Big data-based software product information management system and method

Abstract

The application relates to the technical field of software information management, in particular to a software product information management system and method based on big data, which comprises the following steps: receiving heterogeneous original data streams of each link of the whole life cycle of software and performing standardization cleaning to generate multi-source information units; storing the units into a distributed graph database; constructing a global information knowledge graph taking software product entities as cores and the relationship between data as connecting edges; and realizing unified integration and dynamic correlation of multi-source data. Based on the topological structure of the graph, community discovery and other graph relationship mining algorithms are regularly executed to automatically identify function module clusters with high internal correlation, defect clusters with frequent linkage and user behavior pattern clusters. The method solves the information island problem by explicitly modeling the complex relationship network between data, and can automatically mine the hidden deep correlation and group mode in the data without relying on preset rules, so that data-driven deep insight and intelligent management of software products are realized.

Description

Technical Field

[0001] This invention relates to the field of software information management technology, and in particular to a software product information management system and method based on big data. Background Technology

[0002] Heterogeneous raw data streams are generated at each stage of the software product lifecycle, and current technologies manage this data using separate systems. Different systems have independent data formats and storage structures, resulting in weak connections between data. Conventional data integration methods, such as data warehouses or batch processing tools, rely on predefined data patterns and fixed association rules. This static integration approach cannot adapt to the continuously evolving nature of software information; logical and business relationships between data are often achieved through manual maintenance or simple labeling, lacking in-depth modeling of complex interactions. Therefore, unified representation and efficient querying of cross-system data face obstacles, and relationship networks are difficult to explicitly solidify.

[0003] At the data analysis and correlation mining level, existing technologies focus on processing single data sources or performing statistical correlation analysis on multi-source data. These methods require manually pre-setting association rules or thresholds to identify superficial connections between data points, failing to adaptively explore the inherent topological structure and cluster patterns of the data. Potentially close connections within software products, such as intrinsically dependent sets of functional modules, frequently interconnected defect groups, and regular clusters of user behavior, are therefore difficult to discover automatically and systematically. This limits the ability to extract deep knowledge from massive amounts of data, impacting data-driven product insights and decision-making.

[0004] Current technologies face two key challenges: the lack of a unified structured representation of relational networks in multi-source heterogeneous data hinders flexible relation queries and dynamic updates; and association mining relies excessively on predefined rules, failing to automatically identify hidden clusters of close associations based on the data's own topology. Solving these problems requires explicit modeling of data relationships and automated pattern discovery to support intelligent management of software product information. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a software product information management system and method based on big data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a software product information management method based on big data, comprising: Receive heterogeneous raw data streams from all stages of the software product's lifecycle, including code commit records, defect reports, user behavior logs, performance monitoring metrics, and market feedback text. Parallel preprocessing and standardization cleaning are performed on the heterogeneous raw data stream. The standardization cleaning includes deduplication, format conversion, outlier removal, and missing value imputation to generate standardized multi-source information units for software products. The standardized multi-source information units of the software product are stored in a distributed graph database, and a global information knowledge graph is constructed in the distributed graph database. The global information knowledge graph takes the software product entity as the core node, each information unit as the attribute node, and the logical and business relationships between data as the connection edges. Continuously monitor and capture newly generated software product information events, and dynamically update the corresponding nodes and connection edges in the global information knowledge graph for each newly generated software product information event. Based on the topology of the global information knowledge graph, graph relationship mining and community discovery algorithms are executed periodically to identify highly related functional module clusters, frequently linked defect clusters, and user behavior pattern clusters within the software product.

[0007] As a further aspect of the present invention, parallel preprocessing and standardization cleaning are performed on the heterogeneous raw data stream to generate standardized multi-source information units for software products, including: The code commit record is parsed to extract the commit author, modified file set, number of lines of code changed, commit timestamp and associated defect number, and the parsed information is encapsulated into structured commit information units; Extract the defect title, severity level, reproduction steps, module, assigned engineer, status transition record and repair version number from the defect report, and encapsulate the extracted information into structured defect information units; The user behavior logs are segmented into sessions and restored to identify the user's identity, operation sequence, page dwell time and function usage frequency, and the identified information is encapsulated into structured behavioral information units. From the performance monitoring metrics, CPU utilization, memory usage, network latency, error rate, and throughput are selected, and the collection time and server identifier are added to them, and they are encapsulated into structured performance information units. Sentiment analysis and topic extraction are performed on the market feedback text to identify the emotional polarity of the feedback, the core functional points mentioned, specific improvement suggestions or error descriptions, and the identified information is encapsulated into structured feedback information units.

[0008] As a further aspect of the present invention, a global information knowledge graph is constructed in the distributed graph database, including: In the distributed graph database, a root node representing the software product itself is created; Around the root node, create module child nodes representing the main functional modules, team child nodes representing the development team, version child nodes representing the release version, and user child nodes representing the user group; The submission information unit, defect information unit, behavior information unit, performance information unit, and feedback information unit in the standardized multi-source information unit of the software product are respectively mounted as attribute nodes to the global information knowledge graph; Between the root node, various child nodes, and various attribute nodes, directed connection edges are created based on the actual relationships between entities. The types of directed connection edges include belonging relationship, generating relationship, influencing relationship, repair relationship, triggering relationship, and feedback relationship. Each node and each edge in the global information knowledge graph is assigned a globally unique identifier and a time version label.

[0009] As a further aspect of the present invention, the step of dynamically updating each newly generated software product information event to the corresponding node and connection edge in the global information knowledge graph includes: When a newly generated software product information event is captured, it is transformed into a standardized multi-source information unit of the corresponding category of software product according to the type of the event; In the global information knowledge graph, search for existing entity nodes associated with the multi-source information units of the software product; If the associated existing entity node exists, a new version attribute node for the multi-source information unit of the software product will be created under the corresponding node, and the related connection edge attributes will be updated. If the associated existing entity node does not exist, a new entity node is created in the graph, and the multi-source information unit of the software product is attached as its attribute node, and a new connection edge is established between it and other related nodes. Record the timestamp and operation type of this update operation as a change log for the graph itself.

[0010] As a further aspect of the present invention, the periodic execution of graph relationship mining and community detection algorithms to identify highly interconnected functional module clusters, frequently linked defect clusters, and user behavior pattern clusters within the software product includes: Extract all module sub-nodes and their dependency and call relationship edges from the global information knowledge graph to form a module relationship subgraph; The label propagation algorithm is run on the module relationship subgraph to divide tightly coupled module nodes into the same functional module cluster based on the weight and density of the edges between nodes. All defect information unit nodes and their associated edges generated through common file modification or the same version of introduction are extracted from the global information knowledge graph to form a defect association subgraph; The community detection algorithm is run on the defect association subgraph to identify defect information unit node groups that are frequently repaired simultaneously or caused by the same root cause, forming defect clusters; Extract frequently co-occurring operation nodes from the global information knowledge graph and construct a user behavior co-occurrence matrix. Cluster analysis is performed on the user behavior co-occurrence matrix to group user groups with similar operating habits and paths into the same user behavior pattern cluster.

