Logistics management talent cultivation matching and personalized recommendation system based on big data

Abstract

The invention discloses a logistics management talent cultivation matching and personalized recommendation system based on big data. The system comprises a data acquisition module, a double-portrait construction module, a hybrid intelligent matching recommendation module and a dynamic optimization module. The system respectively constructs a logistics management talent portrait and a logistics post demand portrait by collecting multi-source heterogeneous data; a mixed algorithm of a fusion decision tree algorithm, an improved Apriori association mining algorithm and an attention mechanism BP neural network algorithm is adopted, and bidirectional precise matching from talents to posts and from talents to culture resources is achieved; and the dynamic optimization module monitors the matching effect in real time, adjusts algorithm parameters and feature weights, and continuously improves the system matching precision and the personalized recommendation effect. According to the invention, the problem of accurate matching between talents, posts and resources in the logistics field is effectively solved.

Description

Technical Field

[0001] This invention belongs to the fields of computer technology and human resource management, specifically a big data-based logistics management talent training matching and personalized recommendation system. Background Technology

[0002] With the rapid digital and intelligent transformation of the logistics industry, enterprises are increasingly demanding logistics management talent with comprehensive capabilities. However, there is a significant structural contradiction between talent cultivation and job requirements in the current logistics field: on the one hand, talent struggles to accurately identify positions and learning resources that align with their career development; on the other hand, enterprises face challenges such as high recruitment costs, mismatch between personnel and positions, and low employee training efficiency. The root cause of this problem lies in the information barriers between talent, positions, and resources, as well as the difficulty of traditional methods in effectively integrating and intelligently analyzing multi-source heterogeneous data.

[0003] Currently, existing technical solutions mainly fall into two categories: one is traditional keyword-based recruitment and training platforms, which make recommendations through simple keyword matching of resumes and job descriptions, using relatively simple technical methods; the other is recommendation systems that employ a single machine learning algorithm, which achieves preliminary automated matching to some extent. In addition, some solutions are beginning to explore building user profiles to improve the accuracy of recommendations.

[0004] However, these existing technical solutions still have significant problems. First, they typically rely on a single algorithm, making it difficult to handle complex matching tasks such as classification, association, and deep nonlinear mapping simultaneously, resulting in limited recommendation accuracy. Second, they lack in-depth exploration of the unique two-way matching relationships between talent and positions, and talent and resources specific to the logistics field. Finally, the system's models and parameters are mostly statically set, unable to evolve dynamically based on feedback data and industry trends, leading to lagging recommendation results and difficulty in adapting to the rapidly changing needs of the logistics industry. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a big data-based logistics management talent training matching and personalized recommendation system to solve the problems of insufficient accuracy, poor adaptability and inability to achieve closed-loop personalized training in existing recommendation systems.

[0006] To achieve the above objectives, this invention discloses a big data-based logistics management talent training matching and personalized recommendation system. The system includes a data acquisition module, a dual-profile construction module, a hybrid intelligent matching and recommendation module, and a dynamic optimization module. The specific steps are as follows: S1: Collect multi-source heterogeneous data through the data acquisition module. The multi-source heterogeneous data includes logistics management talent data, job demand data, training resource data, and industry trend data. S2: The dual-profile construction module extracts, cleans, and fuses features from the collected multi-source heterogeneous data to construct profiles of logistics management talents and logistics job requirements, respectively. S3: Based on the profile of logistics management talent and the profile of logistics job requirements, the hybrid algorithm used in the hybrid intelligent matching recommendation module is used to complete the two-way matching of logistics management talent to logistics job and logistics management talent to logistics training resources. S4: Based on the bidirectional matching results, the dynamic optimization module monitors two types of indicators in real time: one is the matching effect indicator, including matching accuracy, job retention rate, training utilization rate, and talent capability improvement; the other is the update data indicator, which is measured by the xin increment of the multi-source heterogeneous data. When the matching effect indicator or the update data indicator is lower than the preset threshold, the optimization process is triggered.

[0007] Furthermore, the industry trend data is obtained by regularly crawling logistics industry policy documents and market analysis reports using web crawlers, and key trend features are extracted using the TF-IDF algorithm, expressed by the following formula: Where t represents a term; d represents a document; and D represents a document set; The word frequency of term t in document d; The inverse document frequency of term t in document set D; The score represents the critical importance of term t in document d relative to document set D.

[0008] Furthermore, the talent profile for logistics management includes four characteristics: basic attributes, explicit abilities, hidden potential, and career aspirations. The logistics job requirements profile includes four characteristics: skill requirements, job responsibilities, implicit fit, and career development opportunities.

[0009] Furthermore, the hybrid algorithm includes a decision tree algorithm, an improved Apriori association mining algorithm, and an attention mechanism BP neural network algorithm; The decision tree algorithm described uses an improved C4.5 decision tree algorithm, which optimizes the information gain ratio by introducing a feature redundancy penalty factor, expressed by the following formula: in, Indicates sample characteristics; Represents a sample set; For sample feature a, the sample set Information gain; This is the inherent value of sample feature a; This is the penalty coefficient; For feature redundancy; Indicates that sample feature 'a' is a sample set. The resulting increase in information.

[0010] The improved Apriori association mining algorithm adopts a vertical data storage format. Based on the structured storage characteristics of the vertical data storage format, it combines a bidirectional scanning mechanism and a frequent itemset pre-pruning strategy to optimize the mining efficiency of frequent itemsets. It also dynamically adjusts the support threshold according to the frequent itemset type to mine association rules between talent and job positions, and between talent and training resources, and to build an association rule base. The attention mechanism BP neural network algorithm is a neural network constructed from an input layer, an attention layer, a hidden layer, and an output layer. The attention layer strengthens the feature weights, and the Adam optimizer is used to optimize the loss function. The output layer outputs a quantized matching score. The mathematical expression for the feature weights is: in, Let be the attention weight for the i-th input feature; Let i be the i-th input feature value; This is the attention layer weight matrix; is the attention layer bias term; z is the total number of features.

