Method for predicting outflow trend of college professional talents based on multi-channel feature fusion and adaptive weighted random forest

Abstract

The application provides a college major talent outflow trend prediction method based on multi-channel feature fusion and adaptive weighting random forest, relates to the talent flow prediction and machine learning application technical field, and comprises the following steps: collecting college major enrollment data and corresponding macroeconomic data of each year, and preprocessing missing data; adopting a multi-channel feature decoupling and fusion mechanism; the multi-channel feature decoupling and fusion mechanism adopted by the application effectively solves the defects of weak feature expression ability and neglecting external economic driving factors in the prior art, decouples internal trends and external driving factors into three channels, comprehensively extracts enrollment sequence characteristics, regional macroeconomic characteristics and regional economic difference characteristics, realizes deep fusion of internal and external information, enhances the model's recognition ability to complex patterns, can fully capture the influence of macroeconomic variables on professional talent flow, and greatly improves the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of talent mobility prediction and machine learning application technology, and in particular to a method for predicting the outflow trend of college professional talents based on multi-channel feature fusion and adaptive weighted random forest. Background Technology

[0002] With the continuous expansion of higher education and the profound transformation of the socio-economic structure, the development of university majors exhibits a significant unevenness in popularity. Popular majors see a surge in applicants and an ample supply of talent, while traditional engineering and basic disciplines face difficulties in enrollment and a mismatch between talent cultivation and social needs, further exacerbating the outflow of university professionals from certain regions or industries. Currently, predicting the outflow trend of university professionals mainly employs simple time series models or static regression models. Some solutions also utilize improved models such as LSTM and grey prediction, all of which are based on historical enrollment data and are widely used in university major planning and talent mobility monitoring.

[0003] Existing forecasting methods neglect external economic drivers, relying solely on historical enrollment data for modeling. They fail to consider the significant impact of macroeconomic variables such as GDP, per capita income, employment salaries, and inflation on major selection and talent mobility. They are also slow to respond to sudden trend changes; when structural shifts occur in major popularity, traditional models, due to their fixed weights, cannot quickly capture these changes. Furthermore, their feature representation capabilities are weak; most models treat the enrollment sequence as a single time series, lacking decoupling and fusion modeling of both "internal trends" and "external drivers" information. Sample weighting is also crude, typically using equal weight training or fixed decay coefficients, failing to dynamically adjust the importance of historical data based on sequence volatility. This results in large prediction biases and delayed responses, making it difficult to effectively support early warning of talent outflow and educational decision-making. Therefore, this invention proposes a method for predicting the trend of talent outflow in higher education majors based on multi-channel feature fusion and adaptive weighted random forest to address the problems existing in the current technology. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a method for predicting the outflow trend of professional talents from universities based on multi-channel feature fusion and adaptive weighted random forest. Employing a multi-channel feature decoupling and fusion mechanism, it effectively overcomes the shortcomings of existing technologies, such as weak feature representation and neglect of external economic drivers. By decoupling internal trends and external driving factors into three channels, it comprehensively extracts the characteristics of the enrollment sequence itself, the macroeconomic characteristics of the region, and the regional economic difference characteristics, achieving deep integration of internal and external information. This enhances the model's ability to identify complex patterns, fully captures the impact of macroeconomic variables on the flow of professional talents, and significantly improves prediction accuracy.

[0005] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for predicting the outflow trend of professional talents from universities based on multi-channel feature fusion and adaptive weighted random forest, comprising the following steps:

[0006] S1: Collect historical enrollment data for various majors in universities and corresponding macroeconomic data, and preprocess the missing data;

[0007] S2: Employ a multi-channel feature decoupling and fusion mechanism to extract internal trend features, local macroeconomic correlation features, and regional economic indicator difference features, and then splice them together to form a fused feature vector;

[0008] S3: By evaluating the historical prediction errors of multiple candidate window lengths through rolling backtesting, the window with the smallest mean absolute error is automatically selected as the optimal sliding window for each specialty.

[0009] S4: Based on the coefficient of variation of the enrollment data within the sliding window, adaptively calculate the alpha parameter of the exponentially weighted sample weights and assign dynamic weights to the training samples;

[0010] S5: Based on the optimal sliding window, training samples are generated, and three independent random forest regression models are constructed to predict the number of students to be enrolled in the next 1-3 years.

[0011] S6: Detect the degree of trend oscillation in the predicted sequence, perform linear extrapolation to correct logically unreasonable prediction results, and ensure the logical consistency of multi-step prediction.

[0012] S7: Based on the prediction results, identify professions at risk of brain drain and output early warning information.

[0013] Further improvements are made in the following aspects: In S1, the macroeconomic data includes regional GDP, GDP per capita, average wage of urban employees, resident population, and consumer price index, and the data are sourced from statistical bulletins and public databases; the enrollment data is macroeconomic summary data that has been anonymized and personal identification information has been removed; missing data is filled using linear interpolation.

