User experience improvement system based on large model technology

Abstract

The invention discloses a user experience improvement system based on a large model technology, and relates to the technical field of artificial intelligence, and the system comprises a data processing unit which is used for obtaining an interaction operation record of a target application, constructing a user operation sequence arranged according to a time sequence according to a session identifier, and obtaining a labeled user operation sequence set; the navigation path expansion unit is used for extracting an interface identification sequence from the labeled user operation sequence set to form a navigation path set, so as to obtain a navigation path expansion set; the topology analysis unit is used for constructing a path structure statistical table on the basis of the navigation path extension set and generating reconstruction suggestion elements for each topology causal problem path cluster; and the large model optimization unit is used for inputting the reconstruction suggestion elements and the path structure statistical table into a large model optimization module, generating a prototype interface file and outputting the prototype interface file for improving user experience. According to the method, the experience problem source can be accurately positioned, and the structured ability and causal judgment ability of path analysis are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a user experience improvement system based on large model technology. Background Technology

[0002] With the widespread adoption of mobile internet and smart devices, numerous applications are used extensively, and users frequently perform tasks in various business scenarios, such as setting personal information, completing payment processes, submitting business applications, and viewing processing results. To improve user experience, various applications commonly employ methods such as event tracking, log analysis, and user satisfaction surveys to monitor and evaluate user behavior. Some existing technologies use visual path analysis methods to display the user's navigation relationships between different interfaces, identify common access paths and interfaces with longer dwell times, and assist in analyzing experience bottlenecks. These methods are typically presented in the form of flowcharts and heatmaps, helping developers understand user flow as a whole. However, these methods are mostly statistical and cannot effectively analyze task stages and user emotional changes under complex interface structures, making it difficult to accurately determine the source of deep-seated experience problems.

[0003] Some existing technologies apply machine learning algorithms to user behavior analysis, such as using clustering algorithms to identify user groups and using classification models to predict users' potential operational tendencies. These methods typically train models based on historical statistical data and construct user feature vectors through feature engineering to predict possible abnormal behaviors or optimize user guidance paths. However, these methods rely on fixed features and model structures, making it difficult to capture structural problems such as loops, bottleneck nodes, and isolated termination nodes in interface transitions, especially in scenarios where interface layouts are constantly changing and business processes are frequently adjusted, resulting in weak generalization capabilities. Meanwhile, large-scale language models have made significant progress in the field of natural language processing, with some applications attempting to use them for interface copy optimization, dialogue guidance, and simple interface content generation. These technologies mainly rely on natural language input and output, automatically generating interface prompts or assisting users in completing specific tasks. Although large-scale language models possess strong language generation capabilities, existing technologies generally lack the ability to combine them with interface structure analysis, failing to convert interface topology and user operation sequences into information that large-scale language models can directly understand and use. Furthermore, these technologies often ignore the user's emotional state during task execution and fail to correlate emotional fluctuations with interface issues, resulting in insufficient accuracy in optimization suggestions. Summary of the Invention

[0004] The purpose of this invention is to provide a user experience improvement system based on large-scale model technology. This system labels user operation sequences with task stage tags and emotion tags, identifies loop segments, bottleneck nodes, and isolated termination nodes through navigation path topology analysis, and constructs user experience indicators such as task completion rate, average number of successful steps, and complaint rate in a path structure statistics table. Furthermore, it forms a path structure causal relationship table based on the correlation between structural features and experience indicators, aggregating multiple topological causal problem path clusters. Finally, relying on reconstruction suggestion elements, the large-scale model generates interface reconstruction description text and automatically generates prototype interface files, realizing an automated user experience improvement process from problem identification and explanation to interface solution output. This system can accurately locate the source of experience problems, significantly improve the structured capability and causal judgment capability of path analysis, and improve interface adjustment efficiency through generative interface optimization technology, enabling user experience optimization in complex business scenarios to have higher accuracy, completeness, and engineering feasibility.

[0005] To address the aforementioned technical problems, this invention provides a user experience improvement system based on large-scale model technology. The system includes: a data processing unit, used to acquire interaction operation records of the target application, construct a user operation sequence arranged chronologically according to session identifiers, and generate task stage tags and emotion tags for each operation record through a large-scale model task stage classification interface and a large-scale model emotion classification interface, thereby obtaining a set of labeled user operation sequences; and a navigation path expansion unit, used to extract interface identifier sequences from the set of labeled user operation sequences to form a navigation path set, establish a navigation topology graph with interface identifiers as nodes and adjacent interface jumps as directed edges, perform topology data analysis on the navigation path set to generate a topology feature combination field for each navigation path, thereby obtaining a navigation path expansion... The system comprises: an expansion set and a topology analysis unit. The expansion set is used to build a path structure statistics table, collect user sessions corresponding to the path structure records to obtain user experience metrics, establish a path structure causal relationship table based on topology feature combination fields and user experience metrics, aggregate multiple topology causal problem path clusters in the path structure causal relationship table, and generate reconstruction suggestion elements for each topology causal problem path cluster. The large model optimization unit is used to input the reconstruction suggestion elements and the path structure statistics table into the large model optimization module, construct interface reconstruction prompt text, generate interface reconstruction description text through the large model, input the interface reconstruction description text and reconstruction suggestion elements together into the user interface generator module, generate prototype interface files, and output them to improve the user experience.

[0006] Furthermore, the data processing unit collects operation record sets from the front-end tracking system and the server-side log system. Each operation record includes: session identifier, interface identifier, control identifier, operation type, timestamp, and optional text content. These are grouped by session identifier and arranged chronologically within each session to form a user operation sequence. An interface function description table and an operation type description table are established on the server. The interface function description table configures one interface function description text for each interface identifier, and the operation type description table configures one operation type description text for each operation type. A task stage tag name set and an emotion tag name set are preset on the server, each set containing at least four tag names. A task stage tag field and an emotion tag field are set for each operation record to store the tag names output by the large model.

[0007] Furthermore, the process by which the data processing unit generates task stage tags and emotion tags for each operation record specifically includes: selecting several operation records before and after the current operation record from the corresponding user operation sequence to form a context operation set; concatenating the interface function description text and operation type description text of each operation in the context operation set in chronological order to form a context description text; generating task stage classification prompt text and emotion classification prompt text based on the context description text, the interface function description text and operation type description text of the current operation record; inputting the task stage classification prompt text into the task stage classification interface of the large model, and inputting the emotion classification prompt text into the emotion classification interface of the large model; and restricting the output content of the large model to select one tag name from the task stage tag name set and the emotion tag name set respectively through the instruction template, and writing them into the task stage tag field and emotion tag field of the current operation record.

[0008] Furthermore, the navigation path extension unit reads the interface identifiers of each user operation sequence in the set of labeled user operation sequences in chronological order, compresses the identical interface identifiers that appear repeatedly into one, and retains only the timestamp of the first operation record in the continuous interval to obtain the interface identifier sequence. The interface identifier sequence is then registered as a navigation path to form a navigation path set. Using all the interface identifiers appearing in the navigation path set as nodes, directed edges are established between two adjacent interface identifiers to obtain a navigation topology graph. A degree segment is established for each node in the server, and the degree segment records the number of incoming edges and outgoing edges of the corresponding interface identifier. Topology feature calculations are performed on the navigation topology graph and the navigation path set using topology data analysis methods to generate topology feature combination fields describing loop structures, bottleneck structures, and isolated termination structures.

[0009] Furthermore, the process of generating the topology feature combination field by the navigation path extension unit specifically includes: detecting interface identifier pairs that appear repeatedly before and after each navigation path's interface identifier sequence and have at least one different interface identifier between them; registering the interface identifier sequence from the previous position to the next position as a loop segment and incrementing the loop segment count field of the navigation path by one; performing degree statistics on nodes on the navigation topology graph; when the number of incoming edges of a node is greater than or equal to a preset first threshold and the number of outgoing edges is greater than or equal to a preset second threshold, writing the node's interface identifier into the bottleneck node list field and writing one in the bottleneck marker field of the navigation path containing the node's interface identifier; when the number of outgoing edges of a node is equal to zero, writing the node's interface identifier into the isolated termination node list field and writing one in the isolated termination marker field of the navigation path with the node's interface identifier as the last interface identifier; combining the loop segment count field, bottleneck marker field, and isolated termination marker field into a topology feature combination field, recording the loop existence status, bottleneck node existence status, and isolated termination node existence status in the topology feature combination field; the navigation path extension set consists of all navigation paths with the topology feature combination field.

[0010] Furthermore, the process of constructing the path structure statistics table by the topology analysis unit specifically includes: merging navigation paths in the extended set of navigation paths according to the condition that the interface identifier sequences are completely identical; generating a path structure record for each combination of interface identifier sequences to form a path structure statistics table; recording the path structure identifier, interface identifier sequence, path length, total loop segment count, bottleneck marker field, and isolated termination marker field in each path structure record; associating all sessions corresponding to the interface identifier sequence; performing statistical processing on the labeled user operation sequence corresponding to each session; statistical processing including successful registration sessions, failed sessions, and complaint sessions; counting the total number of successful steps; and generating a task stage label set field and a sentiment label set field to record the subset of task stage label name set and subset of sentiment label name set related to the sessions associated with the path structure record.

[0011] Furthermore, the process of calculating user experience metrics and classifying path structures by the topology analysis unit specifically includes: for each path structure record in the path structure statistics table, the task completion rate is obtained by dividing the number of successful sessions by the total number of sessions and written into the task completion rate field; the average number of successful steps is obtained by dividing the sum of the number of successful steps by the number of successful sessions and written into the average number of successful steps field; and the complaint rate is obtained by dividing the number of complaint sessions by the total number of sessions and written into the complaint rate field. Multiple task completion rate thresholds, multiple average number of successful steps thresholds, and multiple complaint rate thresholds are set in the server. Based on the relationship between the task completion rate field value, the average number of successful steps field value, the complaint rate field value, and each threshold, the path structure records are divided into at least two categories: high-performance path structure records and problem path structure records.

