Employee performance evaluation method based on behavior data driving

Abstract

The invention discloses an employee performance evaluation method based on behavior data driving, and relates to the technical field of human resource management, and the method comprises the following steps: constructing a multi-channel network fluctuation prediction matrix, generating an expected behavior arrival time scale based on the transmission state of a task switching node, and taking the expected behavior arrival time scale as a baseline for behavior sequence gap recognition; and under the guidance of an expected behavior arrival time scale, performing sliding window scanning on the received behavior sequence, generating a gap identifier when detecting that a time slot exceeds a threshold value, and injecting a confidence coefficient weight descending label. According to the method, a behavior gap is accurately identified through network fluctuation prediction and sliding window scanning, behavior reconstruction is realized in combination with self-supervised generation and residual reasoning, and dynamic evaluation correction is completed by using confidence projection and an intelligent regulation engine, so that the data authenticity is guaranteed from the source, and the objectivity and robustness of performance evaluation and the resource allocation precision are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human resource management, in particular to an employee performance evaluation method based on behavior data driving. BACKGROUND

[0002] The "employee performance evaluation based on behavior data driving" refers to that, in the daily work process of employees, instead of relying on traditional superior scoring, subjective evaluation or stage examination, the behavior data of employees in task execution, cooperation and communication, system operation, learning and growth and other aspects are collected and analyzed in real time, the objective data are comprehensively processed by using data modeling and intelligent algorithm, the key indicators reflecting the work efficiency, contribution, innovation and collaboration of employees are extracted, and the indicators are converted into performance evaluation results. The core is to replace human subjective judgment with data driving, to realize the objectivity, dynamics and refinement of performance evaluation, so as to provide a scientific basis for organization management and improve the accuracy of human resource allocation and decision-making.

[0003] The prior art has the following disadvantages: In the prior art, the employee performance evaluation based on behavior data driving often relies on complete capture and modeling of behavior sequences in the dynamic task switching process. However, in actual application scenarios, due to fluctuations in the network environment, part of the key behavior sequences may be broken, resulting in an indiscernible gap in the continuous data chain. When the evaluation model faces such a gap, it lacks an effective dynamic completion mechanism, and usually takes the forced interpolation way to maintain the sequence continuity. Although such interpolation can repair the data structure in form, it inevitably introduces false fragments that do not conform to the real behavior. Once these false fragments enter the performance evaluation system, they will not only cause serious deviation of the single evaluation result, but also damage the objectivity and fairness of the evaluation result in the long-term accumulation process, ultimately leading to the inability of the prior art to guarantee the accuracy and credibility of performance evaluation.

[0004] The above information disclosed in the background section is only intended to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide an employee performance evaluation method based on behavior data driving to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: an employee performance evaluation method based on behavior data driving, comprising the following steps: A multi-path network fluctuation prediction matrix is constructed, and an expected behavior arrival time scale is generated based on the transmission state of the task switching node as a baseline for behavior sequence gap identification; Under the guidance of the expected behavior arrival time mark, the received behavior sequence is scanned by a sliding window, a gap identification is generated when a time gap exceeding a threshold is detected, and a confidence reduction weight label is injected; Under the constraint of the confidence reduction weight label, adaptive stretching is performed on the timeline, and the sliding window boundary is adjusted according to the gap identification, so that feature extraction is concentrated in the section with high confidence; In the sliding window after adaptive stretching, a self-supervised generation network is introduced combined with the gap identification, the micro-behavior in the gap is reconstructed using a bidirectional residual inference mechanism, and a reconstruction confidence map is generated; Based on the reconstruction confidence map, the reconstructed micro-behavior is mapped back to the complete behavior sequence using a recursive confidence projection mechanism, the performance indicators are incrementally recalculated, and the evaluation drift is captured by difference from the previous results; After identifying the evaluation drift, a multi-agent collaborative regulation engine is started, the task incentive weight is adaptively routed based on the recursive confidence projection result, the re-evaluation is performed on the high-risk nodes, and the reward distribution is frozen in the abnormal section, forming an automatic correction closed-loop regulation mechanism.

[0007] Preferably, a multi-path network fluctuation prediction matrix is constructed, and the expected behavior arrival time mark is generated based on the transmission state of the task switching node as follows: A multi-path set for behavior data collection is established, each path is divided according to different network paths, and the transmission state data of the task switching node is recorded, including user end trigger timestamp, server reception timestamp, page loading completion time, submission response time, interface interaction delay and feedback time difference; Based on the collected data, the minimum delay, maximum delay, average response time and delay fluctuation value of each path at each task node are calculated, and a path delay behavior trajectory graph is drawn to extract a representative node window to generate a prediction stability index; The node-level stability index of multiple paths is fused to generate a prediction matrix, and the path delay interval is aligned combined with the task switching trigger time to construct a behavior expected arrival time window; According to the path stability index, confidence weighted fusion is performed to generate the expected behavior arrival time mark, which is used as the timing baseline for behavior gap identification.

[0008] Preferably, when a time gap exceeding a threshold is detected, a gap identification is generated, and a confidence reduction weight label is injected as follows: Based on the expected behavior arrival time mark, a sliding time window is constructed, and behavior records matching the task node, time, and communication path within the window are extracted to construct a time ordering vector; Calculate the time interval between adjacent behavior records, and compare it with the dynamic time gap threshold calculated according to the task node importance, expected response time, historical response cycle and channel stability. If the time interval exceeds the threshold, insert a gap identifier; Embed confidence reduction tags containing data source confidence level, timestamp integrity evaluation, semantic consistency score, channel stability conversion value and behavior continuity score in the behavior records before and after the gap; Slide the window back by a dynamic step, and repeat the above process with the next expected behavior arrival timestamp as the center to complete the full sequence coverage.

