Behavior data credibility evaluation system based on time sequence line feature fusion

Abstract

The invention relates to the technical field of behavior data credibility evaluation, in particular to a behavior data credibility evaluation system based on time sequence row feature fusion, which is characterized in that an input capture unit captures year and month sub-node input time sequence data of starting and ending time of a project filled by a user in real time; the interrupt monitoring unit detects a system pop-up window and a user active pause as external interrupt, and records an interrupt duration and a continuous filling starting point, and the feature fusion unit performs dynamic time sequence window segmentation and cross-modal alignment; a time sequence behavior feature vector containing a year priority mark, a month lagging time consumption ratio, a cross-year modification time consumption effect difference value and a continued time consumption deviation rate is generated, a credibility evaluation unit judges a behavior mode by means of a four-dimensional decision matrix, a bidirectional threshold comparison and scoring algorithm is detected through a logic fault, a filling behavior authenticity probability value is output, and the filling behavior authenticity probability value is calculated. And the output feedback unit feeds back the probability value to the evaluation system according to sectional risk mapping, so that risk domain differentiation warning is realized, and the credibility evaluation precision is improved.

Description

Technical Field

[0001] This invention relates to the field of behavioral data credibility assessment technology, and more specifically, to a behavioral data credibility assessment system based on temporal line feature fusion. Background Technology

[0002] Behavioral data credibility assessment is an important technology, specifically applied in scenarios such as assessment systems and data collection platforms. It is used to determine the authenticity of behavioral data with temporal logic, such as the start and end times of user-filled items, by analyzing the characteristics of the filling behavior. The core is to identify fictitious filling behavior and ensure the validity of assessment or collection data. Existing behavioral data credibility assessment schemes fail to deeply integrate user input temporal features with interruption context, making it impossible to accurately distinguish subtle behavioral differences between genuine recall and fictitious completion. This results in low assessment accuracy and a high misjudgment rate. In genuine recall scenarios, users typically determine the year sub-node first and then refine the month sub-node. Subsequent completion after an interruption will naturally experience time-consuming deviations due to memory recovery. When modifying across years, users may hesitate to make decisions due to sorting out the time logic. In contrast, fictitious completion often presents a logical contradiction: prioritizing month completion, no reasonable deviation in the time consumption of subsequent completion, and short hesitation when modifying across years. However, existing schemes only analyze single behaviors in isolation, failing to capture the temporal relationship between year and month and the impact of interruptions on subsequent completion. This either misjudges the reasonable interruption time consumption of genuine recall as fictitious or misses the logical gaps in fictitious completion, failing to meet the assessment system's requirement for accurate verification of the authenticity of behavioral data. To solve this technical problem, we provide a behavioral data credibility assessment system based on temporal feature fusion. Summary of the Invention

[0003] The purpose of this invention is to provide a behavioral data credibility assessment system based on temporal feature fusion to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a behavioral data credibility assessment system based on temporal line feature fusion is provided, including: The input capture unit is used to capture the input time sequence data of the year and month sub-nodes filled in by the user at the start and end times of the project in real time; The interruption monitoring unit is used to detect external interruption events during the input process and record the interruption duration and the start point of the resume input. The external interruption events include system pop-ups and user-initiated pauses. The feature fusion unit is used to fuse the input time series data with the interruption event to generate a time series behavior feature vector. The time series behavior feature vector includes year priority mark, month lag time consumption ratio, cross-year modification time efficiency difference value and continuation filling time deviation rate. The credibility assessment unit is used to execute a credibility scoring algorithm based on the temporal behavior feature vector. The credibility scoring algorithm achieves logical fault detection by comparing preset thresholds of real recall patterns and fictitious patterns, and outputs the authenticity probability value of the filling behavior. The real recall patterns and fictitious patterns are determined by the temporal behavior feature vector. The output feedback unit is used to feed back the authenticity probability value to the evaluation system in real time for risk warning.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention captures the year and month sub-nodes of the user's input time sequence data in real time through an input capture unit, detects external interruption events and records the interruption duration and the start point of continuation, and generates a time sequence behavior feature vector that integrates the interruption context through dynamic time sequence window segmentation and cross-modal alignment. The credibility assessment unit outputs the authenticity probability value of the filling behavior based on a four-dimensional decision matrix and logical fault detection. This achieves the effect of accurately integrating the input time sequence features and the interruption context and effectively distinguishing between real recollection and fictitious filling. It effectively solves the problem of low accuracy and high misjudgment rate of behavior data credibility assessment caused by the lack of deep integration of the above features in existing solutions. Attached Figure Description

[0006] Figure 1 This is an overall block diagram of the present invention.

[0007] The meanings of the labels in the diagram are as follows: 1. Input capture unit; 2. Interruption monitoring unit; 3. Feature fusion unit; 4. Credibility evaluation unit; 5. Output feedback unit. Detailed Implementation

[0008] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0009] This invention provides a behavioral data credibility assessment system based on temporal line feature fusion. Please refer to [link / reference]. Figure 1 As shown, it includes: Input capture unit 1 is used to capture the input time sequence data of the year and month sub-nodes filled in by the user at the start and end times of the project in real time; Interruption monitoring unit 2 is used to detect external interruption events during the input process and record the interruption duration and the start point of the continuation. External interruption events include system pop-ups and user-initiated pauses. The feature fusion unit 3 is used to fuse the input time series data with the interruption event to generate a time series behavior feature vector. The time series behavior feature vector includes year priority mark, month lag time consumption ratio, cross-year modification time efficiency difference value and continuation filling time deviation rate. The credibility assessment unit 4 is used to execute a credibility scoring algorithm based on the temporal behavior feature vector. The credibility scoring algorithm achieves logical fault detection by comparing the preset thresholds of the real recall mode and the fiction mode, and outputs the authenticity probability value of the filling behavior. The real recall mode and the fiction mode are determined by the temporal behavior feature vector. The output feedback unit 5 is used to feed back the authenticity probability value to the evaluation system in real time for risk warning.

