Computer-based user behavior data mining method, server and medium

Abstract

The application provides a computer-based user behavior data mining method, a server and a medium, relates to the technical field of computers, and obtains a user behavior record data set containing operation behavior sequences carrying operation timestamp markers; performs operation intention boundary identification processing on the user behavior record data set to obtain an intention switching boundary position set and an operation hesitation interval distribution set; constructs a user intention switching association network with operation intention paragraphs as nodes and hesitation transition relationships as connecting edges based on the two sets; calls an intention depth analysis model to perform intention hierarchical expansion processing on the user intention switching association network to generate a user intention hierarchical distribution graph containing an expansion path and a mounting coordinate; extracts intention drift trajectory chains arranged according to timestamps according to the user intention hierarchical distribution graph, and generates a user behavior pattern mining result based on the intention drift trajectory chains. The application can realize hierarchical analysis of user intentions and accurate capture of drift trajectories, and significantly improves the precision of behavior pattern mining.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a computer-based user behavior data mining method, server, and medium. Background Technology

[0002] In computer application systems, the mining and analysis of user behavior data is fundamental to personalized service recommendations, user profile construction, and user experience optimization. Current technologies for analyzing user behavior data typically employ methods based on frequent itemset mining or sequence pattern matching. This involves extracting frequently occurring operation paths or combinations from a user's historical operation records to infer their behavioral preferences. Some existing technologies also utilize intent recognition methods based on Hidden Markov Models, dividing the user's operation sequence into several discrete states and performing probabilistic inference. However, these existing technologies have significant shortcomings in handling actual user operations: they treat user behavior as a linear sequence composed of several discrete intent states, ignoring the deep semantic information contained in the intermediate transitional behaviors such as hesitation, wavering, and repeated switching between adjacent intents. This results in an understanding of the user's true intent remaining at a superficial state judgment level, failing to effectively capture the dynamic evolution process of the user's intent during operation, which involves gradual changes, shifts, or even drifts. Meanwhile, existing user behavior models are typically flat state transition structures, lacking the ability to depict the hierarchical relationships between different operational intent segments. They are difficult to reveal the jump patterns and progressive evolution characteristics of user intents at different abstract levels when performing complex tasks. As a result, when faced with complex scenarios where user intents are ambiguous or multiple intents are intertwined, the accuracy and interpretability of the output behavior pattern mining results are significantly limited. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a computer-based user behavior data mining method, the method comprising: Acquire a set of user behavior record data, which contains a sequence of operation behaviors marked with operation timestamps, and each sequence of operation behaviors consists of continuously collected multi-dimensional operation events; The user behavior record data set is processed to identify the boundary of the operation intent, and the set of intent switching boundary positions and the set of operation hesitation interval distribution corresponding to the operation behavior sequence are obtained. The set of intent switching boundary positions is used to define the segmentation position of adjacent operation intent segments in the operation behavior sequence, and the set of operation hesitation interval distribution is used to characterize the duration distribution of each operation hesitation event in the operation behavior sequence. A user intent switching association network is constructed based on the set of intent switching boundary locations and the set of operation hesitation interval distributions. The user intent switching association network uses operation intent segments as node units and hesitation transition relationships between operation intent segments as connecting edges. The attributes of node units include the duration of intent segments and the density of operation events within intent segments. The attributes of connecting edges include the duration of hesitation transitions and the hesitation transition type identifier. The preset intent depth parsing model is invoked to perform intent hierarchy expansion processing on the user intent switching association network, and the operation intent segments are mapped to the intent hierarchy space to generate a user intent hierarchy distribution map. The user intent hierarchy distribution map includes the expansion path of each intent level in the intent hierarchy space and the mounting coordinates of each operation intent segment in the intent hierarchy space. The intent drift trajectory chain corresponding to the operation behavior sequence is extracted based on the user intent hierarchy distribution map. The intent drift trajectory chain is composed of an intent hierarchy identifier sequence arranged in the order of operation timestamps. User behavior pattern mining results are generated based on the intent drift trajectory chain.

[0004] In another aspect, embodiments of the present invention also provide a server, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described computer-based user behavior data mining method by executing the machine-executable instructions.

[0005] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described computer-based user behavior data mining method.

[0006] Based on the above, by performing operation intent boundary recognition processing on the user behavior record data set to accurately obtain the set of intent switching boundary locations and the set of operation hesitation interval distribution, a user intent switching association network is further constructed with operation intent segments as node units and hesitation transition relationships as connecting edges. This realizes the transformation from the original operation sequence to a structured intent association topology, so that the mining system is no longer limited to the discrete state classification of user operation behavior, but can characterize the deep semantic association and transition features reflected by hesitation behavior between adjacent operation intent segments. This overcomes the technical defect of the existing technology that ignores the transition behavior between operation intents, resulting in a one-sided understanding of the user's true intent. Meanwhile, this invention calls a preset intent depth parsing model to perform intent hierarchy expansion processing on the constructed user intent switching association network, maps the operation intent segments to the intent hierarchy space and generates a user intent hierarchy distribution map containing expansion paths and mounting coordinates, and then extracts intent drift trajectory chains arranged in the order of operation timestamps. In this way, a progressive analysis link is established between user intent boundary identification, association modeling, hierarchy expansion and trajectory tracking. This not only accurately captures the drift evolution law of user intent between different abstraction levels during operation, but also provides a structured representation with temporal continuity and rich hierarchy for user behavior pattern mining through the output of intent drift trajectory chains, which significantly improves the accuracy of user behavior pattern mining results in complex operation scenarios. Attached Figure Description

[0007] Figure 1 This is a schematic diagram of the execution flow of the computer-based user behavior data mining method provided in an embodiment of the present invention.

[0008] Figure 2 This is a logical schematic diagram of a computer-based user behavior data mining method provided in an embodiment of the present invention.

[0009] Figure 3 This is a schematic diagram of exemplary hardware and software components of a server provided in an embodiment of the present invention. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a computer-based user behavior data mining method according to an embodiment of the present invention. The following is a description of the method in conjunction with the appendix. Figure 2 This method is described in detail. It can be applied to user behavior analysis, product optimization, anomaly detection, and other fields, and is particularly suitable for identifying user intent switching patterns and evolution trajectories from multi-dimensional operational behavior sequences. An example of user operation behavior analysis for a mobile application is used for illustration, but this application scenario is merely illustrative and does not constitute a limitation on the scope of protection of this application.

[0011] Step S110: Obtain a set of user behavior record data, which contains a sequence of operation behaviors marked with operation timestamps. Each sequence of operation behaviors consists of continuously collected multi-dimensional operation events.

[0012] Specifically, the system acquires a series of user behavior records generated within a preset time period through an application background log collection service. This user behavior record data set is stored in a standardized log format, with each record representing a sequence of user behavior events, composed of multiple consecutively collected event records arranged chronologically. Each event includes an event timestamp, recording the precise time of the event; an event type, describing the specific action performed by the user; and event parameters, recording additional information related to the operation. In this embodiment, taking user behavior within a mobile application as an example, the sequence of user behavior events covers various types of operations such as clicking, swiping, long-pressing, and input.

[0013] Step S120: Perform operation intent boundary recognition processing on the user behavior record data set to obtain the intent switching boundary position set and operation hesitation interval distribution set corresponding to the operation behavior sequence. The intent switching boundary position set is used to define the segmentation position of adjacent operation intent segments in the operation behavior sequence, and the operation hesitation interval distribution set is used to characterize the duration distribution state of each operation hesitation event in the operation behavior sequence.

[0014] By analyzing the changing patterns of operation event types in user action sequences, the system identifies the locations where user intent shifts, and simultaneously captures the hesitation and pauses that occur during intent switching. This step comprises several sub-steps that detail the logic for identifying intent boundaries and the method for extracting hesitation intervals.

[0015] Step S121: Extract the timestamps of all operation events for each operation behavior sequence in the user behavior record data set, group the operation events according to the operation event type, and obtain an operation event time sequence group set oriented towards the operation event type. The operation event time sequence group set contains the operation event instance chain corresponding to each operation event type, arranged in ascending order of timestamp.

[0016] For each operation sequence in the user behavior record dataset, firstly, all operation event records in the sequence are traversed. From each record, the operation event type is extracted, denoted as A, and the corresponding operation timestamp, denoted as T. Then, all operation events are categorized according to their operation event type. All operation events with the same operation event type identifier are grouped together and arranged in ascending order of timestamp, forming an operation event instance chain for that operation event type. For example, all "click" events are grouped together and sorted, and all "swipe" events are grouped together and sorted. Through the above grouping operations, a set of parallel time series oriented towards operation event types is obtained, i.e., a set of operation event time series groupings oriented towards operation event types.

[0017] Step S122: Perform time interval calculation on each operation event instance chain in the operation event time sequence grouping set to generate an operation event interval sequence oriented towards operation event type. Each element of the operation event interval sequence oriented towards operation event type represents the time difference between two adjacent operation events under the same operation event type.

[0018] For each chain of operation event instances generated in step S121, traverse adjacent operation events in the chain. Let the timestamps of two adjacent operation events be T1 and T2, respectively, and calculate the time difference between them, DT = T2 - T1. Arrange the differences of all adjacent event pairs in order to form the operation event interval sequence corresponding to this operation event type. Each element in this operation event interval sequence represents the frequency or density of the same type of operation on the time axis.

[0019] Step S123: Based on the operation event interval sequence corresponding to all operation event types, perform cross-type interval mutation point detection processing, identify the time position where the interval sequences of different operation event types simultaneously show a sudden increase in interval, mark the time position as a candidate intent switching boundary, and obtain a candidate intent switching boundary set. Each candidate intent switching boundary in the candidate intent switching boundary set contains the corresponding timestamp position and the combination of operation event types that triggered the interval increase.

[0020] Perform cross-sequence linkage analysis on all operation event interval sequences generated in step S122. For each time point in the sequence, check whether the interval sequence of each operation event type shows a significant increase in interval value (i.e., a sudden decrease in operation frequency) near that time point. If a time point appears simultaneously in multiple operation event interval sequences of different types, and these sequences all show a sudden increase in interval at that point, then mark that time point as a candidate intent switching boundary. Record the timestamp of this boundary and the combination of operation event types involved in the sudden increase. Traverse the entire timeline, collect all candidate boundaries, and form a set of candidate intent switching boundaries.

[0021] Step S124: Perform intent switching coherence verification processing on the candidate intent switching boundary set, extract the operation event type distribution within the adjacent time window before and after each candidate intent switching boundary, calculate the transition probability of the operation event type distribution, and determine the candidate intent switching boundary with the transition probability lower than the preset coherence judgment threshold as the formal intent switching boundary, thereby obtaining the intent switching boundary position set. Each intent switching boundary in the intent switching boundary position set includes the boundary timestamp position and the operation event type transition features on both sides of the boundary.

[0022] For each boundary in the candidate intent switching boundary set, using the timestamp of that boundary as the center, extract the distribution of operation event types within a preset time window before and after it, forming an operation event type probability vector. The probability vector of the preceding window is denoted as P1, and the probability vector of the following window as P2. Calculate the transition probability from P1 to P2; a higher value indicates stronger consistency in intent switching. Set a consistency judgment threshold; candidate boundaries with transition probabilities lower than this threshold are determined as formal intent switching boundaries. These formal boundaries collectively constitute the intent switching boundary location set.