[0009] Preferably, adaptive stretching is performed on the timeline, and the sliding window boundary is adjusted according to the gap identifier, so that feature extraction is concentrated in the high-confidence section as follows: Based on the confidence reduction tag, a confidence continuity distribution graph is constructed, and a confidence structure is extracted for each behavior event and mapped to a unified time axis, marking high, medium and low confidence sections; According to the confidence continuity distribution graph, adjust the sliding time window boundary position, only keep the behavior event section with confidence higher than the preset threshold, and eliminate the time segment with sharp confidence fluctuation; Classify and weight the behavior events in the adjusted window according to confidence, identify strong, ordinary and weak signal events, mark the behavior with context semantic drift and lower the confidence score; According to the signal strength, the timeline is adaptively stretched, the strong signal section time proportion is expanded, the weak signal section time span is compressed, and a nonlinear time axis is formed for subsequent feature extraction and evaluation modeling.

[0010] Preferably, a self-supervised generation network is introduced combined with the gap identifier, and a bidirectional residual reasoning mechanism is used to reconstruct the micro-behavior in the gap, and the reconstructed confidence mapping steps are as follows: Locate the gap identifier position, and extract not less than three behavior events before and after the gap with confidence not less than 70%, and construct a context semantic template; Under the guidance of the context semantic template, generate multiple candidate behavior sequences through the contrastive self-supervised generation network, and generate preliminary confidence scores according to time sequence integrity, behavior logic consistency and response path continuity; Based on the bidirectional residual reasoning mechanism, forward and backward residual calculation is performed on each candidate behavior sequence, outputting behavior content deviation vector, time offset vector and logic transfer probability, and filtering sequences with residual score not higher than the threshold; Insert the behavior sequence that passes the filtering into the gap position, and generate a reconstructed confidence mapping combined with the residual score and context similarity, and construct a reconstructed confidence mapping table containing behavior number, behavior name, predicted time, residual value and confidence score.

[0011] Preferably, in the reconstruction confidence mapping table, only when the reconstruction confidence score of the candidate behavior sequence is higher than 85%, the behavior sequence is included in the behavior chain participation performance indicator calculation, the behavior sequence with a score between 70% and 85% is marked as a backup behavior, only used for behavior trajectory recording, and the behavior sequence lower than 70% is eliminated.

[0012] Preferably, the reconstructed micro-behavior is mapped back to the complete behavior sequence by using the recursive confidence projection mechanism, the performance indicators are incrementally recalculated, and the evaluation drift step is captured by difference as follows: Based on the reconstruction confidence mapping, the reconstructed micro-behavior segment is inserted into the gap position of the original behavior sequence, and the behavior time axis is recalibrated; Using the recursive confidence projection mechanism, the influence range of the added behavior on the behavior chain is identified, and the confidence change state of the affected behavior is marked; Based on the updated behavior sequence, the incremental recalculation of the full-cycle performance indicators of the task section involving the added behavior is performed, and the difference matrix is generated by comparing the original indicators to identify the evaluation drift level.

[0013] Preferably, re-evaluation is performed on high-risk nodes, and reward distribution is frozen in abnormal sections to form an automatic correction closed-loop control mechanism as follows: After identifying the evaluation drift, extract the task nodes with an evaluation offset value exceeding 10%, and construct a drift response start table based on the recursive confidence projection results to determine the control task; Based on the drift response start table, the responsibility strength of the behavior responsible person in the control task is calculated, and the mapping relationship between the responsible person and the behavior is established; Based on the responsibility strength and behavior confidence, the adaptive routing adjustment of the incentive weight is performed, the incentive proportion of high-confidence behavior is increased, and the incentive proportion of low-confidence behavior is decreased; Perform re-evaluation on the task section with high responsibility strength and large reconstruction behavior fluctuation, and identify the behavior with inconsistent semantic structure or time jump anomaly as the behavior that needs secondary correction; Perform incentive freezing operation on the section where the behavior that needs secondary correction is located, and integrate the adjustment results with the re-evaluation information to form a task-level closed-loop control feedback package for subsequent task cycle reference.

[0014] In the above technical solutions, the technical effects and advantages provided by the present application are: The application can identify behavior gaps with high precision and mark them with confidence labels by constructing a multi-channel network fluctuation prediction matrix, generating expected behavior arrival time scales in advance, and cooperating with a sliding window dynamic scanning mechanism, guiding subsequent feature extraction to focus only on high-confidence behavior segments, and suppressing the interference of untrustworthy behavior on the evaluation model from the source. On this basis, a contrastive self-supervised generation network and a bidirectional residual reasoning mechanism are introduced, which can restore the true semantics of the behavior in the gap without interpolation, and realize high-fidelity completion of the original behavior sequence through recursive confidence projection mechanism. Finally, combined with the multi-agent regulation engine, the performance evaluation drift caused by the reconstructed behavior can be intelligently perceived and closed-loop corrected, comprehensively improving the objectivity, dynamics and stability of performance evaluation. In summary, the application significantly enhances the robustness of the performance evaluation system to abnormal behavior data, the adaptability to dynamic changes, and the precise allocation ability of incentive resources while maintaining data authenticity. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief introductions will be given to the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0016] Figure 1 The method flowchart of the employee performance evaluation method based on behavior data driving of the present application. DETAILED DESCRIPTION

[0017] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects of the example implementations to those skilled in the art.