[0012] Furthermore, the process of constructing the path structure causal relationship table and aggregating topological causal problem path clusters by the topology analysis unit specifically includes: taking the loop existence status, bottleneck node existence status, and isolated termination node existence status in the topology feature combination field of each path structure record as structural feature items, and taking the task completion rate field, average number of successful steps field, and complaint rate field as indicator feature items; recording the causal relationship entries between structural feature items and indicator feature items in the path structure causal relationship table; and distinguishing between high-performance path structure records and problem path structure records in the problem label field; classifying and aggregating the problem path structure records according to the label combination of the topology feature combination field, interface identifier sequence feature, task stage label set field, and emotion label set field to form multiple topological causal problem path clusters, including at least loop problem path clusters, bottleneck problem path clusters, and isolated termination problem path clusters.

[0013] Furthermore, the process of generating reconstruction suggestion elements by the topology analysis unit specifically includes: for loop problem path clusters, counting the frequency of each interface identifier in the interface identifier sequence recorded in the problem path structure record within the cluster, selecting the group of interface identifiers with the highest frequency, and merging the input controls and submit controls of the interface corresponding to this group of interface identifiers to form a new interface definition. The new interface definition includes a title area, an input control area, and a submit control area in a top-down order, forming loop merging reconstruction suggestion elements; for bottleneck problem path clusters, selecting the interface identifier with the highest frequency based on the bottleneck node interface identifier recorded in the problem path structure record within the cluster as the target bottleneck interface identifier, generating a leading interface definition for the target bottleneck interface identifier. The leading interface definition includes a task description area and an operation option area. The task description area displays the target bottleneck... The interface identifies the relevant task stage label text, and the operation option area is set with at least two operation buttons. The operation buttons jump to the next interface in the path structure record where the subsequent task completion rate of the target bottleneck interface is higher than the set completion rate threshold, forming a bottleneck improvement reconstruction suggestion element. For isolated termination problem path clusters, a return entry component is added to the interface definition corresponding to the termination interface. The return entry component includes at least two buttons, which are used to jump to the starting interface of the task flow to which the termination interface belongs and the previous navigation home page interface, respectively, forming an isolated path guidance reconstruction suggestion element. The above three types of reconstruction suggestion elements are registered in the reconstruction suggestion element table and associated with loop problem path clusters, bottleneck problem path clusters and isolated termination problem path clusters to obtain the topological causal path aggregation reconstruction result.

[0014] Furthermore, the large model optimization unit writes the topological causal path aggregation and reconstruction results and path structure statistics table into a structured configuration file. In the large model optimization module, interface reconstruction prompt text is constructed for each topological causal problem path cluster. The interface reconstruction prompt text includes at least the path structure identifier, interface identifier sequence, task stage label set field content, emotion label set field content, task completion rate field value, average number of successful steps field value, complaint rate field value, loop merging reconstruction suggestion elements, bottleneck improvement reconstruction suggestion elements, and isolated path guidance reconstruction suggestion elements. The interface reconstruction prompt text is input into the large model generation interface, and the interface reconstruction description text is output through the large model generation interface. The interface reconstruction description text provides natural language descriptions of interface merging, addition of leading interface, and addition of return entry component.

[0015] The user experience improvement system based on large model technology of the present invention has the following beneficial effects: By introducing user operation sequence annotation, navigation path topology analysis, path structure causal analysis, large model generative optimization, and automatic prototype interface generation within the same system, the present invention transforms the user experience improvement process from a method that previously relied on manual experience into a process-oriented system that can automatically identify problems, automatically generate solutions, and automatically output interface prototypes. Based on the extended set of navigation paths, the present invention constructs a path structure statistical table. By identifying structural features such as loop segments, bottleneck nodes, and isolated termination nodes, the system can accurately identify key structures in the task flow that cause users to stop, travel back and forth, or interrupt the process. Furthermore, relying on user task stage tags and emotion tags during the operation process, the present invention can correlate user behavioral intentions with structural features, thereby determining whether users hesitate, feel anxious, or attempt to repeatedly try at specific locations on the path, enhancing the accuracy of problem identification. In the path structure causal relationship table, the present invention associates structural features with user experience indicators such as task completion rate, average number of successful steps, and complaint rate, enabling the system to distinguish between high-frequency paths and paths that truly affect the experience, and enhancing the reliability of causal judgment by comparing differences between paths with the same structure. By aggregating loop problem path clusters, bottleneck problem path clusters, and isolated termination problem path clusters, this invention can automatically extract common experience problem types from a large amount of user operation data, avoiding misjudgments caused by relying solely on a single session or local features. After generating refactoring suggestion elements, this invention utilizes a large model optimization unit to construct interface refactoring prompt text. The large model generates logically clear and highly executable interface refactoring explanation text, making the design adjustment suggestions more reasonable and consistent. Finally, through the user interface generator, prototype interface files are automatically output. This invention achieves an end-to-end closed loop from problem identification and explanation to interface solution output, significantly improving the efficiency and accuracy of the user experience optimization process. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the principle of navigation topology graph construction and node degree analysis provided in an embodiment of the present invention. Figure 2 This is a multi-dimensional analysis diagram showing the impact of task stage label combinations on path performance, provided in an embodiment of the present invention. Detailed Implementation

[0017] A user experience improvement system based on large model technology includes: a data processing unit, used to acquire interaction operation records of the target application, construct a user operation sequence arranged chronologically according to session identifiers, and generate task stage labels and emotion labels for each operation record through large model task stage classification interfaces and large model emotion classification interfaces, resulting in a set of labeled user operation sequences; a navigation path expansion unit, used to extract interface identifier sequences from the set of labeled user operation sequences to form a navigation path set, establish a navigation topology graph with interface identifiers as nodes and adjacent interface jumps as directed edges, perform topology data analysis on the navigation path set to generate topology feature combination fields for each navigation path, resulting in an expanded navigation path set; and topology analysis. The first unit is used to build a path structure statistics table based on the navigation path extension set, collect user sessions corresponding to the path structure records to obtain user experience indicators, establish a path structure causal relationship table based on the combination of topological feature fields and user experience indicators, aggregate multiple topological causal problem path clusters in the path structure causal relationship table, and generate reconstruction suggestion elements for each topological causal problem path cluster. The second unit is used to input the reconstruction suggestion elements and the path structure statistics table into the large model optimization module, construct interface reconstruction prompt text, generate interface reconstruction description text through the large model, input the interface reconstruction description text and reconstruction suggestion elements together into the user interface generator module, generate prototype interface files and output them to improve user experience.

[0018] In one embodiment, each functional unit of the present invention can be implemented using software functional modules on a general-purpose server or cloud host. The server includes at least one processor, a memory, and a network interface. The memory stores a computer program that can run on the processor. The data processing unit, navigation path expansion unit, topology analysis unit, and large model optimization unit can all be implemented using different functional modules or processes within this computer program. Those skilled in the art can implement the functional modules using any common backend development language such as Java, Python, Go, or C++, and interact with the frontend tracking system, server-side logging system, and large model service via interface protocols such as HTTP and gRPC. Therefore, the system structure of the present invention is feasible under existing hardware and software conditions.

[0019] In one implementation, the data processing unit is deployed on the server side and continuously receives interaction operation records from the target application through communication with the front-end application and back-end business services. Each time an operation such as interface loading, control clicking, text input, or scrolling occurs, the front-end application encapsulates a session identifier, interface identifier, control identifier, operation type, timestamp, and optional text content into an operation record and sends it to the server. The back-end business services also generate structurally consistent operation records at key business nodes such as order generation and payment result return, thus forming a unified set of operation records on the server side. To ensure timing accuracy, the timestamp is accurate to the millisecond level. When receiving operation records, the server aligns them according to the reception time and timestamp, discarding records with missing timestamps or incorrect formats to ensure that the subsequently constructed user operation sequence has a strict chronological order.

[0020] In one feasible technical implementation, the user interface generator module takes a structured interface generation description structure as input and automatically generates prototype interface files. The interface generation description structure can be in JSON or XML format, and its fields include interface identifiers, interface titles, layout types, control lists, and navigation relationships between controls. After parsing this description structure, the user interface generator module can output prototype interface files in two ways: one is to generate front-end prototype code that can be directly run in a browser, such as automatically generating a collection of files containing HTML structure, CSS styles, and JavaScript event bindings; the other is to generate files that can be imported by prototyping tools (such as general interface prototyping software), such as describing interface components and their positions according to the tool's open format. For web front-end implementation, the user interface generator module can pre-set several general templates, using placeholders to represent control positions and text content. During actual generation, the placeholders are replaced with specific values ​​from the interface generation description structure to output a previewable prototype page. Those skilled in the art can implement the automatic prototype interface generation function of this invention within existing front-end engineering systems.

[0021] The data processing unit first groups the set of operation records using session identifiers as keys. Each session identifier corresponds to all operation records of a user during a single continuous usage session. Within each session, the data processing unit sorts the operation records by timestamp in ascending order, forming a chronological sequence of user operations. To avoid duplicate records within the same session due to network retransmissions, the data processing unit performs deduplication on adjacent operation records after sorting. If the session identifier, interface identifier, control identifier, and operation type are completely identical, and the timestamp difference is less than a preset time interval (e.g., less than or equal to 200 milliseconds), only the record with the earlier timestamp is retained. This method ensures that the user operation sequence accurately reflects the actual interaction process and avoids the same operation being counted repeatedly in statistics.

