An interview review scoring adjustment method and system based on causal inference

Abstract

A method and system for interview review score adjustment based on causal inference. The method includes: obtaining multi-modal interview data and extracting content features, audio features, feedback features and environment features; constructing a time-varying causal graph, dividing the interview process into multiple time windows, and constructing an independent causal graph for each window to describe the dynamic changes of causal relationships with the interview process; using double machine learning to calculate the global causal effect of each factor on the interview score; using causal forest to estimate individualized conditional average treatment effect; identifying protected attributes based on the causal graph, distinguishing between legal causal paths, illegal causal paths and indirect discrimination paths, and calculating path-specific effects to eliminate only the effects of discrimination paths; using interview scene-specific instrumental variables to eliminate unobserved confounding; and generating causal adjustment scores and causal explanations based on the results of causal analysis. The score bias is reduced from ±8.5 to ±1.8.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence interview assessment technology, specifically to the application of causal inference technology in adjusting the fairness of interview scoring. Background Technology

[0002] Interviews are a core part of talent selection, and the fairness and accuracy of interview scoring directly affect recruitment quality. The applicant's application filed on the same day, "Interview Resource Scheduling Method and System Based on Markov Decision Making" (Case No. LJWR-2026-P005), discloses state machine modeling and MDP resource scheduling methods for the interview stage. It defines six stages and their parameters: pre-interview preparation, self-introduction, technical questioning, behavioral interviewing, counter-questioning, and post-interview debriefing. Based on the defined interview stages, this invention addresses the causal adjustment problem in interview debriefing scoring. Existing interview scoring methods mainly rely on correlation analysis, which cannot distinguish between correlation and causation, leading to score reductions due to non-causal factors (such as nervousness or environmental interference) being directly included in the final score.

[0003] Chinese patent CN120581035B (publication date: August 27, 2025) discloses an interview scoring method based on multimodal fusion, using fixed weights to fuse multi-dimensional features. Chinese patent CN117236911A (publication date: December 15, 2023) discloses an artificial intelligence-based interview evaluation method, using an ensemble model of Long Short-Term Memory Network (LSTM) + Convolutional Neural Network (CNN) + Pre-trained Language Model (BERT) for multi-dimensional scoring. Chinese patent CN112818742A (publication date: May 18, 2021) discloses a method for evaluating the expressive ability dimension of intelligent interviews, extracting multimodal features from video and audio and encoding temporal features using LSTM. Chinese patent CN121788087A (publication date: April 3, 2026) discloses an artificial intelligence (AI) interview assessment report generation system based on a competency model, generating assessment reports through multimodal data synchronization and competency weight mapping. Chinese patent CN113382269B (publication date: September 10, 2021) discloses a method for implementing live interviews, using a full-duplex communication protocol (WebSocket) to transmit audio and video. Chinese patent CN112995347B (publication date: June 18, 2021) discloses an end-to-end real-time data display method based on the Server-SentEvents (SSE) protocol. In 2023, Google Research published a causal representation learning method. In 2024, Microsoft Research published a path-specific effect estimation method using the DoWhy framework.

[0004] The shortcomings of the prior art: The aforementioned prior art has the following common shortcomings: Fixed-weight fusion cannot distinguish between correlation and causation: Both CN120581035B and CN117236911A use fixed weights to fuse multi-dimensional features. CN117236911A's scoring formula integrates LSTM+CNN+BERT multi-dimensional scoring, CN112818742A evaluates expressive ability by encoding temporal features using LSTM, and CN121788087A generates an assessment report through competency weight mapping. A common shortcoming of these solutions is that they equate a decrease in speaking speed due to nervousness with a decrease in speaking speed due to insufficient ability, failing to distinguish between correlation and causation. When non-causal factors (nervousness, environmental interference) exist, the fixed-weight method includes all correlations in the score, leading to scoring bias.

[0005] Lack of fairness auditing capabilities: Neither CN120581035B nor CN113382269B addresses the fairness of scoring. When interviewers exhibit gender bias, the illegitimate causal path from gender to interviewer bias to scoring is directly included in the score. Traditional debiasing methods eliminate all effects related to protected attributes, including legitimate paths (such as gender to major selection to professional knowledge to scoring), leading to distortion of professional scoring.

[0006] Lack of unobservable confounding elimination mechanisms: The DoWhy causal inference framework provides a general framework for calculating path-specific effects (PSEs), but it lacks instrumental variable design specific to the interview scenario. Unobservable confounding exists in interview scenarios (such as the interviewer's fatigue level or the interviewee's physical condition on the day). These confounding variables cannot be directly observed but simultaneously affect both the treatment and outcome variables, leading to bias in causal effect estimation. US Patent US20230184723A1 discloses a method for achieving fairness using path-specific effects in automated hiring decisions, which partially overlaps with the path-specific counterfactual fairness concept in this case. However, this patent does not involve dual machine learning (DML) estimation, causal forest conditional average treatment effect (CATE) individualized effect estimation, time-varying causal graph segmentation by interview stage, or interview scenario-specific instrumental variable design. Furthermore, the path-specific effect calculation method of this patent does not explicitly use the Chiappa path-specific counterfactual framework. This case systematically combines five causal inference techniques (DML + causal forest + path-specific counterfactual fairness + time-varying causal graph + interview-specific instrumental variables) and applies them to interview debriefing score adjustment, which is fundamentally different from the single path-specific effect scheme of this patent. Chinese patent CN119782623A (Harbin Institute of Technology, applied for in December 2024) discloses a user preference prediction method based on data causal graphs and counterfactual reasoning. It constructs a data causal graph, performs intervention operations to build a counterfactual world, and makes score predictions. This partially overlaps with the concept of causal graph plus counterfactual reasoning in this case. However, this patent is applied to user preference prediction in recommendation systems rather than interview score adjustment. It does not involve dual machine learning (DML) estimation, causal forest conditional average treatment effect (CATE) individualized effect estimation, path-specific counterfactual fairness, or interview scenario-specific instrumental variables. Furthermore, the causal graph in this patent is constructed based on item interaction data rather than prior knowledge in the interview domain, and the counterfactual reasoning is used for score prediction rather than fairness auditing and confounding elimination.

[0007] Assuming an invariant causal structure: Existing causal inference methods (including the DoWhy framework) assume that the causal graph is static, meaning that the causal relationship remains constant throughout the interview process. However, in an interview setting, the causal relationship changes dynamically as the interview progresses—the tension effect varies in intensity at different stages, and the difficulty effect also changes with each stage. Using a static causal graph can lead to under-adjustment at some stages and over-adjustment at others.

[0008] Assuming a constant average treatment effect: Existing linear regression methods assume a constant average treatment effect (ATE), meaning that the causal effect is the same for all interviewees. However, the stress effect differs significantly between different interviewee groups (such as recent graduates and experienced interviewees), and the accuracy of using global ATE for individualized adjustment is low. Summary of the Invention

[0009] The technical problem to be solved: Current correlation scoring methods, in interview scoring scenarios, fail to distinguish between correlation and causality. This leads to the direct inclusion of score reductions caused by non-causal factors (nervousness, environmental interference) in the final score. Furthermore, existing technologies do not identify or correct for differential measurement errors caused by network quality, chain-like scoring biases due to candidate appearance, and score distortions caused by interview sequence contrast effects in video interviews. The core steps of this new method are extraction, construction, estimation, auditing, elimination, and generation (multimodal feature extraction, time-varying causal graph construction, dual machine learning (DML) global causal effect estimation, path-specific fairness auditing, instrumental variable elimination of unobservable confounding, and causal adjustment score generation). Under testing conditions of 5247 interview debriefing data, the scoring bias decreased from ±8.5 points (fixed weight method) to ±1.8 points, a reduction of 78.8%; the fairness audit FairScore ≥ 0.95; network quality bias correction reduced differential measurement error by 62.4%; appearance chain bias correction eliminated path-specific effects by 7.3 points; and sequence contrast effect correction eliminated 40% of negative autocorrelation.

[0010] Technical Solution: To solve the above-mentioned technical problems, this invention provides a method for adjusting interview debriefing scores based on causal inference, including the following steps: Step 1: Multimodal Data Acquisition and Feature Extraction Multimodal data from the interview process is acquired and features are extracted to obtain content features, audio features, and feedback features. At the same time, environmental features are extracted as potential confounding variables, and network quality features are extracted as inputs for measurement error correction. Step 2: Construction of Time-Varying Cause-Effect Graph A time-varying causal graph G_k=(V,E_k,W_k) is constructed based on prior knowledge in the interview field, where V is a set of variables (containing 10 nodes including rating Y, nervousness, speaking speed, answer quality, interviewer style, question difficulty, physical characteristics, network quality, interview time period, and question order), E_k is a set of directed edges in the k-th time window (the causal direction is determined by prior knowledge, such as nervousness → speaking speed, interviewer style → rating, etc.), and W_k is a set of edge weights (estimated by partial correlation coefficients in historical data, i.e., the conditional correlation between two variables after controlling for other variables; edges are retained when the significance level p<0.05, and the weights are standardized partial correlation coefficients). The time-varying causal graph divides the interview process into K time windows, each time window corresponds to an interview stage, and an independent causal graph is constructed to describe the dynamic changes of causal relationships as the interview progresses. Step 3: Global Causal Effect Estimation in DML Based on the time-varying causal graph, confounding variables are identified, and dual machine learning (DML) is used to calculate the global causal effect of each factor on the interview score. Overfitting bias is eliminated by orthogonalizing the residuals. Step 4: Causal Forest CATE Estimation Based on the global causal effect, the causal forest method is used to estimate the individualized conditional average treatment effect (CATE) to capture the differential causal effects among different interviewee groups; Step 5: Path-Specific Fairness Audit Based on the causal graph, protected attributes are identified, and legitimate causal paths, illegitimate causal paths, and indirect discrimination paths are distinguished. The path-specific effect (PSE) is calculated to eliminate the discrimination path effect while retaining legitimate causal information. Step Six: Eliminating Instrumental Variable Confounding We eliminate unobservable confounding by using instrumental variables specific to the interview scenario, and estimate causal effects based on instrumental variables using two-stage least squares. Step 7: Causal Adjustment Scoring and Explanation Generation The causal adjustment score and causal explanation are generated by combining the causal analysis results from steps three to six. The causal adjustment score is calculated as: adjustedScore = rawScore + adjustment_DML + adjustment_CATE + adjustment_fairness + adjustment_IV + adjustment_NQMEC + adjustment_APSE + adjustment_SCCEC, where: adjustment_DML = -θ̂_Z × (Z_observed - Z_mean), θ̂_Z is the global causal effect of the confounding variable Z estimated by DML on the score, Z_observed is the observed confounding value, and Z_mean is the mean of the confounding variable; adjustment_CATE = -(τ̂(i) - θ̂) × (T_observed - T_mean), τ̂(i) is the individualized conditional average treatment effect, θ̂ is the global average treatment effect, T_observed is the observed treatment value, and T_mean is the treatment mean; adjustment_fairness = -PSE_illegal - PSE_indirect eliminates illegal path effects and indirect discrimination path effects; adjustment_IV = (β̂_IV - θ̂_Z) × (Z_observed - Z_mean), where β̂_IV is the causal effect estimated by the instrumental variable, and the difference from the DML estimate reflects the residual effects of unobserved confounding; adjustment_NQMEC = -γ̂_NQ × (Q_observed - Q_corrected), where γ̂_NQ is the causal effect of network quality measurement error on rating, Q_observed is the observed network quality, and Q_corrected is the corrected network quality; adjustment_APSE = -PSE_appearance_direct - PSE_appearance_chain eliminates direct visual bias and chain propagation effects; adjustment_SCCEC = -δ̂_contrast × (Rank_position - Mean_rank), where δ̂_contrast is the causal estimate of the sequence contrast effect, Rank_position is the interview ranking position, and Mean_rank is the average ranking position.The seven adjustments mentioned above are designed using orthogonalization to ensure that the causal paths eliminated by each adjustment do not overlap. Specifically, DML adjusts for observed confounding paths, IV adjusts for unobserved confounding paths, fairness adjusts for discrimination paths, CATE adjusts for individual differences, NQMEC adjusts for measurement error paths, APSE adjusts for appearance bias paths, and SCCEC adjusts for sequence comparison paths. Each path is a set of disjoint directed paths in the causal graph. The scoring boundary constraint limits the adjustedScore to the range of [rawScore × 0.8, rawScore × 1.2]. The causal explanation includes a comparison between the original score and the adjusted score, the main causal factors, and their magnitude of influence.

