Method and system for constructing talent portrait based on behavior sequence mining and storage medium

Abstract

The present application provides a kind of talent portrait construction method, system and storage medium based on behavior sequence mining, comprising: collecting talent multi-source heterogeneous behavior data and extracting single behavior feature, the behavior initial polarity is modeled as the two-dimensional probability polarity state vector of positive and negative superposition state;Context rotation operator is constructed relying on behavior-context co-occurrence matrix, and the polarity probability distribution is corrected;The mapping model of ability dimension and measurement base is established, the vector projection is completed in combination with behavior intensity, and the positive and negative behavior components of each ability dimension are obtained;The time sequence correlation is represented by constructing behavior transition probability graph;With the aid of time sequence attenuation function, behavior component is weighted and aggregated, generates each dimension positive and negative cumulative features, constructs the multi-dimensional talent portrait of positive and negative behavior decoupling, realizes the talent behavior tendency distinction.

Description

Technical Field

[0001] This application belongs to the field of talent profiling, and in particular relates to a method, system and storage medium for constructing talent profiles based on behavioral sequence mining. Background Technology

[0002] With the continuous advancement of digitalization in human resource management, the application value of talent profiling technology in corporate recruitment, employee training, performance evaluation, and talent inventory is becoming increasingly prominent. Talent profiling typically involves collecting and analyzing multi-dimensional data generated by employees during their daily work, project participation, and collaborative interactions to identify individual ability characteristics, behavioral tendencies, and development potential. To improve the completeness and accuracy of profiling results, existing systems usually collect multi-source heterogeneous behavioral data of the talent to be profiled within a certain time period. This data covers multiple aspects, including task completion status, communication records, system operation logs, and business approval records. By extracting features and comprehensively analyzing the above behavioral data, enterprises can evaluate talent from multiple preset ability dimensions. Therefore, constructing a scientific and reasonable talent profiling model that fits actual business scenarios is of great significance for optimizing human resource allocation, tapping employee potential, and improving the overall operational efficiency of the organization.

[0003] However, in real-world work scenarios, employee behavior is highly complex and context-dependent. A single behavior may not necessarily reflect only positive or negative meanings in different business contexts, but may simultaneously contain both positive and negative impacts. For example, while a high-intensity, intensive task may boost performance in the short term, it may also be accompanied by potential risks such as friction in teamwork, decreased process standardization, or subsequent efficiency reduction. Existing talent profiling methods often employ linear scoring or single-polarity classification, making it difficult to effectively decouple and separately evaluate the coexisting positive and negative characteristics of the same behavior. Furthermore, employee performance is not a static result but a continuously evolving process over time; there is usually a temporal correlation and influence transmission relationship between historical and current behaviors. Traditional methods often treat each behavior as an independent event and simply add them up, lacking the characterization of behavioral transfer relationships and the attenuation effect of historical influence. The aforementioned problems of insufficient decoupling of multifaceted behaviors and lack of temporal correlation analysis can easily lead to flat talent profiles, making it difficult to accurately represent the positive and negative behavioral performance of talents under different ability dimensions. Therefore, there is an urgent need for a talent profile construction method that can decouple positive and negative behaviors and take into account the characteristics of temporal evolution. Summary of the Invention

[0004] According to a first aspect of the present invention, a method for constructing talent profiles based on behavioral sequence mining is proposed to address the problem that existing methods result in overly flat talent profiles that fail to accurately represent the positive and negative performance of talents across different ability dimensions. The method includes the following steps:

[0005] Acquire multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extract single behavioral features; for each single behavior, model the initial polarity of the single behavior as a two-dimensional probability polarity state vector of positive and negative superposition; construct a behavior-context co-occurrence matrix based on the single behavior and related events, extract two-dimensional context features and calculate a context rotation operator; apply the rotation operator to perform a rotation operation on the two-dimensional probability polarity state vector to adjust the positive and negative probability distribution of the vector;

[0006] Establish a mapping model from N preset capability dimensions to their corresponding measurement bases; combine the rotated two-dimensional probability polarity state vector with the behavior intensity value representing the degree of influence of the behavior, and project it onto the measurement base corresponding to each dimension based on the correlation between the behavior and the capability dimension to decompose and obtain positive and negative behavior components belonging to the N capability dimensions; construct a behavior transition probability graph representing the temporal correlation of a single behavior, wherein the edge weights of the probability graph are determined based on the inner product of the rotated two-dimensional probability polarity state vectors of adjacent behavior nodes;

[0007] A temporal decay function is set, and based on the behavior transition probability map and decay function, the positive and negative behavior components decomposed from all behaviors are time-weighted and aggregated to generate positive and negative cumulative features for each of the N capability dimensions. Based on the positive and negative cumulative features of the N capability dimensions, a talent profile of the candidate to be profiled, with decoupled positive and negative behaviors in multiple dimensions, is constructed.

[0008] According to a second aspect of the present invention, a talent profile construction system based on behavioral sequence mining is proposed, comprising the following modules:

[0009] The extraction module is used to acquire multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extract single behavioral features; for each single behavior, the initial polarity of the single behavior is modeled as a two-dimensional probability polarity state vector of positive and negative superposition; a behavior-context co-occurrence matrix is ​​constructed based on the single behavior and related events, and two-dimensional context features are extracted to calculate the context rotation operator; the rotation operator is applied to perform a rotation operation on the two-dimensional probability polarity state vector to adjust the positive and negative probability distribution of the vector;

[0010] A construction module is used to establish a mapping model from N preset capability dimensions to their corresponding measurement bases; the rotated two-dimensional probability polarity state vector is combined with the behavior intensity value representing the degree of influence of the behavior, and based on the correlation between the behavior and the capability dimension, it is projected onto the measurement base corresponding to each dimension to decompose and obtain positive and negative behavior components belonging to the N capability dimensions; a behavior transition probability graph representing the temporal correlation of a single behavior is constructed, and the edge weights of the probability graph are determined based on the inner product of the rotated two-dimensional probability polarity state vectors of adjacent behavior nodes;

[0011] The generation module is used to set a temporal decay function and, based on the behavior transition probability map and decay function, perform temporal weighted aggregation on the positive and negative behavior components decomposed from all behaviors to generate positive and negative cumulative features for each of the N capability dimensions; based on the positive and negative cumulative features of the N capability dimensions, a talent profile of the candidate to be profiled, with decoupled positive and negative behaviors across multiple dimensions, is constructed.