[0022] To generate more accurate task stage labels and sentiment labels, the data processing unit, when processing each operation record, considers not only the current operation record itself but also constructs categorized prompt text containing contextual information. Specifically, in the corresponding user operation sequence, the data processing unit selects up to three operation records before the current operation record and up to three operation records after the current operation record to form a contextual operation set. When there are fewer than three available forward or backward records, all available records are used. For each operation record in the contextual operation set, the data processing unit reads the interface function description text from the interface function description table and the operation type description text from the operation type description table, and concatenates the interface function description text and operation type description text in chronological order to form the contextual description text. By introducing preceding and following contextual operations, task stage labels can reflect the position of the current operation in the entire task flow. For example, an operation of clicking the settings entry is more suitable to be labeled as the "search settings" stage if multiple function browsing operations have occurred before it and the network configuration interface appears immediately afterward. Sentiment labels can also be judged based on operation patterns over a continuous period of time. For example, repeated input and deletion of content and rapid interface switching are more likely to correspond to anxiety or hesitation.

[0023] The selection of a maximum of 3 forward and 3 backward operation records is determined based on the following considerations: Firstly, from the perspective of user behavior characteristics, the 3-5 operations performed by a user before and after completing a small sub-task (such as filling out a form or clicking a confirmation button) can usually fully reflect the contextual semantics of the current task stage. Secondly, from the perspective of large model inference performance, when the length of a single prompt text is controlled within several hundred characters, the inference latency can be reduced while ensuring the accuracy of the inference result. Therefore, in this embodiment, the length of the context window is set to the first 3 and the last 3 operation records. For scenarios with low operation frequency or simple interfaces, those skilled in the art can also adjust this parameter to the first 2 and the last 2, or the first 5 and the last 5. As long as the large model can obtain sufficient contextual information, the technical effect of the present invention can be achieved, and the specific value of this parameter does not affect the essence of the method of the present invention.

[0024] When generating task stage classification prompt text, the data processing unit sequentially concatenates the interface function description text of the current operation record, the operation type description text of the current operation record, and the context description text, appending a task stage label description paragraph at the end. The task stage label description paragraph lists each label in the task stage label name set, such as search settings, information entry, retry after payment failure, and result viewing, and provides a description of the typical operation scenarios applicable to each label. For example, search settings corresponds to a user finding a configuration entry in a multi-level menu; information entry corresponds to a user entering personal information or payment information in a form or input box; retry after payment failure corresponds to a user attempting another operation after a payment failure prompt appears; and result viewing corresponds to a user viewing query results, order details, or processing status. By explicitly prompting all available labels and their semantic boundaries to the large model's task stage classification interface, the large model can select only from a limited set of labels during output, reducing ambiguity and improving label consistency.

[0025] In one feasible approach, the large-scale model task stage classification interface and the large-scale model sentiment classification interface can be implemented by calling the HTTP inference service of an external large-scale language model. Specifically, the server constructs a JSON-formatted request message for each operation record to be classified. The request message includes the model name, temperature parameter, maximum output length, and the generated task stage classification prompt text or sentiment classification prompt text. The temperature parameter is preferably set to a value between 0 and 0.3, such as 0.1, to reduce randomness and ensure stable label outputs for the same type of input. The maximum output length is preferably set to 64 to 256 tokens, such as 128 tokens, to cover a natural language explanation and a label name. The server sends the request message to the large-scale model inference service address via HTTP POST, and the inference service returns a response message containing a natural language answer.

[0026] To ensure the output can be automatically parsed by the program, the server presets fixed output format constraints in the prompt text. For example, it might add instructions like "Please only output a tag name, do not output any other content" at the end of the prompt text, or require the large model to wrap the result in specific tag pairs, such as "[LABEL]tag name[ / LABEL]". Upon receiving the response message, the data processing unit extracts the string located between the tag pairs according to the preset format, and then matches it within the task-stage tag name set or the emotion tag name set to obtain the final tag. Through this method, those skilled in the art can directly use any large-scale language model that meets the conditions for task-stage classification and emotion classification without relying on the implementation details of a specific model.

[0027] In a preferred embodiment of the present invention, the task stage tag name set may include, but is not limited to, at least four tags such as "search settings", "information filling", "retry after payment failure", "result viewing", and "complaint submission"; the emotion tag name set may include, but is not limited to, at least four tags such as "calm", "hesitant", "anxious", and "disappointed". The specific tags selected can be configured by the business unit according to the actual scenario. The present invention does not limit the specific content of the tag names, but only requires that the task stage tag name set and the emotion tag name set remain relatively stable within a certain period of time in the same deployment for subsequent statistical analysis.

[0028] The data processing unit provides the task stage classification prompt text as input to the large-scale model task stage classification interface. This interface can access a remotely deployed large-scale language model service via the network or call a locally deployed large-scale language model inference program. The request uses a fixed format: the task stage classification prompt text is placed in the text input field, the set of task stage label names is listed in natural language in the instruction section, and a restrictive description is added, requiring the large model to output only one task stage label name in the response. After the interface returns a response text, the data processing unit searches for labels in the task stage label name set within the returned text, matching them in a pre-agreed order, such as starting from the search settings and checking sequentially until the first label present in the returned text is found. This label is then written to the task stage label field of the current operation record. If no label is matched in the returned text, a fallback strategy can be used, such as setting the task stage label field to information input or setting it to a predefined default task stage label to ensure that each operation record has a valid task stage label.

[0029] Once the task stage label and sentiment label fields are completed for all user operation sequences, the data processing unit stores these labeled user operation sequences into a labeled user operation sequence set. The labeled user operation sequence set contains multiple fields, including session identifier, operation sequence number, interface identifier, control identifier, operation type, timestamp, task stage label, sentiment label, and optional text content. In practice, the labeled user operation sequence set can be stored in a relational database table or in a distributed columnar storage. To improve the efficiency of subsequent navigation path construction, a composite index can be created on the labeled user operation sequence set based on the session identifier and timestamp, enabling the navigation path expansion unit to quickly read records in session order during scanning.

[0030] In one implementation, the navigation path expansion unit directly takes the set of labeled user operation sequences as input and generates a navigation path set by performing interface-level extraction on each user operation sequence. Specifically, the navigation path expansion unit traverses the set of labeled user operation sequences according to session identifiers. For each user operation sequence, it reads the interface identifiers in chronological order, starting from the first operation record. When the current interface identifier is the same as the interface identifier of the previous record, it indicates that the user is still on the same interface and may have only performed continuous operations on multiple controls. In this case, the navigation path expansion unit does not add a new interface identifier but skips the current record. Only when the current interface identifier is different from the previous record, indicating that the user interface has changed, does the navigation path expansion unit append the current interface identifier to the interface identifier sequence. In this way, fine-grained user interactions within the same interface can be compressed into a single interface node, focusing attention on the interface-level jump structure, reducing path length, and facilitating subsequent topology analysis.

[0031] When constructing the navigation topology graph, the navigation path expansion unit uses all interface identifiers appearing in the navigation path set as nodes. For each navigation path, starting from the first interface identifier in the interface identifier sequence, the navigation path expansion unit reads two adjacent interface identifiers sequentially and registers the jump relationship between these two interface identifiers as a directed edge. For any pair of adjacent interface identifiers, the navigation path expansion unit checks in its internally maintained edge set whether a corresponding directed edge record already exists. If it does not exist, a new directed edge record is created, and the occurrence count of the edge is initialized to 1; if it already exists, the occurrence count of the edge is incremented by 1. This allows both structural information and jump frequency information to be retained in the navigation topology graph, which is beneficial for subsequently prioritizing the analysis of paths frequently traversed by users.

[0032] The navigation path expansion unit establishes a degree segment for each node in the navigation topology graph. The degree segment includes two subfields: the number of incoming edges and the number of outgoing edges. When processing each directed edge, the navigation path expansion unit increments the number of incoming edges corresponding to the node with the terminating interface identifier by 1, and increments the number of outgoing edges corresponding to the node with the starting interface identifier by 1. After scanning all navigation paths, the degree segment of each node in the navigation topology graph reflects the positional characteristics of that interface within the global navigation structure. For example, when the number of incoming edges for a certain interface identifier is significantly higher than that of other interface identifiers, it indicates that the user has navigated to that interface from multiple different entry points, and this interface is structurally more likely to become a bottleneck node; when the number of outgoing edges for a certain interface identifier is 0, it indicates that the user will not navigate to other interfaces from that interface, and this interface is a terminating node, possibly a results display page or an abnormal interruption page.

[0033] In one specific implementation, Figure 1This section showcases the navigation topology graph structure constructed by the navigation path extension unit and a visual representation of the node degree characteristics. The navigation topology graph consists of interface nodes and directed edges, used to characterize the user's interface-level navigation behavior patterns within the target application. Figure 1 It contains eight interface nodes, labeled Home, Product Details, Shopping Cart, Shipping Address, Payment Method, Payment Result, Settings Home, and Network Configuration. Each interface node is represented by a circle, with the interface name inside the circle and two degree segments above it: the number of entering edges and the number of leaving edges. The number of entering edges indicates how many different navigation paths lead to this interface, and the number of leaving edges indicates how many different subsequent interfaces can be navigated to from this interface. For example, the Home node has 0 entering edges and 3 leaving edges, meaning that the Home node, as the starting node, has no preceding interface, but users can navigate to 3 different subsequent interfaces from the Home node. Interface nodes are connected by directed edges, each represented by an arrow, indicating the direction of interface navigation. The thickness of the directed edge is proportional to the navigation frequency, and the navigation frequency value is marked at the midpoint of the directed edge.