[0018] The present application provides an employee performance evaluation method based on behavior data driving as shown in Figure 1 The method flowchart of the employee performance evaluation method based on behavior data driving of the present application. A multi-channel network fluctuation prediction matrix is constructed, and expected behavior arrival time scales are generated based on the instantaneous transmission state of the task switching node, which provides a reference baseline for subsequent behavior sequence gap identification; To solve the problem of behavior data delay or loss caused by network fluctuation in the process of dynamic task switching, a multi-channel network fluctuation prediction matrix is constructed, and expected behavior arrival time scales are generated based on the instantaneous transmission state of the task switching node, which provides a stable time sequence reference for subsequent identification of behavior sequence gaps. The specific implementation is as follows: In a real business scenario, a set of independent path sets for behavior data collection is established for the task switching behavior that may occur during the execution of tasks by employees. Each path is divided according to different physical or logical network paths and establishes a strict one-to-one mapping relationship with the operation behavior of the employee at a specific task switching node. At the moment of operation behavior, the trigger timestamp of the behavior is recorded, which is derived from the local high-precision clock of the user end and synchronized with the behavior reception timestamp recorded by the server end. To improve sampling consistency, the time granularity is accurately set to milliseconds, and the behavior context information is captured, including the loading completion time of the click source page, the submission response time of the previous task, the interface interaction response delay, the feedback time difference after input submission, and other dimensions. Each dimension is independently stored as a single data field and numbered in chronological order under a unified timeline. In addition, considering the network topology location of the employee's terminal device, such as whether it is accessed through the intranet, VPN, public Wi-Fi or mobile cellular network, the network characteristics under different paths are classified and labeled in detail. The captured instantaneous transmission state data is packaged and stored, and is archived according to the task switching node for subsequent path stability modeling.

[0019] Based on the collected data, the transmission state samples of each path at all task switching nodes are processed for stability modeling. First, all timestamp data in each path are aggregated by task node, and the minimum delay, maximum delay, average response time and delay fluctuation value at the same node are calculated in turn. Among them, the delay fluctuation value is calculated using standard deviation, and the delay characteristics of each path at each task node are expressed in a four-tuple manner. Then, the delay characteristic vectors of all paths at all task nodes are mapped to a two-dimensional delay space, with time as the horizontal axis and delay fluctuation intensity as the vertical axis, to construct a delay behavior trajectory graph for each path. In the trajectory graph, different colors are used to represent different levels of delay stability, such as green for low fluctuation stable channels, yellow for medium fluctuation channels, and red for high fluctuation and easy mutation channels. On each channel trajectory graph, representative task node windows are extracted, and based on the slope of the delay trend line within the window, the node-level prediction stability index of the channel is calculated. Finally, all channel trajectory graphs are projected onto a unified timeline, and by fusing the stability indexes of multiple paths at the same task node, a multi-path prediction matrix is constructed, which can reflect the expected delay distribution interval and stability level of each node under multiple paths.

[0020] After obtaining the multi-path prediction matrix, the delay features of each path are horizontally aligned at the task node granularity, based on the trigger time of the actual task switching behavior. During alignment, the behavior trigger time is used as the origin time point, and the minimum transmission delay of each channel at the corresponding node is used as the lower bound, and the maximum transmission delay as the upper bound, to construct the expected arrival time interval for the behavior. This interval no longer represents the expected occurrence time of the behavior as a single time point, but establishes a continuous time window to encompass the reasonable arrival period of the behavior. To improve the actual representativeness of this time interval, a behavior delay replay mechanism from historical task behavior sequences is further introduced. For historical task switching nodes with similar context structures, their behavior response delays are extracted and aligned using time normalization to generate the time period in which the behavior is most likely to occur in the current context. After obtaining the expected time window for the behavior, a behavior semantic comparison mechanism is used to compare data from different paths but with consistent behavior instructions and different response paths. Semantic duplicate behaviors are filtered out, retaining the behavior records with the highest semantic confidence as valid samples, thereby avoiding the mistaken inclusion of repeatedly reported behavior events in the behavior sequence and causing data redundancy.

[0021] Based on the expected time windows of behaviors across multiple pathways, a confidence-weighted fusion is performed according to their stability indices to form the final behavior arrival timescale. During the weighting process, the stability index is used as the basis for weight allocation, assigning higher weights to channels with lower historical volatility and higher task response consistency, while assigning lower weights to channels with frequent fluctuations and lower prediction confidence. A single behavior prediction timescale is obtained by linearly weighting and fusing the center points of the arrival time windows of all pathways. In practical applications, this timescale will serve as a reference baseline for behavior gap identification. If a behavior is not received within a reasonable offset range before or after this prediction time point, it can be determined that the behavior has been missing.

[0022] Guided by the expected behavior arrival time marker, a dynamic sliding window scan is performed on the actual received behavior sequence. When the time gap between adjacent behaviors is detected to exceed the threshold, a gap marker is generated, and a confidence reduction label is injected at the gap marker position to introduce confidence constraints while maintaining the continuity of the behavior sequence. To accurately identify gaps in behavioral data, after obtaining the expected arrival time of each task node, a dynamic sliding window mechanism needs to be introduced to perform continuity verification and time-gap scanning on the actual received behavioral data sequence, and to label and control abnormal behavior segments using a confidence weight mechanism. The specific implementation steps are as follows: Based on the obtained behavior arrival timestamps, the first expected timestamp in the current task node is selected as the scanning starting point to construct an initial sliding time window. This window establishes a fixed-length time range around the timestamp, for example, extending forward by 500 milliseconds and backward by 500 milliseconds, forming a scanning segment with a total duration of one second. All actually received behavior data records are extracted within this window. These records must simultaneously meet three conditions: first, their behavior type matches the current task node type; second, their operation trigger time falls within the time window; and third, their behavior source path corresponds to a currently known effective communication channel. The extracted behavior records are sorted in ascending order according to their trigger time, and a time sorting vector is constructed for subsequent time difference calculations. The purpose of this step is to establish a realistic behavior trigger time chain and compare it with the theoretical arrival timestamps to identify potential time discontinuities or sudden intervals.

