Methods, apparatus, equipment, readable storage media, and computer program products for determining competitive relationships

Abstract

This application provides a method, apparatus, device, program product, and computer-readable storage medium for determining competitive relationships. The method includes: acquiring a set of competing units, the set of competing units comprising multiple competing units, wherein each competing unit is a media object competing for exposure opportunities; assigning multiple agents to the set of competing units, each agent corresponding one-to-one with a competing unit in the set; establishing competitive relationships for the multiple competing units in multiple rounds through the multiple agents, wherein each round of establishing competitive relationships corresponds to a competitive relationship graph; determining whether the competitive relationship graph of each round meets preset conditions; if the competitive relationship graph of a target round meets the preset conditions, then determining the competitive relationships of the multiple competing units based on the competitive relationship graph of the target round; thus improving the accuracy and practicality of the competitive relationship modeling results.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, readable storage medium, and computer program product for determining competition relationships. Background Technology

[0002] In e-commerce advertising, multiple advertising units compete for limited exposure and click resources within the same traffic pool, and their interactions can be considered a competitive relationship. This competition not only affects advertising effectiveness but also has a profound impact on bidding mechanisms, price distribution, and the overall platform ecosystem. Therefore, accurately modeling the competitive relationships among advertising units is crucial for optimizing advertising strategies. Current technologies typically use static merchant characteristics to classify competitive relationships among advertising units, but this lacks a dynamic portrayal of actual competitive behavior and fails to reflect the complex and constantly changing interactions between advertising units. Summary of the Invention

[0003] This application provides a method, apparatus, device, readable storage medium, and computer program product for determining competitive relationships, which can effectively cope with the complex and dynamically changing competitive relationships in the advertising bidding environment and improve the system's ability to model the competitive relationships of newly joined merchants or advertising units lacking historical data.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a method for determining a competitive relationship, the method comprising: Obtain a set of competing units, which includes multiple competing units, and the competing units are media objects that compete for exposure opportunities; Multiple agents are assigned to the set of competing units, and each of the multiple agents corresponds one-to-one with a competing unit in the set of competing units. Through the multiple intelligent agents, multiple rounds of competitive relationships are established for the multiple competing units, wherein each round of competitive relationship establishment corresponds to a competitive relationship graph; Determine whether the competition relationship diagram of each round meets the preset conditions. If it is determined that the competition relationship diagram of the target round meets the preset conditions, then determine the competition relationship of the multiple competing units based on the competition relationship diagram of the target round.

[0005] This application provides a device for determining competition relationships, including: The data acquisition module is used to acquire a set of competing units, which includes multiple competing units, and the competing units are media objects that compete for exposure opportunities. The data processing module is used to allocate multiple agents to the set of competing units, with each agent corresponding one-to-one with a competing unit in the set. Through these agents, multiple rounds of competition relationships are established between the competing units, with each round's establishment corresponding to a competition relationship graph. The module then determines whether the competition relationship graph for each round meets preset conditions. If the competition relationship graph for a target round meets the preset conditions, the competitive relationships between the competing units are determined based on the target round's competition relationship graph.

[0006] This application provides an electronic device, the electronic device comprising: Memory is used to store executable instructions or computer programs. When a processor executes computer-executable instructions or computer programs stored in the memory, it implements the competition relationship determination method provided in the embodiments of this application.

[0007] This application provides a computer-readable storage medium storing a computer program or computer-executable instructions, which, when executed by a processor, implements the competition relationship determination method provided in this application.

[0008] This application provides a computer program product, including a computer program or computer executable instructions. When the computer program or computer executable instructions are executed by a processor, they implement the competition relationship determination method provided in this application.

[0009] In the technical solution of this application embodiment, a set of competing units is obtained, which includes multiple competing units, each of which is a media object competing for exposure opportunities. Multiple agents are assigned to the set of competing units, with each agent corresponding one-to-one with a competing unit. Through these agents, multiple rounds of competitive relationships are established between the competing units, with each round's establishment corresponding to a competitive relationship graph. It is determined whether the competitive relationship graph for each round meets preset conditions. If the competitive relationship graph for the target round meets the preset conditions, the competitive relationships between the multiple competing units are determined based on the target round's competitive relationship graph. Thus, through a graph generation and optimization mechanism, and by utilizing agents to perform semantic modeling of node attributes and historical interaction data, dynamic modeling and continuous optimization of advertising competitive relationships are achieved. This overcomes the inaccurate characterization problem caused by existing technologies relying on static rule division, thereby improving the accuracy and practicality of competitive relationship modeling. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the system architecture for determining competition relationships provided in the embodiments of this application; Figure 2This is a schematic diagram of the server structure provided in an embodiment of this application; Figure 3 This is a first flowchart illustrating the competition relationship determination method provided in the embodiments of this application; Figure 4 This is a schematic diagram of the second process of the competition relationship determination method provided in the embodiments of this application.

[0011] Figure 5 This is a schematic diagram of the third process of the competition relationship determination method provided in the embodiments of this application.

[0012] Figure 6 This is a schematic diagram of the fourth process of the competition relationship determination method provided in the embodiments of this application.

[0013] Figure 7 This is a schematic diagram of the fifth process of the competition relationship determination method provided in the embodiments of this application.

[0014] Figure 8 This is a schematic diagram of the system principle of the competition relationship determination method provided in the embodiments of this application.

[0015] It should be noted that the terms "first" and "second" mentioned above are only used to distinguish between different options and do not represent the degree of superiority or inferiority of the options or their priority in the implementation process. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0018] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0019] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0020] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0021] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0022] In the current e-commerce advertising field, research on modeling the competitive relationships between advertising units is relatively limited. The mainstream industry approach is largely rule-based, generally assuming that advertising units with the same placement mode, product category, and merchant level are in competition. For example, some existing technologies simply categorize ads based on their category or merchant attributes, lacking in-depth modeling of actual competitive behavior and influencing factors, and have not yet developed a precise and dynamic method for depicting advertising competition relationships. Existing technologies primarily use direct classification based on static merchant characteristics (such as placement mode, product category, and merchant level) to characterize the competitive relationships between advertising units. However, this static rule-based classification method has significant drawbacks, including inaccurate characterization, slow response, and limited predictive ability.

[0023] This application provides a method, apparatus, device, readable storage medium, and computer program product for determining competitive relationships, which can achieve accurate identification and dynamic modeling of advertising competitive relationships, thereby improving the accuracy of competitive relationship modeling. The following describes exemplary applications of the electronic devices provided in this application. The electronic devices provided in this application can be implemented as various types of terminals such as laptops, tablets, desktop computers, set-top boxes, smartphones, smart speakers, smartwatches, smart TVs, and vehicle terminals, or they can be implemented as servers.

[0024] See Figure 1 , Figure 1 This is a schematic diagram of the system architecture of the competition relationship determination method provided in the embodiments of this application. Figure 1The system involves server 100, terminal device 200, and network 300. Terminal device 200 is connected to server 100 through network 300, which can be a wide area network (WAN), a local area network (LAN), or a combination of both.

[0025] In some embodiments, the competition relationship determination method provided in this application can be implemented collaboratively by a server and a terminal device. For example, the terminal device 200 sends a set of competing units to the server 100. The server 100 receives the set of competing units and, using the competition relationship determination method provided in this application, assigns multiple agents to the set of competing units, establishes multiple rounds of competition relationships for the multiple competing units, and obtains a competition relationship graph corresponding to each round of competition relationship establishment. Finally, the competition relationship graph of the target round that meets the preset conditions is sent to the terminal device 200. The terminal device 200 receives the competition relationship graph of the target round and determines the competition relationship of the multiple competing units based on the competition relationship graph of the target round.

[0026] In other embodiments, the competition relationship determination method provided in this application can be implemented independently by a terminal device. The terminal device 200 sends a competition relationship establishment request and a set of competing units to the server 100. The server 100 receives the competition relationship establishment request, allocates multiple agents to the set of competing units, and sends the model (agent) used for competition relationship determination provided in this application to the terminal device 200. The terminal device 200 receives the model sent by the server and downloads it locally. It then establishes multiple rounds of competition relationships for the multiple competing units using the model. Each round of competition relationship establishment corresponds to a competition relationship graph. The server determines whether the competition relationship graph for each round meets preset conditions. If the competition relationship graph for the target round meets the preset conditions, the competition relationship of the multiple competing units is determined based on the competition relationship graph for the target round.

[0027] Here, server 100 can be a single server. In this case, the competition relationship determination method provided in this application embodiment can be implemented by the same server. Server 100 can also be a cluster of servers. In the case that server 100 is a cluster of servers, the competition relationship determination method provided in this application embodiment can be implemented by different servers, and this application embodiment does not impose any limitations.

[0028] See Figure 2 , Figure 2 This is a schematic diagram of the server structure provided in an embodiment of this application. Figure 2The server 100 shown includes at least one processor 110, memory 130, and at least one network interface 120. The various components of server 100 are coupled together via a bus system 140. It is understood that the bus system 140 is used to implement communication between these components. In addition to a data bus, the bus system 140 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 140.

[0029] The processor 110 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0030] The memory 130 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, etc. The memory 130 may optionally include one or more storage devices physically located away from the processor 110.

[0031] The memory 130 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 130 described in this application embodiment is intended to include any suitable type of memory.