[0034] For example, the directed edge from the homepage to product details has a frequency of 85, indicating that this jump path appeared 85 times in all user sessions counted. This edge is shown as a thick solid line in the diagram. The directed edge from product details to shopping cart has a frequency of 78, also shown as a thick solid line. In contrast, the directed edge from settings homepage to payment method has a frequency of only 8, shown as a thinner line in the diagram. In the navigation topology diagram, some directed edges are drawn as curves to represent reverse jumps or loop segments. For example, the directed edges from shipping address to shopping cart and from payment method to shipping address are both curved. These two edges form a loop structure with the forward path, indicating that users have a back-and-forth modification behavior pattern during the payment process. The directed edge from settings homepage back to homepage is also curved, representing the navigation behavior of users returning to the main interface after settings operations.

[0035] Figure 1 The area below displays the legend. The legend first defines three line thicknesses corresponding to different navigation frequency ranges: thin lines represent low-frequency navigation (less than 30 times); medium-thickness lines represent medium-frequency navigation (30 to 60 times); and thick lines represent high-frequency navigation (60 times or more). The visual difference in line thickness allows for quick identification of the user's primary and secondary navigation paths. The legend also explains the meaning of node types and degree segments. Interface nodes are represented by hollow circles with a 3-pixel border. The number of entering edges is indicated in the upper left corner of the node, with the prefix "Enter"; the number of leaving edges is indicated in the upper right corner of the node, with the prefix "Exit".

[0036] The number of incoming edges reflects the degree of convergence of the interface in the navigation structure; the larger the value, the more paths the interface converges at. The number of outgoing edges reflects the flow distribution capability of the interface; the larger the value, the more subsequent paths the user can choose from the interface. Figure 1 The bottom right corner of the interface uses dashed rectangles to mark two types of special nodes. The payment method node is surrounded by a dashed box and marked as a bottleneck node, with the description below indicating ≥3 incoming edges. This means the payment method node has at least three incoming edges, originating from the shipping address, product details, and shopping cart, satisfying the bottleneck node criterion. Bottleneck nodes are prone to causing congestion in the task flow. Users entering this interface from multiple entry points may become confused due to high interface complexity or insufficient guidance, leading to a decrease in task completion rate. The payment result node is surrounded by another dashed box and marked as isolated termination, with the description below indicating 0 outgoing edges. This means the payment result node has zero outgoing edges, and there is no record of users navigating to other interfaces after reaching this screen. Isolated termination nodes may be normal task completion interfaces, such as the result display page after successful payment, or they may be locations where the user's task is prematurely interrupted, such as when the user exits the application without guidance after a failed payment. By identifying isolated termination nodes, entry points or guidance information to return to the main flow can be added to these interfaces to prevent user churn.

[0037] After constructing the navigation topology map, the navigation path extension unit performs topology data analysis on the navigation path set to generate a topology feature combination field describing the topological characteristics of each navigation path. In one implementation, the navigation path extension unit first performs loop detection on each navigation path. During loop detection, starting from the first position of the interface identifier sequence, the navigation path extension unit compares the interface identifier at the current position with the interface identifiers at all subsequent positions one by one. If a subsequent position contains an interface identifier identical to the current position, and there is at least one different interface identifier between them, the navigation path extension unit registers the interface identifier sequence from the previous position to the next position as a loop segment and increments the loop segment count field of that navigation path by 1. This detection method can identify situations where users repeatedly return to a certain interface. For example, if a user first enters the network configuration interface from the settings homepage, then jumps back to the settings homepage, and then goes to the notification settings interface, this path forms a loop segment. The larger the loop segment count field, the more back-and-forth actions the user has performed on that navigation path, which usually means that the information architecture is unclear or the guidance prompts are insufficient.

[0038] The navigation path expansion unit then identifies bottleneck nodes and isolated termination nodes based on the degree segment of the navigation topology graph. In one implementation, the navigation path expansion unit presets a first threshold of 3 and a second threshold of 1. When the number of incoming edges to a node is greater than or equal to 3 and the number of outgoing edges is greater than or equal to 1, it indicates that at least 3 different paths lead to that interface, and the user will continue to jump to the next interface from that interface. Such interfaces are prone to becoming congestion points in the task flow. The navigation path expansion unit writes the interface identifier of that node into the bottleneck node list field and sets the bottleneck marker field to 1 in the navigation path containing that interface identifier. This identification strategy helps to prioritize the location of key interfaces that have a significant impact on task completion rate, facilitating the subsequent provision of targeted interface reconstruction suggestions.

[0039] The determination of the first and second thresholds can be based on the statistical results of historical navigation data. In a typical embodiment, developers first collect all user session data over a period of time (e.g., one month), construct a navigation topology map, and count the number of incoming and outgoing edges for each interface node, obtaining the distribution of the number of incoming and outgoing edges. Then, the number of incoming edges is sorted from smallest to largest, and the value approximately at the 30% to 40% quantile of the distribution is taken as the lower limit threshold for bottleneck determination. In actual samples, this quantile usually falls between 3 and 5. At the same time, considering that a bottleneck node must have at least one subsequent jump path, the lower limit threshold for the number of outgoing edges is set to 1. Therefore, in this embodiment, an incoming edge threshold of 3 and an outgoing edge threshold of 1 are selected as example parameters. For small applications with fewer interfaces, those skilled in the art can lower the incoming edge threshold to 2; for large applications with many interfaces and extremely complex navigation structures, the incoming edge threshold can be increased to 4 or 5, thereby identifying only the most typical convergence node as the bottleneck node. The above threshold adjustment only changes the sensitivity of bottleneck node screening, and does not change the basic idea of ​​this invention to identify key nodes through degree statistics and combine user experience indicators for causal analysis.

[0040] For isolated termination nodes, the navigation path extension unit searches for nodes with zero departure edges on the navigation topology graph and writes the interface identifiers of these nodes into the isolated termination node list field. Within each navigation path, when the last interface identifier in the interface identifier sequence equals an interface identifier in the isolated termination node list field, the navigation path extension unit sets the isolated termination flag field of that navigation path to 1. Isolated termination nodes often correspond to interfaces where user tasks are prematurely interrupted or ended, such as error message pages or result pages that do not guide the user back to the main process. Marking these nodes helps identify locations where users are likely to churn and allows for the addition of entry points and guidance information for returning to the main process in refactoring suggestions.

[0041] When generating the topology feature combination field, the navigation path extension unit combines the loop segment count field, bottleneck marker field, and isolated termination marker field of each navigation path into a structured field. The topology feature combination field records the loop existence status, bottleneck node existence status, and isolated termination node existence status. In a simple implementation, the loop existence status can be represented by an integer count, directly using the value of the loop segment count field; the bottleneck node existence status can be represented by a binary marker, where a bottleneck marker equals 1 indicating the presence of a bottleneck node, and a bottleneck marker equals 0 indicating the absence of a bottleneck node; the isolated termination node existence status also uses a binary marker. The navigation path extension set consists of all navigation paths with the topology feature combination field, preserving the interface identifier sequence while adding topology information for subsequent causal and aggregation analysis.

[0042] In one implementation, the topology analysis unit is deployed on the server side. It takes the navigation path extension set output by the navigation path extension unit as input, performs structured statistics and aggregation on a large number of user sessions, thereby obtaining a path structure statistics table, user experience indicators, path structure causal relationship table, and multiple topology causal problem path clusters. Based on this, it generates reconstruction suggestion elements for use by the subsequent large model optimization unit and user interface generator.

[0043] Each navigation path in the extended navigation path set includes at least a navigation path identifier, an interface identifier sequence, a loop segment count field, a bottleneck marker field, an isolated termination marker field, and a list of session identifiers corresponding to the navigation path. The topology analysis unit first merges the extended navigation path set, grouping navigation paths with identical interface identifier sequences together, generating a path structure record for each group, forming the basic record for the path structure statistics table. The interface identifier sequence is used as the merging condition because it determines the actual access order of users at the interface level. The same interface identifier sequence corresponds to the same navigation structure, i.e., the same task path structure. Success rates, failure rates, and complaint records of different users on this path can be aggregated in the same path structure record, facilitating statistical comparison.

[0044] During the generation of the path structure statistics table, the topology analysis unit sets multiple fields for each path structure record, including path structure identifier, interface identifier sequence, path length, total loop segment count, bottleneck marker field, isolated termination marker field, and subsequent fields to be filled in, such as the total number of sessions, number of successful sessions, number of failed sessions, number of complaint sessions, total number of successful steps, task stage tag set field, and sentiment tag set field. The path length can directly record the number of interface identifiers in the interface identifier sequence, used to measure the complexity of the task path. The total loop segment count, bottleneck marker field, and isolated termination marker field are directly inherited from the navigation path extension set and are used for subsequent topology classification and causal analysis.

[0045] To categorize path structure records into high-performance and problem-problem paths, the topology analysis unit pre-configures several task completion rate thresholds, average number of successful steps thresholds, and complaint rate thresholds on the server. For example, in one specific setting, the task completion rate thresholds can be set to 0.9, 0.8, and 0.4; the average number of successful steps thresholds to 5 and 10; and the complaint rate thresholds to 0.05, 0.1, and 0.3. Based on these thresholds, the topology analysis unit divides the task completion rate, average number of successful steps, and complaint rate values ​​into different intervals, thereby classifying the path structure records into multiple levels. In a simplified implementation, when a path structure record has a task completion rate value greater than or equal to 0.8, an average number of successful steps value less than or equal to 5, and a complaint rate value less than or equal to 0.1, the topology analysis unit registers the path structure record as a high-performance path structure record. When a path structure record has a task completion rate value less than or equal to 0.4, an average number of successful steps value greater than or equal to 10, and a complaint rate value greater than or equal to 0.3, the path structure record is registered as a problem path structure record. Path structure records in the middle range can be considered as general path structure records. In actual deployment, these thresholds can be adjusted according to business objectives and historical statistical results. For example, the proportion of problem path structure records can be controlled within the top 20% of all path structure records to focus on the most serious problem paths.

