A method and system for behavioral assessment based on machine learning in recruitment applications

Abstract

The present invention discloses a method and system for behavioral assessment based on machine learning in recruitment applications, which specifically relates to the field of machine learning technology, and comprises the following steps: generating an assessment data set by collecting and preprocessing candidate assessment behavior data; calculating an assessment task complexity index and an assessment behavior regularity index based on the data set to quantify the complexity and regularity of candidate behavior characteristics; determining the most suitable machine learning strategy through fuzzy reasoning; constructing and training a machine learning model according to the strategy, the trained model outputs the candidate's assessment score, and all scores are summarized and sorted to generate an assessment score form; the present invention can dynamically match the model depth to ensure resource utilization efficiency when a recruitment company makes optimization requirements for a pre-provided general model, so that a shallow model can quickly complete candidate scoring in a simple scenario, and a deep model can accurately extract features in a complex scenario.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and more specifically, to a method and system for machine learning-based behavioral assessment in recruitment applications. Background Art

[0002] In the traditional recruitment process, evaluators often use subjective judgment to evaluate the suitability of candidates. However, this method is easily affected by the personal preferences and emotional factors of the evaluators. The behavioral evaluation method based on machine learning collects the behavioral data of candidates (such as language, reaction, response mode, etc. during the evaluation process) and analyzes these data with machine learning algorithms. It can quantitatively evaluate the candidate's work attitude, emotional intelligence, communication skills and other characteristics, reduce human bias, and improve the accuracy and fairness of recruitment.

[0003] Compared with traditional recruitment methods, the behavioral assessment method based on machine learning has higher objectivity and accuracy. The machine learning model can learn and optimize based on a large amount of data, which can help companies quickly screen out the most suitable personnel from a large number of candidates. However, the differences in the evaluation content that different companies focus on will lead to different scales of machine learning models. If a unified machine learning model construction strategy is adopted, it is easy to have problems such as overperformance, long computing time, and increased computing power costs. How to determine the machine learning model suitable for recruitment is a problem. Therefore, a method and system for behavioral assessment based on machine learning in recruitment applications are proposed here to solve the above problems. Summary of the invention

[0004] To achieve the above object, the present invention provides the following technical solutions: A method for behavioral assessment based on machine learning in recruitment applications, comprising the following steps: Collect behavioral data from the candidate evaluation process and preprocess it to obtain a preprocessed evaluation data set; Based on the evaluation data set, the complexity analysis of evaluation tasks and the regularity analysis of evaluation behaviors were performed to obtain the complexity index of evaluation tasks and the regularity index of evaluation behaviors respectively. Based on the evaluation task complexity index and the evaluation behavior regularity index, the machine learning strategy is determined according to the results of fuzzy reasoning; According to the determined machine learning strategy, a machine learning model is constructed and trained, and the trained machine learning model is applied to the recruitment work of the recruiting unit. The output result of the machine learning model is the candidate evaluation score. Finally, all candidate evaluation scores are summarized and sorted to obtain an evaluation score form.

[0005] In a preferred embodiment, the logic for obtaining the complexity index of the evaluation task is: The complexity of language features is calculated by the complexity of grammatical analysis and the amplitude of emotional changes. The calculation formula is: ; represents the preset language features at time t, is the mean of the preset language features, is the standard deviation of the preset language feature, N is the total number of time points in the evaluation, is the preset non-zero adjustment factor of one, Indicates the complexity of language features; Entropy is used to measure the uncertainty and diversity of emotional changes, and the complexity of facial expression features is obtained through the following calculation formula: ; M is the number of facial expression categories, is the probability of facial expression category i appearing in the evaluation, Indicates the complexity of facial expression features; The complexity of body language features is calculated through the frequency fluctuation and position change of body language. The calculation formula is: ; N is the total number of time points in the evaluation, represents the preset limb features at the tth moment, represents the preset limb features at the t-1th moment, is the preset non-zero adjustment factor of two, Indicates the complexity of body language features; The standard deviation and relative rate of change of reaction time are used to measure the complexity of reaction time characteristics. The calculation formula is: ; is the standard deviation of reaction times, is the mean reaction time, is the average of the difference between two consecutive reaction times. is the preset non-zero adjustment factor three, Represents the complexity of reaction time characteristics; Calculate the language feature complexity corresponding to the single-person evaluation , Facial expression feature complexity , body language feature complexity , and the reaction time characteristic complexity The Euclidean distance between the composed vector 1 and the standard vector 2 is the complexity of the single-person evaluation task. Then the average value of the complexity of all single-person evaluation tasks is calculated to obtain the evaluation task complexity index.

[0006] In a preferred embodiment, the logic for obtaining the regularity index of the evaluation behavior is: First, calculate the evaluation behavior regularity value of a single person evaluation. The calculation formula is: ; m is the total number of evaluation links, is the weight coefficient of the jth link, is the number of behavioral features contained in the jth link, is the standard deviation of the kth behavior characteristic in the jth link, reflecting the degree of fluctuation of the behavior characteristic. is the mean of the kth behavior feature in the jth link, reflecting the average level of the behavior feature. k is the index of the preset behavior feature, and j is the index of the evaluation link. Indicates the regularity value of the evaluation behavior of a single person; The average of the regularity values ​​of the evaluation behaviors of all single-person evaluations was calculated to obtain the evaluation behavior regularity index.

