of risk transmission. Consequently, Granger causality can be used to establish the transmission direction between individual stocks, and a directed risk transmission network can be constructed accordingly, as shown in Fig. 7. It can be observed that compared to undirected networks, this network has significantly sparser edges, but its modularity is further enhanced, and the overall connectivity of the network deteriorates. We argue that socially-driven comovement created vulnerability channels that manifested during a market stress event. When negative news affected one stock in the cluster, the sell-off rapidly propagated through the social linkage network, affecting all connected stocks despite their disparate fundamentals. Most fundamentally, this scenario demonstrates that the SKBG provides a unique window into the social fabric of markets, revealing how investor behavior and information diffusion create latent structures that fundamentally reshape asset price dynamics and systemic risk propagation.

Fig. 7.

Directed risk transmission network generated based on Granger causality.

While the primary knowledge association patterns are derived from the Eastmoney forum data, a potential concern is whether findings based on a single online community are robust and reflect genuine market phenomena, rather than platform-specific artifacts. To verify the important role of social connections in identifying key transmission paths, this study establishes a Quadratic Assignment Procedure (QAP) regression model to confirm the impact of social connections on stock price linkages. QAP regression analysis considers that variables in social networks are usually not independent but interdependent, making it suitable for analyzing matrix data in relationship graphs. Its purpose is to assess the strength and significance of relationships between variables. Therefore, using QAP regression analysis to study the impact of stock price linkages arising from investor social interactions is feasible. This paper draws on the model construction approach of Chen et al. (2024), and the final model expression is as follows:

(3)$COMov=\alpha 0+\alpha 1 SC+\alpha 2 IND+\alpha 3 CIT+\alpha 4 PLA+\alpha 5 MAN+\alpha 6 SHA+\alpha 7 SUP+\alpha 8 DMV+\epsilon$

In this matrix, SC represents the social association matrix, IND represents the industry association matrix, CIT represents the regional association matrix, PLA represents the conceptual association matrix, MAN represents the personnel association matrix, SHA represents the equity association matrix, and SUP represents the supply chain relationship matrix. The elements of the IND, CIT, MAN, and SUP matrices take values of 1 or 0. The element values of the PLA matrix indicate the strength of the conceptual association, the element values of the SHA matrix indicate the shareholding ratio, and the DMV is the size difference matrix of the enterprises, which represents the degree of difference in size between enterprises.

If the social connections identified from the single forum source are merely noise or platform-specific, they should not systematically explain the co-movement of stock prices in the broader market. A statistically significant relationship, however, would provide strong evidence that the social patterns captured by our method transcend the specific data source and are linked to a fundamental market outcome. The regression results are shown in Table 8. It can be seen that at the 5% level, the social association matrix has a significant positive impact on the stock price correlation coefficient matrix, meaning that the social connections among investors can explain the correlation between stock prices.

Furthermore, to address potential concerns regarding data source limitations and to verify the stability of our primary finding that investor social associations explain stock price co-movement, we conducted robustness checks, with results summarized in Table 9. We performed 50 iterations of random sampling (70% of stocks in each iteration) to assess the stability of our model estimates against sample composition variations. We then conducted 100 permutations of the social association matrix to generate empirical p-values, thereby testing whether the observed relationships could arise by chance. The bootstrap analysis reveals exceptional stability of our core finding. The social association coefficient remained statistically significant (p 0.05) in 100% of the 50 resampling iterations. The placebo test further confirms the non-random nature of the observed relationship, with an empirical p-value of 0.000, substantially below conventional significance thresholds.

This analysis demonstrates SKBG’s ability to organize and explain systemic market phenomena through socially-derived associations. Beyond traditional categorical linkages (industry sectors), SKBG incorporates behavioral (sentiment spillover), statistical (investor overlap), and spatiotemporal (temporal lead-lag) associations to construct a multidimensional proximity network. This reveals how social and informational structures create latent risk channels not captured by conventional financial ontologies. The framework thus advances KO from static entity classification to dynamic mechanism modeling, providing explanatory power for complex market behaviors through integrated association reasoning.