[0012] This invention identifies the dual attributes of behavioral polarity in reality by acquiring multi-source heterogeneous behavioral data and modeling the initial polarity of behavior as a two-dimensional probabilistic polarity state vector of positive and negative superposition. By performing a rotation operation on this vector using a behavior-context co-occurrence matrix and a rotation operator, the accuracy of behavioral polarity assessment is improved by correcting the positive and negative probability distributions based on the specific scenario context. Projective decomposition using a measurement basis mapping model enables the calculation, decomposition, and decoupling of the positive and negative impacts of a single behavior on various capability dimensions. Combining the behavior transition probability map determined by the vector inner product and the temporal decay function, the positive and negative behavioral components are time-weighted and aggregated, considering the timeliness value of behavior and the correlation patterns between sequences. This solves the problems of absolutism in polarity determination and isolation of behavior in traditional methods, generating a multi-dimensional, decoupled talent profile of positive and negative aspects. Attached Figure Description

[0013] Figure 1 A flowchart of the first embodiment;

[0014] Figure 2 This is a diagram illustrating the distribution of check-in behavior.

[0015] Figure 3 This is a statistical chart of document submission and approval. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the first embodiment, the present invention proposes a method for constructing talent profiles based on behavioral sequence mining, such as... Figure 1 It includes the following steps:

[0018] S1. Acquire multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extract single behavioral features; for each single behavior, model the initial polarity of the single behavior as a two-dimensional probability polarity state vector of positive and negative superposition; construct a behavior-context co-occurrence matrix based on the single behavior and related events, extract two-dimensional context features and calculate a context rotation operator; apply the rotation operator to perform a rotation operation on the two-dimensional probability polarity state vector to adjust the positive and negative probability distribution of the vector.

[0019] Structured behavioral data such as performance ratings and attendance records are extracted from the enterprise data management system, while unstructured behavioral data such as code submission records are obtained through the code hosting platform interface. For text-based behavioral description data, a text feature extraction model is used to map the text into high-dimensional feature vectors; for numerical behavioral data, interval normalization is performed. Text feature vectors belonging to the same behavioral record, the same business event, or the same preset time window are aligned and concatenated with numerical features to form a single behavioral feature vector.

[0020] The pre-trained sentiment analysis model assigns an initial polarity score to the text description of the single behavior, resulting in a sentiment polarity score between 0 and 1. This score is used as the base probability for positive polarity, and the value obtained by subtracting the score from 1 is used as the base probability for negative polarity. The square roots of both the positive and negative polarity base probabilities are then calculated to obtain the positive and negative probability amplitudes. The positive and negative probability amplitudes are combined as the x-axis and y-axis to form a two-dimensional real-valued column vector with a magnitude of 1, serving as the initial two-dimensional probability polarity state vector, thus completing the transformation from a deterministic polarity to a superposition of positive and negative polarities.

[0021] Collect all individual behaviors and their associated events, such as project development events or team collaboration events, and construct a co-occurrence frequency matrix with behavior categories as rows and associated events as columns. Initialize the co-occurrence frequency matrix and accumulate the occurrence frequencies according to behavior category and associated events. Perform singular value decomposition on the co-occurrence frequency matrix to obtain the left singular vector corresponding to the largest singular value as the context feature vector. Project the context feature vector onto a two-dimensional space through a pre-trained linear dimensionality reduction mapping matrix and perform L2 normalization to obtain a two-dimensional unit vector. Use the two components of this two-dimensional unit vector as the cosine and sine values ​​of the rotation angle, respectively, and construct a 2×2 two-dimensional real number rotation matrix as the context rotation operator.

[0022] The calculated 2×2 context rotation operator is multiplied by the constructed two-dimensional probability polarity state vector. The square of the new vector's x-coordinate is calculated as the updated positive probability, and the square of the new vector's y-coordinate is calculated as the updated negative probability. This process uses contextual information to perform a two-dimensional planar rotation correction on the initial emotional polarity of a single behavior, adjusting the positive or negative polarity expression of the behavior within a specific business context.

[0023] In an optional embodiment, the step of acquiring multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extracting single behavioral features includes:

[0024] Extract attendance records, project delivery documents, and business flow approval data of the talent to be profiled from the company's internal database within a preset time period;

[0025] The attendance record data is parsed into attendance behavior features by timestamp, and the project delivery document is converted into document submission behavior features using word frequency statistics rules. The approver field and approval time value are extracted from the business flow approval data as approval behavior features.

[0026] The attendance behavior characteristics, document submission behavior characteristics, and approval behavior characteristics are uniformly structured into a single set of behavior characteristics that includes timestamps and behavior categories.

[0027] Obtain raw behavioral log data of the candidate to be profiled within a preset time period, such as the past twelve months, i.e., the time window T is set to 365 days. For attendance record data, extract the daily sign-in and sign-out timestamps and compare them with preset standard working hours to obtain attendance behavior features such as on-time attendance, lateness, early departure, or overtime. Among these, behaviors such as being late for more than a preset lateness duration threshold or working overtime for more than a preset overtime duration threshold can be set as high-weight sub-features for extraction.