[0007] In a preferred embodiment, the weight coefficient of the jth link is The acquisition logic is: Get the sum of the standard deviations of all behavioral characteristics of the jth link to get the comprehensive difference value, Then divide the difference comprehensive value of the jth link by the preset difference threshold to get the weight coefficient of the jth link .

[0008] In a preferred embodiment, the logic of fuzzy reasoning is: The complexity index of assessment tasks and the regularity index of assessment behavior in the assessment work of the recruitment unit are taken as input variables, and the standard type of machine learning model construction of the recruitment unit is taken as the output variable. The input variables are fuzzified and the values ​​of the input variables are converted into fuzzy sets. The output variables are fuzzified and the output variables are converted into fuzzy sets. Fuzzy rules are formulated to describe the standard type of machine learning model construction under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to obtain the standard type of machine learning model construction of the recruitment unit, that is, the corresponding machine learning strategy.

[0009] In a preferred embodiment, the machine learning model is a convolutional neural network model.

[0010] In a preferred embodiment, the machine learning strategy refers to the number of layers of the convolutional neural network model.

[0011] In a preferred embodiment, the evaluation score table is sorted in descending order according to the candidate evaluation scores.

[0012] In a preferred embodiment, a system for machine learning-based behavioral assessment in recruitment applications includes: The evaluation data collection module collects behavioral data from the candidate evaluation process and preprocesses it to obtain a preprocessed evaluation data set; The assessment task analysis module performs assessment task complexity analysis and assessment behavior regularity analysis based on the assessment data set, and obtains the assessment task complexity index and assessment behavior regularity index respectively; The learning strategy reasoning module determines the machine learning strategy based on the evaluation task complexity index and the evaluation behavior regularity index according to the results of fuzzy reasoning; Model training module, which builds and trains machine learning models based on the determined machine learning strategy; The model application module applies the trained machine learning model to the recruitment work of the recruiting unit. The output result of the machine learning model is the candidate evaluation score. Finally, all the candidate evaluation scores are summarized and sorted to obtain an evaluation score form.

[0013] Technical effects and advantages of the present invention: To address the problem of overperformance, i.e., using deep models for simple tasks leads to unnecessary computational complexity and overfitting risk, the present invention dynamically selects shallow networks through the regularity index of behavioral data, quickly extracts stable features, and avoids overperformance. For example, for customer service positions with high regularity, a shallow model of 3-5 layers can meet the demand.

[0014] To address the problem of long computation time, i.e., the long computation time of deep networks, which may delay the recruitment process, the present invention dynamically adjusts the number of model layers and determines whether to use a deep network based on the task complexity index. For low-complexity tasks, shallow networks can significantly reduce training and inference time, thereby speeding up the candidate scoring process.

[0015] To address the problem of increased computing power costs, namely that a unified strategy may occupy too many computing resources and increase recruitment costs, the present invention dynamically adjusts the model depth according to task complexity and behavioral regularity so that computing power resources match recruitment needs.

[0016] The present invention can dynamically match the model depth to ensure resource utilization efficiency. The shallow model can quickly complete candidate scoring in simple scenarios, and the deep model can accurately extract features in complex scenarios. Through fuzzy logic reasoning, computing power is dynamically allocated according to data characteristics to avoid unnecessary waste of resources. The most appropriate convolutional neural network depth is selected for different recruitment scenarios to avoid excess performance and ensure feature extraction capabilities in complex scenarios. The dual indicators of task complexity and behavioral regularity are used to ensure that model construction matches candidate characteristics, improve scoring accuracy and scientific decision-making. The model depth can be flexibly adjusted for different positions and different task complexities to meet the needs of enterprises in diversified recruitment scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings; Figure 1 This is a schematic diagram of a method for using machine learning-based behavioral assessment in recruitment applications in Example 1 of the present invention.

[0018] Figure 2 This is a system schematic diagram of a machine learning-based behavioral assessment in recruitment applications in the present invention.

[0019] Figure 3 This is a schematic diagram of a method for using machine learning-based behavioral assessment in recruitment applications in Example 3 of the present invention. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Reference Figure 1 - Figure 2 The following examples are obtained: Embodiment 1: In the traditional recruitment process, evaluators often use subjective judgment to evaluate the suitability of candidates. However, this method is easily affected by the personal preferences and emotional factors of the evaluators. The behavioral evaluation method based on machine learning collects the behavioral data of candidates (such as language, reaction, response mode, etc. during the evaluation process) and analyzes these data with machine learning algorithms. It can quantitatively evaluate the candidate's work attitude, emotional intelligence, communication skills and other characteristics, reduce human bias, and improve the accuracy and fairness of recruitment.