[0023] Step S125: Extract the pairs of operation events in the user behavior record data set whose time interval between adjacent operation events exceeds a preset hesitation judgment threshold, mark the time interval between the operation event pairs as operation hesitation intervals, and collect the duration data and start timestamp data of all operation hesitation intervals to form an operation hesitation interval distribution set. The operation hesitation interval distribution set includes the start timestamp and hesitation duration of each operation hesitation interval.

[0024] The algorithm iterates through adjacent operation event pairs for each operation sequence in the user behavior record dataset, calculating the time interval between each pair. Operation event pairs with time intervals exceeding a preset hesitation threshold are identified, and the time interval between these two operation events is marked as an operation hesitation interval. The start timestamp and duration of this hesitation interval are recorded. After iterating through all sequences, all collected hesitation interval information is aggregated to form an operation hesitation interval distribution set.

[0025] Step S126: Perform intent segmentation processing on each operation behavior sequence according to the intent switching boundary position set. Take the operation event segment between two adjacent intent switching boundaries as an operation intent segment. Calculate the intent segment duration and operation event density within the intent segment for each operation intent segment to obtain the intent segment feature data set.

[0026] Using the set of intent switching boundary locations determined in step S124 as segmentation points, each operation behavior sequence in the user behavior record data set is segmented. An operation event sequence segment located between two adjacent intent switching boundaries is defined as an operation intent segment. For each operation intent segment, its duration (i.e., the difference between the segment's end timestamp and start timestamp) and the total number of operation events contained within the segment are calculated, and the operation event density Rho within the segment is further calculated. These feature data are associated with the corresponding segment identifiers to form an intent segment feature data set.

[0027] Step S127: Perform hesitation type classification processing on each hesitation interval in the set of hesitation interval distributions, extract the direction of change of operation event density of adjacent operation intention segments before and after the hesitation interval, divide the hesitation interval into increasing hesitation intervals and decreasing hesitation intervals according to the direction of change of operation event density, and use the division result of increasing hesitation intervals or decreasing hesitation intervals as hesitation transition type identifiers to obtain a set of hesitation interval distributions containing hesitation transition type identifiers.

[0028] Iterate through each hesitation interval in the set of operational hesitation interval distributions generated in step S125. Based on the start and end timestamps of the hesitation interval, locate its preceding and following operational intent segments in the operational behavior sequence. Calculate the operational event densities of the preceding and following segments, denoted as R1 and R2, respectively. Calculate the density difference between the two, DR = R2 - R1. If DR is greater than 0, mark the hesitation interval as an increasing hesitation interval; otherwise, mark it as a decreasing hesitation interval. Attach this classification result as a hesitation transition type identifier to the attributes of the hesitation interval, and update the set of operational hesitation interval distributions.

[0029] Step S130: Construct a user intent switching association network based on the intent switching boundary location set and the operation hesitation interval distribution set. The user intent switching association network uses operation intent segments as node units and hesitation transition relationships between operation intent segments as connecting edges. The attributes of the node units include the duration of the intent segment and the density of operation events within the intent segment. The attributes of the connecting edges include the duration of the hesitation transition and the hesitation transition type identifier.

[0030] The operational intent segments and hesitation transition relationships identified in step S120 are abstracted into a directed graph structure, namely the user intent switching association network. Each node in the network corresponds to an operational intent segment, and each edge corresponds to the hesitation transition relationship between segments. The attributes of the nodes and the attributes of the edges together describe the behavioral pattern of user intent switching.

[0031] Step S131: Using each operation intent segment defined by the intent switching boundary position set as a node construction unit, extract the intent segment duration and operation event density within the intent segment as node attribute parameters to generate a node set of the user intent switching association network. Each node in the node set corresponds to an operation intent segment and carries the node attribute parameters.

[0032] Iterate through all the operation intent segments generated in step S126, creating a corresponding node for each segment. Extract the duration of the intent segment and the density of operation events within the intent segment from the segment's feature data, using these as attribute parameters for the node. Aggregate all nodes to form a set of nodes in the user intent switching association network.

[0033] Step S132: Based on the timestamp position of each operation hesitation interval in the operation hesitation interval distribution set, determine the preceding operation intention segment and the following operation intention segment connected to each operation hesitation interval, establish a directed connection edge between the node corresponding to the preceding operation intention segment and the node corresponding to the following operation intention segment, and obtain a set of connection edges. The direction of the directed connection edge is from the preceding operation intention segment to the following operation intention segment.

[0034] Iterate through each hesitation interval in the updated hesitation interval distribution set after step S127. Using the start and end timestamps of each hesitation interval, locate its preceding and following segments in the intent segment sequence. Then, in the node set created in step S131, find the nodes corresponding to the preceding and following segments, and establish a directed connection edge between these two nodes, pointing from the preceding node to the following node. Collect all these connection edges to form the connection edge set of the user intent switching association network.

[0035] Step S133: Extract the duration of hesitation for each operation hesitation interval as the duration of hesitation transition for the corresponding directed connection edge, extract the hesitation transition type identifier carried by each operation hesitation interval as the hesitation transition type identifier attribute for the corresponding directed connection edge, and attach the duration of hesitation transition and the hesitation transition type identifier attribute to the corresponding connection edge of the connection edge set.

[0036] For each directed connection edge created in step S132, the hesitation duration and hesitation transition type identifier are extracted from the corresponding operation hesitation interval record and assigned as attributes of the edge to the edge object. In this way, each edge not only represents the connection relationship between nodes, but also carries the strength and type information of the connection relationship.

[0037] Step S134: Perform normalization mapping processing on the operation event density within the intent segment of each node in the node set, map the operation event density within the intent segment to a preset density scale range to obtain the normalized operation event density, and update the normalized operation event density to the node's operation density normalization attribute.

[0038] Since the range of operation event density values ​​may vary significantly within different operation intent segments, it is necessary to normalize the density values ​​for easier subsequent calculations. A target density scale range [0, 1] is set, and the operation event density D of each node is linearly mapped using the formula Dn = (D - Dmin) / (Dmax - Dmin), where Dmin and Dmax are the minimum and maximum densities among all nodes. The normalized value Dn is then updated to reflect the normalized operation density attribute of that node, replacing the original density value.

[0039] Step S135: Based on the hesitant transition type identifier of each connecting edge in the set of connecting edges, the connecting edges are divided into increasing connecting edges and decreasing connecting edges. The increasing connecting edges are assigned positive transition weights, and the decreasing connecting edges are assigned negative transition weights, thus obtaining a set of connecting edge weight attributes.

[0040] Based on the hesitation transition type identifier attached to each edge in step S133, the connecting edges are divided into two categories: increasing connecting edges and decreasing connecting edges. An increasing connecting edge is assigned a positive weight value Wp, and a decreasing connecting edge is assigned a negative weight value Wn. The absolute value of the weight can be set according to the actual situation, and these weight values ​​will be used in subsequent network propagation calculations.

[0041] Step S136: Perform time granularity alignment processing on the duration of the intent segment of each node in the node set, divide the duration of the intent segment into multiple time slice units according to the preset time slice step size, generate the duration slice representation of the node based on the number of time slice units, and attach the duration slice representation to the attributes of the corresponding node.

[0042] The duration Td of a node's intent segment is divided into segments of fixed time slice length Ls, and the number of time slices contained in the node, Ns = floor(Td / Ls), is calculated. This number of slices, Ns, is used as the duration slice representation of the node and appended to the node's attributes. This representation converts continuous time values ​​into discrete counting units, facilitating subsequent hierarchical analysis.

[0043] Step S137: Combine the set of nodes carrying node attribute parameters, operation density normalization attribute and duration slice representation with the set of connection edges carrying hesitation transition duration attribute, hesitation transition type identifier attribute and connection edge weight attribute set to generate user intent switching association network.

[0044] All node and edge objects generated in steps S131 to S136 are integrated into a graph structure, ultimately forming a complete user intent switching association network.

[0045] Step S140: Call the preset intent depth parsing model to perform intent hierarchy expansion processing on the user intent switching association network, map the operation intent segments to the intent hierarchy space, and generate a user intent hierarchy distribution map. The user intent hierarchy distribution map includes the expansion path of each intent level in the intent hierarchy space and the mounting coordinates of each operation intent segment in the intent hierarchy space.

[0046] The user intent switching association network generated in step S137 is input into a pre-trained intent deep parsing model. This intent deep parsing model adopts a graph attention network architecture, which maps the nodes in the association network to a low-dimensional intent hierarchy space through multi-layer aggregation and propagation, and outputs the position coordinates of each node in this space and the level to which it belongs.

[0047] Step S141: Perform neighborhood feature aggregation processing on each node in the user intent switching association network, extract the set of preceding neighbor nodes and the set of succeeding neighbor nodes for each node, and calculate the neighborhood operation density difference vector of the node based on the operation density normalization attribute of the preceding neighbor node set and the operation density normalization attribute of the succeeding neighbor node set. The neighborhood operation density difference vector is used to characterize the gradient of the operation density change of the node in the association network.

[0048] The intent deep parsing model first performs a neighborhood feature aggregation operation on each node v in the network. It extracts the set of all its preceding neighbor nodes Nprev(v) and the set of its following neighbor nodes Nnext(v) from the interconnected network. It then calculates the average values ​​Davg1 and Davg2 of the normalized operation density attributes of these neighbor nodes. Finally, it calculates the difference DV = Davg2 - Davg1, which is the neighborhood operation density difference vector for node v.

[0049] Step S142: Input the normalized operation event density, duration slice representation and neighborhood operation density difference vector of each node in the node set of the user intent switching association network into the intent encoding layer of the intent deep parsing model. Perform intent feature encoding processing on each node through the hierarchical encoder of the intent encoding layer to obtain the node intent encoding vector. The node intent encoding vector contains the local operation density features and neighborhood density change features of the node.

[0050] The three feature vectors generated in steps S131, S136, and S141 are concatenated into a complete input feature vector, which is then input into the intent encoding layer. This intent encoding layer employs a multilayer perceptron or graph neural network structure, extracting higher-dimensional abstract features through nonlinear transformations, and ultimately outputting a node intent encoding vector for each node. This node intent encoding vector integrates the node's local density information and neighborhood change information.

[0051] Step S143: Perform weight propagation path aggregation processing on the weight attributes of each connection edge in the set of connection edges of the user intent switching associated network, and accumulate the weight attributes of the connection edges layer by layer along the direction of the directed connection edges of the user intent switching associated network to generate the cumulative weight propagation value of each node, and use the cumulative weight propagation value as the intent level depth indicator parameter of the node.

[0052] Starting with the node with an in-degree of 0, a topological sort is performed along the direction of the network. For each node v, the cumulative weight propagation value of all its predecessor nodes is added to the corresponding edge weights to obtain the cumulative weight propagation value of that node. This cumulative weight propagation value reflects the depth of the node in the network and is used as an indicator of the node's hierarchy in the intent hierarchy space.

[0053] Step S144: Based on the intent level depth indicator parameter, perform hierarchical clustering processing on the node intent encoding vectors of all nodes, and classify the nodes whose intent level depth indicator parameters fall into the same level interval to the same intent level, thereby obtaining the intent level division result. The intent level division result includes the level identifier of each intent level and the subset of nodes belonging to each intent level.

[0054] Based on the intent level depth indicator parameter for each node, nodes falling within the same numerical range are grouped into the same intent level. Each level is assigned a unique level identifier, and a record is kept of all node subsets contained in that level.