[0011] Furthermore, the bidirectional scanning mechanism includes forward scanning and reverse scanning; the forward scanning is to traverse from low-order frequent itemsets to high-order frequent itemsets layer by layer to generate candidate frequent itemsets; the reverse scanning is to decompose the candidate frequent itemsets generated by the forward scanning in reverse to generate frequent item subsets.

[0012] Furthermore, the pre-pruning strategy uses frequent itemsets for intersection calculation; the intersection calculation of frequent itemsets uses a bitwise AND operation, expressed by the formula: in, The intersection of two frequent itemsets; and Let the frequent itemsets be the ones to be intersected. For binary and operators.

[0013] Furthermore, the matching of logistics management talent to logistics positions specifically includes: Based on the logistics management talent profile and the logistics job requirement profile, profile features are extracted. The profile features include talent features and job features. The profile feature extraction includes feature selection, feature encoding and vector construction. The feature selection is based on the feature weights of the logistics management talent profile and the logistics job requirement profile, prioritizing the selection of core features with a weight ratio greater than a set value, and eliminating redundant features with a weight ratio less than a set value. The feature encoding involves using one-hot encoding or label encoding for categorical core features and standardizing numerical core features. The vector construction involves multiplying the encoded feature values ​​by their corresponding feature weights and combining them in a fixed dimensional order to construct talent feature vectors and job feature vectors, respectively. The value of each dimension in the vector is equal to the encoded feature value multiplied by the corresponding feature weight. The extracted talent feature vector is input into the improved C4.5 decision tree algorithm to classify job types, initially screening out candidate jobs of the same type that match the talent's career aspirations. Then, based on the association rule base constructed by the improved Apriori association mining algorithm, the suitability of the talent's core skill combination with the job skill requirements is compared, further eliminating candidate jobs whose skill combinations do not meet the core job requirements. Finally, the job feature vector and talent feature vector of the remaining candidate jobs are concatenated to form a fused feature vector, which is input into the attention mechanism BP neural network algorithm to output a quantitative matching score. Finally, a job recommendation list is generated in descending order of the score. The matching of logistics management talent with logistics training resources specifically includes: First, the profiles of logistics management personnel and logistics job requirements are compared across feature dimensions, and the differences between each dimension are calculated to identify the personnel's skill gaps. Then, based on these skill gaps, the association rule base constructed using the improved Apriori association mining algorithm is retrieved to initially screen training resources whose skill coverage matches the skill gaps. Subsequently, the skill coverage dimensions, difficulty level, and the personnel's own learning rate of the training resources are combined to narrow down the range of training resources and form multiple candidate learning paths. Finally, the path optimization objective is to minimize the weighted average of learning time and cost, using the formula: in, The overall cost of the learning path; Weighted by learning time; Total study time; Weighted by learning cost; For the total learning cost, and + =1.

[0014] Furthermore, the optimization process steps are as follows: S801: Standardize the updated data, extract the features of each dimension of the standardized logistics management talent data and job demand data, and combine them in a fixed dimension order to generate a job classification feature vector adapted to the improved C4.5 decision tree algorithm and a quantitative matching feature vector adapted to the attention mechanism BP neural network algorithm; extract the skill coverage related features of the standardized training resource data, combine them with the TF-IDF key trend features of the standardized industry trend data, and combine them in a fixed dimension order to generate an association rule mining feature vector adapted to the improved Apriori association mining algorithm; S802: Input the generated job classification feature vector, quantized matching feature vector, and association rule mining feature vector into the corresponding algorithms to perform incremental learning; only update the local key parameters of each algorithm. The specific update logic is as follows: input the job classification feature vector into the improved C4.5 decision tree algorithm to update the penalty coefficient α; input the quantized matching feature vector into the attention mechanism BP neural network algorithm to update the attention layer weight matrix. and bias terms The feature vectors from association rule mining are input into the improved Apriori association mining algorithm to update the support threshold. S803: Based on the incremental learning of S802, multi-dimensional parameter dynamic adjustment is carried out; the learning rate of the attention mechanism BP neural network is adjusted by an exponential decay strategy; combined with the TF-IDF key trend features of the industry trend data, the minimum support threshold of the Apriori association mining algorithm is further optimized and improved; the feature weights of the logistics management talent profile and the logistics job demand profile are adjusted simultaneously, and L1 normalization is used after adjustment to map the weights to the [0,1] interval. S804: Periodically compare the matching accuracy, job retention rate, training utilization rate, and talent capability improvement before and after the S802 parameter update and S803 parameter adjustment. Set thresholds. If all indicators do not decrease after optimization and at least one indicator improves, the optimization is deemed effective, and the updated algorithm parameters and normalized feature weights are fixed. If any indicator decreases, the optimization is deemed ineffective, and the parameters and weights are rolled back to the configuration before optimization. This optimization process is re-executed after the next round of triggering conditions are met.

[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention proposes and constructs a dual-profile analysis model. By integrating multi-source heterogeneous data, it realizes a three-dimensional and multi-dimensional characterization of the characteristics of logistics management talents and job requirements, laying a solid data foundation for accurate matching and effectively solving the problems of single features and one-sided characterization in traditional methods.

[0016] (2) This invention integrates a hybrid intelligent matching algorithm that combines decision trees, improved Apriori association rules, and attention-based BP neural networks. This hybrid model combines classification filtering, association mining, and deep semantic matching capabilities, overcoming the limitations of single algorithms and significantly improving the accuracy, interpretability, and personalization of talent-job and talent-resource bidirectional matching.

[0017] (3) This invention introduces a dynamic optimization module with incremental learning capability. By monitoring key performance indicators in real time and adaptively adjusting algorithm parameters and feature weights, the system has the ability to continuously optimize and evolve, and can respond sensitively to the dynamic changes in the logistics industry. This solves the technical pain points of traditional static systems, such as lagging recommendation effect and poor adaptability. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the framework of the system of the present invention.

[0019] Figure 2 This is the interface design diagram of the system of the present invention.

[0020] Figure 3 This is a heatmap showing the matching and correlation between talent profile features and job requirement features in the system of this invention.

[0021] Figure 4 This is a heatmap showing the changes in dynamic optimization cycle and monitoring indicators of the system of the present invention.