[0014] Further improvements are made in the following: In S2, the multi-channel feature decoupling and fusion mechanism is specifically as follows: the factors affecting the professional enrollment trend are decoupled into one internal trend channel and two external driving channels; the internal trend channel extracts the mean, standard deviation, linear regression slope, total growth, last year value, and autocorrelation coefficient of the enrollment sequence; external driving channel 1 extracts the mean and linear regression slope features of the macroeconomic indicators of the region; external driving channel 2 extracts the difference, ratio, and ranking features of the macroeconomic indicators of the region and the reference region; the feature vectors of the three channels are concatenated to obtain the fused feature vector.

[0015] A further improvement lies in: the autocorrelation coefficient is obtained through... The function calculation introduces a small smoothing factor to prevent calculation failure due to zero variance; the difference feature is the difference between the current regional index and the average value of the reference regional index, and the ratio feature is the ratio between the current regional index and the average value of the reference regional index plus the small smoothing factor.

[0016] A further improvement is made in S3, where the specific process for window selection is as follows: traverse the candidate window lengths. Check whether each candidate window meets the modeling conditions; for each feasible candidate window, calculate its mean absolute error through rolling backtesting; select the candidate window with the smallest mean absolute error as the optimal sliding window for the major, and realize personalized modeling of one window per major.

[0017] A further improvement is made in S4, where the adaptive alpha parameter is calculated as follows: First, the ratio of the standard deviation to the mean of the enrollment sequence within the sliding window is calculated to obtain the coefficient of variation, and a small smoothing factor is introduced to prevent the denominator from being zero; the coefficient of variation is mapped to a preset alpha value range through a nonlinear mapping function to obtain the basic alpha value; the basic alpha value is used as an exponential decay parameter to calculate the time weight of each training sample, and the weight decays exponentially as the sample time is further away from the current time.

[0018] A further improvement is made in S4, where the specific calculation algorithm is as follows:

[0019] Calculate the coefficient of variation (CV): First, calculate the standard deviation of the admission sequence within the current sliding window. with the mean The ratio, and introduce a small smoothing factor. To prevent the denominator from being zero:

[0020] ,

[0021] Mapping computation fundamentals value, In other words, alpha: using a non-linear mapping function, the CV value is mapped to a preset value. Within the value range, the following logical function form is used:

[0022]

[0023] Determine the final Value: The calculated `base` is used as the final exponential decay parameter to calculate the sample weight `wt`, where `T` is the latest sample time point and `t` is the historical sample time point.

[0024] .

[0025] Further improvements are made in the following aspects: In S5, the random forest regression model uses Bayesian optimization for automatic hyperparameter tuning, and the hyperparameters include the number of decision trees, maximum depth, minimum number of samples for internal node splits, minimum number of samples for leaf nodes, and maximum number of features; the training process uses time series cross-validation and early stopping strategies to ensure the model's generalization ability.

[0026] Further improvements are made in the following steps: In step S6, the specific detection process is as follows: a complete time series containing historical data benchmark values ​​and a predicted sequence is constructed, and the first-order difference sequence of adjacent years is calculated; the standard deviation of the difference sequence is calculated as a fluctuation oscillation index, and the mean of the absolute values ​​of the differences is calculated as the average change amplitude; when the fluctuation oscillation index exceeds a preset proportion of the average change amplitude, the predicted sequence is reconstructed using a linear extrapolation method to ensure that the multi-step prediction trend is smooth and coherent, wherein the preset proportion coefficient is 0.5; the linear extrapolation method calculates the average growth slope through the first and last terms of the predicted sequence, and generates a corrected predicted sequence based on the initial predicted value and the slope.

[0027] A further improvement is that, in S7, the early warning of talent outflow is based on the trend changes of the predicted sequence, identifies majors with continuously declining enrollment, and outputs early warning information and corresponding educational decision support suggestions.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. This invention adopts a multi-channel feature decoupling and fusion mechanism, which effectively solves the defects of existing technologies such as weak feature expression ability and neglect of external economic driving factors. By decoupling internal trends and external driving factors into three channels, it comprehensively extracts the characteristics of the enrollment sequence itself, the macroeconomic characteristics of the region, and the regional economic difference characteristics, realizing the deep integration of internal and external information, enhancing the model's ability to identify complex patterns, fully capturing the impact of macroeconomic variables on the flow of professional talents, and significantly improving prediction accuracy.

[0030] 2. This invention adopts an adaptive alpha decay mechanism based on volatility, which effectively solves the defects of existing technologies such as coarse sample weight processing and slow response to trend changes. By dynamically adjusting the alpha parameter through the coefficient of variation, the model can adaptively allocate historical sample weights according to the volatility characteristics of professional enrollment data. When the volatility is severe, it prioritizes recent data and responds quickly to trend changes. When the volatility is stable, it takes into account long-term historical trends, thus improving the flexibility and response speed of the model.

