Intelligent talent evaluation method based on multi-dimensional data visualization

Abstract

The invention relates to the technical field of multi-dimensional data visualization, and discloses an intelligent talent evaluation method based on multi-dimensional data visualization, and the method comprises the steps: constructing a time sequence sensitive visualization engine, and collecting and processing the static attribute and dynamic behavior sequence data of an individual; mapping the high-dimensional behavior data into a low-dimensional dynamic trajectory through an improved nonlinear dimension reduction algorithm; realizing semantic alignment and fusion of static and dynamic data by utilizing a semantic embedding model based on an attention mechanism; a high-performance rendering engine is adopted to realize time sequence continuous visualization and abnormal inflection point interaction detection of the trajectory; and quantitative evaluation, prediction and traceability analysis are carried out on individual potential, performance and adaptability based on a mixed model of a gradient boosting tree and a long short-term memory network. According to the technical scheme, the dynamic evolution of the talent behavior mode can be visually presented, the abnormal inflection point can be accurately identified, and the timeliness, accuracy and decision support capability of talent evaluation are improved.

Description

Technical Field

[0001] This invention belongs to the field of multidimensional data visualization technology, specifically relating to a talent intelligent evaluation method based on multidimensional data visualization. Background Technology

[0002] As human resource management evolves towards data-driven and intelligent approaches, talent evaluation systems are gradually shifting from experience-based judgment to multi-dimensional quantitative analysis. Current talent assessment methods generally rely on static attribute data, such as structured information like education, job level, and certifications. While these reflect an individual's historical accumulation, they struggle to capture their dynamic behavioral characteristics and growth potential in actual work scenarios. Especially in rapidly changing organizational environments, temporal behavioral data such as an individual's collaboration patterns, task response pace, and knowledge update frequency contain crucial potential signals. Therefore, a comprehensive evaluation mechanism that integrates static attributes and dynamic behaviors is urgently needed to support forward-looking talent decisions.

[0003] The talent intelligent evaluation method based on multidimensional data visualization focuses on transforming the behavioral flow of individuals within an organizational system into perceptible and analyzable visual trajectories. This method aims to reveal the intrinsic laws governing the evolution of talent capabilities by constructing time-sensitive visual representations, thereby breaking through the dependence of traditional evaluation models on discrete indicators and achieving dynamic tracking and prediction of talent potential.

[0004] Existing talent evaluation systems generally simplify time-series behavioral data into statistical scalars (such as average response time and task completion rate), losing the temporal structure and contextual relationships within the behavioral sequence. This results in a significant lag in potential identification, often requiring months to reveal trend changes. Furthermore, static attributes and dynamic behaviors belong to heterogeneous data spaces, lacking a unified embedding and mapping mechanism, making semantic alignment difficult at the visualization level. In addition, mainstream visualization tools mostly use static charts or simple animations, failing to construct dynamic trajectories with temporal continuity, making it difficult for managers to intuitively perceive the evolution of individual behavioral patterns and abnormal inflection points. Therefore, there is an urgent need for an intelligent evaluation and visualization solution that can deeply integrate multidimensional heterogeneous data, preserve temporal details, and support dynamic trajectory mapping. Summary of the Invention