[0032] In some embodiments, memory 130 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0033] Operating system 131 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, for implementing various basic business functions and handling hardware-based tasks. The network communication module 132 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 120, exemplary network interfaces 120 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc. In some embodiments, the apparatus provided in this application can be implemented in software. Figure 2A race condition determination device 133 stored in memory 130 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: a data acquisition module 1331 and a data processing module 1332. These modules are logically linked and can therefore be arbitrarily combined or further separated according to their implemented functions. The functions of each module will be described below.

[0034] In some embodiments, the terminal device or server can implement the competition relationship determination method provided in this application by running various computer-executable instructions or computer programs. For example, computer-executable instructions can be microprogram-level commands, machine instructions, or software instructions. Computer programs can be native programs or software modules in an operating system; they can be native applications (APPs), i.e., programs that need to be installed in the operating system to run; or they can be applets that can be embedded in any APP, i.e., programs that only need to be downloaded to a browser environment to run. In summary, the aforementioned computer-executable instructions can be any form of instruction, and the aforementioned computer programs can be any form of application, module, or plugin.

[0035] In other embodiments, the apparatus provided in this application can be implemented in hardware. As an example, the apparatus provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the race relationship determination method provided in this application. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0036] The following will describe the method for determining competition relationships provided in this application, using the server as the execution subject, with exemplary applications and implementations of the server provided in the embodiments of this application. See also Figure 3 , Figure 3 This is a first flowchart illustrating the competition relationship determination method provided in this application embodiment, which will be combined with... Figure 3 The steps shown are explained.

[0037] Step 301: Obtain the set of competing units. The set of competing units includes multiple competing units, which are media objects that compete for exposure opportunities.

[0038] Here, competing entities are the main participants in the competitive relationship modeling. They refer to business entities that participate in the advertising bidding process and have mutual influence relationships, such as merchants placing advertisements on an e-commerce platform, product category combinations (e.g., shop*cid3), ad placement units, and advertisers. Competing entities include identifying information, which can consist of product stores and product categories at the third level, used to accurately distinguish different competing entities. For example, in an e-commerce platform, an advertising bidding request may include multiple competing entities, and these entities have mutual competitive relationships.

[0039] Here, the set of competing units can be the set of target merchants for whom competitive relationships need to be modeled or updated. In other words, the set of competing units can be the set K of competitive relationships to be updated, input at startup. For example, if an e-commerce platform has 100,000 stores, and each store has several third-level categories (cid3), then each shopcid3 combination can be considered a competing unit, and all shopcid3 combinations together constitute a large-scale set of competing units. The set of competing units can be constructed based on historical data mining, user behavior analysis, product attribute matching, etc.

[0040] In some embodiments, structured data of advertisers, including advertiser ID, category, and historical bidding records, are read from the database, and advertiser combinations with potential competitive relationships are selected according to rules or algorithms to form an initial set of competitive units.

[0041] In some embodiments, a set of competitor attributes is obtained. The set of competitor attributes N includes, but is not limited to, advertiser store name, user level / volume, merchant type, main SKU list, fee rate, advertising consumption, gross merchandise volume (GMV), and other related attributes.

[0042] In some embodiments, initial competitive relationship data is obtained, which may refer to the initial competitive relationship data of multiple competing units in a set of competing units. For example, preliminary competitive relationship data obtained using data mining methods includes co-occurrence data and competition intensity coefficients.

[0043] Here, co-occurrence can refer to competing entities appearing simultaneously in a single competition. For example, in the field of e-commerce advertising, co-occurrence can refer to multiple competing advertising entities simultaneously vying for the exposure and clicks of a particular ad within homogeneous traffic (i.e., the same group of users, the same exposure opportunity, and the same traffic pool).

[0044] Step 302: Assign multiple agents to the set of competing units, with each agent corresponding one-to-one with a competing unit in the set.

[0045] In this embodiment of the application, an independent intelligent agent is created, and an intelligent agent is assigned to each competing unit. The intelligent agent has independent decision-making ability and memory system.

[0046] In some embodiments, the agent performs semantic retrieval through Retrieval Augmentation (RAG). For example, the features of the set of competing units K are vectorized and stored in the FAISS vector database. For instance, potential competing relationships are pre-screened using a candidate edge pool; real competing relationships (CompetitionGraph) are obtained through a graph-based competition network; semantic representations of nodes and edges are stored in the database; semantic similarity is calculated through semantic retrieval for semantic matching of vector similarities; multi-strategy results are fused through a re-ranking module; and rich contextual information is provided to the Large Language Model (LLM) through context enhancement.

[0047] In some embodiments, the agent's memory system is initialized by establishing a long memory and a short memory, and storing historical interaction data in vector form to achieve efficient semantic retrieval and fast recall. The long memory stores historical interaction data and historical graph structures, while the short memory stores interaction data and graph structures from the current round. For example, historical interaction data, competition strategies, preference patterns, and other long-term memory data are vectorized to obtain vectorized long-term memory features, which are then stored in the long memory. Similarly, temporary information such as current round interaction information, query results, and decision context is vectorized to obtain short-term memory features, which are then stored in the short-term memory. This enables the agent to achieve efficient semantic retrieval and fast recall.

[0048] In some embodiments, the agent can configure and invoke the LLM model through the large model interface, fully leveraging the semantic understanding and multi-dimensional information integration capabilities of the LLM model to dynamically optimize competitive relationship modeling from multiple dimensions such as merchant attributes, historical competition data, and time. Especially for newly joined merchants, the LLM can infer their potential competitors, enabling automatic construction of competitive relationships even without historical data. For example, when modeling competitive relationships through the agent, the LLM is used to perform semantic modeling of node attributes and historical interaction data, enabling dynamic expansion and improvement of the competitive relationship graph; the agent simulates merchant competitive behavior, achieving automatic discovery and continuous optimization of competitive relationships. Specifically, an Agent is assigned to each merchant node. In each round of interaction, the Agent evaluates node activity using the LLM and selects nodes with similar attributes to update the competitive relationship. After multiple iterations, a comprehensive and dynamically optimized competitive relationship graph is finally generated.

[0049] Step 303: Through multiple intelligent agents, establish competitive relationships for multiple competing units in multiple rounds, wherein each round of competitive relationship establishment corresponds to a competitive relationship graph.

[0050] In this embodiment, the competition relationship graph is a graph structure where nodes represent competing entities and edges represent the competitive relationships between them. In each round, each agent generates new competitive relationships based on the existing competition relationship graph and its own historical memory, using a large language model to infer and update the graph structure. For example, in the first round, some competitive relationships may be initially established based on simple co-occurrence frequencies; in subsequent rounds, more complex semantic information (such as the similarity of competing entity attributes and the matching degree of marketing strategies) will be incorporated to further refine the competition relationship network. Through multiple iterations, the optimal competition relationship graph is gradually approximated, thereby achieving dynamic modeling of the advertising competition landscape.

[0051] Step 304: Determine whether the competition relationship diagram of each round meets the preset conditions. If it is determined that the competition relationship diagram of the target round meets the preset conditions, then determine the competition relationship of multiple competing units based on the competition relationship diagram of the target round.

[0052] In some embodiments, the preset conditions may include multiple factors, such as the completeness of the competition relationship graph (i.e., the relationships between all relevant competing units have been established), stability (i.e., the graph structure does not change much after multiple iterations), and rationality (i.e., the intensity of the competition relationship is consistent with the actual bidding behavior). When the competition relationship graph of a certain round meets the above conditions, the competition relationship graph of that round is used as the final competition relationship graph to guide subsequent adjustments to advertising bidding strategies and optimization of resource allocation.

[0053] For example, after each iteration, the agent uses LLM to determine whether the node edge connections in the competition graph are complete. If incomplete connections exist, the interactive iteration continues until all edges are connected or the preset iteration limit is reached. Combining the updated state of the competing unit set and the state of connected edges after the current iteration, the agent uses LLM to determine whether the node edge connections in the competing unit set are complete. If incomplete connections exist, the competition relationship establishment process continues until all edges are connected or the preset iteration limit is reached.

[0054] In some embodiments, to support zero-sample competitive relationship inference for competitors corresponding to newly joined advertisers, a memory mechanism based on Retrieval Augmentation (RAG) is introduced. The attribute information and competitive relationship records of historical competitors are vectorized and stored in the FAISS vector database. When a competitor corresponding to a newly joined advertiser lacks historical data, the agent corresponding to that competitor can use semantic retrieval technology to find competitors with similar attributes from the existing memory and infer the potential competitive relationships between these competitors. This memory mechanism not only improves the system's adaptability to new competitors but also enhances the generalization and robustness of the overall competitive relationship modeling.

[0055] In summary, the competitive relationship determination method provided in this application, by introducing a multi-agent mechanism and the semantic reasoning capabilities of a large language model, achieves accurate identification and dynamic modeling of advertising competitive relationships. Compared to traditional methods that rely on static rules or historical co-occurrence data, the competitive relationship determination method provided in this application can more comprehensively depict the complex competitive behaviors between competing units. Furthermore, the competitive relationship determination method provided in this application has good real-time adaptability and scalability, making it particularly suitable for practical application scenarios of large-scale e-commerce advertising bidding platforms.

[0056] In some embodiments, see Figure 4 , Figure 3 As shown in step 303, the following steps are also performed during the establishment of competitive relationships in each round.

[0057] Step 401: Select one competing unit from the set of competing units as the target competing unit.