[0031] 3. The present invention adopts a dynamic window length selection mechanism, which effectively solves the defects of the existing technology with fixed window settings that cannot adapt to the heterogeneity of different professions. By automatically selecting the optimal window length for each profession through rolling backtesting, it realizes personalized modeling of "one profession, one window". It can provide sufficient historical data support for stable professions, and capture the latest trend signals for sudden and rapidly growing professions, avoiding the noise sensitivity or trend dilution problems caused by fixed windows.

[0032] 4. This invention employs a prediction consistency verification and linear fallback mechanism, effectively addressing the shortcomings of existing technologies such as chaotic multi-step prediction logic and low reliability. By detecting the oscillation degree of the prediction sequence and performing linear correction, it ensures that the multi-step prediction results are smooth and consistent in the trend direction, eliminates random noise, and improves the logical rationality and reliability of the prediction results. Simultaneously, it combines... Functional and Bayesian optimization further enhances the robustness and generalization ability of the model, providing reliable support for early warning of talent outflow. Attached Figure Description

[0033] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0034] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0035] Example 1

[0036] according to Figure 1 As shown in the figure, this embodiment proposes a method for predicting the outflow trend of professional talents from universities based on multi-channel feature fusion and adaptive weighted random forest, including the following steps:

[0037] Data preparation and preprocessing

[0038] The data used in this embodiment strictly complies with relevant regulations, and the data sources are legal and compliant, specifically including:

[0039] Macroeconomic data: Sourced from the National Economic and Social Development Statistical Communiqué, Statistical Yearbook, and government open databases published by the state and provincial-level administrative regions. This type of data is legally publicly available macroeconomic statistical information and does not involve any personal privacy or trade secrets.

[0040] Educational enrollment data: Sourced from historical professional enrollment statistics compiled by education authorities. This data is an aggregated data that has undergone anonymization, removing all personally identifiable information (PII) and retaining only statistical dimensions such as major codes, years, and enrollment numbers.

[0041] Data processing principles: In the preprocessing stage, this method only performs numerical calculations and feature extraction on the above-mentioned legally obtained macro-statistical data. It does not involve the collection, storage or processing of any personal information, nor does it involve recommendations or decisions for specific individuals, and complies with the ethical norms of artificial intelligence.

[0042] Based on the aforementioned compliant channels, the following two types of data are collected for model training:

[0043] Admissions data: Annual enrollment figures for each major were collected for 11 years, from 2014 to 2024. For example:

[0044] 080901 (Computer Science and Technology): [201,253,222,1425,1683,1390,1530,1869,2164,2382,2498]

[0045] 030101 (Law): [739,668,738,788,792,618,623,846,888,900,1197]

[0046] (Other majors omitted)

[0047] Macroeconomic data (2014–2024):

[0048] GDP: Gross Regional Product (100 million yuan)

[0049] gdp_per_capita: GDP per capita (yuan)

[0050] wage: Average wage of urban employees (yuan)

[0051] population: resident population (ten thousand people)

[0052] CPI: Consumer Price Index

[0053] Missing value handling: For majors with missing values ​​(such as the 2024 data missing for Nursing 630201), linear interpolation is used to fill in the missing values. For example, the number of students admitted to 630201 in 2024 is obtained by linear interpolation of 2805 in 2022 and 2677 in 2023, which is approximately 2641.

[0054] Sliding window sample generation

[0055] Set the sliding window length k=3, that is, use three consecutive years of data to predict future years.

[0056] Taking major 080901 (Computer Science) as an example, its enrollment sequence from 2014 to 2024 is: [201,253,222,1425,1683,1390,1530,1869,2164,2382,2498]

[0057] The following training samples were generated:

[0058]

[0059] A total of 11-3=8 samples were generated.

[0060] Part 1: Multi-channel feature extraction (taking the 2021 window as an example)

[0061] For each window, extract three types of features:

[0062] Channel A: Internal Trend Characteristics (Based on Enrollment Sequence [1869,2164,2382])

[0063] Mean = (1869 + 2164 + 2382) / 3 ≈ 2138.3

[0064] Standard deviation ≈ 257.5

[0065] The linear regression slope is approximately 256.5 (showing a rapid upward trend).

[0066] Total growth over three years = (2382 - 1869) / 3 ≈ 171.0

[0067] The value in the last year was 2382.

[0068] Autocorrelation coefficient (lag=1) ≈ 0.98 (highly autocorrelated)

[0069] Channel B: External Driving Feature 1 - Provincial Economic Indicator Data (Based on Economic Data from 2021–2023)

[0070] Average GDP ≈ (45713.5 + 47558.6 + 50012.9) / 3 ≈ 47761.7

[0071] Average salary mean ≈ (85438 + 91413 + 97388) / 3 ≈ 91413.0

[0072] The average annual GDP growth rate is approximately 2149.7%.