[0022] This method generally includes three main steps: data collection, feature extraction, and model training. First, the behavioral data of candidates is collected through various channels (such as online tests, assessment processes, work simulations, etc.). Then, relevant features such as language usage habits, emotional fluctuations, reaction speed, etc. are extracted from the data using technologies such as natural language processing, image recognition, and sentiment analysis. Finally, a machine learning algorithm (such as decision tree, support vector machine, deep learning, etc.) is used to build a prediction model, and the model parameters are adjusted through training data to accurately predict the performance and behavior patterns of candidates at work.

[0023] Compared with traditional recruitment methods, the behavioral assessment method based on machine learning has higher objectivity and accuracy. The machine learning model can learn and optimize based on a large amount of data, which can help companies quickly screen out the most suitable personnel from a large number of candidates. However, the differences in the assessment content that different companies focus on will lead to different scales of machine learning models. If a unified machine learning model construction strategy is adopted, it is easy to have problems such as overperformance, long computing time, and increased computing power costs. Based on this, the present invention proposes the following content: A method for behavioral assessment based on machine learning in recruitment applications, comprising the following steps: Collect behavioral data from the candidate assessment process and preprocess it to obtain the preprocessed assessment data set; collecting behavioral data is the most basic part of the recruitment process, and the purpose is to extract the key features displayed by the candidate in the assessment. These behavioral data may include: language features: speech speed, intonation, keywords, etc.; emotional features: voice emotion analysis, facial expression changes, etc.; body language: gestures, body inclination, movements, etc.; reaction features: answer time, logic, etc. The significance of data preprocessing is to clean, normalize and format the original data for subsequent analysis and processing by machine learning algorithms. This includes: removing noisy data (such as background noise or irrelevant actions). Standardizing data features (such as normalizing feature values ​​to unify them to the same range), and structuring data (such as dividing the data in the assessment stage into time series or feature matrices). Through preprocessing, data quality can be ensured, redundant information can be reduced, and the accuracy of analysis results can be improved.

[0024] Based on the assessment data set, the complexity analysis of assessment tasks and the regularity analysis of assessment behaviors are respectively performed to obtain the complexity index of assessment tasks and the regularity index of assessment behaviors. The complexity analysis of assessment tasks is used to evaluate the difficulty of assessment tasks such as answering questions and the characteristic complexity of candidates in dealing with these tasks. Through complexity analysis, we can understand the adaptability and performance stability of candidates in dealing with complex tasks. Assessment task complexity index: The performance index of candidates in complex scenarios is calculated by analyzing behavioral data. The higher the complexity index, the more difficult the task and the greater the fluctuation of the candidate's performance. The regularity analysis of assessment behaviors is used to evaluate the consistency and stability of candidates' behaviors in different links. If the behavioral regularity is high, it means that the candidate can maintain a stable performance during the assessment process, reflecting his emotional stability, adaptability and ability to cope with stress. The regularity index of assessment behaviors is calculated by analyzing the variation range of the candidate's behavioral characteristics (such as tone, expression, and action) between assessment links. The higher the index, the more consistent the candidate's performance and the more stable the behavior.

[0025] Based on the evaluation task complexity index and the evaluation behavior regularity index, the machine learning strategy is determined according to the results of fuzzy reasoning; the purpose of fuzzy reasoning is to determine which number of layers of convolutional neural network (CNN) model is more suitable for the evaluation work of this unit based on the evaluation task complexity index and the evaluation behavior regularity index, and to determine the final machine learning strategy (that is, the number of CNN layers). The following is the significance and role of the fuzzy reasoning process and the key value of determining the machine learning strategy: The complexity index of the assessment task reflects the difficulty of the assessment task and the fluctuation of the candidate's performance in complex task scenarios. A higher complexity index usually means that the candidate's behavior pattern is more complex and the data features are more abstract, requiring deeper network layers (such as a higher number of CNN layers) to capture these features.

[0026] The evaluation behavior regularity index reflects the stability and consistency of the candidate's behavior. A higher regularity index usually means that the candidate's behavioral characteristics are relatively stable, the data pattern is relatively simple, and a shallower network (fewer CNN layers) can be used to effectively extract features.

[0027] Solve the problem of model layer selection: When building a CNN model, the number of layers (depth) directly determines the model's learning ability and computational complexity. The purpose of fuzzy reasoning is to automatically determine the required number of CNN layers based on the employer's focus (for example, whether the candidate's stability or ability to adapt to complex tasks is more important), combined with the complexity index and regularity index, so that the model is neither overfitting nor too simple.

[0028] Adjust machine learning strategies: Fuzzy reasoning can make reasonable suggestions for the number of network layers based on the input index (complexity and regularity), making the CNN model more adaptable to recruitment tasks. The number of layers (depth) of a convolutional neural network has a direct impact on the model's feature extraction capabilities, ability to learn complex patterns, and training time: Shallow network (few layers): Advantages: high computational efficiency, suitable for scenarios with simple data and strong regularity. Disadvantages: unable to capture complex behavioral features, may lead to underfitting. Applicable scenarios: when the regularity index is high and the complexity index is low, the shallow network can meet the needs.