[0005] The purpose of this invention is to provide a talent intelligent evaluation method based on multidimensional data visualization, which can effectively solve the problems in the background technology mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A talent intelligent evaluation method based on multidimensional data visualization includes the following specific steps: Step 1: Build a time-sensitive visualization engine to map behavioral flow into dynamic trajectories. This is achieved by collecting multi-source heterogeneous data of individuals in the organizational environment, including static attribute data and dynamic behavioral sequence data, and cleaning, aligning, and normalizing the data to form a standardized input data stream. Step 2: Design and implement a dynamic trajectory mapping algorithm. Based on the standardized input data stream processed in Step 1, extract the temporal features and contextual information, and use nonlinear dimensionality reduction technology to project the high-dimensional behavioral data onto a low-dimensional visualization space to generate a dynamic trajectory that can reflect the evolution of individual behavioral patterns. Step 3: Develop a semantic alignment and fusion mechanism. To address the issue that static attribute data and dynamic behavior data belong to different data spaces, a unified semantic embedding model is constructed to map the two types of data to the same semantic space, thereby achieving alignment and fusion at the feature level and providing a consistent semantic foundation for visualization. Step 4: Achieve visualization and interaction of dynamic trajectory. Based on the dynamic trajectory generated in Step 2 and the semantic alignment results achieved in Step 3, use temporally continuous dynamic visualization technology to render the individual's behavioral trajectory and provide interactive functions, allowing users to explore the evolution process and abnormal inflection points of the trajectory from different perspectives and granularities. Step 5: Establish an intelligent evaluation and prediction model. Based on the dynamic trajectory and semantic alignment features, construct a machine learning model to quantitatively evaluate and predict an individual's potential, performance, and adaptability, generate a comprehensive evaluation report, and support source analysis of the evaluation results.

[0007] Preferably, the multi-source heterogeneous data collection in step 1 includes automatically acquiring static attribute data such as basic information, educational background, and work experience of individuals through internal system interfaces of the organization, as well as capturing dynamic behavioral sequence data such as task response time, collaboration frequency, knowledge contribution, and communication patterns of individuals in real time through behavior logs, task management systems, and collaboration platforms. The data collection frequency can be configured from once per minute to once per day according to the evaluation needs.

[0008] Preferably, the data cleaning, alignment, and normalization process in step 1 specifically includes: filling missing values ​​in the collected raw data; using a time-series-based linear interpolation method to handle missing values ​​in continuous variables, and using the mode to fill in missing values ​​for categorical variables; standardizing and converting inconsistent data formats and units to ensure that all numerical data are unified to the same dimension; and aligning dynamic behavioral sequence data and static attribute data with timestamps based on individual unique identifiers to form a multidimensional data stream with a consistent time base.

[0009] Preferably, the dynamic trajectory mapping algorithm in step 2 adopts an improved t-distribution random neighborhood embedding technique. This technique enhances the capture of time dependencies in the behavior sequence by introducing a temporal weight factor. The temporal weight factor is dynamically adjusted according to the time interval and contextual importance of the behavior events. The low-dimensional space dimension after projection is set to 2 or 3 dimensions to facilitate visualization.

[0010] Preferably, the temporal feature extraction in step 2 includes calculating the autocorrelation function and partial autocorrelation function of the behavior sequence to identify periodic patterns, extracting moving average and exponential smoothing values ​​to capture trend changes, and using a hidden Markov model to identify the probability of behavior state transitions. These features together constitute a high-dimensional behavior feature vector, which serves as the input to the dimensionality reduction algorithm.

[0011] Preferably, in step 3, the semantic embedding model adopts a deep neural network architecture based on an attention mechanism. This network simultaneously receives static attribute features and dynamic behavioral features as input, and learns the shared representation of the two types of features through multi-layer nonlinear transformation. The attention mechanism is used to dynamically weight the contribution of different feature dimensions to the final semantic representation, ensuring that key information is preserved.

[0012] Preferably, the alignment and fusion at the feature level in step 3 is achieved by calculating the cosine similarity between static attributes and dynamic behaviors in the shared semantic space. The similarity threshold is set to 0.7. Features with a similarity higher than this threshold are considered to have achieved effective alignment and participate in subsequent fusion calculations. The dimension of the fused feature vector remains at 128.

[0013] Preferably, the time-series continuous dynamic visualization technology in step 4 adopts a WebGL-based rendering engine. This engine can smoothly transition the position and shape changes of trajectory points at a refresh rate of 60 frames per second, while providing interactive functions such as timeline dragging, hovering over trajectory points to view detailed information, and multi-trajectory comparison views, to help users intuitively perceive the evolution of individual behavior patterns.