[0055] Step S145: Perform hierarchical center anchoring processing on the node subset of each intent level, extract the mean of the node intent encoding vector of all nodes in the node subset as the hierarchical cluster center position of the intent level, and obtain the set of cluster center positions of each intent level in the intent hierarchy space.

[0056] For each intent level, extract the node intent encoding vectors of all nodes it contains, calculate the arithmetic mean of these vectors, and obtain the cluster center position for that level. Collect the center positions of all levels to form the set of cluster center positions for each intent level in the intent hierarchy space.

[0057] Step S146: Perform hierarchical attribution mapping processing on the node corresponding to each operation intent segment. Calculate the distance between the node intent encoding vector of the node and the hierarchical cluster center position of each intent level. Select the intent level with the smallest distance as the attribution level of the node. Mount the coordinates of the node in the intent hierarchy space to the hierarchical expansion path corresponding to the attribution level to obtain the mounted coordinates of each operation intent segment in the intent hierarchy space.

[0058] For each node, calculate the Euclidean distance between its node intent encoding vector and the cluster center location of each level generated in step S145. Select the level with the smallest Euclidean distance as the node's belonging level, and map the node to the corresponding position of that level in the intent hierarchy space.

[0059] Step S147: Combine the cluster center locations of each intent level in the intent hierarchy space, the hierarchical expansion path of each intent level, and the mounting coordinates of each operation intent segment in the intent hierarchy space to generate a user intent hierarchy distribution map.

[0060] The outputs of steps S145 and S146 are combined to form a user intent hierarchy distribution map. This user intent hierarchy distribution map is a set of points and paths in a three-dimensional or two-dimensional space, which intuitively shows the distribution and connection relationships of user operation intentions in the hierarchy space.

[0061] Step S150: Extract the intent drift trajectory chain corresponding to the operation behavior sequence according to the user intent hierarchy distribution map. The intent drift trajectory chain is composed of an intent hierarchy identifier sequence arranged in the order of operation timestamps. Generate user behavior pattern mining results based on the intent drift trajectory chain.

[0062] Using the user intent hierarchy distribution map generated in step S147, the user's operation behavior sequence is mapped to a movement trajectory in the intent hierarchy space, and patterned behavioral features are extracted from it.

[0063] Step S151: Based on the mounting coordinates of each operation intent segment in the intent hierarchy space in the user intent hierarchy distribution map, determine the intent hierarchy identifier corresponding to each operation intent segment, arrange each operation intent segment in chronological order in the operation behavior sequence, and generate an original intent hierarchy identifier sequence composed of intent hierarchy identifiers.

[0064] Traverse all the mounting coordinates in the user intent hierarchy distribution map, and retrieve the intent hierarchy identifier to which each operation intent segment belongs based on its mounting coordinates. Arrange the operation intent segments in order according to their corresponding operation timestamps to obtain a sequence of intent hierarchy identifiers, i.e., the original intent hierarchy identifier sequence.

[0065] Step S152: Perform hierarchical transition point detection processing on the original intent hierarchical identifier sequence, identify the sequence position where adjacent intent hierarchical identifiers change in the original intent hierarchical identifier sequence as hierarchical transition points, extract the intent hierarchical identifier before the transition and the intent hierarchical identifier after the transition corresponding to each hierarchical transition point, and obtain a set of hierarchical transition points.

[0066] Traverse the original intent level identifier sequence and compare the values ​​of two adjacent elements. If the values ​​of two elements are different, record the corresponding position as a level transition point and record the level identifier before and after the transition.

[0067] Step S153: Based on the set of hierarchical transition points, extract the node attributes of the operation intent segment corresponding to each hierarchical transition point in the user intent switching association network, obtain the duration of the intent segment before the transition and the duration of the intent segment after the transition, and calculate the duration ratio parameter of the duration of the intent segment before and after the transition.

[0068] For each level transition point, identify its corresponding preceding and following operation intent segments. Extract the duration of these two segments from the node attributes of the user intent switching association network, denoted as Tb and Ta. Calculate the proportional parameter RP = Ta / Tb.

[0069] Step S154: Based on the pre-jump intention level identifier and post-jump intention level identifier corresponding to each level jump point in the set of level jump points, determine the jump direction vector of the level jump in the intention level space of the user intention level distribution map. The jump direction vector points from the level cluster center position of the pre-jump intention level to the level cluster center position of the post-jump intention level.

[0070] For each hierarchical transition point, extract the center coordinates Cb of the intent level before the transition and the center coordinates Ca of the intent level after the transition from the set of cluster center locations in the user intent hierarchy distribution map. Calculate the vector pointing from Cb to Ca, which serves as the transition direction vector.

[0071] Step S155: Extract the hesitation transition type identifier and hesitation transition duration of the connection edges of the pre-transition operation intent segment and post-transition operation intent segment corresponding to the connection level transition point in the user intent switching association network. Use the hesitation transition type identifier as the transition transition type of the level transition and the hesitation transition duration as the transition transition duration of the level transition.

[0072] Locate the directed edge connecting the pre-transition and post-transition segments in the user intent switching network. Extract the hesitant transition type identifier and hesitant transition duration from the attributes of this edge.

[0073] Step S156: Combine the duration ratio parameter, transition direction vector, transition type and transition duration into hierarchical transition description data, and concatenate the hierarchical transition description data of all hierarchical transition points in chronological order to generate an intention drift trajectory chain. The intention drift trajectory chain contains the transition path information of the operation behavior sequence in the intention hierarchy space.

[0074] The information extracted in steps S153, S154, and S155 is merged into a structured data record. The data records of all hierarchical transition points are arranged in chronological order to form the intended drift trajectory chain.

[0075] Step S157: Perform pattern induction processing on the intent drift trajectory chain, extract the recurring jump path segments in the intent drift trajectory chain, mark the recurring jump path segments as user behavior pattern units, and collect all user behavior pattern units to form the user behavior pattern mining results.

[0076] A sequence pattern mining algorithm was used to analyze the intent drift trajectory chain and identify the recurring transition path sequences in different time periods. These recurring path segments were defined as user behavior pattern units, and the user behavior pattern mining results were summarized and output.

[0077] Step S210: Obtain the auxiliary behavior environment data set, which includes the operation interface layout status data and operation trigger area distribution data during the operation behavior sequence collection period.

[0078] The application's user interface rendering engine acquires snapshot data of the user interface synchronized with the user action sequence collection period. The user interface layout state data is stored in the application's view hierarchy cache as a hierarchical graph structure, recording the type identifier of each interface element, its relative position coordinates within its parent container, its width and height values, and the parent-child containment relationships and sibling stacking order between elements. The action trigger area distribution data is stored as a heatmap layer overlaid on the layout graph, recording the set of boundary polygon coordinates of all triggerable interactive controls such as buttons, links, and input boxes in the interface, as well as the number of times each interactive control is triggered in the current action sequence. All of the above data was collected with the user's informed consent and has undergone data anonymization to ensure it does not contain personally identifiable information.

[0079] Step S220: Perform interface operation hotspot analysis on the auxiliary behavior environment data set to generate an operation interface hotspot distribution map. The operation interface hotspot distribution map includes the operation frequency density distribution of each operation triggering area on the operation interface and the spatial adjacency relationship between the operation triggering areas.

[0080] The operation trigger area distribution data obtained in step S210 is mapped to the coordinate space of the operation interface layout state data. For each operation trigger area (e.g., button, icon, link), the total number of times it is triggered by the user in the operation behavior sequence is counted, and divided by the total length of the operation behavior sequence to obtain the operation frequency density of that area. Then, the Euclidean distance between the center points of all trigger areas is calculated, and spatial adjacency relationships between trigger areas are established using the reciprocal of the distance as the weight. The operation frequency density is mapped to color depth, and the spatial adjacency relationships are mapped to connecting lines to generate a heat map of the operation interface.

[0081] Step S230: Based on the hotspot distribution map of the operation interface and the set of intent switching boundary positions, perform spatial intent association analysis and process, map each intent switching boundary to the corresponding spatial position of the operation interface, and generate an intent switching spatial distribution map. The intent switching spatial distribution map includes the spatial coordinates of each intent switching boundary on the operation interface and the operation trigger area identifier corresponding to the spatial coordinates.

[0082] Iterate through each boundary in the intent switching boundary position set generated in step S124 and extract the operation timestamp corresponding to that boundary. Based on this timestamp, extract the interface snapshot at the corresponding moment from the operation interface layout state data. Locate the regions triggered by the preceding and following operation events in the operation interface hotspot distribution map, calculate the midpoint coordinates of the line connecting the geometric centers of these two regions, use these midpoint coordinates as the mapping position of the intent switching boundary on the operation interface, and record the operation triggering area identifier corresponding to this position. After the traversal is complete, generate the intent switching spatial distribution mapping.

[0083] Step S240: Perform spatial constraint fusion processing on the spatial distribution mapping of intent switching and the user intent switching association network, add spatial operation trigger area identifier to the node attributes of the user intent switching association network, and add spatial jump distance parameter to the connection edge attributes to obtain a user intent switching association network with spatial constraints.

[0084] For each node in the user intent switching association network, based on the time range of its corresponding operation intent segment, the intent switching boundary overlapping with that time period is found in the intent switching spatial distribution map, and the operation triggering region identifier corresponding to the boundary is extracted as the spatial operation triggering region identifier of that node. For each directed connection edge in the network, the Euclidean distance between the geometric centers corresponding to the operation triggering region identifiers of its two end nodes is calculated, and this distance is used as the spatial jump distance parameter of that edge.

[0085] Step S250: Based on the user intent switching association network carrying spatial constraints, perform spatial weighted correction processing on the user intent hierarchy distribution map. Adjust the mounting coordinates of the nodes in the intent hierarchy space according to the spatial operation trigger area identifier of the nodes and the spatial jump distance parameter of the connecting edges to obtain the spatially corrected user intent hierarchy distribution map. Extract the spatial constraint intent drift trajectory chain corresponding to the operation behavior sequence based on the spatially corrected user intent hierarchy distribution map. Generate user behavior pattern mining results that integrate spatial factors based on the spatial constraint intent drift trajectory chain.

[0086] The spatial operation triggering region identifier and spatial jump distance parameter added in step S240 are input as additional feature dimensions into the intent hierarchy expansion model in step S140. When calculating the node mounting coordinates, the model converts the spatial region identifier into a one-hot encoded vector and concatenates it with the original node attribute feature vector. Simultaneously, the spatial jump distance parameter is used as a weight constraint for the connecting edges, adjusting the relative distance between the nodes at both ends of the edge in the intent hierarchy space. After recalculation, a spatially corrected user intent hierarchy distribution map is obtained. Based on this distribution map, the intent drift trajectory chain is re-extracted to obtain the mining results that incorporate spatial factors.

[0087] Step S310: Obtain a set of historical user behavior records, which includes a sequence of historical operation behaviors of multiple users and a user identity identifier associated with each historical operation behavior sequence.

[0088] The system extracts behavioral data from other users over a past period from a behavioral database. Each record contains a complete sequence of actions and the identifier of the user who initiated that sequence. The collection and use of this behavioral data have been authorized and anonymized.

[0089] Step S320: Perform operation intent boundary recognition processing on the historical user behavior record data set to obtain the historical intent switching boundary position set and the historical operation hesitation interval distribution set corresponding to the historical operation behavior sequence of each user.

[0090] For each user's action sequence in the historical data, repeat step S120 for intent boundary identification and hesitation interval extraction. Generate a corresponding set of historical intent switching boundary positions and a set of historical operation hesitation interval distributions for each user.