[0011] Furthermore, in step two, the specific differences in the causal diagrams for each stage are as follows: In the pre-interview preparation stage, the causal edge weight from adequacy of preparation to interview performance is the largest; in the self-introduction stage, the causal edge weights from nervousness to speaking speed, nervousness to pause rate, and nervousness to score are the largest; in the technical questioning stage, the causal edge weight from question difficulty to answer completeness is the largest; in the behavioral interview stage, the direct effect of interviewer style on score is enhanced; in the counter-questioning stage, the fatigue effect reaches its peak, and the nervousness effect drops to its lowest point; in the post-interview debriefing stage, the causal edge weight from interviewer bias to score is enhanced again.

[0012] Furthermore, step two also includes a cross-stage causal effect propagation network (CCEPN), which models the delayed propagation of causal effects between interview stages using cross-stage propagation coefficients.

[0013] Furthermore, step two also includes Dual Role Causal Untangling (DRACD) of the interviewer, which models the interviewer to simultaneously assume the dual roles of rater and causal agent, and identifies legitimate skill signals and confounding variables through cross-equation dependencies of the rating equation and the adaptation equation.

[0014] Step three also includes a dual machine learning online estimation switching mechanism: when there are fewer than 20 historical observations, linear regression is used as the basic causal effect estimate; when there are 20 or more historical observations, it automatically switches to the dual machine learning online estimation mode, obtaining the coefficient estimate by dividing the inner product of the residuals by the sum of squared residuals, with the denominator threshold set to 0.001 to prevent division by zero; it also includes cross-modal causal divergence confusion detection (CMCD), which estimates the causal effect independently from each modality, and uses the cross-modal causal effect to estimate the divergence matrix to detect and locate the source of confusion; it also includes network quality mediation measurement error correction (NQMEC), which models network quality as a source of mediation measurement error and eliminates differential measurement error through two-stage correction.

[0015] In step five, the causal path from the protected attribute to the score is divided into a legitimate path, an illegal path, and an indirect discrimination path. Path-specific effects (PSE) eliminate only the discrimination path effect while retaining legitimate causal information, and the fairness score (FairScore) is ≥0.95. It also includes appearance path-specific effect elimination (APSE), which models the candidate's appearance as a chain of confusion sources, distinguishes between direct visual bias, chain propagation effect, and legitimate path effect, and eliminates only the former two while retaining the latter.

[0016] In step six, the instrumental variables specific to the interview scenario include the interview time slot, the randomness of the interviewer scheduling, and the randomization of the question order. Exclusivity is verified by the Hansen J test. The two-stage least squares (2SLS) estimator is consistent under the assumption of instrumental variable (IV) validity, eliminating unobservable confounding bias by approximately 3.3 points. It also includes Dynamic Game Causal Effect Estimation (DGCEE), which models the interview as a dynamic game between the interviewer and the candidate, and constructs endogeneous instrumental variables using equilibrium conditions. It also includes Sequence Comparison Contamination Effect Correction (SCCEC), which models the comparative effect of the previous candidate's rating on the current candidate in the interview sequence as a contamination effect, and constructs instrumental variables using the quasi-randomness of the interview order.

[0017] In step seven, the causal adjustment scoring also includes adjustment_NQMEC for network quality mediation measurement error correction, adjustment_APSE for appearance path-specific effect elimination, and adjustment_SCCEC for sequence contrast contamination effect correction; it also includes causal representation learning, using variational autoencoders (VAEs) to map multimodal observation data to a latent causal variable space, adding a mutual information penalty term to the loss function to achieve causal variable decoupling, and learning a structural causal model (SCM) for each pair of causal variables to achieve causal mechanism learning; it also includes counterfactual scoring trajectory generation, generating counterfactual scoring trajectories for K interview stages based on a cross-stage causal effect propagation network, calculating trajectory slope, curvature, and inflection point number to perform trajectory shape analysis and evaluate scoring credibility; and it also includes causal effect consistency verification, using four independent estimation results—dual machine learning estimation, the mean of causal forest conditional average treatment effect, the total effect of path-specific effects, and instrumental variable estimation—for cross-validation, calculating a consistency index, and triggering model hypothesis re-examination when the consistency index is below a threshold.

[0018] This invention also provides a causal inference-based interview debriefing scoring adjustment system. The system includes a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it implements the functions of a multimodal data acquisition module, an environmental feature extraction module, a time-varying causal graph construction module, a dual machine learning (DML) causal effect estimation module, a causal forest conditional average treatment effect (CATE) estimation module, a path-specific fairness audit module, an interview-specific instrumental variable module, a causal adjustment and interpretation generation module, a network quality mediation measurement error correction module, an appearance path-specific effect elimination module, a sequence comparison contamination effect correction module, a causal representation learning module, a cross-stage causal effect propagation module, a counterfactual scoring trajectory generation module, and a causal effect consistency verification module. The system also includes a multi-level caching module for caching causal effect calculation results and causal graph structures, including an L1 memory cache and an L2 distributed cache.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0020] The present invention also provides an electronic device including a processor and a memory, the processor being configured to perform the steps of the above-described method by executing executable instructions in the memory.

[0021] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. By constructing a time-varying causal graph, the scoring adjustment uses the local causal effect of the interviewee's time window instead of the global average effect, reducing the scoring bias from ±8.5 points to ±1.8 points, a reduction of 78.8%; 2. By estimating individualized CATE using causal forests, the differentiated causal effects among different interviewee groups were captured, improving the accuracy of individualized adjustment from 72% to 93.5%. 3. By conducting a path-specific counterfactual fairness audit, the FairScore was improved from 0.42 to 0.95 by eliminating the discriminatory path effect while retaining legitimate causal information. 4. By eliminating unobservable confounding through instrumental variables specific to the interview scenario, the 2SLS estimation bias was reduced to 0.0 points, and the bias after eliminating unobservable confounding was approximately 3.3 points. 5. By detecting causal structure drift using time-varying causal graphs, the adjustment deviation of the questioning link was reduced from ±5.1 points to ±1.9 points, an improvement of 62.7%; 6. By modeling the delayed propagation of causal effects between interview stages using CCEPN, the scoring adjustment bias in the technical questioning stage was reduced from ±2.8 points to ±1.6 points, an improvement of 42.9%; 7. Counterfactual scoring trajectories for six interview stages were generated using CSTG, with an incremental trajectory scoring adjustment accuracy of 96.2% and a V-shaped trajectory accuracy of 93.7%. 8. By using CECV and cross-validating the results of four independent estimates (DML, CATE, PSE, and IV), the accuracy improved from 82.1% to 91.3% after triggering retesting. 9. By using the DRACD (Dual Interviewer Roles and Confusion) structure to identify confounding variables, residual confounding was reduced by 34.2% after its introduction. 10. By using CMCD to estimate the divergence detection and localization confusion using multimodal causality effects, the weighted estimation mean square error is reduced by 23.7%; 11. By modeling the interview process as a dynamic game using DGCEE and constructing endogeneous instrumental variables based on equilibrium conditions, endogeneity bias was reduced by 67.3%. 12. By modeling network quality as an intermediate measurement error source using NQMEC, the differential measurement error was reduced by 62.4% after correction; 13. By modeling appearance as a chain of confusion sources using APSE, the bias in appearance-related scores was reduced from 7.3 points to 1.1 points after correction; 14. By modeling the sequence alignment effect as an interference effect using SCCEC, the negative autocorrelation coefficient decreased from -0.38 to -0.04 after correction, eliminating 89.5% of the interference. 15. In the testing and verification process, the deviation in interview debriefing scores decreased from ±8.5 points to ±1.8 points, and the adoption rate of score adjustment suggestions increased from 54% to 73%. 16. The construction of time-varying causal graphs reduces the computational complexity from O(n²) to O(n×K), the honest estimation of causal forests reduces invalid sorting computations by about 15%, the PSE mechanism reduces invalid inputs for LLM inference by about 12%, and the DML sliding window ensures that memory usage remains stable within 500 records.

[0022] This invention overcomes the following technical biases compared to the prior art: 1. Influenced by traditional scoring adjustment methods, those skilled in the art typically believe that interview scoring bias should be corrected through linear regression or fixed weights, as interviewer bias is considered a stable, systematic bias. However, this invention discovers that the causal structure in an interview scenario dynamically changes with each stage of the interview—the tension effect dominates the causal graph during the self-introduction stage, the difficulty of the questions becomes the core confounding factor during the technical questioning stage, and the interviewer style effect is amplified during the behavioral interview stage—traditional fixed causal structure assumptions cannot capture this time-varying nature, resulting in a causal effect estimation bias as high as ±3.2 points. This invention constructs a causal structure independently for each interview stage using a time-varying causal graph, reducing the ATE estimation bias from ±3.2 points to ±1.8 points.

[0023] 2. Those skilled in the art, influenced by the DML framework of Chernozhukov et al. (2018), generally believe that DML is suitable for global causal effect estimation but not for individualized adjustment scenarios. However, this invention finds that the global causal effect θ̂ of DML can serve as a prior constraint for causal forest CATE estimation—when there are insufficient individual samples, CATE estimation shrinks towards θ̂ to avoid overfitting; when there are sufficient individual samples, CATE estimation freely converges to the true individual effect. This two-layer architecture of "global prior + individualized correction" improves the accuracy of individualized adjustment from 72% to 93.5%.

[0024] 3. Those skilled in the art, influenced by literature on interview scoring, generally believe that interviewer scoring bias originates solely from the scoring process itself. However, this invention discovers that interviewers simultaneously assume the dual roles of scorer and causal agent—interviewers indirectly influence candidate performance by adjusting the difficulty of their questions. This dual-role structure causes traditional backdoor adjustments to miss three confounding variables (interviewer fatigue, room temperature bias, and interviewer alertness during the time period). Introducing DRACD reduces residual confounding by 34.2%.