[0010] Feature fusion unit 3 performs dynamic temporal window segmentation, which cuts the input temporal data into independent temporal segments based on the starting point of the interruption duration. Each segment contains a complete sequence of sub-node operations for a single continuous filling. Cross-modal alignment is performed on each segment. Specifically, the input actions of the year sub-node and the input actions of the month sub-node are bound into logical event groups through recursive matching of timestamps. At the same time, the interruption event is transformed into a context association marker between segments. Finally, the spatiotemporal correlation between the last year sub-node value before the interruption and the month sub-node at the start point of the continuation is extracted in the cross-segment scenario through the memory anchor backtracking mechanism, generating a temporal behavior feature vector that integrates the interruption context information.

[0011] The generation of year-priority tags employs a two-stage verification mechanism, specifically including: The first stage compares the action sequence of event groups within a time segment. When the timestamp of the first input event of any year sub-node in the start and end time of the same project is earlier than that of the associated month sub-node, a primary marker is generated. The second stage verifies cross-segment association. When the first action filled after the interruption is a month sub-node and forms a logical event group with the last input year sub-node before the interruption, an anchor break detector is triggered. If no new year input event is detected before the input of the month sub-node, the primary marker is upgraded to a reinforced year-priority marker; otherwise, it is downgraded to a weakened marker and a fictitious mode determination is triggered.

[0012] The calculation of the monthly lag time ratio incorporates dynamic benchmark calibration, specifically including: First, the net input time of the year sub-nodes within the logical event group is extracted. This time needs to exclude the time occupied by modification events. For the associated month sub-nodes, a cognitive load compensation algorithm is used. The baseline value is adjusted based on whether there are cross-year modification events before the input of the associated month sub-node. When there are cross-year modifications, the input time of the month after the modification is completed and refocused is used as the baseline. The final output ratio is fused with the interruption context flag. If the month input occurs in the post-interruption filling stage, the continuation filling time deviation rate is used for weight compensation.

[0013] The time difference in modification across years is quantified by modifying the source tree, specifically including: When a month sub-node modification event is detected, the year value corresponding to the original input of that month in the matching logic event group is traced back. If the modified month value causes a change in the year, the spatiotemporal correlation detector is activated. The detector performs three operations, including recording the total time spent on the modification operation, stripping the mechanical time spent on the user interface operation, and extracting the decision hesitation time in the modification confirmation stage. Finally, the decision hesitation time is compared with the baseline hesitation time of the month modification scenario in the same year, and a cross-year correlation time gain value is generated as the efficiency difference value output.

[0014] The deviation rate of the continued filling time is located using the interruption memory anchor point positioning, specifically including: The last complete input month sub-node before the interruption is used as the baseline anchor point, and its original input time is recorded. After the continuation is started, the input action of the first month sub-node is monitored. If the action belongs to the same logical event group as the baseline anchor point, the memory coherence analyzer is triggered. The analyzer extracts the cognitive recovery features in the entire cycle of the continuation month input, including the initial input delay time, the frequency of deletion and modification during the input process, and the hovering time before final confirmation. Finally, the above features are fused into a composite time value. The ratio of the composite time value to the original time of the baseline anchor point is standardized by dynamic baseline calibration and output as the continuation time deviation rate.

[0015] The distinction between real and fictional memory patterns is based on a four-dimensional decision matrix, which specifically includes: The first dimension receives the year-priority label strength value; the second dimension loads the month lag time ratio calibration value; the third dimension associates the cross-year modification time efficiency difference value; and the fourth dimension integrates the continuation filling time deviation rate. A pattern mapping engine is used, which projects the four-dimensional feature vectors to a preset behavior pattern space. When the four-dimensional features simultaneously meet the real memory feature interval, a strong realism mode is activated. If any dimension falls into the fictional feature interval, a pattern conflict detection is triggered. For conflict scenarios, logical discontinuity detection is used for arbitration.

[0016] Logical tomography detection implements bidirectional threshold comparison, specifically including: The forward comparison chain is based on a four-dimensional decision matrix. When at least two dimensions deviate from the real memory feature range at the same time, the fictional pattern marker is activated. The reverse verification chain uses a behavior consistency checker. This checker calls the spatiotemporal distribution characteristics of all modification events within the dynamic time window and calculates the fluctuation variance of the modification time efficiency difference. When the variance value is lower than the fictional pattern threshold and is negatively correlated with the cross-year modification time efficiency difference, the existence of a logical fault is confirmed. Finally, the arbitration result is output through the fault weighted decision module.

[0017] The credibility scoring algorithm includes a tomography-weighted decision module, which receives the logical tomography detection results and performs three-dimensional weighted fusion, specifically including: The first weight dimension is based on the year-priority labeling strength. The second weight dimension is associated with the cognitive recovery feature density in the continuation filling time deviation rate. The third weight dimension is based on the modified source tree depth. Through a dynamic confidence allocator, the above weights are convolved with the pattern matching degree of the four-dimensional decision matrix to generate an initial confidence score. Finally, the variance value of the behavior consistency checker is introduced as a decay factor to perform nonlinear correction on the initial score and output the authenticity probability value.

