3.1. System Architecture of the HRA-TS Model

Traditional job recommendation systems typically match job seekers with positions based on static features, such as education, gender, and professional background. However, job seekers’ needs are determined not only by static attributes but also by dynamic behavioral factors such as skill development, job search activities, and career progression. Because traditional methods fail to capture these dynamic changes, such systems cannot adapt effectively to the evolving needs of job seekers. In addition, existing recommendation systems generally optimize short-term objectives such as matching accuracy, while overlooking the long-term potential for career development. Furthermore, with the growing importance of privacy protection regulations, the provision of personalized recommendations while preserving the privacy of job seekers has become a critical challenge.

To address these challenges, we propose a HRA-TS for job recommendations. This framework integrates LSTM networks to model user dynamic behaviors, employs GATs to capture the relationships between job seekers and job positions, and introduces a multimodal reward mechanism to balance short-term matching quality with long-term developmental objectives. Additionally, federated learning was incorporated to support the privacy-preserving model training. The framework consists of three modules.

Dynamic Profiling and Feature Encoding Module: This module employs an LSTM–GAT architecture to jointly encode the static attributes and dynamic behaviors of job seekers, thereby providing a comprehensive representation of their states.

Reward Function Design and Progressive Policy Optimization Module: This module designs a multimodal reward function that integrates matching relevance, developmental potential, and behavioral feasibility and progressively optimizes the recommendation policy through a curriculum learning mechanism.

Privacy Protection and Federated Learning Module: This module preserves jobseeker privacy through federated learning while improving the efficiency of distributed model training.

The following sections present the detailed mathematical formulation of each module together with rigorous derivations. Fig. 1 illustrates the overall architecture of the HRA-TS framework.

3.2. Dynamic Profiling and Feature Encoding

In job recommendation tasks, job seekers’ needs are shaped not only by static attributes (such as education, gender, and professional background), but also by evolving behavioral patterns (e.g., skill development and job search activities). Traditional methods that focus primarily on static features are insufficient to effectively capture such dynamic changes. To address this limitation, LSTM networks were employed to model sequential data and capture the temporal evolution of job-seeker behaviors. In addition, GATs were introduced to model the relational dependencies between job seekers and recommended positions.

3.2.2. LSTM-Based Dynamic Behavior Encoding

The goal of dynamic behavior encoding is to model evolving behavioral features of job seekers such as skill development, job search activities, and other career progression behaviors. These dynamic features change over time. To capture these temporal dependencies, we used LSTM networks, which are well suited for modeling time-series data.

Job seekers’ dynamic behavior is represented as sequential data \(S_{\textit{dynamic}}=\{s_t\}_{t=1}^T\), where \(T\) is the time step, and \(s_t=\mathbb{R}^k\) is the dynamic state at step \(t\). The LSTM recurrence relationships are as follows.

LSTM networks were used to process the time-series data and compute the hidden state \(h_t\) and cell state \(c_t\) at each time step. The LSTM network recursively updates these states based on the previous hidden state and current input feature \(s_t\).

Input Gate: Controls the extent to which the current input \(s_t\) is allowed to affect the cell state.

\begin{equation} \label{eq:3} i_t=\sigma\left(W_i s_t+U_i h_{t-1}+b_i\right). \tag{3} \end{equation}

Forget Gate: Decides how much of the previous cell state \(c_{t-1}\) should be forgotten.

\begin{equation} \label{eq:4} f_t=\sigma\left(W_f s_t+U_f h_{t-1}+b_f\right). \tag{4} \end{equation}

Output Gate: Determines how much of the current cell state \(c_t\) should be used to compute the hidden state \(h_t\).

\begin{equation} \label{eq:5} o_t=\sigma\left(W_o s_t+U_o h_{t-1}+b_o\right). \tag{5} \end{equation}

Update Cell State: The cell state \(c_t\) is updated based on the input and forget gates.

\begin{equation} \label{eq:6} c_t=f_t\cdot c_{t-1}+i_t\cdot \tanh\left(W_c s_t+U_c h_{t-1}+b_c\right). \tag{6} \end{equation}

Hidden State: The output hidden state \(h_t\) is computed based on the output gate and cell state.

\begin{equation} \label{eq:7} h_t=o_t\cdot \tanh\left(c_t\right), \tag{7} \end{equation}

where \(i_t\) denotes the input gate, \(f_t\) the forget gate, \(o_t\) the output gate, \(c_t\) the cell state, and \(h_t\) the hidden state. These recurrence relations enable LSTM to capture long-term dependencies in jobseeker behaviors.