[0011] As a further aspect of the present invention, it also includes: using an offline batch processing computing framework to perform a full analysis of the historical state changes of all nodes and edges in the global information knowledge graph, and extracting the long-term trends, periodic patterns and abrupt change points of the evolution of software product information; By capturing dynamic event patterns with the offline batch processing computing framework through a real-time stream processing engine, the captured dynamic event patterns are fused and aligned with the long-term evolution patterns extracted by the offline batch processing computing framework to construct a dynamic prediction model of software product information status. The dynamic prediction model of the information status of the software product is used to quantitatively predict the future defect incidence rate of a specified functional module, the behavioral change trend of a specific user group, or the fluctuation range of key performance indicators. Based on the results of the quantitative prediction, a targeted software product information management instruction set is automatically generated. The software product information management instruction set includes code review suggestions, resource scheduling schemes, or user notification strategies. A feedback execution closed loop is established, and the actual execution effect data of the software product information management instruction set is transmitted back to the system as a new heterogeneous raw data stream for incremental learning and continuous optimization of the global information knowledge graph and the dynamic prediction model.

[0012] As a further aspect of the present invention, a full analysis is performed on the historical state changes of all nodes and edges in the global information knowledge graph to extract the long-term trends, periodic patterns, and abrupt change points of software product information evolution, specifically including: The global information knowledge graph is snapshotted according to time windows to form a series of graph state sequences sorted by time. By comparing the graph state sequences at adjacent time points, calculate the growth rate of lines of code for each functional module, the activity change of each defect cluster, and the average offset of each performance metric. Time series decomposition was performed on the code line count growth rate sequence of each functional module to separate its long-term trend component, seasonal cycle component and residual component; Using a mutation point detection algorithm, the residual components and defect cluster activity sequences are analyzed to identify time points where statistical characteristics change significantly, and these are marked as mutation points in the evolution of software product information.

[0013] As a further aspect of the present invention, the dynamic event patterns captured by the real-time stream processing engine are fused and aligned with the long-term evolution patterns extracted by the offline batch processing computing framework to construct a dynamic prediction model for the information status of software products, including: The real-time substructure of the global information knowledge graph is used as the dynamic graph input, and the long-term trend component and seasonal cycle component are used as temporal context features. Construct a hybrid model combining a graph neural network and a recurrent neural network, wherein the graph neural network is used to encode the topological and node features of the dynamic graph input, and the recurrent neural network is used to encode the temporal context features; The output of the graph neural network and the output of the recurrent neural network are concatenated and then input into a multilayer perceptron. Using historical snapshots and corresponding future real states as training samples, the hybrid model is trained with the goal of minimizing the difference between the predicted state and the real state, so that it becomes a dynamic prediction model of the information state of the software product. The quantitative prediction of the future defect incidence rate of a specified functional module, the behavioral change trend of a specific user group, or the fluctuation range of key performance indicators specifically includes: Extract the subgraph structure and node features of a specified functional module from the global information knowledge graph over a recent period; The subgraph structure and its node features, along with the long-term trends and seasonal cycle features of the functional modules, are input into the dynamic prediction model of the software product information status. The dynamic prediction model outputs the probability distribution of the number of new defect information units associated with the functional module within several future time windows, as a prediction of its future defect occurrence rate. Using the same process, by inputting the subgraph and features of a specific user behavior pattern cluster, we can obtain the trend vector of future operation path changes for the user group, which serves as a prediction of its behavior change trend. By inputting the characteristics of key performance indicator nodes and their related topologies, the confidence intervals of future values ​​of key performance indicators are obtained, which serve as predictions of their fluctuation range.

[0014] As a further aspect of the present invention, the step of automatically generating a targeted software product information management instruction set based on the results of the quantitative prediction includes: When the predicted future defect incidence rate exceeds a preset risk threshold, an instruction is generated that includes a list of modules requiring enhanced code review, a set of recommended test cases, and an early warning notification. When the predicted behavioral change trend indicates that a user group will use a new feature on a large scale, instructions are generated that include pre-allocating computing resources for the new feature, preparing user onboarding documentation, and adjusting load balancing strategies. When the predicted fluctuation range of the key performance indicator shows that its lower boundary touches the alarm line, an instruction is generated that includes activating the performance optimization plan, performing capacity assessment, and setting monitoring alarms. All generated instructions are sorted and packaged according to priority and execution time to form a set of instructions that can be directly called by external systems; The establishment of a feedback execution closed loop involves transmitting the actual execution effect data of the software product information management instruction set back to the system as a new heterogeneous raw data stream. This data is used for incremental learning and continuous optimization of the global information knowledge graph and the dynamic prediction model. Specifically, this includes: Deploy a data acquisition agent to continuously monitor and record the actual execution process and results of the software product information management instruction set in the external environment; New code commit records, defect status changes, user behavior responses, and performance metric changes generated during execution are input into the system again as new heterogeneous raw data streams. The global information knowledge graph is updated using the new heterogeneous raw data stream to include the new states generated after instruction execution. Using the updated global information knowledge graph and its historical sequence, the dynamic prediction model of the software product information status is incrementally trained, and the model parameters are adjusted to absorb the knowledge of system state changes caused by the execution of instructions.

[0015] As a further aspect of the present invention, the present invention also includes a software product information management system based on big data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the software product information management method based on big data as described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Standardized multi-source information units are stored in a distributed graph database, and connecting edges are constructed based on their inherent logic and business relationships to form a global information knowledge graph centered on software product entities. This technology enables unified and explicit modeling of all entities and their relationships throughout the software lifecycle. Complex relationships between various types of data are directly embedded in the storage structure, supporting real-time, flexible, and dynamic queries and updates of cross-stage and cross-type information. Data silos are broken down, and multi-source heterogeneous information forms an interconnected organic whole, thereby improving data consistency and the traceability of the evolution process.

[0017] Based on the topology of the constructed global information knowledge graph, graph relationship mining algorithms such as community detection are executed periodically. Based on the density and pattern of edges connecting nodes, highly cohesive clusters are automatically identified. This process does not rely on any preset association rules or manually defined pattern labels. The algorithm can autonomously reveal hidden, tightly coupled patterns from the inherent network structure of the data. This achieves purely data-driven, automated, deep mining of complex relationships and group patterns within software.