[0018] The output of the authenticity probability value adopts a segmented risk mapping mechanism, specifically including: The authenticity probability value is input into the risk spectrum mapping engine, which divides the system into three decision domains. The high-confidence domain directly outputs the authenticity probability value to the evaluation system interface element display module. The medium-risk domain activates the risk warning module, generates a floating prompt box with the authenticity probability value and associates it with suspicious fields. The high-risk domain triggers a deep verification chain, calls the modification event sequence and modification tracing tree in the time sequence segment for visual reconstruction, generates a spatiotemporal behavior heatmap and embeds it into the evaluation report appendix, and sends the authenticity probability value and associated evidence chain data package to the back-end audit system.

[0019] It needs further explanation that the specific implementation of the two-stage verification method of dynamic temporal window segmentation and year-priority marking in the feature fusion unit is as follows: After the input capture unit 1 captures the input temporal data of the year and month sub-nodes of the start and end times of the user's input items in real time, and the interruption monitoring unit 2 detects and records the interruption duration and the start point of the continuation of the input, the feature fusion unit 3 needs to integrate the scattered input data with the interruption context through dynamic temporal window segmentation to generate a temporal behavior feature vector that reflects the user's input behavior logic. If the unsegmented data is directly fused, the behavior fragmentation caused by the interruption may easily obscure the true input pattern. Therefore, it is necessary to first complete the accurate division and alignment of the temporal segments. The specific implementation method is as follows: The dynamic temporal window segmentation operation performed by feature fusion unit 3 is based on the logical segmentation of the filling behavior using interruption events as boundaries. First, the system divides the input temporal data acquired by input capture unit 1 into multiple independent temporal segments, using the interruption duration start point recorded by interruption monitoring unit 2 as the segmentation boundary. Each segment corresponds to a continuous filling process, including the complete sequence of sub-node operations completed by the user during that process, ensuring the logical coherence of the behavior within each segment. Then, cross-modal alignment is performed on each independent temporal segment. Here, cross-modal refers to two different types of input actions: year sub-nodes and month sub-nodes. The purpose of alignment is to establish the logical relationship between the two, specifically through recursive timestamp matching. Starting with the first input action of a segment, all actions are recursively traversed in chronological order of timestamps. Upon detecting a year-related input action, the system continues to track subsequent adjacent month-related input actions. If the timestamp interval between two actions is less than a preset association threshold, they are bound together as a logical event group. If no subsequent month-related input action or the interval exceeds the threshold, the year-related action is temporarily stored until a matching month-related action is detected, ensuring that each year-related child node can find a corresponding month-related child node, forming a complete year-month logical association. Simultaneously, interrupt events are converted into inter-segment context association markers, with each interrupt event corresponding to two adjacent time segments. The system adds association markers to the pre-interruption and post-interruption segments, which include interruption duration and interruption type information for subsequent cross-segment behavior analysis. Finally, the spatiotemporal correlation between the segments is extracted through a memory anchor backtracking mechanism. The memory anchor backtracking mechanism refers to using the interruption event as an anchor point to backtrack the last complete year sub-node value before the interruption, then locating the continuation starting point of segment B after the interruption, analyzing the correlation between the month sub-node of the continuation starting point and the year sub-node before the interruption. If a correlation exists, the correlation information is incorporated into the features, ultimately generating a temporal behavior feature vector that includes dimensions such as year priority markers and month lag time ratio, and integrates the interruption context.

[0020] After the initial construction of the temporal behavior feature vector is completed, a year-priority tag needs to be generated first. This tag is used to determine whether the user follows the true recall logic of determining the year first and then refining the month when filling in the information. In true recall, people usually remember the year of the event first and then recall the specific month. When filling in information fictitiously, the order may be arbitrarily reversed. The generation of the tag adopts a two-stage verification mechanism to ensure the accuracy of the tag. The specific implementation method is as follows: The first stage is the comparison of the action sequence within a time segment, focusing on the behavioral logic in a single continuous filling process. The system traverses all logical event groups within each independent time segment, extracts the first input event timestamp of the year sub-node and the first input event timestamp of the associated month sub-node in each event group, and performs a sequence comparison. If, in the filling of the start and end time of the same project, the first input timestamp of the year sub-node of any logical event group is earlier than the associated month sub-node, a primary marker is generated for that event group. If the month sub-node timestamp is earlier than the year sub-node, a primary marker is not generated temporarily and will be further verified in the second stage. The second stage is cross-segment association verification, which focuses on analyzing whether the interrupted subsequent filling behavior conforms to the logic of memory continuity. In real recall, interrupted subsequent filling usually continues the memory of the previous year and does not require re-entering the year. However, when filling in fictional data, the year may be forgotten and re-entered. When the interruption monitoring unit 2 detects that the first action of the interrupted subsequent filling is the month sub-node, and that the month sub-node and the last entered year sub-node before the interruption can form a logical event group, the system automatically triggers the anchor point breakage detector. The detector will backtrack all input events before the June action in the continuation phase to determine if there are any new year input events. If no new year input is detected, the primary marker generated in the first phase will be upgraded to a reinforced year-priority marker, indicating that the filling behavior is more in line with the real memory pattern. If a new year input is detected, it means that the user may have forgotten the previous content and there is suspicion of fabrication. The primary marker will be downgraded to a weakened marker, and the fabrication pattern judgment process will be triggered simultaneously to provide evidence of abnormal behavior for subsequent credibility assessment.

