Exploring new AI methodologies that enable value creation through knowledge discovery facilitates management decision-making. PHOTO: TUCHONG

In the context of the digital economy and innovation-driven development empowered by digital intelligence, research on management decision-making has expanded significantly in both breadth and depth. Increasingly, scholars and practitioners are stressing the exploration of new artificial intelligence methodologies in order to enable value creation through knowledge discovery, thereby injecting fresh momentum into management decision-making. Within this process, “relationships” constitute an important category of knowledge. Relevant literature often refers to this as relational knowledge, focusing on its representational forms and key characteristics, the intelligent processes through which it can be discovered or identified, and its application across diverse management decision-making scenarios.

Limitations of traditional approaches

The limitations of traditional research and relationship discovery methods have become increasingly apparent.

First, constrained by data accessibility, traditional studies rely on datasets that are limited in scale, granularity, scope, sources, types, and timeliness, making many relationships difficult to detect and represent.

Second, the forms of relationships explored in traditional research remain relatively narrow, making it difficult to meet the demands of management innovation, emerging business models, and new behavioral patterns that require the discovery and characterization of diverse, novel relationships.

Third, limited effectiveness in mapping methods—particularly in algorithm and model design, implementation strategies, and levels of intelligence—restrict the ability to capture and identify complex relationships. Novel relationships extend beyond the set of known relationships within a given problem context and can support more up-to-date and diversified marketing strategies. Accordingly, the development of intelligent methods for discovering such knowledge offers new value for management decision-making.

Novel relationships and their discovery

Conceptually, “novel relationships” can be understood as potential, unknown, and valuable relationships embedded within the high dimensional space of big data. Here, “novel” refers to either new forms of relationships or new instances of relationships; “unknown” indicates that these relationships are latent within data and require intelligent methods to uncover; “valuable” underscores their relevance for supporting management decision-making; and the “high dimensional space of big data” emphasizes the extraction of more abstract and complex semantics from observable datasets.

In practice, novel relationships may appear either as new instances of existing relationships or as entirely new relationship forms.

“Novel relationship discovery,” in turn, refers to the innovative design of intelligent methods based on big data analytics to identify such relationships for management decision-making. Research in this area emphasizes methodological innovation, as well as the significance and practical value of these methods in digital-intelligent management contexts. In knowledge-driven management decision-making, data are first collected and preprocessed based on an understanding of real-world management problems. Advanced analytical techniques are then developed to map data into knowledge, which is ultimately applied to support decision-making. This process involves three key dimensions: data, knowledge, and mapping.

Compared with traditional research on relationship discovery, the characteristics of novel relationship discovery can be summarized across these three dimensions and their interconnections.

In terms of data, observable big data collections increasingly display features such as macro-micro scale granularity, diverse internal and external sources, rich multimodality, and cross-domain and time varying dynamics. These characteristics enable intelligent methods to introduce new relationship objects or instances based on comprehensive, diverse, and timely datasets, while methodological innovation further facilitates the discovery of novel relationships and their value creation.

In terms of knowledge, novel relationships tend to be more diverse, encompassing direct and indirect, explicit and implicit, static and dynamic, and single- and cross-domain forms. These properties provide decision-makers with varied insights tailored to specific problems.

In terms of mapping, novel relationship discovery relies on innovation in intelligent methods. These approaches integrate algorithmic learning, statistical, and optimization techniques, while increasingly incorporating constructs from management and behavioral sciences according to specific problem contexts. In particular, novel relationships can be captured, modeled, and characterized from perspectives such as sequential learning, collaborative optimization, vector embedding, and latent variables.

As real-world decision-making scenarios become increasingly rich and complex, research on novel relationship discovery continues to deepen. Data usage has evolved from single-granularity, single-source, single-modality, and single-time-point datasets to integrated big data collections featuring macro-micro scale granularity, diverse sources, rich multimodality, and cross-domain and time-varying dynamics. Meanwhile, research focus has shifted from identifying direct, explicit, static, and single-domain relationships to uncovering knowledge concerning indirect, implicit, dynamic, and cross-domain relationships. On this basis, research designs new mapping paradigms through innovations in intelligent methods, particularly by constructing high-order spaces over observable big data to extract more abstract semantics and thereby discover desired novel relationships.

Two key challenges arise in this process.The first challenge lies in effectively integrating richer datasets and accurately representing the information embedded within them.