[0023] For the sorted time chain of actions, the time interval between any two adjacent action events is calculated item by item. This calculation is based on the timestamp difference between adjacent actions, in milliseconds. For each group of adjacent actions, the time difference is calculated, and the result is compared with the dynamic time gap threshold introduced in this invention. This time gap threshold is not a fixed constant, but is calculated by combining multiple factors, including the importance level of the task node, the expected response time of the action operation, the average response cycle in historical tasks, and the stability score of the current communication channel in the network fluctuation prediction matrix. For example, in critical operation nodes such as "task submission" or "approval confirmation," the time gap threshold can be set to 800 milliseconds; while in regular information viewing nodes, it can be extended to more than 1,500 milliseconds. Once it is found that the time interval between a group of adjacent actions exceeds the corresponding threshold, a gap identifier is immediately inserted between the two actions. This gap identifier includes not only the time range of the gap location, but also the preceding and following action numbers used to trigger the gap determination, the time interval value, the threshold source parameter, the task node information, the action type category, and the stability label of the channel to which it belongs.

[0024] After the gap identifier is inserted, to further control the misleading impact of this area on model feature calculations during subsequent data processing, this invention embeds confidence-weighted labels in the behavior records on both sides of the gap. This label is not a simple Boolean variable, but a confidence weight structure containing multiple dimensions, mainly including five fields: behavior data source confidence level, timestamp integrity evaluation, behavior semantic consistency score, channel stability conversion value, and context behavior continuity score. The values ​​of these fields are generated based on the preceding network fluctuation prediction results, the current time series structure integrity, task context relevance, and the automatic scoring mechanism of historical behavior semantic models. For behavior segments with gaps, the weighted label reduces the overall behavior confidence from the original 100% to between 40 and 70%, with the specific value depending on the minimum value of the aforementioned five indicators. The weighted label is attached to the behavior record structure for reference in subsequent feature extraction processes. During sequence analysis, pattern recognition, or model training, the influence factor of this data segment will be automatically adjusted to avoid unreasonable bias or amplification of the final performance evaluation results.

[0025] After each sliding window execution, gap identification, and weighted label injection is completed, the time window advances by a set step size to ensure complete coverage of the entire behavior sequence. The step size is dynamically adjusted based on the behavior density within the current window. If the current window contains only two behavior events, the step size is set to half the length of the current window; if the behavior events are dense, the step size is reduced to one-third of the total window duration to avoid missing potential fine-grained gaps. During each window advancement, the center of the sliding window is aligned to the arrival time of the next expected behavior, and the above scanning, comparison, labeling, and weighting process is repeated until the entire task node behavior sequence has been traversed.

[0026] Under the constraint of confidence-weighted labels, the timeline is adaptively scaled and adjusted in real time according to the gap identifier, so that the feature extraction process is focused only on the high-confidence segment and avoids triggering continuous interpolation operations in the low-confidence segment. After identifying behavioral data gaps and injecting confidence-weighted labels, to prevent low-confidence segments from interfering with feature calculation, the timeline needs to be adaptively scaled based on confidence constraints, and the sliding window boundaries need to be adjusted in real time. This guides the feature extraction process to focus on trustworthy behavioral segments, improving analysis quality. The specific implementation steps are as follows: Based on the constructed confidence-weighted labels, a continuous confidence distribution map is established in the behavioral sequence. Specifically, for each behavioral event in the timeline, its associated confidence structure is extracted, containing five sub-items: data source score, timestamp integrity score, behavioral semantic consistency score, network path stability score, and contextual continuity score. Using the event timestamp as the x-axis, the weighted average of the five sub-items is calculated as the total confidence value for the behavior, and this value is mapped onto a unified timeline. After mapping all events, the confidence values ​​of adjacent time points are connected to form a confidence curve. On this curve, time periods with confidence above 80% are marked as high-confidence segments, time periods below 50% are marked as low-confidence segments, and the remaining segments are intermediate confidence regions. This curve serves as the benchmark for subsequent adjustments to window boundaries and timeline reconstruction.

[0027] Based on the annotation results of the continuous confidence distribution map, the boundaries of the sliding time window used for current feature extraction are dynamically adjusted. Assuming the original sliding window width is two seconds, it is symmetrically extended forward and backward by one second each, with the initial window center aligned with the expected behavior timescale of the current task node. In actual execution, the average confidence level between the time point of the left boundary and the center point is checked. If the built-in confidence level in this segment is continuously below 50%, the left boundary is moved closer to the center until the confidence level at the new boundary rises above 60%. The adjustment logic for the right boundary is the same as for the left, ensuring that the entire window only covers the confidence level-compliant area. During the adjustment process, if multiple confidence level fluctuations are encountered, segmented corrections can be performed to ensure that the window boundary does not cross multiple high-low frequency switching zones. The width of the adjusted new window should not be less than one second. If the adjustment results in the window's effective duration being less than one second, the current window processing is skipped, and the window is re-established at the next timescale.

[0028] Within the adjusted window segment, the confidence data undergoes further in-depth filtering and labeling. All behavioral events located in the high-confidence segment are labeled as "strong signal events," with their feature participation weight set to 120% of the standard value to enhance their representational ability in sequence modeling. All events located in the low-confidence segment are labeled as "weak signal events," with their weight reduced to 40% of the original value. Events in the intermediate region are treated as "ordinary signal events" and participate in subsequent calculations. Furthermore, for behavioral events immediately adjacent to the gap marker, even if their confidence scores are slightly higher, a contextual semantic comparison must be performed to determine if the behavior is affected by the gap and thus experiences semantic drift. If drift exists, such as a mismatch between the behavior content and the expected task, an abnormally early or late behavior position, or a discontinuous behavior type with adjacent events, its confidence score is forcibly lowered, and it is removed from the strong signal pool. This process effectively avoids erroneous reinforcement caused by misjudgments based on high confidence levels.