[0022] Figure 5 This is a heat map showing the cost of matching training resources to address the talent capability gap in the system of this invention.

[0023] Figure 6 This is a heatmap of the feature weights of the hybrid algorithm-attention mechanism in the system of this invention.

[0024] Figure 7 This is a talent-job matching degree distribution diagram for the system of this invention. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7This application provides a detailed description of the technical solutions provided in each embodiment.

[0027] This application provides a flowchart illustrating the specific framework of a big data-based logistics management talent training matching and personalized recommendation system. Figure 1 As shown.

[0028] The system includes a data acquisition module, a dual-profile construction module, a hybrid intelligent matching and recommendation module, and a dynamic optimization module. The specific steps are as follows: S1: Collect multi-source heterogeneous data through the data acquisition module. The multi-source heterogeneous data includes logistics management talent data, job demand data, training resource data, and industry trend data. S2: The dual-profile construction module extracts, cleans, and fuses features from the collected multi-source heterogeneous data to construct profiles of logistics management talents and logistics job requirements, respectively. S3: Based on the profile of logistics management talent and the profile of logistics job requirements, the hybrid algorithm used in the hybrid intelligent matching recommendation module is used to complete the two-way matching of logistics management talent to logistics job and logistics management talent to logistics training resources. S4: Based on the bidirectional matching results, the dynamic optimization module monitors two types of indicators in real time: one is the matching effect indicator, including matching accuracy, job retention rate, training utilization rate, and talent capability improvement; the other is the update data indicator, which is measured by the xin increment of the multi-source heterogeneous data. When the matching effect indicator or the update data indicator is lower than the preset threshold, the optimization process is triggered.

[0029] The data sources for logistics management talent include enterprise HR systems, resume interfaces of job search platforms such as Zhaopin.com and 51job.com, university employment management systems, certificate databases of vocational skills assessment centers, talent past performance evaluation systems, and psychological testing platforms.

[0030] Data types are divided into two categories: structured and unstructured. Structured data includes age, gender, education, major, years of work experience, skills certificates, salary history, work location history, performance rating, and number of project participations. Unstructured data includes personal resume texts, job performance reports, peer evaluations, project achievement descriptions, and career planning questionnaires.

[0031] In terms of data collection methods, data is synchronized daily incrementally by connecting to structured data source systems via API interfaces; web crawlers are used to crawl publicly available resumes from job search platforms, and OCR technology is used to identify key information in scanned certificates and resumes; natural language processing technology is used to perform word segmentation and keyword extraction operations on unstructured text data.

[0032] The sources of job demand data include enterprise recruitment management systems, job posting interfaces of recruitment websites such as BOSS Zhipin and Liepin, internal job description databases of enterprises, and human resource planning reports.

[0033] Data types are also divided into structured and unstructured: Structured data includes job title, department, salary range, work location, number of openings, education requirements, major requirements, years of work experience requirements, skill certificate requirements, and job responsibility number; Unstructured data includes job responsibility description, job qualification details, team atmosphere description, and career path description.

[0034] The data collection method uses an API interface to synchronize enterprise recruitment needs in real time, crawls new job information from recruitment websites daily, and transforms unstructured job descriptions into standardized skill tags through text structuring.

[0035] The sources of training resources include online education platforms such as Coursera and China University MOOC, corporate internal training systems, vocational school training base resource databases, industry association training project databases, and logistics textbook publishing databases.

[0036] Data types are divided into structured and unstructured: Structured data includes resource name, type, skill coverage tags, difficulty level, learning time, training cost, pass rate, and update time; Unstructured data includes resource description, course outline, practical training project task list, and textbook catalog.

[0037] The data collection method involves connecting to the open interface of the education platform to obtain basic resource information, and using PDF parsing technology to extract the skill coverage dimensions from textbooks and course outlines to generate standardized tags.

[0038] The sources of industry trend data include the State Post Bureau's policy release platform, the Ministry of Transport's industry dynamics column, logistics industry reports from iResearch and Analysys, and annual reports of listed companies' logistics businesses.

[0039] The data collection method uses the Scrapy web crawler framework to capture policy documents and market analysis reports from the target platform and store them as a document set D.

[0040] Key trend feature extraction is implemented based on the TF-IDF algorithm. The specific steps are as follows: First, each document d in the document set D is segmented, stop words are removed, and meaningless words are eliminated to obtain the term set { The second step is to calculate the term frequency (TF) of each term t in document d, which is the number of times t appears in d divided by the total number of terms in d, denoted as TF(t,d); the third step is to calculate the inverse document frequency (IDF) (t,D), expressed by the formula: log(total number of documents in document set D / (number of documents containing term t + 1)) Adding 1 is to avoid the parents being 0; fourth step calculation Sorted in descending order of TF-IDF values, the top 50 terms were selected as key trend features; among them, The score represents the critical importance of term t in document d relative to document set D.

[0041] The dual-profile construction module extracts, cleans, and fuses features from the collected multi-source heterogeneous data to construct profiles of logistics management talents and logistics job requirements, respectively. In the feature extraction stage, structured data directly extracts fields as basic features, including education level, years of work experience, skill certificates, salary range, etc.; unstructured data uses natural language processing technology to extract features, such as extracting skill tags from resume text through keyword matching, extracting potential features from performance comments through sentiment analysis, and extracting core responsibility tags from job descriptions through topic modeling (LDA).

[0042] In the data cleaning process, for missing value handling, numerical data such as years of work experience and training costs are filled with the mean, while categorical data such as professional and skill certificates are filled with the mode. Samples with missing key features, such as core skills, are directly removed. For outlier handling, the IQR method is used to identify outliers. For example, samples with more than 30 years of work experience or negative training costs are marked as outliers and removed if they exceed the range of [Q1-1.5IQR, Q3+1.5IQR].

[0043] The data fusion adopts a weighted fusion strategy, which assigns weights to data from different sources according to their credibility. The weight of data from the enterprise HR system is 0.8, the weight of data from the recruitment platform is 0.6, and the weight of data from third-party assessments is 0.7. The data is processed to the [0,1] interval through Min-Max normalization to achieve scale uniformity and generate the fused feature vector.