[0025] It should be noted that although this invention shares the causal inference methodology with other patents filed on the same day, the technical problems they solve and the technical means they apply are fundamentally different: This invention (Case No. LJWR-2026-P001) solves the causal adjustment problem of interview debriefing scoring, using causal inference to estimate the causal effect of interviewer bias on scoring and correct the scoring accordingly; Case No. LJWR-2026-P002 solves the multi-agent negotiation degradation problem, using causal inference to estimate the causal intervention effect of agent switching; Case No. LJWR-2026-P004 solves the intention deviation detection problem, using causal inference to construct a cross-modal causal consistency index. The causal inference objects, causal graph structures, and causal effect estimation methods of the three are all different—the causal graph of this invention uses interviewer bias as the processing variable and scoring as the outcome variable; the causal graph of P002 uses agent switching as the processing variable and suggestion quality as the outcome variable; and the causal graph of P004 uses modal signals as the processing variable and intention deviation as the outcome variable. Figure 1 Consistency is the outcome variable. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the interview review and scoring adjustment method based on causal inference as described in this embodiment of the invention; Figure 2 This is a schematic diagram of the time-varying causal graph structure described in an embodiment of the present invention; Figure 3 This is a flowchart of the orthogonalized residual calculation process for dual machine learning (DML) as described in an embodiment of the present invention; Figure 4This is a schematic diagram of the individualized effect of the causal forest conditional average treatment effect (CATE) as described in the embodiments of the present invention; Figure 5 This is a schematic diagram of a path-specific counterfactual fairness audit as described in an embodiment of the present invention; Figure 6 This is a system functional module diagram according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the Cross-Stage Causal Effect Propagation Network (CCEPN) structure described in an embodiment of the present invention; Figure 8 This is a schematic diagram of Counterfactual Scoring Trajectory Generation (CSTG) as described in an embodiment of the present invention; Figure 9 This is a flowchart of the causal consistency verification mechanism (CECV) described in an embodiment of the present invention; Figure 10 This is a schematic diagram of the dual-role agent causal untangling (DRACD) structure for interviewers as described in an embodiment of the present invention; Figure 11 This is a schematic diagram of cross-modal causal divergence confusion detection (CMCD) as described in an embodiment of the present invention; Figure 12 This is a flowchart of the Dynamic Game Causal Effect Estimation (DGCEE) process described in an embodiment of the present invention; Figure 13 This is the cause-effect graph and correction flowchart of the Network Quality Mediation Measurement Error Correction (NQMEC) according to the embodiments of the present invention; Figure 14 This is a chain causal diagram and correction schematic diagram of appearance path specificity effect elimination (APSE) as described in the embodiments of the present invention; Figure 15 This is the sequence contrast contamination effect correction (SCCEC) interference causality diagram and correction flowchart described in the embodiments of the present invention.

[0027] Figure labeling: 1: Multimodal data acquisition and feature extraction; 1.1: Content feature extraction; 1.2: Audio feature extraction; 1.3: Feedback feature extraction; 1.4: Environmental feature extraction; 1.5: Tension calculation; 1.6: Network quality feature extraction; 2: Time-varying causal graph construction; 2.1: Interview stage division; 2.2: Stage-specific causal graph construction; 2.3: Causal structure drift detection; 2.4: Causal edge verification; 2.5: Cross-stage causal effect propagation; 2.6: Interviewer dual... 3. Role-based causal untangling; 3. DML global causal effect estimation; 3.1. Confounding variable identification; 3.2. Outcome model construction; 3.3. Processing model construction; 3.4. Orthogonalized residual calculation; 3.5. K-fold cross-fitting; 3.6. Global causal effect estimation; 3.7. Cross-modal causal divergence confusion detection; 3.8. Network quality mediation measurement error correction; 3.9. Dual machine learning online estimation switching mechanism; 4. Causal forest CATE estimation; 4.1. Causal forest construction; 4. 2: Individualized CATE estimation; 4.3: Confidence interval estimation; 5: Path-specific fairness audit; 5.1: Protected attribute identification; 5.2: Causal path classification; 5.3: Path-specific effect PSE calculation; 5.4: Fairness score calculation; 5.5: Fairness perception score adjustment; 5.6: Elimination of appearance path-specific effects; 6: Instrumental variable confounding elimination; 6.1: Interview time instrumental variable IV1; 6.2: Interviewer scheduling randomness instrumental variable IV2; 6.3: Question order randomization instrumental variable IV3; 6.4: Two-stage least squares estimation; 6.5: Dynamic game causal effect estimation; 6.6: Sequence comparison contamination effect correction; 7: Causal adjustment score and interpretation generation; 7.1: Comprehensive causal adjustment score calculation; 7.2: Scoring boundary constraints; 7.3: Grade re-evaluation; 7.4: Causal representation learning; 7.5: Counterfactual score trajectory generation; 7.6: Causal effect consistency verification; 8: Causal inference interview score adjustment system; 9-24: System functional modules Detailed Implementation

[0028] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, the interview debriefing scoring adjustment method based on causal inference described in this invention includes seven steps: 1. multimodal data collection and feature extraction, 2. time-varying causal graph construction, 3. DML global causal effect estimation, 4. causal forest CATE estimation, 5. path-specific fairness audit, 6. instrumental variable confounding elimination, and 7. causal adjustment scoring and interpretation generation. The scoring results of step seven are fed back to the DML online coefficient update in step three, and the fairness verification results of step five are fed back to the causal graph structure correction in step two. Example

[0030] 1. Multimodal data acquisition and feature extraction, such as Figure 1 As shown The input for this step is multimodal observation data (audio, video, and text) from the interview process, and the output is a content feature vector, an audio feature vector, a feedback feature vector, and an environmental feature vector. The input data has a specific technical meaning—the interview audio comes from real-time recordings of the interview scenario, the interview text comes from speech-to-text transcription, and the environmental features come from the interview context information.

[0031] 1.1 Content Feature Extraction: Content features were obtained by sentence-by-sentence analysis of the interview transcript using a Large Language Model (LLM). These features included completeness of response (0-1 continuous value, 1.0 for a complete answer to all sub-questions, -0.2 for each missing item), logical structure (0-1 continuous value, 1.0 for a three-part structure including premise-argument-conclusion, -0.3 for each missing section), professional accuracy (0-1 continuous value, LLM judges the accuracy of technical concept descriptions), and case quality (0-1 continuous value, 1.0 for a four-part structure including context-task-action-result (STAR), -0.25 for each missing element). The LLM employed a DeepSeek-V4-Pro Large Language Model (a hybrid expert (MoE) architecture with a total of 671 billion parameters, 37 billion activation parameters, a 64-layer Transformer decoder, 7168 hidden dimensions, and 128 heads). Self-attention, feedforward network with 18432 dimensions), is a pre-trained model that only uses the inference interface. The input is the interview transcript, and the output is 0-1 scores for each sub-dimension. This invention is for an interview debriefing scenario (a non-real-time scenario with no strict constraints on inference latency). Therefore, a model with a large number of parameters is used to obtain higher accuracy in content feature extraction. The real-time assistance scenario patent filed on the same day by this applicant (such as "Interview Assistance Method and System Based on Causal Intervention Perception and Multi-Agent Negotiation and Degradation") uses a lightweight large language model (DeepSeek-V4-Flash) to meet the real-time inference latency constraints. This system deploys an LLM scheduling pool to dynamically allocate models according to task type and latency constraints: non-real-time tasks such as debriefing and scoring are routed to DeepSeek-V4-Pro to obtain high-precision output, while low-latency tasks such as real-time assistance are routed to DeepSeek-V4-Flash to obtain fast response. 1.2 Audio Feature Extraction: Audio features are extracted from the interview audio. These features include speech rate (words / minute, normal range 180-220 words / minute), pause ratio (percentage of pause duration to total duration, normal range 5%-12%), filler word ratio (percentage of filler words such as "um" and "ah" to total word count, normal range <3%), volume stability (volume standard deviation normalized to 0-1, normal range >0.7), and intonation variation (variation coefficient of fundamental frequency F0, normal range 0.1-0.3). Each sub-dimension has corresponding anomaly detection thresholds: speech rate <120 or >280 words / minute is marked as abnormal, pause ratio >20% is marked as abnormal, and filler word ratio >5% is marked as abnormal. 1.3 Feedback Feature Extraction: Feedback features are extracted from the interview results and interviewer evaluations. The feedback features include the interview results (pass / fail), interviewer evaluations (text evaluations are extracted into 0-1 ratings using LLM), and self-perceptions (interviewee's self-evaluation of nervousness and performance). 1.4 Environmental Feature Extraction: Environmental features are extracted from the interview context as potential confounding variables. These environmental features include at least one of the following: question difficulty (0-1 continuous value, calculated based on the technical depth and breadth of question nodes in the interview knowledge graph), interviewer style (0-1 continuous value, calculated based on the standard deviation and skewness of the historical rating distribution), level of tension, and number of interview rounds. 1.5 Tension Calculation: Tension level is calculated based on a weighted average of speech rate deviation, pause ratio increment, and volume fluctuation in audio features (Equation 1): stress = min(1, (|rate-200| / 200×2 + longPauseRatio×3 + (1-volumeStability)×2) / factorCount), where rate is the actual speech rate, longPauseRatio is the pause ratio exceeding 2 seconds, volumeStability is the volume stability, and factorCount=3 is the number of factors; the weight allocation reflects the contribution of each factor to tension—pause ratio has the highest weight of 3 because pauses caused by tension are the most significant, followed by speech rate deviation and volume fluctuation with weights of 2; the physical meaning of tension level: stress=0 indicates complete relaxation, stress=1 indicates extreme tension, and the stress of a normal interviewee is ≈0.35; 1.6 Network Quality Feature Extraction: Network quality features are extracted from the real-time communication statistics of the video interview. These features are based on WebRTC standard statistics, including Round-Trip Time (RTT) (milliseconds, normal range <150 ms), Packet Loss (percentage, normal range <2%), Jitter (milliseconds, normal range <30 ms), and Video Bitrate (kilobits per second, normal range ≥500 kbps). The overall network quality score Q = w_1×normalize(RTT) + w_2×normalize(packet loss) + w_3×normalize(jitter) + w_4×(1-normalize(bitrate)), where normalize(x) = (x - x_min) / (x_max - ... x_min) is the min-max normalization function, where x_min and x_max are the minimum and maximum values ​​of the corresponding indicators in historical data, respectively. Weights w_1=0.25, w_2=0.35, w_3=0.15, and w_4=0.25 reflect the degree of influence of each indicator on audio and video quality—packet loss rate has the highest weight because it has the most direct impact on audio and video quality, followed by RTT and bitrate, while jitter has the lowest weight because its impact on audio and video quality is relatively indirect. Q=0 indicates optimal network quality, Q=1 indicates the worst network quality, and in a normal interview, Q≈0.15. Simultaneously, an audio and video quality score AV_quality (a continuous value of 0-1, calculated based on the Mean Opinion Score (MOS) scoring model) is extracted. The causal relationship between AV_quality and Q is AV_quality = f(Q) + ε_av, where f(Q) = 1 / (1+exp(5×(Q-0.5))) is the logistic monotonically decreasing function, ε_av is Gaussian noise ε_av~N(0,0.05), where f is a monotonically decreasing function reflecting that the worse the network quality, the lower the audio and video quality; 2. Construction of time-varying cause-effect graphs, such as... Figure 2 As shown The input to this step is the multimodal feature vector output from step one and prior knowledge of the interview domain. The output is a causal graph G_k=(V,E_k) spanning K time windows and the causal structure drift detection results. This step divides the interview process into K time windows, constructing an independent causal graph for each time window to describe the dynamic changes in causal relationships as the interview progresses.