[0058] Here, target competitive unit extraction can involve selecting several competitive units from the set of competitive units based on a sampling strategy. This sampling strategy includes, but is not limited to, time factors, business cycle time, and historical node activity. Specifically, the extraction of target competitive units can be based on certain strategies, such as random sampling, priority ranking based on activity level, or dynamic adjustment based on time-periodic changes. The number of target competitive units extracted in each round can be adjusted according to system load and iteration efficiency. For example, in the first round, highly active competitive units might be selected as target competitive units; while in subsequent rounds, more potential competitors will be introduced to gradually refine the competitive relationship graph. It should be noted that the number of target competitive units extracted in each round can be one or more.

[0059] Step 402: Add the target competing unit as a new node to the first competition relationship graph.

[0060] Here, the first competitive relationship graph refers to the graph structure used to represent the competitive relationships between various competing entities. Nodes in the first competitive relationship graph represent competing entities, and edges represent the competitive relationships between the entities represented by the nodes and edges. In each iteration, the currently extracted target competing entity is added as a new node to the first competitive relationship graph, thereby gradually constructing a complete competitive relationship network. Thus, by introducing target competing entities as new nodes into the first competitive relationship graph, dynamic expansion of the graph can be achieved, making the system modeling process more flexible and adaptable to the ever-changing market competition environment.

[0061] In some embodiments, the first competition relationship graph includes multiple nodes, each node representing a competing unit, and the connecting edges between different nodes represent the competition relationship between different nodes; the competition relationship of the target competing unit in the first competition relationship graph refers to the competition relationship between the target competing unit as a new node and other nodes in the first competition relationship graph.

[0062] Here, in the interaction and iteration process, each iteration extracts several competing units from the set of competing units to become the target competing units for the current round. The target competing units for the current round can be abstracted as the first node in a graph structure. The first node is defined as a basic unit in the graph structure during each iteration, representing a specific target competing unit (such as an advertiser, merchant, or placement unit). The first node contains attribute information related to the target competing unit, such as product category (cid3), shop identifier (shop), user level, and ad spend. This attribute information is used to describe the behavioral characteristics and potential competitive capabilities of the target competing unit during the bidding process.

[0063] Graph nodes corresponding to other competing entities that are in competition with the first node are defined as second nodes. These other competing entities also contain informational attributes related to the competing entities, such as product type, merchant level, and historical bidding behavior. The edges connecting the graph nodes represent whether a competitive relationship exists between two nodes and the strength of that edge relationship. Edge relationships can be unidirectional or bidirectional, and their numerical weights can reflect the intensity of competition. Specifically, the intensity of traffic competition between two competing entities can be quantified using a set competition intensity coefficient.

[0064] Step 403: Establish the competitive relationship of the target competitor in the first competitive relationship graph through the agent corresponding to the target competitor.

[0065] For example, the agent corresponding to each target competitor can recall candidate competitors using historical co-occurrence data, vector similarity retrieval, and LLM semantic reasoning to generate a set of candidate competitors for the current competitor. This set includes multiple candidate competitors. Then, in each round, the agent corresponding to the target competitor analyzes the competitive behavior and intensity between the target competitor and the set of candidate competitors, determines whether a competitive relationship exists between the target competitor and the candidate competitors in each round, and provides the basis for this judgment and a confidence score.

[0066] Step 404: Based on the competitive relationships of the target competitors in the first competitive relationship graph, update the first competitive relationship graph to obtain the competitive relationship graph for the current round; wherein, the first competitive relationship graph is either the default competitive relationship graph or the competitive relationship graph obtained in the previous round.

[0067] In some embodiments, based on the competitive relationship of the target competing unit in the first competitive relationship graph, a connecting edge corresponding to the target competing unit is added to the first competitive relationship graph to update the first competitive relationship graph; wherein, the connecting edge corresponding to the target competing unit represents the competitive relationship between the node corresponding to the target competing unit and other nodes in the first competitive relationship graph.

[0068] Here, after establishing the competitive relationships, the first competitive relationship graph is updated based on the newly added edge relationships to form the competitive relationship graph for the current round. The update process includes adding new edges to the graph structure and simultaneously updating the memory state and vector embedding of relevant nodes to reflect the latest competitive situation. By continuously updating the first competitive relationship graph, sensitivity to dynamic market changes can be maintained, and the expression of competitive relationships can be adjusted in a timely manner, improving the timeliness and accuracy of prediction and modeling.

[0069] As described above, in this embodiment, target competitors are extracted in rounds and added as new nodes to the first competitive relationship graph. Furthermore, an agent simulates competitive behavior and establishes corresponding competitive relationships. Subsequently, the entire graph structure is updated based on these newly added competitive relationships. This enables dynamic modeling of advertising competition relationships, allowing for a more accurate depiction of complex market competition patterns.

[0070] In some embodiments, during the establishment of competitive relationships in a round other than the first round, the first competitive relationship diagram is the competitive relationship diagram obtained in the previous round.

[0071] In some embodiments, the competitive relationship graph obtained in the previous round is obtained through the agent corresponding to the target competing unit in the current round.

[0072] Here, "previous round" refers to the previous iteration stage completed in the iterative competition graph construction process. In each round, new competition relationships are generated based on the current data and agent behavior, forming a new competition graph. Therefore, the competition graph obtained in the previous round is the competition relationship structure retained after the previous iteration, serving as the basis for the current round's construction.

[0073] In practice, there is a direct relationship between target competitors and agents. Each target competitor corresponds to an agent, which is responsible for actively acquiring the competitive relationship graph obtained from the previous round in the current round to support the updating and optimization of the current competitive relationship. On the one hand, by introducing the competitive relationship graph obtained from the previous round as the basis for the current round, continuous iteration and optimization of the competitive relationship can be achieved, avoiding modeling from scratch each time, thereby improving modeling efficiency and accuracy. On the other hand, using agents to automatically acquire historical graph structures helps maintain the consistency and stability of the model, especially in the face of complex and ever-changing advertising bidding environments, enabling rapid adaptation to market changes and improving prediction and optimization effects.

[0074] In some embodiments, during the establishment of competitive relationships in the first round, the first competitive relationship graph is the default competitive relationship graph.

[0075] For example, when the i-th round is the first round, before executing the interaction and iteration process, different graph initialization strategies are executed according to the task type to obtain the default competition graph P before iteration. 0 The task types include seed graph construction mode and graph update mode. When the task type is graph update mode, the first competitive relationship graph in the first round is obtained based on the historical competitive relationship graph, that is, by reading the historically constructed competitive relationship graph P. -1 P serves as the base graph for the 0th iteration. 0When the task type is seed graph construction mode, the first competition relationship graph in the first round is obtained based on historical co-occurrence data.

[0076] In some embodiments, a default competition graph is determined by the agent corresponding to the target competing unit in the first round, based on the historical competition graph and / or historical co-occurrence data.

[0077] For example, if the agent corresponding to the target competing unit in the first round is determined to have no historical competitive relationship graph, then the default competitive relationship graph is determined based on the historical co-occurrence data.

[0078] Here, historical co-occurrence data refers to data records of multiple competing entities jointly participating in the same exposure opportunity during the advertising bidding process. For example, in a single ad request (sid), multiple shopcid3 combinations may participate in the bidding simultaneously. This joint participation of multiple shopcid3 combinations reflects the potential competitive relationship between them. By statistically analyzing a large amount of historical co-occurrence data, we can identify which competing entities frequently appear in the same bidding scenarios, thereby inferring a potentially strong traffic competition relationship between multiple competing entities. Methods based on co-occurrence frequency can serve as the foundation for constructing a default competitive relationship graph, especially when direct competitive relationship records are lacking. By introducing historical co-occurrence data as an alternative information source, a reasonable default competitive relationship graph can still be constructed even without a historical competitive relationship graph, thereby improving the accuracy of competitive relationship modeling for new merchants or low-frequency participants.

[0079] For example, if the agent corresponding to the target competing unit in the first round is determined to have a historical competition relationship graph, then the historical competition relationship graph is determined to be the default competition relationship graph.

[0080] Here, the historical competition relationship graph refers to the competition relationship graph structure generated and continuously optimized by the system in past iterations. The historical competition relationship graph uses nodes to represent individual competing units and edges to represent the competitive relationships between them, and may include metadata such as competition intensity scores. When a reliable historical competition relationship graph exists for a target competing unit, it can be directly used as the default competition relationship graph for the current round without needing to be recalculated or constructed, thus saving computational resources and improving system response speed. By reusing existing high-quality historical competition relationship graphs, a new round of competition relationship modeling can be quickly initiated, avoiding unnecessary redundant calculations and thus improving overall operational efficiency.

[0081] Specifically, refer to Figure 5 If the first competition graph in the first round is obtained based on historical co-occurrence data, then the default competition graph is determined based on the historical co-occurrence data. This can be achieved through the following steps 501 to 505, which are explained in detail below: Step 501: Obtain the historical co-occurrence data of each competing unit in the set of competing units.

[0082] In some embodiments, obtaining historical co-occurrence data for each competitor in the competitor set can be achieved by analyzing the historical display data of the competitors. Historical display data comprises information about competitors that appeared during the ad bidding process, including the unique SID code of the ad bidding request and competitor identification information.