[0091] Step S330: Construct a historical user intent switching association network for each user based on the historical intent switching boundary location set and the historical operation hesitation interval distribution set. Extract common node structures and common connection edge structures from all historical user intent switching association networks to construct a group intent switching benchmark network. Perform deviation measurement processing on the user intent switching association network according to the group intent switching benchmark network. Calculate the attribute deviation vector between the node attributes of the user intent switching association network and the node attributes of the group intent switching benchmark network, as well as the structural deviation parameter between the connection edge structure of the user intent switching association network and the connection edge structure of the group intent switching benchmark network.

[0092] For each user's historical intent switching boundary location set and historical operation hesitation interval distribution set, execute steps S130 to S137 to construct each user's historical user intent switching association network. Perform union processing on all historical networks to extract node attributes (i.e., common intent segment features) and connection edge structures (i.e., common hesitation transition patterns) that appear in all networks, forming a group intent switching baseline network. Then, calculate the Euclidean distance between the target user's user intent switching association network and the group baseline network in terms of node attributes, as the attribute deviation vector; simultaneously calculate the difference between the two networks in the adjacency matrix, as the structural deviation parameter.

[0093] Step S340: Input the attribute deviation vector and the structural deviation parameter into a preset user intent deviation classification model. The user intent deviation classification model classifies users into preset intent deviation categories based on the attribute deviation vector and the structural deviation parameter to obtain user intent deviation category identifiers. The user intent deviation category identifiers are then appended to the user behavior pattern mining results to generate user behavior pattern mining results carrying intent deviation category identifiers.

[0094] The user intent deviation classification model employs multi-class logistic regression or support vector machine algorithms, accepting attribute deviation vectors and structural deviation parameters as input, and outputting a category label, such as "regular user," "exploratory user," or "interrupted user." This category label is then appended to the user behavior pattern mining results generated in step S157.

[0095] Step S410: Perform network topology measurement processing on the user intent switching association network, extract the node degree distribution sequence and the connection edge betweenness distribution sequence of the user intent switching association network, identify the hub nodes in the user intent switching association network based on the node degree distribution sequence, the hub nodes are nodes whose node degree value is higher than the preset hub determination threshold, extract the intent segment duration and the operation event density within the intent segment corresponding to each hub node, and obtain the hub node intent feature data set.

[0096] Calculate the in-degree and out-degree of each node in the user intent switching association network, and sum them to obtain the node degree value Deg. Traverse all nodes, calculate the node degree value of each node, and generate a node degree distribution sequence. Set a hub determination threshold Th, and mark nodes with a node degree value Deg greater than Th as hub nodes. For each node marked as a hub node, extract the duration Td of the intent segment and the density Rho of the operation events within the intent segment from the feature data set of its corresponding operation intent segment. Associate these two values ​​with the node identifier and store them to form a hub node intent feature data set.

[0097] Step S420: Identify key connection edges in the user intent switching association network based on the connection edge betweenness distribution sequence. The key connection edge is the connection edge whose betweenness value is higher than the preset key edge judgment threshold. Extract the hesitation transition duration and hesitation transition type identifier of the operation hesitation interval corresponding to each key connection edge to obtain the key connection edge hesitation feature data set.

[0098] Calculate the betweenness value of each connection edge in the user intent switching network. This betweenness value is defined as the ratio of the number of all shortest paths passing through that connection edge to the total number of all shortest paths. Iterate through all connections, calculating the betweenness value for each edge to generate a betweenness value distribution sequence. Set a critical edge judgment threshold Tc, marking connections with betweenness values ​​greater than Tc as critical connections. For each critical connection edge, extract the hesitation transition duration Th and hesitation transition type identifier Lt from its corresponding operation hesitation interval attribute to form a critical connection edge hesitation feature data set.

[0099] Step S430: Perform hub stability analysis on each hub node in the hub node intent feature data set, extract the preceding and following connection edge sets of the hub node in the user intent switching association network, calculate the hesitant transition type identifier distribution of the preceding and following connection edge sets, and determine the type switching stability parameter of the hub node based on the hesitant transition type identifier distribution.

[0100] For each hub node in the hub node intent feature dataset, extract the set of all preceding edges Ep ending at that node and the set of all following edges Es starting at that node from the user intent switching association network. Calculate the proportion Ri of increasing hesitant edges in the preceding edge set Ep and the proportion Rd of decreasing hesitant edges in the following edge set Es. Calculate the hub node stability parameter St = 1 - abs(Ri - Rd), where abs represents the absolute value. The value of St ranges from 0 to 1; a larger value indicates higher node stability.

[0101] Step S440: Perform key edge connectivity analysis on each key edge in the key connection edge hesitation feature data set, extract the preceding and following hub nodes connected by the key connection edge, obtain the operation event density within the intent segment of the preceding hub node and the operation event density within the intent segment of the following hub node, and calculate the jump amplitude of the operation density between the preceding and following hub nodes.

[0102] For each key connection edge in the key connection edge hesitation feature data set, locate the preceding and following hub nodes connected by that edge in the user intent switching association network. Read the operation event density R1 within the intent segment of the preceding hub node and the operation event density R2 within the intent segment of the following hub node from the hub node intent feature data set generated in step S410. Calculate the operation density jump amplitude J = R2 - R1, that is, the density change between the two nodes.

[0103] Step S450: Construct user behavior rhythm stability assessment data based on the type switching stability parameter of the hub node and the operation density jump amplitude of the key connection edge, and fuse the user behavior rhythm stability assessment data with the user behavior pattern mining results to generate user behavior pattern mining results carrying rhythm stability assessment data.

[0104] The type switching stability parameter St of each hub node calculated in step S430 is combined with the operation density jump amplitude J of each key connection edge calculated in step S440. The rhythm stability evaluation value Rs of the entire user intent switching association network is calculated as Rs = sum(St) / Nh - alpha*sum(abs(J)) / Ne, where Nh is the number of hub nodes, Ne is the number of key connection edges, and alpha is the adjustment coefficient. Rs is appended as an additional feature field to the user behavior pattern mining results generated in step S157.

[0105] Step S510: Extract the spatial operation trigger area identifier of each node in the user intent switching association network carrying spatial constraints, obtain the regional center coordinates and operation frequency density of the operation trigger area corresponding to each spatial operation trigger area identifier based on the operation interface hot zone distribution map, and construct a spatial trigger area coordinate mapping table.

[0106] For each node in the associated network carrying a user intent to switch, its attached spatial operation trigger area identifier is read. The corresponding operation trigger area is located in the operation interface hotspot distribution map generated in step S220, and the geometric center coordinates and operation frequency density of that area are extracted. The spatial operation trigger area identifiers, center coordinates, and operation frequency densities of all nodes are stored in tabular form to construct a spatial trigger area coordinate mapping table.

[0107] Step S520: Map each node to the two-dimensional coordinate space of the operation interface according to the spatial triggering area coordinate mapping table to obtain the operation interface mapping coordinates of each node, calculate the Euclidean distance between the operation interface mapping coordinates of the node pairs with connecting edge relationships in the user intent switching association network, and generate a set of spatial distances between node pairs.

[0108] Iterate through each record in the spatial trigger region coordinate mapping table, extract the center coordinates corresponding to the spatial operation trigger region identifier of the node, and use these coordinates as the operation interface mapping coordinates for that node. Iterate through all directed edges in the user intent switching network, and for each preceding and succeeding node connected by an edge, obtain their operation interface mapping coordinates and calculate the Euclidean distance between these two points. Associate and store the distance calculation results of all edges with the edge identifiers to generate a set of node pair spatial distances.

[0109] Step S530: Extract the spatial jump distance parameter of each connection edge in the user intent switching association network carrying spatial constraints, perform a consistency comparison between the spatial jump distance parameter and the Euclidean distance of the corresponding node pair in the node pair spatial distance set, generate a spatial jump distance correction coefficient based on the comparison result, and update the spatial jump distance parameter of each connection edge using the spatial jump distance correction coefficient.

[0110] Read the current spatial jump distance parameter Lc from the attributes of each connection edge in the user-intended switching network. Find the corresponding Euclidean distance Le for that edge in the node pair spatial distance set. Calculate the consistency difference DL = abs(Lc - Le). If DL is less than a preset threshold, assign a correction coefficient Lam close to 1 to the edge; otherwise, assign a correction coefficient less than 1. Multiply the original spatial jump distance parameter by the correction coefficient Lam to obtain the updated spatial jump distance parameter Ln = Lam * Lc.

[0111] Step S540: Calculate the spatial offset of the mounting coordinates of each operation intent segment in the intent hierarchy space in the user intent hierarchy distribution map, extract the operation interface mapping coordinates of the corresponding node of the operation intent segment, project the operation interface mapping coordinates onto the intent hierarchy space according to a preset spatial scaling ratio, and generate a spatial offset vector.

[0112] For each operation intent segment in the user intent hierarchy distribution map, locate its corresponding node and extract the operation interface mapping coordinates of that node. Set a spatial scaling parameter S to project the operation interface coordinates onto the intent hierarchy space to obtain the projected coordinates. Calculate the difference vector between the mounted coordinates and the projected coordinates, which is the spatial offset vector.

[0113] Step S550: Superimpose the spatial offset vector onto the mounting coordinates of the corresponding operation intention segment in the intention hierarchy space to generate preliminary spatial correction mounting coordinates. Based on the spatial jump distance parameter between node pairs, adjust the relative position of the preliminary spatial correction mounting coordinates of operation intention segment pairs with connected edge relationships to ensure that the distance between operation intention segment pairs in the intention hierarchy space is proportional to the spatial jump distance parameter, thus obtaining the spatial correction mounting coordinates.

[0114] The spatial bias vector generated in step S540 is superimposed on the original mounting coordinates to obtain the preliminary spatially corrected mounting coordinates. Then, for each pair of directed connections in the user intent switching network, the updated spatial jump distance parameter Ln is used as a constraint. The relative positions of the preliminary spatially corrected mounting coordinates of the two end nodes are adjusted using an energy minimization method, so that the adjusted coordinate distance is proportional to Ln. After the adjustment is completed, the final spatially corrected mounting coordinates are obtained.

[0115] Step S560: Replace the initial spatial correction mount coordinates of all operation intent segments with spatial correction mount coordinates, update the hierarchical cluster center position of each intent level in the intent hierarchy space and the hierarchical expansion path of each intent level, and generate a spatial correction user intent hierarchy distribution map.

[0116] Replace the original mount coordinates with the spatially corrected mount coordinates generated in step S550. Re-execute steps S145 and S146 to calculate the updated cluster center locations and hierarchical expansion paths for each intent level. Finally, generate a spatially corrected user intent hierarchy distribution map.

[0117] Step S610: Perform intent segment type labeling processing on the nodes in the user intent switching association network. Based on the density of operation events within the intent segment and the duration of the intent segment corresponding to each node, the nodes are divided into dense operation nodes and sparse operation nodes to obtain a node type label set.

[0118] For each node in the user intent switching association network, obtain the operation event density Dd and duration Td within the corresponding operation intent segment. Set a density threshold Td1 and a duration threshold Tt1. If Dd is greater than Td1 and Td is less than Tt1, mark the node as a dense operation node; otherwise, mark it as a sparse operation node. Store the classification results of all nodes in a node type label set.