[0018] Through the construction of the dynamic knowledge graph and the automated relationship mining process described above, the management paradigm of software product information is transformed. Data is no longer an isolated record, but rather constitutes a continuously evolving relationship network. This network provides a solid structured data foundation for analyzing the internal structure of software, tracing the propagation of problems, and understanding user behavior patterns. The automatically discovered clusters of relationships provide objective insights into the data's own topology for product optimization and decision-making, enabling management activities to be based on real patterns emerging from the data rather than human assumptions. Attached Figure Description

[0019] Figure 1 This is a flowchart of the software product information management method based on big data as described in this invention; Figure 2 A flowchart for constructing a global information knowledge graph; Figure 3 A flowchart for dynamically updating information events related to new software products; Figure 4 Line chart for predicting high-frequency user behavior trends; Figure 5 A line graph illustrating the process of building a global information knowledge graph. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] See Figure 1 The system receives heterogeneous raw data streams from all stages of the software product's lifecycle. These streams include various data types such as code commit records, defect reports, user behavior logs, performance monitoring metrics, and market feedback text. The system then performs parallel preprocessing and standardization cleaning operations on these heterogeneous raw data streams. This standardization cleaning process includes data deduplication, format unification conversion, outlier removal, and missing value imputation, ultimately generating a series of standardized multi-source information units for the software product. These standardized information units are stored in a distributed graph database, where a global information knowledge graph is constructed. This global information knowledge graph uses the software product entity itself as the core root node, each information unit as an attribute node, and logical and business relationships between data as connecting edges, thus forming an interconnected data network. The system also deploys an event monitoring mechanism that continuously listens for and captures newly generated software product information events, dynamically updating the corresponding nodes and connecting edges in the global information knowledge graph after each new event is transformed, maintaining the graph's real-time performance. Based on the completed dynamic knowledge graph, the system periodically executes graph relationship mining and community discovery algorithms. By analyzing the graph's topology, it identifies highly interconnected functional module clusters, frequently linked defect clusters, and common user behavior pattern clusters within the software product.

[0023] See Figure 2In one embodiment of the present invention, the original code commit record received by the system may be a log entry containing commit information, a list of changed files, and modified content. A parsing operation is performed on the code commit record, and the parsed information includes the commit author's identity, a list of paths to the modified file set, the number of lines of code changed (added and deleted), the commit timestamp, and the defect number explicitly associated with this commit in the information. The processing procedure for defect reports is as follows: The original defect report may come from a project management tool and includes a title, description, attachments, and status transition history. An information extraction operation is performed on the defect report, and the extracted elements include the defect title text, severity level classification label, a list of steps for reproduction, the name of the module to which it belongs, the employee ID of the assigned engineer, the timestamp and status sequence of the status transition record, and the string of the fix version number. The extracted information is encapsulated into structured defect information units.

[0024] In specific implementation, the preprocessing process for user behavior logs is as follows: The original user behavior logs contain clickstream data generated by users accessing the server. Session segmentation and path restoration operations are performed on the user behavior logs. Session segmentation is based on the user session timeout mechanism, and path restoration concatenates discrete page access records. Through the session segmentation and path restoration operations, the anonymous ID of the user's identity, the page URL order of the operation sequence, the number of milliseconds of page dwell time, and the count of function usage frequency are identified. The identified information is encapsulated into structured behavior information units. The data structure of the behavior information unit includes "User ID: User_XYZ", "Session Sequence: / home-> / search-> / product / 1001", "Dwell Time: [1200,5000,8000]ms", and "Number of times Function A was used: 5".

[0025] In specific implementation, the processing procedure for performance monitoring metrics is as follows: The original performance monitoring metrics are a stream of values ​​periodically reported by the server cluster. From these metrics, the percentage of CPU utilization, bytes of memory usage, milliseconds of network latency, percentage of error rate, and requests per second (RPQ / s) of throughput are selected. Each selected metric is appended with a timestamp of its collection time and the IP address of the server that generated the metric. This data with added contextual information is encapsulated into structured performance information units. For example, a performance information unit might contain "CPU utilization: 65%", "Memory usage: 2.3GB", "Network latency: 150ms", "Error rate: 0.1%", "Throughput: 1200 requests / s", "Collection time: 2026-05-10 15:00:00", and "Server: 10.0.0.1".

[0026] In practice, the processing of market feedback text is as follows: The original market feedback text comes from app store reviews and customer service tickets. Sentiment analysis and topic extraction are performed on the market feedback text. Sentiment analysis determines the text's emotional polarity as positive, negative, or neutral. Topic extraction identifies the names of core functionalities mentioned in the text, as well as specific improvement suggestions or error descriptions. The identified information is encapsulated into structured feedback information units, each containing "Sentiment Polarity: Negative," "Core Functionality: Payment Function," and "Description: When using a bank card to pay, the page freezes at the last step and becomes unresponsive."

[0027] In practice, the process of constructing a global information knowledge graph is as follows: In the distributed graph database, a unique root node representing the software product itself is created. This root node has the attributes "Product ID: Product_X" and "Product Name: XX System". Around this root node, multiple module child nodes representing main functional modules, team child nodes representing the development team, version child nodes representing release versions, and user child nodes representing user groups are created. The various standardized multi-source information units of the software product generated previously, including submission information units, defect information units, behavior information units, performance information units, and feedback information units, are respectively attached as attribute nodes to the global information knowledge graph. Each attribute node is connected to its corresponding entity node through a directed edge. Directed connections are created between the root node, various child nodes, and various attribute nodes based on the actual relationships between entities. The types of directed connections include: "belonging" relationships (e.g., the module child node "M_Login" belongs to the root node "Product_X"); "generating" relationships (e.g., a commit operation generates a new commit information unit); "affecting" relationships (e.g., a defect may affect a functional module); "fixing" relationships (e.g., a code commit fixes a specific defect); "triggering" relationships (e.g., a user's operation triggers the collection of a performance metric); and "feedback" relationships (e.g., a user submits feedback regarding a functional module). Each node and each edge in the global information knowledge graph is assigned a globally unique identifier and a time version label. The globally unique identifier is a system-generated UUID string, and the time version label records the time when the node or edge was created or updated.

[0028] See Figure 3In one embodiment of the present invention, each newly generated software product information event is dynamically updated to the corresponding node and connection edge in the global information knowledge graph. The specific process is as follows: When the system's event listening mechanism captures a newly generated software product information event, such as a defect report titled "Payment Interface Timeout" being submitted, it is converted into a standardized multi-source information unit of the corresponding category of software product, i.e., a structured defect information unit, according to the event type. In the global information knowledge graph, the system searches for the existing entity node corresponding to the "Payment Gateway" module, i.e., the module child node "M_Payment", based on the "belonging module: Payment Gateway" attribute in the defect information unit. Since the associated existing entity node "M_Payment" exists, the system creates a new version attribute node for this new defect information unit under the "M_Payment" module child node, with the node identifier "Defect_Node_New", and creates a directed connection edge of "generation" relationship from the "M_Payment" node to the "Defect_Node_New" node, while updating the "most recently associated defect time" attribute of the "M_Payment" node itself. In practice, if the captured event is the first code commit for a brand-new functional module, "AI Recommendation Engine," searching for related existing entity nodes in the global information knowledge graph will reveal that there is no module child node named "AI Recommendation Engine." In this case, the system will create a new entity node in the global information knowledge graph, namely a new module child node "M_AI_Recommend." The commit information unit generated by this code commit will then be attached as an attribute node for the "M_AI_Recommend" node. Furthermore, a "belongs to" relationship connection will be established between the "M_AI_Recommend" node and the root node (software product), as well as a "responsible for" relationship connection between the "M_AI_Recommend" node and other relevant team child nodes. Regardless of whether a node is updated or created, the system will record the timestamp of this update operation, "2026-08-22 10:05:31," and the operation type, "node addition," as a change log for the graph itself.