[0083] Reference Figure 6 Taking shop*cid3 as an example of a competing unit, firstly, shop*cid3 is used as the unique identifier of the competing unit to ensure accurate differentiation of different advertisers in subsequent data processing. The co-occurrence relationship between shop*cid3 can be obtained by acquiring historical display data. The co-occurrence relationship between shop*cid3 can be calculated at the display granularity; if two shop*cid3 appear in the same display, they are considered to be participating in one competition.

[0084] Step 502: Based on historical co-occurrence data, obtain the co-occurrence data for each competing unit in each exposure opportunity, and the total number of participations for each competing unit in all exposure opportunities.

[0085] Here, co-occurrence data for each exposure opportunity refers to the frequency of co-occurrence between each competing unit and other units in a single bidding request; while the total number of bids represents the total number of times a competing unit participates in bidding throughout the entire time period. Each ad competition can be identified by the unique code sid of the ad bidding request. In each ad request (using sid as the unique request identifier), the co-occurrence of all shop*cid3s is counted to obtain the traffic competition situation for advertisers. Then, the total number of times shop*cid3s participates in bidding over a past period is calculated, that is, for each competing unit, the total number of times it participates in bidding in all historical requests is counted.

[0086] Reference Figure 6 The co-occurrence data for each competing unit in each advertising competition can be obtained by calculating the number of times shop*cid3 co-occurs in a single sid. Obtained. Specifically, ,in, This indicates the number of times shop*cid3_1 and shop*cid3_2 co-occur in the k-th ad bidding request; This represents the number of times shop*cid3_1 co-occurs in the k-th ad bidding request; This represents the number of times shop*cid3_2 co-occurs in the k-th ad bidding request. For example, in a single sid k, if the bidding queue contains ads {sc1, sc1, sc1, sc2}, then the number of times (sc1, sc2) co-occurs is... =3, co-occurrence count of (sc2,sc1) =3.

[0087] Step 503: For each competing unit, calculate the competition intensity coefficient based on the co-occurrence data and total number of participations for each competing unit.

[0088] Here, the Competition Intensity Coefficient is a quantitative indicator used to measure the intensity of competition between two competing units, calculated using co-occurrence frequency and participation frequency. For example, for competing unit a and competing unit b, the Competition Intensity Coefficient of b over a is calculated. , can be defined as: = Co-occurrence counts a,b / Total number of competition counts a. This coefficient reflects the intensity of competition unit b in the traffic competition for competition unit a.

[0089] Step 504: Based on the competition intensity coefficient, screen the competitive relationships of each competitive unit to obtain the first competitive group. The first competitive group includes competitive units with a competition intensity coefficient greater than or equal to the first threshold.

[0090] Here, refer to Figure 6 Calculate the level of competition among each shop*cid3. It can be obtained through the following formula (1): , (1) For example, sc1 appears 10,000 times and sc2 appears 10 times. The co-occurrence count of (sc1, sc2) is... =3, co-occurrence count of (sc2,sc1) =3. sc1 appeared 10000 times, sc2 appeared 10 times, then... =3 / 10000, =3 / 10. Next, a first threshold, the 'thershold', is set. This first threshold is the competition intensity coefficient threshold. By filtering out competitive unit pairs whose competition intensity coefficient exceeds the threshold, the first competitive group is obtained. , The competing entities in the first competitive group were identified as having a strong competitive relationship.

[0091] Step 505: Based on the first competing group, obtain the default competition relationship diagram.

[0092] In this embodiment, to address the problem that traditional strategies and methods struggle to accurately depict competitive intensity, this approach analyzes advertisers' historical display data to uncover actual traffic competition relationships and dynamically constructs competitive groups. By acquiring historical co-occurrence data of competing units, calculating the competition intensity coefficient, and filtering out ad units with strong competitive relationships, a more accurate default competitive relationship graph can be constructed. By obtaining historical co-occurrence data of competing units, calculating the competition intensity coefficient, and filtering out ad units with strong competitive relationships, the actual competitive behavior between advertisers can be more accurately reflected, thereby optimizing advertising bidding strategies and ultimately improving the overall operational efficiency and user satisfaction of the advertising platform.

[0093] In some embodiments, see Figure 7 , Figure 4 Step 403, which shows the determination of the competitive relationship of the target competing units in each round based on the first graph structure of each round, can be achieved through the following steps 701 to 702.

[0094] Step 701: Using the agent corresponding to the target competing unit, perform candidate competing node recall processing on the target competing unit to obtain a set of candidate competing nodes.

[0095] Here, candidate competitor node recall refers to the process by which the agent corresponding to the target competitor actively searches for and filters other competitors that may have a competitive relationship with the target competitor, based on the target competitor's attribute characteristics, historical interaction data, and current environmental information. Candidate competitor node recall combines semantic retrieval (such as vector similarity matching) with rule filtering (such as co-occurrence frequency and merchant category matching) to generate a highly relevant list of candidate competitors. Candidate competitor node recall can employ LLM inference and RAG-enhanced retrieval to ensure high accuracy and coverage of the recall results.

[0096] The candidate competitor set refers to the final set of potential competitors identified after the aforementioned candidate competitor node recall process. Each node in the candidate competitor set may be an actual or potential competitor of the target competitor and can be used for subsequent judgment of competitive relationships and updating of the graph structure. The construction of the candidate competitor set is a crucial preliminary step in establishing a dynamic competitive relationship graph, and its construction directly affects the comprehensiveness and accuracy of competitive relationship identification.

[0097] By implementing a candidate competitor node recall process, other advertisers or ad placement units that may compete with the target unit can be identified more accurately, avoiding the omission of key competitors. Simultaneously, the construction of the candidate competitor node set provides a foundational input for subsequent competition relationship modeling, contributing to improved completeness and dynamic adaptability of the competition relationship graph.

[0098] In some embodiments, before performing candidate competing node recall processing on the target competing unit, the method further includes: Obtain the first competitive relationship data of the target competitor, including the first competition intensity coefficient; The agent corresponding to the target competing unit performs semantic retrieval to obtain the context data, node attributes, and historical interaction data of the nodes in the first competitive relationship graph.

[0099] Here, the first competitive relationship data refers to the historical posterior true competitive relationship data of the target competing units in the current round. The first competition intensity coefficient reflects the intensity of competition between competing units in the historical data. The first competition intensity coefficient can be calculated based on co-occurrence frequency and participation frequency, and generated through statistical or algorithmic models.

[0100] Semantic retrieval enables in-depth mining of complex semantic relationships. For example, for newly joined merchants, in the absence of direct historical co-occurrence data, potential competitors can be inferred through semantic parsing of the new merchants' attributes. Semantic parsing-based competitive analysis methods significantly improve the coverage and accuracy of competitive relationship modeling, and perform exceptionally well in zero-sample scenarios.

[0101] Contextual data includes environmental information for the current round, such as market trends, holiday activities, and changes in advertising strategies; node attributes refer to the static characteristics of competitors, such as merchant type, main category, and user base; historical interaction data reflects a competitor's behavioral trajectory over multiple rounds, such as the number of bids, success rate, and price distribution. All three types of data are obtained through semantic retrieval from the database by an intelligent agent.

[0102] In practice, the process of an agent retrieving data from a database using semantic retrieval technology can be achieved by calling the `search_advertisers` tool and searching the RAG (Registered Area of ​​Interest) based on the query content. Specifically, the agent first encodes the attributes of the target node, then uses efficient vector databases such as FAISS to perform similarity retrieval, finding other competing units that are semantically similar to the target node. Subsequently, the agent further verifies and optimizes the retrieval results based on a large language model combined with contextual data and historical interaction data, ensuring that the acquired information is accurate, comprehensive, and representative. This step of verifying and optimizing retrieval results by combining contextual data and historical interaction data not only improves data quality but also enhances the agent's ability to capture complex competitive relationships.

[0103] In some embodiments, the first node of each round undergoes a candidate competing node recall process, i.e., a Query process, which includes: Based on the first competition intensity coefficient, candidate node recall processing is performed on the target competing unit to obtain the first candidate competing node set; Based on contextual data, node attributes, and historical interaction data, semantic retrieval and vector similarity calculation are performed on the target competing unit to obtain the second candidate competing node set. The first set of candidate competing nodes and the second set of candidate competing nodes are merged to obtain the final set of candidate competing nodes.

[0104] Here, target competitors are initially screened based on a first competition intensity coefficient, prioritizing those with a higher coefficient as potential rivals. This allows for the rapid identification of potentially highly competitive businesses and the construction of a first set of candidate competitive nodes, providing a foundation for more refined matching in the future.

[0105] Next, a large language model is used to semantically understand and encode the aforementioned multidimensional data, transforming this information into high-dimensional vector form. Subsequently, vector similarity calculation is used to identify other nodes that are semantically close to the target competitor, forming a second set of candidate competing nodes. This process not only considers the direct bidding relationship between advertisers but also combines node attributes and historical interaction data to achieve a more comprehensive competitive relationship modeling. By using semantic retrieval and vector similarity calculation, potential competitors can be identified even in the absence of direct co-occurrence data. This capability is particularly suitable for newly registered merchants or data-sparse scenarios, thereby improving the breadth and accuracy of competitive relationship modeling.

[0106] Fusion processing refers to integrating candidate node sets from different recall strategies, removing duplicates, and ranking and filtering candidate nodes based on confidence scores to generate a final set of competing candidate nodes. Fusion processing typically employs a weighted fusion algorithm, which comprehensively considers factors such as the initial competition intensity coefficient, semantic similarity, and historical interaction frequency of each candidate node, thereby ensuring that the final set of competing candidate nodes is both highly relevant and diverse.