[0073] Average annual wage growth slope ≈ 5975.0

[0074] (Other economic variables are similar)

[0075] Channel C: External Driving Feature 2 - Difference in Regional Economic Indicators

[0076] For each macroeconomic variable, three types of differences were calculated. The provincial data used for comparison were the economic indicators of five developed provinces: Beijing, Shanghai, Jiangsu, Zhejiang, and Guangdong. The average economic indicators of these five developed provinces were used for the difference and ratio characteristics.

[0077] Difference characteristics =

[0078] Ratio characteristics =

[0079] Ranking characteristics: The current province's ranking in terms of economic indicators among N comparable provinces (e.g., ranking 3rd out of 10 in wages).

[0080] Feature vector fusion: Concatenate all features from channels A, B, and C to form the input features of the sample.

[0081] 2. Adaptive Alpha Sample Weighting Mechanism

[0082] When training a random forest model, time weights are assigned to each training sample, following the exponential decay principle that "the more recent the sample, the higher its weight." Traditional methods typically use a fixed decay coefficient (e.g., α=0.3), which is difficult to adapt to the fluctuating characteristics of different disciplines.

[0083] This invention proposes an adaptive method based on the coefficient of variation (CV). Calculation method. This method dynamically adjusts the time decay coefficient based on the dispersion of enrollment data within the sliding window. The core logic is: when professional enrollment data fluctuates drastically (high CV value), a larger percentage is assigned... A higher CV value makes the model more reliant on recent data to respond quickly to sudden changes; when the data is stable (small CV value), a smaller CV value is assigned. This value allows the model to take into account long-term historical trends.

[0084] The specific adaptive calculation process is as follows:

[0085] Calculate the coefficient of variation (CV): First, calculate the standard deviation of the admission sequence within the current sliding window. with the mean The ratio, and introduce a small smoothing factor. To prevent the denominator from being zero:

[0086]

[0087] Mapping computation fundamentals Value: Using a non-linear mapping function, the CV value is mapped to a preset value. Within the value range. This embodiment adopts the following logical function form:

[0088]

[0089] Determine the final Value: The calculated `base` is used as the final exponential decay parameter to calculate the sample weight `wt`, where `T` is the latest sample time point and `t` is the historical sample time point.

[0090]

[0091] For example:

[0092] For the major code 080901 (Computer Science), the CV for the past three years [1869, 2164, 2382] is approximately 257.5 / 2138.3, which is approximately 0.12, indicating relatively small fluctuations. (Relatively gradual decline)

[0093] For the professional code 080202 (Mechanical) [2038,2224,2264], the CV is minimal → alpha≈0.1, and the weights of historical data are more balanced.

[0094] For professional code 630302 (Financial Accounting) [3870,4324,159], the CV is extremely large (sudden drop) → alpha≈0.7, which is highly dependent on the latest data and responds quickly to anomalies.

[0095] Model training and prediction

[0096] For each profession, this invention constructs and trains three independent random forest regression models (M1, M2, M3) to predict talent demand in the first, second, and third years, respectively. To ensure the model's prediction accuracy and generalization ability, this embodiment configures the model's architecture parameters, training and validation process, and termination strategy as follows:

[0097] Hyperparameter Configuration and Optimization Strategy: This invention employs a Bayesian optimization-based automatic hyperparameter tuning mechanism. The goal is to maximize the model's determination coefficient on the validation set, dynamically determining the optimal model parameter combination for each specialty. The specific hyperparameter search space and configuration are as follows:

[0098] Number of decision trees: Define the search range as [50, 500]. This parameter controls the number of base learners in the forest. A larger number of base learners usually results in a more stable model, but also higher computational costs.

[0099] Maximum Depth: Defines the search range as [3, 20]. Used to limit the maximum depth of each decision tree to prevent the model from overfitting the training data.

[0100] Minimum number of samples for internal node splitting: The search range is defined as [2, 20]. This parameter controls the minimum number of samples required for a node to split; a larger value helps smooth the model and reduce overfitting.

[0101] Minimum number of samples per leaf node: Define the search range as [1, 10]. Ensure that each leaf node contains the minimum number of samples to improve the model's robustness to noise.

[0102] Maximum number of features: The sqrt{n} strategy (i.e., the square root of the total number of features) is adopted. This is a classic configuration for random forest regression, which aims to ensure diversity among decision trees.

[0103] The training, validation, and termination mechanism strictly follows the cross-validation principle during model training and includes a clear early stopping strategy. The specific process is as follows:

[0104] Data partitioning and cross-validation: Historical data for each specialty were partitioned into a training set (first 80%) and a test set (last 20%) based on time series data. During the training phase, 5-fold time series cross-validation was used to evaluate candidate parameters to ensure the model's generalization performance over time.

[0105] Optimization objective: The Bayesian optimizer seeks the optimal combination of the above hyperparameters by taking the root mean square error (RMSE) as the joint optimization objective and the coefficient of determination (R^2) as the maximum.