[0014] Preferably, the abnormal inflection point detection in step 4 is achieved by analyzing the rate of change of curvature and the sudden change of velocity vector of the dynamic trajectory. When the local rate of change of curvature of the trajectory exceeds a preset threshold of 2.5 or the direction of velocity vector changes by more than 90 degrees within three consecutive time points, the system automatically marks the position as a potential abnormal inflection point and highlights it in a color in the visualization interface.

[0015] Preferably, in step 5, the machine learning model adopts a hybrid architecture of gradient boosting decision tree and long short-term memory neural network. The gradient boosting decision tree is used to process static attributes and fused semantic features, while the long short-term memory neural network is used to model the temporal dependencies of dynamic trajectories. The outputs of the two sub-models are integrated through a fully connected layer to finally output the individual's scores in three dimensions: potential, performance, and adaptability.

[0016] Preferably, the specific process of quantitative assessment and prediction in step 5 includes: inputting the individual's dynamic trajectory sequence and semantic alignment features into a trained hybrid model, and the model outputs potential score, performance score and adaptability score, respectively. Each score range is standardized to 0 to 100 points and calibrated based on historical data to ensure that the score distribution conforms to the actual talent distribution within the organization.

[0017] Preferably, in step 5, the comprehensive evaluation report generation module automatically integrates information such as model output scores, key behavioral events, and trajectory anomalies to generate a structured evaluation report. The report content includes a score summary, strengths analysis, improvement suggestions, and development forecasts for the next three months. The report format supports both PDF and interactive web page formats.

[0018] Preferably, it also includes a feedback learning mechanism, which collects user feedback on the evaluation report and subsequent actual talent development data, and uses this information to regularly update the machine learning model in step 5 to achieve online optimization of model parameters and continuously improve the evaluation accuracy. The model update cycle can be configured to be once a month or once a quarter.

[0019] Preferably, the method is deployed on a cloud service platform, which processes large-scale concurrent requests through a distributed computing framework. The average response time for a single evaluation task is less than 5 seconds. The platform supports the simultaneous management of evaluation data for more than 100,000 individuals. The data storage adopts a hybrid architecture of columnar database and time-series database to ensure efficient data read, write and query performance.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: By constructing a time-sensitive visualization engine and mapping behavioral flows to dynamic trajectories, we have achieved deep integration and intuitive presentation of multidimensional talent data. This effectively preserves the temporal structure and contextual relationships in the behavioral sequence, enabling managers to capture the evolution trend and abnormal inflection points of individual behavioral patterns in a timely manner. This improves the timeliness and accuracy of talent potential identification and provides reliable data support for the organization's forward-looking talent decision-making. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the talent intelligent evaluation method based on multidimensional data visualization proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the dynamic trajectory mapping and time-sensitive visualization engine in this invention; Figure 3 This is a logical flowchart of the semantic alignment and fusion mechanism for static attributes and dynamic behavior data in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of dynamic trajectory visualization and interactive analysis in this invention; Figure 5 This is a logical flowchart of the intelligent evaluation, prediction, and feedback learning mechanism based on a hybrid model in this invention. Detailed Implementation

[0022] This invention proposes to construct a time-sensitive visualization engine to map behavioral flows into dynamic trajectories and develop a semantic alignment and fusion mechanism to achieve deep integration and intuitive presentation of multidimensional talent data, thereby improving the timeliness and accuracy of talent potential identification. It is then applied to a talent intelligent evaluation method based on multidimensional data visualization.

[0023] refer to Figure 1 The overall technical architecture of this invention comprises five core functional modules: a data acquisition and preprocessing module 101, a dynamic trajectory mapping module 102, a semantic alignment and fusion module 103, a visualization and interaction module 104, and an intelligent evaluation and prediction module 105. These modules work collaboratively to form a closed-loop data processing and decision support workflow. Specifically, the data acquisition and preprocessing module 101 is responsible for acquiring and standardizing input data; the dynamic trajectory mapping module 102 generates low-dimensional trajectories reflecting the evolution of behavioral patterns; the semantic alignment and fusion module 103 unifies the semantic space of heterogeneous data; the visualization and interaction module 104 provides an intuitive exploration interface; and the intelligent evaluation and prediction module 105 outputs quantitative evaluation results and supports feedback learning.