The second concerns how to precisely characterize more complex novel relationships and construct high dimensional spaces that bridge data and knowledge.

To address the first challenge, existing studies propose several approaches, including designing collaborative optimization methods to integrate datasets with macro-micro scale granularity, diverse internal and external sources, and rich multimodality, adopting sequential learning strategies to characterize cross-domain and time-varying dynamics, and modeling vector embeddings and latent variables to extract key information from complex and heterogeneous datasets.

To tackle the second challenge, effective approaches include developing intelligent methods based on vector embedding and latent variables to represent high dimensional semantics and uncover indirect or implicit relationships, applying sequential learning techniques to reveal dynamic relationships, constructing and applying collaborative optimization frameworks to abstract high dimensional spaces and identify cross-domain relationships.

When observable data exhibit macro-micro scale granularity, diverse sources, rich multimodality, and cross-domain and time-varying dynamics, and when target relationships are indirect, implicit, dynamic, and cross-domain, mapping design becomes much more complex, and typically requires the combined application of multiple methodological strategies.

Directions for novel relationship discovery

Seven representative research directions across different domains can be identified, reflecting current lines of exploration in novel relationship discovery.

The first concerns complementary or substitutive relationships among products.

This direction employs datasets characterized by macro-micro scale granularity, diverse internal and external sources, rich multimodality, and cross-domain and time-varying dynamics to identify direct, explicit, static, and single-domain novel relationships. One representative study constructs neural network models based on product embeddings to identify complementary or substitutive relationships among products.

The second focuses on product bundling relationships. Here, innovative mapping methods based on latent variables model users’ purchasing motivations to uncover bundling relationships among products and generate bundled recommendations.

The third involves discovering synergistic relationships among keywords. Another line of research uses rich and complex data to identify direct, explicit, static, and single-domain novel relationships by designing mapping methods based on collaborative optimization. For example, multi-objective constrained genetic algorithms have been proposed that incorporate consumer interests into product titles, thereby identifying synergistic relationships among promotional keywords in optimal titles.

The fourth direction examines competitive relationships among brands. This line of research typically relies on single-granularity, single-source, single-modality, and single-time-point data to uncover indirect, implicit, dynamic, and cross-domain novel relationships.

One representative study designs mapping methods based on collaborative optimization, constructing bipartite graph models from search log data to identify competitive relationships among brands and analyze the degree of competition.

The fifth explores mutual learning relationships among embeddings. This direction employs datasets characterized by macro-micro scale granularity, diverse sources, rich multimodality, and cross-domain and time-varying dynamics to discover indirect, implicit, dynamic, and cross-domain novel relationships. A representative study proposes mapping methods that integrate vector embedding and collaborative optimization, designing social collaborative mutual-learning approaches for product recommendation tasks.

The sixth direction identifies sequential relationships among products. A related study proposes mapping methods based on sequential learning and latent variables, using graph-based embedding smoothing to design sequential recommendation algorithms and uncover product ordering relationships from the perspective of user interactions.

The seventh direction investigates “psychology–interest–behavior” driving relationships among consumers. Another line of research uses rich and complex big data to discover indirect, implicit, dynamic, and cross-domain novel relationships. Specifically, dynamic Bayesian network models are constructed based on sequential learning and latent variables to characterize the driving relationships among consumers’ psychology, interests, and behaviors across multi-stage purchasing processes—drawing on the marketing funnel theory—thereby enabling dynamic product recommendations.

Collectively, these approaches demonstrate how novel relationship discovery, supported by advanced intelligent methods, can generate insights for business model innovation and offer valuable technical support for enterprises in areas such as digital-intelligent marketing, operations, and competitive strategy.

Future directions

Looking ahead, interpretability represents an important challenge. A key direction for future research is to explore interpretable technical methods that integrate closely with specific management scenarios, projecting high dimensional data spaces into management semantic spaces to form more transparent, interpretable mapping paradigms.Another important issue concerns the identification of bias relationships and the design of corresponding debiasing algorithms.

Biased algorithms can lead to biased human behavior, while biased user behavior in turn leads to biased algorithms, creating dynamic and cyclical bias relationships. As human-machine interaction becomes increasingly frequent and integrated, addressing these issues will become essential.

Wei Qiang is a professor from the School of Economics and Management at Tsinghua University. This article has been edited and excerpted from Management World, Issue 11, 2025.