[0119] Step S620: Perform hesitation interval context extraction processing on each hesitation interval in the set of hesitation interval distributions, extract the node type markers of the preceding and following operation intent paragraphs corresponding to the hesitation interval, and generate hesitation interval context type pairs, which are composed of preceding node type markers and following node type markers.

[0120] Iterate through each hesitation interval in the set of hesitation interval distributions, and locate the corresponding preceding and following operation intent paragraphs based on its timestamp. Read the node type tags of these two paragraphs from the node type tag set, and combine them into an ordered pair, which is the hesitation interval context type pair.

[0121] Step S630: Based on the context type of the hesitation interval, divide the operation hesitation interval into hesitation intervals of the same type and hesitation intervals of different types. The preceding node type label and the following node type label of the hesitation interval of the same type are the same, and the preceding node type label and the following node type label of the hesitation interval of different types are different.

[0122] For each hesitant interval, check if the two tags in its hesitant interval context type pair are equal. If the preceding tag and the following tag are equal, classify the interval as a hesitant interval of the same type; otherwise, classify it as a hesitant interval of a different type.

[0123] Step S640: Perform distribution statistical processing on the duration of hesitation in the same type of hesitation interval and the hesitation in the opposite type of hesitation interval to obtain the hesitation duration distribution of the same type of hesitation interval and the hesitation duration distribution of the opposite type of hesitation interval. Calculate the hesitation duration difference parameter between the types based on the hesitation duration distribution of the same type of hesitation interval and the hesitation duration distribution of the opposite type of hesitation interval.

[0124] The durations of all hesitant intervals of the same type are collected into a set, and the mean M1 and standard deviation S1 of this set are calculated to form a hesitant duration distribution of the same type. Similarly, the mean M2 and standard deviation S2 of hesitant intervals of different types are calculated to form a hesitant duration distribution of different types. The hesitant duration difference parameter Dv between types is calculated as Dv = (M2 - M1) / (S1 + S2). Here, M1 represents the mean of the hesitant duration of the same type, S1 represents the standard deviation of the hesitant duration of the same type, M2 represents the mean of the hesitant duration of different types, and S2 represents the standard deviation of the hesitant duration of different types. This parameter is used to quantify the degree of difference in hesitant duration between different node types.

[0125] Step S650: Extract the intent level identifier sequence from the intent drift trajectory chain, align and match the positions where adjacent intent level identifiers in the intent level identifier sequence transition with the positions of the heterogeneous hesitation intervals, and determine the hesitation interval context type pair and hesitation duration of the operation hesitation interval corresponding to each intent level transition.

[0126] Traverse the intent drift trajectory chain, recording the location (i.e., transition point) where each adjacent intent level has a different identifier. For each transition point, find the heterogeneous hesitation interval with the closest timestamp to that transition point in the operation hesitation interval distribution set, and establish a correspondence. Record the hesitation interval context type pair and hesitation duration for that hesitation interval.

[0127] Step S660: Generate intent transition cost evaluation data based on the difference parameter of hesitation duration between the types and the context type pair and hesitation duration of each intent level transition, and attach the intent transition cost evaluation data to the corresponding transition path segment of the intent drift trajectory chain.

[0128] For each transition point, calculate its transition cost Ct = Dv * Tv, where Dv is the difference in hesitation duration between types defined in step S640, and Tv is the hesitation duration corresponding to that transition point. Append this cost value to the metadata of the corresponding transition path segment in the intended drift trajectory chain.

[0129] Step S670: Perform transition cost annotation processing on the user behavior pattern mining results based on the intent drift trajectory chain with the intent transition cost evaluation data attached, and generate user behavior pattern mining results with transition cost annotation.

[0130] The transition cost values ​​generated in step S660 are used as additional annotation fields to update the description of each transition path segment in the user behavior pattern mining results. The final output is the mining results carrying transition cost annotations.

[0131] Step S710: Obtain the operation interface screen recording data corresponding to the user behavior record data set. The operation interface screen recording data includes a sequence of interface screen frames synchronized with the operation behavior sequence time. Each interface screen frame carries a frame timestamp aligned with the operation event timestamp.

[0132] The application's screen recording interface is used to acquire screen recording data synchronized with the user's action sequence. This screen recording data is stored in video stream format, with each frame accompanied by a frame timestamp accurate to milliseconds, perfectly aligned with the operation timestamps of the operation events in the action sequence. Each frame contains complete pixel information of the current interface, used for subsequent interface control layout parsing. The acquisition of screen recording data has been explicitly authorized by the user and employs encrypted storage and access control mechanisms to ensure data security and privacy.

[0133] Step S720: Perform interface control layout parsing processing on the interface screen frame sequence, extract the control boundary outline data and control type identifier of the operation control in each interface screen frame, and generate an interface control layout snapshot sequence. Each frame interface control layout snapshot in the interface control layout snapshot sequence contains a set of control boundary outline data and control type identifier of all operation controls in the interface at the corresponding frame time.

[0134] For each frame in the interface frame sequence, the RGB pixel data is first converted to grayscale and normalized to ensure pixel values ​​fall within a standard range. Then, a region proposal network-based object detection model is used to identify interactive controls in the screen. This object detection model uses a residual network structure as its backbone, extracting multi-scale feature maps from the feature pyramid network and generating candidate bounding boxes through the region proposal network. The candidate bounding boxes are input into a classification regression head, which outputs the boundary coordinates of the controls and their corresponding control type labels (e.g., buttons, input boxes, sliders, checkboxes, etc.). For each identified control, its boundary contour data in the image coordinate system is recorded (i.e., the top-left X-coordinate, top-left Y-coordinate, width, and height of the bounding box). The boundary data and type labels of all identified controls in the same frame are summarized to form a snapshot of the interface control layout for that frame. All frame snapshots are arranged chronologically to form a sequence of interface control layout snapshots.

[0135] Step S730: Perform control state transition detection processing based on the interface control layout snapshot sequence, identify the operation controls that have undergone control state changes between adjacent interface control layout snapshots, the control state changes include control position movement changes and control visual attribute changes, extract the control state data before and after the change of the operation controls in the interface control layout snapshot, and generate a control state transition event sequence.

[0136] Traverse the snapshot sequence of adjacent frames in the UI control layout snapshot sequence, denoted as St1 and St2. For each control detected in St1, find the corresponding control in St2 by matching the control type identifier and spatial position. Calculate the offsets of these two controls in the bounding box coordinates: DX=abs(X1-X2) and DY=abs(Y1-Y2), as well as the width change DW=abs(W1-W2) and height change DH=abs(H1-H2). Simultaneously, calculate the differences in visual attributes between the two controls, such as the difference DC between their average pixel color values. If any of DX, DY, DW, DH, or DC exceeds a preset threshold, the control is considered to have undergone a state transition. Record the control's identifier, the bounding box data before the change, and the bounding box data after the change, and associate this record with a timestamp t to form a control state transition event. Traverse all adjacent frames, arrange all detected transition events in chronological order, and generate a control state transition event sequence.

[0137] Step S740: Perform event timing correlation analysis on the sequence of control state transition events and the sequence of operation behaviors, establish a causal relationship between each control state transition event and the operation event with the closest timestamp, and generate an association mapping table between operation events and control state transition events. The association mapping table between operation events and control state transition events includes the timestamp of each operation event, the operation event type, and the control state data before and after the change of the control state transition event associated with that operation event.

[0138] Align each event in the control state transition event sequence with the operation event in the operation behavior sequence using its timestamp. For each operation event in the operation behavior sequence, using its operation timestamp Te as the center, search for the control state transition event in the control state transition event sequence with the closest timestamp and a time difference not exceeding a preset window length (e.g., 500 milliseconds). If such an event is found, establish a causal relationship between the two and record it in a mapping table. Each row of the mapping table contains the operation event's timestamp Te, operation event type Type, and the associated control state transition event's before-change bounding box data B1 and after-change bounding box data B2. If no associated event is found, mark the operation event as unassociated.

[0139] Step S750: Based on the association mapping table between the operation events and control state transition events, perform control context enhancement processing on the user intent switching association network. For each node attribute in the user intent switching association network, append the sequence of control state transition events associated with all operation events within the corresponding operation intent paragraph to obtain a set of node enhanced attributes carrying control state transition context information. Perform control operation semantic encoding processing on the node enhanced attribute set. Input the control state transition event sequence of each node into a preset control semantic encoder to generate a node control semantic encoding vector. The node control semantic encoding vector is used to characterize the interface control operation semantic features corresponding to the operation intent paragraph.

[0140] For each node in the user intent switching association network, based on the time range of its corresponding operation intent segment, the control state transition event records corresponding to all operation events within that time period are extracted from the association mapping table. These records are arranged into a sequence in chronological order and appended to the node's attributes to form a node augmented attribute set. Then, this node augmented attribute set is input into a pre-trained control semantic encoder, which consists of three stacked Transformer encoder layers, each containing a multi-head self-attention mechanism and a feedforward neural network. The encoder transforms each event in the state transition event sequence into an embedding vector, then captures the temporal dependencies between events through the self-attention mechanism, and finally outputs a fixed-length feature vector as the node control semantic encoding vector.

[0141] Step S760: Use the node control semantic encoding vector to perform control semantic weighting correction processing on the mounting coordinates of each operation intent segment in the intent hierarchy space of the user intent hierarchy distribution diagram. Superimpose the node control semantic encoding vector onto each dimension component of the mounting coordinate according to the preset control semantic weighting coefficient to generate control semantic corrected mounting coordinates.

[0142] Multiply the semantic encoding vector of each node control by a preset semantic weighting coefficient (e.g., 1.0), and then add the resulting vector element-wise to each dimension of the node's mounting coordinates in the intent hierarchy space. Specifically, if the mounting coordinates are (Xm, Ym, Zm) and the encoding vector is (V1, V2, V3), then the corrected coordinates are (Xm+V1, Ym+V2, Zm+V3). This yields the corrected semantic mounting coordinates for the control.

[0143] Step S770: Update the mounting coordinates of each operation intent segment in the user intent hierarchy distribution map to the control semantic correction mounting coordinates, recalculate the hierarchical cluster center position of each intent level in the intent hierarchy space, obtain the control semantic enhanced user intent hierarchy distribution map, extract the control semantic constraint intent drift trajectory chain corresponding to the operation behavior sequence based on the control semantic constraint intent drift trajectory chain, and generate the user behavior pattern mining result that integrates interface control semantics based on the control semantic constraint intent drift trajectory chain.

[0144] Replace the original mounting coordinates with the control semantic correction coordinates generated in step S760. Re-execute the operation in step S145 to calculate the cluster center positions for each intent level; that is, for each intent level, take the arithmetic mean of the corrected mounting coordinates of all nodes within that level as the new cluster center position. Then, re-execute step S146 to reassign each node to the nearest level, generating a control semantic enhancement user intent hierarchy distribution map. Based on this distribution map, re-extract the intent drift trajectory chain, and finally generate the user behavior pattern mining results that integrate interface control semantics.

[0145] Step S810: Perform network snapshot serialization processing on the user intent switching association network. Using the timestamp of each intent switching boundary in the intent switching boundary position set as the snapshot segmentation point, the user intent switching association network is divided into multiple network time snapshots along the time axis. Each network time snapshot corresponds to a user intent switching association network subgraph during the duration of an operation intent segment, thus obtaining a network time snapshot sequence.