Output: The output of the LSTM network is the hidden state \(h_t\), which captures the temporal dynamics of the job-seeker behavior at time step \(t\): \(h_t\in \mathbb{R}^d\).

Figure 2 shows a flowchart of the dynamic behavior encoding process. Dynamic behavior encoding using LSTM networks captures temporal dependencies and long-term relationships in the job seeker’s actions, such as skill improvements and job search activities. This provides a dynamic representation that complements the static feature encoding, allowing for a more complete profile of job seekers, including their evolving skills and career progression.

3.3. Multimodal Reward Function Design and Progressive Strategy Optimization

To optimize the long-term effects of the recommendation system, this module introduces a multimodal reward function that integrates matching degree, growth potential, and feasibility. These reward components are used to guide the system not only in short-term matches but also in ensuring the long-term career growth of job seekers.

The multimodal reward function is a critical component in the HRA-TS framework, guiding the system to make recommendations that not only match the job seeker’s current status, but also contribute to their long-term career growth. The design of the function reflects the joint optimization of both short-term goals (e.g., job matching accuracy) and long-term objectives (e.g., career growth potential). By incorporating both static feature encoding and dynamic behavior modeling, the reward function allows the system to adapt to the evolving needs of job seekers. Moreover, it is closely integrated with the framework’s policy optimization process and federated learning mechanism to ensure that job recommendations are tailored to users while respecting privacy constraints. Furthermore, a curriculum learning mechanism was introduced to progressively increase the complexity of the task during training, thereby preventing overfitting during the early stages.

3.4. Privacy Protection and Federated Learning

As privacy concerns increase, traditional centralized training methods are no longer sufficient to meet modern privacy protection requirements. Therefore, federated learning was adopted to train the models without exposing sensitive user data, thereby ensuring jobseeker privacy.

Within the HRA-TS framework, federated learning is introduced to enable model training, while safeguarding the privacy of job seekers. In contrast to traditional centralized training, federated learning distributes training tasks across multiple clients, thereby avoiding the centralization of sensitive data. Each client trains a model locally and updates the global model by passing gradients.

Client: The primary task of the client is to train the models using local data and to compute gradients. The client does not upload raw data but instead uploads encrypted gradients to ensure data privacy. Each client contains an LSTM module and a GAT module, processing dynamic behavioral and static features, respectively. LSTM module: Processes a job seeker’s time-series data, such as skill evolution and job search behavior, to generate a dynamic behavioral feature representation. GAT module: Models the relationship between job seekers and recommended positions, capturing interactions between job seekers and positions through a graph attention mechanism. The client’s gradient update formula is

Server: The server’s task is to aggregate all gradients uploaded by clients and use the federated averaging method to aggregate the gradients, thereby updating the global model.

LSTM Module (Dynamic Behavior Encoding): The LSTM network receives time-series features, such as skill development and job search activities, to generate a dynamic behavioral feature representation, \(h_t\). The LSTM captures long-term dependencies in job seekers’ behaviors and learns skill improvements and job search activities over time. Input features \(s_t\) include information such as the job seeker’s skills and job search activities at time step \(t\). The LSTM calculates the current hidden state \(h_t\) using the following formula:

This process generates a dynamic behavioral representation \(h_t\) of the job seeker at time step \(t\) for use by subsequent modules.

GAT Module (Job-Applicant Relationship Modeling): GAT models the relationship between job applicants and recommended positions by calculating the weights between nodes via a graph attention mechanism and updating node features. Each node represents an applicant or position, with the edges denoting their relationship. Attention coefficient \(\alpha\) between each applicant-position pair is computed through a self-attention mechanism to update node features:

This process generates the final representation \(h_{\textit{GAT}}\) for each node (job seeker and job position) through the graph attention mechanism, providing the basis for recommendation decisions.