[0106] Early stopping strategy: During Bayesian optimization, if the performance of the validation set does not improve after K consecutive iterations, or if the preset maximum number of iterations is reached, the search is automatically terminated and the historical optimal parameters are returned.

[0107] Through the above configuration, the present invention ensures that the M1, M2, and M3 models trained for each profession are in optimal condition, providing a solid technical foundation for subsequent accurate year-by-year predictions.

[0108] 3. Dynamic window length selection mechanism (taking Computer Science and Technology, major 080901, as an example)

[0109] In traditional time series forecasting, the sliding window length k is usually used as a fixed hyperparameter (e.g., k=3), but enrollment data from different majors exhibit significant heterogeneity:

[0110] Stable majors (such as Law 030101): Historical trends are stable, and a longer window can improve statistical reliability;

[0111] Abruptly changing majors (such as Finance and Accounting 630302): The number of students enrolled in 2024 dropped sharply from 4,324 to 159, requiring an extremely short window to respond quickly to the anomaly;

[0112] Accelerating growth majors (such as Computer Science 080901): These majors have experienced continuous high-speed growth in recent years, but an excessively long window of opportunity will dilute the latest trend signals.

[0113] To address the aforementioned issues, this invention proposes a dynamic window length selection mechanism based on minimizing historical backtesting errors, which automatically determines the optimal window length for each specialty. .

[0114] Implementation process (taking major 080901 as an example):

[0115] Candidate window traversal: For each candidate Check whether the data meets the modeling conditions (at least one training sample can be generated to predict t+3).

[0116] Rolling backtesting evaluation:

[0117] For each feasible starting year t, take a window Construct features, train a temporary model, and predict the next year (t+k+2).

[0118] Calculate the absolute error (MAE) between the predicted value and the actual value.

[0119] The average MAE of all feasible t is used as the comprehensive error score for k.

[0120] Choose the best The final window length for this major is selected based on the minimum average MAE value, k.

[0121] Example calculation (major 080901, data from 2014–2024):

[0122] Candidate k=2: Backtesting average MAE = 210.4

[0123] Candidate k=3: Backtesting average MAE = 185.7 ← Optimal

[0124] Candidate k=4: Backtesting average MAE=203.1

[0125] Candidate k=5: Skip due to data length limitations that prevent complete backtesting.

[0126] → Ultimately, k=3 was chosen for the computer science major, which not only captured the accelerated growth trend over the past 3 years (2022–2024: 2164→2382→2498), but also avoided noise sensitivity caused by an excessively short window.

[0127] 4. Predictive Consistency Verification Mechanism

[0128] M1 (forecast for 2025), M2 (forecast for 2026), and M3 (forecast for 2027) should have logical consistency (such as consistent trend direction); otherwise, a correction will be triggered.

[0129] Multi-step forecasting often produces irrational results such as "2025↑, 2026↓, 2027↑";

[0130] Post-processing consistency constraints can be introduced to improve the reliability of the results.

[0131] Implementation process

[0132] Calculating trend volatility and determining threshold First, construct a complete time series containing the current baseline value ybase (i.e., the historical data of the last year) and the future prediction sequence P={p1,p2,...,pN}, and calculate the first-order difference sequence D (i.e., the increment for each year) between adjacent years:

[0133]

[0134] To determine whether the predicted trend is chaotic, the volatility oscillation indicator Iosc is defined as the standard deviation of the difference sequence, and the average change amplitude Aavg is defined as the mean of the absolute values ​​of the differences. If the oscillation indicator exceeds a certain proportion of the average change amplitude (a proportion coefficient is set in this embodiment), the trend will be reversed. If the trends are inconsistent, corrections are needed.

[0135] Judgment conditions:

[0136] Linear trend extrapolation correction When the above criteria are met, it indicates that the original prediction results exhibit logical oscillations. In this case, the specific values ​​of the original predictions are discarded, and the prediction sequence is reconstructed using linear extrapolation.

[0137] Calculate the overall average growth slope k using the first term p1 and the last term pN of the predicted sequence:

[0138]

[0139] Subsequently, based on the initial predicted value p1p1 and the slope kk, a corrected smoothed prediction sequence is generated. :

[0140]

[0141] Technical effect description:

[0142] This mechanism acts as a "logic filter." For professions experiencing steady growth or decline, it effectively eliminates random noise generated by the model, forcing the predicted curve to conform to linear evolution. For professions with obvious trends, since the oscillation criteria are not met, the original prediction results will be retained, thus ensuring the diversity and accuracy of the prediction results.

[0143] Prediction phase:

[0144] Using the latest enrollment data (e.g., Computer Science: [2164,2382,2498]) and economic data from the latest windows [2022,2023,2024], multi-channel features are extracted. Input M1 / M2 / M3 to obtain the predicted values ​​for 2025 / 2026 / 2027.