[0029] Based on the confidence classification results, an adaptive scaling and reconstruction operation is performed on the entire timeline. Specifically: for segments consisting of all strong signal events, time expansion processing is performed, increasing the corresponding time span by 30% from the original record, and increasing the density of behavioral sequence detail analysis within the expanded time interval, such as adding fine-grained timestamp difference extraction, behavioral event duration decomposition, and micro-action interval comparison indicators; for segments consisting of all weak signal events, time compression processing is performed, compressing their original time length to 50%, and prohibiting behavioral classification, behavioral clustering, and time series modeling processes within this compressed segment, retaining only basic existence judgments as behavioral time placeholders; for ordinary signal segments, the timeline remains unchanged, and only the standard feature extraction process is performed. After completing the scaling and reconstruction processing of all segments, the entire timeline is reassembled to generate a confidence-driven nonlinear time axis. This time axis is used for the entire process of subsequent feature extraction, behavioral pattern recognition, behavioral prediction modeling, and performance indicator calculation, and forms a natural isolation barrier for low-quality data segments, effectively blocking the diffusion path of data noise.

[0030] In the sliding window after adaptive scaling, a contrastive self-supervised generative network is introduced in combination with the gap identifier in the context. The micro-behaviors inside the gap are reconstructed using a bidirectional residual inference mechanism, and the reconstructed confidence map is output to achieve interpolation-free restoration of the behavior sequence. To achieve true semantic completion of missing segments in behavioral sequences, after adjusting the sliding window and adaptively scaling the timeline, a contrastive self-supervised generative network is introduced. Combined with contextual gap identifiers, a bidirectional residual inference mechanism is used to reconstruct the missing behaviors without interpolation, and a reconstructed confidence map is output to support subsequent computation. The specific implementation steps are as follows: Within the current sliding window, locate the start and end positions of the gap identifier and extract at least three valid behavioral events before and after the gap to form a complete contextual semantic reference structure. The gap identifier has been generated in the previous steps through timeout detection and confidence labeling, and includes the start and end timestamps of the abnormal time interval segment, the types of behaviors before and after, the task stage, the path stability score, and the semantic similarity index. The selection of behavioral events is based on the criterion that the confidence score is not less than 70%, ensuring that the contextual behaviors have high semantic completeness and temporal accuracy. Each behavioral event needs to record detailed operation action name, trigger time, behavior description information, user interaction attributes, historical behavior frequency, and behavior execution duration. After selection, this set of data is constructed into a set of contextual semantic templates to provide behavioral semantic boundaries and temporal evolution basis for gap reconstruction.

[0031] Guided by semantic templates, a contrastive self-supervised generative network is constructed as its training input to generate candidate behavior sequences for the gap region. This generative network is trained using a contrastive learning strategy, without relying on manual labeling. Positive sample pairs consist of consecutive high-confidence behaviors under the same task, while negative sample pairs consist of behavior pairs with significant time shifts in triggering, semantic incoherence, or high volatility in the source channel. During training, temporal sequence integrity, behavioral logical consistency, and behavioral response path continuity are used as multi-dimensional references. The encoder extracts contextual semantic features, and the decoder attempts to reconstruct possible micro-behavioral fragments within the gap. After completing one generation process, three sets of candidate sequences are output, each containing one or more behavioral events. Each event includes a predicted timestamp, operation type, behavioral target, behavioral context location number, and preliminary confidence score. All candidate sequences proceed to the next step of residual comparison processing to evaluate their suitability in terms of semantic continuity and logical rationality.

[0032] A bidirectional residual inference mechanism is initiated to perform forward and backward inference analysis on each candidate sequence, calculate its residual value, and then filter it. Forward inference starts from the pre-gap behavior sequence to predict the behavior trajectory under natural development; backward inference derives the pre-gap behavior trend from the post-gap behavior sequence. Each direction outputs a behavior content deviation vector, a behavior time offset vector, and a behavior logical transition probability. The behavior content deviation vector is measured based on the semantic distance between the behavior description word vectors; the time offset vector is solved by the difference between the actual prediction time and the average time of the context behavior; the logical transition probability is obtained by modeling the frequency of occurrence and transition probability of the behavior in the evolution path of the historical behavior sequence. Finally, by weighted integration of the residual values ​​of the three dimensions, a total residual score is calculated for each candidate behavior sequence. Candidate sequences with residual scores below a preset threshold are considered reconstructed reliable sequences and are retained, while the rest are discarded. This mechanism can simultaneously evaluate the rationality of behavior at both the semantic and temporal levels, effectively avoiding meaningless behavior fragments that may be introduced by interpolation methods.

[0033] The selected reconstructed behavior sequences are inserted into the gaps in the original behavior sequences, and a reconstruction confidence mapping table is calculated based on the residual score and context similarity. Behaviors with a reconstruction confidence score of at least 85% are included in the formal behavior chain and participate in feature extraction and evaluation calculations; behaviors with scores between 70% and 85% are considered as backup behaviors, stored only in the behavior trajectory database, and do not affect the calculation of the main indicators; candidate behaviors with scores below 70% are completely discarded. The completed reconstruction confidence mapping table must include the behavior number, behavior name, prediction time, context fragment identifier, forward residual value, backward residual value, behavior semantic offset, time offset degree, and final confidence score. This table will be used to guide the dynamic weighting of behavior confidence in subsequent behavior evaluation paths.