Privacy Protection Mechanism: Within the federated learning framework, differential privacy techniques are introduced to ensure privacy protection. When each client uploads a gradient, privacy is safeguarded by the addition of noise.

Federated Learning Aggregation and Update: Clients upload locally trained gradients to the server. The server aggregates the gradients using federated averaging to update the global model.

This module ensures the protection of job seeker data privacy via federated learning, encryption techniques, and gradient aggregation to optimize the model training efficiency. Differential privacy further strengthens user privacy protection.

The proposed HRA-TS framework combines LSTM–GAT networks for dynamic feature modeling, designs a multimodal reward function with progressive strategy optimization, and incorporates federated learning for privacy protection. This framework addresses challenges in traditional job recommendation systems related to dynamic profiling, long-term value guidance, and privacy protection and demonstrates significant application potential for personalized job recommendations.

4.1. Explanation and Processing of Experimental Data

This study used employment-tracking data from 100 graduates who completed their studies at a domestic university in 2020 as the research sample, as presented in Table 1 and Fig. 3. This dataset contains information on students’ comprehensive performance and job search behavior during their time at university, including structured data from two dimensions: static attributes and dynamic behavior. In the static attribute dimension, the dataset integrated multiple basic characteristic indicators. Demographic characteristics include basic information such as gender and age, while educational background characteristics encompass academic performance indicators such as the field of study and grades in core courses. These static indicators form the basic profiles of graduates and provide important feature dimensions for personalized recommendation models.

The dynamic behavior dimension focuses on recording students’ skill development trajectories and specifically includes three types of time-series data. The first is professional skill development, which describes the students’ progress in professional competencies, language proficiency, and computer application skills. The second is honor awards, which systematically record students’ achievements in academic competitions, innovation and technology activities, and social practices. The third category comprises career preparation activities that comprehensively track internship experience, project participation, and career training, all of which are directly related to employment. These dynamic data are organized in chronological order to form time-series features that reflect the evolution of students’ professional competencies.

During the data preprocessing stage, standardized procedures were applied to clean and transform raw data. Static attribute data were standardized to eliminate differences in measurement units, whereas categorical variables were converted into numerical representations using one-hot encoding. Dynamic behavioral data were processed using time-series modeling to construct behavioral sequences that characterized the development of students’ abilities. For low-frequency behavioral features, statistical smoothing methods were employed to effectively alleviate data sparsity while preserving the underlying feature information.

To comprehensively evaluate the performance of the recommendation system, with particular emphasis on validating the applicability of the proposed model in data-scarce scenarios, this study employed a stratified random sampling method to partition the dataset. The training set comprised 80% of the samples and retained complete feature information, whereas the test set comprised the remaining 20% and simulated the cold-start conditions by masking partial feature information. The final experimental dataset covers seven academic disciplines, five categories of professional certification information, and five levels of honorary awards. All experimental data were strictly anonymized in compliance with the research ethics standards and data security management requirements.

In a study of a personalized career recommendation method for college students based on hierarchical temporal reinforcement learning, a detailed analysis of competency requirements across different occupational fields provided an important foundation for precise model matching. By illustrating the distribution of demand across dimensions, such as programming skills, foreign language proficiency, management capabilities, design capabilities, and engineering capabilities in various industries, the radar chart clearly reveals the distinct competency requirements of different sectors. This visual analysis not only provides an intuitive understanding of inter-industry differences in skill demands but also offers valuable guidance for the design of personalized career recommendation models. During model development, by incorporating the heterogeneity of industry-specific requirements, the recommendation algorithm can consider students’ foundational backgrounds more comprehensively while generating career development paths aligned with the skill priorities of different industries.

Figure 4 illustrates the differences in demand across five key competency dimensions—programming skills, foreign language proficiency, management skills, design skills, and engineering skills—in the five major industries: IT, finance, manufacturing, education, and healthcare. Specifically, the IT industry places the greatest emphasis on programming skills (0.85), whereas the demand for foreign language proficiency and design skills is relatively low. In contrast, the finance industry prioritizes management skills (0.80) and foreign language proficiency, while showing limited demand for technical skills. The manufacturing industry exhibited the strongest demand for engineering skills (0.90), with other competencies being comparatively less emphasized. The education industry shows a particularly strong demand for foreign language proficiency (0.90) but a comparatively weaker demand for technical skills. The healthcare industry places a strong emphasis on engineering skills (0.80) and demonstrates a greater reliance on technical competencies than most other industries.