[0024] In the aforementioned talent intelligent evaluation method based on multidimensional data visualization, step 1 involves constructing a time-sensitive visualization engine to map behavioral flows into dynamic trajectories. This is achieved by collecting multi-source heterogeneous data of individuals within the organizational environment, including static attribute data and dynamic behavioral sequence data. The data is then cleaned, aligned, and normalized to form a standardized input data stream. Specifically, step 1 involves the automatic acquisition of static attribute data such as basic information, educational background, and work experience of individuals through internal organizational system interfaces. It also involves the real-time capture of dynamic behavioral sequence data such as task response time, collaboration frequency, knowledge contribution, and communication patterns of individuals through behavioral logs, task management systems, and collaboration platforms. The data collection frequency can be configured from once per minute to once per day, depending on the evaluation requirements.

[0025] The data collection process is executed by a distributed data broker, deployed behind the API gateway of various business systems within the organization's intranet. It uses the OAuth 2.0 protocol for secure authentication to ensure controlled data access permissions. For static attribute data, collected fields include, but are not limited to, name, employee ID, department, job level, education, major, start date, and historical performance rating, in JSON or XML format. For dynamic behavioral sequence data, collected event types include task creation, task assignment, task completion, document editing, meeting participation, instant message sending, and code submission. Each event record includes metadata such as event type, timestamp, target, result, and a list of associated personnel.

[0026] The data cleaning, alignment, and normalization processes in step 1 specifically include: filling missing values ​​in the collected raw data; using a time-series-based linear interpolation method to handle missing values ​​in continuous variables, and using the mode to fill in missing values ​​for categorical variables; standardizing inconsistent data formats and units to ensure that all numerical data are unified to the same dimension; and aligning dynamic behavioral sequence data and static attribute data with timestamps based on unique individual identifiers to form a multidimensional data stream with a consistent time base.

[0027] The data cleaning phase begins with format validation, removing invalid records that do not conform to the predefined schema. Subsequently, for continuous variables such as task response time (in seconds) and daily active time (in minutes), if a time point is missing, linear interpolation is performed using values ​​from adjacent valid time points. For categorical variables such as department name and project role, the category with the highest historical frequency is used as the filler value. Data alignment uses individual unique identifiers (e.g., employee ID) as keys, broadcasting static attribute data to all corresponding dynamic behavior time points, forming composite records where each time point contains complete static attributes and dynamic behavior features. Normalization uses the Min-Max scaling method, mapping all numerical features to the [0, 1] interval, with the formula: ; in, The original value, and These are the minimum and maximum values ​​of the feature in the training set, respectively. The processed, standardized input data stream is stored in a distributed file system in Parquet columnar format for efficient subsequent retrieval.

[0028] In the aforementioned talent intelligent evaluation method based on multidimensional data visualization, step 2 involves designing and implementing a dynamic trajectory mapping algorithm. Based on the standardized input data stream processed in step 1, it extracts temporal features and contextual information, and uses nonlinear dimensionality reduction techniques to project high-dimensional behavioral data into a low-dimensional visualization space, generating a dynamic trajectory that reflects the evolution of individual behavioral patterns. Specifically, step 2's temporal feature extraction includes calculating the autocorrelation function and partial autocorrelation function of the behavioral sequence to identify periodic patterns, extracting moving averages and exponential smoothing values ​​to capture trend changes, and using a hidden Markov model to identify behavioral state transition probabilities. These features collectively constitute a high-dimensional behavioral feature vector, which serves as the input to the dimensionality reduction algorithm. Temporal feature extraction is performed within a sliding time window, with a window length of 7 days and a step size of 1 day. For task completion frequency sequences... Calculate its lag Autocorrelation coefficient of order : ; in, The mean within the window. The window length is specified. The partial autocorrelation function is used to eliminate the influence of intermediate lag terms and identify direct periodicity. A 7-day simple moving average with an exponential smoothing factor is used. The value is set to 0.3. The Hidden Markov Model is trained to identify the behavioral states of an individual (such as "high-productivity period," "cooperative period," "learning period," and "inefficient period"). The state transition probability matrix and observation probability distribution are learned from historical data using the Baum-Welch algorithm. Finally, a high-dimensional behavioral feature vector of 64 dimensions is generated for each time window, containing statistical features, temporal features, and state probabilities.