[0034] Based on the reconstructed confidence mapping, the recursive confidence projection mechanism is used to map the reconstructed micro-behavioral fragments back to the complete behavioral sequence, and the full-cycle performance indicators are incrementally recalculated. The new calculation results are then compared with the previous calculation results to capture the evaluation drift caused by gap repair. To effectively integrate the reconstructed micro-behavioral fragments into the original behavioral sequence and correct evaluation bias caused by missing data, this invention employs a recursive confidence projection mechanism based on reconstructed confidence mapping to achieve behavioral sequence structure restoration, performance indicator increment recalculation, and evaluation drift identification and feedback. The specific implementation steps are as follows: The system reads the reconstruction confidence mapping table, identifies all reconstruction behavior fragments, and inserts them one by one into the gap positions of the original behavior sequence in chronological order. Each reconstruction behavior must meet the following three admission criteria: its confidence score is higher than 85%; both bidirectional residual values ​​are less than a set threshold; and the behavior content is consistent with the semantic structure of the gap context. During insertion, the timestamp, behavior content, context anchor, confidence score, generation source identifier, and behavior label of the behavior are all written into the behavior sequence, and the entire behavior sequence is recalibrated on the timeline to ensure that there is no time overlap or reverse order misalignment between preceding and following behaviors. After insertion, a recursive confidence projection mechanism is initiated. Taking each new behavior as the base point, the system identifies its logical dependencies on the preceding behavior chain line forward and recursively projects the behavior response chain that it may affect backward. During the recursive process, the confidence adjustment value of each affected behavior is calculated, considering factors including: the impact of the new behavior on the interval between behaviors, the reconstruction of the behavior collaboration relationship by the new behavior, and the degree to which the new behavior completes the behavior context structure. The projection mechanism will mark the affected behaviors as "confidence enhanced", "confidence weakened" or "confidence maintained" based on the level of impact, and update the corresponding fields in the behavior sequence for reference in the subsequent indicator recalculation process.

[0035] Based on the behavioral sequence that has completed recursive confidence projection, incremental recalculation of full-cycle performance indicators is performed. This calculation does not recalculate all indicators from scratch, but rather identifies the time segments of tasks where changes have occurred and only updates the affected indicator dimensions. First, the start and end times of the task phase to which the reconstructed behavior belongs are extracted, along with the set of key performance indicators within that segment. Common indicators include: task response timeliness rate, task processing quality score, behavioral synergy index, cross-departmental collaboration frequency, knowledge transfer chain integrity, and innovative behavior trigger rate. During incremental recalculation, for each indicator, it is first determined whether it directly depends on the new behavior. For example, if the new behavior is "cross-team comment feedback," it directly affects "cross-departmental collaboration frequency" and "behavioral synergy index." Then, based on the confidence level of the behavior, its participation weight is weighted: behaviors with a confidence level above 90% are weighted at 120%, those between 85% and 90% are weighted at 100%, and behaviors below this threshold are not included in the indicator calculation. During the recalculation process, the logical dependencies between indicators also need to be adjusted synchronously. For example, if the new behavior restructures the continuity of the knowledge transfer chain, the connection segment score and chain integrity score of the "Knowledge Transfer Chain Integrity" indicator need to be updated synchronously. Finally, an updated full-cycle performance indicator report is generated, with the addition of three items: "Indicator Fields Changed Due to Restructuring Behavior", "Direction of Indicator Change", and "Amount of Change".

[0036] To determine whether the current gap repair has caused a systemic drift in the overall performance evaluation, this invention constructs a differential comparison mechanism. It compares the current version of the performance indicator report with the previous version item by item to capture the degree of evaluation shift caused by data repair. Specifically, a performance indicator differential matrix is ​​established. The horizontal axis represents all indicator dimensions, and the vertical axis includes the original value, updated value, absolute change value, direction of change, relative rate of change, ranking change, change in contribution weight, and the number of tasks affected by the data. Changes in each dimension are quantified as a percentage, with the direction of change being "increased" or "decreased." A change rate greater than 20% is marked as "significant shift." Next, the indicator changes are comprehensively scored, and a weighted average method is used to integrate the degree of change of each indicator with its weight contribution in the overall score, resulting in an overall evaluation drift value. This value is divided into five levels based on the scoring results: "No drift" (change rate less than 2%), "Mild drift" (2% to 5%), "Moderate drift" (5% to 10%), "Significant drift" (10% to 20%), and "Severe drift" (greater than 20%). If the system identifies a "significant drift" or "serious drift," it automatically triggers a behavior reconstruction backtracking check mechanism to re-examine the contextual logic consistency, timestamp rationality, and semantic path coherence of the reconstructed behavior. If no reconstruction error is found, the drift value is included in the performance evaluation deviation field for reference in human resource optimization, performance feedback review, and incentive allocation rebalancing strategies.

[0037] After identifying evaluation drift, the multi-agent collaborative control engine is activated. Based on the results of recursive confidence projection, the task-level incentive weights are adaptively routed. At the same time, real-time re-evaluation is performed on high-risk nodes, and reward allocation is frozen in abnormal sections, thus forming a closed-loop control mechanism that can automatically correct deviations. To prevent the performance indicator drift caused by behavioral restructuring from continuing to expand, a control mechanism based on a multi-agent structure needs to be activated. This mechanism involves dynamically adjusting incentive weights, re-evaluating high-risk behavior areas, and implementing a freeze strategy for abnormal behavior areas, ultimately forming a self-correcting and adaptive evaluation closed loop. The specific implementation steps are as follows: After identifying evaluation drift, all task nodes with evaluation offset values ​​exceeding 10% are extracted from the differential index matrix generated in the previous stage. This, combined with the behavioral weight changes from recursive confidence projection, generates a "Drift Response Initiation Table." This table includes the following fields: task number, start and end time, drift level, drift direction, number of involved behaviors, average confidence of reconstructed behaviors, rate of change of key indicators, and the entity contributing to the behavior. Tasks meeting the following three conditions will be marked as "triggering intervention": at least three core indicators have a rate of change greater than 10%; two or more reconstructed behaviors are attributed to the same responsible person; and the behavioral change affects the task completion cycle or scoring ranking. This triggering mechanism ensures that intervention resources are accurately allocated to key task nodes that have a substantial impact on the organization's evaluation structure, improving collaborative response efficiency.