4.2. Evaluation of a Recommendation Model Based on HRA-TS for Personalized Career Recommendations for College Students

To validate the performance advantages of the HRA-TS in employment and entrepreneurship platform recommendations, we selected multidimensional employment data from 100 students as the experimental sample. MATLAB was used as the development platform combined with Excel for data preprocessing, and the experiments was performed on a CPU-based platform. Through a recommendation system comparison experiment, the proposed HRA-TS algorithm was compared with the DQN, A3C, and PPO algorithms for performance analysis.

As shown in Figs. 5–8, the HRA-TS algorithm (blue markers) exhibits a widely dispersed distribution, covering the solution space more uniformly and with slight clustering observed in certain regions, indicating its ability to explore diverse career paths and provide personalized recommendations. In contrast, the DQN (purple markers) exhibited a more concentrated distribution, suggesting that it focused on a relatively limited range of career paths. A3C (red markers) and PPO (green markers) showed broader distributions than DQN, but still failed to achieve the level of diversity demonstrated by HRA-TS. Although both methods improve upon the DQN, they do not fully match the adaptability and flexibility of the HRA-TS in delivering personalized career recommendations.

Overall, HRA-TS outperformed the other models in terms of recommendation diversity and personalization, making it more suitable for providing students with comprehensive career path options.

Figure 9 illustrates the distribution characteristics of graduates’ career preferences grouped by major and gender. On the horizontal axis, the student groups are coded as follows: numbers 1–7 correspond to computer science, business administration, mechanical engineering, foreign languages and literature, data science, intelligent medical engineering, and aviation transportation, respectively, and the suffixes M and F denote male and female students, respectively. The vertical axis covers five career types: other technical positions, engineering and technical positions, further education abroad, self-employment, and civil service. The bubble size is positively correlated with the intensity of career preferences, thereby visually illustrating the differences in preferences across groups.

As shown in Fig. 9, the occupational preferences exhibited significant structural characteristics. Preferences for the civil service are most prominent among female students majoring in foreign languages, literature, and intelligent medical engineering, forming a distinct clustering pattern. Preferences for self-employment were highly concentrated among male students majoring in business administration, with a significantly higher preference intensity than the other groups. Preferences for further education abroad demonstrated cross-gender consistency among foreign language and literature majors, with both male and female students exhibiting comparable levels of preference intensity. Engineering and technical positions, however, exhibit significant gender differentiation, with male students in mechanical engineering and aviation transportation showing substantially stronger preferences than their female counterparts in the same majors, thereby creating a clear contrast. Notably, preferences for other technical positions were widely observed among students majoring in computer science and data science, indicating strong cross-gender generality. These structured preference patterns provide an important modeling basis for the hierarchical recommendation mechanism of the HRA-TS. By establishing a dynamic weight allocation strategy across major- and gender-related dimensions, the algorithm effectively enhanced the adaptability of the career recommendation system to complex scenarios.

The F1 score balances the precision and recall of a classification model, and is commonly used to evaluate the overall performance of recommendation systems. In this study, the F1 score was used to assess the effectiveness of the career recommendation system in terms of precision and coverage. A high F1 score indicates that the recommendation system not only accurately matches suitable positions but also effectively recommends multiple relevant positions, thereby demonstrating a strong overall recommendation performance. Therefore, the F1 score plays a crucial role in personalized career recommendations as it provides a comprehensive reflection of recommendation quality.

As shown in Figs. 10 and 11, the F1 score of the HRA-TS algorithm is generally higher than that of the DQN algorithm for different numbers of recommended positions and candidate users. In particular, as the number of recommended candidates increased, the F1 score of the HRA-TS algorithm steadily improved, demonstrating its superior performance when the recommendation set increased. In contrast, although the F1 score of the DQN algorithm improves as the number of recommended candidates increases, its overall performance remains inferior to that of HRA-TS. This indicates that HRA-TS demonstrates stronger precision and recall performance in personalized career recommendation tasks, enabling the system to provide users with more comprehensive and personalized recommendation results.