[0029] In step 2, the dynamic trajectory mapping algorithm employs an improved t-distributed random neighborhood embedding technique. This technique enhances the capture of temporal dependencies in behavioral sequences by introducing a temporal weight factor. This temporal weight factor is dynamically adjusted based on the time interval and contextual importance of behavioral events. The projected low-dimensional space is set to 2 or 3 dimensions for easier visualization. The standard t-SNE algorithm only considers the Euclidean distance between data points, ignoring temporal order. This invention introduces temporal weights into its similarity calculation. Corrected conditional probability : ; in, = , and They are time points respectively and timestamp, The attenuation coefficient is set to 0.1. For adjacent time points... The weights are close to 1; for time points that are far apart, the weights decay exponentially. Furthermore, contextual importance is weighted by the importance score of the event type; for example, the weight of a "critical task completed" event is higher than that of "normal message sent". The dimensionality reduction target space is fixed at 2 dimensions to facilitate trajectory rendering on a flat screen. The algorithm iterations are set to 1000, the perplexity to 30, and the learning rate to 200.

[0030] In the aforementioned talent intelligent evaluation method based on multidimensional data visualization, step 3 involves developing a semantic alignment and fusion mechanism. Addressing the issue of static attribute data and dynamic behavior data belonging to heterogeneous data spaces, a unified semantic embedding model is constructed to map the two types of data to the same semantic space, achieving feature-level alignment and fusion, and providing a consistent semantic foundation for visualization. Specifically, the semantic embedding model in step 3 adopts a deep neural network architecture based on an attention mechanism. This network simultaneously receives static attribute features and dynamic behavior features as input, learns the shared representation of the two types of features through multi-layer nonlinear transformations, and uses the attention mechanism to dynamically weight the contribution of different feature dimensions to the final semantic representation, ensuring that key information is preserved. This neural network contains two independent encoder branches: a static encoder and a dynamic encoder. The static encoder consists of three fully connected layers, each with 128 neurons, using ReLU activation. The dynamic encoder consists of one LSTM layer (64 hidden units) followed by two fully connected layers (128 neurons each). The outputs of the two encoders are concatenated and input into a shared attention layer. Attention weights are used to further refine the model. Calculated using a small neural network: ; ; in, and These are the i-th dimension codes for the static and dynamic features, respectively. and These are learnable parameters. The final semantic embedding vector is a weighted sum. The dimension is 256.

[0031] In step 3, feature alignment and fusion are achieved by calculating the cosine similarity between static attributes and dynamic behaviors in the shared semantic space. A similarity threshold of 0.7 is set; features exceeding this threshold are considered effectively aligned and participate in subsequent fusion calculations. The fused feature vector maintains a dimension of 128. The cosine similarity calculation formula is: ; in, Embed vectors for static attributes. This is a dynamic behavior embedding vector. The system iterates through the historical data of all individuals and calculates the average similarity between their static and dynamic embeddings. If the similarity is below 0.7, a model retraining process is triggered, adjusting the encoder parameters until the alignment requirements are met. The fusion calculation uses a gating mechanism to generate a 128-dimensional fusion feature vector. : ; in, For the Sigmoid function, For the gated weight matrix, This is an element-wise multiplication. The fused vector serves as the core input to the subsequent evaluation model.