[0046] The aforementioned task completion rate threshold, average number of successful steps threshold, and complaint rate threshold can be determined by combining statistical analysis of historical data with business goal setting. Specifically, the topology analysis unit can perform a rough statistical analysis of all path structure records at the initial stage of system deployment, calculating the task completion rate, average number of successful steps, and complaint rate for each path, and then plotting the empirical distribution curves for these three indicators. Taking task completion rate as an example, sorting all paths from high to low by task completion rate reveals that approximately 20%–30% of paths in common business systems have a completion rate higher than 0.8, which typically represents a well-designed main process; approximately 10%–20% of paths have a completion rate lower than 0.4, which typically corresponds to abnormal or marginal processes with poor user experience. Combining the business team's subjective expectations for "high-performance paths" and "problem paths," 0.8 and 0.4 can be used as reference thresholds for high-performance and problem paths, respectively; if an application wants to more strictly filter problem paths, the problem path threshold can be raised to 0.5.

[0047] Similarly, the average number of successful steps threshold can be determined by statistically analyzing the distribution of steps across all successful sessions, selecting a threshold from the median and several quantiles above it. For example, when most successful sessions involve 3 to 7 steps, and sessions exceeding 10 steps account for only about 15% of all successful sessions, 5 steps can be used as the upper limit for a "better path," and 10 steps as the lower limit for a "poorer path." The complaint rate threshold can be determined based on the company's complaint rate control objectives. For example, in e-commerce payment scenarios, the overall complaint rate is typically required to be below 5%, so 0.05 can be used as the first threshold for the complaint rate. When the complaint rate for a certain path exceeds 0.3, it is significantly higher than the overall system level and can be considered a serious problem path. The above thresholds can be fixed empirical values ​​or can be adjusted by business line and time using configuration files. Those skilled in the art can configure them according to actual needs.

[0048] After completing the calculation of user experience metrics and the classification of path structure records, the topology analysis unit begins to build a path structure causal relationship table. This table records the association between topology feature combination fields and user experience metrics. The topology analysis unit breaks down the topology feature combination fields in each path structure record into multiple structural feature items, such as loop existence status, bottleneck node existence status, and isolated termination node existence status. It also uses the task completion rate, average number of successful steps, and complaint rate fields as metric feature items. For each path structure record, the topology analysis unit writes a causal relationship entry into the path structure causal relationship table. This entry includes a path structure identifier, structural feature items, metric feature items, and a problem marker field to distinguish between high-performance and problematic path structure records. The problem marker field can be represented by integers; a value of 0 is set for high-performance path structure records, a value of 1 for problematic path structure records, and other values ​​for general path structure records. In this way, the path structure causal relationship table can centrally display the performance of a certain type of topology combination under different user experience metrics, which is beneficial for subsequently identifying which structural feature combinations are more strongly associated with unfavorable situations in user experience metrics.

[0049] In one specific implementation, Figure 2 A heatmap is used to illustrate the quantitative relationship between topological feature combinations and the overall user experience score. The horizontal axis of the heatmap represents the number of loop segments, ranging from 0 to 5, while the vertical axis represents the number of bottleneck nodes, also ranging from 0 to 5. The heatmap consists of 36 cells in 6 rows and 6 columns. Each cell represents a specific combination of topological features, and the cell displays the corresponding overall user experience score, ranging from 15 to 95. The overall user experience score is calculated using a weighted average: Overall User Experience Score = Task Completion Rate multiplied by 50, plus 1 minus the average number of successful steps divided by 20, multiplied by 30, plus 1 minus the complaint rate, multiplied by 20. This formula comprehensively considers three core user experience indicators: task completion rate, average number of successful steps, and complaint rate, with weights of 50, 30, and 20 respectively, reflecting the dominant role of task completion rate in experience evaluation.

[0050] Figure 2The cell in the bottom left area, corresponding to 0 loop segments and 0 bottleneck nodes, has a user experience score of 95, the highest score on the heatmap. This cell is filled with pure white without any texture, indicating an excellent user experience level. When the path structure has no loop segments and no bottleneck nodes, users can complete tasks along a clear linear path with fewer interaction steps, avoiding backtracking and hesitation. Task completion rates are typically above 80%, and complaint rates are typically below 10%. As the number of loop segments and bottleneck nodes increases, the overall user experience score gradually decreases.

[0051] For example, when the number of loop segments is 1 and the number of bottleneck nodes is 0, the score drops to 87 points; when the number of loop segments is 0 and the number of bottleneck nodes is 1, the score drops to 88 points. This indicates that loops and bottlenecks have independent negative impacts on user experience; even the presence of only one topology problem can lead to a drop in experience score of approximately 7 to 8 points. When the number of loop segments is 3 and the number of bottleneck nodes is 3, the overall user experience score drops to 49 points, with the cells filled with a dense dotted texture. Dense dotted textures represent a low to medium level of user experience, corresponding to a score range of 40 to 59 points. Under this topology combination, users need to make multiple back-and-forth jumps during navigation, and face complex path choices at multiple bottleneck nodes. The average number of successful steps usually exceeds 10, the task completion rate drops to around 50%, and the complaint rate rises to between 20% and 30%. Figure 2 The cell in the upper right corner with 5 loop segments and 5 bottleneck nodes has a user experience score of 15, which is the lowest score on the heatmap.

[0052] This cell uses a diagonal line texture with an 8-pixel spacing, indicating a poor user experience. When the path structure contains numerous loop segments and bottleneck nodes, users will experience severe disorientation and frustration during navigation, with a task completion rate potentially below 30%, an average number of successful steps exceeding 15, and a complaint rate exceeding 40%. This is a critically problematic path requiring priority reconstruction. The heatmap uses four different fill textures to differentiate user experience levels. Cells scoring between 80 and 100 are filled with pure white without texture, labeled as excellent; cells scoring between 60 and 79 use a sparse dotted texture with a dot radius of 2 pixels and a dot spacing of 20 pixels, labeled as good; cells scoring between 40 and 59 use a dense dotted texture with a dot radius of 2.5 pixels and a dot spacing of 12 pixels, labeled as average; and cells scoring between 15 and 39 use a diagonal line texture with a diagonal line width of 1.5 pixels and a diagonal line spacing of 8 pixels, labeled as poor.

[0053] Figure 2 The right side displays a color-coded legend of the overall user experience score. The legend uses a vertical layout, showing four score ranges and their corresponding texture styles and level labels from top to bottom. Each legend cell is 60×70 pixels in size and filled with the same texture as the heatmap cells, facilitating a quick understanding of the heatmap's visual coding rules. In the lower left corner of the heatmap, a dashed rectangle marks the excellent path area. This area covers four cells with 0 to 1 loop segments and 0 to 1 bottleneck nodes; all of these cells score above 85 points.

[0054] Excellent path zones exhibit simple and clear path structures and superior user experience metrics, serving as a benchmark for path design. In the upper right corner of the heatmap, problem path zones are marked with dashed rectangles. This area covers nine cells with 3 to 5 loop segments and 3 to 5 bottleneck nodes, all scoring below 55 points. Problem path zones represent complex and chaotic path structures with poor user experience metrics, requiring targeted optimization through refactoring suggested elements. Figure 2 The notes below explain the additional impact of isolated terminal nodes on the overall user experience score. When isolated terminal nodes exist in the path structure, the overall user experience score needs to be reduced by an additional 10 to 15 points beyond what is shown on the heatmap.

[0055] For example, when the number of loop segments and bottleneck nodes is 2, the heatmap shows a score of 66. If the path also has an isolated termination node, the actual score should be adjusted to between 51 and 56, and the level will be downgraded from good to average. The annotation further points out that when three topology problems occur simultaneously—isolated termination nodes, three or more loop segments, and two or more bottleneck nodes—the overall user experience score is usually below 25, indicating a severely problematic path. This path combination means that users repeatedly try different paths within a complex navigation structure, face difficulty in making choices at key nodes, and are ultimately forced to interrupt their task at an isolated termination node, unable to smoothly return to the main flow.

[0056] When the topology analysis unit identifies such paths, it assigns them the highest reconstruction priority, ensuring that improvements are implemented first in product iterations. The overall distribution pattern of the heatmap reveals a negative correlation between the overall user experience score and the number of loop segments and bottleneck nodes, with an interactive effect between the two. When both the number of loop segments and bottleneck nodes increase simultaneously, the score decreases faster than when either factor increases. For example, when the number of loop segments increases from 0 to 5, with 0 bottleneck nodes, the score drops from 95 to 55, a decrease of 40 points; however, with 5 bottleneck nodes, the score drops from 60 to 15, a decrease of 45 points. This non-linear relationship indicates a cumulative effect between topology problems; the coexistence of multiple problems significantly increases the damage to the user experience.