[0146] The user intent switching association network is serialized using network snapshots. Using the boundary timestamp Tb of each intent switching boundary in the intent switching boundary position set as the snapshot segmentation point, the user intent switching association network is divided into multiple network time snapshots along the time axis. Each network time snapshot corresponds to a subgraph of the user intent switching association network during the duration of an operation intent segment, resulting in a network time snapshot sequence. Each boundary timestamp in the intent switching boundary position set is obtained and denoted as T1, T2, ..., Tk. For each time interval [Ti, Ti+1] (where T0 is the start time of the operation sequence and Tk+1 is the end time of the operation sequence), the nodes and edges existing within this interval are extracted to form a network time snapshot Gi. All snapshots are arranged in chronological order to form a network time snapshot sequence.

[0147] Step S820: Perform network structure difference quantization processing on adjacent network time snapshots in the network time snapshot sequence, calculate the number of newly added nodes, the number of missing nodes, the number of newly added connections, and the number of missing connections between adjacent network time snapshots, and generate a network structure difference vector sequence. Each network structure difference vector in the network structure difference vector sequence is used to characterize the structural evolution magnitude of the user's intention to switch the associated network within the corresponding time interval.

[0148] For two adjacent snapshots Gi and Gj in the network time snapshot sequence, compare the node set in Gj with the node set in Gi. Count the number of nodes present in Gj but not in Gi as the number of newly added nodes, and count the number of nodes present in Gi but not in Gj as the number of disappeared nodes. Perform the same operation on the edge set, counting the number of newly added and disappeared edges. Combine these four quantities into a four-dimensional vector, denoted as Vi=(Na,Nr,Ea,Er), as the network structure difference vector for the i-th time interval. Arrange all difference vectors in chronological order to form the network structure difference vector sequence.

[0149] Step S830: Based on the network structure difference vector sequence, perform network evolution rhythm analysis and processing, extract the magnitude fluctuation curve of the network structure difference vector in the network structure difference vector sequence, identify the network time snapshot position corresponding to the local maximum of the magnitude in the magnitude fluctuation curve, mark the network time snapshot position as the evolution acceleration point of the user intention to switch the associated network, and obtain the set of network evolution acceleration points.

[0150] Calculate the magnitude M = sqrt(Na^2 + Nr^2 + Ea^2 + Er^2) of each vector Vi in the network structure difference vector sequence to obtain the magnitude sequence. Find local maxima points in the magnitude sequence, i.e., the index positions that satisfy M > Mi-1 and M > Mi+1. Mark the network time snapshots corresponding to these positions as evolution acceleration points, forming a set of network evolution acceleration points.

[0151] Step S840: Perform acceleration cause tracing processing on each evolution acceleration point in the set of network evolution acceleration points, extract the newly added nodes and newly added connection edges in the network time snapshot corresponding to the evolution acceleration point, determine the combination of network evolution acceleration cause features based on the operation event density within the intention segment of the operation intention segment corresponding to the newly added node and the hesitation transition type identifier of the operation hesitation interval corresponding to the newly added connection edge, and generate an evolution acceleration cause feature data set.

[0152] For each evolution acceleration point, obtain the set of newly added nodes and the set of newly added edges from the corresponding network time snapshot. For each node in the set of newly added nodes, extract the operation event density Dd within its corresponding operation intent segment. For each edge in the set of newly added edges, extract the hesitation transition type identifier Lt of its corresponding operation hesitation interval. Combine the above features into a trigger feature vector and store it in the evolution acceleration trigger feature dataset.

[0153] Step S850: Based on the evolution acceleration trigger feature data set, perform evolution stage division processing on the user intent switching association network, merge adjacent network evolution acceleration points with similar trigger feature combinations into the same evolution stage, and generate network evolution stage division results. The network evolution stage division results include the start and end timestamps of each evolution stage and the stage evolution trigger features.

[0154] The DBSCAN clustering algorithm is used to cluster feature vectors in the evolutionary acceleration causal feature dataset, grouping feature points with a distance less than a preset threshold into the same cluster. On the timeline, acceleration points belonging to the same cluster and with adjacent times are merged into one evolutionary stage. For each evolutionary stage, the earliest timestamp of the acceleration points within that stage is recorded as the stage start time, and the latest timestamp as the stage end time. The average of the causal features of all acceleration points within that stage is used as the stage's evolutionary causal feature.

[0155] Step S860: Perform stage intent pattern extraction processing on the network time snapshot subsequence within each evolution stage, statistically analyze the distribution of node attributes and connection edge attributes in the network time snapshot within that evolution stage, extract the central trend of node attribute distribution and the connection pattern features of connection edge attribute distribution, and obtain stage intent behavior pattern description data.

[0156] For each evolutionary stage, the node sets from all network time snapshots within that stage are aggregated, and the average value of node attributes (operation event density Dd and duration Td) is calculated as the central trend of node attribute distribution. Simultaneously, the proportions of increasing and decreasing connections among all connections within that stage are statistically analyzed as connection pattern features. These features are then combined into stage-specific intentional behavior pattern description data.

[0157] Step S870: Combine the network evolution stage division results and the stage intention behavior pattern description data in chronological order to generate a user intention evolution stage sequence, and attach the user intention evolution stage sequence to the user behavior pattern mining results to generate user behavior pattern mining results carrying intention evolution stage information.

[0158] The network evolution stage segmentation results generated in step S850 (including the start and end timestamps of each evolution stage and the characteristics of the evolutionary causes of each stage) are combined with the stage intention behavior pattern description data generated in step S860 in chronological order to form a continuous evolution stage sequence. This evolution stage sequence is then added as an additional information layer to the user behavior pattern mining results.

[0159] Step S910: Perform network community discovery processing on the user intent switching associated network. Based on the connection edge weight attributes between nodes and the normalized operation event density similarity of node attributes, divide the node set into communities. Nodes that are closely connected and have similar operation event densities are grouped into the same network community to obtain the network community division result. The network community division result includes multiple network communities and a subset of nodes contained in each network community.

[0160] A community detection algorithm based on modularity optimization (such as the Louvain algorithm) is used to partition the user intent switching association network into communities. The algorithm's input is the adjacency matrix of a weighted directed graph and the feature vector of each node (composed of normalized operation event density and duration slices). The algorithm maximizes the modularity value Qm=(1 / 2m)*Σ[Aij-(ki*kj) / (2m)]*δ(Ci,Cj) by iteratively moving nodes and merging communities, where m is the number of edges, Aij is the adjacency matrix element, ki is the node degree, and Ci is the community to which the node belongs. After iteration, the algorithm outputs the community partitioning results, where each community contains a group of nodes that are closely connected and have similar operation event densities.

[0161] Step S920: Perform community intent semantic aggregation processing on the node subset of each network community, extract the node intent encoding vector of all nodes in the network community, calculate the community semantic center vector of the node intent encoding vector, and use the intent level corresponding to the community semantic center vector in the intent hierarchy space as the community intent level identifier of the network community to obtain network community intent semantic annotation data.

[0162] For each network community defined in step S910, extract the node intent encoding vectors of all nodes within that community. Calculate the arithmetic mean of these vectors to obtain the community semantic center vector Cm = (1 / N)*sum(Vi), where Vi is the node intent encoding vector of the i-th node, and N is the number of nodes in the community. Input the community semantic center vector Cm into the intent depth parsing model in step S140. The model calculates the mounting coordinates of this vector in the intent hierarchy space and determines its intent level. This level identifier is used as the community intent level identifier for that community.

[0163] Step S930: Based on the semantic annotation data of the network community intent, perform inter-community migration pattern mining processing on the user intent switching association network, extract the migration path sequence of the operation intent segment in the intent drift trajectory chain that migrates between different network communities, count the occurrence frequency of each migration path sequence, extract the frequent migration path with an occurrence frequency exceeding a preset frequency threshold, and obtain a set of frequent inter-community migration patterns.

[0164] Traverse the intent drift trajectory chain generated in step S157, recording the event of each operation intent segment migrating from one network community to another. For each migration, record the source community identifier and the destination community identifier to form a migration path. Count the frequency of all migration paths, and mark the paths with a frequency exceeding a preset frequency threshold as frequent migration paths, forming a set of frequent migration patterns between communities.

[0165] Step S940: Perform migration cost analysis on each frequent migration path in the set of frequent migration patterns between communities, extract the connection edges between adjacent network communities on the frequent migration path, obtain the duration of hesitation transition and the hesitation transition type identifier of the connection edge, calculate the community migration cost based on the duration of hesitation transition and the hesitation transition type identifier, and obtain the set of frequent migration patterns between communities carrying migration costs.

[0166] For each frequent migration path in the set of frequent migration patterns between communities, migration cost analysis is performed. The connection edges between adjacent network communities on the frequent migration path are extracted, and the duration of hesitant transition (Th) and the hesitant transition type identifier (Lt) of the connection edge are obtained. Based on the duration of hesitant transition and the hesitant transition type identifier, the community migration cost Cm = Th * (1 + B * Dlt) is calculated, where B is an adjustment coefficient, and Dlt is 1 when the hesitant transition type is increasing and -1 when it is decreasing. The calculated cost is appended to the attributes of the frequent migration path to obtain the set of frequent migration patterns between communities carrying migration costs.

[0167] Step S950: Construct a user community migration behavior profile based on the set of frequent migration patterns between communities carrying migration costs. The user community migration behavior profile includes the user's preferred migration path, preferred migration direction, and migration cost distribution of each preferred migration path between different network communities.

[0168] The set of frequent inter-community migration patterns with migration costs generated in step S940 is summarized and analyzed. For each source community, the target communities of all its outgoing edges and their corresponding migration costs are counted, and the target community with the lowest migration cost is marked as the preferred migration direction of the source community. All preferred migration paths and their migration cost distributions are organized into a structured profile data.

[0169] Step S960: Map the node subsets of each network community in the network community division result back to the time axis of the operation behavior sequence to generate the community affiliation timeline of the user operation behavior. The community affiliation timeline is used to represent the active network community identifiers corresponding to each time interval in the operation behavior sequence.

[0170] For each time point in the sequence of user actions, the node correspondence in the associated network is switched according to the user's intent to determine the network community identifier to which that time point belongs. These community identifiers are arranged in chronological order to form a timeline of community affiliation for the user's actions.

[0171] Step S970: Integrate the user community migration behavior profile and the community affiliation timeline of the user operation behavior with the user behavior pattern mining results to generate user behavior pattern mining results carrying community migration behavior characteristics.

[0172] The user community migration behavior profile generated in step S950 and the community affiliation timeline generated in step S960 are used as additional output information and fused and packaged with the user behavior pattern mining results to generate the final output results.

[0173] Step S1010: For each operation hesitation interval in the operation hesitation interval distribution set, calculate the change in operation density before and after hesitation, extract the operation event density within the intention segment of the preceding operation intention segment and the operation event density within the intention segment of the following operation intention segment, calculate the density change difference between the operation event density within the intention segment of the following operation intention segment and the operation event density within the intention segment of the preceding operation intention segment, and obtain the set of density change amounts of the hesitation interval.

[0174] Traverse each hesitation interval in the set of hesitation interval distributions, and locate its preceding and following operation intent segments based on timestamps. Read the operation event density values ​​R1 and R2 within these two segments from the intent segment feature data set. Calculate the density change difference DR = R2 - R1 to obtain the set of hesitant interval density changes.

[0175] Step S1020: Perform density change direction determination processing on the set of density change amounts in the hesitation interval. Divide the operation hesitation interval into positive density change hesitation interval and negative density change hesitation interval according to the positive or negative sign of the density change difference, and generate a set of density change direction labels for the hesitation interval.

