An AI technology-based sports event data mining method and system

Abstract

The application discloses a sports event data mining method and system based on AI technology, and relates to the technical field of machine learning, comprising the following steps: aligning each multi-source sports event data in space-time through a dynamic time warping algorithm, and outputting a multi-modal event data package; identifying sports entity coordinates and tactical events from the multi-modal event data package, and outputting a tactical mode report; fusing and instantiating the tactical mode report and preset sports field rules through neural symbols, and outputting a tactical knowledge graph; obtaining tactical instances in the tactical knowledge graph through logical deduction retrieval, and performing causal reasoning analysis on the tactical instances to generate a tactical causal graph; and performing tactical mode matching and effect evaluation on real-time space-time hypergraphs and the tactical causal graph, and outputting a sports event tactical instruction set. The application realizes the deep conversion from multi-modal data to high-order tactical decision by constructing a reasoning tactical knowledge graph and a knowledge completion mechanism.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for mining sports event data based on AI technology. Background Technology

[0002] In recent years, the field of sports event data analysis has deeply integrated artificial intelligence and big data mining, forming an evolutionary path from multi-source perception to intelligent decision-making. Computer vision is widely used for athlete posture estimation and behavior recognition, combined with inertial measurement units and optical tracking to collect high-precision spatiotemporal coordinates. Graph structure data models are used to characterize player interactions. Deep learning models, especially temporal neural networks and graph neural networks, are applied to detect basic tactical events and identify fixed tactical patterns from game videos and sensor data streams. Knowledge graph methods have also been introduced for structured storage of event entities and rules, while causal inference methods provide a preliminary framework for analyzing tactical correlations. Together, these technologies form the foundational technical system for current sports big data analysis.

[0003] However, existing methods have limitations in deep knowledge discovery and real-time decision support. Multi-source heterogeneous data often fails to achieve accurate spatiotemporal alignment, resulting in compromised analytical accuracy. Tactical pattern recognition often relies on preset templates or shallow statistics, making it difficult to automatically discover complex and high-level team collaboration patterns. It lacks symbolic reasoning capabilities that integrate domain knowledge, leading to poor interpretability and the inability to make causal inferences. Existing methods typically focus on post-match analysis and cannot build a queryable and reasonable tactical knowledge system, making it even more difficult to support real-time tactical decision-making during matches, thus hindering the effective transformation of data value into practical command capabilities. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a sports event data mining method based on AI technology to solve the limitations in deep knowledge discovery and real-time decision support.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a sports event data mining method based on AI technology, comprising: collecting multi-source sports event data; aligning the multi-source sports event data in time and space using a dynamic time warping algorithm to output a multimodal event data package; identifying sports entity coordinates and tactical events from the multimodal event data package, constructing a spatiotemporal hypergraph, defining meta-path templates for the spatiotemporal hypergraph, and performing walk sampling and sequence pattern mining to output a tactical pattern report; performing neural symbol fusion and instantiation transformation on the tactical pattern report and preset sports domain rules to output a preliminary neural graph, and performing knowledge completion on the preliminary neural graph to output a tactical knowledge graph; retrieving tactical instances from the tactical knowledge graph through logical deduction, and performing causal reasoning analysis on the tactical instances to generate a tactical causal graph; and performing tactical pattern matching and effect evaluation on the real-time spatiotemporal hypergraph and the tactical causal graph to output a sports event tactical instruction set.

[0008] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the specific steps for outputting the multimodal event data package are as follows:

[0009] Establish a time reference axis for multi-source sports event data based on the timestamps of sports events;

[0010] Multi-source sports event data is time-aligned and normalized according to a time baseline to generate spatiotemporally aligned event data.

[0011] The spatiotemporally aligned event data is integrated and standardized under a unified spatiotemporal coordinate system to form a multimodal event data package.

[0012] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the specific steps for identifying sports entity coordinates and tactical events from multimodal event data packets and constructing a spatiotemporal hypergraph are as follows:

[0013] Extract the spatiotemporal location information of athletes, referees, and the ball from multimodal event data packets, and integrate them to generate a set of sports entity coordinates;

[0014] Identify the movement and interaction patterns of each sports entity in the sports entity coordinate set, and generate tactical event data;

[0015] A spatiotemporal hypergraph is constructed by using sports entities as dynamic nodes in the sports entity coordinate set and the interaction relationships in the tactical event data as sports hyperedges.

[0016] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the specific steps for outputting the tactical pattern report are as follows:

[0017] Based on the relationship between dynamic nodes and sports hyperedges in the spatiotemporal hypergraph, a meta-path template reflecting tactical logic is defined.

[0018] Biased transfer sampling is performed on the spatiotemporal hypergraph according to the meta-path template, and the dynamic nodes and sports hyperedges passed through in each sampling are recorded as a walking path sequence.

[0019] Sequence pattern mining is performed on the wandering path sequences to identify high-frequency path patterns and integrate them to form a tactical pattern report.

[0020] As a preferred embodiment of the AI-based sports event data mining method of the present invention, the specific steps of fusing and instantiating tactical pattern reports with preset sports domain rules to output a preliminary neural map are as follows:

[0021] Extract the dynamic node sequence and sports superedge type of the path pattern in the tactical mode report, and convert the preset sports domain rules into a set of logical rules.

[0022] The dynamic node sequence and sports hyperedge type are transformed into numerical vector representations, and the numerical vector representations are combined with the set of logical rules for constraint optimization to generate regularized vector representations.

[0023] The regularized vector representation is decoded into corresponding entity nodes, node relationship edges, and relationship reasoning edges derived from logical rules, and then integrated to output a preliminary neural graph.

[0024] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the steps of decoding the regularized vector representation into corresponding entity nodes, node relationship edges, and relationship reasoning edges derived from logical rules, and integrating them to output a preliminary neural graph, are as follows:

[0025] The tensor decomposition method is used to separate the sub-vectors corresponding to entity features and the sub-vectors corresponding to relation features from the regularized vector representation, and then integrate them to generate a set of entity feature vectors and a set of relation feature vectors.