[0029] In some embodiments, the process of periodically executing graph relationship mining and community discovery algorithms to identify functional module clusters is as follows: At a fixed time every Sunday, the system extracts all module child nodes and their dependency and call relationship edges from the global information knowledge graph, forming a module relationship subgraph. In the module relationship subgraph, each node represents a functional module such as "user authentication," "order processing," or "inventory query," and each edge represents the call relationship between modules and has a weight, the weight value of which is determined based on the call frequency. The label propagation algorithm is run on the module relationship subgraph. During the algorithm iteration, each module node determines its label assignment based on the labels of its neighboring nodes and the weights of the connecting edges. Based on the weights and connection density of the edges between nodes, tightly coupled module nodes are ultimately assigned to the same functional module cluster. For example, the three module nodes "shopping cart management," "coupon calculation," and "order generation" are assigned to the same functional module cluster called "transaction core" by the label propagation algorithm because the weights of the call edges between them are high and the connections are tight.

[0030] In some embodiments, the process of identifying frequently linked defect clusters is as follows: The system extracts all defect information unit nodes and their associated edges generated through shared file modifications or the same introduced version from the global information knowledge graph, forming a defect association subgraph. In the defect association subgraph, nodes represent specific defects. If two defects were historically fixed by the same code commit (i.e., associated with the same commit information unit node), then there is an associated edge between them. The community detection algorithm is run on the defect association subgraph. This algorithm searches for groups of nodes with tightly connected internal connections and sparse external connections based on the graph's topology, identifying groups of defect information unit nodes that are frequently fixed simultaneously or caused by the same root cause, forming defect clusters. A defect cluster identified by the community detection algorithm may contain defects "#BUG-1001 (payment timeout)," "#BUG-1002 (payment status out of sync)," and "#BUG-1005 (callback failure)," because they are all associated with the same problematic third-party payment service interface file.

[0031] The process of identifying user behavior pattern clusters can be understood as follows: The system extracts frequently co-occurring operation nodes from the global information knowledge graph and constructs a user behavior co-occurrence matrix. The rows and columns of the user behavior co-occurrence matrix represent different user operation nodes, such as "search for products," "view details," "add to cart," and "click to pay." The element values ​​in the matrix represent the frequency with which these two operations appear sequentially in the same session across a large number of user sessions. Cluster analysis is then performed on the user behavior co-occurrence matrix, for example, using the K-means clustering algorithm, to group user groups with similar operation habits and paths into the same user behavior pattern cluster. The user behavior pattern clusters that may be generated after cluster analysis include: "browsing user cluster," characterized by extremely high co-occurrence frequencies of "search for products" and "view details" in the operation sequence, but very few occurrences of "click to pay"; and "quick transaction user cluster," characterized by the dense co-occurrence of the operation nodes "search for products," "add to cart," and "click to pay" within a short period.

[0032] Optionally, during community discovery, a modularity metric can be used to evaluate the quality of partitioning the module relationship subgraph into functional module clusters. The formula for calculating the modularity metric is as follows: Where: symbol Value representing modularity; symbol This represents the weight of the edge between node i and node j in the adjacency matrix; the weight is 0 if there is no connection. and These represent the sum of the weights of all connecting edges between nodes i and j, respectively; (symbols omitted) Represents half the sum of the edge weights in the entire module relational subgraph; symbol It is a function whose value is 1 when node i and node j are assigned to the same functional module cluster, and 0 otherwise. Optimization improves the modularity metric. Maximizing the value can yield better community partitioning results for the module relationship subgraph.

[0033] In one embodiment of the present invention, an offline batch processing computing framework is used to perform a full analysis of the historical state changes of all nodes and edges in the global information knowledge graph, extracting the long-term trends, periodic patterns, and abrupt change points of software product information evolution. The offline batch processing computing framework starts during the low-load period in the early morning of each day, using a complete state snapshot of the global information knowledge graph at 0:00 every day for the past 365 days as input data. It calculates the monthly cumulative value of the number of submitted information units attached to each functional module node, the average time taken for the status of active defect information units within each defect cluster to change to "closed," and the daily average response time of key performance indicator nodes. Through the offline batch processing computing framework's full analysis of these time-dimension sequences, slowly changing long-term trend components, seasonal periodic components with weekly or monthly cycles, and residual components representing sudden fluctuations are separated. The offline batch processing computing framework uses a change point detection algorithm to scan the residual component sequence, identifying time points where statistical characteristics change significantly. For example, around July 15, 2026, the code submission activity residual of the payment gateway module showed a large positive shift, and this time point was marked as an abrupt change point.

[0034] In some embodiments, a dynamic prediction model for the information status of a software product is constructed by fusing and aligning dynamic event patterns captured by a real-time stream processing engine with long-term evolution patterns extracted by an offline batch processing computing framework. The real-time stream processing engine continuously monitors the message queue, capturing dynamic events such as "a surge in user logins," "new defect report submissions," and "server CPU utilization exceeding a threshold," and transforms these events into real-time update operations on the substructure of the global information knowledge graph. The long-term trend components and seasonal periodic components extracted by the offline batch processing computing framework serve as the temporal context features input to the model. A hybrid model combining a graph neural network and a recurrent neural network is constructed. The graph neural network encodes the topological and node features of the dynamic knowledge graph substructure updated by real-time events, while the recurrent neural network encodes the historical sequence patterns of the temporal context features. The output vectors of the graph neural network and the recurrent neural network are concatenated and input into a multilayer perceptron for dimensionality reduction and mapping. Using historical graph snapshots and their corresponding future real states as training samples, for example, using daily graph snapshots from January to June 2026 as input features, the graph state changes for the next 7 days from January to June 2026 as the prediction target, and mean squared error as the loss function, a multilayer perceptron and its upstream graph neural network and recurrent neural network are trained through backpropagation algorithm, so that the hybrid model can minimize the difference between the predicted state and the real state. The trained hybrid model becomes the dynamic prediction model of the information state of the software product.