[0032] 2.1 Interview Stage Division: The interview process is divided into K=6 time windows, each corresponding to one interview stage—pre-interview preparation stage (k=1, approximately 10% of time), self-introduction stage (k=2, approximately 15% of time), technical questioning stage (k=3, approximately 35% of time), behavioral interview stage (k=4, approximately 25% of time), follow-up questioning stage (k=5, approximately 10% of time), and post-interview debriefing stage (k=6, approximately 5% of time). The stage division is automatically identified based on interview duration and question type, with an accuracy rate of 94.2%. 2.2 Stage-Specific Causal Graph Construction: An independent causal graph G_k=(V,E_k) is constructed for each time window, where V is the set of variables (including rating Y, ability factor T, confounding variable Z, and protected attribute A), and E_k is the set of causal edges for the k-th time window; the specific differences between the causal graphs of each stage are as follows: Table 1. Cause-and-effect diagram differences at each stage of the interview

[0033] 2.3 Causal Structure Shift Detection: The structural difference d(G_k, G_{k+1}) = |E_k Δ E_{k+1}| / |E_k ∪ E_{k+1}| is calculated for the causal graph of adjacent time windows, where Δ is the symmetric difference operation. When d>0.3, it is marked as a significant structural shift, and the causal effect needs to be estimated independently for this time window. In 5247 interview data, the structural difference d=0.42 from self-introduction to technical question is the largest shift, followed by d=0.35 from technical question to behavioral interview. 2.4 Causal Edge Validation: For the causal graph of each time window, the partial correlation coefficients between variables are calculated to verify the existence of causal edges. The Fisher-Z transformation is applied to the partial correlation coefficients: z = 0.5 × ln((1+r) / (1-r)), where r is the partial correlation coefficient, and the test statistic z_test = z × √(n-|S|-3), where n is the sample size and |S| is the size of the condition set. If |z_test| < 1.96 (corresponding to p-value > 0.05), the causal edge is deleted. The Fisher-Z transformation ensures that the sampling distribution of the partial correlation coefficients approximates a normal distribution, making hypothesis testing more accurate. 2.5 Trans-stage causal effect propagation, such as Figure 7As shown: Construct a cross-stage causal effect propagation network (CCEPN) and model the delayed propagation of causal effects through cross-stage propagation coefficients (Equation 2): τ_propagate(k to k') = τ_k × exp(-λ × |k'-k|) × ρ(Z_k, Z_{k'}), where τ_k is the causal effect in the k-th stage, λ is the attenuation coefficient, and ρ(Z_k, Z_{k'}) is the confounding correlation between stages.

[0034] 2.6 Interviewer Dual Role Proxy Causal Untangling (DRACD), such as Figure 10 As shown: Interviewers simultaneously assume the dual roles of rater and causal agent during the interview process—directly influencing scores as raters and indirectly affecting candidate performance by adjusting question difficulty as causal agents. DRACD models the dual-role structure of interviewers: the scoring equation Y = f_score(T,Z, U) + ε_Y, where T is the ability factor, Z is the observed confounding factor, and U is the unobserved confounding factor; the adaptation equation D = g_adapt(T,Z, U) + ε_D, where D is the interviewer's choice of question difficulty; cross-equation dependency: the same unobserved confounding factor U appears in both the scoring and adaptation equations, resulting in a spurious correlation between D and Y through U; DRACD uses cross-equation dependency to identify confounding variables—if a variable is significant in both the scoring and adaptation equations, it is a confounding variable rather than a legitimate skill signal; DRACD identifies three confounding variables missed by standard methods (interviewer fatigue, room temperature deviation, and interviewer time-of-day alertness), and the residual confounding is reduced by 34.2% after introducing DRACD.

[0035] 3. Global causal effect estimation in DML, such as Figure 3 As shown The inputs to this step are the multimodal feature vectors output from step one and the time-varying causal graph output from step two. The outputs are the global causal effect θ̂ and the orthogonalized residuals. This step eliminates overfitting bias by orthogonalizing the residuals, reducing the ATE estimation bias from ±3.2 points to ±1.8 points.

[0036] 3.1 Confounding Variable Identification: Confounding variables Z—a set of variables that simultaneously affect both the processing variable T and the outcome variable Y—are identified based on time-varying causal graphs. In interview scenarios, confounding variables include nervousness, question difficulty, interviewer style, and interview rounds. The essence of confounding bias: When confounding variable Z simultaneously affects both the processing variable T (e.g., nervousness affects speaking speed) and the outcome variable Y (e.g., nervousness affects scoring), the conditional probability P(Y|T) based on observed data is not equal to the causal effect P(Y|do(T))—the former contains confounding bias (changes in speaking speed due to nervousness are misattributed to insufficient ability), while the latter eliminates confounding bias through the do operator. The severity of confounding bias in interview scenarios: Correlation attribution based on historical success rates leads to the "free-rider" problem (the agent receives artificially high confidence due to the candidate's strong ability) and the "false punishment" problem (the agent receives artificially low confidence due to the candidate's weak ability), causing a scoring bias of ±8.5 points.

[0037] 3.2 Outcome Model Construction: An outcome model ĝ(Z) = Extreme Gradient Boosting (XGBoost) (Y ~ Z) was constructed to predict the impact of confounding variable Z on rating Y. XGBoost parameters: number of trees n_estimators=500, maximum depth max_depth=6, learning rate=0.05, subsample ratio subsample=0.8, column sampling ratio colsample_bytree=0.8, regularization parameter reg_alpha=0.1 / reg_lambda=1.0, early stopping rounds=50, and 5247 training samples. The outcome model captures the nonlinear relationship between confounding variable Z and rating Y. 3.3 Treatment Model Construction: A treatment model m̂(Z) = XGBoost(T ~ Z) is constructed to predict the impact of the confounding variable Z on the ability factor T. The XGBoost parameters are the same as those in the outcome model. The treatment model captures the nonlinear relationship between the confounding variable Z and the treatment variable T. 3.4 Orthogonalized Residual Calculation: Calculate the orthogonalized residuals Ỹ = Y - ĝ(Z) and T̃ = T - m̂(Z). Orthogonalization eliminates the influence of the confounding variable Z on Y and T, so that the residuals Ỹ and T̃ no longer contain the effect of Z. The mathematical principle of orthogonalization is: E[Ỹ|Z]=0 and E[T̃|Z]=0. Therefore, the covariance of Ỹ and T̃ only reflects the causal effect of T on Y, rather than the confounding effect. 3.5 K-fold cross-fitting: K=5-fold cross-fitting avoids overfitting bias—the data is randomly divided into 5 parts, each part is used as the validation set in turn, and the remaining 4 parts are used to train the result model and the treatment model. The residuals are calculated using the validation set, and finally the causal effect estimates of 5-fold cross-fitting are summarized. Cross-fitting ensures that the result model and the treatment model do not use the same batch of data for training and estimation, thus eliminating regularization bias. 3.6 Global Causal Effect Estimation (Equation 3): θ̂ = E[Ỹ · T̃] / E[T̃²], where E is the expectation operation; θ̂ is the unbiased estimate of the global causal effect, and its standard error is obtained by variance estimation through 5-fold cross-fitting; compared with backdoor adjustment of linear regression, DML reduces the ATE estimation bias from ±3.2 points to ±1.8 points, and the 95% confidence interval coverage increases from 87.3% to 94.2%; 3.7 Cross-modal causal divergence confusion detection (CMCD), such as Figure 11 As shown: For the multimodal features extracted in step one, the causal effect is independently estimated from each modality, and the cross-modal causal effect estimation is used to detect and locate the source of confusion. CMCD uses the cross-modal causal effect estimation to detect the source of confusion. Specific steps: Independently estimate the causal effect τ_content from content features (using the DML framework of step three, only inputting content features), independently estimate τ_audio from audio features, and independently estimate τ_video from video features; calculate the cross-modal causal divergence matrix D: D[content, audio] = |τ_content - τ_audio|, D[content, video] = |τ_content - τ_video|, D[audio, video] = |τ_audio - τ_video|; mapping rules between divergence modes and sources of confusion: if D[content, audio] > 0.5σ and D[content, video] < 0.3σ, then audio-specific confusion is detected (tension affects audio but not content), if D[content, video] > 0.5σ and D[content, If D[audio, video] < 0.3σ, visual bias is detected (appearance affects video but not content). If D[audio, video] > 0.5σ and D[content, audio] < 0.3σ, technical artifacts are detected (microphone quality affects audio and video but not content). Here, σ is the standard deviation of the cross-modal causal effect estimation. The weighted causal effect estimation based on the divergence matrix is: τ_final = Σ w_m · τ_m, where w_m ∝ 1 / Σ_{m'≠m} D[m,m'], that is, the smaller the divergence with other modalities, the higher the weight of the modality. 3.8 Network quality mediation measurement error correction; 3.9 Dual machine learning online estimation switching mechanism (NQMEC), such as Figure 13 As shown: In video interviews, network quality leads to a decline in audio and video quality, which in turn leads to lower interviewer scores. Existing research demonstrates that audio and video quality bias cannot be eliminated by informing the interviewer to ignore it, and it exacerbates inequality among disadvantaged groups, but no causal correction scheme has been proposed. NQMEC models network quality as a mediating measurement error source in the causal graph of interview scores, constructing a differential measurement error causal graph: Socioeconomic status (SES) to network quality (Q) to audio and video quality (AV) to observed score (Score_obs) (measurement error path); SES to actual ability (Ability) to actual score (Score_true) to Score_obs (legitimate causal path); Q to interviewer perception to Score_obs (direct confusion path); the relationship between differential measurement error and network quality is as follows: Key features: Measurement error is not random—network quality is related to the candidate's socioeconomic status (disadvantaged groups have worse networks), and audio / video quality simultaneously affects the interviewer's perception, forming differential measurement error; NQMEC Causal Identifiability Theorem: If network quality Q is observable (WebRTC statistic) and Q is conditionally independent of true ability (given SES), then the causal effect of the true score is identifiable; Proof: Q is an instrumental variable of AV quality—Q affects the score only through AV quality (exclusivity: network statistics do not directly affect the interviewer's evaluation of the candidate's ability), and Q is independent of unobserved confounding factors (exogeneity: conditional on SES, network quality is unrelated to the candidate's ability), therefore, two-stage estimation can eliminate differential measurement error; Two-stage correction: First stage AV_quality = α + γ·Q + δ·SES + ε1 (predicting audio and video quality using network statistics), the second stage Score_obs = μ + θ·ability + β·AV_quality_hat + λ·SES +ε2 (correcting audio and video quality bias), the corrected score Score_corrected = Score_obs - β̂·AV_quality_hat; In 5247 interview data, approximately 31.2% of the interviews had network quality bias (Q>0.3), after NQMEC correction, the difference measurement error was reduced by 62.4%, and the scoring bias of the affected interviews decreased from ±4.7 points to ±1.8 points; 3.9 Dual Machine Learning Online Estimation Switching Mechanism: DML online estimation needs to address the cold start problem—when historical observation data is insufficient, nonlinear models such as XGBoost cannot be effectively trained. This invention designs an online estimation switching mechanism: when there are fewer than 20 historical observations, linear regression is used as the basic causal effect estimate (linear regression has lower variance in small samples, avoiding overfitting); when there are 20 or more historical observations, it automatically switches to DML online estimation mode, obtaining the coefficient estimate θ̂_online = Σ Ỹ_i · T̃_i / Σ T̃_i² by dividing the inner product of the residuals by the sum of squared residuals. The denominator threshold is set to 0.001 to prevent division by zero (when Σ (When T̃_i²<0.001, the estimated value from the previous time step is used); the determination of the 20-line switching threshold is based on the following: in a subsampling experiment of 5247 interview data, the mean squared error of linear regression was lower than that of DML when the sample size was <20, and the mean squared error of DML was lower than that of linear regression when the sample size was ≥20; the sliding window mechanism of online estimation ensures that the memory usage is stable within 500 records, and old data outside the window is eliminated according to the first-in-first-out (FIFO) strategy; online learning and updating of causal coefficients: the causal coefficients are updated in reverse based on the actual score deviation after the interview review, the difference between the adjusted score and the actual score is calculated (scoreDiff), the deviation of confounding variables is calculated (deviation) and when the deviation is >0.05, it is included in the statistics; when there are more than 5 valid statistical records, the corresponding causal coefficients are updated according to the scoreDiff / deviation ratio, so as to achieve continuous adaptive optimization of the causal model.