[0026] Each entity feature vector in the entity feature vector set is mapped to an entity node with a unique identifier and attributes, generating an entity node set;

[0027] Each relation feature vector in the relation feature vector set is mapped to a node relation edge connecting different entity nodes, and the relation type is labeled for the node relation edge according to the semantic information of the relation feature vector, thus generating a node relation edge set;

[0028] The set of logical rules is applied to the set of entity nodes and the set of node relationship edges to perform logical derivation reasoning, generating a set of relationship reasoning edges;

[0029] A preliminary neural graph is constructed by topologically connecting the set of entity nodes, the set of node relationship edges, and the set of relationship reasoning edges.

[0030] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the knowledge completion of the preliminary neural graph refers to adding entity nodes and node relationships to the preliminary neural graph based on the cosine similarity between regularized vector representations, and using a set of logical rules to verify the consistency of entity nodes and node relationships, thereby forming a tactical knowledge graph with complete and rigorous structure and logic.

[0031] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the specific steps for generating the tactical causal graph are as follows:

[0032] Acquire tactical query data, retrieve tactical instances that satisfy the tactical query data from the tactical knowledge graph through logical deduction, and integrate them to generate a set of tactical instances;

[0033] Each tactical instance in the tactical instance set is analyzed to identify the dependencies between entity nodes and node relationship edges on the path from initial conditions to final result, and a causal path set is generated.

[0034] The dependencies in the set of causal paths are encoded as causal relationships and integrated into the structure of the tactical knowledge graph to form a tactical causal graph.

[0035] As a preferred embodiment of the AI-based sports event data mining method described in this invention, the specific steps for outputting the sports event tactical instruction set are as follows:

[0036] Tactical pattern similarity matching is performed between the real-time spatiotemporal hypergraph and the tactical causal graph to generate tactical pattern matching data.

[0037] The tactical pattern matching data is evaluated based on the causal relationships in the tactical causal graph to generate tactical effect data.

[0038] The tactical effect data is structured and compiled for decision-making, outputting a set of tactical instructions for sports events.

[0039] Secondly, this invention provides a sports event data mining system based on AI technology, comprising,

[0040] The acquisition module is used to collect multi-source sports event data, and uses a dynamic time warping algorithm to align the multi-source sports event data in time and space, and output multimodal event data packets.

[0041] The reporting module is used to identify sports entity coordinates and tactical events from multimodal event data packets, construct a spatiotemporal hypergraph, define meta-path templates for the spatiotemporal hypergraph, perform walk sampling and sequence pattern mining, and output a tactical pattern report.

[0042] The graph module is used to perform neural symbol fusion and instantiation transformation between tactical pattern reports and preset sports domain rules, output a preliminary neural graph, and complete the preliminary neural graph with knowledge to output a tactical knowledge graph.

[0043] The reasoning module is used to retrieve tactical instances from the tactical knowledge graph through logical deduction, and to perform causal reasoning analysis on the tactical instances to generate a tactical causal graph.

[0044] The instruction module is used to perform tactical pattern matching and effect evaluation between the real-time spatiotemporal hypergraph and the tactical causal graph, and output the tactical instruction set for sports events.

[0045] The beneficial effects of this invention are as follows: By constructing a reasonable tactical knowledge graph and a knowledge completion mechanism, a deep transformation from multimodal data to high-order tactical decisions is achieved. Through neural symbolic fusion and instantiation of tactical pattern reports and preset sports domain rules, data-driven complex patterns are deeply integrated with symbolic domain logic, constructing a structured tactical knowledge system that combines statistical learning and logical reasoning capabilities. This solves the problem of the separation between data models and domain knowledge in traditional methods, significantly improving the interpretability and logical consistency of tactical representations. By performing knowledge completion on the initial neural graph based on the cosine similarity between rule-based vector representations, semantically related entity nodes and node relationships are automatically inferred and added, making the structure of the tactical knowledge graph more complete and rigorous, and enhancing the completeness of the knowledge system and the robustness of reasoning. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a sports event data mining method based on AI technology.

[0048] Figure 2 This is a schematic diagram of a sports event data mining system based on AI technology.

[0049] Figure 3 This is a flowchart for outputting a tactical mode report.

[0050] Figure 4A flowchart for constructing a preliminary neural atlas. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for mining sports event data based on AI technology, including the following steps:

[0055] S1. Collect multi-source sports event data, and use a dynamic time warping algorithm to align the multi-source sports event data in time and space, and output a multimodal event data packet.

[0056] Collect multi-source sports event data and establish a time reference axis for the multi-source sports event data based on the timestamps of the sports events.

[0057] Specifically, the collection of multi-source sports event data includes obtaining structured match event logs from official event data interfaces, video streams from high-definition cameras deployed in the venue, biomechanical sensor signals containing acceleration and angular velocity from athletes' wearable devices, and audio streams containing whistles, ball-hitting sounds, and commentary voices from live broadcasts. Multi-source sports event data has high dimensionality, high frequency, and large scale characteristics, which meet the requirements of big data mining for data diversity and volume. During the collection process, the timestamps of sports events embedded in the main camera video stream are used as the benchmark. The event timestamps in the official event data interface, the sensor sampling timestamps output by wearable devices, and the audio frame timestamps in the audio stream are converted to the same time reference system. The time granularity of each data stream is aligned through interpolation and resampling, and asynchronous data is calibrated based on landmark audio events such as whistles or kick-offs to establish a time reference axis for multi-source sports event data.

[0058] Multi-source sports event data is time-aligned and normalized according to a time baseline to generate spatiotemporally aligned event data.