[0176] For each density change difference DR calculated in step S1010, if DR is greater than 0, the hesitation interval is marked as a positive density change hesitation interval, indicating that the operational density increases after hesitation; if DR is less than 0, it is marked as a negative density change hesitation interval, indicating that the operational density decreases after hesitation; if DR is equal to 0, it is marked as a no-change hesitation interval. The classification results are associated with the hesitation interval labels to form a set of hesitant interval density change direction labels.

[0177] Step S1030: Extract the duration of hesitation for each hesitation interval in the set of hesitation interval distributions, normalize the duration of hesitation and the magnitude of density change difference, and combine the normalized duration of hesitation and the normalized magnitude of density change difference with the corresponding density change direction label in the set of density change direction labels of hesitation intervals to construct a standardized feature vector of hesitation intervals.

[0178] For each operational hesitation interval, extract its hesitation duration Th and the absolute value of the density change difference |DR|. Calculate the normalized value of Th in the entire set of operational hesitation interval distributions: Thn = (Th - min(Th)) / (max(Th) - min(Th)) and the normalized value of |DR| in the entire set: DRn = (|DR| - min(|DR|)) / (max(|DR|) - min(|DR|)). Combine the normalized hesitation duration, the normalized density change difference magnitude, and the density change direction label as three components into a three-dimensional feature vector, i.e., the standardized feature vector of the hesitation interval.

[0179] Step S1040: Perform cluster analysis on the standardized feature vectors of all operational hesitation intervals, and cluster the standardized feature vectors of hesitation intervals into multiple hesitation pattern clusters in the feature space. Each hesitation pattern cluster corresponds to a hesitation behavior pattern. Extract the cluster center vector of each hesitation pattern cluster as the pattern prototype vector of the hesitation behavior pattern. The dimension of the pattern prototype vector corresponds to the dimension of the standardized feature vector of the hesitation interval.

[0180] The K-means clustering algorithm is used to cluster all the standardized feature vectors generated in step S1030. The number of clusters K is determined according to the elbow rule. The algorithm iteratively updates the cluster centers and assigns data points, ultimately outputting K clusters. The cluster center vector of each cluster is the pattern prototype vector of that hesitant behavior pattern.

[0181] Step S1050: Perform hesitation mode classification processing on the hesitation interval corresponding to each connection edge in the user intent switching association network, calculate the similarity between the standardized feature vector of the hesitation interval corresponding to the operation hesitation interval and the prototype vector of each pattern in the hesitation behavior pattern prototype set, select the hesitation behavior pattern corresponding to the pattern prototype vector with the highest similarity as the hesitation mode label of the connection edge, and obtain the connection edge hesitation mode label set.

[0182] For each connection edge in the user intent switching network, the corresponding operational hesitation interval is found from the operational hesitation interval distribution set, and its standardized feature vector is extracted. The Euclidean distance between this vector and each pattern prototype vector is calculated, and the pattern label corresponding to the pattern prototype vector with the smallest distance is selected as the hesitation pattern label of this connection edge. The labels of all connection edges are summarized to form a set of connection edge hesitation pattern labels.

[0183] Step S1060: Based on the set of hesitant mode labels of the connection edges, perform hesitant mode network reconstruction processing on the user intent switching association network, merge connection edges with the same hesitant mode label into the same hesitant mode sub-network, generate multiple hesitant mode sub-networks, each hesitant mode sub-network consists of connection edges with the same hesitant mode label and nodes associated with the connection edges.

[0184] Based on the label information in the set of hesitation pattern labels for connected edges, the edges in the network are grouped by label. For each label, all connected edges corresponding to that label and the nodes involved by these edges are extracted to form a sub-network. This ultimately results in multiple hesitation pattern sub-networks, each corresponding to a specific hesitation behavior pattern.

[0185] Step S1070: Extract the intention transfer features of each hesitant pattern subnetwork, and statistically analyze the intention hierarchy distribution of nodes and the transition direction distribution of connecting edges in the hesitant pattern subnetwork to generate hesitant pattern subnetwork intention transfer feature data.

[0186] For each hesitant pattern subnetwork, the hierarchical distribution of its nodes in the intent hierarchy distribution graph is statistically analyzed, and the proportion of nodes at each level is calculated. Simultaneously, the direction vectors of all connecting edges in the subnetwork are statistically analyzed, and the average direction of the transition directions is calculated. These statistical features are combined to form the intent transition feature data for the hesitant pattern subnetwork.

[0187] Step S1080: Combine the intention transfer feature data of all hesitant pattern subnetworks with the corresponding pattern prototype vectors in the set of hesitant behavior patterns to generate hesitant pattern behavior parsing results, and attach the hesitant pattern behavior parsing results to the user behavior pattern mining results.

[0188] The feature data of all hesitant pattern subnetworks generated in step S1070 are associated with the pattern prototype vectors generated in step S1040 to form the hesitant pattern behavior analysis result. This result is then appended to the user behavior pattern mining result as the final output.

[0189] Based on the same inventive concept, please refer to Figure 3 The diagram shows a schematic block diagram of the structure of a server 100 provided in an embodiment of this application, which may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0190] In this embodiment, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the computer-based user behavior data mining method provided in the aforementioned method embodiments.

[0191] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A computer-based user behavior data mining method, characterized in that, The method includes: Acquire a set of user behavior record data, which contains a sequence of operation behaviors marked with operation timestamps, and each sequence of operation behaviors consists of continuously collected multi-dimensional operation events; The user behavior record data set is processed to identify the boundary of the operation intent, and the set of intent switching boundary positions and the set of operation hesitation interval distribution corresponding to the operation behavior sequence are obtained. The set of intent switching boundary positions is used to define the segmentation position of adjacent operation intent segments in the operation behavior sequence, and the set of operation hesitation interval distribution is used to characterize the duration distribution of each operation hesitation event in the operation behavior sequence. A user intent switching association network is constructed based on the set of intent switching boundary locations and the set of operation hesitation interval distributions. The user intent switching association network uses operation intent segments as node units and hesitation transition relationships between operation intent segments as connecting edges. The attributes of node units include the duration of intent segments and the density of operation events within intent segments. The attributes of connecting edges include the duration of hesitation transitions and the hesitation transition type identifier. The preset intent depth parsing model is invoked to perform intent hierarchy expansion processing on the user intent switching association network, and the operation intent segments are mapped to the intent hierarchy space to generate a user intent hierarchy distribution map. The user intent hierarchy distribution map includes the expansion path of each intent level in the intent hierarchy space and the mounting coordinates of each operation intent segment in the intent hierarchy space. The intent drift trajectory chain corresponding to the operation behavior sequence is extracted based on the user intent hierarchy distribution map. The intent drift trajectory chain is composed of an intent hierarchy identifier sequence arranged in the order of operation timestamps. User behavior pattern mining results are generated based on the intent drift trajectory chain.

2. The computer-based user behavior data mining method according to claim 1, characterized in that, The step of performing operation intent boundary recognition processing on the user behavior record data set to obtain the set of intent switching boundary positions and the set of operation hesitation interval distributions corresponding to the operation behavior sequence includes: Extract the timestamps of all operation events for each operation behavior sequence in the user behavior record data set, group the operation events according to the operation event type, and obtain an operation event time sequence group set oriented by operation event type. The operation event time sequence group set contains a chain of operation event instances arranged in ascending order of timestamps corresponding to each operation event type. The time interval between adjacent operation events is calculated for each chain of operation event instances in the operation event time sequence grouping set to generate an operation event interval sequence oriented towards operation event type. Each element of the operation event interval sequence oriented towards operation event type represents the time difference between two adjacent operation events under the same operation event type. Based on the operation event interval sequence corresponding to all operation event types, cross-type interval mutation point detection processing is performed to identify the time position where the interval sequence of different operation event types simultaneously shows a sudden increase in interval. The time position is marked as a candidate intention switching boundary to obtain a set of candidate intention switching boundaries. Each candidate intention switching boundary in the set of candidate intention switching boundaries contains a corresponding timestamp position and a combination of operation event types that triggered the interval increase. The set of candidate intent switching boundaries is subjected to intent switching coherence verification processing. The distribution of operation event types in the adjacent time windows before and after each candidate intent switching boundary is extracted. The transition probability of the operation event type distribution is calculated. The candidate intent switching boundaries with a transition probability lower than a preset coherence judgment threshold are determined as formal intent switching boundaries, and the set of intent switching boundary positions is obtained. Each intent switching boundary in the set of intent switching boundary positions includes the boundary timestamp position and the operation event type transition features on both sides of the boundary. Extract pairs of operation events from the user behavior record data set whose time interval between adjacent operation events exceeds a preset hesitation judgment threshold, mark the time interval between the operation event pairs as operation hesitation intervals, and collect the duration data and start timestamp data of all operation hesitation intervals to form an operation hesitation interval distribution set. The operation hesitation interval distribution set includes the start timestamp and hesitation duration of each operation hesitation interval. Based on the set of intent switching boundary positions, each operation behavior sequence is segmented into intent segments. The operation event segment between two adjacent intent switching boundaries is taken as an operation intent segment. The duration of the intent segment and the density of operation events within the intent segment are calculated for each operation intent segment to obtain the intent segment feature data set. Each operational hesitation interval in the operational hesitation interval distribution set is processed by hesitation type classification. The direction of change of operational event density of adjacent operational intent segments before and after the operational hesitation interval is extracted. Based on the direction of change of operational event density, the operational hesitation interval is divided into increasing hesitation interval and decreasing hesitation interval. The division result of increasing hesitation interval or decreasing hesitation interval is used as hesitation transition type identifier, and an operational hesitation interval distribution set containing hesitation transition type identifier is obtained.

3. The computer-based user behavior data mining method according to claim 1, characterized in that, The construction of the user intent switching association network based on the intent switching boundary location set and the operation hesitation interval distribution set includes: Each operation intent segment defined by the intent switching boundary position set is used as a node construction unit. The duration of the intent segment and the density of operation events within the intent segment are extracted as node attribute parameters to generate a node set of the user intent switching association network. Each node in the node set corresponds to an operation intent segment and carries the node attribute parameters. Based on the timestamp position of each operation hesitation interval in the operation hesitation interval distribution set, determine the preceding operation intention segment and the following operation intention segment connected to each operation hesitation interval, establish a directed connection edge between the node corresponding to the preceding operation intention segment and the node corresponding to the following operation intention segment, and obtain a set of connection edges. The direction of the directed connection edge is from the preceding operation intention segment to the following operation intention segment. Extract the duration of hesitation for each operation hesitation interval as the duration of hesitation transition for the corresponding directed connection edge, and extract the hesitation transition type identifier carried by each operation hesitation interval as the hesitation transition type identifier attribute for the corresponding directed connection edge. Then, attach the duration of hesitation transition and the hesitation transition type identifier attribute to the corresponding connection edge of the connection edge set. Normalize the operation event density within the intent segment of each node in the node set, map the operation event density within the intent segment to a preset density scale range to obtain the normalized operation event density, and update the normalized operation event density to the node's operation density normalization attribute. Based on the hesitation transition type identifier of each connection edge in the connection edge set, the connection edges are divided into increasing connection edges and decreasing connection edges. The increasing connection edges are assigned positive transition weights, and the decreasing connection edges are assigned negative transition weights, thus obtaining the connection edge weight attribute set. The duration of the intent segment of each node in the node set is aligned with time granularity. The duration of the intent segment is divided into multiple time slice units according to a preset time slice step size. A duration slice representation of the node is generated based on the number of time slice units. The duration slice representation is then appended to the attributes of the corresponding node. The set of nodes carrying node attribute parameters, operation density normalization attribute, and duration slice representation is combined with the set of connection edges carrying hesitation transition duration attribute, hesitation transition type identifier attribute, and connection edge weight attribute set to generate the user intent switching association network.