[0107] In this embodiment, an intelligent weight allocation mechanism assigns higher weights to nodes with high first-level competition intensity coefficients in the first candidate competitive node set, while appropriately reducing the weights of nodes with high semantic similarity in the second candidate competitive node set, thereby balancing the impact of short-term behavior and long-term trends. Through fusion processing, the advantages of multiple recall strategies can be effectively integrated, avoiding biases caused by a single dimension, improving the overall quality of candidate competitive nodes, and laying the foundation for the subsequent construction and optimization of the competitive relationship graph.

[0108] In some embodiments, a confidence mechanism is introduced, in which the agent labels the source of each node, for example, by marking it as recall based on a first competition intensity coefficient or recall based on semantic similarity, so that the system can refer to the confidence level in subsequent decision-making processes. This multi-path recall fusion mechanism can avoid the bias caused by a single recall strategy, and enhance the robustness and diversity of the results by the system.

[0109] In this embodiment, by fusing the results of multiple recall strategies, the advantages of various strategies can be combined to compensate for the shortcomings of a single strategy, thereby generating a more comprehensive and accurate set of candidate competing nodes, further improving the stability and generalization ability of competitive relationship modeling. In summary, by employing a multi-dimensional candidate node recall method based on a first competition intensity coefficient, semantic retrieval, and vector similarity calculation, combined with a fusion processing mechanism, the accuracy of competitive relationship modeling can be improved. This allows the system to optimize advertising bidding strategies and further enhance the overall resource allocation efficiency of the platform.

[0110] Step 702: Based on the set of candidate competing nodes, establish the competitive relationship of the target competing unit in the first competitive relationship graph to obtain the competitive relationship of the target competing unit in the first competitive relationship graph.

[0111] Here, based on the set of candidate competing nodes, a large language model is used to perform semantic reasoning on each candidate competing unit to determine whether a significant competitive relationship exists between each candidate competing unit and the target competing unit. The judgment criteria include, but are not limited to, multi-dimensional factors such as product category similarity, user group overlap, bidding behavior patterns, and historical co-occurrence frequency. For candidate competing units determined to have a competitive relationship, edge connections are established between them and the target competing unit, and the structure of the competitive relationship graph is updated synchronously. Furthermore, since the competitive relationship is bidirectional, the method also synchronously updates the perceptual memory of other competing units to achieve bidirectional confirmation of the competitive relationship.

[0112] In some embodiments, establishing the competitive relationship between the target competitor and the candidate competitor set in the first competitive relationship graph based on the candidate competitor set includes: obtaining the connection edge relationship, historical interaction data and node attributes of the target competitor through the agent corresponding to the target competitor; and establishing the competitive relationship between the target competitor and the candidate competitor set based on the connection edge relationship, historical interaction data and node attributes of the target competitor.

[0113] In this embodiment, an agent is assigned to each competing entity. The agent corresponding to the target competing entity is responsible for simulating and deciding on the target competing entity's behavior in the advertising bidding environment. The agent corresponding to the target competing entity can proactively acquire the target competing entity's historical interaction data, connection edge relationships (i.e., which competing entities it has known competitive relationships with), and node attributes (such as store name, product category, user level, etc.). Historical interaction data, connection edge relationships, and node attributes constitute the basis for the agent corresponding to the target competing entity to make subsequent competitive relationship judgments.

[0114] For example, in practical applications, when advertiser A's competitor is selected as the target competitor, the agent corresponding to the target competitor invokes the retrieval enhancement generation module to search relevant records in the vector database. This extracts key information such as advertiser A's target competitor's historical bidding behavior, historical interactions with other advertisers, and the node attributes of the target competitor. This data is used not only for modeling the current competitive relationship but also provides a reference for future dynamic adjustments. By introducing the agent mechanism, the system achieves autonomous perception and multi-dimensional data analysis capabilities of competitors, thereby improving the accuracy and real-time performance of competitive relationship modeling.

[0115] After acquiring the above three types of information, LLM is used to perform semantic understanding and reasoning on this data to determine whether there is a potential or actual competitive relationship between the target competing unit and the candidate competing nodes, and the competitive relationship graph is updated accordingly. This process not only relies on static historical interaction data, but also combines semantic similarity analysis of node attributes and the changing trends of historical interaction patterns, thereby achieving more refined competitive relationship identification.

[0116] For example, in a certain iteration, if the agent of advertiser A's target competitor finds that advertiser B's candidate competitor is highly similar to it in terms of product category, advertising strategy, and user group, then LLM will infer that there may be a strong competitive relationship between the two and add it to the competition relationship graph. Furthermore, the system will assess the confidence level of this competitive relationship and introduce a two-way confirmation mechanism when necessary to ensure the rationality of the competitive relationship. By combining LLM's semantic understanding capabilities with multi-dimensional data input, it can infer the competitive relationships of newly joined merchants even without historical data, while simultaneously achieving continuous optimization and dynamic updates to existing competitive relationships.

[0117] In this embodiment, by introducing an intelligent agent mechanism and combining it with a large language model to analyze and reason about multidimensional data, a more accurate competitive relationship graph can be established, thereby enabling the system to dynamically adapt to market changes and improve the efficiency and fairness of the advertising bidding system.

[0118] In some embodiments, the agent obtains connection edge relationships, historical interaction data, and node attributes from the database based on a random walk algorithm.

[0119] Here, the random walk algorithm is a graph traversal algorithm. It moves by randomly selecting adjacent nodes in the graph, and performs path sampling and feature extraction based on the first competitive relationship graph in each round, obtaining the connection edges and historical interaction information of the node in the current iteration. Due to the synergy between the agent and the random walk algorithm, it can not only identify existing competitive relationships but also predict new competitive relationships caused by changes in the external environment (such as holiday promotions or new merchants), thereby achieving dynamic modeling and continuous optimization.

[0120] In some embodiments, when a competitive relationship is determined between the current target competitor and the candidate competitor, a two-way perception mechanism is also implemented. When establishing a competitive relationship, the agents corresponding to the target competitor and the candidate competitor simultaneously update the perception memory of both agents, thereby achieving two-way confirmation of the competitive relationship.

[0121] In some embodiments, a first competition relationship graph is stored in a database of agents corresponding to the competing units; wherein the agents are used to update the database using the first competition relationship graph in each round.

[0122] Here, the agent sets up a memory reflection mechanism, which refines node memory through LLM-driven memory summarization and reflection, thereby reducing memory storage costs and improving retrieval efficiency.

[0123] The competitive relationship determination method provided in this application extracts target competing units in multiple rounds and constructs a graph structure. It then utilizes intelligent agents for semantic reasoning and competitive relationship judgment, achieving accurate identification and dynamic modeling of complex relationships between competing units. This competitive relationship determination method not only handles large-scale competitive unit relationship networks but also adapts to dynamic changes in the advertising bidding environment, thereby improving the intelligence level and operational efficiency of the advertising bidding system.

[0124] In some embodiments, when determining competitive relationships through an agent, it is achieved by the agent invoking a large language model. Semantic modeling of node attributes and historical interaction data is performed using an LLM (Local Language Model) to dynamically expand and improve the competitive relationship graph. The agent simulates merchant competitive behavior to achieve automatic discovery and continuous optimization of competitive relationships. Specifically, an Agent is assigned to each merchant node. In each round of interaction, the agent, based on its own attributes and memory stream (historical interaction records), updates the competitive relationship by selecting nodes with similar attributes based on node activity evaluated by the LLM. After multiple iterations, a comprehensive and dynamically optimized competitive relationship graph is finally generated.

[0125] In some embodiments, the large language model can be a general large language model.

[0126] In other embodiments, the large language model is a fine-tuned version of the large language model.

[0127] For example, the second competitive relationship data is obtained, and positive sample pairs and negative sample pairs are constructed based on the second competitive relationship data; the agent is fine-tuned based on the positive sample pairs and negative sample pairs to obtain the fine-tuned agent.

[0128] Here, we construct fine-tuning data by collecting and organizing real online advertising competition data based on a general pre-trained large language model. Using this data, we label the competitive relationships between ads as edges, constructing positive samples (pairs of ads with competitive relationships) and negative samples (pairs of ads without competitive relationships). Then, we perform fine-tuning training, using the labeled data to refine the large language model, enabling it to more accurately understand and analyze the competitive relationships between ads.

[0129] In some embodiments, after determining the competitive relationships of multiple competing units based on the competitive relationship graph of the target round, the method further includes: evaluating the competitive relationship graph of the target round based on a first model to obtain a first reward value; wherein the first model is a reward model trained using an expert feedback dataset, and the expert feedback dataset is obtained by evaluating the competitive relationships; and adjusting the agent based on the first reward value to obtain an adjusted agent.

[0130] Here, edge quality assessment is performed by human experts who label the quality and reasonableness of edges (i.e., competitive relationships) generated by the model, forming an expert feedback dataset. Next, a reward model is trained based on the expert feedback data to evaluate the quality of the edges generated by the model. This reward model can quantitatively score the reasonableness and accuracy of edges. The system also optimizes the edge generation strategy. During edge generation, the agent continuously calls the reward model to evaluate the generation results in real time and dynamically adjusts the edge generation strategy based on reward feedback. Through continuous iteration, RLHF closed-loop optimization is achieved, thereby improving the accuracy of new edge exploration and the model's ability to discriminate advertising competitive relationships.