[0059] Specifically, the structured match event logs, video streams from high-definition cameras, biomechanical sensor signals from athletes' wearable devices, and audio streams from on-site broadcasts in the official event data interface are time-mapped frame-by-frame or sample-by-sample point according to the established time reference axis of multi-source sports event data. The dynamic time warping algorithm is used to non-linearly stretch or compress each data stream in the time dimension, so that video frames, sensor signal sampling points, audio frames, and match event records correspond precisely under a unified time scale. The timing deviation is corrected by combining iconic audio events such as whistles and ball-hitting sounds as anchor points to generate spatiotemporally aligned event data.

[0060] Furthermore, the input to the dynamic time warping algorithm is multi-source sports event data streams from different acquisition devices. Each multi-source sports event data stream contains timestamps recorded at the original sampling frequency and corresponding observation content. The dynamic time warping algorithm constructs a cumulative distance matrix between the reference time series and each sequence to be aligned, and uses dynamic programming to search for time mapping paths that satisfy boundary, monotonicity, and continuity constraints, thereby flexibly aligning asynchronous and variable-rate time series to a unified time reference axis. The output is spatiotemporally aligned event data aligned frame by frame under a unified time reference, with time synchronization accuracy controlled at the millisecond level to ensure the temporal consistency of subsequent multimodal fusion.

[0061] The spatiotemporally aligned event data is integrated and standardized under a unified spatiotemporal coordinate system to form a multimodal event data package.

[0062] Specifically, a three-dimensional spatial coordinate system is established based on the physical dimensions of the competition venue. The position coordinates of athletes, referees, and the ball extracted from the video streams acquired by high-definition cameras are transformed into the three-dimensional spatial coordinate system. The biomechanical sensor signals acquired by the athletes' wearable devices are aligned with the timestamps and associated with the spatial locations of the corresponding entities. The event types and occurrence locations in the structured competition event logs from the official event data interface are mapped to spatiotemporal points in the same coordinate system. Key audio event markers from the audio streams acquired by the live broadcast are embedded in the corresponding spatiotemporal coordinates. Field naming is unified, physical quantity units are standardized, and data structures are aligned for the video streams acquired by high-definition cameras, the acceleration and angular velocity signals acquired by the athletes' wearable devices, the structured competition event logs acquired by the official event data interface, and the audio streams acquired by the live broadcast, forming a multimodal event data package.

[0063] S2. Identify sports entity coordinates and tactical events from multimodal event data packets, construct a spatiotemporal hypergraph, define meta-path templates for the spatiotemporal hypergraph, perform walk sampling and sequence pattern mining, and output a tactical pattern report.

[0064] The spatiotemporal location information of athletes, referees and the ball is extracted from multimodal event data packets and integrated to generate a set of sports entity coordinates.

[0065] Specifically, frame-by-frame target detection and multi-target tracking are performed on the video streams acquired by high-definition cameras in the multimodal event data package to obtain the spatial coordinates and corresponding timestamps of athletes, referees, and the ball in a three-dimensional spatial coordinate system in each frame; the player numbers, referee identifiers, and ball event positions recorded in the structured match event logs obtained from the official event data interface are mapped to the same three-dimensional spatial coordinate system; combined with the acceleration and angular velocity signals obtained from the athletes' wearable devices, the athletes' posture and position are corrected with the help of inverse kinematics; the spatiotemporal position information of athletes, referees, and the ball is aligned by timestamps and then uniformly merged into a structured record containing entity identifiers, spatial coordinates, and timestamps, and integrated to generate a sports entity coordinate set.

[0066] Identify the movement and interaction patterns of each sports entity in the sports entity coordinate set to generate tactical event data.

[0067] Specifically, based on the spatiotemporal position information of athletes, referees, and the ball in the sports entity coordinate set, the velocity vector, acceleration vector, and trajectory curvature of each athlete are calculated to identify movement patterns such as acceleration breakthroughs, sudden stops and changes of direction, and retreat defense. By analyzing the rate of change of distance and relative azimuth between athletes and the ball, interaction behaviors such as ball possession, passing, shooting, and stealing are determined. For example, when the distance between an athlete and the ball increases rapidly and the direction of the ball's movement is consistent with the direction of the passer, it is determined to be a pass. Based on the spatial topological relationship characteristics between multiple athletes, group interaction patterns such as pick-and-roll, double-team, switching, and cross-cutting are identified. That is, by calculating the relative position, spacing changes, and consistency of movement direction of multiple athletes in consecutive time frames, when two offensive athletes are detected to briefly obscure the path of a defender and then separate, it is identified as a pick-and-roll; when two or more defensive athletes simultaneously shrink towards the ball handler, it is identified as a double-team. The identified movement patterns and interaction behaviors are structured and encoded according to the occurrence time, participating entities, and spatial position to generate tactical event data.

[0068] The formula for calculating the velocity vector is:

[0069] ;

[0070] The formula for calculating the acceleration vector is:

[0071] ;

[0072] The formula for calculating the curvature of the motion trajectory is:

[0073] ;

[0074] in, The time identifier representing the current moment. Indicates the time interval between two adjacent frames. Indicates at a point in time The velocity vector at time, Indicates at a point in time The position at that time Indicates at a point in time The position at that time Indicates at a point in time The acceleration vector at time, Indicates at a point in time The velocity vector at time, It represents the curvature of the motion trajectory.

[0075] A spatiotemporal hypergraph is constructed by using sports entities as dynamic nodes in the sports entity coordinate set and the interaction relationships in the tactical event data as sports hyperedges.

[0076] Specifically, the athletes, referees, and balls at each moment in the sports entity coordinate set are represented as dynamic nodes with unique identifiers and spatiotemporal coordinates. Based on the ball-handling, passing, screen-setting, and double-teaming interactions recorded in the tactical event data, multiple dynamic nodes participating in the same tactical event are linked together to form a sports hyperedge connecting different dynamic nodes. The sports hyperedge includes event type, occurrence time, and spatial location attributes. The dynamic nodes at each moment and their corresponding sports hyperedges are organized in chronological order into a graph structure that evolves over time, forming a spatiotemporal hypergraph.