4. The computer-based user behavior data mining method according to claim 1, characterized in that, The process of calling a preset intent depth parsing model to perform intent hierarchy expansion processing on the user intent switching association network, mapping the operation intent segments to the intent hierarchy space, and generating a user intent hierarchy distribution map includes: For each node in the user intent switching association network, neighborhood feature aggregation processing is performed to extract the previous neighbor node set and the subsequent neighbor node set of each node. Based on the operation density normalization attribute of the previous neighbor node set and the operation density normalization attribute of the subsequent neighbor node set, the neighborhood operation density difference vector of the node is calculated. The neighborhood operation density difference vector is used to characterize the gradient of the operation density change of the node in the association network. The normalized operation event density, duration slice representation, and neighborhood operation density difference vector of each node in the node set of the user intent switching association network are input into the intent encoding layer of the intent depth parsing model. The hierarchical encoder of the intent encoding layer performs intent feature encoding processing on each node to obtain the node intent encoding vector. The node intent encoding vector contains the local operation density features and neighborhood density change features of the node. The weight attributes of each connection edge in the set of connection edges of the user intent switching association network are aggregated by weight propagation path. The weight attributes of the connection edges are accumulated layer by layer along the direction of the directed connection edges of the user intent switching association network to generate the cumulative weight propagation value of each node. The cumulative weight propagation value is used as the intent level depth indicator parameter of the node. Based on the intent level depth indicator parameter, hierarchical clustering is performed on the node intent encoding vectors of all nodes. Nodes whose intent level depth indicator parameters fall into the same level interval are classified into the same intent level to obtain the intent level division result. The intent level division result includes the level identifier of each intent level and the node subset belonging to each intent level. For each intent level, a subset of nodes is subjected to hierarchical center anchoring processing. The mean of the node intent encoding vectors of all nodes in the subset is extracted as the hierarchical cluster center position of that intent level, thus obtaining the set of cluster center positions of each intent level in the intent hierarchy space. For each node corresponding to an operation intent segment, perform hierarchical attribution mapping processing. Calculate the distance between the node intent encoding vector of the node and the hierarchical cluster center of each intent level. Select the intent level with the smallest distance as the attribution level of the node. Mount the coordinates of the node in the intent hierarchy space to the hierarchical expansion path corresponding to the attribution level to obtain the mounted coordinates of each operation intent segment in the intent hierarchy space. The user intent hierarchy distribution map is generated by combining the cluster center locations of each intent level in the intent hierarchy space, the hierarchical expansion path of each intent level, and the mounting coordinates of each operation intent segment in the intent hierarchy space.

5. The computer-based user behavior data mining method according to claim 1, characterized in that, The step of extracting the intent drift trajectory chain corresponding to the operation behavior sequence based on the user intent hierarchy distribution map, wherein the intent drift trajectory chain is composed of an intent hierarchy identifier sequence arranged in order of operation timestamp, and generating user behavior pattern mining results based on the intent drift trajectory chain, includes: Based on the mounting coordinates of each operation intent segment in the intent hierarchy space in the user intent hierarchy distribution diagram, the intent hierarchy identifier corresponding to each operation intent segment is determined, and the operation intent segments are arranged in chronological order in the operation behavior sequence to generate an original intent hierarchy identifier sequence composed of intent hierarchy identifiers. The original intent hierarchy identifier sequence is subjected to hierarchy transition point detection processing. The sequence position where adjacent intent hierarchy identifiers change in the original intent hierarchy identifier sequence is identified as a hierarchy transition point. The intent hierarchy identifier before the transition and the intent hierarchy identifier after the transition corresponding to each hierarchy transition point are extracted to obtain a set of hierarchy transition points. Based on the set of hierarchical transition points, extract the node attributes of the operation intent segment corresponding to each hierarchical transition point in the user intent switching association network, obtain the duration of the operation intent segment before the transition and the duration of the operation intent segment after the transition, and calculate the duration ratio parameter of the duration of the intent segment before and after the transition. Based on the pre-jump intention level identifier and post-jump intention level identifier corresponding to each level jump point in the set of level jump points, the jump direction vector of the level jump is determined in the intention level space of the user intention level distribution map. The jump direction vector points from the level cluster center position of the pre-jump intention level to the level cluster center position of the post-jump intention level. Extract the hesitation transition type identifier and hesitation transition duration of the connection edges of the pre-transition operation intent segment and post-transition operation intent segment corresponding to the connection level transition point in the user intent switching association network. Use the hesitation transition type identifier as the transition transition type of the level transition and the hesitation transition duration as the transition transition duration of the level transition. The duration ratio parameter, transition direction vector, transition type and transition duration are combined into hierarchical transition description data. The hierarchical transition description data of all hierarchical transition points are concatenated in chronological order to generate an intention drift trajectory chain. The intention drift trajectory chain contains the transition path information of the operation behavior sequence in the intention hierarchy space. The intent drift trajectory chain is subjected to pattern induction processing. Repeated transition path segments are extracted from the intent drift trajectory chain and marked as user behavior pattern units. All user behavior pattern units are collected to form the user behavior pattern mining results.

6. The computer-based user behavior data mining method according to claim 1, characterized in that, The method further includes: Acquire an auxiliary behavior environment data set, which includes operation interface layout status data and operation trigger area distribution data during the operation behavior sequence collection period; The auxiliary behavior environment data set is subjected to interface operation hot zone analysis and processing to generate an operation interface hot zone distribution map. The operation interface hot zone distribution map includes the operation frequency density distribution of each operation trigger area on the operation interface and the spatial adjacency relationship between the operation trigger areas. Based on the hotspot distribution map of the operation interface and the set of intent switching boundary positions, spatial intent association analysis is performed to map each intent switching boundary to the corresponding spatial position of the operation interface, generating an intent switching spatial distribution map. The intent switching spatial distribution map includes the spatial coordinates of each intent switching boundary on the operation interface and the operation trigger area identifier corresponding to the spatial coordinates. The spatial distribution mapping of intent switching and the user intent switching association network are subjected to spatial constraint fusion processing. Spatial operation trigger area identifiers are added to the node attributes of the user intent switching association network, and spatial jump distance parameters are added to the connection edge attributes to obtain a user intent switching association network with spatial constraints. Based on the user intent switching association network carrying spatial constraints, the user intent hierarchy distribution map is spatially weighted and corrected. The mounting coordinates of the nodes in the intent hierarchy space are adjusted according to the spatial operation trigger area identifier of the nodes and the spatial jump distance parameter of the connecting edges to obtain the spatially corrected user intent hierarchy distribution map. Based on the spatially corrected user intent hierarchy distribution map, the spatially constrained intent drift trajectory chain corresponding to the operation behavior sequence is extracted. Based on the spatially constrained intent drift trajectory chain, the user behavior pattern mining result integrating spatial factors is generated.

7. The computer-based user behavior data mining method according to claim 1, characterized in that, The method further includes: Obtain a historical user behavior record data set, which includes a sequence of historical operation behaviors of multiple users and a user identity identifier associated with each historical operation behavior sequence; Perform operation intent boundary recognition processing on the historical user behavior record data set to obtain the set of historical intent switching boundary positions and the set of historical operation hesitation interval distributions corresponding to the historical operation behavior sequence of each user; Based on the set of historical intent switching boundary positions and the set of historical operation hesitation interval distribution, a historical user intent switching association network is constructed for each user. Common node structures and common connection edge structures are extracted from all historical user intent switching association networks to construct a group intent switching benchmark network. The user intent switching association network is subjected to deviation measurement processing based on the group intent switching benchmark network. The attribute deviation vector between the node attributes of the user intent switching association network and the node attributes of the group intent switching benchmark network is calculated, as well as the structural deviation parameter between the connection edge structure of the user intent switching association network and the connection edge structure of the group intent switching benchmark network. The attribute deviation vector and the structural deviation parameter are input into a preset user intent deviation classification model. The user intent deviation classification model classifies users into preset intent deviation categories based on the attribute deviation vector and the structural deviation parameter, and obtains a user intent deviation category identifier. The user intent deviation category identifier is attached to the user behavior pattern mining result to generate a user behavior pattern mining result carrying the intent deviation category identifier.

8. The computer-based user behavior data mining method according to claim 6, characterized in that, The user intent hierarchy distribution map is spatially weighted and corrected based on the user intent switching association network carrying spatial constraints. This involves adjusting the node's mounting coordinates in the intent hierarchy space according to the node's spatial operation trigger region identifier and the spatial jump distance parameter of the connecting edges, resulting in a spatially corrected user intent hierarchy distribution map. This includes: Extract the spatial operation trigger area identifier of each node in the user intent switching association network carrying spatial constraints, obtain the regional center coordinates and operation frequency density of the operation trigger area corresponding to each spatial operation trigger area identifier based on the operation interface hot zone distribution map, and construct a spatial trigger area coordinate mapping table. Based on the spatial triggering area coordinate mapping table, each node is mapped to the two-dimensional coordinate space of the operation interface to obtain the operation interface mapping coordinates of each node. The Euclidean distance between the operation interface mapping coordinates of the node pairs with connecting edge relationships in the user intent switching association network is calculated to generate a set of spatial distances between node pairs. Extract the spatial jump distance parameter of each connection edge in the user intent switching association network carrying spatial constraints, compare the spatial jump distance parameter with the Euclidean distance of the corresponding node pair in the node pair spatial distance set, generate a spatial jump distance correction coefficient based on the comparison result, and update the spatial jump distance parameter of each connection edge using the spatial jump distance correction coefficient. Spatial offset calculation is performed on the mounting coordinates of each operation intent segment in the intent hierarchy space of the user intent hierarchy distribution map. The operation interface mapping coordinates of the corresponding node of the operation intent segment are extracted. The operation interface mapping coordinates are projected onto the intent hierarchy space according to a preset spatial scaling ratio to generate a spatial offset vector. The spatial bias vector is superimposed on the mounting coordinates of the corresponding operation intention segment in the intention hierarchy space to generate preliminary spatial correction mounting coordinates. Based on the spatial jump distance parameter between node pairs, the relative position constraint adjustment is performed on the preliminary spatial correction mounting coordinates of operation intention segment pairs with connecting edge relationships to ensure that the distance between operation intention segment pairs in the intention hierarchy space is consistent with the spatial jump distance parameter, thus obtaining the spatial correction mounting coordinates. Replace the initial spatial correction mount coordinates of all operation intent segments with spatial correction mount coordinates, update the hierarchical cluster center positions of each intent level in the intent hierarchy space and the hierarchical expansion paths of each intent level, and generate a spatial correction user intent hierarchy distribution map.

9. A server, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the computer-based user behavior data mining method of any one of claims 1 to 8 by executing the machine-executable instructions.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions, the server's processor reads the machine-executable instructions from the machine-readable storage medium, and the processor executes the machine-executable instructions, causing the computer device to perform the computer-based user behavior data mining method as described in any one of claims 1 to 8.

Images