[0038] For task nodes marked as "triggered regulation," all responsible parties within that node are extracted, and the intensity of their behavioral responsibility in the task is calculated. The intensity of responsibility is calculated based on four parameters: the ratio of the number of actual behavioral events of the participant to the total number of behaviors in the task; the average confidence level of their behavior; the contribution of the participant's behavior to the changes in evaluation indicators before and after reconstruction; and the criticality of the participant's behavior at the beginning or end of the task's logical sequence. Based on this intensity of responsibility, a four-dimensional mapping structure of "responsible party—behavior—indicator—weight" is established within the task. Adaptive routing of incentive weights is performed on this structure, that is, the incentive ratio originally statically allocated according to task roles is fine-tuned based on behavior quality and behavior confidence. Specifically, if a participant exhibits multiple high-risk reconstruction behaviors in their responsible segment, and their average behavior confidence level is below 80%, their incentive weight will be reduced by 10% to 20%; if a participant's behavior confidence level is consistently above 95%, and the continuity and contextual consistency of the behavior sequence are good, their incentive weight can be increased by 5% to 10% from the original level. The adjusted weights will be synchronously written into the incentive allocation list and marked as "adaptive control source".

[0039] For task node behavior segments with high responsibility intensity scores but large fluctuations in the confidence level of reconstruction behaviors or unreasonable jumps in the behavior sequence structure, they are marked as "re-evaluation candidate segments". During the re-evaluation process, all reconstruction behaviors and their original contextual behavior events for this segment are re-extracted to construct a behavior sequence comparison structure diagram. This diagram uses time sequence as the horizontal axis and the behavior semantic continuity score as the vertical axis, marking the sequence differences before and after each behavior reconstruction to form a behavior chain path diagram. Each reconstruction behavior is analyzed sequentially to determine whether it introduces semantic disjointness, temporal leaps, or overlapping repetitive behaviors in its context. If the timestamp offset of a behavior is greater than twice the expected context interval, or its semantic label is inconsistent with the classification of preceding and following events, or if the behavior logic experiences a non-linear break, then the behavior will be marked as "requiring secondary correction". The re-evaluation report will embed detailed structural change annotations in the task node behavior structure for subsequent behavior chain repair and reverse adjustment of indicators.

[0040] For sections of behavior marked as "requiring secondary correction" during reassessment, an incentive freeze mechanism will be implemented. The criteria for freezing are: the confidence level of the reconstructed behavior is below 75%, there is a sudden change in the semantic structure before and after the behavior, or the reconstructed behavior affects two or more core performance indicators. The freeze includes: freezing the incentive allocation processing for the task section containing the behavior; freezing the corresponding contribution reward for the person responsible for the behavior; and freezing the scoring push permission for that section in the task result submission chain. Freeze records will be generated in the form of a "freeze index table," with record fields including behavior number, freeze time, freeze reason, responsible person identification number, corresponding task identifier, expected freeze duration, and unfreeze trigger conditions. Frozen data will not participate in the current period's performance statistics but will be retained in the behavior history for subsequent review. Unfreezing requires a new round of semantic correction or confidence level recovery by the behavior reconstruction mechanism, and confirmation from the responsible person before it can be lifted.

[0041] The adaptive adjustment results of task incentive weights, the reassessment results of high-risk nodes, and the list of frozen behaviors are summarized to form a "task-level closed-loop control feedback package." This feedback package will serve as a reference for regulation in the next task cycle, and a behavior quality threshold model will be established in advance for similar tasks. For example, for task types that experience two consecutive rounds of incentive rollback or freezing, a behavior confidence monitoring zone will be set in advance for their behavior trigger path; and those responsible for behaviors whose weights are continuously reduced will be guided to participate in behavior standard optimization training to improve data behavior quality.

[0042] This invention constructs a multi-path network fluctuation prediction matrix to generate expected behavior arrival timestamps in advance. Combined with a sliding window dynamic scanning mechanism, it can accurately identify behavioral gaps and label them with confidence levels, guiding subsequent feature extraction to focus only on high-confidence behavioral segments, thus suppressing the interference of untrusted behaviors on the evaluation model from the source. Building upon this, a contrastive self-supervised generative network and a bidirectional residual inference mechanism are introduced, enabling the realistic semantic reconstruction of behaviors within the gaps without interpolation. A recursive confidence projection mechanism achieves high-fidelity completion of the original behavioral sequences. Finally, combined with a multi-agent control engine, it can intelligently perceive and close-loop correct performance evaluation drift caused by reconstructed behaviors, comprehensively improving the objectivity, dynamism, and stability of performance evaluation. In summary, this invention significantly enhances the robustness of performance evaluation systems to abnormal behavioral data, their adaptability to dynamic changes, and their ability to accurately allocate incentive resources while maintaining data authenticity.

[0043] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for evaluating employee performance based on behavior data driving, characterized in that, The method comprises the following steps: Construct a multi-path network fluctuation prediction matrix, generate an expected behavior arrival time scale based on the transmission state of the task switching node, and use it as the baseline for behavior sequence gap identification; Under the guidance of the expected behavior arrival time scale, perform sliding window scanning on the received behavior sequence, generate a gap identification when the time gap is detected to exceed the threshold, and inject a confidence reduction weight label; Under the constraint of the confidence reduction weight label, perform adaptive stretching on the timeline, and adjust the sliding window boundary according to the gap identification to make the feature extraction concentrate on the section with high confidence; In the adaptive stretched sliding window, introduce a self-supervised generation network combined with the gap identification, reconstruct the micro-behaviors in the gap using a bidirectional residual reasoning mechanism, and generate a reconstruction confidence map; Based on the reconstruction confidence map, use a recursive confidence projection mechanism to map the reconstructed micro-behaviors back to the complete behavior sequence, recalculate the performance indicators incrementally, and capture the evaluation drift by differentiating the previous results; After identifying the evaluation drift, start the multi-agent collaborative regulation engine, adaptively route the task incentive weight based on the recursive confidence projection result, perform reevaluation on the high-risk nodes, and freeze the reward distribution in the abnormal section to form an automatic correction closed-loop regulation mechanism.