[0044] A profile of logistics management talent is constructed based on four characteristics: basic attributes, explicit abilities, hidden potential, and career aspirations. The basic attributes are set with a weight of 0.2; age is quantified as 1 for 20-30 years old, 2 for 31-40 years old, and 3 for 41-50 years old; gender is quantified as 1 for male and 0 for female; education level is quantified as 1 for junior college, 2 for bachelor's degree, and 3 for master's degree and above; major is quantified as 3 for logistics management, 2.8 for supply chain management, 2.5 for transportation, 2 for other related majors, and 1 for unrelated majors; place of origin is quantified as 3 for local, 2 for surrounding provinces, and 1 for other places based on the degree of matching with the target job location.

[0045] The explicit competencies are weighted at 0.4; professional skills are quantified as follows: logistics planning, warehouse management, transportation scheduling, data analysis, WMS system operation, and customs declaration and inspection, with weights of 3, 2.8, 2.7, 2.6, 2.5, and 2.4 respectively; foreign language proficiency is quantified as follows: English CET-6 and above, CET-4, and others, with weights of 3, 2, and 1 respectively; computer skills are quantified as follows: Python / SQL, advanced Excel, and basic operations, with weights of 3, 2, and 1 respectively; and certificate holdings are quantified as follows: logistics professional level 1, level 2, level 3, and other certificates, with weights of 3, 2.5, 2, and 1 respectively, based on industry recognition.

[0046] The hidden potential is weighted at 0.2; learning ability is calculated by the sum of past training pass rate × 0.3 and performance evaluation learning score × 0.7; communication and coordination ability is calculated by the sum of project collaboration evaluation × 0.5 and interview score × 0.5; problem-solving ability is calculated by the sum of past fault handling cases × 0.4 and performance score × 0.6; innovation ability is calculated by the sum of the number of optimized solutions × 0.6 and leadership evaluation × 0.4.

[0047] The weight of career aspirations is set at 0.2; salary expectations are quantified by matching degree with the salary range of the position, with complete matching, close matching, and mismatch as 3, 2, and 1 respectively; work location is quantified by consistency with the position location, with consistent, nearby, and different locations as 3, 2, and 1 respectively; job type is quantified by management, technical, and operational positions as 3, 2.5, and 2 respectively; and development path is quantified by promotion orientation, stability orientation, and skills improvement orientation as 3, 2, and 2.8 respectively.

[0048] A profile of logistics job requirements is constructed based on four characteristics: skill requirements, job responsibilities, implicit fit, and career development opportunities. The skill requirements are weighted at 0.4; corresponding to the explicit competency dimension of talent, the minimum requirements for each skill for the position are quantified. For example, a warehouse supervisor needs to meet the requirements of warehouse management. 2.8 WMS System Operation 2.5 Team Management 2.6.

[0049] The required responsibilities are weighted at 0.3; quantified according to the core responsibilities of the position, for example, logistics solution design = 3, inventory optimization = 2.8, supply chain coordination = 2.7, team management = 2.6.

[0050] The weight of the stealth adaptation setting is 0.15; the stress resistance requirement is quantified as 3, 2, and 1 for high, medium, and low respectively; the teamwork requirement is quantified as 3, 2, and 1 for high, medium, and low respectively; and the sense of responsibility requirement is quantified as 3, 2, and 1 for high, medium, and low respectively.

[0051] The development space is weighted at 0.15; the clarity of the promotion channel is quantified as 3 for clear, 2 for relatively clear, and 1 for unclear; the training opportunity frequency is calculated per year. Three promotions, one to two promotions, and none are quantified as 3, 2, and 1 respectively; career growth rate is quantified as 3, 2, and 1 for promotions in 1 to 2 years, promotions in 3 to 5 years, and no promotions.

[0052] Based on the aforementioned profiles of logistics management talent and logistics job requirements, a hybrid algorithm employed in the hybrid intelligent matching and recommendation module is used to achieve a two-way matching process between logistics management talent and logistics positions, and between logistics management talent and logistics training resources. The hybrid intelligent matching and recommendation module employs a hybrid algorithm that combines a decision tree algorithm, an improved Apriori association mining algorithm, and an attention-based backpropagation neural network algorithm. The decision tree algorithm described employs an improved C4.5 decision tree algorithm, aiming to classify job types. It optimizes the information gain ratio calculation by introducing a feature redundancy penalty factor, expressed by the following formula: in, Indicates sample characteristics; Represents a sample set; For features of the sample set Information gain; This is the intrinsic value of feature a; This is the penalty coefficient; For feature redundancy; Indicates that sample feature 'a' is a sample set. The resulting increase in information.

[0053] Regarding sample set and feature selection, sample set We selected 100,000 talent-to-job matching data points from the past three years, of which 60,000 were successful and 40,000 were unsuccessful. The features selected were the core dimensions of the profile, including education, major, years of work experience, core skills, and job type requirements.

[0054] In the key parameter settings, the penalty coefficient The optimal value was determined to be 0.15 through 5-fold cross-validation, which is used to balance information gain and feature redundancy; feature redundancy The correlation between features is calculated using the Pearson correlation coefficient, expressed by the following formula: in, Let be the correlation coefficient between features i and j; k be the total number of features; the higher the correlation, the higher the correlation coefficient. The larger the value, the stronger the penalty; inherent value in, The percentage of samples for each value of feature a.

[0055] Here's an example of information gain ratio: when the sample feature is core skills, the sample set... The percentage of successfully matched samples was 60%. , , ,but .

[0056] Decision tree construction process according to Select splitting features from high to low, recursively construct a decision tree, and when the number of samples at a node... 50 or information gain rate The splitting process stops at 0.05, ultimately generating 12 job category nodes, including warehouse management, transportation scheduling, logistics planning, etc.