[0145] Example 2

[0146] according to Figure 1 As shown, this embodiment proposes a method for predicting the outflow trend of professional talents in colleges and universities based on multi-channel feature fusion and adaptive weighted random forest. On the basis of the above sliding window sample generation, the generalization ability of the model is verified and the model is backtested.

[0147] Overall solution verification (comparison with existing technologies)

[0148] This experiment selected a single LSTM model, an ARIMA model, a grey prediction model, and a fixed-parameter random forest (FixedRF, with a fixed window size of 3 and a fixed weight of 0.3) as benchmarks for comparison. The evaluation metrics were the RMSE, MAE, and MAPE of the predicted values ​​for 2024-2025.

[0149]

[0150] Analysis of experimental results:

[0151] As shown in Table 1, for stable disciplines (law), the differences between the models are small, but the proposed solution still achieves the highest accuracy. For rapidly growing disciplines (computer science) and sudden-growth disciplines (finance), existing technologies (especially LSTM and ARIMA) suffer from a sharp increase in error because they cannot quickly respond to trend changes. The proposed solution, through dynamic windowing and adaptive weighting, effectively captures sudden-growth signals, achieving an average MAE of only about half that of the comparative models, demonstrating the unexpected technical effectiveness of this solution.

[0152] Innovative point ablation experiment

[0153] To demonstrate the independent contribution of each innovative component in this invention, this experiment removed the core components of this invention one by one from the "Computer Science and Technology" professional dataset and observed the changes in model performance (MAE). To further demonstrate the independent contribution of each innovative component in this invention, five control experiments were set up on the "Computer Science and Technology" professional dataset. The specific configurations of each group are defined as follows:

[0154] Group A (Benchmark Group / Complete Solution): Employs all the technical features of this invention, namely multi-channel feature fusion (channels A+B+C) + adaptive alpha weighting + dynamic window selection + consistency verification.

[0155] Group B (Removal of Multi-Channel Feature Fusion): Only channel A (internal trend features) is retained as input, and channel B (external economic indicators) and channel C (regional difference features) are completely removed. Other parameters remain consistent with Group A.

[0156] Group C (Adaptive Alpha Weighting Removed): Uses a fixed attenuation coefficient Perform sample weighting, and follow the same steps as for group A.

[0157] Group D (Remove Dynamic Window Selection): Uses a fixed sliding window length. No window length optimization is performed; the other steps are the same as those for Group A.

[0158] Group E (Removal of Consistency Check): No post-processing correction is performed after model prediction; the original predicted values ​​of M1 / M2 / M3 are directly output. Other steps are the same as in Group A.

[0159] The experiment quantifies the technological contribution of each innovation point by comparing the rate of increase in error of each group relative to group A.

[0160] The error increase rate after removing the component is shown in the table below:

[0161]

[0162] Analysis of experimental results:

[0163] Multi-channel feature fusion (Group B): The error increased the most (217.8%). This proves that external economic drivers (such as GDP and salaries) are crucial for predicting the popularity of majors, and historical enrollment data alone (single channel) cannot explain the changes brought about by external shocks.

[0164] Dynamic window selection (Group D): The error increased by 164.8%. This proves that a "one-size-fits-all" fixed window cannot adapt to the life cycle characteristics of different professions (e.g., the rapid growth of computer science requires a short window, while the stable period of law requires a long window).

[0165] Adaptive Alpha Weighting (Group C): Error increased by 130.5%. This demonstrates that when trends change abruptly, more recent data must be given higher weights, and fixed weights will cause the model to lag.

[0166] Consistency check (Group E): Although the error increase was relatively small (53.3%), in multi-step prediction (M1 / M2 / M3), this mechanism effectively smoothed the prediction curve, prevented logical oscillations of "2025 rise, 2026 fall, 2027 rise", and improved the interpretability and credibility of the prediction results.

[0167] Empirical Optimization of Dynamic Windows and Adaptive Coefficients

[0168] To further verify the effectiveness of the dynamic window length selection mechanism and adaptive alpha calculation method proposed in this invention, this embodiment selects representative professional data for backtesting and comparison experiments. The experiment aims to demonstrate that, compared with the traditional fixed parameter scheme, the adaptive mechanism of this invention can significantly reduce prediction errors, especially when dealing with highly volatile data.

[0169] Adaptive Mechanism verification (comparison with fixed) =0.3)

[0170] To rule out the possibility that "similar results can be obtained by adjusting and fixing α", this invention selects two types of extreme features for comparison:

[0171] Highly volatile majors (mutation type): Taking finance (0203) as an example, this major experienced a sudden drop in demand in 2020 due to macroeconomic adjustments.

[0172] Low volatility majors (stable type): Taking law (0301) as an example, the demand for this major has shown a steady and slow increase over the past five years.

[0173] Experimental design: Using a fixed attenuation coefficient =0.3 (a commonly used traditional value) and the adaptive α algorithm proposed in this invention are used for prediction, and the root mean square error (RMSE) is calculated.