[0077] Based on the relationship between dynamic nodes and sports hyperedges in the spatiotemporal hypergraph, a meta-path template reflecting tactical logic is defined.

[0078] Specifically, this study analyzes frequently occurring sequences of dynamic node types and connected sports hyperedge types in the spatiotemporal hypergraph to extract tactically significant structural patterns, such as "ball handler - passing sports hyperedge - receiving player - screen sports hyperedge - screener". These structural patterns are then abstracted into path rules composed of alternating dynamic node types and sports hyperedge types. Each entity role (e.g., ball handler) in the structural pattern is mapped to a dynamic node type, and each interaction behavior (e.g., passing) is mapped to a sports hyperedge type, arranged alternately in chronological order. Semantic roles and connection constraints are labeled for each type of dynamic node and sports hyperedge (connection constraints refer to the constraints on dynamic nodes). The semantic constraints that must be satisfied when a sports hyperedge appears adjacent to another sports hyperedge in the metapath template, such as "the starting point of the passing sports hyperedge must be a dynamic node of the ball-handling player type, and the ending point must be a dynamic node of the non-ball-handling player type", are used to complete the abstraction and clarify the semantic roles and order constraints of each link (the order constraint means that the dynamic node type and the sports hyperedge type in the metapath template must be arranged in chronological or logical order, such as "the screen sports hyperedge must appear after the passing sports hyperedge" or "the screen action must be completed before receiving the ball"). Multiple path rules are formalized into configurable metapath templates to guide the biased transfer sampling in the spatiotemporal hypergraph.

[0079] Furthermore, the meta-path template is a semantic path pattern defined in a spatiotemporal hypergraph. Essentially, it is a directed sequence composed of alternating dynamic node types and sports hyperedge types, used to characterize the interaction flow of entities with tactical significance. The specific construction method is as follows: based on typical tactical logic in the sports field (such as pick-and-roll and fast break), determine the entity type sequence of participating roles (such as ball handler, screener, and receiver) and the interaction behavior type sequence (such as passing, pick-and-roll, and positioning), and arrange the entity type sequence and interaction behavior type sequence alternately according to time sequence and causal dependency to form a structured path rule in the form of "dynamic node type - sports hyperedge type - dynamic node type", which serves as a semantic guidance template for subsequent walk sampling.

[0080] Biased transition sampling is performed on the spatiotemporal hypergraph according to the metapath template, and the dynamic nodes and sports hyperedges passed through in each sampling are recorded as a walking path sequence.

[0081] Specifically, starting from any dynamic node in the spatiotemporal hypergraph that matches the type of the starting dynamic node in the metapath template (matching means that the entity type of the current dynamic node, such as "ball-carrying player", is completely consistent with the type of the dynamic node specified at the starting position of the metapath template), only sports hyperedges whose type matches the requirements of the next position of the metapath template are selected for transfer among the adjacent sports hyperedges of the current node. Among the dynamic nodes connected by the current sports hyperedge, dynamic nodes that meet the subsequent node type constraints of the metapath template are selected as the next sampling target. Biased transfer sampling is repeated, strictly following the alternation order of dynamic node type and sports hyperedge type defined by the metapath template, and advancing jump by jump within the time slice of the spatiotemporal hypergraph until the traversal of the entire metapath template is completed. The dynamic nodes and sports hyperedges visited in sequence during the traversal are spliced ​​together according to the alternation order of dynamic node type and sports hyperedge type defined by the metapath template to form a complete walking path sequence.

[0082] Sequence pattern mining is performed on the wandering path sequences to identify high-frequency path patterns and integrate them to form a tactical pattern report.

[0083] Specifically, all dynamic nodes and sports hyperedges in the entire walking path sequence are converted into symbolic path strings composed of node type identifiers and hyperedge type identifiers according to their order in the sequence; all symbolic path strings are traversed, and the occurrence frequency of each continuous subsequence in the entire dataset is counted; continuous subsequences that occur more than once are selected as candidate path patterns, and short sequences containing only a single node are excluded; for the retained candidate path patterns, the complete node-hyperedge alternation structure, the types of participating athletes, the time stage of occurrence, and the corresponding spatial region information are recorded; the node-hyperedge alternation structure, the types of participating athletes, the time stage of occurrence, and the corresponding spatial region information are classified and summarized according to the path pattern structure to form a tactical pattern report containing path pattern description, occurrence frequency, participating entity type, and spatial context.

[0084] S3. Perform neural symbol fusion and instantiation transformation on the tactical mode report and the preset sports domain rules to output a preliminary neural map. Then, perform knowledge completion on the preliminary neural map to output a tactical knowledge map.

[0085] Extract the dynamic node sequence and sports superedge type of the path pattern from the tactical mode report, and convert the preset sports domain rules into a set of logical rules.

[0086] Specifically, each path pattern in the tactical pattern report is traversed, and the dynamic nodes and corresponding sports hyperedges contained within the path pattern are parsed according to the chronological order of events occurring in the path pattern. The entity type identifier of each dynamic node and the interaction type identifier of each sports hyperedge are recorded respectively, forming a structured sequence composed of alternating entity types and interaction types. The natural language descriptions or formal statements in the preset sports domain rules are converted into first-order logical predicate expressions one by one according to the content of the sports domain rules. Each expression clearly includes the predicate name, participating entity variables and constraints. All the converted first-order logical predicate expressions are integrated to form a set of logical rules.

[0087] Furthermore, the pre-defined sports domain rules are set based on the competition regulations and event manuals issued by sports federations, specifically including three categories: The first category is legality rules, used to determine whether an event is compliant, such as "in a basketball game, if the offensive team holds the ball in the backcourt for more than 8 seconds without advancing it to the frontcourt, it constitutes a violation"; the second category is role behavior constraint rules, which stipulate the behavioral boundaries of specific entities, such as "in football, the goalkeeper may touch the ball with his hands in his own penalty area, except when receiving a back pass"; the third category is tactical validity rules, which describe the structural conditions of effective tactical combinations, such as "in a basketball pick-and-roll, if the screener bounces out to the three-point line and is open after completing the screen, it is considered a valid bounce out." All sports domain rules are organized in the form of formal statements or in the form of first-order logical predicates.