[0032] In the aforementioned talent intelligent evaluation method based on multidimensional data visualization, step 4 visualizes and interacts with dynamic trajectories. Based on the dynamic trajectory generated in step 2 and the semantic alignment results achieved in step 3, a temporally continuous dynamic visualization technique is used to render the individual's behavioral trajectory and provides interactive functions, allowing users to explore the trajectory's evolution and abnormal inflection points from different perspectives and granularities. Specifically, the temporally continuous dynamic visualization technique in step 4 uses a WebGL-based rendering engine. This engine can smoothly transition the position and shape changes of trajectory points at a refresh rate of 60 frames per second, while providing interactive functions such as timeline dragging, hovering over trajectory points to view detailed information, and multi-trajectory comparison views, helping users intuitively perceive the evolution of individual behavioral patterns. (Reference) Figure 4 The visualization system's front end uses the Three.js library to build a 3D scene. Each trajectory point corresponds to a behavioral state within a time slice, and its position is determined by the 2D coordinates generated in step 2. The Z-axis height can be mapped to performance score or activity level. The color of the trajectory point encodes its corresponding behavioral state (e.g., blue represents "high productivity period," green represents "collaboration period"), and its size encodes the number of tasks completed. The timeline control allows users to specify any start and end time, and the system redraws the trajectory within that time period in real time. Hovering over any trajectory point pops up an information box displaying a detailed list of behavioral events, a snapshot of static attributes, and model confidence for that time point.

[0033] In step 4, abnormal inflection point detection is achieved by analyzing the rate of change of curvature and abrupt changes in velocity vector of the dynamic trajectory. When the local rate of change of curvature of the trajectory exceeds a preset threshold of 2.5 or the velocity vector direction changes by more than 90 degrees within three consecutive time points, the system automatically marks this location as a potential abnormal inflection point and highlights it with a bright color in the visualization interface. (Track curvature) Approximate calculation at discrete points using the three-point method: ; in Let be the area of ​​the triangle formed by the three points. The rate of change of curvature is defined as... Velocity vector Its direction angle .like If a point is 90 degrees away from the center, and all three points meet this condition, it is considered a directional abrupt change. All detected abnormal inflection points are marked with flashing red circles on the trajectory, and an event summary is generated in the sidebar, such as "Behavioral pattern mutation: Collaboration frequency drops by 50%".

[0034] In the aforementioned talent intelligent evaluation method based on multidimensional data visualization, step 5 involves establishing an intelligent evaluation and prediction model. Based on the dynamic trajectory and semantic alignment features, a machine learning model is constructed to quantitatively assess and predict an individual's potential, performance, and adaptability, generating a comprehensive evaluation report and supporting source analysis of the evaluation results. Specifically, in step 5, the machine learning model adopts a hybrid architecture of gradient boosting decision trees and long short-term memory neural networks. The gradient boosting decision tree is used to process static attributes and fused semantic features, while the long short-term memory neural network is used to model the temporal dependencies of the dynamic trajectory. The outputs of the two sub-models are integrated through a fully connected layer, ultimately outputting the individual's scores in the three dimensions of potential, performance, and adaptability.

[0035] refer to Figure 5 The Gradient Boosting Decision Tree (GBDT) sub-model takes a 128-dimensional fused feature vector as input, with 500 trees and a maximum depth of 6. The LSTM sub-model takes a dynamic trajectory coordinate sequence from the past 30 days (one point per day, for a total of 30 2D points), with 128 hidden layer units. The outputs of the two sub-models (each 3-dimensional) are concatenated and fed into a 2-layer fully connected network (64 neurons per layer), ultimately outputting 3 scalar scores.

[0036] The specific process of quantitative assessment and prediction in step 5 includes: inputting the individual's dynamic trajectory sequence and semantic alignment features into a trained hybrid model; the model outputs potential score, performance score, and adaptability score, respectively, with each score range standardized from 0 to 100, and calibrated based on historical data to ensure that the score distribution conforms to the actual talent distribution within the organization. The calibration process uses the Platt Scaling method to convert the original model output... Mapping to probability : ; A and B were obtained by fitting logistic regression on the validation set. Subsequently, the probability distribution was linearly transformed to a score of 0-100, ensuring that the top 10% of individuals within the organization scored above 90 points, and the bottom 10% scored below 10 points. In step 5, the comprehensive evaluation report generation module automatically integrates information such as model output scores, key behavioral events, and trajectory anomalies to generate a structured evaluation report. The report includes a score summary, strengths analysis, improvement suggestions, and a development forecast for the next three months. The report supports both PDF and interactive web page formats.