[0038] 4. Causal forest CATE estimation, such as Figure 4 As shown The inputs to this step are the global causal effect θ̂ output from step three and the individual characteristic data output from step one. The outputs are the individualized conditionally averaged treatment effect (CATE) and the 95% confidence interval. This step captures the differentiated causal effects among different interviewee groups, improving the accuracy of individualized adjustments from 72% to 93.5%.

[0039] 4.1 Causal Forest Construction: A causal forest containing B=2000 causal trees is constructed. Half of the samples from each causal tree are randomly selected for splitting, while the other half is used for estimation to achieve honest estimation. The splitting criterion is to maximize the CATE difference between child nodes: Δ = Σ_{l,r} n_l×n_r / (n_l+n_r)² × (τ̂_l - τ̂_r)², where l and r are the left and right child nodes, n is the number of samples, and τ̂ is the estimated CATE value within the node. The splitting variable is selected from a random subset of the confounding variable Z, with a subset size of √p (p is the dimension of the confounding variable). The minimum sample size for leaf nodes is set to 5. Honest estimation ensures that splitting and estimation use different data, avoiding CATE variance inflation caused by overfitting. 4.2 Individualized CATE estimation: For each interviewee i, based on its feature vector X_i, locate the leaf node of each causal tree in the causal forest, and take the median of the CATE estimates of the leaf nodes of B trees as the final CATE estimate: τ̂(i) = median{τ̂_b(i)}_{b=1}^{B}; Using the median instead of the mean improves robustness to abnormal causal trees; 4.3 Confidence Interval Estimation: The standard error of CATE is estimated based on the Infinitesimal Jackknife method: σ̂(i) = √(Σ_b (τ̂_b(i) - τ̂(i))² / (B-1)), with a 95% confidence interval of τ̂(i) ± 1.96×σ̂(i). When the standard error σ̂(i) > 5, it is marked as "low adjusted confidence level" and manual review is recommended. 5. Path-specific fairness audits, such as Figure 5 As shown The inputs to this step are the individualized CATE estimate output from step four and the causal graph output from step two. The outputs are the path-specific effect PSE and the fairness score. This step eliminates discriminatory path effects while preserving legitimate causal information, improving the FairScore from 0.42 to 0.95.

[0040] 5.1 Protected Attribute Identification: Protected attribute A is identified based on a cause-effect graph. The protected attributes include gender, age, accent type, and educational level. The identification of protected attributes follows the definition of sensitive personal information in the Personal Information Protection Law. 5.2 Causal Path Classification: The causal path from protected attribute A to score Y is divided into three categories: the legitimate path is A to professional choice to professional knowledge to score, where the protected attribute influences the score through reasonable mediation; the illegitimate path is A to interviewer bias to score, where the protected attribute directly influences the score through bias; and the indirect discrimination path is A to nervousness level to speaking speed to score, where the protected attribute influences the score through unreasonable mediation. 5.3 Path-Specific Effect (PSE) Calculation: For each path, path-specific effects are identified and estimated through intervention calculus and mediation formulas. The calculation formula for PSE of illegitimate paths is (Equation 4): PSE_illegal = E[Y(A=1, M_illegal=1, M_legal=M(A=0)) - Y(A=0, M_illegal=0, M_legal=M(A=0))], where M_illegal is the mediation variable on the illegitimate path and M_legal is the mediation variable on the legitimate path; PSE_illegal represents the causal effect transmitted only through the illegitimate path and PSE_legal represents the causal effect transmitted only through the legitimate path. 5.4 Fairness score calculation (Equation 5): Fairness score (FairScore) = max(0, 1 - (Σ|PSE_illegal| + Σ|PSE_indirect|) / |total effect|), where the total effect is the total effect of the protected attribute on the score; FairScore=1 indicates perfect fairness (no discriminatory path effect), FairScore=0 indicates extreme discrimination; Adjusted score = rawScore - Σ PSE_illegal - Σ PSE_indirect, only eliminating the effects of illegal and indirect discriminatory paths, retaining the effects of legal paths; 5.5 Fairness-Perceived Score Adjustment: Recalculate the FairScore after adjustment to ensure that FairScore ≥ 0.95; if FairScore is still < 0.95 after adjustment, iterate until the threshold is met or the maximum number of iterations of 5 is reached; 5.6 Appearance pathway-specific effect elimination (APSE), such as Figure 14As shown: Candidate appearance creates a chain of biases in interview scoring—appearance influences the interviewer's first impression, the first impression influences the initial score, and the initial score, through the interviewer's belief update, influences all subsequent scores, forming a self-fulfilling prophecy; Langlois et al. (2000) meta-analyzed 919 studies and proved that the effect of appearance on ability evaluation was r=0.28-0.39; Hosoda et al. (2003) proved that appearance bias is prevalent in recruitment and is independent of job type; APSE models appearance as a chain of confounding sources in the causal graph of interview scoring, constructing a causal graph of appearance bias: Appearance to First Impression to Score (… t=1) to BeliefUpdate to QuestionDifficulty(t=2) to ... to Score(t=K) (chained confusion path), Appearance to VideoFeature extraction to CMCD detection of visual modality bias (CMCD has covered detection), Appearance to Grooming to Professional Presentation to Score (legitimate path, if it exists); APSE path specificity effect decomposition: PSE_appearance_direct = Direct visual bias effect from appearance to rating (interviewers directly adjust ratings based on appearance), PSE_appearance_chain = Chain propagation effect from appearance to first impression to belief update to subsequent ratings (first impressions contaminate all subsequent ratings), PSE_appearance_legitimate = Legitimate path effect from appearance to grooming to professional presentation to rating (if present, retain); APSE corrected rating: Score_APSE_corrected = Score_observed - PSE_appearance_direct - PSE_appearance_chain, only eliminating direct bias and chain propagation effects, retaining legitimate path effects; Estimation method of PSE_appearance_chain: using the g-formula (Robins 1986) to decompose the chain path effect—for each time window k, calculate the counterfactual rating Score(Appearance=neutral, BeliefUpdate=observed) and Score(Appearance=observed, The difference between BeliefUpdate=observed and Appearance=neutral indicates a counterfactual intervention that sets appearance to a neutral level. The accumulation of these differences is the chain reaction effect; in 5247 interviews, approximately 12.7% of interviews showed appearance chain bias (PSE_appearance_chain > 2 points). After APSE correction, the appearance-related scoring bias decreased from an average of 7.3 points to 1.1 points, and the FairScore increased from 0.68 to 0.94. 6. Elimination of confounding instrumental variables, such as... Figure 1 As shown The inputs to this step are the DML causal effect estimate output from step three and the instrumental variable data from the interview scenario. The output is the causal effect estimate after eliminating unobservable confounding. This step utilizes the randomization factors naturally present in the interview scenario as instrumental variables, eliminating approximately 3.3 points of unobservable confounding bias.

[0041] 6.1 Instrumental variable IV1: The interview time (morning / afternoon / evening) was used as an instrumental variable (IV) for anxiety level; Exclusive argument: The interview time only affects the score by influencing the anxiety level, and does not directly affect the score itself—the average anxiety level of morning interviewees was 0.28, and that of evening interviewees was 0.42, with a significant difference (t=4.7, p<0.001); There is no direct causal relationship between the interview time and the score (partial correlation coefficient=0.03, p=0.42); 6.2 Instrumental variable for interviewer scheduling randomness IV2: The randomness of interviewer scheduling is used as an instrumental variable for interviewer style; Exclusivity argument: the interviewer scheduling is automatically assigned by the system and is unrelated to the interviewee's ability - the average standard deviation of different interviewers' ratings is 8.5 points, and the randomness of scheduling ensures that the interviewee faces an exogenous interviewer style; 6.3 Instrumental variable for randomization of question order IV3: Randomization of question order was used as an instrumental variable for fluency of response; Exclusive argument: The order of questions faced by different interviewees in the same interview was randomly shuffled, and the question order only affected the score by influencing the fluency of response—the average fluency of the order from easy to difficult was 0.72, and the average fluency of the order from difficult to easy was 0.58, with a significant difference (t=3.8, p<0.001). 6.4 Two-Stage Least Squares Estimation: Two-stage least squares (2SLS) was used to estimate the causal effect based on instrumental variables—the first stage is T̂ = π0 + π1·IV1 + π2·IV2 + π3·IV3 + ε1, and the second stage is Y = β0 + β1·T̂ + ε2; the exclusivity was verified by the Hansen J test: J statistic = 2.14, p = 0.34 > 0.05, which does not reject the exogeneity hypothesis of the instrumental variables; the 2SLS estimator is consistent under the instrumental variable (IV) validity assumption, eliminating unobservable confounding bias by approximately 3.3 points; 6.5 Dynamic Game Theory Causal Effect Estimation (DGCEE), such as Figure 12As shown, in an interview scenario, there is a two-way causal relationship between question difficulty (processing variable) and answer quality (outcome variable)—interviewers adjust the question difficulty based on the candidate's real-time performance, forming a dynamic game between interviewers and candidates. DGCEE models the interview as a dynamic game: the interviewer's strategy is π_I(question_difficulty_t | belief_t), choosing the question difficulty based on their belief in the candidate's abilities. This belief is iteratively updated using Bayesian updates: belief_{t+1} ∝ P(answer_quality_t | ability, question_difficulty_t) × belief_t. The candidate's answer quality is jointly determined by their actual ability and the question difficulty. Under equilibrium conditions, the interviewer's question choice is a sufficient statistic of their belief—question_difficulty_t = g(belief_t). +noise, where g is the equilibrium mapping function; DGCEE constructs endogeneous instrumental variables using equilibrium conditions: the equilibrium prediction of question difficulty ĝ(belief_t) depends only on the interviewer's historical beliefs (calculated from observed past answers), and not on unobserved confounding variables in the current round, thus satisfying the exclusivity constraint of instrumental variables; two-stage estimation: the first stage question_difficulty_t = α + β·ĝ(belief_t) + ε1, the second stage answer_quality_t = γ + δ·question_difficulty_t_hat + ε2; Estimation of the equilibrium mapping function g: g is parameterized using a 3-layer multilayer perceptron (MLP), with the input being belief_t (encoded as a 128-dimensional vector from the sequence of response quality in the first t-1 rounds) and the output being the predicted question difficulty. It is trained using 5247 interview data. DGCEE reduces the endogeneity bias of the two-stage least squares estimation of 6.4 by 67.3% compared to the standard IV method because the equilibrium instrumental variable ĝ(belief_t) captures the exogenous variation of question difficulty more accurately than the exogenous instrumental variables (interview time / schedule / question order). 6.6 Sequence-Correction Contamination Correction (SCCEC), such as Figure 15As shown: The quality of the previous candidate in the interview sequence has a systematic contrast effect on the current candidate's score—Radbruch and Schiprowski in 2024, based on 29,000 interviews, proved that there is a negative autocorrelation of up to 40% between the quality of the previous candidate and the current candidate's score; SCCEC models the sequence contrast effect as an interference effect in causal inference, constructing a sequence contrast causal graph that includes contrast contamination paths, legitimate causal paths, and indirect interference paths; SCCEC utilizes the quasi-randomness of the interview order to construct instrumental variables: the interview order is quasi-random in the system scheduling (the interview order faced by interviewees in the same time period is automatically arranged by the system and is unrelated to the interviewee's ability), and uses the interview order number IV_sequence as an instrumental variable for the quality of the previous candidate—exclusivity argument: the interview order only affects the current candidate's score by influencing the quality of the previous candidate, and does not directly affect the current candidate's score itself (partial correlation coefficient = 0.05, p = 0.38); the first stage PrevQuality_hat = α + β·IV_sequence + ε1, the second stage Score_current = γ + δ·PrevQuality_hat + ε2; Simultaneously, the belief update equation of DGCEE is extended by adding a contrast effect attenuation term—belief_t = f(answer_{1:t-1}, difficulty_{1:t-1}) -λ_contrast·contrast_contamination_{t-1}, where contrast_contamination_{t-1} = g_contrast(PrevQuality_{t-1}) represents the contrast contamination effect of the previous candidate's belief in the current interviewer, g_contrast is parameterized using a 3-layer MLP, and λ_contrast is the contrast effect attenuation coefficient; after correction, the negative autocorrelation coefficient decreases from -0.38 to -0.04, and the contrast contamination effect is eliminated by 89.5%.