[0035] This can be understood as the process of quantitatively predicting the future defect incidence rate of a specified functional module using a dynamic prediction model of the software product's information status. The subgraph structure and node characteristics of the specified functional module "User Authentication" for the past 30 days are extracted from the global information knowledge graph. The subgraph contains the "User Authentication" module node, its connected defect information unit nodes, submission information unit nodes, and the edges between them. The subgraph structure and node characteristics, along with the long-term trend of the number of defects in the "User Authentication" module over the past year and its quarterly seasonal cyclical characteristics obtained from offline batch processing analysis, are input into the trained dynamic prediction model of the software product's information status. The dynamic prediction model outputs a sequence of Poisson distribution parameters λ for the number of new defect information units that the "User Authentication" module may be associated with each day for the next 7 days, for example, [0.5, 0.7, 1.2, 1.1, 0.9, 0.6, 0.4]. This probability distribution serves as the prediction of its future defect incidence rate. Using the same process, inputting the subgraph and historical behavioral characteristics of a specific user behavior pattern cluster "high-value customers," the dynamic prediction model for the software product information status outputs a vector of the estimated number of times this group visits the "Advanced Features" page daily in the coming week [152, 148, 155, 162, 158, 145, 150], as a prediction of their behavioral change trend. Inputting the key performance indicator "database query latency" node and its related topology characteristics, the model outputs a 95% confidence interval of [120ms, 180ms] for the indicator's value in the next 24 hours, as a prediction of its fluctuation range.

[0036] In some embodiments, a targeted software product information management instruction set is automatically generated based on the results of quantitative prediction. When the predicted future defect incidence rate of the "User Authentication" module shows that the λ value of 1.2 on the third day exceeds the preset risk threshold of 1.0, the system automatically generates an instruction containing a list of modules requiring enhanced code review ["M_Login", "M_AuthToken"], a set of test cases recommended for the "Password Reset Process" related interfaces ["TC_AUTH_101", "TC_AUTH_105"], and a template for sending a warning notification to the development team. When the predicted behavioral trend of the "High-Value Customer" group indicates that the estimated number of times this group visits the "Advanced Features" page on the fourth day is 162, a significant increase from the baseline of 120 times, the system generates an instruction containing pre-allocating 20% ​​additional computing resources to the "Advanced Features" page, preparing a user guide pop-up document for this feature, and increasing the load balancing strategy weight by 15%. When the predicted fluctuation range of the "database query latency" metric indicates that its lower boundary of 120ms has reached the preset alarm threshold of 120ms, the system generates instructions. These instructions include activating "Database Index Optimization Plan 004," performing "Database Sharding and Table Sharding Capacity Assessment," and temporarily lowering the monitoring alarm threshold from 150ms to 110ms. All generated instructions are sorted and packaged according to preset priority rules and calculated execution times, forming a JSON-formatted instruction set that can be directly invoked by external operation and maintenance systems or project management tools.

[0037] Optionally, a feedback execution loop can be established, sending the actual execution effect data of the software product information management instruction set back to the system as a new heterogeneous raw data stream. A data acquisition agent deployed in the external environment monitors and records the execution process of the instruction set. For example, it records that "Database Index Optimization Plan 004" was executed from 03:00 to 05:00 on August 25, 2026, after which the CPU utilization of the database master node decreased from 85% to 65%. The data acquisition agent encapsulates new code commit records, defect status changes, user behavior responses, and performance indicator changes generated during the execution process into a new heterogeneous raw data stream and pushes it to the system's message queue entry. The system receives these new heterogeneous raw data streams again, transforms them into standardized multi-source information units of the software product, and dynamically updates the global information knowledge graph, ensuring that the global information knowledge graph includes the new performance information unit nodes generated after "Database Index Optimization" and their connection edges with related modules. The updated global information knowledge graph, which includes the new state after instruction execution and its historical sequence, is used to incrementally train the dynamic prediction model of the software product information state. The incremental training uses a small learning rate and adjusts the weights based on the existing model parameters, so that the dynamic prediction model of the software product information state can absorb the knowledge of system state changes caused by the execution of the "database index optimization" instruction.

[0038] It is understandable that in offline full-data analysis, an additive model can be used to decompose time series data to extract periodic patterns, and its formula is expressed as: Where: symbol Represents a point in time Observed raw sequence values, such as the number of code commits for a specific functional module in a given month; symbols Represents a point in time The long-term trend component reflects the long-term upward or downward direction of the sequence; the symbol... Represents a point in time The seasonal periodic component reflects fluctuations that repeat at fixed periods in the sequence; symbol Represents a point in time The residual component is the irregular variation that remains after removing trends and cycles.

[0039] In one embodiment of the present invention, snapshots of the global information knowledge graph are taken according to time windows to form a series of graph state sequences ordered by time. The time window is set to midnight every day. The system automatically captures the complete state of the global information knowledge graph and stores it as a snapshot. For example, for the functional module "Payment Gateway", the daily code lines, number of submitted information units, number of defect information units, and average response time are recorded in the snapshot sequence of the first week of August 2026. By comparing the graph state sequences of adjacent time points, the growth rate of code lines for each functional module, the change in activity of each defect cluster, and the average offset of each performance indicator are calculated. The growth rate of code lines is obtained by dividing the difference between the code lines of the current day and the code lines of the previous day by the code lines of the previous day. The change in activity of defect clusters is calculated by the difference in the number of defects with the status of "open" within the cluster. The average offset of performance indicators is obtained by the absolute difference between the average values ​​of indicators of two adjacent days. See Table 1.

[0040] Table 1: Status Sequence Data of the Functional Module "Payment Gateway" in the First Week of August 2026 Based on the data in Table 1, the code line count growth rate of the functional module "Payment Gateway" from August 1, 2026 to August 2, 2026 was (15200-15000) / 15000≈1.33%, the defect cluster activity change was (4-3) / 3≈33.33%, and the average response time offset was |118-120|=2ms. A time series decomposition was performed on the code line count growth rate sequence of each functional module to separate its long-term trend component, seasonal periodic component, and residual component. Using a seasonal decomposition method, the code line count growth rate of the functional module "Payment Gateway" from January to July 2026 was analyzed. The monthly daily code line growth rate sequence was decomposed to obtain a long-term trend component showing a slow increase, a seasonal cycle component showing that weekly code submission activity peaked on Wednesdays, and a residual component reflecting random fluctuations. Using a mutation point detection algorithm, the residual components and defect cluster activity sequences were analyzed to identify time points where statistical characteristics changed significantly, which were marked as mutation points in the evolution of software product information. For example, applying the mutation point detection algorithm to the residual component of the code line growth rate of the functional module "Payment Gateway" identified a mutation point on June 15, 2026, after which the mean of the residual sequence shifted significantly.

[0041] In some embodiments, dynamic event patterns captured by the real-time stream processing engine are fused and aligned with long-term evolution patterns extracted by the offline batch processing computing framework to construct a dynamic prediction model of the software product's information status. The real-time stream processing engine continuously monitors the message queue and captures events such as "cache service error rate spike," which are then transformed into real-time updates to the "cache service" module nodes in the global information knowledge graph. The long-term trend components and seasonal periodic components extracted by the offline batch processing computing framework are used as temporal context features. For example, the long-term trend of the "cache service" module error rate is a 2% decrease per month, and the seasonal period is a higher error rate every evening. The real-time substructure of the global information knowledge graph is used as the dynamic graph input, including, for example, the "cache service" module nodes and their associated performance indicator nodes, and the error rate spikes. For the subgraph of the trapped node, a hybrid model combining a graph neural network and a recurrent neural network is constructed. The graph neural network uses a graph convolutional network to encode the topological and node features of the dynamic graph input, while the recurrent neural network uses a gated recurrent unit network to encode the temporal context features. The output vectors of the graph neural network and the recurrent neural network are concatenated and input into a multilayer perceptron. Historical graph snapshots and corresponding future real states are used as training samples. For example, daily graph snapshots from March 1, 2026 to August 31, 2026 are used as input features, and the graph state of the next three days is used as the prediction target. The mean absolute error is used as the loss function to train the hybrid model, minimizing the difference between the predicted state and the real state. After training, a dynamic prediction model of the software product information state is obtained.