[0028] For delivery data such as project delivery documents, text processing documents, or code commit records, the text content, commit time, number of commits, and deliverable type are parsed, and delivery behavior features are extracted by combining preset keyword rules or word frequency statistics rules. For example, the number of code commits, the number of document commits, or the frequency of key deliverables within a preset period are counted and converted into document commit behavior features that characterize the document commit activity, delivery completeness, or project participation level.

[0029] For business workflow approval data, the node information in the approval process is parsed to obtain the approver field, approval initiation time, approval completion time and workflow status, and an approval timeliness value is generated based on the time difference between workflow creation and processing completion; for example, records with approval timeliness less than the preset efficient approval threshold are marked as efficient approval behavior, and records with approval timeliness greater than the preset delayed approval threshold are marked as delayed approval behavior.

[0030] Subsequently, the aforementioned structured, unstructured, or semi-structured behavioral features are uniformly converted into a standardized data structure. Each individual behavioral record must include at least a behavior timestamp, behavior subject identifier, behavior category, behavior duration or processing time, and the corresponding standardized behavior category code. Based on this standardized data structure, a set of single behavioral features with temporal and business classification attributes is constructed for subsequent behavior polarity modeling, capability dimension projection, and temporal aggregation analysis. The distribution of employee time-series clock-in / out behavior is shown below. Figure 2 As shown in the figure, the statistics of employee document submission and approval behavior are as follows: Figure 3 As shown.

[0031] In an optional embodiment, the step of modeling the initial polarity of each individual action as a two-dimensional probabilistic polarity state vector of positive and negative superposition states includes:

[0032] Extract the sentiment polarity label value of the single behavior, and obtain the positive polarity base probability and negative polarity base probability of the single behavior;

[0033] The normalization factor is calculated as the square root of the sum of the positive polarity base probability and the negative polarity base probability. If the normalization factor is 0, then the normalization factor is assigned a preset small positive number.

[0034] Construct a two-dimensional column vector, setting the first element as the square root of the positive polarity base probability divided by the normalization factor, and the second element as the square root of the negative polarity base probability divided by the normalization factor. Output a two-dimensional column vector that satisfies the L2 norm normalization condition, which serves as the two-dimensional probability polarity state vector of the positive and negative superposition states.

[0035] To model the superposition of positive and negative states of a single action polarity, a pre-trained domain sentiment classification deep model is invoked, such as a text convolutional neural network fine-tuned based on enterprise-specific management corpora. The input of the domain sentiment classification deep model is the context text corresponding to the single action or the extracted structured numerical parameters. The network structure of the domain sentiment classification deep model includes an embedding layer, a one-dimensional convolutional layer, a global max pooling layer, and a fully connected layer. The embedding layer maps the discrete text words into continuous dense word vector representations; the one-dimensional convolutional layer uses various convolutional kernels of different sizes to perform sliding window convolution calculations on the word vector sequence to extract local contextual features; the global max pooling layer selects the maximum value from the feature maps extracted by each convolutional kernel as the global feature and concatenates the maximum values ​​corresponding to different convolutional kernels into a fixed-length feature vector; the fully connected layer receives this feature vector and maps it to a two-dimensional output space through a linear transformation, connecting it to a normalized exponential function layer.

[0036] The output of the domain sentiment classification deep model is the classification probability value of the model's normalized exponential function layer. This probability value is used as the positive polarity base probability (variable p) and the negative polarity base probability (variable q), respectively. For example, for a complex behavior involving a project delay but urgent repair of a high-risk vulnerability, the sentiment model calculates a positive probability p = 0.25 and a negative probability q = 0.75. The polarity normalization factors for both are then calculated. To prevent division by zero errors caused by missing data or network anomalies resulting in both p and q being 0, a constraint check is performed. When a constraint is detected... At that time, Z is assigned a system-preset small positive constant. To ensure the stability of the mathematical calculation process, a 2×1 two-dimensional column vector is initialized based on these calculation parameters. Set the first matrix element of this vector to represent the amplitude of the positive state. Let the second matrix element of this vector represent the negative state amplitude. Taking input variables p=0.25 and q=0.75 as an example, since p+q=1 and Z=1, performing the square root operation yields... =0.5, ≈0.866. At this point, the two-dimensional probabilistic polarity state vector generated and returned by the system is... The resulting vector satisfies The L2 norm normalization constraint is applied. This processing allows the same recorded behavior in a workplace scenario to account for both positive and negative tendencies with differentiated probability amplitudes.

[0037] In an optional embodiment, the step of constructing a behavior-context co-occurrence matrix based on a single behavior and associated events, extracting two-dimensional context features, and calculating a context rotation operator includes:

[0038] Collect all behavioral categories that appear within a specified event, and construct a behavior-to-context co-occurrence frequency matrix with behavioral categories as rows and event identifiers as columns;

[0039] Singular value decomposition is performed on the co-occurrence frequency matrix of the behavior to the context, and the left singular vector corresponding to the maximum singular value is obtained as the context feature vector.

[0040] The context feature vector is reduced to a two-dimensional space through linear mapping and then L2 normalized to obtain a two-dimensional unit vector containing the first and second components.

[0041] The first component of the two-dimensional unit vector is set as the cosine value of the context rotation angle, and the second component is set as the sine value. A second-order rotation orthogonal transformation matrix is ​​generated with the cosine value and negative sine value as the first row and the sine value and cosine value as the second row, which serves as the context rotation operator.

[0042] Extract all behavioral node data occurring within a specified business event time window period, where the specified business event time window period is, for example, one standard agile iteration cycle, and the event identifier set is set to... to The system collects data on all behavioral nodes occurring within a given time period, and statistically analyzes K different standardized behavioral categories generated during that period, such as business requirement review, core code retrieval, and online defect fixing. Based on this mapping and statistics, a K×M behavior-to-context co-occurrence frequency matrix C is constructed in memory, and each element is defined... Record the cumulative total frequency of the i-th type of canonical behavior actually occurring in the j-th specified event, for example, record the frequency in the first iteration. Fifteen core code commits occurred during the event cycle. Singular Value Decomposition (SVD) was performed on the sparse feature matrix C, decomposing it into... The form is given, and the first eigenvector corresponding to the maximum singular value is extracted from the left singular orthogonal matrix U obtained by decomposition. This serves as a context feature vector, containing the main variance information and context feature distribution of the business cycle.