[0057] Based on the logistics management talent profile and the logistics job requirement profile, profile features are extracted. The profile features include talent features and job features. The profile feature extraction includes feature selection, feature encoding and vector construction. The feature selection is based on the feature weights of the logistics management talent profile and the logistics job requirement profile, prioritizing the selection of core features with a weight ratio greater than a set value, and eliminating redundant features with a weight ratio less than a set value. The feature encoding involves using one-hot encoding or label encoding for categorical core features and standardizing numerical core features. The vector construction involves multiplying the encoded feature values ​​by their corresponding feature weights and combining them in a fixed dimensional order to construct talent feature vectors and job feature vectors, respectively. The value of each dimension in the vector is equal to the encoded feature value multiplied by the corresponding feature weight. The extracted talent feature vector is input into the improved C4.5 decision tree algorithm to classify job types, initially screening out candidate jobs of the same type that match the talent's career aspirations. Then, based on the association rule base constructed by the improved Apriori association mining algorithm, the suitability of the talent's core skill combination with the job skill requirements is compared, further eliminating candidate jobs whose skill combinations do not meet the core job requirements. Finally, the job feature vector and talent feature vector of the remaining candidate jobs are concatenated to form a fused feature vector, which is input into the attention mechanism BP neural network algorithm to output a quantitative matching score. Finally, a job recommendation list is generated in descending order of the score. The extracted profile features are preprocessed and discretized into Boolean variables. The improved Apriori association mining algorithm uses a vertical data storage format, where each row stores one feature and each column stores the sample ID containing that feature. A bidirectional scanning mechanism and a frequent itemset pre-pruning strategy are combined to optimize the mining efficiency of frequent itemsets. The support threshold is dynamically adjusted based on the frequent itemset type to mine association rules between talent and job positions, and between talent and training resources, thus constructing an association rule base.

[0058] In the application embodiment, "Logistics Specialist Certificate = Yes" corresponds to the sample ID set {1,3,5,...,9999}, and "Warehouse Management Experience ≥ 3 Years = Yes" corresponds to the sample ID set {1,4,6,...,9998}. The bidirectional scanning mechanism generates 1-itemsets and 2-itemsets through forward scanning and prunes infrequent itemsets through reverse scanning, reducing the number of scans by more than 30%. The dynamic support threshold is adjusted according to the itemset type; the support threshold for popular positions such as logistics planning is set to 0.15, while for less popular positions such as cold chain logistics it is set to 0.08. The support threshold for training resource association rules is set to 0.12. Frequent itemset pre-pruning stipulates that if a subset of a k-itemset is not a frequent itemset, the k-itemset is directly removed without calculating support. The binary bitwise AND operation for intersection converts the sample ID set corresponding to the itemset into a binary vector. For example, if sample ID = 1 corresponds to the first bit being 1, the intersection is calculated using the & operator. ={1,3,5}, which is represented as 10101 in binary. ={1,4,6}, binary representation is 100110, intersection =10101&100110=00000, indicating no intersection.

[0059] The association rule base was constructed using a confidence level of at least 0.7 as the criterion for selection, ultimately generating over 500 valid rules. For example, when an employee possesses a logistics specialist certificate, has at least four years of warehouse management experience, and is familiar with WMS system operations, their suitability for the warehouse supervisor position has a confidence level as high as 0.85.

[0060] In the network structure design of the attention mechanism BP neural network algorithm, the number of nodes in the input layer equals the total dimension of the profile features, which include talent features and job features, totaling 50 nodes. The hidden layer consists of two fully connected layers: the first layer has 64 nodes using ReLU activation, and the second layer has 32 nodes using ReLU activation. The attention layer is located between the input layer and the first hidden layer, and is activated by a weight matrix. and bias terms Calculate the attention weight for each feature, as shown in the mathematical table below: in, Let be the attention weights assigned to the input features for the i-th input feature; Let i be the i-th input feature value; This is the attention layer weight matrix; is the attention layer bias term; z is the total number of features, z=50; the output layer has 1 node, the activation function is Sigmoid, and the output is a quantized matching score of 0-100.

[0061] In the training parameter settings, the Adam optimizer was selected, with an initial learning rate of 0.001 and a decay coefficient of [missing value]. , ,epsilon= The loss function uses the mean squared error loss, and the formula is as follows: in The total number of samples; Score the model predictions; The actual matching result is used; a score of 100 is given when the matching result is successful, and a score of 0 is given when the matching result fails; the training rounds are set to 100 rounds, the batch size is 32, and an early stopping strategy is adopted, which is to stop training when the loss of the validation set does not decrease after 3 rounds.

[0062] In the two-way matching process for logistics management talent to logistics positions, the first step is classification and screening. The talent feature vector is input into an improved C4.5 decision tree, which outputs the position type and filters out all candidate positions under that type. The second step is association rule matching. The core features in the talent profile are extracted and matched with the position matching rules in the association rule base. Positions that do not meet the rules are eliminated, leaving 20 candidate positions. The third step is quantitative scoring and ranking. The talent feature vector is concatenated with the position feature vectors of the 20 candidate positions to form a fusion vector. The fusion vector is input into a trained attention mechanism BP neural network, which outputs a matching score for each position. A job recommendation list is generated in descending order of the scores. In this implementation case, the Top 10 are selected, and core matching points are marked, such as a 95% matching score for warehouse management experience.