[0037] The report generation engine is based on template filling technology. The strengths analysis section uses behavioral characteristics of high-contribution individuals, while improvement suggestions are based on behavioral gaps with high-performing individuals. Development prediction uses an LSTM model to extrapolate trajectory coordinates for the next 30, 60, and 90 days, and estimates score change trends accordingly.

[0038] In addition, a feedback learning mechanism is included. This mechanism collects user feedback on the evaluation reports and subsequent actual talent development data. This information is used to periodically update the machine learning model in step 5, enabling online optimization of model parameters and continuously improving evaluation accuracy. The model update cycle can be configured to be once a month or once a quarter. Feedback information includes user ratings of the report's accuracy (1-5 stars) and text comments. Actual development data includes subsequent promotions, job transfers, and resignations. The system uses this data as new training samples, employing an incremental learning strategy to fine-tune the model, avoiding the computational overhead of full retraining.

[0039] The entire methodology is deployed on a cloud service platform, handling large-scale concurrent requests through a distributed computing framework. The average response time for a single evaluation task is less than 5 seconds, and the platform supports managing evaluation data for over 100,000 individuals simultaneously. Data storage employs a hybrid architecture of columnar and time-series databases to ensure efficient data read / write and query performance. The columnar database (such as Apache Parquet) stores static attributes and aggregate features, while the time-series database stores the raw behavioral event stream; the two are linked by individual IDs.

[0040] To illustrate the effects of this invention more specifically, an application scenario is constructed: A large technology company with 5,000 R&D personnel needs to conduct potential assessments for junior engineers to develop training plans. The system automatically collects behavioral data from the previous day every morning, including Git code commits, Jira task completions, Confluence document edits, and Slack message interaction frequency. Step 1 processes this data to form a standardized data stream. Step 2 generates a 2D dynamic trajectory for each engineer, with the horizontal axis representing task output efficiency and the vertical axis representing collaboration breadth. Step 3 aligns and integrates static attributes such as education, major, and years of service with dynamic behaviors. Step 4 renders the trajectory in the management backend. HR discovered that Engineer A's trajectory showed a significant inflection point on the 45th day of employment—the curvature change rate reached 3.1, and the trajectory suddenly shifted from a high-collaboration zone to an isolated zone. Upon inspection, it was found that the engineer did not participate in any team meetings that day, and code commits decreased by 80%. Step 5's model showed that their potential score dropped from 75 to 60. After HR intervention and communication, it was learned that the engineer was experiencing low mood due to family reasons, and after counseling, their trajectory returned to normal. Three months later, the engineer successfully led a key project, validating the model's early warning value.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A talent intelligent evaluation method based on multidimensional data visualization, characterized in that: The specific steps include the following: Step 1: Build a time-sensitive visualization engine to map behavioral flow into dynamic trajectories. This is achieved by collecting multi-source heterogeneous data of individuals in the organizational environment, including static attribute data and dynamic behavioral sequence data, and cleaning, aligning, and normalizing the data to form a standardized input data stream. Step 2: Design and implement a dynamic trajectory mapping algorithm. Based on the standardized input data stream processed in Step 1, extract the temporal features and contextual information, and use nonlinear dimensionality reduction technology to project the high-dimensional behavioral data onto a low-dimensional visualization space to generate a dynamic trajectory that can reflect the evolution of individual behavioral patterns. Step 3: Develop a semantic alignment and fusion mechanism. To address the issue that static attribute data and dynamic behavior data belong to different data spaces, a unified semantic embedding model is constructed to map the two types of data to the same semantic space, thereby achieving alignment and fusion at the feature level and providing a consistent semantic foundation for visualization. Step 4: Achieve visualization and interaction of dynamic trajectory. Based on the dynamic trajectory generated in Step 2 and the semantic alignment results achieved in Step 3, use temporally continuous dynamic visualization technology to render the individual's behavioral trajectory and provide interactive functions, allowing users to explore the evolution process and abnormal inflection points of the trajectory from different perspectives and granularities. Step 5: Establish an intelligent evaluation and prediction model. Based on the dynamic trajectory and semantic alignment features, construct a machine learning model to quantitatively evaluate and predict an individual's potential, performance, and adaptability, generate a comprehensive evaluation report, and support source analysis of the evaluation results.

2. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: The multi-source heterogeneous data collection in step 1 includes automatically acquiring static attribute data such as basic information, educational background, and work experience of individuals through internal system interfaces of the organization, as well as capturing dynamic behavioral sequence data such as task response time, collaboration frequency, knowledge contribution, and communication patterns of individuals in real time through behavior logs, task management systems, and collaboration platforms. The data collection frequency is configured to be once per minute to once per day.

3. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: The data cleaning, alignment, and normalization processes in step 1 specifically include: filling missing values ​​in the collected raw data; using a time-series-based linear interpolation method to handle missing values ​​in continuous variables, and using the mode to fill in missing values ​​for categorical variables; standardizing inconsistent data formats and units to ensure that all numerical data are unified to the same dimension; and aligning dynamic behavioral sequence data and static attribute data with timestamps based on unique individual identifiers to form a multidimensional data stream with a consistent time base.

4. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: In step 2, the dynamic trajectory mapping algorithm adopts an improved t-distribution random neighborhood embedding technique. This technique enhances the capture of time dependencies in the behavior sequence by introducing a temporal weight factor. The temporal weight factor is dynamically adjusted according to the time interval and contextual importance of the behavior events, and the low-dimensional space dimension after projection is set to 2 or 3.

5. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: Step 2, the extraction of time-series features, includes calculating the autocorrelation function and partial autocorrelation function of the behavior sequence to identify periodic patterns, extracting moving average and exponential smoothing values ​​to capture trend changes, and using a hidden Markov model to identify the probability of behavior state transitions. These features together constitute a high-dimensional behavior feature vector, which serves as the input to the dimensionality reduction algorithm.

6. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: In step 3, the semantic embedding model adopts a deep neural network architecture based on the attention mechanism. This network simultaneously receives static attribute features and dynamic behavioral features as input, and learns the shared representation of the two types of features through multi-layer nonlinear transformation. The attention mechanism is used to dynamically weight the contribution of different feature dimensions to the final semantic representation.

7. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: In step 3, the alignment and fusion at the feature level are achieved by calculating the cosine similarity between static attributes and dynamic behaviors in the shared semantic space. The similarity threshold is set to 0.

7. Feature pairs with similarity values ​​higher than this threshold participate in subsequent fusion calculations. The fusion feature vector dimension remains at 128 dimensions.

8. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: The time-series continuous dynamic visualization technology in step 4 uses a WebGL-based rendering engine. This engine can smoothly transition the position and shape changes of trajectory points at a refresh rate of 60 frames per second, and provides interactive functions such as dragging the timeline, hovering over trajectory points to view detailed information, and comparing multiple trajectories.

9. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: In step 4, abnormal inflection point detection is achieved by analyzing the rate of curvature change and velocity vector mutation of the dynamic trajectory. When the local rate of curvature change of the trajectory exceeds 2.5 or the velocity vector direction changes by more than 90 degrees within three consecutive time points, the system automatically marks the position as a potential abnormal inflection point and highlights it in a color in the visualization interface.

10. The talent intelligent evaluation method based on multidimensional data visualization according to claim 1, characterized in that: In step 5, the machine learning model adopts a hybrid architecture of gradient boosting decision tree and long short-term memory neural network. Gradient boosting decision tree is used to process static attributes and fused semantic features, while long short-term memory neural network is used to model the temporal dependency of dynamic trajectory. The outputs of the two sub-models are integrated through a fully connected layer, and finally the individual's score is output in three dimensions: potential, performance and adaptability. Each score range is standardized to 0 to 100.

Images