[0043] To input the extracted feature vectors into the calculation process, a pre-trained dimensionality reduction projection matrix W of size 2×K is initialized, and matrix multiplication is performed. Projecting the feature column vector into a two-dimensional space using linear dimensionality reduction yields a two-dimensional coordinate vector. Calculate the Euclidean magnitude of the vector and perform L2 normalization, outputting a vector containing the first dimension component. Second dimension components A two-dimensional unit vector. In the operator parameter instantiation operation, the extracted normalized component c is set to the cosine value of the polar rotation angle θ generated by the current external context, i.e., cosθ, and the component s is set to the sine value, i.e., sinθ. For example, when c=0.866 and s=0.5, it is equivalent to the specific project event cycle causing a counterclockwise 30° polar rotation offset effect on the current behavioral polarity of the individual participants. According to the planar rotation matrix formula, a second-order rotation orthogonal transformation matrix is ​​instantiated and output with the vector group [c,-s] as the first row and the vector group [s,c] as the second row. This data dimensionality reduction mapping and matrix operator construction process reduces the parameter oscillation risk caused by traditional subjective weight assignment. Relying on data dimensionality reduction and algebraic geometric projection relationship, while preserving the environmental symbiotic statistical characteristics, a rotation control operator is generated to quantitatively determine how external events such as business fluctuations affect a single behavioral attribute.

[0044] S2, establish a mapping model from N preset capability dimensions to their corresponding measurement bases; combine the rotated two-dimensional probability polarity state vector with the behavior intensity value representing the degree of influence of the behavior, and project it onto the measurement base corresponding to each dimension based on the correlation between the behavior and the capability dimension, decomposing it into positive and negative behavior components belonging to the N capability dimensions; construct a behavior transition probability graph representing the temporal correlation of a single behavior, wherein the edge weights of the probability graph are determined based on the inner product of the rotated two-dimensional probability polarity state vectors of adjacent behavior nodes.

[0045] N preset capability dimensions are defined, including leadership, execution, innovation, and teamwork. A set of two-dimensional measurement bases is configured for each capability dimension in a two-dimensional real-valued polar state space. Each set of two-dimensional measurement bases includes one positive measurement basis vector and one negative measurement basis vector, and the positive and negative measurement basis vectors within the same capability dimension are orthogonal to each other. Different capability dimensions are distinguished by capability dimension name, measurement basis configuration, and associated weights. It is not required that all measurement basis vectors corresponding to all capability dimensions be orthogonal to each other or linearly independent in the same two-dimensional space. A mapping relationship is established between capability dimension names and their corresponding two-dimensional measurement basis vectors, forming a mapping model from capability dimensions to measurement basis vectors.

[0046] The execution time of a single behavior is obtained by reading system logs. This execution time is then divided by a preset global baseline time constant to calculate a scalar value representing the behavior intensity. The cosine similarity between the single behavior description text vector and the setting text vectors for each capability dimension is calculated, and this cosine similarity is non-negative to obtain the correlation between the behavior and the capability dimension. Each component of the rotated two-dimensional probability polarity state vector is squared to obtain non-negative positive and negative probability components. These positive and negative probability components are then scaled using scalar multiplication with the behavior intensity value and the correlation. The scaled two-dimensional state vector is then multiplied by the positive and negative measurement bases corresponding to the capability dimension in the mapping model. The resulting positive and negative projections represent the positive and negative behavior components of the single behavior under the corresponding capability dimension.

[0047] Using the individual actions of the talent to be profiled as graph nodes, a directed acyclic graph structure is constructed according to the chronological order of timestamps. For any two single action nodes that occur adjacently in time, the two nodes are connected as a directed edge. The two-dimensional probability polarity state vectors of the source node and the target node after rotation are extracted, and the inner product of these two real vectors is calculated. The inner product result is incremented by 1 and divided by 2 to obtain a non-negative mapping value. This non-negative mapping value is assigned to the weight attribute of the corresponding directed edge, representing the continuity and transition probability of the preceding and following actions in the positive and negative polarity states.

[0048] In an optional embodiment, the process involves establishing a mapping model from N preset capability dimensions to their corresponding measurement bases; combining the rotated two-dimensional probability polarity state vector with a behavior intensity value representing the degree of influence of the behavior; and projecting the behavior onto the measurement base corresponding to each dimension based on the correlation between the behavior and the capability dimensions to decompose and obtain positive and negative behavior components belonging to the N capability dimensions, including:

[0049] Extract the execution time of a single behavior, divide the execution time by a preset global reference time constant, and obtain the behavior intensity scalar value;

[0050] Calculate the product of the context rotation operator and the two-dimensional probabilistic polarity state vector before rotation to obtain the two-dimensional probabilistic polarity state vector after rotation;

[0051] Squaring each component of the rotated two-dimensional probability polarity state vector yields non-negative positive and negative probability components. Multiplying the behavior intensity scalar value with the positive and negative probability components yields the behavior vector carrying the intensity.

[0052] For each target dimension among the N capability dimensions, a two-dimensional positive measurement basis vector and a two-dimensional negative measurement basis vector are set for the dimension. The two-dimensional positive measurement basis vector and the two-dimensional negative measurement basis vector are non-negative unit vectors, and the association weight between the single behavior and the target dimension is obtained.