[0063] In the two-way matching process for logistics management talent and logistics training resources, the first step is to identify skill gaps. This involves comparing the profiles of logistics management talent and the requirements of logistics positions, calculating the differences in each feature dimension, and identifying dimensions with a difference ≥ 0.5 as skill gaps. For example, if a talent's data analysis ability is 2.0 and the target position requirement is 2.6, the difference is 0.6, thus identifying it as a gap. The second step is to select suitable resources. Based on the identified skill gap characteristics, the association rule base constructed using the improved Apriori association mining algorithm is retrieved, initially selecting 15 resources. Subsequently, the skill coverage dimensions, difficulty levels, and the talent's own learning rate of the training resources are combined to narrow down the resource range and form multiple candidate learning paths. Finally, the path optimization objective is to minimize the weighted average of learning time and cost, expressed as the objective function expression: in, The overall cost of the learning path; Weighting by learning time =0.5; Total study time; As a weight for learning costs, ; For the total learning cost, and + =1. The constraints are that the skill coverage dimension of the learning resources is greater than 90% of the weakness dimension, and the difficulty level is less than the talent's current ability level plus 1. Path generation calculates the comprehensive cost (Cost) by traversing all resource combinations and selecting the combination with the lowest cost as the optimal learning path. In the example, Talent B has a weakness in data analysis. Candidate resources include: A. Course: 25 hours, cost 600 yuan; B. Practical training: 30 hours, cost 1200 yuan; C. Textbook: 15 hours, cost 80 yuan. Combination 1, the cost of A plus B = 0.5 × (25 + 30) + 0.5 × (600 + 1200) = 27.5 + 900 = 927.5; Combination 2, the cost of A plus C is 0.5×(25+15)+0.5×(600+80)=20+340=360, but the skill coverage dimension is only 85%, which does not meet the constraint; Combination 3, the cost of B plus C is 0.5×(30+15)+0.5×(1200+80)=22.5+640=662.5, and the skill coverage dimension is 92%. Therefore, combination 3 is chosen as the optimal path.

[0064] Finally, the dynamic optimization module monitors the matching effect and data update in real time, and dynamically adjusts the profile feature weights and recommendation algorithm parameters to improve the system's matching accuracy.

[0065] The dynamic optimization module is specifically implemented as follows: The updated data is standardized using Min-Max to unify the units of measurement. Outliers are identified and removed using the IQR method. Numerical missing values ​​are filled with the mean, and categorical missing values ​​are filled with the mode. Samples lacking core features are directly removed. Subsequently, the basic attributes, explicit abilities, hidden potential, and career aspirations of the logistics management talent data are extracted after standardization. The skill requirements, responsibilities, implicit fit, and development space of the job demand data are extracted after standardization. Categorical features are processed using one-hot encoding or label encoding, and numerical features are standardized. They are then combined in a fixed dimensional order to generate job classification feature vectors adapted to the improved C4.5 decision tree algorithm and quantitative matching feature vectors adapted to the attention mechanism BP neural network algorithm. At the same time, skill coverage-related features are extracted from the standardized training resource data. Combined with the TF-IDF key trend features of the standardized industry trend data, after the above encoding and standardization, they are combined in a fixed dimensional order to generate association rule mining feature vectors adapted to the improved Apriori association mining algorithm. The generated job classification feature vector, quantization matching feature vector, and association rule mining feature vector are input into the corresponding algorithms to perform incremental learning. Only the local key parameters of each algorithm are updated, without retraining the entire model. The parameter update formula is as follows: in, The updated parameters; The parameters were before the update; For incremental learning rate, ; This represents the loss gradient for newly added data; in the example, if 1000 new talent-job matching data points are added, the loss gradient between the model's predicted score and the actual hiring result is calculated. Original parameters After the update .

[0066] The specific update logic involves inputting the job classification feature vector into the improved C4.5 decision tree algorithm and updating the feature redundancy penalty coefficient. The quantized matching feature vector is input into the attention mechanism BP neural network algorithm to update the attention layer weight matrix. and bias terms The feature vectors from association rule mining are input into the improved Apriori association mining algorithm to update the dynamic support threshold. Based on incremental learning, multi-dimensional parameter dynamic adjustment is carried out. An exponential decay strategy is used to adjust the learning rate of the attention mechanism BP neural network, with the following formula: in, Learning rate ; The initial learning rate; For the number of training rounds; This is the decay step size; This is the attenuation coefficient. , =100, and the learning rate decays to 90% of the original value after every 100 training rounds; the minimum support threshold of the Apriori association mining algorithm is further optimized and improved by combining the key trend features of TF-IDF with industry trend data to enhance the sensitivity of emerging demand adaptation. The minimum support threshold of Apriori is adjusted according to industry trend features. For example, when smart logistics becomes popular, the support threshold of related job or resource association rules is reduced to improve the matching sensitivity of emerging demand; the profile feature weights are mapped to the [0,1] interval using L1 normalization to ensure that the sum of the weights is 1. For example, after adjusting the weight of skill demand features from 0.3 to 0.35, the weight of basic attributes is simultaneously reduced from 0.2 to 0.15 to maintain the balance of the total weight.

[0067] To compare the matching accuracy, job retention rate, training utilization rate, and talent capability improvement rate before and after parameter updates and adjustments each quarter; the matching accuracy rate is calculated as the number of talents who accept and successfully join recommended jobs divided by the total number of recommended jobs, multiplied by 100%, with a preset threshold of 80%; the job retention rate is calculated as the number of people still employed 6 months after joining divided by the total number of people who joined through system recommendations, multiplied by 100%, with a preset threshold of 75%; the training utilization rate is calculated as the number of recommended training resources completed by talents divided by the total number of recommended training resources, multiplied by 100%, with a preset threshold of 60%; the talent capability improvement rate is calculated as the skill score after learning training resources minus the skill score before learning, divided by the skill score before learning, multiplied by 100%, with a preset threshold of 30%; the data update threshold is set to trigger optimization when the proportion of new data to the total data volume is ≥20%.

[0068] An optimization effect evaluation is conducted quarterly, comparing the changes in monitoring indicators before and after optimization. If the improved indicators are greater than 5%, the adjusted parameters are retained; if the improvement is less than 3%, the parameters are rolled back to those before optimization, and the optimization strategy is readjusted.

[0069] The interface diagram of the present invention is as follows: Figure 2 As shown, it mainly includes the HR user interface, the talent user interface, and the administrator interface.

[0070] Figure 3A heatmap visualizing the matching strength between talent profile features and job requirement features is presented. The horizontal axis represents the eight core features of the logistics job requirement profile, and the vertical axis represents the eight core features of the logistics management talent profile. The heatmap value ranges from 0 to 1, with higher values ​​indicating stronger matching strength between features. The graph clearly shows that the correlation strengths between education level and education requirements, and work experience and experience requirements, are 0.92 and 0.90 respectively, belonging to hard matching dimensions. The correlation strengths between logistics planning skills and logistics solution design, warehouse management skills and warehouse management responsibilities, and data analysis ability and data tool ability all exceed 0.85, belonging to core skill matching dimensions. The correlation strengths between salary expectations and salary range, and between job type requirements and job type, are 0.91 and 0.96 respectively, belonging to career requirement matching dimensions. This heatmap visually verifies the accuracy of the dual-profile construction and provides a clear direction for improving the Apriori association mining algorithm, ensuring the effectiveness of core rules in the association rule base.