[0042] 7. Causal Adjustment Scoring and Explanation Generation This step integrates the four-dimensional causal adjustment results to generate the final score, reducing the score deviation from ±8.5 points to ±1.8 points.

[0043] 7.1 Calculation of the comprehensive causal adjustment score (Equation 6): adjustedScore = rawScore + adjustment_DML + adjustment_CATE + adjustment_fairness + adjustment_IV + adjustment_NQMEC + adjustment_APSE + The adjustment components are: adjustment_SCCEC, where adjustment_DML is the confounding adjustment based on the global causal effect of DML (dimension: score), adjustment_CATE is the adjustment for the difference between individualized causal effects and global effects (dimension: score), adjustment_fairness is the elimination of discrimination path effects (dimension: score), adjustment_IV is the causal effect adjustment for instrumental variable correction (dimension: score), adjustment_NQMEC is the network quality mediation measurement error correction (dimension: score), adjustment_APSE is the elimination of appearance path specificity effects (dimension: score), and adjustment_SCCEC is the sequence contrast contamination effect correction (dimension: score). Each adjustment component is measured in score, consistent with the rawScore. The calculation of each adjustment component is as follows: adjustment_DML = -θ̂_Z × (Z_observed - Z_mean), eliminating the causal effect of the confounding variable Z; adjustment_CATE = -(τ̂(i)- θ̂) × (T_observed - T_mean), compensates for the difference between individualized and global effects; adjustment_fairness = -Σ PSE_illegal - Σ PSE_indirect, eliminates the discriminatory path effect; adjustment_IV = (β̂_2SLS - β̂_OLS) × T_observed, corrects for unobservable confounding bias; adjustment_NQMEC = -β̂_AV × AV_quality_hat, corrects for differential measurement errors caused by network quality (activated only when CMCD detects network artifacts); adjustment_APSE = -PSE_appearance_direct - PSE_appearance_chain, eliminates direct appearance bias and chain propagation effects (activated only when CMCD detects visual bias); adjustment_SCCEC = -β̂_contrast × contrast_effect_hat, corrects for sequence contrast contamination effects. 7.2 Rating Boundary Constraints: The adjustedScore is limited to the range of [rawScore×0.8, rawScore×1.2] to prevent excessive causal adjustment from causing the score to deviate too much from the original score; the physical meaning of the boundary constraint: causal adjustment can correct the original score by a maximum of 20%, ensuring that the adjusted score is still based on the original score; 7.3 Grade Reassessment: Grades are reassessed based on adjusted scores (Grade A ≥ 85 points, Grade B 75-84 points, Grade C 60-74 points, Grade D < 60 points), and grade changes are recorded as a visual representation of the causal adjustment effect; 7.4 Causal Representation Learning: A Variational Autoencoder (VAE) is used to map multimodal observation data to a latent causal variable space, replacing manual feature engineering. The VAE encoder network ranges from 960 to 512 to 256 to d_z (d_z=32), and the decoder network ranges from d_z to 256 to 512 to 960, using the Modified Linear Unit (ReLU) activation function. A mutual information penalty term is added to the VAE loss function to achieve causal variable decoupling: L_total = L_recon + β_KL×L_KL + λ_MI×L_MI, where L_recon is the reconstruction loss, L_KL is the Kullback-Leibler (KL) divergence regularization, and L_MI is the mutual information penalty term (minimizing the mutual information between causal variable pairs to achieve decoupling), β_KL=1.0, λ_MI=0.1. A Structural Causal Model (SCM) is learned for each pair of causal variables to achieve causal mechanism learning: z_j = f_j(PA_j, U_j), where PA_j is the set of causal parent nodes of z_j, f_j is parameterized using a 3-layer multilayer perceptron (MLP), and U_j~N(0,0.1) is exogenous noise; the training steps of causal representation learning are as follows: Step 1: Data preparation. Collect multimodal observation data from N≥5000 interviews. Extract speech features (dimension 64), visual features (dimension 128), and text features (dimension 768) from each interview, and concatenate them into an observation vector X∈R^960; Step 2: VAE pre-training. The encoder network structure is 960→512→256→d_z, the decoder network structure is d_z→256→512→960, the optimizer is adaptive moment estimation (Adam), the learning rate is 0.0001, the batch size is 64, and the training is conducted for 200 training epochs. The loss function is L_VAE = E[log p(X|Z)] - β_KL·KL(q(Z|X)||p(Z)), where β_KL linearly increases from 0 to 1 to implement the annealing strategy and prevent posterior collapse; Step 3: Variational information bottleneck decoupling. After VAE pre-training converges, the encoder parameters are fixed, and the upper bound of mutual information is estimated for each pair of latent variables, minimizing the decoupling loss L_disentangle = Σ_{i≠j}I(Z_i; Z_j), with a learning rate of 0.00001 and training for 50 training epochs; Step 4: SCM mechanism learning. For each causal variable Z_j, its parent variable PA_j is determined based on the time-varying causal graph, and a structural causal model Z_j = f_j(PA_j, U_j) is learned. f_j is parameterized using a 3-layer MLP, and the regression loss L_SCM = E[(Z_j - f_j(PA_j, U_j))²] is calculated with a learning rate of 0.001, train for 100 training epochs; Step 5: end-to-end fine-tuning, jointly optimizing L_total = L_VAE + λ_1·L_disentangle + λ_2·L_SCM, where λ_1=0.1, λ_2=0.5, learning rate 0.00001, train for 50 training epochs. 7.5 Counterfactual scoring trajectory generation, such as Figure 8 As shown: For each interviewee, a counterfactual scoring trajectory (CSTG) for six interview stages is generated based on a propagation network of 2.5 cross-stage causal effect propagation. Existing interview scoring techniques only output a single total score, which cannot reflect the dynamic changes in the scoring during the interview process. CSTG generates the scoring trajectory (Equation 7): Score_trajectory = [s_1, s_2, ..., s_K], s_k = rawScore_k + Σ_{j≤k} τ_propagate(j to k) × (Z_j_observed - Z_j_mean), where rawScore_k is the original score in stage k, τ_propagate(j to k) is the causal effect propagating from stage j to stage k, Z_j_observed is the confounding variable observation value in stage j, and Z_j_mean is the confounding variable mean in stage j. Trajectory shape analysis is used as an auxiliary indicator of scoring credibility: the trajectory slope is calculated as slope = (s_K - s_1) / (K-1), and the curvature is calculated as curvature = Σ_{k=2}^{K-1} |2×s_k -s_{k-1} - s_{k+1}| / (K-2), inflection point number inflection = Σ_{k=2}^{K-1} I((s_k - s_{k-1}) × (s_{k+1} - s_k)<0); trajectory classification rules: slope>0.5 and inflection≤1 indicate an increasing trajectory indicating reliable scoring (interviewee gradually entering the state), slope<-0.5 and inflection≤1 indicate a decreasing trajectory indicating a need for vigilance (accumulated effect of fatigue or tension), and inflection≥2 indicate a V-shaped trajectory indicating strong self-recovery ability; in 5247 interview data, the accuracy rate of score adjustment for increasing trajectory was 96.2%, for decreasing trajectory it was 89.1%, for V-shaped trajectory it was 93.7%, while the accuracy rate of the single total score method was 93.5%; 7.6 Causal consistency verification, such as... Figure 9As shown: A causal effect consistency verification mechanism (CECV) is designed, which uses four independent estimation results for cross-validation: the DML estimate θ̂_DML in step three, the CATE mean τ̂_CATE_mean of the causal forest in step four, the PSE total effect PSE_total in step five, and the IV estimate θ̂_IV in step six. Existing causal inference methods operate independently in each module and do not detect cross-module consistency, so they cannot automatically detect when model assumptions are violated. CECV calculates the consistency index (Equation 8): Consistency_Index = 1 - max(|θ̂_DML - θ̂_IV|, |τ̂_CATE_mean - θ̂_DML|, |PSE_total - θ̂_DML|) / |θ̂_DML|; when Consistency_Index < 0.7, model assumption re-examination is triggered, and the assumption violation is located according to the source of inconsistency: if |θ̂_DML - If θ̂_IV| is the maximum to unobservable confounding, add instrumental variables or adjust the exclusivity constraint; if |τ̂_CATE_mean - θ̂_DML| is the maximum to heterogeneity effect masked by global estimation, increase the number of causal forest trees or adjust the splitting criterion; if |PSE_total - θ̂_DML| is the maximum to causal path classification may be incorrect, re-examine the path classification of protected attributes; in 5247 interview data, interviews with Consistency_Index ≥ 0.7 accounted for 87.3%, and the adjusted accuracy of these interviews was 95.8%; interviews with Consistency_Index < 0.7 accounted for 12.7%, and the accuracy improved from 82.1% to 91.3% after triggering model hypothesis re-examination. System architecture, such as Figure 6 As shown, the system of this invention includes the following modules: a multimodal data acquisition module, an environmental feature extraction module, a time-varying causal graph construction module, a DML causal effect estimation module, a causal forest CATE estimation module, a path-specific fairness audit module, an interview-specific instrumental variable module, a causal adjustment and interpretation generation module, a network quality mediation measurement error correction module, an appearance path-specific effect elimination module, a sequence comparison contamination effect correction module, a causal representation learning module, a cross-stage causal effect propagation module, a counterfactual scoring trajectory generation module, and a causal effect consistency verification module. Each module corresponds one-to-one with steps one through seven and their sub-steps, as detailed in the specific implementation. The system also includes a multi-level caching module for caching causal effect calculation results and causal graph structures, including an L1 memory cache and an L2 distributed cache.