[0057] The aforementioned formula for calculating the overall user experience score is an example of normalizing and weighting multiple indicators. The selection of weights 50, 30, and 20 is primarily based on the following considerations: the task completion rate directly reflects whether users can complete the core task and is the primary indicator for measuring the quality of the path, therefore it is given the highest weight; the average number of successful steps reflects the efficiency and operational burden of completing the task, which has a significant impact on the experience but is usually secondary to whether the task can be completed, therefore it is given a medium weight; although the complaint rate is generally low in numerical terms, once a complaint occurs, it often indicates an extremely poor user experience, therefore it is given a certain weight to lower the score. To simplify implementation, the sum of the weights of the three is set to 100, so that the overall user experience score naturally falls within the range of 0 to 100. For different business systems, those skilled in the art can adjust the relative weights of the three indicators according to the actual situation. For example, in a customer service system, the weight of the complaint rate can be increased, and in a configuration system with extremely complex operation paths, the weight of the average number of successful steps can be increased. As long as the aggregation is still carried out in the manner of "linear weighting after normalization of each indicator", it falls within the scope of implementation of this invention.

[0058] After the path structure causal relationship table is constructed, the topology analysis unit aggregates the problem path structure records to obtain multiple topological causal problem path clusters. The aggregation process comprehensively considers the topological feature combination field, interface identifier sequence features, and tag combinations in the task stage tag set field and sentiment tag set field. For loop problem path clusters, the topology analysis unit selects path structure records from the path structure causal relationship table with a problem tag field value of 1 and a loop existence status indicating the existence of a loop, forming a loop problem candidate set. Within the loop problem candidate set, when the difference in interface identifier sequence length between two path structure records is less than or equal to 2, and the starting and ending interface identifiers of the interface identifier sequences are the same, and the task stage tag set field of the two path structure records contains the same task stage tag combination, the topology analysis unit groups these two path structure records into the same loop problem path cluster. This aggregation method can group together loop problems that are structurally highly similar and have consistent task stage scenarios, avoiding the need to generate reconstruction suggestion elements separately for each path with subtle differences, thereby improving the generalization ability of the reconstruction suggestion elements.

[0059] For bottleneck problem path clusters, the topology analysis unit selects path structure records from the path structure causal relationship table where the problem marker field value is 1 and the bottleneck node exists, forming a bottleneck problem candidate set. Within this candidate set, the topology analysis unit searches for the interface identifier following the bottleneck node interface identifier in the interface identifier sequence based on the position of the bottleneck node interface identifier in the path structure record. If two path structure records have the same bottleneck node interface identifier, and the first two interface identifiers following that bottleneck node interface identifier are identical, and the proportion of the anxiety label in the emotion label set field of both path structure records exceeds a preset threshold (e.g., greater than or equal to 0.5), then the topology analysis unit classifies these two path structure records into the same bottleneck problem path cluster. This classification method groups together paths that structurally resemble each other in behavior near the same bottleneck node and where user emotions are generally inclined towards anxiety, indicating that there are relatively consistent problems in the interaction design near this bottleneck node, requiring targeted optimization.

[0060] For isolated termination problem path clusters, the topology analysis unit selects path structure records from the path structure causality table where the problem marker field value is 1 and the isolated termination node exists, forming an isolated termination problem candidate set. Within this candidate set, when two path structure records have the same termination interface identifier sequence and contain the same task stage label set field, the topology analysis unit groups these two path structure records into the same isolated termination problem path cluster. This aggregation method groups paths where user tasks end at the same termination interface and have similar task stage environments. This typically indicates that the user is finding it difficult to continue the task or return to the main flow on a result interface, error interface, or interruption interface, requiring the addition of a return entry point and guidance content.

[0061] After classifying the path clusters into loop problem clusters, bottleneck problem clusters, and isolated termination problem clusters, the topology analysis unit generates corresponding reconstruction suggestion elements for each type of topological causal problem path cluster. For loop problem path clusters, the topology analysis unit statistically analyzes the interface identifier sequence of all path structure records within the cluster, extracts loop segment intervals from the interface identifier sequence, counts the frequency of interface identifiers appearing within the loop segment intervals, sorts them from high to low frequency, and selects the group of interface identifiers with the highest frequency, for example, the top 5 interface identifiers, as the candidate merged interface set. Subsequently, the topology analysis unit reads the interface definitions corresponding to these interface identifiers from the interface definition library and merges the input controls and submit controls in the interface into a new interface definition. The new interface definition is usually designed in a top-down order as a title area, an input control area, and a submit control area. The title area is used to describe the current task environment and expected results, the input control area is used to accommodate form items from multiple interfaces, and the submit control area is used to centrally place confirmation buttons. In this way, multiple interfaces that previously required multiple jumps and frequent user back-and-forth can be merged into a single centralized interface, reducing the number of interface switching and minimizing user disorientation within the loop, thereby improving task completion rate and efficiency. The topology analysis unit organizes this information into loop merging and reconstruction suggestion elements, including a set of candidate merged interfaces, a recommended layout structure, and a set of path structure identifiers bound to the problem path cluster of the loop.

[0062] For bottleneck path clusters, the topology analysis unit counts the frequency of bottleneck node interface identifiers within each cluster and selects the most frequently occurring identifier as the target bottleneck interface identifier. For the target bottleneck interface identifier, the topology analysis unit generates a leading interface definition. This definition typically includes a task description area in the upper half and an operation option area in the lower half. The task description area displays task stage label text related to the target bottleneck interface identifier. Combining this with the task stage label set field in the path structure statistics table, the unit selects the task stage label most closely related to the bottleneck node, such as "retry after information entry or payment failure," and generates a concise task description to help users understand the next operation. The operation option area provides at least two operation buttons, for example, three buttons, which jump to the next interface identifier in the corresponding path structure record where the subsequent task completion rate is higher than a set completion rate threshold. This design aims to significantly reduce user hesitation and trial-and-error behavior near bottleneck nodes by adding a leading interface before the bottleneck node, providing clear task descriptions and a few high-success-rate path options, thereby reducing the frequency of anxiety labels and improving the overall experience. The topology analysis unit organizes information such as the target bottleneck interface identifier, task description content, number of operation options, and corresponding jump interface identifiers into bottleneck improvement and reconstruction suggestion elements.

[0063] For isolated termination problem path clusters, the topology analysis unit reads the corresponding interface definition from the interface definition library based on the termination interface identifier and adds a return entry component to the termination interface definition. The return entry component includes at least two buttons: one button jumps to the starting interface identifier of the task flow to which the termination interface belongs, and the other button jumps to the previous navigation homepage interface identifier. This design helps prevent users from getting lost on isolated termination interfaces, unable to continue the task, or unable to return to the main function entry point. By providing users with an intuitive return path, the interruption rate on this path can be reduced, the task completion rate can be improved, and the complaint rate can be lowered. The topology analysis unit organizes the structural information of the termination interface identifier, the task flow starting interface identifier, the previous navigation homepage interface identifier, and the newly added return entry component into isolated path guidance refactoring suggestion elements.

[0064] In one implementation, the large model optimization unit runs on the server side, directly reads the reconstruction suggestion element table and path structure statistics table output by the topology analysis unit, calls the large model optimization module to generate interface reconstruction description text by constructing interface reconstruction prompt text, and then combines the interface reconstruction description text and reconstruction suggestion elements to drive the user interface generator module to generate prototype interface files, thereby realizing automated reconstruction and prototype output of the user interface.

[0065] In the specific implementation process, the large model optimization unit first loads the refactoring suggestion element table and the path structure statistics table from persistent storage. In the refactoring suggestion element table, each refactoring suggestion element record includes at least a refactoring suggestion element identifier, a refactoring suggestion type field, a candidate interface identifier set field, a layout suggestion field, and the associated topological causal problem path cluster identifier. The refactoring suggestion type field can be explicitly marked as a loop merging refactoring suggestion element, a bottleneck improvement refactoring suggestion element, or an isolated path guiding refactoring suggestion element. In the path structure statistics table, each path structure record includes at least a path structure identifier, an interface identifier sequence, a task completion rate field, an average number of successful steps field, a complaint rate field, a task stage tag set field, and a sentiment tag set field.

[0066] The large model optimization unit internally maintains the mapping relationship between topological causal problem path clusters and path structure records. For each topological causal problem path cluster, the large model optimization unit extracts all path structure records belonging to that path structure path cluster from the path structure statistics table, forming a path structure record set. To ensure that the interface reconstruction description text generated by the large model accurately reflects the severity of the problem and provides targeted interface adjustment solutions, the large model optimization unit summarizes and extracts features from the user experience indicators in the path structure record set. For example, it calculates the median value of the task completion rate field, the median value of the average number of successful steps field, and the median value of the complaint rate field within the topological causal problem path cluster. It also counts the most frequently occurring task stage tag combinations in the task stage tag set fields and the proportion of anxiety tags, hesitation tags, etc., in the emotion tag set fields.

[0067] The large model optimization unit constructs interface reconstruction prompt text based on the summarized information and reconstruction suggestion elements mentioned above. The interface reconstruction prompt text is typically divided into several parts so that the large model can clearly understand the current interface structure, user behavior, and expected changes. The first part is a problem background description. Based on the interface identifier sequence field in the path structure statistics table, the large model optimization unit provides a natural language description of the user's navigation process on the current path. For example, it explains which interface the user starts from, which intermediate interfaces they pass through, and which final interface they finally arrive at. To enhance comprehensibility, the large model optimization unit mentions the path length, whether there are obvious loop segments, whether there are bottleneck nodes, and whether the path ends at an isolated terminal node, enabling the large model to grasp the overall complexity of the path structure and potential risks.

[0068] The second part is the overview of user experience metrics. The large model optimization unit writes the values ​​of the task completion rate, average number of successful steps, and complaint rate into the interface redesign prompt text. For example, it can be described as follows: on average, only 25 out of every 100 sessions complete the key task on this path; on average, each successful session requires 12 interactive steps; and on average, 40 out of every 100 sessions generate complaints or feedback. Using specific numerical descriptions helps the large model grasp the severity of the problem, infer the burden the current structure places on the user, and thus tend to provide structural adjustments rather than just minor text tweaks when generating interface redesign instructions.