[0053] The dot product of the behavior vector carrying intensity and the two-dimensional positive measurement basis vector is calculated and multiplied by the associated weight, which is used as the positive behavior component of the single behavior in the target dimension.

[0054] The dot product of the behavior vector carrying intensity and the two-dimensional negative measurement basis vector is calculated and multiplied by the associated weight, which serves as the negative behavior component of the single behavior in the target dimension. The decomposition is completed by traversing N dimensions.

[0055] The system business logs for this single behavior are analyzed to extract the execution time. For example, if a code refactoring behavior takes 4 hours, this time is divided by the company's preset standard daily working hours, such as eight hours, as a global baseline time constant, to calculate the behavior intensity scalar value of 0.5. The previously extracted second-order context rotation operator matrix is ​​multiplied with the initial two-dimensional probability polarity state vector to complete the environment correction of the polarity phase. Each component of the rotated two-dimensional polarity vector is squared to obtain non-negative positive and negative probability components. The behavior intensity scalar value of 0.5 is multiplied by the positive and negative probability components respectively to obtain a behavior vector that reflects the carrying intensity of actual business input and effort. In the capability dimension mapping stage, a loop is started to traverse N preset capability dimensions, such as execution ability and code quality, and the two-dimensional positive measurement basis vector of the current target dimension, such as an absolutely orthogonal basis, is retrieved from the mapping model dictionary. And two-dimensional negative measurement basis vectors, such as absolutely orthogonal basis vectors. Simultaneously, the cosine similarity between the behavior category and the text set for the current target dimension is calculated, and the cosine similarity is non-negatively processed to obtain the association weight. For example, the association weight between code refactoring and the "code quality" dimension is 0.9. The inner product of the behavior vector carrying strength and the two-dimensional positive measurement basis vector is calculated, and the result is multiplied by the association weight of 0.9. The resulting scalar value is the positive behavior component of the single behavior in the "code quality" dimension, representing the positive contribution of the capability strength. Similarly, the inner product of the behavior vector carrying strength and the two-dimensional negative measurement basis vector is calculated and multiplied by the association weight to obtain the negative behavior component, representing the negative loss that may introduce potential defects. By traversing N target dimensions one by one through the above calculation process, the positive and negative features of a single behavior in the multi-dimensional capability space are orthogonally decoupled and decomposed.

[0056] In an optional embodiment, the construction of a behavior transition probability graph representing the temporal association of a single behavior, wherein the edge weights of the probability graph are determined based on the inner product of the two-dimensional probability polarity state vectors after rotation of adjacent behavior nodes, includes:

[0057] Using each extracted individual behavior as a node, construct the initial structure of the behavior transition probability graph;

[0058] Directed edges are established between adjacent source and target behavior nodes according to the chronological order of the behavior occurrence.

[0059] Obtain the rotated two-dimensional probability polarity state vector corresponding to the source behavior node and the rotated two-dimensional probability polarity state vector corresponding to the target behavior node;

[0060] Calculate the dot product of two vectors;

[0061] Adding 1 to the inner product and then dividing by 2 yields a non-negative mapping value;

[0062] The non-negative mapping value is used as the initial edge weight of the directed edge from the source behavior node to the target behavior node, thus completing the construction of the behavior transition probability graph.

[0063] All single behavior instances generated by the talent to be profiled within a time period are instantiated as graph nodes, and a blank directed acyclic graph structure is initialized. Based on the chronological order of the behavior records' timestamps, a sliding window mechanism is used to extract any pair of behavior nodes that occur in close proximity in time. The earlier occurrence is designated as the source behavior node, and the later occurrence as the target behavior node. A directed edge connection is established between the two, and this topological network represents the trajectory of the business behavior. To measure the polarity smoothness between behaviors, the rotated two-dimensional probability polarity state vectors of the source and target behavior nodes after context matrix correction are read from memory space, and the inner product of these two real vectors is calculated. Since the theoretical value of the inner product of normalized polarity vectors is distributed in the interval [-1, 1], where 1 represents that the polarity states are completely consistent and -1 represents that the polarity has changed abruptly or reversed, in order to meet the requirement of non-negative edge weights in the graph algorithm, a linear normalization mapping operation is performed. The calculated inner product result is uniformly added by 1 and then divided by 2, smoothing and compressing the result to the non-negative real number interval [0, 1]. This non-negative mapping value represents the probability consistency and situational continuity of two consecutive business behaviors during the polarity evolution process. This mapping value is assigned to the weight attribute of the corresponding directed graph edge as the initial edge weight from the source node to the target node, thus completing the construction of a behavior transition probability graph that combines temporal topological coherence and two-dimensional polarity evolution characteristics. The behavior transition probability graph is used to characterize the transition correlation strength between adjacent behaviors through edge weights in the range of 0 to 1. These edge weights are not limited to a probability distribution where the sum of all outgoing edge weights from the same source behavior node equals 1.

[0064] S3, set a time-series decay function and, based on the behavior transition probability map and decay function, perform time-series weighted aggregation on the positive and negative behavior components decomposed from all behaviors to generate positive and negative cumulative features for each of the N capability dimensions; based on the positive and negative cumulative features of the N capability dimensions, construct a talent profile of the candidate to be profiled with decoupled positive and negative behaviors in multiple dimensions.

[0065] An exponential decay model is chosen as the temporal decay function, and a decay coefficient based on a time interval is calculated, where the time interval is the non-negative time difference obtained by subtracting the behavior occurrence time from the target evaluation time. For each capability dimension, the directed edges in the behavior transition probability graph are traversed, and the initial edge weight of the directed edge is multiplied by the decay coefficient of the corresponding source behavior node to obtain the temporally adjusted edge weight; then, the temporally adjusted edge weights pointing to the same target behavior node are summed as the aggregation coefficient of that target behavior node. For the starting behavior node without an in-degree directed edge, the temporal decay function value corresponding to the starting behavior node is used as its aggregation coefficient, so that the starting behavior node within the evaluation period participates in the calculation of positive and negative cumulative features. All target behavior nodes in the graph are traversed, and the positive behavior component of each target behavior node in that dimension is multiplied by the corresponding aggregation coefficient and summed to obtain the positive cumulative feature of that dimension; similarly, the negative behavior component is multiplied by the corresponding aggregation coefficient and summed to generate the negative cumulative feature of that dimension.