[0131] The following describes an embodiment of the competition relationship modeling method provided in this application.

[0132] In the current e-commerce advertising field, multiple competing ad units exist for advertising campaigns targeting homogeneous traffic (i.e., the same group of users, the same exposure opportunities, and the same traffic pool). The interaction between these competing ad units can be viewed as a competitive group relationship. On mainstream advertising bidding platforms, as different ad units compete for exposure and clicks, their bidding behavior directly affects their traffic acquisition and price distribution, forming a complex volume-price interaction mechanism.

[0133] Currently, research on modeling competitive relationships among advertising units is relatively limited. The mainstream industry approach is largely rule-based, generally assuming that advertising units with the same placement mode, product category, and merchant level are in competition. For example, some existing technologies simply categorize ads based on their category or merchant attributes, lacking in-depth modeling of actual competitive behavior and influencing factors, and have not yet developed a precise and dynamic method for depicting advertising competitive relationships.

[0134] Existing technologies for characterizing competitive relationships among advertising entities primarily employ methods that directly segment based on static merchant characteristics (such as advertising patterns, product categories, and merchant ratings). However, this static rule-based segmentation method, which directly segments based on merchant static characteristics, has the following significant drawbacks: First, existing methods cannot accurately reflect the true and complex competitive relationships between advertising entities. In reality, the competitive relationships between advertising entities are highly dynamic, and this dynamism can change significantly due to various factors such as time periods, holiday events, advertising materials, placement costs, and marketing strategies.

[0135] Secondly, traditional methods lack a deep semantic understanding of merchant attributes and competitive behavior. They can only perform coarse-grained classification based on surface features, making it difficult to capture potential competitive relationships and complex interaction patterns among merchants.

[0136] Furthermore, because existing methods rely solely on static features and lack effective modeling of the aforementioned dynamic influencing factors, they struggle to adapt to real-time changes in competitive relationships within the advertising bidding environment. When market conditions or advertising strategies are adjusted, existing technologies cannot promptly identify and predict potential new or reduced competitive relationships between advertising units.

[0137] Furthermore, existing technologies primarily rely on historical co-occurrence data to mine competitive relationships. For newly joined merchants or advertising units lacking historical data, existing technologies cannot effectively construct competitive relationship models for these advertising units.

[0138] Therefore, existing technologies have significant shortcomings in dealing with the complex and dynamic competitive relationships in advertising bidding platforms, such as inaccurate characterization, untimely response, and limited predictive capabilities.

[0139] To address the aforementioned issues and achieve accurate and dynamic modeling of advertising competition relationships, this paper proposes a competition relationship modeling method based on the application of a large language model, a multi-agent collaboration mechanism, and a memory enhancement system. This improves the accuracy and timeliness of advertising competition relationship identification and supports competition relationship reasoning in zero-shot scenarios. See also Figure 8 , Figure 8 This is a schematic diagram of the system principle of a method for determining competitive relationships provided in an embodiment of this application. As shown in the figure, the competitive relationship modeling system includes a system startup layer 801, an execution engine layer 802, an agent interaction execution layer 803, an LLM reasoning process 804, a RAG retrieval strategy 805, and a decision node 806.

[0140] In some embodiments, the system startup layer 801 is used to start the system, start the distributed service, and initialize AgentScope; it is also used to load task configuration and LLM interface configuration.

[0141] Here, after the system startup layer 801 starts the system startup main.py, starts the distributed Launcher service, and initializes AgentScope, it interacts with the execution engine layer 802 by creating an Executor execution engine. Specifically, after the system startup layer 801 loads the task configuration config.yaml and the LLM interface configuration first model, it interacts with the execution engine layer 802 by creating ManagerAgent data.

[0142] In some embodiments, the execution engine layer 802 is used to create an execution engine, initialize the environment, create ManagerAgent data, retrieve and manage data, and simulate loop control.

[0143] Here, the execution engine layer 802 interacts with the system startup layer 801 to create the Executor execution engine, initialize the GeneralEnvironment, and execute the simulation loop control executor.run(). It also configures the first model to create ManagerAgent data and manage retrieval through the LLM interface of the system startup layer 801. The LLM model called can vary depending on the configuration; for example, it can call an external large language model via API or a finely tuned large language model after training.

[0144] In some embodiments, the Agent interaction execution layer 803 includes: an Agent interaction layer, used to update the graph structure, obtain active Agents, initialize Agent instance creation, update the interaction result graph structure and memory, and save simulation data trajectories and graph data.

[0145] In some embodiments, the Agent interaction execution layer 804 further includes Agent interaction execution, which is used for Agent interaction execution, context preparation node information and memory, Query phase LLM inference query, tool invocation, Action phase LLM inference decision, trajectory recording and saving interaction results.

[0146] Here, the Agent interaction layer constructs the first graph structure for each round using the `update_graph()` function, obtains the `get_active_agent_ids` of the active Agent, initializes the `GeneralAgent` instance, and then interacts with Agent Interaction Execution 804. After the Agent Interaction Execution obtains and saves the trajectory recording interaction results, the Agent interaction layer updates the interaction result graph structure, memory, and saves the simulation data trajectory and graph data. Agent Interaction Execution interacts with the Agent interaction layer, performing Agent Interaction Execution `interact()`, preparing context node information and memory; in the Query phase, it calls the LLM inference process 804 to perform LLM inference query and uses the tool `search_advertisers` to retrieve relevant content from the RAG retrieval strategy 805; in the Action phase, it calls the LLM inference process 804 to perform LLM inference decision-making, and then saves the trajectory recording interaction results.

[0147] In some embodiments, the LLM inference process 804 is used to perform LLM interface calls, Prompt templates, LLM inference process competition analysis, output parsing JSON formatting, and memory updates of long short-term memory.

[0148] Here, the Query and Action phases of the Agent Interaction Execution Layer 803 call the LLM API or the fine-tuned LLM model through the LLM interface, obtain the Prompt template acg_query / action, perform LLM inference process competition analysis, output JSON formatting parsing, update the long short-term memory, and send the updated long short-term memory to the Agent Interaction Execution Layer 703 for use by the Agent Interaction Execution Layer 803 in the Action process.

[0149] In some embodiments, the RAG retrieval strategy 805 is used for candidate edge pooling, real competitive relationships, vector repository, semantic retrieval semantic similarity calculation, reordering module, and context enhancement.

[0150] Here, the Candidate Pool is used to pre-screen potential competitive relationships, the Competition Graph is a graph-based competitive network, the Vector Database is used to store the semantic representations of nodes and edges, the Semantic Retrieval and Semantic Similarity Calculation is used to perform semantic matching of vector similarity, the Re-ranking module is used to perform multi-strategy result fusion, and the Context Enhancement module is used to provide rich contextual information for LLM.

[0151] For example, the system creates an independent intelligent agent instance for each competing unit through the system startup layer 801 and the execution engine layer 802, and endows the intelligent agent instance with independent decision-making capabilities and a memory system for simulating and inferring the relationship between competing units and other competing units. Furthermore, the system integrates the RAG retrieval strategy 805 to perform vectorization processing on the attribute information of all competing units, and stores these vector data in the FAISS vector database to construct an efficient semantic retrieval system, thereby supporting similarity matching and context-aware functions in subsequent tasks.

[0152] The system initializes the Agent memory system through the system startup layer 801 and the execution engine layer 802, and configures a long-term memory area and a short-term memory area for each Agent to store historical interaction records and the decision context of the current round, respectively, so as to quickly call relevant information during the process of judging competitive relationships.

[0153] Here, the system updates the graph structure via the Agent interaction layer 803 (update_graph()) to construct or retrieve the seed graph. Specifically, different graph initialization strategies are selected based on the task type. In seed graph construction mode, the system constructs an initial competition graph based on posterior co-occurrence relationships; in graph update mode, the system reads an existing competition graph as the initial structure. Simultaneously, all node and edge information is synchronously loaded into the RAG memory system to ensure the Agent has complete context awareness capabilities.

[0154] Here, the system expands the active node library through the Agent interaction layer and Agent interaction execution. Specifically, in each iteration, the system intelligently extracts several nodes from the set of competing units to be updated and merges them into the current graph. The sampling strategy comprehensively considers time factors, business cycle, and historical node activity to improve iteration efficiency. Secondly, the system performs a Query process. The Agent recalls candidate nodes, and the system sorts them in descending order according to the competition intensity coefficient based on historical competition relationship data, recalling the top n nodes. The system can also interact with the LLM inference process 804. The execution module recalls the top n nodes from the set through LLM semantic understanding and vector similarity calculation. The system then fuses the results of the multi-path recall to generate the final set of candidate competing nodes.

[0155] Subsequently, the system interacts with the LLM inference process 804 through Agent interaction execution 803, completing the competition relationship establishment Action process. The Agent performs the following operations: the system extracts connection edges and historical interaction information from its own memory; the LLM analyzes whether there is a competition relationship between nodes based on the candidate set generated in the Query phase, and outputs the judgment criteria and confidence score; after establishing the competition relationship, the two Agents synchronously update their own memory content, thereby realizing bidirectional perception and confirmation between the two Agents.