[0088] Neural symbolic fusion refers to the process of co-modeling vector embedding representations based on neural networks with domain rules based on symbolic logic. The core lies in transforming discrete first-order logic rules into differentiable soft constraints and solving them jointly with the optimization objective of neural embedding. In specific implementation, dynamic node sequences and sports hyperedge types are embedded and encoded to obtain numerical vector representations. At the same time, the first-order logic predicate expressions in the preset sports domain rules (such as "If A passes the ball to B, then B must be an attacking player") are transformed into logic constraint functions. The logic constraint functions take the embedded vectors as input and output the degree of rule violation as a soft penalty term. Under the unified optimization framework, the goal is to minimize the sum of path reconstruction error and logic constraint loss. The embedding parameters are updated iteratively through gradient descent so that the final vector representation retains the data-driven structural features and satisfies the semantic constraints of symbolic rules.

[0089] The dynamic node sequence and sports hyperedge type are transformed into numerical vector representations, and the numerical vector representations are combined with the set of logical rules for constraint optimization to generate regularized vector representations.

[0090] Specifically, each entity type identifier in the dynamic node sequence and each interaction type identifier in the sports hyperedge type are embedded and encoded to obtain the corresponding entity embedding vector and relation embedding vector, which are then concatenated in the order of the path pattern to form a path-level numerical vector representation. At the same time, the first-order logical predicate expressions in the logical rule set are transformed into differentiable logical constraint functions. For example, "If A passes the ball to B, then B must be an attacking player" is transformed into a soft constraint loss term based on vector semantics. During the optimization process, the path-level numerical vector representation is used as input to jointly minimize the reconstruction error and logical constraint loss. The embedding parameters are adjusted iteratively through gradient descent so that the output numerical vector representation retains the structural features of the tactical pattern and satisfies the semantic constraints of the sports domain rules, forming a regularized vector representation.

[0091] The tensor decomposition method is used to separate sub-vectors corresponding to entity features and sub-vectors corresponding to relation features from the regularized vector representation, and then integrates them to generate a set of entity feature vectors and a set of relation feature vectors.

[0092] Specifically, the regularized vector representation is organized into a three-dimensional tensor structure, with the dimensions of the three-dimensional tensor structure corresponding to the dynamic node positions, sports hyperedge positions, and embedded feature dimensions in the path, respectively. A higher-order singular value decomposition (SVFD) method is used to factorize the three-dimensional tensor, extracting the factor matrix along the dynamic node dimension as entity feature components and the factor matrix along the sports hyperedge dimension as relation feature components. That is, a higher-order singular value decomposition method is used to factorize the three-dimensional tensor constructed from the regularized vector representation. Specifically, the regularized vector representation is organized into a third-order tensor according to the dynamic node positions, sports hyperedge positions, and embedded feature dimensions in the path. Then, matrix expansions are performed on the third-order tensor along the dynamic node mode, sports hyperedge mode, and feature mode, respectively. For each expanded matrix, standard singular value decomposition is performed to extract the left singular vector matrix corresponding to each mode. The left singular vector matrix of each mode is then used to perform multilinear projection on the third-order tensor to obtain the core tensor and three mode factor matrices. Each row of the mode factor matrix corresponding to the dynamic node mode is an entity feature sub-vector, and each row of the factor matrix corresponding to the sports hyperedge mode is a relation feature sub-vector, thus separating entity feature components and relation feature components from the regularized vector representation. The entity feature components corresponding to each dynamic node position are normalized to form a set of entity feature vectors; the relation feature components corresponding to each sports hyperedge position are also normalized to form a set of relation feature vectors.

[0093] Each entity feature vector in the entity feature vector set is mapped to an entity node with a unique identifier and attributes, generating an entity node set.

[0094] Specifically, each entity feature vector in the entity feature vector set is traversed, and a globally unique entity identifier is assigned to the entity feature vector based on the dynamic node position and path context corresponding to the entity feature vector in the regularized vector representation; the role type, team, jersey number, and spatial location attributes corresponding to the current entity are extracted from the sports entity coordinate set, and the role type, team, jersey number, and spatial location attributes are bound to the entity identifier; the entity identifier, attribute information, and corresponding entity feature vector are encapsulated into structured data as entity nodes, and all entity nodes are aggregated to generate an entity node set.

[0095] Each relation feature vector in the relation feature vector set is mapped to a node relation edge connecting different entity nodes, and the relation type is labeled for the node relation edge according to the semantic information of the relation feature vector, thus generating a node relation edge set.

[0096] Specifically, each relation feature vector in the relation feature vector set is traversed. Based on the sports hyperedge position corresponding to the relation feature vector in the regularized vector representation and the associated dynamic node context, the two entity nodes connected by the current relation are determined. The two entity nodes are connected by a directed or undirected node relation edge. Based on the semantic similarity between the direction of the relation feature vector in the embedding space and the cluster centers of nearby known interaction types (such as passing, screen, and steal), the interaction semantic category is determined. That is, by calculating the Euclidean distance between the relation feature vector and the cluster center vectors of each interaction type, the interaction type corresponding to the cluster center with the shortest distance is selected as the interaction semantic category of the current relation feature vector, and the interaction semantic category is labeled as the relation type on the node relation edge. All node relation edges labeled with relation types are collected to generate a node relation edge set.

[0097] The set of logical rules is applied to the set of entity nodes and the set of node relationship edges to perform logical derivation reasoning, generating a set of relationship reasoning edges.