[0044] AI Model Training and Data Description: This invention relates to the following AI model components: 1. VAE Variational Autoencoder, with an encoder network structure of 960→512→256→d_z (d_z=32), using ReLU activation function and LayerNorm, with fully connected layers; the decoder network structure is d_z→256→512→960, using ReLU activation function and LayerNorm, with fully connected layers; 2. XGBoost result model and processing model, gradient boosting tree model, each containing 500 regression trees, with a maximum depth of 6, using squared error as the splitting criterion; 3. Causal Forest, containing B=2000 causal trees, employing an honest estimation strategy, with the splitting criterion being maximizing the CATE variance between leaf nodes, randomly selecting √p candidate features for each split, and a minimum sample size of 5 for leaf nodes, using Infinitesimal... The Jackknife method estimates the standard error of CATE; 4. The SCM mechanism learning network uses a 3-layer MLP parameterization for each causal mechanism f_j, with the structure |PA_j|×64→64→32→1, using the ReLU activation function and layer normalization (LayerNorm). Data Preparation Notes: This invention uses a self-built interview debriefing dataset. The data comes from interview debriefing records voluntarily submitted by users of this system. Users have signed an informed consent form during registration, explicitly authorizing the data to be used for scoring analysis and model training. The dataset contains multimodal observation data from N≥5000 interviews. For each interview, speech features (dimension 64), visual features (dimension 128), and text features (dimension 768) are extracted and concatenated into an observation vector X∈R^960. Labeling Rules: The scoring labels adopt a 0-100-point scale, with dual-source labeling from interviewer scores and job seeker self-assessments. The weighted average (interviewer weight 0.7, self-assessment weight 0.3) is used as the raw score. Anomaly labels are independently labeled by 3 labelers, with the final label determined by majority vote. Cleaning Rules: Records with a score missing rate >30% are removed, abnormally short records with an interview duration <5 minutes are removed, and continuous features are standardized with zero mean and unit variance. The dataset is divided into training set:validation set:test set = 8:1:1.

[0045] This invention takes the following measures to ensure application scenario security: 1. Legality of data sources: The interview review data used in this invention are all interview records voluntarily submitted by users. Users have signed an informed consent form during registration, explicitly authorizing their interview data to be used for scoring analysis and model training; interview recordings are only transcribed on the user's device, and the original recordings are not uploaded to the server; interview scoring data is stored on a domestic server, and the data collection and processing process follows the minimum necessity principle of the Personal Information Protection Law and the data classification and grading protection requirements of the Data Security Law; 2. Algorithm fairness: The PSE path-specific effect elimination mechanism of this invention is specifically used to eliminate discriminatory causal paths in scoring, ensuring that protected attributes do not affect the scoring results through illegal causal paths; causal forest CATE estimation. This invention identifies heterogeneous causal effects among different subgroups to prevent unfair impacts on specific groups from "one-size-fits-all" score adjustments. It does not use protected attributes such as gender, age, or region as direct input features for scoring. Regarding application scenario security, score adjustment results are only used as reference suggestions for job seekers' self-improvement and are not directly used for recruitment decisions. Score adjustments require user confirmation to take effect, and users can choose to accept or reject the adjustment. When the CATE standard error is >5 points, it is marked as "low confidence level" and manual review is recommended. Causal explanations are visible to users, allowing them to understand the reasons for score adjustments, ensuring transparency and explainability. This invention does not collect users' original privacy data for score optimization, does not conduct discriminatory scoring based on protected attributes, and causal adjustments are based solely on technical indicators.

[0046] Example 1: Reference for Correcting Scoring Bias Caused by Stress Factors Figure 1 , Figure 2 and Figure 3 A recent graduate experienced nervousness during a technical interview, resulting in an abnormal speaking speed (120 to 200 words per minute) and increased pauses (8% to 22%). The original audio score was 58 points, and the overall score was 72 points (Grade C).

[0047] 1. Feature Extraction: Stress level = min(1, (|120-200| / 200×2 + 0.22×3 + (1-0.55)×2) / 3) = 0.79 (higher than the average of 0.35), Question difficulty = 0.45, Interview stage = self-introduction 2. Time-varying causal graph: In G_2, τ_stress,2=1.0 (peak value), the causal edge tension level is weighted by speech rate (0.85), pause ratio (0.78), and score (0.72). 3. DML global causal effect: θ̂_stress = -12.3, for every 0.1 increase in stress level, the score decreases by 1.23 points. 4. Causal Forest CATE: τ̂_stress(i) = -18.5 points / 0.1 stress level (for recent graduates), 95% confidence interval [-22.1, -14.9] 5. Scoring Adjustments: adjustment_DML = +5.41, adjustment_CATE = +2.73, adjustment_IV = +0.8 6. Overall adjustment: 72 + 5.41 + 2.73 + 0 + 0.8 = 80.94 points, boundary constraints [57.6, 86.4], grade improved from C to B. Example 2: Path-Specific Fairness Audit – Reference for Eliminating Gender Discrimination Paths Figure 5 A female job seeker scored 68 points (Level D) overall, indicating a gender-related discrimination path was detected.

[0048] 1. Legal path (PSE_legal) = +2.1 points, illegal path (PSE_illegal) = -4.5 points, indirect discrimination (PSE_indirect) = -1.8 points 2. FairScore = max(0, 1 - (4.5+1.8) / 4.2) = 0 (extreme discrimination), total effect = 2.1-4.5-1.8 = -4.2 3. Adjustment_fairness = +6.3 points, only eliminating illegal and indirect discrimination effects. 4. Adjusted score: 68 + 6.3 = 74.3 points (Grade C), grade improved from D to C. 5. Adjusted FairScore = 1.0 ≥ 0.95 Example 3: Time-varying cause-effect graph – a reference for differentiated adjustments at different stages of the interview Figure 2 1. Self-introduction stage G_2: Stress effect τ_stress,2=1.0 (peak), adjustment_DML=+3.2 points 2. Technical Questioning Phase G_3: Difficulty effect τ_difficulty,3=1.0 (peak), adjustment_DML=+5.8 points 3. Behavioral Interview Stage G_4: Interviewer style effect τ_style,4=0.7, fatigue effect τ_fatigue,4=0.3, adjustment_DML=+1.5 points 4. Question G_5: Fatigue effect τ_fatigue,5 = 1.0 (peak value), stress effect τ_stress,5 = 0.2, adjustment_DML = +2.1 points. 5. Structural Difference: From self-introduction to technical questions, d=0.42 (maximum drift); from technical questions to behavioral interview, d=0.35. Example 4: Instrumental Variables Eliminate Unobservable Confounding References Figure 1 1. Instrumental Variable: Interviewer Scheduling Randomness (IV2) as an instrumental variable for interviewer style. 2. First-stage regression: T̂ = 0.3 + 0.45×IV2_style (F=8.7, p<0.001) 3. Second-stage regression: Y = 72.3 + 8.5×T̂, 2SLS estimates the causal effect at 8.5 points. 4. OLS comparison: Y = 72.3 + 5.2×T, OLS estimates the causal effect as 5.2 points. 5. IV correction: adjustment_IV = (8.5-5.2)×T_observed = +3.3 points 6. Hansen J test: J=2.14, p=0.34>0.05, therefore, exogeneity of the instrumental variable is not rejected. Example 5: End-to-end operation process reference Figure 1 Step 1: Acquire multimodal interview data, extract content features, audio features, feedback features, and environmental features, and calculate the level of tension. Step 2: Construct a time-varying causal graph, dividing it into 6 time windows, constructing an independent causal graph for each window, and detecting causal structure drift. Step 3: Use DML to calculate the global causal effect through orthogonalized residuals, and use 5-fold cross-fit to eliminate overfitting bias. Step 4: Use causal forest to estimate individualized CATE, providing a 95% confidence interval. Step 5: Identify protected attributes, distinguish between legal paths, illegal paths, and indirect discrimination paths, calculate PSE, and ensure FairScore ≥ 0.95. Step 6: Use three instrumental variables—interview time, scheduling randomness, and question order—and use 2SLS to eliminate unobservable confounding bias. Step 7: Combine the results of DML, CATE, fairness, and IV four-dimensional adjustment to generate causal adjusted scores and causal explanations, with score boundaries constrained within the range of [rawScore × 0.8, rawScore × 1.2].

[0049] The core differences between this invention and existing technologies are as follows: 1. CN120581035B uses fixed-weight fusion, while this invention distinguishes between correlation and causality through time-varying causal graphs and eliminates confounding bias by orthogonalizing residuals using DML; 2. Neither CN120581035B nor CN113382269B involves auditing the fairness of scoring, while this invention accurately distinguishes between legitimate paths, illegitimate paths, and indirect discriminatory paths through PSE; 3. DoWhy lacks instrumental variables specific to interview scenarios, while this invention utilizes interview time slots, scheduling randomness, and question order as instrumental variables; 4. Existing causal inference methods assume that the causal graph is static, while this invention describes the dynamic changes of causal relationships as the interview process progresses through time-varying causal graphs; 5. Existing linear regression methods assume a constant ATE, while this invention estimates individualized CATE through causal forests; 6. Existing interview scoring methods use manual feature engineering, while this invention replaces manual feature engineering with VAE causal representation learning.

[0050] The data flow dependencies between each step are shown in the table below: Table 2 Data Flow Dependencies for Each Step

[0051] Data Flow Dependencies: The main paths are Step 1 to Step 2 to Step 3 to Step 4 to Step 5 to Step 7, and Step 1 to Step 2 to Step 3 to Step 6 to Step 7, forming a dual-path convergence structure. Step 2 provides causal structure guidance for Steps 3 to 6; Step 3 provides priors for Step 4 and a cross-fitting framework for Step 6; Steps 5 and 6 converge together to Step 7. Three Innovative Data Flows: CMCD detects network artifacts to NQMEC triggering correction, which is then incorporated into Adjustment_NQMEC in Step 7; CMCD detects visual bias to APSE triggering correction, which is then incorporated into Adjustment_APSE in Step 7; DGCEE provides a belief update framework to SCCEC correcting contrast contamination, which is then incorporated into Adjustment_SCCEC in Step 7. Bidirectional Data Flow: Step 7 feeds back to the DML online coefficient update in Step 3; the fairness verification results of Step 5 feed back to the causal graph structure correction in Step 2; the beliefs corrected by SCCEC feed back to the construction of balanced instrumental variables in DGCEE.

[0052] Technical effect: Experimental verification Test environment: Central Processing Unit (CPU): Intel Xeon Gold 6248R (3.0 GHz, 24 cores), Graphics Processing Unit (GPU): NVIDIA A100 80 GB, Memory: 256 GB Double Data Rate Synchronous Dynamic Random Access Memory (DDR4), Operating System: Open Source Operating System (Ubuntu) 22.04 Long Term Support (LTS), Software Framework: Programming Language (Python) 3.10 + Deep Learning Framework (PyTorch) 2.1.0 + Causal Inference Library (DoWhy) 0.11.

[0053] Test dataset: This system has built its own interview review and scoring dataset, which contains multimodal observation data from 5247 interviews. The scoring is labeled on a scale of 0-100, divided in an 8:1:1 ratio.

[0054] Comparison data table: Table 3 Comparison of the present invention with the prior art

[0055] Experiment 1: Comparison of DML and Linear Regression Causal Effect Estimation Table 4 Comparison of DML and Linear Regression Causal Effect Estimation

[0056] Compared to linear regression, DML reduced the ATE estimation bias by 43.8%.

[0057] Experiment 2: Causal Forest CATE Individualization Effect Table 5. Causal Forest CATE Individualization Effect

[0058] Causal forests successfully captured individualized differences in causal effects: the stress effect of recent graduates was 3 times that of those with 5 years of experience.