[0029] Deeper networks (more layers): Advantages: Able to extract higher-level and more abstract features, suitable for complex task scenarios. Disadvantages: High computational cost and may lead to overfitting, especially when there is less data or strong regularity. Applicable scenarios: When the complexity index is high and the regularity index is low, a deeper network is more suitable.

[0030] By using fuzzy reasoning to determine the number of CNN layers, the model depth can be dynamically adjusted to adapt to data complexity and recruitment needs. Fuzzy reasoning can combine the index to make reasonable suggestions for the number of CNN layers, thereby avoiding manual repeated adjustment of model parameters and improving model development efficiency. By setting a reasonable number of model layers, the performance characteristics of candidates can be learned more accurately, thereby outputting more reliable evaluation scores.

[0031] According to the determined machine learning strategy, a machine learning model is constructed and trained, and the trained machine learning model is applied to the recruitment work of the recruiting unit. The output of the machine learning model is the candidate evaluation score. Finally, all the candidate evaluation scores are summarized and sorted to obtain the evaluation score form. The behavioral data of new candidates is analyzed through the trained model to output the evaluation score. The evaluation score integrates the performance of candidates in multiple dimensions and provides an objective evaluation basis for the recruiting unit. The model can quickly and accurately complete the scoring of a large number of candidates, reduce the deviation that may be introduced by human subjective judgment, and make the recruitment process more fair and scientific.

[0032] The logic for obtaining the complexity index of the evaluation task is: The complexity of language features is calculated by the complexity of grammatical analysis and the amplitude of emotional changes. The calculation formula is: ; Represents one of the preset language features at the tth moment, such as speaking speed, pause frequency, keyword richness, sentence length, etc. The preset language feature refers to a quantitative indicator extracted from the candidate's speech or text through language analysis, which is usually used to measure the candidate's language expression ability, tone and emotion. is the mean of the preset language features, is the standard deviation of the preset language feature, N is the total number of time points in the evaluation, is a preset non-zero adjustment factor of one, controlling the effect of fluctuations on complexity. Indicates the complexity of language features; Entropy is used to measure the uncertainty and diversity of emotional changes, and the complexity of facial expression features is obtained through the following calculation formula: ; M is the number of facial expression categories, is the probability of facial expression category i appearing in the evaluation, Indicates the complexity of facial expression features; The complexity of body language features is calculated through the frequency fluctuation and position change of body language. The calculation formula is: ; N is the total number of time points in the evaluation, Represents one of the preset body features at the tth moment, such as gesture frequency, movement amplitude, body tilt angle, etc. The preset body feature refers to a quantitative indicator extracted from the candidate's body movements or posture behaviors, which is used to measure the candidate's body language consistency. represents the preset limb features at the t-1th moment, is the preset non-zero adjustment factor 2, which controls the influence of the difference between the previous and next moments on the complexity of body language features. Indicates the complexity of body language features; The standard deviation and relative rate of change of reaction time are used to measure the complexity of reaction time characteristics. The calculation formula is: ; is the standard deviation of reaction times, is the mean reaction time, is the average of the difference between two consecutive reaction times. is the preset non-zero adjustment factor three, which controls the influence of reaction time difference on complexity. Represents the complexity of reaction time characteristics; Calculate the language feature complexity corresponding to the single-person evaluation , Facial expression feature complexity , body language feature complexity , and the reaction time characteristic complexity The Euclidean distance between the composed vector 1 and the standard vector 2 is the complexity of the single-person evaluation task. Then the average value of the complexity of all single-person evaluation tasks is calculated to obtain the evaluation task complexity index.

[0033] The Assessment Task Complexity Index is a key indicator used to measure the complexity and coping ability of candidates in assessment tasks. By comprehensively analyzing the complexity of multi-dimensional features such as language, facial expressions, body language and reaction time, the Assessment Task Complexity Index can quantify the relationship between the candidate's behavioral performance and task difficulty during the assessment process. The Assessment Task Complexity Index quantifies the candidate's performance in the following four aspects: Language feature complexity: measures the diversity, flexibility and emotional adaptability of the candidate's language expression through the complexity of grammatical analysis and the amplitude of emotional changes. The higher the complexity of language features, the more capable the candidate is of adapting to complex tasks in terms of language expression, but this may also be accompanied by a certain degree of instability.

[0034] Complexity of facial expression features: Entropy is used to calculate the uncertainty and diversity of facial expressions, and to assess the candidate's diverse emotional expressions and self-emotional regulation ability. A high level of facial expression complexity indicates that the candidate has rich or fluctuating emotions during the assessment, which may indicate that the candidate has high uncertainty in a stressful environment.

[0035] Body language feature complexity: Assess the amplitude and stability of the candidate's body language through the frequency fluctuations and position changes of body language. Higher body language complexity may reflect that the candidate shows more confidence or nervousness in the assessment.