[0098] Specifically, each first-order logical predicate expression in the set of logical rules is traversed, and the predicate variables in the first-order logical predicate expression are matched and instantiated with the entity nodes in the set of entity nodes and the node relationship edges in the set of node relationship edges. When any pair of entity nodes and existing node relationship edges satisfy the logical rules (a logical rule is a first-order logical predicate expression consisting of premises and conclusions, where the premises describe the constraint relationships between several entity node types, attributes, and node relationship edge types; satisfying the logical rules means that there exists a specific combination of elements in the set of entity nodes and the set of node relationship edges, and the semantic roles, relationship types, and structural relationships of the element combinations are completely consistent with the constraints specified by the premises), a new relational fact is derived based on the conclusion part of the current logical rule. The new relational fact is represented as a directed edge connecting the corresponding entity nodes, and the relationship type is labeled as the derived relationship type specified by the logical rule. All new relational edges derived through the logical rules are collected to generate a relational reasoning edge set.

[0099] A preliminary neural graph is constructed by topologically connecting the set of entity nodes, the set of node relationship edges, and the set of relationship reasoning edges.

[0100] Specifically, each entity node in the entity node set is taken as a vertex. Each node relationship edge in the node relationship edge set is established as a directed or undirected edge between the corresponding vertices according to the two entity nodes connected and the labeled relationship type. Similarly, each relationship reasoning edge in the relationship reasoning edge set is added with a corresponding reasoning edge according to the entity node connected and the derived relationship type. All entity nodes, node relationship edges, and relationship reasoning edges together constitute a heterogeneous graph structure with a unified identifier space, semantic relationship labels, and vector embedding attributes, i.e., a preliminary neural graph.

[0101] Complete the preliminary neural network map with knowledge and output a tactical knowledge map.

[0102] Specifically, the cosine similarity between the regularized vector representations of all entity nodes in the preliminary neural graph is calculated. When the cosine similarity between the regularized vector representations of any two entity nodes is higher than a set similarity threshold, it is determined that the two entity nodes have a potential semantic or functional relationship. Based on the potential relationship, new entity nodes (such as inferred implicit roles) or new node relationship edges (such as potential cooperation relationships) are added to the preliminary neural graph. The newly added entity nodes and node relationships are input into the logical rule set for consistency verification, that is, to verify whether the legality, role constraints and tactical effectiveness conditions defined by the preset sports domain rules are met simultaneously. If the verification passes, the newly added content is retained. After the expansion based on vector similarity and the verification iteration based on the logical rule set, a tactical knowledge graph with complete and rigorous structure is formed.

[0103] The formula for calculating cosine similarity is:

[0104] ;

[0105] in, This represents the cosine similarity between the regularized vector representations. The dimension identifier representing the regularized vector. This represents the total number of dimensions of the regularized vector. The first regularized vector represents the vector in the first... Values ​​in each dimension The second regularized vector represents the vector in the first... The values ​​in each dimension.

[0106] Furthermore, the set similarity threshold is not a fixed constant, but is dynamically determined based on the characteristics of the sports project, the distribution of historical tactical data, and the strength of rule constraints: the cosine similarity distribution between the regularized vector representations of corresponding entity nodes is calculated based on the labeled tactical knowledge graph samples (such as real-world cooperation relationships), and the lower quartile of the cosine similarity distribution or the critical point that optimizes the combined accuracy and recall of knowledge completion is selected as the initial threshold through cross-validation; the tolerance for false additions is adjusted in conjunction with the rule set in the sports domain, such as increasing the similarity threshold (e.g., above 0.85) in defensive cooperation tasks due to the high convergence of behavioral patterns, while appropriately decreasing it (e.g., around 0.70) in offensive creation scenarios due to the diversity of strategies; the similarity threshold usually falls in the range of 0.70–0.90.

[0107] By deeply integrating tactical pattern reports with sports rules, this approach organically combines representational capabilities with the advantages of symbolic logic reasoning. While ensuring tactical semantic compliance, it significantly improves the accuracy and interpretability of mining high-level tactical knowledge from massive, multi-source event data. Based on a knowledge completion mechanism using rule-based vector representation and logical consistency verification, it effectively overcomes the generalization bottleneck of traditional big data mining methods in sparse interaction and implicit association scenarios. This enables sports event data mining to stably extract structured, reasonable, and executable tactical knowledge from complex, heterogeneous, and high-dimensional sports event big data, providing a high-quality knowledge foundation for subsequent causal analysis and real-time decision-making.

[0108] S4. Retrieve tactical instances from the tactical knowledge graph through logical deduction, and perform causal reasoning analysis on the tactical instances to generate a tactical causal graph.

[0109] Acquire tactical query data, retrieve tactical instances that satisfy the tactical query data from the tactical knowledge graph through logical deduction, and integrate them to generate a set of tactical instances.

[0110] Specifically, the system receives tactical query data expressed in first-order logic or structured statements. The tactical query data includes the target tactical type, participating entity roles, and constraints. Constraints refer to the explicitly specified target tactical type (e.g., a pick-and-roll followed by a pop), participating entity roles (e.g., the ball-handling guard and the screening center), and additional constraints (e.g., the screen occurs beyond the three-point line and the receiver is open). Constraints are input by the user or automatically generated by a higher-level task and expressed as first-order logic predicates. The predicates and variables in the tactical query data are mapped to entity nodes, node relationship edges, and relationship reasoning edges in the tactical knowledge graph. Forward chain reasoning is performed on the tactical knowledge graph based on the set of logical rules, traversing all possible combinations of entities and relationships to select subgraph structures that satisfy the constraints of the tactical query data. Each subgraph structure that meets the conditions is treated as a tactical instance, and its contained entity nodes, relationship edges, and attribute information are extracted. All tactical instances are then aggregated and integrated to generate a tactical instance set.

[0111] Each tactical instance in the tactical instance set is analyzed to identify the dependencies between entity nodes and node relationship edges on the path from initial conditions to final result, and a causal path set is generated.