[0071] Figure 4 This is a heatmap showing the changing trends of the four core monitoring indicators over 12 optimization cycles. The horizontal axis represents the optimization cycle, and the vertical axis represents the indicator value. The intensity of the heatmap value corresponds to the level of the indicator value. Monitoring data shows that the matching accuracy gradually improved from 80% in Q1 to 89% in Q12, the job retention rate increased from 78% to 85%, the training utilization rate increased from 60% to 72%, and the improvement in talent capabilities increased from 28% to 38%. Furthermore, all indicators tended to stabilize after Q8. This trend fully validates the effectiveness of the dynamic optimization module: by incrementally learning and updating model parameters, and dynamically adjusting feature weights and algorithm thresholds, the system's matching accuracy and resource utilization efficiency have continuously improved, solving the technical pain point of lagging recommendation performance in traditional static systems.

[0072] Figure 5 This heatmap visualizes the matching cost between talent skill gaps and training resources. The horizontal axis represents five core training resources, and the vertical axis represents six common talent skill gaps. The heatmap values ​​are matching cost coefficients, with 1 representing low cost and high matching, and 5 representing high cost and low matching. The graph clearly shows that: data analysis skills gaps are matched with practical training projects; customs declaration and inspection skills gaps are matched with professional textbooks; logistics planning and warehouse management skills gaps are matched with online courses; and certification exams have a matching cost coefficient of 5 for all gaps. This heatmap provides direct data support for learning path optimization. The hybrid intelligent matching recommendation module can quickly generate personalized learning paths with the lowest overall cost based on the cost coefficients in the heatmap, combined with learning time weights and cost weights.

[0073] Figure 6A heatmap showing the feature weight allocation for the three core algorithms in the hybrid algorithm is presented. The horizontal axis represents the 10 system input features, and the vertical axis represents the three algorithms: improved C4.5 decision tree, improved Apriori association mining, and attention-based BP neural network. The heatmap values ​​show that the attention-based BP neural network assigns the highest weight to explicit abilities and skill requirements, significantly higher than the other two algorithms; the improved Apriori association mining assigns slightly higher weight to career aspirations and job requirements; and the improved C4.5 decision tree assigns relatively prominent weight to basic attributes and job type. This graph verifies the complementarity of the hybrid algorithms: by strengthening the weight of core skill features through the attention mechanism, and combining the classification ability of the decision tree with the rule mining ability of the association algorithm, the accurate utilization of multi-dimensional features is achieved, supporting the accuracy and personalization of bidirectional matching.

[0074] Figure 7 A histogram showing the distribution of talent-job matching scores is presented, with the horizontal axis representing the matching score and the vertical axis representing the number of talents with the corresponding score. The distribution results show that matching scores are concentrated in the 80-90 score range, accounting for 48.2%, with an average matching score of 85.2. The low matching score range of 60-70 points accounts for only 2.3%. This distribution characteristic fully demonstrates the accurate matching effect of the hybrid intelligent matching recommendation module: through a three-level matching process of decision tree classification screening, association rule matching, and neural network quantification scoring, low-fit jobs are effectively filtered out, ensuring a high proportion of highly matched jobs in the recommended list, reducing recruitment costs for enterprises and job search costs for talents, and verifying the technical advantages of this invention in improving the efficiency of talent-job matching.

[0075] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data-based logistics management talent training matching and personalized recommendation system, characterized in that, The system includes a data acquisition module, a dual-profile construction module, a hybrid intelligent matching and recommendation module, and a dynamic optimization module. The specific steps are as follows: S1: Collect multi-source heterogeneous data through the data acquisition module. The multi-source heterogeneous data includes logistics management talent data, job demand data, training resource data, and industry trend data. S2: The dual-profile construction module extracts, cleans, and fuses features from the collected multi-source heterogeneous data to construct profiles of logistics management talents and logistics job requirements, respectively. S3: Based on the profile of logistics management talent and the profile of logistics job requirements, the hybrid algorithm used in the hybrid intelligent matching recommendation module is used to complete the two-way matching of logistics management talent to logistics job and logistics management talent to logistics training resources. S4: Based on the bidirectional matching results, the dynamic optimization module monitors two types of indicators in real time: one is the matching effect indicator, including matching accuracy, job retention rate, training utilization rate, and talent capability improvement; the other is the update data indicator, which is measured by the xin increment of the multi-source heterogeneous data. When the matching effect indicator or the update data indicator is lower than the preset threshold, the optimization process is triggered.

2. The big data-based logistics management talent training matching and personalized recommendation system according to claim 1, characterized in that, The industry trend data is obtained by regularly crawling logistics industry policy documents and market analysis reports using web crawlers, and key trend features are extracted using the TF-IDF algorithm. The formula is expressed as: Where t represents a term; d represents a document; and D represents a document set; The word frequency of term t in document d; The inverse document frequency of term t in document set D; The score represents the critical importance of term t in document d relative to document set D.

3. The big data-based logistics management talent training matching and personalized recommendation system according to claim 1, characterized in that, The profile of logistics management talent includes four characteristics: basic attributes, explicit abilities, hidden potential, and career aspirations. The logistics job requirements profile includes four characteristics: skill requirements, job responsibilities, implicit fit, and career development opportunities.