[0069] The third part is the description of task stages and emotion labels. The large model optimization unit selects the combination of task stage labels most closely related to the path cluster of causal problems in the topology from the task stage label set field, such as the combination of search settings and information entry, or the combination of information entry and retrying after payment failure, and describes the approximate location of these task stages in the path. Simultaneously, the large model optimization unit extracts the occurrence ratio of labels such as anxiety and hesitation from the emotion label set field, and describes the distribution of these emotions during user operation through text. For example, when the proportion of anxiety labels in this path reaches 50%, the interface reconstruction prompt text can clearly indicate that the user experiences a large amount of anxiety in this navigation path, significantly impacting the overall experience. Through this part, the large model can understand that the current task not only has structural obstacles but also accumulates significant negative feedback at the emotional level, giving it more motivation to propose design suggestions to reduce burden and uncertainty in the interface reconstruction prompt text.

[0070] Part Four is the description of the refactoring suggestion elements. Based on the different fields in the refactoring suggestion type field, the large model optimization unit converts the candidate interface identifier set field and layout suggestion field contained in the refactoring suggestion elements into natural language descriptions. For example, for the loop merging refactoring suggestion element, the large model optimization unit would explain that there are currently multiple functionally highly related interfaces with a high probability of repetition. It would suggest merging the most frequently occurring interfaces into a single centralized operation interface, stating the task objective in the title area, arranging all necessary form items in the input control area, and providing a single confirmation button in the submission control area. Through this descriptive method, the large model not only understands which interfaces need to be merged but also comprehends the design intent of the merge, namely, reducing user back-and-forth navigation and shortening the number of steps required to complete the task.

[0071] Regarding the bottleneck improvement and restructuring suggestions, the large model optimization unit explains in the interface restructuring prompt text the location of the target bottleneck interface marker, typical user behavior patterns before and after the bottleneck node, and the frequency of anxiety tags. It suggests adding a preparatory interface before the bottleneck node, using concise text to explain the user's current task stage and providing several operation option buttons with statistically validated high success rates, guiding users directly to a better path. The reason for this design is that when users repeatedly try near complex nodes, they actually need the system to proactively provide the optimal path, rather than having them fumble around in a complex structure. By clearly stating this in the prompt text, the large model can generate more specific text guidance content and button naming suggestions in the interface restructuring description text.

[0072] Regarding the suggested elements for refactoring isolated path guidance, the large model optimization unit explains in the interface refactoring prompt text that the interface corresponding to the current termination interface will cause users to remain in a state with no further options, easily leading to task interruption. It suggests adding buttons to the termination interface to return to the starting point of the task flow and to return to the previous navigation homepage, along with brief explanations next to these buttons, allowing users to quickly return to a state where they can continue operating. The large model can then use this information to provide specific button text suggestions in the interface refactoring prompt text, such as prompting users to return to the main function page to restart the operation, reducing feelings of disorientation.

[0073] After receiving the interface reconstruction instructions, the large model optimization unit parses the text, structuring the key content into interface adjustment commands. The parsing process can utilize rule-based text analysis methods, such as locating the current status summary, problem description, and reconstruction suggestion sections according to the expected paragraph order. Within the reconstruction suggestion sections, operations associated with specific interface identifiers are further extracted, such as updating title text, merging input controls, adding new buttons, or adjusting control positions. The large model optimization unit compares this parsed content with the corresponding reconstruction suggestion elements to ensure that the suggestions proposed by the large model are consistent with the structural adjustment direction output by the previous topology analysis unit. For example, when a reconstruction suggestion element points to a loop merging reconstruction suggestion element, the large model optimization unit checks whether the interface reconstruction instructions contain descriptions of the number of merged interfaces, the layout after merging, and the title text; when a reconstruction suggestion element points to a bottleneck improvement reconstruction suggestion element, the large model optimization unit checks whether it contains task description text and suggested operation button names from the preceding interface.

[0074] In one implementation, after obtaining the candidate merged interface set, the topology analysis unit extracts the input controls and submission controls from each interface according to a predefined interface component library. The interface component library predefines unified data structures and attribute fields for common control types (single-line text input boxes, multi-line text input boxes, drop-down selection boxes, radio button groups, checkbox groups, main operation buttons, secondary operation buttons, etc.). The merging process follows these rules: First, all main operation buttons from the candidate interfaces are merged into the submission control area of ​​the new interface, arranged in the order of "main task-related buttons first, secondary task-related buttons last"; second, input controls from the candidate interfaces are logically grouped by field, for example, recipient name, contact number, and delivery address are grouped into the "Delivery Information Group," and payment method and invoice information are grouped into the "Payment Information Group," and sub-blocks are created within each group in the input control area of ​​the new interface; finally, a task description text is automatically generated for the title area of ​​the new interface, such as "Please fill in the following information at once to complete the current task." This text can be generated or fine-tuned by the large model optimization module based on task stage tags and user sentiment tags. Those skilled in the art can map the above components to actual renderable control instances according to specific UI frameworks (such as web front-ends, mobile native control libraries, etc.), thereby achieving automatic merging and reconstruction of the interface.

[0075] Under the premise that the interface reconstruction description text and reconstruction suggestion elements are consistent and complementary, the large model optimization unit constructs an interface generation description structure that the user interface generator module can recognize. The interface generation description structure typically includes interface identifiers, interface types, layout information, a control list, and control attributes. Layout information can be described using a grid layout, such as dividing the interface into several rows and columns, specifying the row and column position, width percentage, and height percentage of each control. The control list assigns a control identifier, control type, and text content from the interface reconstruction description text to each control. Control types include title text, plain text, input boxes, buttons, etc. For buttons added by the large model suggestion, the large model optimization unit adds a target interface identifier to the control attributes, enabling the user interface generator module to correctly establish the jump relationships between interfaces when generating prototype interface files.

[0076] Taking the payment process in an e-commerce application as an example, the specific implementation process of the system of the present invention is explained. In this embodiment, the target application includes interfaces such as homepage, product details, shopping cart, shipping address, payment method, and payment result. The data processing unit obtains approximately 5 million operation records from the front-end tracking system and the server-side log system within the past month, of which approximately 800,000 are session identifiers. The data processing unit groups by session identifier and sorts by timestamp to form user operation sequences; after deduplication, each session contains an average of approximately 15 operation records. The task stage label name set is configured as four labels: "Product Browsing", "Information Filling", "Retry After Payment Failure", and "Result Viewing", and the emotion label name set is configured as five labels: "Calm", "Hesitant", "Anxious", "Excited", and "Disappointed". For each operation record, the data processing unit uses three operation records before and after as context windows to construct task stage classification prompt text and emotion classification prompt text, and calls the large model task stage classification interface and large model emotion classification interface with a temperature parameter of 0.1 and a maximum output length of 128 to obtain the label results, finally obtaining a set of labeled user operation sequences.

[0077] The navigation path expansion unit extracts each user operation sequence at the interface level, resulting in a set of navigation paths. A common path is "Homepage → Product Details → Shopping Cart → Shipping Address → Payment Method → ​​Payment Result". When constructing the navigation topology graph, the number of entry edges for the payment method interface is 3, and the number of exit edges is 1; the number of entry edges for the payment result interface is 2, and the number of exit edges is 0. Based on the rule that the number of entry edges is greater than or equal to 3 and the number of exit edges is greater than or equal to 1, the payment method interface is marked as a bottleneck node; based on the rule that the number of exit edges is equal to 0, the payment result interface is marked as an isolated termination node. The topology analysis unit statistically analyzes all path structure records and finds that the total number of sessions for the above path structure records is 60,000, the number of successful sessions is 24,000, the number of failed sessions is 12,000, the number of complaint sessions is 18,000, and the total number of successful steps is approximately 288,000. Therefore, the task completion rate is calculated to be 0.4%, the average number of successful steps is 12, and the complaint rate is 0.3%. Based on the criteria of task completion rate ≤ 0.4, average number of successful steps ≥ 10, and complaint rate ≥ 0.3, this path structure record is classified as a problem path structure record and categorized into a topological causal problem path cluster that simultaneously contains loop segments (users repeatedly jump between the shopping cart and shipping address interfaces in some sessions), bottleneck nodes, and isolated termination nodes. The topology analysis unit then generates loop merging and reconstruction suggestions (suggesting merging some input controls on the shipping address and payment method interfaces), bottleneck improvement and reconstruction suggestions (suggesting adding a pre-screening interface before the payment method to guide users to choose a payment method with a higher success rate), and isolated path guidance and reconstruction suggestions (suggesting adding "Return to Homepage" and "Re-initiate Payment" buttons to the payment result interface).

[0078] The large model optimization unit constructs interface reconstruction prompt text from information such as the interface identifier sequence, task completion rate, average number of successful steps, complaint rate, and task stage and sentiment label distribution of the aforementioned problem path cluster. It then calls the large model to generate interface reconstruction explanation text, resulting in natural language suggestions such as "Add the explanation 'Currently, payment has failed; we recommend prioritizing quick payment methods' before the payment method interface" and "Add a 'Re-initiate Payment' button below the payment results page." The user interface generator module generates a new prototype interface file based on the interface reconstruction explanation text and reconstruction suggestion elements. After a brief manual review, the original interface is replaced. A retest one month after launch revealed that the task completion rate for this path increased to 0.7, the average number of successful steps decreased to 7, and the complaint rate dropped to 0.08.