[0021] The specific implementation of the dynamic benchmark calibration of the monthly lag time ratio and the quantification of the time difference in cross-year modification time is as follows: After completing the two-stage verification of the year-priority marking, the feature fusion unit 3 needs to further calculate the monthly lag time ratio. This ratio is used to measure the degree of delay of the user's filling in the month sub-node relative to the year sub-node. An excessively long delay may reflect hesitation in thinking when filling in the information. However, simply calculating the time ratio is easily affected by modification operations, changes in cognitive load, and interruptions. Therefore, dynamic benchmark calibration needs to be introduced to eliminate interference and ensure that the ratio can truly reflect the filling behavior logic. The specific implementation is as follows: The calculation of the monthly lag time ratio is based on logical event groups as the basic unit. First, the net input time of the year sub-nodes within each logical event group is extracted. Net input time refers to the total time from when the user starts inputting the year sub-node to when the user first confirms the year. The time occupied by modification events during this period needs to be excluded. Then, the cognitive load compensation algorithm is used to adjust the benchmark value of the associated month sub-nodes. The cognitive load compensation algorithm dynamically adjusts the monthly time benchmark based on the user's cognitive state before entering the month. The core logic is that the higher the cognitive load, the more relaxed the monthly input time benchmark should be. In specific implementation, it is first determined whether there are cross-year modification events before the month sub-node is input. If there are no cross-year modifications, it means that the user's cognitive load is low, and the original input time of the month sub-node is used as the benchmark value. If there are cross-year modifications, the user needs to re-sort out the logical relationship between the year and the month, and the cognitive load increases. At this time, the monthly input time before the modification is not used. Instead, the time spent by the user refocusing on the monthly input after the cross-year modification is completed is used as the benchmark value to ensure that the benchmark value can adapt to the changes in cognitive load.

[0022] Finally, the monthly lag time ratio is calculated (monthly sub-node baseline time ÷ yearly sub-node net input time), and weight compensation is performed by integrating the interruption context flag. The system first checks whether the input of the monthly sub-node occurred in the interruption follow-up filling stage. It judges the start point of the follow-up filling recorded by the interruption monitoring unit (2). If it did not occur in the follow-up filling stage, the calculated ratio is directly output. If it occurred in the follow-up filling stage, the user may experience memory connection problems due to the interruption. It is necessary to enable the follow-up filling time deviation rate for weight compensation. The follow-up filling time deviation rate is the deviation ratio between the monthly input time in the follow-up filling stage and the baseline time before the interruption. The deviation rate is added to the original ratio according to a preset weight, and the final output is the monthly lag time ratio of the fusion interruption impact. This ensures that the ratio can comprehensively reflect the filling delay characteristics under different scenarios. After obtaining the monthly lag time ratio, it is also necessary to quantify the time difference value of cross-year modification. This value is used to capture the degree of decision-making hesitation when users modify the month, resulting in a change of year. In real recollection, cross-year modifications usually involve adjustments to the event time logic, and the hesitation time is usually longer. Fictional filling may lead to abnormal hesitation due to logical confusion. By modifying the source tree, the modification process can be accurately traced to achieve the scientific quantification of the time difference value. The specific implementation method is as follows: The calculation of the time difference in cross-year modification begins with the monitoring of month sub-node modification events. Input capture unit 1 records the user's modification operations on month sub-nodes in real time. When such modification events are detected, the system automatically calls the modification source tree. The modification source tree is a data structure that records the correlation between all modification operations. Each node corresponds to one modification action, the parent node is the original input, and the child node is the modified content. Through this tree, the original input year value of the currently modified month sub-node in the logical event group can be back-matched. If the modified month value causes a year change, the spatiotemporal correlation detector is immediately activated. This detector is specifically used to analyze the temporal logical correlation and decision consumption in cross-year modifications. The detector performs three core operations: The first step is to record the total time spent on the modification operation, that is, the complete time from when the user clicks the month input box to when the modification is finally confirmed. The second step is to remove the mechanical time spent on the user interface operation, which refers to the pure interface interaction time, and remove it from the total time, retaining the effective time that includes the thinking process. The third step is to extract the decision hesitation time in the modification confirmation stage, that is, the hovering time from when the user enters the new month value to when they click the confirm button. This time directly reflects the degree of hesitation of the user in judging the cross-year time logic. Finally, the extracted decision hesitation time is compared with the baseline hesitation time of the modification scenario in the same year. The baseline hesitation time is obtained by statistical analysis of a large amount of real data. It is the average decision hesitation time of the modification month in the same year. The difference between the two is calculated. This difference is the cross-year correlation time gain value, which is output as the cross-year modification time efficiency difference value. If the cross-year hesitation time is significantly higher than the baseline value, it means that the user needs to frequently sort out the time logic when modifying, which is more consistent with the characteristics of real memory. If the difference is too small, it may be that there is arbitrary modification when fictitious filling, providing a key basis for subsequent credibility assessment.