2. The behavior data driven employee performance evaluation method of claim 1, wherein, The step of constructing a multi-path network fluctuation prediction matrix and generating an expected behavior arrival time scale based on the transmission state of the task switching node is as follows: A multi-path set for behavior data collection is established, each path is divided according to different network paths, and the transmission state data of the task switching node is recorded; Based on the collected data, the minimum delay, maximum delay, average response time and delay fluctuation value of each path at each task node are calculated, and a path delay behavior trajectory graph is drawn to extract a representative node window to generate a prediction stability index; The node-level stability indexes of multiple paths are fused to generate a prediction matrix, and the path delay interval is aligned combined with the task switching trigger time to construct a behavior expected arrival time window; According to the path stability index, perform confidence weighted fusion to generate an expected behavior arrival time scale as the timing baseline for behavior gap identification.

3. The behavior data driven employee performance evaluation method of claim 2, wherein, The step of generating a gap identification when the time gap is detected to exceed the threshold, and injecting a confidence reduction weight label is as follows: Based on the expected behavior arrival time scale, construct a sliding time window, and extract the behavior records in the window that match the task node, time, and communication path, and construct a time ordering vector; Calculate the time interval between adjacent behavior records, and compare it with the dynamic time interval threshold calculated based on the importance of the task node, the expected response time, the historical response period and the path stability. If the time interval exceeds the threshold, insert a gap identification; Embed a confidence reduction weight label in the behavior records before and after the gap; The sliding window is pushed back by a dynamic step, and the above process is repeated with the next expected behavior arrival time scale as the center to complete the full sequence coverage.

4. The behavior data driven employee performance evaluation method of claim 3, wherein, The step of performing adaptive stretching on the timeline and adjusting the sliding window boundary according to the gap identification to make the feature extraction concentrate on the section with high confidence is as follows: Based on the confidence weight reduction label, a confidence continuous distribution graph is constructed. The confidence structure of each behavior event is extracted and mapped to a unified time axis, and high, medium and low confidence sections are marked. According to the confidence continuous distribution graph, the boundary position of the sliding time window is adjusted, only the behavior event section with confidence higher than the preset threshold is retained, and the time segment with dramatic confidence fluctuation is removed. The confidence of the behavior event in the adjusted window is classified and weighted to identify strong, ordinary and weak signal events, mark the context semantic drift behavior and reduce the confidence score. According to the signal intensity, the timeline is adaptively stretched and contracted, the time proportion of strong signal section is expanded, the time span of weak signal section is compressed, and the nonlinear time axis is formed for subsequent feature extraction and evaluation modeling.

5. The behavior data driven employee performance evaluation method of claim 4, wherein, The self-supervised generation network is introduced combined with the gap identification. The micro behavior in the gap is reconstructed by using the bidirectional residual reasoning mechanism, and the reconstructed confidence mapping is as follows: Locate the gap identification position, and extract not less than three behavior events before and after the gap with confidence not less than 70%, and construct the context semantic template. Under the guidance of the context semantic template, a plurality of candidate behavior sequences are generated by the contrastive self-supervised generation network, and the initial confidence score is generated according to the time sequence integrity, behavior logic consistency and response path continuity. Based on the bidirectional residual reasoning mechanism, the forward and backward residual calculation is performed on each candidate behavior sequence, and the behavior content deviation vector, time offset vector and logic transfer probability are output, and the sequence with residual score not higher than the threshold is screened. The behavior sequence passing the screening is inserted into the gap position, and the reconstructed confidence mapping is generated combined with the residual score and the context similarity, and the reconstructed confidence mapping table is constructed.

6. The behavior data driven employee performance evaluation method of claim 5, wherein, In the reconstructed confidence mapping table, only when the reconstructed confidence score of the candidate behavior sequence is higher than 85%, the behavior sequence is included in the behavior chain to participate in the performance indicator calculation, the behavior sequence with score between 70% and 85% is marked as standby behavior, and only used for behavior trajectory recording, and the behavior sequence lower than 70% is removed.

7. The behavior data driven employee performance evaluation method of claim 5, wherein, The reconstructed micro behavior is mapped back to the complete behavior sequence by using the recursive confidence projection mechanism, the performance indicator is incrementally recalculated, and the evaluation drift is captured by difference from the previous result. The steps are as follows: Based on the reconstructed confidence mapping, the reconstructed micro behavior segment is inserted into the gap position of the original behavior sequence, and the behavior time axis is recalibrated. Using the recursive confidence projection mechanism, the influence range of the added behavior on the behavior chain is identified, and the confidence change state of the affected behavior is marked. Based on the updated behavior sequence, the incremental recalculation of the full cycle performance indicator is performed on the task section involving the added behavior, and the difference matrix is generated by comparing the original indicator to identify the evaluation drift level.

8. The behavior data driven employee performance evaluation method of claim 7, wherein, The high-risk nodes are re-evaluated, and the reward distribution is frozen in the abnormal section to form an automatic correction closed-loop regulation mechanism. The steps are as follows: After the evaluation drift is identified, the task node with evaluation offset value exceeding 10% is extracted, and the drift response start table is constructed combined with the recursive confidence projection result to determine the regulation task. The responsibility intensity of the behavior person in the regulation task is calculated based on the drift response starting table, and a mapping relationship between the responsibility person and the behavior is established; The adaptive routing adjustment of the incentive weight is performed based on the responsibility intensity and the behavior confidence, the incentive proportion is increased for the high-confidence behavior, and the incentive proportion is decreased for the low-confidence behavior; For the task segment with high responsibility intensity and large reconstruction behavior fluctuation, re-evaluation is performed, and behaviors with inconsistent semantic structure or time jump anomaly are identified as behaviors requiring secondary correction; The incentive freezing operation is performed on the section where the behavior requiring secondary correction is located, and the adjustment result and the re-evaluation information are integrated to form a task-level closed-loop control feedback package for reference in the subsequent task period.

Images