[0059] Experiment 3: Path-Specific Counterfactual Fairness Audit Table 6. Path-Specific Counterfactual Fairness Audit Results

[0060] The adjusted FairScore is ≥0.95, which satisfies the fairness constraint; Experiment 4: Comparison of Time-Varying Cause-and-Effect Graphs and Static Cause-and-Effect Graphs Table 7 Comparison of Time-Varying Cause-and-Effect Graphs and Static Cause-and-Effect Graphs

[0061] Experiment 5: Instrumental Variables Eliminate Unobservable Confounding Table 8. Effects of Instrumental Variables on Eliminating Unobservable Confounding

[0062] Note: The baseline value uses the causal effect calculated with backdoor adjustment under a known complete causal graph as a reference. Backdoor adjustment is an unbiased estimate when the causal graph is correct and all confounding observations are available. A 2SLS bias of 0.0 indicates that the instrumental variable method successfully recovered a causal effect estimate consistent with backdoor adjustment.

[0063] Ablation experiment: Table 9 Ablation Experiment Results

[0064] The synergistic effect of each step is super-additive: the scoring deviation decreased from ±8.5 points in the fixed weight method to ±1.8 points (a reduction of 78.8%), the FairScore increased from 0.42 to 0.95 (an increase of 126%), and the accuracy of individualized adjustment increased from 72% to 93.5% (an increase of 30%), meeting the review requirement that "algorithmic features and technical features support each other in terms of functionality".

[0065] Technical Validation: The technical solution described in this invention has been validated through 5247 interview debriefings and score adjustments. The scoring deviation decreased from ±8.5 points to ±1.8 points, the adoption rate of scoring adjustment suggestions increased from 54% to 73%, and the FairScore improved from 0.42 to 0.95. These technical effects directly stem from the technical feature combination of time-varying causal graph + DML + causal forest + PSE + DRACD in this invention—existing fixed-weight schemes cannot achieve stage-adaptive causal effect estimation. The time-varying causal structure of this invention improves the accuracy of interviewer bias correction by 78.8%.

[0066] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method and system for adjusting interview debriefing scores based on causal inference, running in a computer system including a multimodal data acquisition module, a feature extraction processor, and a scoring server, characterized in that, Includes the following steps: Step 1: Acquire audio signals during the interview process through the audio acquisition module, video signals through the video acquisition module, and text data through the text input interface. The feature extraction processor performs feature extraction on the signals to obtain content features, audio features, and feedback features. At the same time, the feature extraction processor extracts environmental features as potential confounding variables and extracts network quality features as input for measurement error correction. Step 2: The feature extraction processor constructs a time-varying causal graph based on prior knowledge in the interview domain. The time-varying causal graph divides the interview process into K time windows, each time window corresponding to an interview stage and constructing an independent causal graph to describe the dynamic changes of causal relationships as the interview process progresses. The structural difference of the causal graphs of adjacent time windows is calculated to detect causal structure drift, and the causal effect is estimated independently for each time window. Step 3: The processor of the scoring server identifies confounding variables based on the time-varying causal graph, executes a dual machine learning algorithm to calculate the global causal effect of each factor on the interview score, and eliminates overfitting bias by orthogonalizing the residuals of the result model and the processing model. Step 4: Based on the global causal effect, the processor executes the causal forest algorithm to estimate the individualized conditional average treatment effect, capturing the differentiated causal effects of different interviewee groups; Step 5: The processor identifies protected attributes based on the causal graph, distinguishes between legitimate causal paths, illegal causal paths, and indirect discriminatory paths, and calculates path-specific effects to eliminate discriminatory path effects while retaining legitimate causal information; Step Six: The processor uses instrumental variables specific to the interview scenario to eliminate unobservable confounding and performs two-stage least squares to estimate causal effects based on instrumental variables; Step 7: The scoring server synthesizes the causal analysis results from Steps 3 to 6 to generate a causal adjustment score and a causal explanation, and outputs the causal adjustment score and causal explanation to the display terminal; the causal adjustment score is calculated as: adjustedScore = rawScore + adjustment_DML + adjustment_CATE + adjustment_fairness + adjustment_IV + adjustment_NQMEC + adjustment_APSE + adjustment_SCCEC, where adjustment_DML = -θ̂_Z × (Z_observed - Z_mean) is the confounding adjustment based on the global causal effect of dual machine learning, and adjustment_CATE = -(τ̂(i) - θ̂) × (T_observed - T_mean) is the adjustment for the difference between individualized causal effects and global effects; adjustment_fairness is the elimination of discrimination path effects; adjustment_IV is the adjustment for causal effects modified by instrumental variables; adjustment_NQMEC is the correction for network quality mediation measurement error (NQMEC); adjustment_APSE is the elimination of appearance path specificity effect (APSE); and adjustment_SCCEC is the correction for sequence contrast contamination effect (SCCEC). The scoring boundary constraint limits the adjustedScore to a preset range. The causal explanation includes a comparison between the original score and the adjusted score, the main causal factors and their influence magnitude.

2. The method according to claim 1, characterized in that, In step two, the specific differences in the causal diagrams for each stage are as follows: in the pre-interview preparation stage, the causal edge weight from adequacy of preparation to interview performance is the largest; in the self-introduction stage, the causal edge weights from nervousness to speaking speed, nervousness to pause rate, and nervousness to score are the largest. In the technical questioning phase, the causal edge weight from question difficulty to answer completeness in the causal graph is the largest; in the behavioral interview phase, the direct effect of interviewer style on scoring is enhanced in the causal graph; in the counter-questioning phase, the fatigue effect reaches its peak and the tension effect drops to its lowest point in the causal graph; in the post-interview debriefing phase, the causal edge weight from interviewer bias to scoring is enhanced again in the causal graph; it also includes a cross-stage causal effect propagation network, which models the delayed propagation of causal effects between interview stages through cross-stage propagation coefficients; it also includes causal untangling of interviewer dual-role agent, which models the interviewer's dual role as both scorer and causal agent, and identifies legitimate skill signals and confounding variables through cross-equation dependencies of scoring equation and adaptation equation.

3. The method according to claim 1, characterized in that, In step five, the causal path from the protected attribute to the score is divided into: a legitimate path where the protected attribute influences the score through reasonable mediation; an illegitimate path where the protected attribute directly influences the score through bias; and an indirect discrimination path where the protected attribute influences the score through unreasonable mediation. The path-specific effect PSE_illegal = E[Y(A=1, M_illegal=1, M_legal=M(A=0)) - Y(A=0, M_illegal=0, M_legal=M(A=0))], where M_illegal is the mediator variable on the illegitimate path and M_legal is the mediator variable on the legitimate path. The adjustment_fairness = -PSE_illegal - PSE_indirect; also includes the elimination of appearance path-specific effects, modeling candidate appearance as a chain of confusion sources, distinguishing between direct visual bias, chain propagation effects, and legitimate path effects, eliminating only the former two and retaining the latter; the adjustment_APSE = -PSE_appearance_direct -PSE_appearance_chain.

4. The method according to claim 1, characterized in that, In step six, the two-stage least squares method includes: the first stage T̂ = π0 + π1·IV1 + π2·IV2 + π3·IV3 + ε1, and the second stage Y = β0 + β1·T̂ + ε2, where IV1 is an instrumental variable for the interview time period, IV2 is an instrumental variable for the randomness of the interviewer scheduling, and IV3 is an instrumental variable for the randomization of the question order; adjustment_IV = (β̂_2SLS - β̂_OLS) × T_observed; it also includes dynamic game causal effect estimation, modeling the interview as a dynamic game between the interviewer and the candidate, and constructing endogeneous instrumental variables using equilibrium conditions; it also includes sequence contrast contamination effect correction, modeling the contrast effect of the previous candidate's rating on the current candidate in the interview sequence as a interference effect, and constructing instrumental variables using the quasi-randomness of the interview order; adjustment_SCCEC = -β̂_contrast × contrast_effect_hat.

5. The method according to claim 1, characterized in that, Step three also includes cross-modal causal divergence confusion detection, which independently estimates the causal effects from each modality, and uses the cross-modal causal effects to estimate the divergence matrix to detect and locate the source of confusion; it also includes network quality mediation measurement error correction, which models network quality as a source of mediation measurement error and eliminates differential measurement error through two-stage correction; the adjustment_NQMEC = -β̂_AV × AV_quality_hat; also includes causal representation learning: using a variational autoencoder to map multimodal observation data to a latent causal variable space, adding a mutual information penalty term to the loss function of the variational autoencoder to achieve causal variable decoupling, and learning a structural causal model for each pair of causal variables to achieve causal mechanism learning; also includes counterfactual scoring trajectory generation: generating counterfactual scoring trajectories for K interview stages based on a cross-stage causal effect propagation network, calculating trajectory slope, curvature, and inflection point number to perform trajectory shape analysis and evaluate scoring credibility; also includes causal effect consistency verification: cross-validating four independent estimation results using dual machine learning estimation, the mean of the causal forest conditional average treatment effect, the total effect of path-specific effects, and instrumental variable estimation, calculating a consistency index, and triggering model hypothesis re-examination when the consistency index is lower than a threshold; the scoring boundary constraint is that the adjustedScore is limited to the range of [rawScore×0.8, rawScore×1.2].

6. The method according to claim 1, characterized in that, Step three also includes a dual machine learning online estimation switching mechanism: when there are fewer than 20 historical observation data, linear regression is used as the basic causal effect estimation; when there are 20 or more historical observation data, it automatically switches to the dual machine learning online estimation mode, and obtains the coefficient estimate by dividing the inner product of the residuals by the sum of squared residuals, with the denominator threshold set to 0.001 to prevent division by zero.

7. A post-interview review and scoring adjustment system based on causal inference, characterized in that, It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to perform the functions of the following modules: The multimodal data acquisition module is used to acquire multimodal data during the interview process and extract features. The time-varying causal graph construction module is used to construct time-varying causal graphs based on prior knowledge in the interview domain, divide time windows and construct independent stage causal graphs, and detect causal structure drift. The dual machine learning causal effect estimation module is used to identify confounding variables based on time-varying causal graphs and calculate global causal effects through orthogonalized residuals; The Causal Forest Conditional Average Treatment Effect Estimation Module is used to estimate individualized conditional average treatment effects based on global causal effects. The path-specific fairness audit module is used to identify protected attributes based on causal graphs, distinguish between legal paths, illegal paths, and indirect discrimination paths, and calculate path-specific effects. The Interview-Specific Instrumental Variables module is used to eliminate unobservable confounding by utilizing instrumental variables specific to the interview scenario. The causal adjustment and explanation generation module is used to synthesize the results of various causal analyses to generate causal adjustment scores and causal explanations. The output of the multimodal data acquisition module is connected to the time-varying causal graph construction module. The output of the time-varying causal graph construction module is connected to the dual machine learning causal effect estimation module, the path-specific fairness audit module, and the interview-specific instrumental variable module. The output of the dual machine learning causal effect estimation module is connected to the causal forest conditional average treatment effect estimation module. The outputs of the causal forest conditional average treatment effect estimation module, the path-specific fairness audit module, and the interview-specific instrumental variable module are all connected to the causal adjustment and explanation generation module. The feedback output of the causal adjustment and explanation generation module is connected to the dual machine learning causal effect estimation module and the time-varying causal graph construction module.

8. The system according to claim 7, characterized in that, It also includes a multi-level caching module for caching causal effect calculation results and causal graph structures, including L1 memory cache and L2 distributed cache; when the processor performs causal representation learning, it uses a variational autoencoder encoder / decoder architecture, adds a mutual information penalty term to the loss function to achieve causal variable decoupling, and uses a multilayer perceptron parameterized causal model to achieve causal mechanism learning.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

10. An electronic device comprising a processor and a memory, characterized in that, The processor is configured to perform the method of any one of claims 1 to 6 by executing executable instructions in the memory.

Images