[0066] The positive and negative cumulative features of N capability dimensions are integrated into a structured table, and a radar chart with two overlay layers is drawn. The first layer maps the positive cumulative feature values ​​of each dimension, and the second layer maps the negative cumulative feature values ​​of each dimension. The generated talent profile chart is exported in JPEG format and stored together with the cumulative feature data in JSON format in a MongoDB NoSQL database.

[0067] In an optional embodiment, the step of setting a temporal decay function and, based on the behavior transition probability map and the decay function, performing temporal weighted aggregation on the positive and negative behavior components decomposed from all behaviors to generate positive and negative cumulative features for each of the N capability dimensions, including:

[0068] The time-series decay function is set to be the power of the product of the negative preset decay coefficient of the natural constant and the time difference, where the time difference is the time difference between the occurrence time of the behavior node and the target evaluation time.

[0069] Traverse each directed edge in the behavior transition probability graph, multiply the initial edge weight corresponding to the directed edge by the temporal decay function value corresponding to the source behavior node, and obtain the temporally adjusted edge weight.

[0070] For each of the N capability dimensions, the sum of the time-adjusted edge weights corresponding to the directed edges pointing to each target behavior node is taken as the aggregation coefficient of the target behavior node. The positive behavior components of each target behavior node under the dimension are weighted and summed to obtain the positive cumulative feature of the dimension.

[0071] The negative behavior components of each target behavior node under the given dimension are weighted and summed to obtain the negative cumulative feature of the given dimension.

[0072] For behavior nodes that do not have incoming edges, the time decay function value corresponding to the behavior node is used as its aggregation coefficient.

[0073] Constructing an exponential time-decay function The time difference Δt is defined as the non-negative difference between the baseline timestamp of the system triggering the current talent assessment action and the timestamp generated by the behavior node, e.g., in days. When Δt is measured in days, a pre-defined attenuation coefficient λ has the reciprocal dimension of days, making λ⋅Δt a dimensionless quantity. λ can also be set to a constant, such as 0.01, restoring the business logic that "the older the behavior, the lower its reference value for the current ability evaluation." All directed edges in the constructed behavior transition probability graph are traversed. The temporal attenuation function value of the source behavior node corresponding to each directed edge is calculated, and this value is multiplied by the existing initial edge weight of the directed edge, outputting the temporally adjusted edge weight with time penalty characteristics. Based on this graph network, feature aggregation calculation is initiated for each independent dimension of the N ability dimensions. For each target behavior node in the graph, search for all in-degree directed edges pointing to that node, arithmetically sum the temporally adjusted edge weights of these edges, and set this sum as the aggregation coefficient of the target behavior node under the current local graph network and time span, i.e., the node influence weight. For starting behavior nodes without in-degree directed edges, use the temporal decay function value corresponding to the starting behavior node as its aggregation coefficient; or, treat the starting behavior node as having a virtual self-connected in-edge with an initial weight of 1, and multiply the virtual self-connected in-edge by the temporal decay function value corresponding to the starting behavior node as the aggregation coefficient of the starting behavior node. Multiply the previously decomposed positive behavior components of each target behavior node under the capability dimension with their respective aggregation coefficients and sum them to output the positive cumulative feature value of this dimension within the overall evaluation period; similarly, multiply the negative behavior components of each node with their corresponding aggregation coefficients and sum them to output the negative cumulative feature value of this dimension. Iterate through all N dimensions to complete the bipolar feature aggregation that takes into account both the spatial linkage and temporal decay constraints of the graph.

[0074] In an optional embodiment, the step of constructing a talent profile of the candidate to be profiled, based on the positive and negative cumulative features of the N ability dimensions and decoupling positive and negative behaviors across multiple dimensions, includes:

[0075] The positive cumulative features of the N capability dimensions are mapped to the coordinates of each vertex of a polygonal positive capability radar chart, and the first visualization layer is generated and rendered.

[0076] The negative cumulative features of the N capability dimensions are mapped to the coordinates of each vertex of a polygonal negative risk radar chart, and a second visualization layer is generated and rendered.

[0077] The first visualization layer and the second visualization layer are aligned in coordinate system and superimposed and merged to generate a dual-layer radar superimposed image, which is used as a talent profile output for decoupling positive and negative behaviors.

[0078] The feature data is visualized graphically using a front-end chart rendering engine. The canvas is initialized based on a polar coordinate system. The calculated positive cumulative feature values ​​of the N capability dimensions are mapped to their respective capability ray axes at equal central angles (360° / N) to form a set of vertex coordinates. A color scheme representing positive contributions is used, such as light blue fill with dark blue outline, connecting each vertex sequentially to generate and render a polygonal positive capability radar chart representing the strengths of the candidate's overall qualities. This is stored as the first visualization layer at the bottom of the canvas in memory. Reusing the same set of polar coordinate system origin and ray angle configurations, the corresponding negative cumulative feature values ​​of the N capability dimensions are extracted and mapped to another set of vertex coordinates at the same scale. A warning color scheme representing business risks or efficiency delays is used, such as semi-transparent red fill with dark red dashed outline, connecting to generate a polygonal negative risk radar chart, serving as the second visualization layer representing the candidate's behavioral shortcomings and management risks. Maintaining the absolute origin of the polar coordinates and the spatial alignment of the N attribute dimension axes, the first and second visualization layers are overlaid and rendered with alpha channel transparency along the Z-axis. A dual-layer radar overlay image is calculated and synthesized, simultaneously presenting outward positive contributions and embedded negative risks. This image is then exported as a JPEG image or a web-interactive JSON configuration package, outputting as a high-dimensional talent profile with decoupled positive and negative behaviors to the human resources dashboard.