[0042] It is understandable that quantitative predictions are made regarding the future defect incidence rate of a specified functional module, the behavioral trends of a specific user group, or the fluctuation range of key performance indicators. This involves extracting the subgraph structure and node characteristics of the "Order Processing" functional module from the global information knowledge graph for the past thirty days. The subgraph contains "Order Processing" module nodes, connected defect information unit nodes, submission information unit nodes, and connecting edges. The subgraph structure and node characteristics, along with the long-term trend and seasonal cyclical characteristics of the "Order Processing" functional module, are input into a dynamic prediction model of the software product's information status. The dynamic prediction model outputs the probability score of the number of new defect information units associated with the "Order Processing" functional module within the next five working days. For example, the sequence of Poisson distribution parameters λ [1.1, 1.3, 1.0, 0.8, 0.9] is output as a prediction of its future defect incidence rate. Using the same process, the subgraph and features of a specific user behavior pattern cluster "high-frequency users" are input to obtain the trend vector of the user group's future operation path changes. For example, the estimated number of times the "Personal Center" page is accessed daily for the next five days [210, 205, 215, 208, 212] is used as a prediction of its behavior change trend. The features of the key performance indicator node "API gateway latency" and its related topology are input to obtain the confidence interval of the future value of the key performance indicator. For example, the 90% confidence interval of the latency for the next six hours is [95ms, 110ms], which is used as a prediction of its fluctuation range.

[0043] See Figure 4 This is a line chart predicting high-frequency user behavior trends, clearly showing the changing patterns of three core user operation types over the next five days. It represents a visualization result from the user behavior pattern cluster analysis phase in software product information management. The number of visits to the personal center fluctuates at a high level overall, first slightly decreasing then significantly increasing, then slightly decreasing again before rising again, consistently remaining the highest among the three operation types, reaching its peak on the third day. The number of order queries first decreases then increases, then slightly decreases again before rising slightly, bottoming out on the second day and rising to the second-highest point on the third day. The trend of payment operations is highly synchronized with order queries, with a slightly smaller amplitude, and remains the lowest among the three operation types, exhibiting the same fluctuation pattern. All three operation types will reach their peak on the third day, allowing for targeted expansion of server resources for relevant functional modules in advance to avoid performance bottlenecks.

[0044] In one embodiment of the present invention, a targeted software product information management instruction set is automatically generated based on the results of quantitative prediction. When the predicted value of the future defect occurrence rate exceeds a preset risk threshold, an instruction containing a list of modules that need to strengthen code review, a set of test cases to be executed, and an early warning notification is generated. For example, the dynamic prediction model of the software product information status outputs the defect occurrence rate prediction value of the functional module "Message Push" for the next three days as [1.5, 1.8, 2.1]. The preset risk threshold is 1.5 defects per day. Since the predicted values ​​for the second and third days exceed the threshold, the system automatically generates an instruction. The module list in the instruction content is ["Message Queue Processing Service", "Push Gateway Service"], the test case set to be executed is identified as ["TEST_PUSH_STRESS_01", "TEST_PUSH_FAILOVER_03"], and the recipients of the early warning notification are "Backend Development Team A" and "Test Team B". When behavioral trend predictions indicate that a user group will use a new feature on a large scale, instructions are generated that include pre-allocating computing resources for the new feature, preparing user onboarding documentation, and adjusting load balancing strategies. For example, if behavioral trend predictions for the user group "internal beta testers" show that the number of requests to access the newly launched feature "intelligent voice assistant" is expected to increase by 300% in the next 48 hours, the system generates an instruction to pre-allocate 20 additional container instances to the "intelligent voice assistant" backend service, prepare and release version V1.2 of the user onboarding documentation and frequently asked questions, and increase the weight of the "intelligent voice assistant" service cluster in the load balancer strategy from 10% to 30%. When the predicted fluctuation range of a key performance indicator shows that its lower boundary is approaching the alarm line, an instruction is generated that includes initiating a performance optimization plan, performing a capacity assessment, and setting monitoring alarms. For example, if the predicted fluctuation range of the key performance indicator "database master connection pool utilization" shows that its lower boundary for the next 12 hours is 88%, while the preset alarm line is 90%, since the predicted lower boundary is close to the alarm line, the system generates an instruction to immediately initiate the performance optimization plan numbered "PERF_DB_CONN_OPT_004". This plan includes cleaning up idle connections and adjusting timeout parameters, performing a capacity assessment scan of the database connection pool, and temporarily adjusting the threshold of real-time monitoring alarms from 90% to 85%. All generated instructions are sorted and packaged according to priority and execution time to form a set of instructions that can be directly called by external systems. The priority of the instructions is calculated based on the severity and scope of the predicted threshold, and the execution time is determined based on the instruction type and system maintenance window. The final packaged instruction set is a structured JSON array that can be directly called and executed by external continuous integration / continuous deployment platforms or operation and maintenance automation systems through RESTful API.

[0045] In some embodiments, a feedback execution closed loop is established, which transmits the actual execution effect data of the software product information management instruction set as a new heterogeneous raw data stream back to the system for incremental learning and continuous optimization of the global information knowledge graph and dynamic prediction model. A data acquisition agent is deployed to continuously monitor and record the actual execution process and results of the software product information management instruction set in the external environment. The data acquisition agent can be a lightweight daemon deployed on the application server, which listens to instruction execution logs, system monitoring interfaces, and callbacks from deployment tools. For example, it records that the instruction "Start performance optimization plan PERF_DB_CONN_OPT_004" was triggered at 03:00 on 2026-09-10 and completed execution at 03:25. After execution, the "database master connection pool utilization rate" metric dropped to 78% at 04:00. New code commit records, defect status changes, user behavior responses, and performance metric changes generated during execution are input into the system as new heterogeneous raw data streams. For example, the execution of a performance optimization plan triggers a code commit for a configuration change, with the commit message "Optimize database connection pool parameters." Simultaneously, the "database master connection pool utilization rate" performance metric generates new time-series data. This data is encapsulated by the data acquisition agent into new code commit records and performance monitoring metric data streams, and pushed back to the system's data receiving endpoint. The new heterogeneous raw data streams are used to update the global information knowledge graph, making it include the new states generated after the command execution. After receiving the code commit record for "Optimize database connection pool parameters," the system parses it into new commit information units according to the standard process, finds the corresponding "database service" module sub-node in the global information knowledge graph, creates new commit information unit attribute nodes under it, and establishes connection edges. Using the updated global information knowledge graph and its historical sequences, the dynamic prediction model of the software product information state is incrementally trained to adjust the model parameters and absorb the knowledge of system state changes caused by the execution of instructions. The incremental training process uses a smaller learning rate than the original training and takes the graph snapshot sequence containing the new state after instruction execution as part of the training samples and inputs it into the dynamic prediction model of the software product information state. The model parameters are fine-tuned through backpropagation so that the model can take into account the possible impact of similar performance optimization instructions when making predictions in the future.