[0156] The system updates the graph structure through Agent interaction layer 803. The system performs the following operations: the system adds newly established competitive relationship edges to the graph incrementally; the system updates the node memory bidirectionally and further confirms the competitive relationships between nodes; through an LLM-driven memory reflection mechanism, the system refines the node memory content, thereby reducing storage burden and improving retrieval efficiency.

[0157] Finally, the system uses decision node 806 to determine whether to stop the iteration. Combining the node states and edge connectivity information after the current iteration, the system performs an LLM evaluation to determine if the convergence condition has been met. If the nodes are not fully connected, the system continues the graph update operation; otherwise, the system terminates the iteration process.

[0158] The above constitutes the competitive relationship modeling and optimization process based on LLM and multi-agent mechanisms in this application. Simultaneously, the system also introduces a memory enhancement mechanism based on RAG, further improving the historical experience reuse capability and semantic matching accuracy of the LLM and multi-agent mechanism-based model. Furthermore, to enhance the model's discriminative ability, a reinforcement learning (RLHF) training mechanism based on human feedback is introduced. Researchers conduct manual quality assessments of the competitive relationships generated by the model and optimize the reward model and edge generation strategy based on the results of the manual quality assessment, forming a closed-loop optimization process to continuously improve model performance.

[0159] In summary, this application, by integrating several key technologies such as large language models, multi-agent collaboration, RAG memory enhancement, and RLHF training, achieves accurate identification, dynamic modeling, and efficient prediction of advertising competition relationships. It overcomes the shortcomings of existing technologies in terms of inaccurate characterization and untimely response, demonstrating promising application prospects and significant value for technology promotion. Specifically, First, the competitive relationship determination method provided in this application utilizes the semantic understanding and information integration capabilities of a large language model (LLM) to dynamically and comprehensively model and optimize the competitive relationships between merchants. Unlike traditional methods that rely on historical data mining, this method combines multi-dimensional information such as merchant attributes, historical competition records, and time to accurately identify and predict real-time competitive relationships between merchants. Specifically, LLM performs semantic encoding and feature extraction on merchant attributes and historical data, leveraging its powerful semantic reasoning capabilities to dynamically generate and optimize the competitive relationship structure, achieving accurate characterization and real-time updates of competitive relationships. Particularly, for newly joined merchants, the system can infer potential competitive relationships through semantic similarity analysis even without historical display data. Thus, it supports zero-sample competitive relationship reasoning, possesses powerful semantic understanding and generalization capabilities, and enables dynamic real-time updates of competitive relationships.

[0160] Second, the competitive relationship determination method provided in this application abstracts merchants and their competitive relationships into a graph structure through a multi-agent collaborative competitive relationship graph iterative optimization mechanism. Using an LLM-driven multi-agent mechanism, it simulates the interactive behavior of merchants in the environment, gradually improving the competitive relationship graph. Each merchant node acts as an intelligent agent, and based on its own attributes and historical interaction data, the LLM determines its activity level and interaction objects, achieving dynamic iteration and expansion of competitive relationships. Specifically, in each iteration, the LLM evaluates the activity of nodes based on node attributes and memory flow, retrieves relevant historical experience based on vector similarity, filters potential competitors through attribute similarity, and dynamically generates or optimizes edges between nodes. This mechanism can effectively simulate changes in competitive relationships in real-world scenarios, continuously improving the completeness and accuracy of the competitive relationship graph. Thus, it achieves distributed multi-agent parallel processing capabilities, an adaptive graph structure optimization algorithm, support for large-scale merchant network modeling, and possesses good scalability and fault tolerance.

[0161] Third, the competitive relationship determination method provided in this application utilizes RAG-based intelligent memory and retrieval enhancement technology for advertising competition groups. It applies RAG (Retrieval Enhanced Generation) technology to model the relationships between advertising competition groups, constructing a long-short-term memory combined with intelligent memory system for advertising bidding scenarios. By storing semantic embeddings of advertising unit attributes, bidding behavior, and historical interactions of competition groups in a FAISS vector database, and combining this with a random walk algorithm for competitive relationship memory retrieval, it achieves efficient utilization of historical experience and semantic matching for dynamic changes in advertising competition groups. Specifically, the system vectorizes merchant attributes and historical interaction data and stores them in a high-dimensional semantic space. During the decision-making process, the agent quickly obtains relevant historical experience through semantic retrieval and makes inferences based on the current context. A memory reflection mechanism periodically summarizes and refines the stored information, maintaining the efficiency of the memory system. Thus, it achieves efficient semantic retrieval and matching capabilities, supports rapid querying of large-scale vector data, possesses a memory self-optimization and refinement mechanism, and enables intelligent reuse of historical experience.

[0162] The following description continues to illustrate the exemplary structure of the competition relationship determination device 133 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2 As shown, the software module stored in the contention determination device 133 of the memory 130 may include: The data acquisition module 1331 is used to acquire a set of competing units, which includes multiple competing units, and the competing units are media objects that compete for exposure opportunities.

[0163] The data processing module 1332 is used to allocate multiple agents to the set of competing units, wherein each agent corresponds one-to-one with a competing unit in the set of competing units; through the multiple agents, multiple rounds of competition relationships are established for the multiple competing units, wherein each round of competition relationship establishment corresponds to a competition relationship graph; it is determined whether the competition relationship graph of each round meets preset conditions; if it is determined that the competition relationship graph of the target round meets the preset conditions, the competition relationship of the multiple competing units is determined based on the competition relationship graph of the target round.

[0164] In some embodiments, the data processing module 1332 is further configured to: extract a competitive unit from the set of competitive units as a target competitive unit during the competitive relationship establishment process in each round; add the target competitive unit as a new node to the first competitive relationship graph; establish the competitive relationship of the target competitive unit in the first competitive relationship graph through the agent corresponding to the target competitive unit; update the first competitive relationship graph based on the competitive relationship of the target competitive unit in the first competitive relationship graph to obtain the competitive relationship graph of the current round; wherein, the first competitive relationship graph is a default competitive relationship graph or a competitive relationship graph obtained in the previous round.

[0165] In some embodiments, during the establishment of competition relationships in the first round, the first competition relationship graph is the default competition relationship graph.

[0166] In some embodiments, the data processing module 1332 is further configured to determine the default competition relationship graph based on the historical competition relationship graph and / or historical co-occurrence data of the agent corresponding to the target competing unit in the first round.

[0167] In some embodiments, during the establishment of competitive relationships in a round other than the first round, the first competitive relationship diagram is the competitive relationship diagram obtained in the previous round.

[0168] In some embodiments, the data processing module 1332 is further configured to obtain the competition relationship graph obtained in the previous round through the agent corresponding to the target competing unit in the current round.

[0169] In some embodiments, the first competition graph includes multiple nodes, each node representing a competing unit, and the connecting edges between different nodes represent the competition relationships between different nodes; the competition relationship of the target competing unit in the first competition graph refers to the competition relationship between the target competing unit as a new node and other nodes in the first competition graph.

[0170] In some embodiments, the data processing module 1332 is further configured to add connecting edges corresponding to the target competing unit to the first competing relationship graph based on the competing relationship of the target competing unit in the first competing relationship graph, so as to update the first competing relationship graph; wherein, the connecting edges corresponding to the target competing unit represent the competing relationship between the node corresponding to the target competing unit and other nodes in the first competing relationship graph.

[0171] In some embodiments, the data processing module 1332 is further configured to determine the default competition relationship graph based on the historical co-occurrence data if the agent corresponding to the target competing unit in the first round determines that the historical competition relationship graph does not exist; or, if the agent corresponding to the target competing unit in the first round determines that the historical competition relationship graph exists, determine the historical competition relationship graph as the default competition relationship graph.

[0172] In some embodiments, the data acquisition module 1331 is further configured to acquire historical co-occurrence data of each competing unit in the set of competing units.

[0173] In some embodiments, the data processing module 1332 is further configured to: obtain co-occurrence data corresponding to each of the competing units in each exposure opportunity and the total number of participations corresponding to each of the competing units in all exposure opportunities based on the historical co-occurrence data; calculate the competition intensity coefficient of each of the competing units based on the co-occurrence data and the total number of participations for each of the competing units; perform competitive relationship filtering on each of the competing units based on the competition intensity coefficient to obtain a first competitive group, wherein the first competitive group includes competing units whose competition intensity coefficient is greater than or equal to a first threshold; and obtain the default competitive relationship graph based on the first competitive group.

[0174] In some embodiments, the data processing module 1332 is further configured to perform candidate competition node recall processing on the target competing unit through the agent corresponding to the target competing unit to obtain a set of candidate competition nodes; and establish the competition relationship of the target competing unit in the first competition relationship graph based on the set of candidate competition nodes to obtain the competition relationship of the target competing unit in the first competition relationship graph.

[0175] In some embodiments, the data acquisition module 1331 is further configured to acquire first competitive relationship data of the target competing unit, the first competitive relationship data including a first competition intensity coefficient; the agent corresponding to the target competing unit performs semantic retrieval to acquire context data, node attributes and historical interaction data of nodes in the first competitive relationship graph.

[0176] In some embodiments, the data processing module 1332 is further configured to perform candidate node recall processing on the target competing unit based on the first competition intensity coefficient to obtain a first candidate competing node set; perform semantic retrieval and vector similarity calculation on the target competing unit based on the context data, the node attributes and the historical interaction data to obtain a second candidate competing node set; and perform fusion processing on the first candidate competing node set and the second candidate competing node set to obtain the candidate competing node set.