[0079] In the second embodiment, the present invention also proposes a talent profile construction system based on behavioral sequence mining, comprising the following modules:

[0080] The extraction module is used to acquire multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extract single behavioral features; for each single behavior, the initial polarity of the single behavior is modeled as a two-dimensional probability polarity state vector of positive and negative superposition; a behavior-context co-occurrence matrix is ​​constructed based on the single behavior and related events, and two-dimensional context features are extracted to calculate the context rotation operator; the rotation operator is applied to perform a rotation operation on the two-dimensional probability polarity state vector to adjust the positive and negative probability distribution of the vector;

[0081] A construction module is used to establish a mapping model from N preset capability dimensions to their corresponding measurement bases; the rotated two-dimensional probability polarity state vector is combined with the behavior intensity value representing the degree of influence of the behavior, and based on the correlation between the behavior and the capability dimension, it is projected onto the measurement base corresponding to each dimension to decompose and obtain positive and negative behavior components belonging to the N capability dimensions; a behavior transition probability graph representing the temporal correlation of a single behavior is constructed, and the edge weights of the probability graph are determined based on the inner product of the rotated two-dimensional probability polarity state vectors of adjacent behavior nodes;

[0082] The generation module is used to set a temporal decay function and, based on the behavior transition probability map and decay function, perform temporal weighted aggregation on the positive and negative behavior components decomposed from all behaviors to generate positive and negative cumulative features for each of the N capability dimensions; based on the positive and negative cumulative features of the N capability dimensions, a talent profile of the candidate to be profiled, with decoupled positive and negative behaviors across multiple dimensions, is constructed.

[0083] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for constructing talent profiles based on behavioral sequence mining, characterized in that, Includes the following steps: Acquire multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extract single behavioral features; For each single action, the initial polarity of the single action is modeled as a two-dimensional probabilistic polarity state vector of positive and negative superposition states; an action-context co-occurrence matrix is ​​constructed based on the single action and associated events, and two-dimensional context features are extracted to calculate the context rotation operator; The rotation operator is applied to the two-dimensional probability polarity state vector to perform a rotation operation, thereby adjusting the positive and negative probability distributions of the vector; Establish a mapping model from N preset capability dimensions to their corresponding measurement bases; combine the rotated two-dimensional probability polarity state vector with the behavior intensity value representing the degree of influence of the behavior, and project it onto the measurement base corresponding to each dimension based on the correlation between the behavior and the capability dimension to decompose and obtain positive and negative behavior components belonging to the N capability dimensions; construct a behavior transition probability graph representing the temporal correlation of a single behavior, wherein the edge weights of the probability graph are determined based on the inner product of the rotated two-dimensional probability polarity state vectors of adjacent behavior nodes; A temporal decay function is set, and based on the behavior transition probability map and decay function, the positive and negative behavior components decomposed from all behaviors are time-weighted and aggregated to generate positive and negative cumulative features for each of the N capability dimensions. Based on the positive and negative cumulative features of the N capability dimensions, a talent profile is constructed that decouples the positive and negative behaviors of the talent to be profiled across multiple dimensions.

2. The method according to claim 1, characterized in that, The process of acquiring multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extracting single behavioral features includes: Extract attendance records, project delivery documents, and business flow approval data of the talent to be profiled from the company's internal database within a preset time period; The attendance record data is parsed into attendance behavior features by timestamp, and the project delivery document is converted into document submission behavior features using word frequency statistics rules. The approver field and approval time value are extracted from the business flow approval data as approval behavior features. The attendance behavior characteristics, document submission behavior characteristics, and approval behavior characteristics are uniformly structured into a single set of behavior characteristics that includes timestamps and behavior categories.

3. The method according to claim 1, characterized in that, For each individual action, the initial polarity of the individual action is modeled as a two-dimensional probabilistic polarity state vector of positive and negative superposition states, including: Extract the sentiment polarity label value of the single behavior, and obtain the positive polarity base probability and negative polarity base probability of the single behavior; The normalization factor is calculated as the square root of the sum of the positive polarity base probability and the negative polarity base probability. If the normalization factor is 0, then the normalization factor is assigned a preset small positive number. Construct a two-dimensional column vector, setting the first element as the square root of the positive polarity base probability divided by the normalization factor, and the second element as the square root of the negative polarity base probability divided by the normalization factor. Output a two-dimensional column vector that satisfies the L2 norm normalization condition, which serves as the two-dimensional probability polarity state vector of the positive and negative superposition states.

4. The method according to claim 2, characterized in that, The step of constructing a behavior-context co-occurrence matrix based on single behaviors and associated events, extracting two-dimensional context features, and calculating the context rotation operator includes: Collect all behavioral categories that appear within a specified event, and construct a behavior-to-context co-occurrence frequency matrix with behavioral categories as rows and event identifiers as columns; Singular value decomposition is performed on the co-occurrence frequency matrix of the behavior to the context, and the left singular vector corresponding to the maximum singular value is obtained as the context feature vector. The context feature vector is reduced to a two-dimensional space through linear mapping and then L2 normalized to obtain a two-dimensional unit vector containing the first and second components. The first component of the two-dimensional unit vector is set as the cosine value of the context rotation angle, and the second component is set as the sine value. A second-order rotation orthogonal transformation matrix is ​​generated with the cosine value and negative sine value as the first row and the sine value and cosine value as the second row, which serves as the context rotation operator.