[0023] The specific implementation method for determining the continuation filling time deviation rate, interruption memory anchor point location, and four-dimensional decision-making of real and fictitious modes: After completing the dynamic benchmark calibration of the monthly lag time ratio and the quantification of the cross-year modification time efficiency difference, feature fusion unit 3 also needs to calculate the continuation filling time deviation rate. This deviation rate is used to evaluate the continuity between the filling behavior after the interruption and the filling behavior before the interruption. If the memory is continuous, the probability of real recall is high; otherwise, there may be fictitious behavior. The core of its calculation is to establish a comparison benchmark between the behavior before and after the interruption by locating the interruption memory anchor point. The specific implementation method is as follows: The calculation of the continuation time deviation rate uses the interruption event as the core anchor point. First, the system locates the last fully entered month sub-node before the interruption as the benchmark anchor point. The fully entered month sub-node refers to the month data that the user has completed and confirmed, and is not in the modification state. The month sub-node is chosen as the anchor point because its entry relies on year memory, which better reflects the continuity of memory. The system will extract the original input time of the benchmark anchor point from the input time sequence data, that is, the total time from when the user starts entering the month sub-node to the first confirmation, and store the time as the benchmark for subsequent comparison. When the interruption monitoring unit 2 detects the user's interruption, the system will continue to monitor the month data. After a user restarts the filling process (i.e., resumes filling), the system monitors the input actions of the first month's sub-node in the resume phase in real time and determines whether this action belongs to the same logical event group as the baseline anchor point. The same logical event group refers to the start and end time filling logic of the same project. If they belong to the same logical event group, it means that the resume filling behavior is directly related to the behavior before the interruption, and the memory coherence analyzer needs to be triggered. If they do not belong to the same logical event group, it means that it is a new filling task, and the resume filling time deviation rate is not calculated at this time. The core of the memory coherence analyzer is to extract the cognitive recovery features in the entire cycle of the resume month input. These features can reflect the user's reaction after the interruption. The state of memory recovery is assessed through three aspects: first, the initial input delay duration, i.e., the time interval between the user focusing on the month input box and actually starting to input after the continuation of input is initiated; a longer delay indicates slower memory recovery. Second, the frequency of deletions and modifications during input, i.e., the number of times content is deleted or modified while inputting the month; a higher frequency indicates more significant input hesitation due to memory lapse. Third, the hovering time before final confirmation, i.e., the time the user's mouse or finger remains on the confirmation button without clicking after inputting the month; a longer hovering time indicates higher uncertainty about the content to be entered. The analyzer integrates these three types of cognitive recovery features according to preset weights into a composite consumption. For example, input the initial delay duration of 2 seconds (weight 0.3), the deletion / modification frequency of 1 time (corresponding to a time consumption of 1 second, weight 0.2), and the hover duration of 3 seconds (weight 0.5). The composite time consumption value = 2×0.3 + 1×0.2 + 3×0.5 = 2.3 seconds. Then, calculate the ratio between the composite time consumption value and the original time consumption of the benchmark anchor point (4 seconds) (2.3 ÷ 4 = 0.575). After standardization processing by the dynamic benchmark calibration mentioned above, the final output ratio is used as the continuation filling time deviation rate. The closer the deviation rate is to 1, the smaller the difference between the filling time before the interruption and the filling time before the interruption, and the stronger the memory continuity.

[0024] After feature fusion unit 3 generates a complete temporal behavioral feature vector including year priority marker, month lag time ratio, cross-year modification time efficiency difference, and continuation filling time deviation rate, credibility assessment unit 4 needs to determine whether the user's filling behavior belongs to the real memory mode or the fictitious mode based on these features. Single feature judgment is prone to misjudgment. Therefore, a four-dimensional decision matrix is ​​established to achieve multi-dimensional comprehensive judgment. The specific implementation method is as follows: The determination of real and fictional memory modes is based on a four-dimensional decision matrix. The four dimensions of the matrix correspond to four key features, ensuring comprehensiveness in the determination. The first dimension receives the year-priority label strength value, generated by the aforementioned two-stage verification mechanism. A strong year-priority label corresponds to a strength value of 1.0, a primary label to 0.7, and a weak label to 0.3. A higher value indicates a stronger real memory logic prioritizing the year over the month. The second dimension loads the month lag time ratio calibration value, i.e., the final ratio after dynamic benchmark calibration. In real memory mode, this ratio is typically stable at 0. The time taken for filling in the year and month is moderate, ranging from 0.8 to 1.5 seconds. In a fictional scenario, this range may be exceeded. The third dimension relates to the time difference for cross-year modifications, i.e., the difference between the decision-making hesitation time for cross-year modifications and the baseline time for modifications within the same year. In real recall, due to the need to organize the time logic, this difference is usually greater than 1 second. In a fictional scenario, due to logical confusion, the difference may be less than 0.5 seconds. The fourth dimension integrates the continuation filling time deviation rate, i.e., the standardized ratio of the time spent on continuation filling to the time spent before interruption. In real recall, this deviation rate is mostly in the 0.8-1.2 range (memory continuity, stable time), while in a fictional scenario, it may be less than 0.5 seconds. For scores below 0.6 or above 1.5 (where memory breaks cause sudden drops or increases in time), the system employs a pattern mapping engine to perform feature matching. This engine pre-stores the criteria for dividing real and fictional feature intervals, derived from a large number of real and fictional completion samples. It projects a four-dimensional feature vector onto a preset behavioral pattern space, where the horizontal and vertical axes represent different feature dimensions. Each dimension marks the boundary between real and fictional intervals. When all four features simultaneously satisfy the real memory feature interval, the engine immediately activates a strong realism mode, determining that the current completion behavior highly conforms to real memory logic. If any dimension falls into a fictional feature interval, pattern conflict detection is triggered, indicating a feature contradiction that requires further arbitration. For pattern conflict scenarios, the system uses logic break detection for arbitration. Logic break detection backtracks to the original behavioral data corresponding to the four-dimensional features, analyzing whether the contradictory features are caused by accidental factors. It also combines behavioral consistency checks to finally output the arbitration result. If the contradiction is determined to be caused by accidental factors, it is still classified as a real memory mode; if the contradiction is confirmed to reflect a real logical break, it is determined to be a fictional mode, providing a clear pattern determination basis for subsequent credibility scoring.