[0036] Reaction time characteristic complexity: measures the adaptability and thinking speed of candidates when answering questions through the standard deviation and rate of change of reaction time. High reaction time complexity means that the candidate may need more time to adapt to complex tasks, but it may also mean that the flexibility is high.

[0037] Vector 1, the behavioral complexity vector, consists of the complexity of language features, the complexity of facial expression features, the complexity of body language features, and the complexity of reaction time features, which represents the comprehensive performance of each candidate in multiple dimensions. The Euclidean distance measures the degree of deviation of the candidate's complexity vector from the standard vector. A larger Euclidean distance indicates that the candidate's behavioral characteristics deviate more from the standard and the performance is more complex. A smaller Euclidean distance indicates that the candidate's behavior is closer to the standard behavior, the performance is more stable or the pattern is simpler. A larger Euclidean distance indicates that the candidate's behavior is highly complex, and there may be nonlinear or multidimensional interactions between features. This requires the construction of a deep CNN model (such as more than 10 layers) to extract low-level to high-level features layer by layer and capture the complex correlation between features. A smaller Euclidean distance indicates that the candidate's behavior is relatively stable or close to the standard pattern, and the relationship between features is simple or linear. This is suitable for building a shallow CNN model (such as about 3 layers) to efficiently process simple feature relationships.

[0038] In the recruitment process, the Euclidean distance distribution of different candidates reflects the overall complexity of the evaluation data. Based on this distribution, the recruiting unit can dynamically adjust the number of layers of the CNN model: The Euclidean distance is generally large (candidates generally have complex behaviors): a deeper CNN model (such as 10 layers or more) is needed to capture the high-order relationships of multi-dimensional features. Deep networks can gradually extract more abstract features from raw data (such as language features, expression features), thereby more accurately evaluating complex candidate behaviors.

[0039] The overall Euclidean distance is small (candidates' behavior is generally close to the standard): A shallower CNN model (such as 3-5 layers) can complete the feature extraction and scoring tasks. Shallow networks can quickly identify stable behavior patterns and avoid overfitting.

[0040] A higher complexity index indicates that the candidate's behavior pattern is complex, and there may be nonlinear and high-dimensional interactions between features. For example, there may be a complex relationship between language expression and emotional fluctuations, and the candidate's body movements may also be highly correlated with emotional expression. A deeper convolutional network (such as 6 layers or more) may be required to extract high-level features. Deep CNN can gradually learn from low-level features (such as edge features, simple fluctuations) to high-level features (such as complex emotional patterns, multi-dimensional feature interactions). A lower complexity index indicates that the candidate's behavior pattern is simple, and the relationship between features is relatively linear. The model only needs to capture the basic pattern, and a shallower convolutional network (such as 2-4 layers) may be enough to meet the needs and avoid over-computation.

[0041] The logic for obtaining the evaluation behavior regularity index is: first calculate the evaluation behavior regularity value of a single person's evaluation, and the calculation formula is: ; m is the total number of evaluation links, is the weight coefficient of the jth link, is the number of behavioral features contained in the jth link, is the standard deviation of the kth behavior characteristic in the jth link, reflecting the degree of fluctuation of the behavior characteristic. is the mean of the kth behavior feature in the jth link, reflecting the average level of the behavior feature. k is the index of the preset behavior feature, and j is the index of the evaluation link. Represents the regularity value of the evaluation behavior of a single person; calculate the average value of the regularity value of the evaluation behavior of all single people to get the regularity index of the evaluation behavior. The weight coefficient of the jth link The acquisition logic is: obtain the sum of the standard deviations of all behavioral characteristics of the jth link to obtain the difference comprehensive value, and then divide the difference comprehensive value of the jth link by the preset difference threshold to obtain the weight coefficient of the jth link .

[0042] The evaluation behavior regularity index measures the stability and consistency of the candidate's behavioral characteristics during the evaluation process. It reflects whether the candidate can maintain stable performance in a variety of evaluation tasks by analyzing the fluctuations of behavioral characteristics in different links. This index is related to the design of the number of layers of the convolutional neural network (CNN) model in the following ways: High regularity (large index): The behavioral characteristics of candidates in the assessment fluctuate less, and the performance is stable and close to the standard pattern. The correlation between data features is low, and the relationship is more linear. Shallow networks (such as 3-5 layers) are competent because stable data features are easy to extract and learn through simple networks. Shallow networks can quickly extract features and score without complex nonlinear relationship modeling.

[0043] Low regularity (small index): Candidates' performance fluctuates greatly in different evaluation stages, their behavioral characteristics lack consistency, and there may be complex nonlinear relationships between data features (such as large emotional fluctuations in some stages, but stable in other stages). Deep networks (such as 6-10 layers or more) are suitable for processing such complex data and gradually extracting high-dimensional interaction patterns between features. Deep networks can capture the potential patterns and regularities between different behavioral characteristics, improving the model's ability to score complex candidates.