5. The method according to claim 1 or 3, characterized in that, The process involves establishing a mapping model from N preset capability dimensions to their corresponding measurement bases; combining the rotated two-dimensional probability polarity state vector with the behavior intensity value representing the degree of influence of the behavior; and projecting the behavior onto the measurement base corresponding to each dimension based on the correlation between the behavior and the capability dimensions to decompose and obtain positive and negative behavior components belonging to the N capability dimensions, including: Extract the execution time of a single behavior, divide the execution time by a preset global reference time constant, and obtain the behavior intensity scalar value; Calculate the product of the context rotation operator and the two-dimensional probabilistic polarity state vector before rotation to obtain the two-dimensional probabilistic polarity state vector after rotation; Squaring each component of the rotated two-dimensional probability polarity state vector yields non-negative positive and negative probability components. Multiplying the behavior intensity scalar value with the positive and negative probability components yields the behavior vector carrying the intensity. For each target dimension among the N capability dimensions, a two-dimensional positive measurement basis vector and a two-dimensional negative measurement basis vector are set for the dimension. The two-dimensional positive measurement basis vector and the two-dimensional negative measurement basis vector are non-negative unit vectors, and the association weight between the single behavior and the target dimension is obtained. The dot product of the behavior vector carrying intensity and the two-dimensional positive measurement basis vector is calculated and multiplied by the associated weight, which is used as the positive behavior component of the single behavior in the target dimension. The dot product of the behavior vector carrying intensity and the two-dimensional negative measurement basis vector is calculated and multiplied by the associated weight, which serves as the negative behavior component of the single behavior in the target dimension. The decomposition is completed by traversing N dimensions.

6. The method according to claim 1, characterized in that, The construction of the behavior transition probability graph representing the temporal correlation of a single behavior, wherein the edge weights of the probability graph are determined based on the inner product of the two-dimensional probability polarity state vectors after rotation of adjacent behavior nodes, includes: Using each extracted individual behavior as a node, construct the initial structure of the behavior transition probability graph; Directed edges are established between adjacent source and target behavior nodes according to the chronological order of the behavior occurrence. Obtain the rotated two-dimensional probability polarity state vector corresponding to the source behavior node and the rotated two-dimensional probability polarity state vector corresponding to the target behavior node; Calculate the dot product of two vectors; Adding 1 to the inner product and then dividing by 2 yields a non-negative mapping value; The non-negative mapping value is used as the initial edge weight of the directed edge from the source behavior node to the target behavior node, thus completing the construction of the behavior transition probability graph.

7. The method according to claim 1, characterized in that, The set temporal decay function, based on the behavior transition probability map and the decay function, performs temporal weighted aggregation on the positive and negative behavior components decomposed from all behaviors, generating positive and negative cumulative features for each of the N capability dimensions, including: The time-series decay function is set to be the power of the product of the negative preset decay coefficient of the natural constant and the time difference, where the time difference is the time difference between the occurrence time of the behavior node and the target evaluation time. Traverse each directed edge in the behavior transition probability graph, multiply the initial edge weight corresponding to the directed edge by the temporal decay function value corresponding to the source behavior node, and obtain the temporally adjusted edge weight. For each of the N capability dimensions, the sum of the time-adjusted edge weights corresponding to the directed edges pointing to each target behavior node is taken as the aggregation coefficient of the target behavior node. The positive behavior components of each target behavior node under the dimension are weighted and summed to obtain the positive cumulative feature of the dimension. The negative behavior components of each target behavior node under the given dimension are weighted and summed to obtain the negative cumulative feature of the given dimension. For behavior nodes that do not have incoming edges, the time decay function value corresponding to the behavior node is used as its aggregation coefficient.

8. The method according to claim 1, characterized in that, The process of constructing a talent profile based on the positive and negative cumulative features of the N capability dimensions, decoupling the positive and negative behaviors of the candidate across multiple dimensions, includes: The positive cumulative features of the N capability dimensions are mapped to the coordinates of each vertex of a polygonal positive capability radar chart, and the first visualization layer is generated and rendered. The negative cumulative features of the N capability dimensions are mapped to the coordinates of each vertex of a polygonal negative risk radar chart, and a second visualization layer is generated and rendered. The first visualization layer and the second visualization layer are aligned in coordinate system and superimposed and merged to generate a dual-layer radar overlay image, which is used as a talent profile output for decoupling positive and negative behaviors.

9. A talent profile construction system based on behavioral sequence mining, characterized in that, Includes the following modules: The extraction module is used to acquire multi-source heterogeneous behavioral data of the person to be profiled within a preset time period and extract single behavioral features; For each single action, the initial polarity of the single action is modeled as a two-dimensional probabilistic polarity state vector of positive and negative superposition states; an action-context co-occurrence matrix is ​​constructed based on the single action and associated events, and two-dimensional context features are extracted to calculate the context rotation operator; The rotation operator is applied to the two-dimensional probability polarity state vector to perform a rotation operation, thereby adjusting the positive and negative probability distributions of the vector; A construction module is used to establish a mapping model from N preset capability dimensions to their corresponding measurement bases; the rotated two-dimensional probability polarity state vector is combined with the behavior intensity value representing the degree of influence of the behavior, and based on the correlation between the behavior and the capability dimension, it is projected onto the measurement base corresponding to each dimension to decompose and obtain positive and negative behavior components belonging to the N capability dimensions; a behavior transition probability graph representing the temporal correlation of a single behavior is constructed, and the edge weights of the probability graph are determined based on the inner product of the rotated two-dimensional probability polarity state vectors of adjacent behavior nodes; The generation module is used to set the temporal decay function and, based on the behavior transition probability map and decay function, perform temporal weighted aggregation on the positive and negative behavior components decomposed from all behaviors to generate positive and negative cumulative features for each of the N capability dimensions. Based on the positive and negative cumulative features of the N capability dimensions, a talent profile is constructed that decouples the positive and negative behaviors of the talent to be profiled across multiple dimensions.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-8.

Images