[0046] Optionally, a priority scoring function can be defined during instruction sorting to determine the execution order of instructions. The formula for calculating the priority scoring function is as follows: Where: symbol Represents the final priority score of an instruction; the higher the score, the higher the priority. (Symbol) This represents a quantitative prediction value, such as the specific value of 1.8 for the predicted future defect incidence rate; [symbol] This represents the corresponding preset risk threshold of 1.5; symbol This represents an estimated number of modules or users affected by the instruction; symbol Represents the total number of system modules or the total number of users; symbol Represents the time interval from the current time to when the predicted problem may occur; symbol Represents a maximum time interval constant; symbol , and These are weighting coefficients used to balance the contributions of the three dimensions: the degree of exceeding the predicted limit, the scope of impact, and the urgency. This is achieved by calculating the priority score for each generated instruction. The system can standardize the sorting of all instructions.

[0047] In some embodiments, the data acquisition agent records the instruction execution process and results in the following ways: When an external system begins executing an instruction, the data acquisition agent records the instruction's unique identifier, execution start timestamp, and execution environment context; during instruction execution, the data acquisition agent captures system change events triggered by instruction execution, such as a configuration release, server expansion, or user notification sending, through hook functions or log parsing; after instruction execution is completed, the data acquisition agent collects key result indicators, such as the performance indicator difference after the execution of performance optimization instructions, the resource utilization change after the execution of resource scheduling instructions, and the click-through rate of user notification instructions. The data acquisition agent packages these structured process and result data, along with the original instruction content, into a feedback record. The feedback record is sent back to the central management system through a secure transmission channel. After parsing the feedback record, the central management system injects the valid data portion, such as new performance indicator data and new user behavior logs, into the original heterogeneous raw data stream entry point, thereby initiating a new round of data processing and knowledge update cycle.

[0048] It is understandable that the incremental learning and continuous optimization of the global information knowledge graph and the dynamic prediction model is an iterative process. Each execution and feedback of the software product information management instruction set generates new data, which is used to update the global information knowledge graph. The updated global information knowledge graph reflects the latest state of the system and contains new causal relationships, such as "after executing optimization instruction A, indicator B improved." By using the historical snapshot sequence of the graph containing new causal relationships to incrementally train the dynamic prediction model, the model can gradually learn the correlation between instruction intervention and changes in system state. Thus, when making predictions in the future, it can not only base them on natural evolution laws but also take potential intervention measures into account as influencing factors, making predictions and recommendations more adaptive and accurate.

[0049] See Figure 5 This is a line graph showing the construction process of a global information knowledge graph, clearly illustrating the changes in graph size and construction efficiency over five weeks. It represents a visualization of the graph construction and dynamic update phase in software product information management. The number of nodes grew linearly and steadily, increasing from approximately 1,000 in week 1 to approximately 10,000 in week 5, a tenfold increase. The number of edges grew superlinearly and rapidly, with a growth rate significantly exceeding that of nodes, increasing from approximately 2,000 in week 1 to approximately 30,000 in week 5, a fifteenfold increase. Construction time was highly synchronized with the number of edges, exhibiting superlinear growth, increasing from approximately 60 minutes in week 1 to approximately 300 minutes in week 5, a fivefold increase. This superlinear growth in the number of edges indicates that as the number of nodes increases, the density of business relationships between entities significantly improves, and the information richness and complexity of the graph rapidly increase.

[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A software product information management method based on big data, characterized in that, include: Receive heterogeneous raw data streams from all stages of the software product's lifecycle, including code commit records, defect reports, user behavior logs, performance monitoring metrics, and market feedback text. Parallel preprocessing and standardization cleaning are performed on the heterogeneous raw data stream. The standardization cleaning includes deduplication, format conversion, outlier removal, and missing value imputation to generate standardized multi-source information units for software products. The standardized multi-source information units of the software product are stored in a distributed graph database, and a global information knowledge graph is constructed in the distributed graph database. The global information knowledge graph takes the software product entity as the core node, each information unit as the attribute node, and the logical and business relationships between data as the connection edges. Continuously monitor and capture newly generated software product information events, and dynamically update the corresponding nodes and connection edges in the global information knowledge graph for each newly generated software product information event. Based on the topology of the global information knowledge graph, graph relationship mining and community discovery algorithms are executed periodically to identify highly related functional module clusters, frequently linked defect clusters, and user behavior pattern clusters within the software product.

2. The software product information management method based on big data as described in claim 1, characterized in that, Parallel preprocessing and standardization cleaning are performed on the heterogeneous raw data streams to generate standardized multi-source information units for software products, including: The code commit record is parsed to extract the commit author, modified file set, number of lines of code changed, commit timestamp and associated defect number, and the parsed information is encapsulated into structured commit information units; Extract the defect title, severity level, reproduction steps, module, assigned engineer, status transition record and repair version number from the defect report, and encapsulate the extracted information into structured defect information units; The user behavior logs are segmented into sessions and restored to identify the user's identity, operation sequence, page dwell time and function usage frequency, and the identified information is encapsulated into structured behavioral information units. From the performance monitoring metrics, CPU utilization, memory usage, network latency, error rate, and throughput are selected, and the collection time and server identifier are added to them, and they are encapsulated into structured performance information units. Sentiment analysis and topic extraction are performed on the market feedback text to identify the emotional polarity of the feedback, the core functional points mentioned, specific improvement suggestions or error descriptions, and the identified information is encapsulated into structured feedback information units.

3. The software product information management method based on big data as described in claim 2, characterized in that, Construct a global information knowledge graph in the distributed graph database, including: In the distributed graph database, a root node representing the software product itself is created; Around the root node, create module child nodes representing the main functional modules, team child nodes representing the development team, version child nodes representing the release version, and user child nodes representing the user group; The submission information unit, defect information unit, behavior information unit, performance information unit, and feedback information unit in the standardized multi-source information unit of the software product are respectively mounted as attribute nodes to the global information knowledge graph; Between the root node, various child nodes, and various attribute nodes, directed connection edges are created based on the actual relationships between entities. The types of directed connection edges include belonging relationship, generating relationship, influencing relationship, repair relationship, triggering relationship, and feedback relationship. Each node and each edge in the global information knowledge graph is assigned a globally unique identifier and a time version label.