[0044] The regularity index of the evaluation behavior reflects the complexity of the data from the perspective of the fluctuation of candidate characteristics. If the regularity index is high, the input data is relatively simple and the fluctuation is small, which is suitable for using shallow CNN to extract simple patterns. If the regularity index is low, the input data fluctuates greatly and has complex nonlinear characteristics, which requires deep CNN to extract features layer by layer to capture subtle changes.

[0045] The significance of the weight coefficient: reflects the importance of the link: if the behavioral characteristics of a certain evaluation link are more volatile (the sum of standard deviations is higher), the weight of the link If the characteristic volatility in a link is small (the sum of standard deviations is low), the weight The weight coefficient is dynamically adjusted by the ratio of the standard deviation to the difference threshold, so that the impact of different links on the behavior regularity index is more realistic. In some high-weight links (such as the stress assessment link), small fluctuations may be considered more significant, thus having a greater impact on the regularity index.

[0046] The logic of fuzzy reasoning is: The complexity index of assessment tasks and the regularity index of assessment behavior in the assessment work of the recruitment unit are taken as input variables, and the standard type of machine learning model construction of the recruitment unit is taken as the output variable. The input variables are fuzzified and the values ​​of the input variables are converted into fuzzy sets. The output variables are fuzzified and the output variables are converted into fuzzy sets. Fuzzy rules are formulated to describe the standard type of machine learning model construction under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to obtain the standard type of machine learning model construction of the recruitment unit, that is, the corresponding machine learning strategy.

[0047] In the construction of machine learning models of recruitment units, the core idea of ​​fuzzy logic is to infer the appropriate machine learning strategy (that is, the number of layers of the convolutional neural network model) based on the complexity index of the assessment task and the regularity index of the assessment behavior (input variables).

[0048] The fuzzy logic process can be divided into the following steps: Fuzzy input variables: Convert the values ​​of the evaluation task complexity index and the evaluation behavior regularity index into fuzzy sets. For example, task complexity can be divided into three fuzzy sets: "low", "medium", and "high". Behavior regularity can be divided into three fuzzy sets: "weak regularity", "medium regularity", and "strong regularity".

[0049] Fuzzify output variables: Divide the machine learning strategy (number of layers of the convolutional neural network) into fuzzy sets. For example, the number of layers can be divided into three fuzzy sets: "shallow network", "medium depth network", and "deep network".

[0050] Formulate fuzzy rules: Fuzzy rules are used to describe the relationship between the combination of input variables and the output variables. For example: If the task complexity is "high" and the behavior regularity is "weakly regular", use "deep network". If the task complexity is "medium" and the behavior regularity is "strongly regular", use "medium deep network".

[0051] Fuzzy reasoning: Reasoning based on the fuzzy set of input variables and fuzzy rules to obtain the fuzzy set of output variables.

[0052] Defuzzification: Convert the results of fuzzy reasoning into specific machine learning strategies (the specific number of layers of the convolutional neural network).

[0053] Fuzzy logic can dynamically adjust machine learning strategies based on the complexity of the test data: When the task complexity is high and the behavioral regularity is weak, it means that the data complexity is high and a deep convolutional neural network needs to be built to extract complex feature patterns. When the task complexity is low and the behavioral regularity is strong, it means that the data is simple and stable, and it is suitable to use a shallow convolutional neural network to reduce computational overhead and avoid overfitting. The needs of recruiting units and the characteristics of evaluation data are diverse, and a fixed number of model layers cannot meet the needs in different situations. Fuzzy logic dynamically determines the appropriate depth of the convolutional neural network by evaluating two key indicators: task complexity and behavioral regularity, so that the model can flexibly adapt to different scenarios. The evaluation task complexity and behavioral regularity index itself may contain a certain degree of ambiguity. For example: The complexity of the task may be between "medium" and "high". The regularity of the behavior may have certain rules, but also show instability in some links. Fuzzy logic can handle these ambiguities and give reasonable inference results using fuzzy sets and fuzzy rules.

[0054] According to the results of fuzzy reasoning, the recruiting unit can choose the appropriate number of convolutional neural network layers to avoid the problem of too many layers in the model causing waste of computing resources, or too few layers causing insufficient feature extraction. Fuzzy logic introduces a combination of quantitative and qualitative analysis methods. The complexity of the assessment task and the behavioral regularity index are quantitative data, which are converted into qualitative evaluations through fuzzy logic. Fuzzy rules are set according to the experience and needs of the recruiting unit. Through this combination method, the construction of the machine learning model is more in line with actual needs. The machine learning model is a convolutional neural network model. The machine learning strategy refers to the number of layers of the convolutional neural network model. The assessment score form is arranged in descending order according to the candidate's assessment score.

[0055] Examples of fuzzy logic applications: Evaluation task complexity index: low (small complexity index, simple task); medium (moderate complexity); high (large complexity index, complex task). Evaluation behavior regularity index: weak regularity (unstable behavior); medium regularity (relatively stable behavior); strong regularity (very stable behavior); output variables (number of convolutional neural network layers): shallow network (such as 3-5 layers); medium-depth network (such as 6-8 layers); deep network (such as 9-12 layers).