[0112] Specifically, for the subgraph structure corresponding to each tactical instance, the entity nodes and node relationship edges are topologically sorted according to the timestamp and the order of events to determine the temporal evolution path from the initial state (such as the start of ball possession) to the final state (such as a shot score or a turnover). Based on the action sequence constraints and causal semantics defined in sports rules (such as passing must precede receiving and screen must precede bounce), the system analyzes whether there are causal dependencies such as triggering, enabling, or hindering between adjacent or intermittent steps. Each temporal path that fully satisfies causal logic is extracted into an ordered sequence composed of alternating entity nodes and node relationship edges, and the causal type between each link is labeled. The ordered sequences extracted from all tactical instances are collected to generate a causal path set.

[0113] The dependencies in the set of causal paths are encoded as causal relationships and integrated into the structure of the tactical knowledge graph to form a tactical causal graph.

[0114] Specifically, each ordered sequence in the causal path set is traversed, and adjacent entity nodes or node relationship edge pairs with causal semantic associations are identified. Based on the dependency type (such as trigger, enable, and hinder), corresponding causal relationship triples are formed in the form of <causal element, causal relationship type, result element>, where the causal element and result element are existing entity nodes or node relationship edges in the tactical knowledge graph. A unified relation predicate identifier is defined for each type of causal relationship, and directed causal edges labeled with relation predicate identifiers are added to the tactical knowledge graph. All the newly added causal edges, entity nodes, node relationship edges, and relation reasoning edges are organized together into an extended graph structure containing causal semantics, namely the tactical causal graph.

[0115] S5. Perform tactical pattern matching and effect evaluation on the real-time spatiotemporal hypergraph and tactical causal graph, and output the tactical instruction set for sports events.

[0116] Tactical pattern similarity matching is performed between the real-time spatiotemporal hypergraph and the tactical causal graph to generate tactical pattern matching data.

[0117] Specifically, the dynamic node sequence and sports hyperedge sequence corresponding to the current game segment are extracted from the real-time spatiotemporal hypergraph to form the current tactical path. The dynamic node type and sports hyperedge type in the current tactical path are converted into symbolic representations consistent with those in the tactical causal graph. All encoded causal paths in the tactical causal graph are traversed, and the similarity between the current tactical path and each causal path in terms of structure, node role, hyperedge type, and temporal order is calculated. The path edit distance is used to quantify the similarity, that is, the allowed replacement operations (such as inserting, deleting, or replacing a dynamic node or sports hyperedge) are defined, and different costs are assigned to the replacement operations of different types of dynamic nodes or sports hyperedges (e.g., the cost of replacing the same role type is 0, and the cost of replacing different role types is 1). The minimum total edit cost required to convert the current tactical path into a causal path is calculated, and the normalized value of the minimum total edit cost is used as the similarity between the current tactical path and each causal path in terms of structure, node role, hyperedge type, and temporal order. The causal path with the highest similarity is selected as the matching result, and the corresponding tactical name and causal structure are recorded. All matching results and related information are integrated to generate tactical pattern matching data.

[0118] The tactical pattern matching data is evaluated based on the causal relationships in the tactical causal graph to generate tactical effect data; the tactical effect data is then structured and compiled for decision-making to output a set of tactical instructions for sports events.

[0119] Specifically, based on each matched tactical path in the tactical pattern matching data, the corresponding actual game result is searched in the real-time spatiotemporal hypergraph. The expected result node of the current path is compared with that in the tactical causal graph to determine whether the causal relationship is valid. The actual outcome, such as whether a point is scored, a mistake is made, or the player is defended, is recorded to form tactical effect data. The tactical name, causal achievement status, and game result of each record in the tactical effect data are mapped to a structured instruction template. Effective and high-success-rate tactical patterns are automatically compiled into executable natural language or coded instructions, such as organizing a high pick-and-roll followed by a pop shot or avoiding unscreened isolation plays inside the three-point line, and outputting a sports event tactical instruction set.

[0120] This embodiment also provides a sports event data mining system based on AI technology, including:

[0121] The acquisition module is used to collect multi-source sports event data, and uses a dynamic time warping algorithm to align the multi-source sports event data in time and space, and output multimodal event data packets.

[0122] The reporting module is used to identify sports entity coordinates and tactical events from multimodal event data packets, construct a spatiotemporal hypergraph, define meta-path templates for the spatiotemporal hypergraph, perform walk sampling and sequence pattern mining, and output a tactical pattern report.

[0123] The graph module is used to perform neural symbol fusion and instantiation transformation between tactical pattern reports and preset sports domain rules, output a preliminary neural graph, and complete the preliminary neural graph with knowledge to output a tactical knowledge graph.

[0124] The reasoning module is used to retrieve tactical instances from the tactical knowledge graph through logical deduction, and to perform causal reasoning analysis on the tactical instances to generate a tactical causal graph.

[0125] The instruction module is used to perform tactical pattern matching and effect evaluation between the real-time spatiotemporal hypergraph and the tactical causal graph, and output the tactical instruction set for sports events.

[0126] In summary, this invention achieves a deep transformation from multimodal data to high-level tactical decisions by constructing a reasonable tactical knowledge graph and a knowledge completion mechanism. By fusing and instantiating tactical pattern reports with preset sports domain rules through neural symbols, it deeply integrates data-driven complex patterns with symbolic domain logic, constructing a structured tactical knowledge system that combines statistical learning and logical reasoning capabilities. This solves the problem of the separation between data models and domain knowledge in traditional methods, significantly improving the interpretability and logical consistency of tactical representations. Furthermore, by performing knowledge completion on the initial neural graph based on the cosine similarity between rule-based vector representations, it automatically infers and adds semantically related entity nodes and node relationships, making the structure of the tactical knowledge graph more complete and rigorous, and enhancing the completeness of the knowledge system and the robustness of reasoning.