[0177] In some embodiments, the data acquisition module 1331 is further configured to acquire the connection edge relationships, historical interaction data, and node attributes of the target competing unit through the agent corresponding to the target competing unit.

[0178] In some embodiments, the data processing module 1332 is further configured to establish a competitive relationship between the target competing unit and the set of candidate competing nodes based on the connection edge relationship of the target competing unit, the historical interaction data, and the node attributes.

[0179] In some embodiments, the data acquisition module 1331 is further configured to acquire second competitive relationship data and construct positive sample pairs and negative sample pairs based on the second competitive relationship data.

[0180] In some embodiments, the data processing module 1332 is further configured to adjust the agent based on the positive sample pairs and the negative samples to obtain an adjusted agent.

[0181] In some embodiments, the data processing module 1332 is further configured to evaluate the competition relationship graph of the target round based on the first model to obtain a first reward value; wherein the first model is a reward model trained through an expert feedback dataset, and the expert feedback dataset is obtained by evaluating the competition relationship; and to adjust the agent based on the first reward value to obtain the adjusted agent.

[0182] This application provides a computer program product including a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the competition relationship determination method described above in this application.

[0183] This application provides a computer-readable storage medium storing computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the processor will execute the competition relationship modeling method provided in this application. For example, ... Figure 3 The method for determining competitive relationships is shown.

[0184] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0185] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0186] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0187] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0188] In summary, this application provides a method for determining intelligent competitive relationships based on a large language model (LLM). Leveraging the semantic understanding and multi-dimensional information integration capabilities of LLM, it achieves accurate identification and dynamic modeling of advertising competitive relationships. A distributed simulation system with multi-agent collaboration is constructed, using intelligent agents to simulate merchant competitive behavior, thereby enabling automatic discovery and continuous optimization of competitive relationships.

[0189] A memory mechanism based on RAG (Retrieval Augmentation) is established, combining vector retrieval technology and historical interaction data to improve the accuracy and timeliness of competitive relationship prediction. Zero-sample competitive relationship inference for newly joined merchants is achieved, and a competitive relationship model is built for advertising units lacking historical data through semantic analysis of merchant attributes.

[0190] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A method for determining competitive relationships, characterized in that, The method includes: Obtain a set of competing units, which includes multiple competing units, and the competing units are media objects that compete for exposure opportunities; Multiple agents are assigned to the set of competing units, and each of the multiple agents corresponds one-to-one with a competing unit in the set of competing units. Through the multiple intelligent agents, multiple rounds of competitive relationships are established for the multiple competing units, wherein each round of competitive relationship establishment corresponds to a competitive relationship graph; Determine whether the competition relationship diagram of each round meets the preset conditions. If it is determined that the competition relationship diagram of the target round meets the preset conditions, then determine the competition relationship of the multiple competing units based on the competition relationship diagram of the target round.

2. The method according to claim 1, characterized in that, The step of establishing multi-round competitive relationships among the multiple competing units through the multiple intelligent agents includes: During the establishment of competitive relationships in each round: Select one competing unit from the set of competing units as the target competing unit; Add the target competing unit as a new node to the first competition relationship graph; The competitive relationship of the target competing unit in the first competitive relationship graph is established by the agent corresponding to the target competing unit; Based on the competitive relationship of the target competing unit in the first competitive relationship graph, the first competitive relationship graph is updated to obtain the competitive relationship graph for the current round; The first competition graph is either the default competition graph or the competition graph obtained in the previous round.

3. The method according to claim 2, characterized in that, In the process of establishing competitive relationships in the first round, the first competitive relationship graph is the default competitive relationship graph; the method further includes: The default competition graph is determined by the agent corresponding to the target competing unit in the first round, based on the historical competition graph and / or historical co-occurrence data.

4. The method according to claim 2, characterized in that, In the process of establishing competitive relationships in rounds other than the first round, the first competitive relationship graph is the competitive relationship graph obtained in the previous round; the method further includes: Obtain the competition relationship graph from the previous round by using the agent corresponding to the target competitor in the current round.

5. The method according to claim 2, characterized in that, The first competition relationship graph includes multiple nodes, each node represents a competing unit, and the connecting edges between different nodes represent the competition relationship between different nodes; the competition relationship of the target competing unit in the first competition relationship graph refers to the competition relationship between the target competing unit as a new node and other nodes in the first competition relationship graph; The step of updating the first competition relationship graph based on the competitive relationship of the target competitor in the first competition relationship graph includes: Based on the competitive relationship of the target competitor in the first competitive relationship graph, add the connection edge corresponding to the target competitor in the first competitive relationship graph to update the first competitive relationship graph; The connecting edge corresponding to the target competing unit represents the competitive relationship between the node corresponding to the target competing unit and other nodes in the first competitive relationship graph.

6. The method according to claim 3, characterized in that, The step of determining the default competition graph based on the historical competition graph and / or historical co-occurrence data by the agent corresponding to the target competing unit in the first round includes: If the agent corresponding to the target competing unit in the first round determines that the historical competition relationship graph does not exist, then the default competition relationship graph is determined based on the historical co-occurrence data; or... If the agent corresponding to the target competing unit in the first round determines that the historical competition relationship graph exists, then the historical competition relationship graph is determined to be the default competition relationship graph.

7. The method according to claim 6, characterized in that, Determining the default competition graph based on the historical co-occurrence data includes: Obtain the historical co-occurrence data for each competing unit in the set of competing units; Based on the historical co-occurrence data, co-occurrence data for each of the competing units in each exposure opportunity is obtained, as well as the total number of participations for each of the competing units in all exposure opportunities; For each of the competing units, the competition intensity coefficient for each of the competing units is calculated based on the co-occurrence data and the total number of participations. Based on the competition intensity coefficient, the competitive relationship of each competitive unit is filtered to obtain a first competitive group, which includes competitive units whose competition intensity coefficient is greater than or equal to a first threshold. Based on the first competing group, the default competition relationship graph is obtained.

8. The method according to claim 2, characterized in that, The step of establishing the competitive relationship of the target competing unit in the first competitive relationship graph through the agent corresponding to the target competing unit includes: Using the intelligent agent corresponding to the target competing unit, a candidate competing node recall process is performed on the target competing unit to obtain a set of candidate competing nodes; Based on the set of candidate competing nodes, the competitive relationship of the target competing unit in the first competitive relationship graph is established, thereby obtaining the competitive relationship of the target competing unit in the first competitive relationship graph.

9. The method according to claim 8, characterized in that, Before performing candidate competing node recall processing on the target competing unit, the method further includes: Obtain first competitive relationship data of the target competing unit, wherein the first competitive relationship data includes a first competition intensity coefficient; The agent corresponding to the target competing unit performs semantic retrieval to obtain the context data, node attributes, and historical interaction data of the nodes in the first competition relationship graph.

10. The method according to claim 9, characterized in that, The process of recalling candidate competing nodes for the target competing unit includes: Based on the first competition intensity coefficient, candidate node recall processing is performed on the target competing unit to obtain a first candidate competing node set; Based on the context data, the node attributes, and the historical interaction data, semantic retrieval and vector similarity calculation are performed on the target competing unit to obtain a second set of candidate competing nodes. The first set of candidate competing nodes and the second set of candidate competing nodes are merged to obtain the set of candidate competing nodes.

11. The method according to claim 8, characterized in that, The step of establishing the competitive relationship of the target competing unit in the first competitive relationship graph based on the candidate competing node set includes: The connection edge relationships, historical interaction data, and node attributes of the target competing unit are obtained through the intelligent agent corresponding to the target competing unit. Based on the connection edge relationships of the target competing unit, the historical interaction data, and the node attributes, the competitive relationship between the target competing unit and the set of candidate competing nodes is established.

12. The method according to any one of claims 1 to 11, characterized in that, The method further includes: Obtain second competitive relationship data, and construct positive sample pairs and negative sample pairs based on the second competitive relationship data; The agent is adjusted based on the positive sample pairs and the negative samples to obtain the adjusted agent.

13. The method according to claim 12, characterized in that, After determining the competitive relationships among the multiple competing units based on the competitive relationship graph of the target round, the method further includes: The competition relationship graph of the target round is evaluated based on the first model to obtain a first reward value; wherein, the first model is a reward model trained with an expert feedback dataset, and the expert feedback dataset is obtained by evaluating the competition relationship; The agent is adjusted based on the first reward value to obtain the adjusted agent.

14. A device for determining competitive relationships, characterized in that, The device includes: The data acquisition module is used to acquire a set of competing units, which includes multiple competing units, and the competing units are media objects that compete for exposure opportunities. The data processing module is used to allocate multiple agents to the set of competing units, with each agent corresponding one-to-one with a competing unit in the set. Through these agents, multiple rounds of competition relationships are established between the competing units, with each round's establishment corresponding to a competition relationship graph. The module then determines whether the competition relationship graph for each round meets preset conditions. If the competition relationship graph for a target round meets the preset conditions, the competitive relationships between the competing units are determined based on the target round's competition relationship graph.

15. An electronic device, characterized in that, The electronic device includes: Memory is used to store executable instructions or computer programs. A processor, when executing computer-executable instructions or computer programs stored in the memory, implements the method according to any one of claims 1 to 13.

16. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, they implement the method described in any one of claims 1 to 13.

17. A computer program product comprising computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or computer program are executed by a processor, they implement the method according to any one of claims 1 to 13.

Images