[0056] Fuzzy rule example: If the task complexity is "high" and the behavior regularity is "weakly regular", use "deep network". If the task complexity is "medium" and the behavior regularity is "medium regular", use "medium deep network". If the task complexity is "low" and the behavior regularity is "strongly regular", use "shallow network".

[0057] Fuzzy reasoning process: Assume that the evaluation task complexity index of a candidate is medium to high, and the behavior regularity index is medium to low. After fuzzification, the task complexity index may belong to both "medium" and "high", and the behavior regularity index may belong to both "weak regularity" and "medium regularity". According to the fuzzy rules, it is inferred that the number of network layers should be between "medium depth network" and "deep network". After defuzzification, the number of layers of the convolutional neural network is determined to be 8.

[0058] Embodiment 2: A system for behavioral assessment based on machine learning in recruitment applications, comprising: The evaluation data collection module collects behavioral data from the candidate evaluation process and preprocesses it to obtain a preprocessed evaluation data set; The assessment task analysis module performs assessment task complexity analysis and assessment behavior regularity analysis based on the assessment data set, and obtains the assessment task complexity index and assessment behavior regularity index respectively; The learning strategy reasoning module determines the machine learning strategy based on the evaluation task complexity index and the evaluation behavior regularity index according to the results of fuzzy reasoning; Model training module, which builds and trains machine learning models based on the determined machine learning strategy; The model application module applies the trained machine learning model to the recruitment work of the recruiting unit. The output result of the machine learning model is the candidate evaluation score. Finally, all the candidate evaluation scores are summarized and sorted to obtain an evaluation score form.

[0059] Embodiment 3: It should be noted that the following description is made based on Embodiment 1 to illustrate the background and significance of the present invention: Many recruitment companies face a common problem when recruiting talents: the company has clear requirements for specific positions and talent characteristics, but due to the complexity of each person's characteristics, it is often difficult to effectively screen through traditional interviews. Therefore, recruitment companies often rely on general models provided by machine learning model providers, which may not accurately reflect the needs of specific positions of recruitment companies and the complexity of candidate characteristics.

[0060] Since recruiting companies usually lack sufficient initial information or are unwilling to disclose their recruitment demand information to external machine learning companies, machine learning model providers can only design preliminary machine learning models based on general data sets and general business needs. When recruiting companies use these general models for the first time, there is a high probability that the output scoring results will be slightly different, resulting in insufficient prediction accuracy of interview scores and candidate performance. For example, if the candidate is suitable for a sales position, even with a general model, the scoring result will fall above the passing line of the screening. For example, the passing line is 60 points, and the general model may output a score of 70 points, but in actual work, it may reach 80 points. The effect of using a general model will not be bad, but after collecting and optimizing corporate data, the design of the model will be more targeted. For example, some companies are targeting the B-end and some are targeting the C-end. There will be some differences in the sales attributes for the B-end and the C-end, while the general model does not distinguish between the BC end. For example, the general model will consider more content in sales attributes, and will not highlight the focus of the content. Therefore, in order to improve the effect and accuracy of recruitment, recruiting companies will want to optimize the existing machine learning models. The key to optimization is to determine the machine learning strategy that best suits the company's needs, that is, the number of layers of the convolutional neural network, and based on this strategy, customize and train a machine learning model suitable for recruitment needs, so as to better provide data support for recruitment decisions.

[0061] The specific steps include: When the recruiting company proposes optimization requirements for the general model provided by the provider that is obtained through initial machine learning model training, the optimization mechanism is triggered; When the optimization mechanism is triggered, behavioral data is collected from the candidate evaluation process and preprocessed to obtain a preprocessed evaluation data set; Based on the evaluation data set, the complexity analysis of evaluation tasks and the regularity analysis of evaluation behaviors were performed to obtain the complexity index of evaluation tasks and the regularity index of evaluation behaviors respectively. Based on the evaluation task complexity index and the evaluation behavior regularity index, the machine learning strategy is determined according to the results of fuzzy reasoning; According to the determined machine learning strategy, a machine learning model is constructed and trained, and the trained machine learning model is applied to the recruitment work of the recruiting unit. The output result of the machine learning model is the candidate evaluation score. Finally, all candidate evaluation scores are summarized and sorted to obtain an evaluation score form.

[0062] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0063] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0064] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0065] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0066] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for behavioral assessment based on machine learning in recruitment applications, characterized in that: The following steps are involved: Collect behavioral data from the candidate evaluation process and preprocess it to obtain a preprocessed evaluation data set; Based on the evaluation data set, the complexity analysis of evaluation tasks and the regularity analysis of evaluation behaviors were performed to obtain the complexity index of evaluation tasks and the regularity index of evaluation behaviors respectively. Based on the evaluation task complexity index and the evaluation behavior regularity index, the machine learning strategy is determined according to the results of fuzzy reasoning; According to the determined machine learning strategy, a machine learning model is constructed and trained, and the trained machine learning model is applied to the recruitment work of the recruiting unit. The output result of the machine learning model is the candidate evaluation score. Finally, all candidate evaluation scores are summarized and sorted to obtain an evaluation score form.