[0127] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for mining sports event data based on AI technology, characterized in that: include, Collect multi-source sports event data, align the multi-source sports event data in time and space using a dynamic time warping algorithm, and output a multimodal event data package; Identify sports entity coordinates and tactical events from multimodal event data packets, construct a spatiotemporal hypergraph, define meta-path templates for the spatiotemporal hypergraph, perform walk sampling and sequence pattern mining, and output a tactical pattern report; The tactical pattern report is fused and instantiated with the preset sports domain rules through neural symbol fusion, and a preliminary neural map is output. The preliminary neural map is then supplemented with knowledge to output a tactical knowledge map. Tactical instances in the tactical knowledge graph are retrieved through logical deduction, and causal reasoning analysis is performed on the tactical instances to generate a tactical causal graph. The system performs tactical pattern matching and effect evaluation by combining real-time spatiotemporal hypergraphs with tactical causal graphs, and outputs tactical instruction sets for sports events.

2. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The specific steps for outputting the multimodal event data packet are as follows: Establish a time reference axis for multi-source sports event data based on the timestamps of sports events; Multi-source sports event data is time-aligned and normalized according to a time baseline to generate spatiotemporally aligned event data. The spatiotemporally aligned event data is integrated and standardized under a unified spatiotemporal coordinate system to form a multimodal event data package.

3. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The specific steps for constructing the spatiotemporal hypergraph are as follows: Extract the spatiotemporal location information of athletes, referees, and the ball from multimodal event data packets, and integrate them to generate a set of sports entity coordinates; Identify the movement and interaction patterns of each sports entity in the sports entity coordinate set, and generate tactical event data; A spatiotemporal hypergraph is constructed by using sports entities as dynamic nodes in the sports entity coordinate set and the interaction relationships in the tactical event data as sports hyperedges.

4. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The specific steps for outputting the tactical mode report are as follows: Based on the relationship between dynamic nodes and sports hyperedges in the spatiotemporal hypergraph, a meta-path template reflecting tactical logic is defined. Biased transfer sampling is performed on the spatiotemporal hypergraph according to the meta-path template, and the dynamic nodes and sports hyperedges passed through in each sampling are recorded as a walking path sequence. Sequence pattern mining is performed on the wandering path sequences to identify high-frequency path patterns and integrate them to form a tactical pattern report.

5. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The specific steps for outputting the preliminary neural map are as follows: Extract the dynamic node sequence and sports superedge type of the path pattern in the tactical mode report, and convert the preset sports domain rules into a set of logical rules. The dynamic node sequence and sports hyperedge type are transformed into numerical vector representations, and the numerical vector representations are combined with the set of logical rules for constraint optimization to generate regularized vector representations. The regularized vector representation is decoded into corresponding entity nodes, node relationship edges, and relationship reasoning edges derived from logical rules, and then integrated to output a preliminary neural graph.

6. The sports event data mining method based on AI technology as described in claim 5, characterized in that: The specific steps for decoding the regularized vector representation into corresponding entity nodes, node relationship edges, and relationship reasoning edges derived from logical rules are as follows: The tensor decomposition method is used to separate the sub-vectors corresponding to entity features and the sub-vectors corresponding to relation features from the regularized vector representation, and then integrate them to generate a set of entity feature vectors and a set of relation feature vectors. Each entity feature vector in the entity feature vector set is mapped to an entity node with a unique identifier and attributes, generating an entity node set; Each relation feature vector in the relation feature vector set is mapped to a node relation edge connecting different entity nodes, and the relation type is labeled for the node relation edge according to the semantic information of the relation feature vector, thus generating a node relation edge set; The set of logical rules is applied to the set of entity nodes and the set of node relationship edges to perform logical derivation reasoning, generating a set of relationship reasoning edges; A preliminary neural graph is constructed by topologically connecting the set of entity nodes, the set of node relationship edges, and the set of relationship reasoning edges.

7. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The knowledge completion of the preliminary neural graph refers to adding entity nodes and node relationships to the preliminary neural graph based on the cosine similarity between regularized vector representations, and using a set of logical rules to verify the consistency of entity nodes and node relationships, thus forming a tactical knowledge graph with complete and rigorous structure and logic.

8. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The specific steps for generating the tactical causal graph are as follows: Acquire tactical query data, retrieve tactical instances that satisfy the tactical query data from the tactical knowledge graph through logical deduction, and integrate them to generate a set of tactical instances; Each tactical instance in the tactical instance set is analyzed to identify the dependencies between entity nodes and node relationship edges on the path from initial conditions to final result, and a causal path set is generated. The dependencies in the set of causal paths are encoded as causal relationships and integrated into the structure of the tactical knowledge graph to form a tactical causal graph.

9. The sports event data mining method based on AI technology as described in claim 1, characterized in that: The specific steps for outputting the sports event tactical instruction set are as follows: Tactical pattern similarity matching is performed between the real-time spatiotemporal hypergraph and the tactical causal graph to generate tactical pattern matching data. The tactical pattern matching data is evaluated based on the causal relationships in the tactical causal graph to generate tactical effect data. The tactical effect data is structured and compiled for decision-making, outputting a set of tactical instructions for sports events.

10. A sports event data mining system based on AI technology, based on the sports event data mining method based on AI technology according to any one of claims 1 to 9, characterized in that: include, The acquisition module is used to collect multi-source sports event data, and uses a dynamic time warping algorithm to align the multi-source sports event data in time and space, and output multimodal event data packets. The reporting module is used to identify sports entity coordinates and tactical events from multimodal event data packets, construct a spatiotemporal hypergraph, define meta-path templates for the spatiotemporal hypergraph, perform walk sampling and sequence pattern mining, and output a tactical pattern report. The graph module is used to perform neural symbol fusion and instantiation transformation between tactical pattern reports and preset sports domain rules, output a preliminary neural graph, and complete the preliminary neural graph with knowledge to output a tactical knowledge graph. The reasoning module is used to retrieve tactical instances from the tactical knowledge graph through logical deduction, and to perform causal reasoning analysis on the tactical instances to generate a tactical causal graph. The instruction module is used to perform tactical pattern matching and effect evaluation between the real-time spatiotemporal hypergraph and the tactical causal graph, and output the tactical instruction set for sports events.

Images