[0025] The specific implementation of the logical fault detection bidirectional threshold comparison, credibility scoring algorithm, and authenticity probability value risk mapping is as follows: When the credibility assessment unit 4 determines the behavior pattern through the four-dimensional decision matrix, if any dimension falls into the fictitious feature interval and triggers the pattern conflict detection, a bidirectional threshold comparison must be performed for arbitration through logical fault detection. This detection cross-validates from two dimensions: the degree of feature deviation and the consistency of behavior, to avoid misjudgment from a single dimension and ensure the accuracy of logical fault determination. The specific implementation is as follows: The core of logical discontinuity detection lies in the collaborative work of the forward comparison chain and the reverse verification chain. First, the forward comparison chain is initiated. This chain is entirely based on the four-dimensional decision matrix constructed earlier. Each dimension of the matrix has a preset real recall feature interval. The system checks whether the values ​​of each dimension in the four-dimensional feature vector deviate from this interval. If only a single dimension deviates, such as a deviation rate of 0.7 for filling in the blanks only (below 0.8), a logical discontinuity is not immediately determined. If at least two dimensions deviate simultaneously, such as a year priority marker strength of 0.4 (below 0.6) or a time difference of 0.6 seconds for cross-year modification (below 1.0 second), it indicates multiple contradictions in the behavioral logic. The fictitious pattern marker is immediately activated, initially indicating a tendency towards logical discontinuity. Next, the reverse verification chain is initiated, using a behavior consistency checker to verify the rationality of the forward determination. The behavior consistency checker is a module used to analyze the overall coherence of the user's filling behavior. Its core logic is that the modification behavior of real recall has reasonable fluctuations, while the modification behavior of fictitious filling often exhibits abnormal stability or chaos. This checker will call all modification events within the dynamic time sequence window. The spatiotemporal distribution characteristics of the data are analyzed, including the time consumption, modification type, and timestamp of each modification event. The variance of the modification time consumption efficiency difference is then calculated. This variance reflects the dispersion of the modification time consumption efficiency difference; a larger variance indicates more reasonable fluctuations in modification time consumption, while a smaller variance indicates abnormally stable modification time consumption. The reverse verification chain must simultaneously meet two conditions to confirm the existence of a logical discontinuity: first, the calculated variance is lower than the fictitious pattern threshold; second, the variance is negatively correlated with the cross-year modification time consumption efficiency difference. This negative correlation indicates that the smaller the variance, the smaller the cross-year modification time consumption efficiency difference. This relationship suggests that when users modify data across years, they deliberately control the time consumption to be stable while lacking genuine recall and decision-making hesitation, further confirming the authenticity of the logical discontinuity. If the forward comparison chain activates the fictitious pattern flag and the reverse verification chain meets the above two conditions, the final arbitration result confirming the existence of the logical discontinuity is output through the discontinuity weighted decision module. If the reverse verification chain does not meet the conditions, the logical discontinuity is output as invalid, overturning the initial forward judgment and ensuring that the arbitration result balances feature deviation and overall behavioral consistency.

[0026] After the logical fault detection outputs the arbitration result, the credibility assessment unit 4, through the fault weighted decision module in the credibility scoring algorithm, fuses multi-dimensional features with the fault results to generate a quantified authenticity probability value. This module needs to balance the credibility contribution of different features to avoid a single feature dominating the score. The specific implementation method is as follows: The fault-weighted decision module first receives the logical fault detection results and performs three-dimensional weighted fusion based on them. The first weight dimension is based on the intensity of the year-priority label. This intensity value directly reflects the logical strength of the true recall logic of year first and then month. The reinforced year-priority label (intensity 1.0) corresponds to the highest weight (e.g., 0.4), the primary label (intensity 0.7) corresponds to the medium weight (e.g., 0.3), and the weakened label (intensity 0.3) corresponds to the lowest weight (e.g., 0.1). The higher the label intensity, the greater the positive contribution to the credibility score, because in real recall, users tend to determine the year first and then refine the month. The second weight dimension is related to the cognitive recovery feature density in the continuation time deviation rate. The cognitive recovery feature density is the result of combining the three types of features in the continuation stage: initial input delay time, deletion and modification frequency, and hovering time. The density formula integrates various metrics, reflecting the smoothness of memory recovery after interruption. A density higher than a preset threshold (e.g., 0.8) corresponds to a weight of 0.3, indicating coherent memory and a high probability of accuracy. A density between 0.5 and 0.8 corresponds to a weight of 0.2, and a density lower than 0.5 corresponds to a weight of 0.1, indicating sluggish memory recovery and a low probability of accuracy. The third weight dimension uses the modified trace tree depth, which refers to modifying the longest path length from the original input node to the final confirmation node in the trace tree. The shallower the depth, the less modification is required, and the more decisive the filling behavior (consistent with accurate memory): Depth 1 (no modification) corresponds to a weight of 0.3, Depth 2 (1 modification) corresponds to a weight of 0.2, and Depth 3 and above (multiple modifications) corresponds to a weight of 0.1, avoiding overestimation of credibility due to frequent modifications. Subsequently, a dynamic confidence allocator adjusts the allocation ratio of the three-dimensional weights based on the logical discontinuity detection results. If there is "no logical discontinuity," the proportion of the first and second weights is increased; if a logical discontinuity exists, the proportion of the first weight is decreased and the proportion of the third weight is increased to ensure that the weights adapt to the behavioral logic state. Next, the adjusted three-dimensional weights are convolved with the pattern matching degree of the four-dimensional decision matrix. The calculation process involves multiplying each weight by the pattern matching contribution value of its corresponding feature and summing the results to generate an initial confidence score. Finally, the fluctuation variance value calculated by the behavioral consistency checker is introduced as a decay factor for nonlinear correction. If the fluctuation variance is higher than the real pattern threshold, the decay factor is set to 0.95 (slight decay); if the fluctuation variance is lower than the fictitious pattern threshold (e.g., 0.3, indicating abnormally stable behavior), the decay factor is set to 0.7 (significant decay). The initial confidence score is multiplied by the decay factor (e.g., 0.675 × 0.95 ≈ 0.64), and the final output is a probability value of authenticity in the 0-1 range. The closer this value is to 1, the higher the authenticity of the submitted behavior. After obtaining the authenticity probability value, the output feedback unit 5 needs to transform the quantified probability into risk response actions that the evaluation system can execute through a segmented risk mapping mechanism. Different risk levels correspond to different processing strategies. It is necessary to ensure the smooth flow of high-credibility behaviors while strictly verifying high-risk behaviors. The specific implementation method is as follows: First, the authenticity probability value is input into the risk spectrum mapping engine. This engine pre-divides three decision domains based on clinical assessment data and auditing experience. The high-confidence domain has an authenticity probability value greater than or equal to 0.8, corresponding to logically coherent and uncontradictory behavior, consistent with genuine recall characteristics. The medium-risk domain has an authenticity probability value of 0.5-0.8, corresponding to minor feature contradictions (such as a single dimension deviating from the true range), but no confirmed logical gaps. The high-risk domain has an authenticity probability value less than 0.5, corresponding to obvious logical gaps or multi-dimensional feature contradictions, with a high suspicion of fabrication. For the high-confidence domain, the risk spectrum mapping engine directly outputs the authenticity probability value (e.g., 0.85) to the assessment system interface element display module. This module displays the probability value in real time next to the corresponding field on the user's input page in the form of confidence level: 85%, without additional warnings, ensuring that the user's input process is not disturbed, while providing assessment personnel with an intuitive confidence reference.