[0174] The experimental results are shown in the table below:

[0175]

[0176] Analysis Conclusion: The data in the table above shows that for finance-related majors with high volatility, adaptive... The proposed solution reduced the prediction error (RMSE) by 45.1% compared to the fixed solution. This demonstrates that the adaptive mechanism can keenly capture data mutations and automatically adjust historical data weights, thereby significantly correcting prediction bias. For stationary data, the adaptive solution also maintained a slight optimization. This experiment proves that the adaptive algorithm of this invention has substantial prediction improvements when dealing with mutated data, a technical effect that cannot be achieved by simply adjusting a fixed α value.

[0177] Verification of the dynamic window length mechanism (compared to fixed windows k=2,3,4,5)

[0178] To verify the superiority of the "one specialty, one window" strategy, this invention compares the dynamic window scheme (i.e., automatically selecting the optimal kkk according to the AIC / BIC criteria) with the unified fixed window scheme.

[0179] Experimental Design:

[0180] a) Test set: N majors with different characteristics are randomly selected.

[0181] b) Evaluation index: Calculate the average backtesting error for all disciplines.

[0182] c) Scheme comparison: Dynamic window scheme vs. fixed window (k=2,k=3,k=4,k=5).

[0183] d) Experimental results:

[0184] Fixed window k=2: Average RMSE=89.5

[0185] Fixed window k=3: Average RMSE=82.3

[0186] Fixed window k=4: Average RMSE=78.6

[0187] Fixed window k=5: Average RMSE=85.1

[0188] Dynamic window scheme: Average RMSE = 65.4

[0189] Analysis Conclusion: Experimental data shows that any single fixed window length (e.g., k=2,3,4,5) cannot achieve optimal results across all disciplines. The dynamic windowing scheme of this invention, by automatically matching the most suitable historical window length for each discipline, results in a significantly lower overall average prediction error (RMSE=65.4) compared to all fixed windowing schemes. The non-uniform distribution of the optimal window length across different disciplines demonstrates, from a data perspective, that the personalized selection strategy of "one window per discipline" is more scientific and advanced than the general strategy of "uniform window length."

[0190] Forecast Results and Early Warning of Brain Outflow

[0191] For example, here are the projected results for some majors from 2025 to 2027:

[0192]

[0193] Decision support:

[0194] The Ministry of Education can initiate a "talent loss early warning" for computer science and technology majors, increase enrollment subsidies, and improve employment security;

[0195] Increase investment in faculty for the nursing profession;

[0196] Promote the reform of "English+" interdisciplinary talent training for English majors.

[0197] This invention achieves high-precision and robust prediction of college major enrollment trends through multi-channel feature fusion and adaptive weighting mechanism, effectively supporting talent mobility monitoring and education policy formulation, and has significant innovation and application value.

[0198] This invention employs a multi-channel feature decoupling and fusion mechanism, effectively addressing the shortcomings of existing technologies, such as weak feature expression capabilities and neglect of external economic drivers. By decoupling internal trends and external driving factors into three channels, it comprehensively extracts the characteristics of the enrollment sequence itself, the macroeconomic characteristics of the region, and the regional economic difference characteristics, achieving deep integration of internal and external information. This enhances the model's ability to identify complex patterns, fully captures the impact of macroeconomic variables on the flow of professional talent, and significantly improves prediction accuracy. Furthermore, this invention employs a volatility-based adaptive alpha decay mechanism, effectively addressing the shortcomings of existing technologies, such as coarse sample weight processing and slow response to trend changes. By dynamically adjusting the alpha parameter through the coefficient of variation, the model can adaptively allocate historical sample weights according to the volatility characteristics of professional enrollment data. During periods of severe volatility, it prioritizes recent data for rapid response to trend changes, while during periods of stability, it considers long-term historical trends, improving the model's flexibility and response speed. This invention employs a dynamic window length selection mechanism, effectively addressing the shortcomings of existing technologies with fixed window settings that cannot adapt to the heterogeneity of different disciplines. Through rolling backtesting, it automatically selects the optimal window length for each discipline, achieving personalized modeling with "one window per discipline." This provides sufficient historical data support for stable disciplines while capturing the latest trend signals for abrupt or rapidly growing disciplines, avoiding noise sensitivity or trend dilution problems caused by fixed windows. Furthermore, this invention utilizes a prediction consistency verification and linear fallback mechanism, effectively addressing the shortcomings of existing technologies with chaotic multi-step prediction logic and low reliability. By detecting the oscillation degree of the prediction sequence and performing linear correction, it ensures that the multi-step prediction results are smooth and consistent in the trend direction, eliminating random noise and improving the logical rationality and reliability of the prediction results. Simultaneously, it combines... Functional and Bayesian optimization further enhances the robustness and generalization ability of the model, providing reliable support for early warning of talent outflow.