4. The software product information management method based on big data as described in claim 3, characterized in that, The step of dynamically updating each newly generated software product information event to the corresponding node and connection edge in the global information knowledge graph includes: When a newly generated software product information event is captured, it is transformed into a standardized multi-source information unit of the corresponding category of software product according to the type of the event; In the global information knowledge graph, search for existing entity nodes associated with the multi-source information units of the software product; If the associated existing entity node exists, a new version attribute node for the multi-source information unit of the software product will be created under the corresponding node, and the related connection edge attributes will be updated. If the associated existing entity node does not exist, a new entity node is created in the graph, and the multi-source information unit of the software product is attached as its attribute node, and a new connection edge is established between it and other related nodes. Record the timestamp and operation type of this update operation as a change log for the graph itself.

5. The software product information management method based on big data as described in claim 4, characterized in that, The periodically executed graph relationship mining and community detection algorithms identify highly interconnected functional module clusters, frequently linked defect clusters, and user behavior pattern clusters within the software product, including: Extract all module sub-nodes and their dependency and call relationship edges from the global information knowledge graph to form a module relationship subgraph; The label propagation algorithm is run on the module relationship subgraph to divide tightly coupled module nodes into the same functional module cluster based on the weight and density of the edges between nodes. All defect information unit nodes and their associated edges generated through common file modification or the same version of introduction are extracted from the global information knowledge graph to form a defect association subgraph; The community detection algorithm is run on the defect association subgraph to identify defect information unit node groups that are frequently repaired simultaneously or caused by the same root cause, forming defect clusters; Extract frequently co-occurring operation nodes from the global information knowledge graph and construct a user behavior co-occurrence matrix. Cluster analysis is performed on the user behavior co-occurrence matrix to group user groups with similar operating habits and paths into the same user behavior pattern cluster.

6. The software product information management method based on big data as described in claim 5, characterized in that, It also includes: using an offline batch processing computing framework to perform a full analysis of the historical state changes of all nodes and edges in the global information knowledge graph, and extracting the long-term trends, periodic patterns and abrupt change points of the evolution of software product information; By capturing dynamic event patterns with the offline batch processing computing framework through a real-time stream processing engine, the captured dynamic event patterns are fused and aligned with the long-term evolution patterns extracted by the offline batch processing computing framework to construct a dynamic prediction model of software product information status. The dynamic prediction model of the information status of the software product is used to quantitatively predict the future defect incidence rate of a specified functional module, the behavioral change trend of a specific user group, or the fluctuation range of key performance indicators. Based on the results of the quantitative prediction, a targeted software product information management instruction set is automatically generated. The software product information management instruction set includes code review suggestions, resource scheduling schemes, or user notification strategies. A feedback execution closed loop is established, and the actual execution effect data of the software product information management instruction set is sent back to the system as a new heterogeneous raw data stream for incremental learning and continuous optimization of the global information knowledge graph and the dynamic prediction model.

7. The software product information management method based on big data as described in claim 6, characterized in that, A full analysis of the historical state changes of all nodes and edges in the global information knowledge graph is performed to extract the long-term trends, periodic patterns, and abrupt change points in the evolution of software product information, specifically including: The global information knowledge graph is snapshotted according to time windows to form a series of graph state sequences sorted by time. By comparing the graph state sequences at adjacent time points, calculate the growth rate of lines of code for each functional module, the activity change of each defect cluster, and the average offset of each performance metric. Time series decomposition was performed on the code line count growth rate sequence of each functional module to separate its long-term trend component, seasonal cycle component and residual component; Using a mutation point detection algorithm, the residual components and defect cluster activity sequences are analyzed to identify time points where statistical characteristics change significantly, and these are marked as mutation points in the evolution of software product information.

8. The software product information management method based on big data as described in claim 7, characterized in that, By fusing and aligning the dynamic event patterns captured by the real-time stream processing engine with the long-term evolution patterns extracted by the offline batch processing computing framework, a dynamic prediction model for the information status of software products is constructed, including: The real-time substructure of the global information knowledge graph is used as the dynamic graph input, and the long-term trend component and seasonal cycle component are used as temporal context features. Construct a hybrid model combining a graph neural network and a recurrent neural network, wherein the graph neural network is used to encode the topological and node features of the dynamic graph input, and the recurrent neural network is used to encode the temporal context features; The output of the graph neural network and the output of the recurrent neural network are concatenated and then input into a multilayer perceptron. Using historical snapshots and corresponding future real states as training samples, the hybrid model is trained with the goal of minimizing the difference between the predicted state and the real state, so that it becomes a dynamic prediction model of the information state of the software product. The quantitative prediction of the future defect incidence rate of a specified functional module, the behavioral change trend of a specific user group, or the fluctuation range of key performance indicators specifically includes: Extract the subgraph structure and node features of a specified functional module from the global information knowledge graph over a recent period; The subgraph structure and its node features, along with the long-term trends and seasonal cycle features of the functional modules, are input into the dynamic prediction model of the software product information status. The dynamic prediction model outputs the probability distribution of the number of new defect information units associated with the functional module within several future time windows, as a prediction of its future defect occurrence rate. Using the same process, by inputting the subgraph and features of a specific user behavior pattern cluster, we can obtain the trend vector of future operation path changes for the user group, which serves as a prediction of its behavior change trend. By inputting the characteristics of key performance indicator nodes and their related topologies, the confidence intervals of future values ​​of key performance indicators are obtained, which serve as predictions of their fluctuation range.

9. The software product information management method based on big data as described in claim 8, characterized in that, Based on the results of the quantitative prediction, a targeted software product information management instruction set is automatically generated, including: When the predicted future defect incidence rate exceeds a preset risk threshold, an instruction is generated that includes a list of modules requiring enhanced code review, a set of recommended test cases, and an early warning notification. When the predicted behavioral change trend indicates that a user group will use a new feature on a large scale, instructions are generated that include pre-allocating computing resources for the new feature, preparing user onboarding documentation, and adjusting load balancing strategies. When the predicted fluctuation range of the key performance indicator shows that its lower boundary touches the alarm line, an instruction is generated that includes activating the performance optimization plan, performing capacity assessment, and setting monitoring alarms. All generated instructions are sorted and packaged according to priority and execution time to form a set of instructions that can be directly called by external systems; The establishment of a feedback execution closed loop involves transmitting the actual execution effect data of the software product information management instruction set back to the system as a new heterogeneous raw data stream. This data is used for incremental learning and continuous optimization of the global information knowledge graph and the dynamic prediction model. Specifically, this includes: Deploy a data acquisition agent to continuously monitor and record the actual execution process and results of the software product information management instruction set in the external environment; New code commit records, defect status changes, user behavior responses, and performance metric changes generated during execution are input into the system again as new heterogeneous raw data streams. The global information knowledge graph is updated using the new heterogeneous raw data stream to include the new states generated after instruction execution. Using the updated global information knowledge graph and its historical sequence, the dynamic prediction model of the software product information status is incrementally trained, and the model parameters are adjusted to absorb the knowledge of system state changes caused by the execution of instructions.

10. A software product information management system based on big data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the software product information management method based on big data as described in any one of claims 1 to 9.

Images