[0027] For medium-risk areas, the engine immediately activates the risk warning module, generating a floating prompt box with a probability value of authenticity: the prompt box displays "The credibility of this field is moderate (e.g., 65%), with slightly suspicious behavior (e.g., large deviation in the time spent on continuing the data entry)", and through field association technology, the prompt box is anchored to the field corresponding to the suspicious feature (e.g., the month field of "Project A End Time" filled in during the continuing data entry stage), making it convenient for assessors to quickly locate suspicious points; at the same time, the prompt box provides a "View Details" button, which can be clicked to expand the input time sequence segment and feature analysis of the field (e.g., year priority marking, month lag time ratio), to assist assessors in manual judgment. For high-risk domains, the engine triggers a deep verification chain, executing multi-dimensional evidence retention and audit push: First, it calls the modification event sequence and modification source tree in the time segment obtained by the dynamic time window segmentation mentioned above, and generates a spatiotemporal behavior heatmap through visualization reconstruction technology. The heatmap uses the time axis as the horizontal axis and the filled field as the vertical axis, and uses the color depth to represent the modification frequency and time consumption, intuitively presenting the abnormal modification trajectory. This heatmap is automatically embedded in the evaluation report appendix to provide visual evidence for subsequent audits. Second, the system automatically packages the authenticity probability value and related evidence chain data package. The data package includes the original input time sequence data, interruption records, four-dimensional feature vectors, logical fault detection reports, etc., and sends them to the background audit system through an encrypted transmission protocol. After receiving the data, the audit system generates an audit work order and assigns it to a dedicated person for review to ensure that no high-risk behavior is missed and to achieve full lifecycle traceability of the evidence chain.

[0028] In this invention, the input capture unit 1 captures the input time-series data of the year and month sub-nodes of the start and end times of the user's filled-in items in real time. The interruption monitoring unit 2 detects system pop-ups and user-initiated pauses as external interruptions, records the interruption duration and the start point of the resume filling. The feature fusion unit 3 generates a time-series behavior feature vector containing year priority markers, month lag time ratio, cross-year modification time efficiency difference value, and resume filling time deviation rate through dynamic time-series window segmentation and cross-modal alignment. The credibility assessment unit 4 determines the behavior pattern using a four-dimensional decision matrix, and outputs the authenticity probability value of the filling behavior through logical fault detection, bidirectional threshold comparison and scoring algorithm. The output feedback unit 5 feeds back the probability value to the evaluation system according to the segmented risk mapping, realizes differentiated warning of risk domains, and improves the accuracy of credibility assessment.

[0029] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A behavioral data credibility assessment system based on temporal line feature fusion, characterized in that, include: The input capture unit (1) is used to capture the input time sequence data of the year sub-node and month sub-node filled in by the user at the start and end time of the project in real time; Interruption monitoring unit (2) is used to detect external interruption events during the input process and record the interruption duration and the start point of the continuation. The external interruption events include system pop-ups and user-initiated pauses. The feature fusion unit (3) is used to fuse the input time series data with the interruption event to generate a time series behavior feature vector. The time series behavior feature vector includes year priority mark, month lag time consumption ratio, cross-year modification time efficiency difference value and continuation filling time deviation rate. The credibility assessment unit (4) is used to execute a credibility scoring algorithm based on the temporal behavior feature vector. The credibility scoring algorithm achieves logical fault detection by comparing the preset thresholds of the real recall mode and the fiction mode, and outputs the authenticity probability value of the filling behavior. The real recall mode and the fiction mode are determined by the temporal behavior feature vector. The output feedback unit (5) is used to feed back the authenticity probability value to the evaluation system in real time for risk warning.

2. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 1, characterized in that: Feature fusion Unit (3) performs a dynamic temporal window segmentation operation, which cuts the input temporal data into independent temporal segments based on the starting point of the interruption duration. Each segment contains a complete sequence of sub-node operations for a single continuous filling. Cross-modal alignment is performed on each segment. Specifically, the input actions of the year sub-node and the input actions of the month sub-node are bound into a logical event group through recursive matching of timestamps. At the same time, the interruption event is converted into a context association marker between segments. Finally, the spatiotemporal association between the last year sub-node value before the interruption and the month sub-node at the start point of the continuation is extracted in the cross-segment scenario through the memory anchor backtracking mechanism, and a temporal behavior feature vector that integrates the interruption context information is generated.

3. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 2, characterized in that: The generation of the year-priority marker employs a two-stage verification mechanism, specifically including: The first stage compares the action sequence of event groups within a time segment. When the timestamp of the first input event of any year sub-node in the start and end time of the same project is earlier than that of the associated month sub-node, a primary marker is generated. The second stage verifies cross-segment association. When the first action filled after the interruption is a month sub-node and forms a logical event group with the last input year sub-node before the interruption, an anchor break detector is triggered. If no new year input event is detected before the input of the month sub-node, the primary marker is upgraded to a reinforced year-priority marker; otherwise, it is downgraded to a weakened marker and a fictitious mode determination is triggered.

4. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 3, characterized in that: The calculation of the monthly lag time ratio incorporates dynamic benchmark calibration, specifically including: First, the net input time of the year sub-nodes in the logical event group is extracted. This time needs to exclude the time occupied by modification events. A cognitive load compensation algorithm is used for the associated month sub-nodes. The baseline value is adjusted by whether there is a cross-year modification event before the input of the associated month sub-node. When there is a cross-year modification, the input time of the month after the modification is completed and refocused is used as the baseline. The final output ratio is fused with the interruption context flag. If the month input occurs in the interruption follow-up filling stage, the follow-up filling time deviation rate is used for weight compensation.

5. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 4, characterized in that: The time efficiency difference of cross-year modifications is quantified by modifying the source tree, specifically including: When a month sub-node modification event is detected, the year value corresponding to the original input of that month in the logical event group is backtracked and matched. If the modified month value causes a change in the year, the spatiotemporal correlation detector is activated. The detector performs three steps, including recording the total time spent on the modification operation, stripping the mechanical time spent on the user interface operation, and extracting the decision hesitation time in the modification confirmation stage. Finally, the decision hesitation time is compared with the baseline hesitation time of the month modification scenario in the same year, and a cross-year correlation time gain value is generated as the efficiency difference value output.

6. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 5, characterized in that: The deviation rate of the continued filling time is located using the interruption memory anchor point, specifically including: The last complete input month sub-node before the location interruption is used as the reference anchor point, and its original input time is recorded. After the continuation is started, the input action of the first month sub-node is monitored. If the action belongs to the same logical event group as the reference anchor point, the memory coherence analyzer is triggered. The analyzer extracts the cognitive recovery features in the entire cycle of the continuation month input, including the initial input delay time, the frequency of deletion and modification during the input process, and the hovering time before final confirmation. Finally, the above features are fused into a composite time value. The ratio of the composite time value to the original time of the reference anchor point is standardized by the dynamic reference calibration and output as the continuation time deviation rate.

7. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 6, characterized in that: The determination of the real memory mode and the fiction mode is based on a four-dimensional decision matrix, specifically including: The first dimension receives the year-priority label strength value; the second dimension loads the month lag time ratio calibration value; the third dimension associates the cross-year modification time efficiency difference value; and the fourth dimension integrates the continuation filling time deviation rate. A pattern mapping engine is used, which projects the four-dimensional feature vectors to a preset behavior pattern space. When the four-dimensional features simultaneously meet the real memory feature interval, a strong realism mode is activated. If any dimension falls into the fictional feature interval, a pattern conflict detection is triggered. For conflict scenarios, logical discontinuity detection is used for arbitration.

8. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 7, characterized in that: The logical tomography detection implements bidirectional threshold comparison, specifically including: The forward comparison chain, based on the four-dimensional decision matrix, activates a fictitious pattern marker when at least two dimensions simultaneously deviate from the real memory feature range. The reverse verification chain employs a behavioral consistency checker, which calls upon the spatiotemporal distribution characteristics of all modification events within the dynamic time window to calculate the fluctuation variance of the modification time efficiency difference. When the variance value is lower than the fictitious pattern threshold and negatively correlated with the cross-year modification time efficiency difference, the existence of a logical fault is confirmed. Finally, the arbitration result is output through the fault weighted decision module.

9. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 8, characterized in that: The credibility scoring algorithm includes a tomography-weighted decision module, which receives the logical tomography detection results and performs three-dimensional weighted fusion, specifically including: The first weight dimension is based on the year-priority labeling strength. The second weight dimension is associated with the cognitive recovery feature density in the continuation filling time deviation rate. The third weight dimension is based on the modified source tree depth. Through a dynamic confidence allocator, the above weights are convolved with the pattern matching degree of the four-dimensional decision matrix to generate an initial confidence score. Finally, the variance value of the behavior consistency checker is introduced as a decay factor to perform nonlinear correction on the initial score and output the authenticity probability value.

10. The behavioral data credibility assessment system based on temporal line feature fusion according to claim 9, characterized in that: The output of the authenticity probability value adopts a segmented risk mapping mechanism, specifically including: The authenticity probability value is input into the risk spectrum mapping engine, which divides the system into three decision domains. The high-confidence domain directly outputs the authenticity probability value to the evaluation system interface element display module. The medium-risk domain activates the risk warning module, generates a floating prompt box with the authenticity probability value and associates it with suspicious fields. The high-risk domain triggers a deep verification chain, calls the modification event sequence and modification tracing tree in the time segment for visual reconstruction, generates a spatiotemporal behavior heatmap and embeds it into the evaluation report appendix, and sends the authenticity probability value and associated evidence chain data package to the back-end audit system.

Images