4. The big data-based logistics management talent training matching and personalized recommendation system according to claim 1, characterized in that, The hybrid algorithm includes a decision tree algorithm, an improved Apriori association mining algorithm, and an attention mechanism BP neural network algorithm. The decision tree algorithm described uses an improved C4.5 decision tree algorithm, which optimizes the information gain ratio by introducing a feature redundancy penalty factor, expressed by the following formula: in, Indicates sample characteristics; Represents a sample set; For sample feature a, the sample set Information gain; This is the inherent value of sample feature a; This is the penalty coefficient; For feature redundancy; Indicates that sample feature 'a' is a sample set. The resulting increase in information. The improved Apriori association mining algorithm adopts a vertical data storage format. Based on the structured storage characteristics of the vertical data storage format, it combines a bidirectional scanning mechanism and a frequent itemset pre-pruning strategy to optimize the mining efficiency of frequent itemsets. It also dynamically adjusts the support threshold according to the frequent itemset type to mine association rules between talent and job positions, and between talent and training resources, and to build an association rule base. The attention mechanism BP neural network algorithm is a neural network constructed from an input layer, an attention layer, a hidden layer, and an output layer. The attention layer strengthens the feature weights, and the Adam optimizer is used to optimize the loss function. The output layer outputs a quantized matching score. The mathematical expression for the feature weights is: in, Let be the attention weight for the i-th input feature; Let i be the i-th input feature value; This is the attention layer weight matrix; is the attention layer bias term; z is the total number of features.

5. The big data-based logistics management talent training matching and personalized recommendation system according to claim 4, characterized in that, The bidirectional scanning mechanism includes forward scanning and reverse scanning; the forward scanning is to traverse from low-order frequent itemsets to high-order frequent itemsets layer by layer to generate candidate frequent itemsets; the reverse scanning is to decompose the candidate frequent itemsets generated by the forward scanning in reverse to generate frequent item subsets.

6. The big data-based logistics management talent training matching and personalized recommendation system according to claim 4, characterized in that, The pre-pruning strategy uses frequent itemsets for intersection calculation; the intersection calculation of frequent itemsets uses a bitwise AND operation, expressed by the following formula: in, The intersection of two frequent itemsets; and Let the frequent itemsets be the ones to be intersected. For binary and operators.

7. The big data-based logistics management talent training matching and personalized recommendation system according to claim 1 or 4, characterized in that, The matching of logistics management talent to logistics positions specifically includes: Based on the logistics management talent profile and the logistics job requirement profile, profile features are extracted. The profile features include talent features and job features. The profile feature extraction includes feature selection, feature encoding and vector construction. The feature selection is based on the feature weights of the logistics management talent profile and the logistics job requirement profile, prioritizing the selection of core features with a weight ratio greater than a set value, and eliminating redundant features with a weight ratio less than a set value. The feature encoding involves using one-hot encoding or label encoding for categorical core features and standardizing numerical core features. The vector construction involves multiplying the encoded feature values ​​by their corresponding feature weights and combining them in a fixed dimensional order to construct talent feature vectors and job feature vectors, respectively. The value of each dimension in the vector is equal to the encoded feature value multiplied by the corresponding feature weight. The extracted talent feature vector is input into the improved C4.5 decision tree algorithm to classify job types, initially screening out candidate jobs of the same type that match the talent's career aspirations. Then, based on the association rule base constructed by the improved Apriori association mining algorithm, the suitability of the talent's core skill combination with the job skill requirements is compared, further eliminating candidate jobs whose skill combinations do not meet the core job requirements. Finally, the job feature vector and talent feature vector of the remaining candidate jobs are concatenated to form a fused feature vector, which is input into the attention mechanism BP neural network algorithm to output a quantitative matching score. Finally, a job recommendation list is generated in descending order of the score. The matching of logistics management talent with logistics training resources specifically includes: First, the profiles of logistics management personnel and logistics job requirements are compared across feature dimensions, and the differences between each dimension are calculated to identify the personnel's skill gaps. Then, based on these skill gaps, the association rule base constructed using the improved Apriori association mining algorithm is retrieved to initially screen training resources whose skill coverage matches the skill gaps. Subsequently, the skill coverage dimensions, difficulty level, and the personnel's own learning rate of the training resources are combined to narrow down the range of training resources and form multiple candidate learning paths. Finally, the path optimization objective is to minimize the weighted average of learning time and cost, using the formula: in, The overall cost of the learning path; Weighted by learning time; Total study time; Weighted by learning cost; For the total learning cost, and + =1.

8. The big data-based logistics management talent training matching and personalized recommendation system according to claim 1, characterized in that, The optimization process steps are as follows: S801: Standardize the updated data, extract the features of each dimension of the standardized logistics management talent data and job demand data, and combine them in a fixed dimension order to generate a job classification feature vector adapted to the improved C4.5 decision tree algorithm and a quantitative matching feature vector adapted to the attention mechanism BP neural network algorithm; extract the skill coverage related features of the standardized training resource data, combine them with the TF-IDF key trend features of the standardized industry trend data, and combine them in a fixed dimension order to generate an association rule mining feature vector adapted to the improved Apriori association mining algorithm; S802: Input the generated job classification feature vector, quantized matching feature vector, and association rule mining feature vector into the corresponding algorithms to perform incremental learning; only update the local key parameters of each algorithm. The specific update logic is as follows: input the job classification feature vector into the improved C4.5 decision tree algorithm to update the penalty coefficient α; input the quantized matching feature vector into the attention mechanism BP neural network algorithm to update the attention layer weight matrix. and bias terms The feature vectors from association rule mining are input into the improved Apriori association mining algorithm to update the support threshold. S803: Based on the incremental learning of S802, it carries out dynamic adjustment of multi-dimensional parameters; The learning rate of the attention mechanism BP neural network is adjusted by an exponential decay strategy; the minimum support threshold of the Apriori association mining algorithm is further optimized and improved by combining the TF-IDF key trend features of the industry trend data; the feature weights of the logistics management talent profile and the logistics job demand profile are adjusted simultaneously, and L1 normalization is used after adjustment to map the weights to the [0,1] interval. S804: Regularly compare the matching accuracy, job retention rate, training utilization rate and talent capability improvement before and after S802 parameter update and S803 parameter adjustment. Set thresholds. If all indicators do not decrease after optimization and at least one indicator improves, the optimization is deemed effective. Fix the updated algorithm parameters and normalized feature weights. If any metric decreases, the optimization is deemed invalid, and the parameters and weights are rolled back to their original configuration. The optimization process will be re-executed once the triggering conditions for the next round are met.

Images