2. The method of using machine learning-based behavior assessment in recruitment applications according to claim 1, characterized in that: The logic for obtaining the complexity index of the evaluation task is: The complexity of language features is calculated by the complexity of grammatical analysis and the amplitude of emotional changes. The calculation formula is: ； represents the preset language features at time t, is the mean of the preset language features, is the standard deviation of the preset language feature, N is the total number of time points in the evaluation, is the preset non-zero adjustment factor of one, Indicates the complexity of language features; Entropy is used to measure the uncertainty and diversity of emotional changes, and the complexity of facial expression features is obtained through the following calculation formula: ； M is the number of facial expression categories, is the probability of facial expression category i appearing in the evaluation, Indicates the complexity of facial expression features; The complexity of body language features is calculated through the frequency fluctuation and position change of body language. The calculation formula is: ； N is the total number of time points in the evaluation, represents the preset limb features at the tth moment, represents the preset limb features at the t-1th moment, is the preset non-zero adjustment factor of two, Indicates the complexity of body language features; The standard deviation and relative rate of change of reaction time are used to measure the complexity of reaction time characteristics. The calculation formula is: ； is the standard deviation of reaction times, is the mean reaction time, is the average of the difference between two consecutive reaction times. is the preset non-zero adjustment factor three, Represents the complexity of reaction time characteristics; Calculate the language feature complexity corresponding to the single-person evaluation , Facial expression feature complexity , body language feature complexity , and the reaction time characteristic complexity The Euclidean distance between the composed vector 1 and the standard vector 2 is the complexity of the single-person evaluation task. Then the average value of the complexity of all single-person evaluation tasks is calculated to obtain the evaluation task complexity index.

3. The method of using machine learning-based behavior assessment in recruitment applications according to claim 1, characterized in that: The logic for obtaining the regularity index of the evaluation behavior is: First, calculate the evaluation behavior regularity value of a single person evaluation. The calculation formula is: ； m is the total number of evaluation links, is the weight coefficient of the jth link, is the number of behavioral features contained in the jth link, is the standard deviation of the kth behavior characteristic in the jth link, reflecting the degree of fluctuation of the behavior characteristic. is the mean of the kth behavior feature in the jth link, reflecting the average level of the behavior feature. k is the index of the preset behavior feature, and j is the index of the evaluation link. Indicates the regularity value of the evaluation behavior of a single person; The average of the regularity values ​​of the evaluation behaviors of all single-person evaluations was calculated to obtain the evaluation behavior regularity index.

4. The method of using machine learning-based behavior assessment in recruitment applications according to claim 3, characterized in that: The weight coefficient of the jth link The acquisition logic is: Get the sum of the standard deviations of all behavioral characteristics of the jth link to get the comprehensive difference value, Then divide the difference comprehensive value of the jth link by the preset difference threshold to get the weight coefficient of the jth link .

5. The method of using machine learning-based behavior assessment in recruitment applications according to claim 4, characterized in that: The logic of fuzzy reasoning is: The complexity index of assessment tasks and the regularity index of assessment behavior in the assessment work of the recruitment unit are taken as input variables, and the standard type of machine learning model construction of the recruitment unit is taken as the output variable. The input variables are fuzzified and the values ​​of the input variables are converted into fuzzy sets. The output variables are fuzzified and the output variables are converted into fuzzy sets. Fuzzy rules are formulated to describe the standard type of machine learning model construction under different data type combinations. The fuzzified input variables are inferred through fuzzy rules to obtain the standard type of machine learning model construction of the recruitment unit, that is, the corresponding machine learning strategy.

6. The method of using machine learning-based behavior assessment in recruitment applications according to claim 5, characterized in that: The machine learning model is a convolutional neural network model.

7. The method of using machine learning-based behavior assessment in recruitment applications according to claim 6, characterized in that: The machine learning strategy refers to the number of layers in the convolutional neural network model.

8. The method of using machine learning-based behavior assessment in recruitment applications according to claim 7, characterized in that: The assessment score table is sorted in descending order according to the candidate's assessment score.

9. A system for machine learning-based behavior assessment in recruitment applications, implemented based on the method for machine learning-based behavior assessment in recruitment applications as claimed in any one of claims 1 to 8, characterized in that: include: The evaluation data collection module collects behavioral data from the candidate evaluation process and preprocesses it to obtain a preprocessed evaluation data set; The assessment task analysis module performs assessment task complexity analysis and assessment behavior regularity analysis based on the assessment data set, and obtains the assessment task complexity index and assessment behavior regularity index respectively; The learning strategy reasoning module determines the machine learning strategy based on the evaluation task complexity index and the evaluation behavior regularity index according to the results of fuzzy reasoning; Model training module, which builds and trains machine learning models according to the determined machine learning strategy; The model application module applies the trained machine learning model to the recruitment work of the recruiting unit. The output result of the machine learning model is the candidate evaluation score. Finally, all the candidate evaluation scores are summarized and sorted to obtain an evaluation score form.

Images