[0199] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest, characterized in that, Includes the following steps: S1: Collect historical enrollment data for various majors in universities and corresponding macroeconomic data, and preprocess the missing data; S2: Employ a multi-channel feature decoupling and fusion mechanism to extract internal trend features, local macroeconomic correlation features, and regional economic indicator difference features, and then splice them together to form a fused feature vector; S3: By evaluating the historical prediction errors of multiple candidate window lengths through rolling backtesting, the window with the smallest mean absolute error is automatically selected as the optimal sliding window for each specialty. S4: Based on the coefficient of variation of the enrollment data within the sliding window, adaptively calculate the alpha parameter of the exponentially weighted sample weights and assign dynamic weights to the training samples; S5: Based on the optimal sliding window, training samples are generated, and three independent random forest regression models are constructed to predict the number of students to be enrolled in the next 1-3 years. S6: Detect the degree of trend oscillation in the predicted sequence, perform linear extrapolation to correct logically unreasonable prediction results, and ensure the logical consistency of multi-step prediction. S7: Based on the prediction results, identify professions at risk of brain drain and output early warning information.

2. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S1, macroeconomic data includes regional GDP, per capita GDP, average wage of urban employees, resident population, and consumer price index. The data are sourced from statistical bulletins and public databases. Enrollment data is macroeconomic summary data that has been anonymized to remove personally identifiable information. Missing data is filled using linear interpolation.

3. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S2, the multi-channel feature decoupling and fusion mechanism is specifically as follows: the factors affecting the professional enrollment trend are decoupled into one internal trend channel and two external driving channels; the internal trend channel extracts the mean, standard deviation, linear regression slope, total growth, last year value, and autocorrelation coefficient of the enrollment sequence; external driving channel 1 extracts the mean and linear regression slope features of the macroeconomic indicators of this region; external driving channel 2 extracts the difference, ratio, and ranking features of the macroeconomic indicators of this region and the reference region. The feature vectors of the three channels are concatenated to obtain the fused feature vector.

4. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 3, characterized in that: The autocorrelation coefficient is obtained through In function computation, a small smoothing factor is introduced to prevent computation failure due to zero variance; The difference feature is the difference between the current regional indicator and the average value of the reference regional indicator, and the ratio feature is the ratio between the current regional indicator and the average value of the reference regional indicator plus a small smoothing factor.

5. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S3, the specific process for window selection is as follows: traverse the candidate window lengths. Check whether each candidate window meets the modeling conditions; For each feasible candidate window, its mean absolute error is calculated through rolling backtesting; The candidate window with the smallest mean absolute error is selected as the optimal sliding window for that major, thus realizing personalized modeling with one window per major.

6. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S4, the calculation process of the adaptive alpha parameter is as follows: First, the ratio of the standard deviation to the mean of the enrollment sequence within the sliding window is calculated to obtain the coefficient of variation, and a small smoothing factor is introduced to prevent the denominator from being zero; the coefficient of variation is mapped to the preset alpha value range through a nonlinear mapping function to obtain the basic alpha value. Using the base alpha value as an exponential decay parameter, the time weight of each training sample is calculated. The weight decays exponentially as the sample time becomes further away from the current time.

7. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 6, characterized in that: The specific calculation algorithm in S4 is as follows: Calculate the coefficient of variation (CV): First, calculate the standard deviation of the admission sequence within the current sliding window. with the mean The ratio, and introduce a small smoothing factor. To prevent the denominator from being zero: ， Mapping computation fundamentals value, In other words, alpha: using a non-linear mapping function, the CV value is mapped to a preset value. Within the value range, the following logical function form is used: Determine the final Value: The calculated `base` is used as the final exponential decay parameter to calculate the sample weight `wt`, where `T` is the latest sample time point and `t` is the historical sample time point. 。 8. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S5, the random forest regression model uses Bayesian optimization to automatically tune hyperparameters, including the number of decision trees, maximum depth, minimum number of samples for internal node splits, minimum number of samples for leaf nodes, and maximum number of features. The training process employs time-series cross-validation and an early stopping strategy to ensure the model's generalization ability.

9. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S6, the specific detection process is as follows: construct a complete time series containing historical data benchmark values ​​and predicted sequences, calculate the first-order difference series of adjacent years; calculate the standard deviation of the difference series as a fluctuation and oscillation index, and calculate the mean of the absolute values ​​of the differences as the average change amplitude. When the fluctuation index exceeds the preset proportion of the average change range, the linear extrapolation method is used to reconstruct the prediction sequence to ensure that the multi-step prediction trend is smooth and coherent. The preset proportion coefficient is 0.

5. The linear extrapolation method calculates the average growth slope through the first and last terms of the prediction sequence, and generates a corrected prediction sequence based on the initial prediction value and the slope.

10. The method for predicting the outflow trend of college professionals based on multi-channel feature fusion and adaptive weighted random forest according to claim 1, characterized in that: In S7, the early warning of brain drain is based on the trend changes of the predicted sequence, identifies majors with continuously declining enrollment, and outputs early warning information and corresponding educational decision support suggestions.

Images