[0079] The present invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1.A user experience improvement system based on large model technology, characterized by, The system includes: a data processing unit, used to acquire interaction operation records of the target application, construct user operation sequences arranged chronologically according to session identifiers, and generate task stage labels and emotion labels for each operation record through large model task stage classification interfaces and large model emotion classification interfaces, resulting in a set of labeled user operation sequences; a navigation path expansion unit, used to extract interface identifier sequences from the set of labeled user operation sequences to form a navigation path set, establish a navigation topology graph with interface identifiers as nodes and adjacent interface jumps as directed edges, perform topology data analysis on the navigation path set to generate topology feature combination fields for each navigation path, resulting in an expanded navigation path set; and a topology analysis unit, used for navigation... Based on the path extension set, a path structure statistics table is built. User sessions corresponding to the path structure records are counted to obtain user experience indicators. A path structure causal relationship table is established based on the combination of topological feature fields and user experience indicators. Multiple topological causal problem path clusters are aggregated in the path structure causal relationship table, and reconstruction suggestion elements are generated for each topological causal problem path cluster. The large model optimization unit is used to input the reconstruction suggestion elements and the path structure statistics table into the large model optimization module to construct interface reconstruction prompt text. The interface reconstruction description text is generated through the large model. The interface reconstruction description text and reconstruction suggestion elements are input into the user interface generator module to generate prototype interface files and output them to improve the user experience. 2.The large model technology-based user experience improvement system of claim 1, wherein, The data processing unit collects operation record sets from the front-end tracking system and the server-side log system. Each operation record includes: session identifier, interface identifier, control identifier, operation type, timestamp, and optional text content. These are grouped by session identifier and arranged chronologically within each session to form a user operation sequence. An interface function description table and an operation type description table are established on the server. The interface function description table configures one interface function description text for each interface identifier, and the operation type description table configures one operation type description text for each operation type. A task stage tag name set and an emotion tag name set are preset on the server, each set containing at least four tag names. A task stage tag field and an emotion tag field are set for each operation record to store the tag names output by the large model. 3.The large model technology-based user experience improvement system of claim 2, wherein, The process of generating the task stage label and the emotion label for each operation record by the data processing unit specifically comprises: selecting a plurality of operation records before and after the current operation record from the corresponding user operation sequence to form a context operation set, splicing interface function description texts and operation type description texts of operations in the context operation set into context description texts in time sequence; generating task stage classification prompt texts and emotion classification prompt texts based on the context description texts and the interface function description text and the operation type description text of the current operation record, inputting the task stage classification prompt texts into a large model task stage classification interface, inputting the emotion classification prompt texts into a large model emotion classification interface, limiting the large model output content to select one label name from the task stage label name set and the emotion label name set respectively through an instruction template, and writing the task stage label and the emotion label into the task stage label field and the emotion label field of the current operation record. 4.The large model technology-based user experience improvement system of claim 1, wherein, The navigation path expansion unit reads the interface identifiers in time sequence for each user operation sequence in the labeled user operation sequence set, compresses the same interface identifiers that continuously and repeatedly appear into one, only retains the time stamp of the first operation record in the continuous interval, obtains an interface identifier sequence, and registers the interface identifier sequence as a navigation path to form a navigation path set; Taking all the interface identifiers appearing in the navigation path set as nodes, a directed edge is established between two adjacent interface identifiers to obtain a navigation topology graph, a degree field is established for each node in the server, and the degree field records the number of entering edges and the number of leaving edges of the corresponding interface identifier; topological feature calculation is performed on the navigation topology graph and the navigation path set by using a topological data analysis method to generate a topological feature combination field describing loop structure, bottleneck structure and isolated termination structure. 5.The large model technology-based user experience improvement system of claim 4, wherein, The process of generating the topological feature combination field by the navigation path expansion unit specifically comprises: detecting, for the interface identifier sequence of each navigation path, an interface identifier pair that appears repeatedly before and after and has at least one different interface identifier between them, registering the interface identifier sequence from the previous position to the next position as a loop segment, and adding one to the loop segment count field of the navigation path; performing degree statistics on the nodes on the navigation topology graph, when the number of entering edges of a node is greater than or equal to a preset first threshold value and the number of leaving edges is greater than or equal to a preset second threshold value, writing the node interface identifier into a bottleneck node list field and writing one into a bottleneck marker field in the navigation path containing the node interface identifier; when the number of leaving edges of a node is equal to zero, writing the node interface identifier into an isolated termination node list field and writing one into an isolated termination marker field in the navigation path with the node interface identifier as the last interface identifier; combining the loop segment count field, the bottleneck marker field and the isolated termination marker field into the topological feature combination field, recording the loop existence state, the bottleneck node existence state and the isolated termination node existence state in the topological feature combination field, and the navigation path expansion set is composed of all the navigation paths with the topological feature combination field. 6.The large model technology-based user experience improvement system of claim 1, wherein, The process of constructing the path structure statistics table by the topology analysis unit specifically includes: merging the navigation paths in the navigation path expansion set according to the condition that the interface identifier sequences are completely identical, generating a path structure record for each interface identifier sequence combination, and forming the path structure statistics table; recording the path structure identifier, the interface identifier sequence, the path length, the loop segment count sum, the bottleneck marker field and the isolated termination marker field in each path structure record, and associating all the sessions corresponding to the interface identifier sequence, performing statistical processing on the annotated user operation sequence corresponding to each session, which includes registering successful sessions, failed sessions and complaint sessions, counting the total number of successful steps, and generating the task phase label set field and the emotion label set field, which are used to record the task phase label name set subset and the emotion label name set subset related to the sessions associated with the path structure record. 7.The large model technology-based user experience improvement system of claim 6, wherein, The process of calculating the user experience indicators and classifying the path structures by the topology analysis unit specifically includes: for each path structure record in the path structure statistics table, obtaining the task completion rate value by dividing the number of successful sessions by the total number of sessions and writing it into the task completion rate field, obtaining the average number of successful steps value by dividing the total number of successful steps by the number of successful sessions and writing it into the average number of successful steps field, and obtaining the complaint rate value by dividing the number of complaint sessions by the total number of sessions and writing it into the complaint rate field; setting multiple task completion rate thresholds, multiple average number of successful steps thresholds and multiple complaint rate thresholds in the server, and dividing the path structure records into at least two categories of high-performance path structure records and problem path structure records according to the relationship between the task completion rate field value, the average number of successful steps field value and the complaint rate field value and each threshold. 8.The large model technology-based user experience improvement system of claim 7, wherein, The process of constructing the path structure causal relationship table and aggregating the topology causal problem path clusters by the topology analysis unit specifically includes: taking the loop existence state, the bottleneck node existence state and the isolated termination node existence state in the topology feature combination field of each path structure record as structure feature items, taking the task completion rate field, the average number of successful steps field and the complaint rate field as indicator feature items, recording the causal association entries of the structure feature items pointing to the indicator feature items in the path structure causal relationship table, and distinguishing the high-performance path structure records from the problem path structure records in the problem marker field; classifying and aggregating the problem path structure records according to the topology feature combination field, the interface identifier sequence feature and the label combination of the task phase label set field and the emotion label set field, forming multiple topology causal problem path clusters, including at least loop problem path clusters, bottleneck problem path clusters and isolated termination problem path clusters. 9.The large model technology-based user experience improvement system of claim 8, wherein, The process of generating the reconstruction suggestion element by the topology analysis unit specifically includes: for the loop problem path cluster, counting the occurrence frequency of each interface identifier in the interface identifier sequence of the problem path structure record in the cluster, selecting a group of interface identifiers with the highest occurrence frequency, merging the input control and the submission control of the interfaces corresponding to the group of interface identifiers to form a new interface definition, and the new interface definition sequentially includes a title area, an input control area and a submission control area in a top-down order to form a loop merging reconstruction suggestion element; for the bottleneck problem path cluster, selecting an interface identifier with the highest occurrence frequency as a target bottleneck interface identifier according to the bottleneck node interface identifier of the problem path structure record in the cluster, generating a leading interface definition for the target bottleneck interface identifier, the leading interface definition including a task description area and an operation option area, the task description area displaying a task stage label text related to the target bottleneck interface identifier, and the operation option area being provided with at least two operation buttons, the operation buttons being respectively connected to a next interface identifier of a path structure record in which the target bottleneck interface identifier is followed by a path in which the subsequent task completion rate is higher than a set completion rate threshold, to form a bottleneck improvement reconstruction suggestion element; for the isolated termination problem path cluster, adding a return entry component in the interface definition corresponding to the termination interface identifier, the return entry component including at least two buttons for respectively jumping to a starting interface identifier of a task flow to which the termination interface belongs and a navigation home page interface identifier of a previous layer, to form an isolated path guide reconstruction suggestion element; and registering the three types of reconstruction suggestion elements in the reconstruction suggestion element table and establishing an association identifier with the loop problem path cluster, the bottleneck problem path cluster and the isolated termination problem path cluster to obtain the topology causal path aggregation reconstruction result. 10.The large model technology-based user experience improvement system of claim 9, wherein, The large model optimization unit writes the topology causal path aggregation reconstruction result and the path structure statistics table into a structured configuration file, constructs an interface reconstruction prompt text for each topology causal problem path cluster in the large model optimization module, the interface reconstruction prompt text at least including a path structure identifier, an interface identifier sequence, a task stage label set field content, an emotion label set field content, a task completion rate field value, an average successful step number field value, a complaint rate field value, and a loop merging reconstruction suggestion element, a bottleneck improvement reconstruction suggestion element and an isolated path guide reconstruction suggestion element, inputs the interface reconstruction prompt text into a large model generation interface, outputs an interface reconstruction description text via the large model generation interface, and gives a natural language description of interface merging, leading interface addition and return entry component addition in the interface reconstruction description text.

Images