HR@5 measures the proportion of correct job positions among the top-5 recommendations. A higher HR@5 value indicates that the recommendation system can accurately match the more relevant job positions within the top-ranked results. In personalized career recommendations, HR@5 reflects the hit accuracy of the recommendation system and helps evaluate its performance under a fixed recommendation list length. Through the HR@5 value, the effectiveness of job matching can be assessed more directly. In particular, when more positions are recommended, an increase in HR@5 indicates that the system can retrieve more relevant jobs within the top-ranked results, thereby improving the overall matching effectiveness.

As shown in the HR@5 comparison results presented in Fig. 12, the HRA-TS consistently outperformed the DQN across all recommendation counts, with HR@5 improvements ranging from 4% to 7%. Specifically, HRA-TS achieved a 6.02% improvement in HR@5 compared to DQN (\(p < 0.01\)), demonstrating its superior ability to rank relevant positions within the top recommendations.

In contrast, although the DQN’s HR@5 value improves as the number of recommendations increases, it never reaches the level achieved by the HRA-TS, indicating an inferior top-K matching performance in personalized recommendation tasks. A3C and PPO both outperformed DQN but remained slightly behind HRA-TS, with improvements of 2%–3% for A3C and 1%–2% for PPO compared to DQN. These results highlight that HRA-TS is the most effective model for providing accurate and diverse career recommendations.

The job matching score (JMS) is an indicator used to measure the degree of alignment between the recommended positions and users’ career backgrounds, thereby reflecting the system’s adaptability. A higher JMS value indicates greater alignment between the recommended positions and the users’ career interests, skills, and backgrounds. In personalized career recommendation tasks, the JMS is used to assess the relevance and matching quality of recommended positions and serves as a key metric for evaluating whether the recommendation system can provide users with suitable positions.

As shown in Fig. 13, the JMS values of the HRA-TS (blue area) were consistently higher than those of the DQN (red area) across all recommendation counts. The JMS values of HRA-TS steadily increased as the number of recommendations increased, indicating its strong ability to align recommended positions with the students’ background characteristics. In contrast, the DQN (red area) shows only limited improvement, with JMS values ranging from 73.81% at 10 recommendations to 82.23% at 100 recommendations, highlighting its limitations in job matching. Both A3C (green area) and PPO (yellow area) outperformed DQN but remained below HRA-TS. A3C reached 85.6% at 100 recommendations, whereas PPO reached 84.7%, indicating better job-matching capability than DQN, although neither achieved the performance level of HRA-TS.

In summary, HRA-TS consistently outperformed DQN, A3C, and PPO in terms of job matching quality and career path suitability, thereby offering more personalized recommendations.

User course exploration path analysis tracks and analyzes students’ course selections and learning trajectories throughout the learning process to assess whether they follow a reasonable progression of “foundation \(\to\) major \(\to\) interdisciplinary extension.” Adherence to this path can reveal student learning behavior patterns and help optimize course recommendation systems to provide personalized and adaptive learning pathways. By analyzing these data, educational platforms can better understand students’ learning needs and design course recommendation strategies that are more closely aligned with the career development needs of students from different backgrounds and majors.

As shown in Table 2, students across different academic disciplines generally followed the “foundation \(\to\) specialization \(\to\) interdisciplinary” pathway in their course exploration. For example, 80.0% of computer science students followed a complete path, and students from other disciplines demonstrated a high level of path adherence (82.0%). Overall, 80.7% of students followed the recommended learning path, indicating that most students exhibited a relatively consistent pattern of course selection. Course recommendation systems effectively support students’ academic growth and development by optimizing their learning paths.

CTR@10 measures the proportion of clicks in the top ten recommended positions in a cold-start scenario. Cold-start issues typically arise when the system lacks historical data or when user preferences are not sufficiently established. A higher CTR@10 value indicates that the system can adapt better to scenarios with limited historical data and recommend more relevant positions to users. In personalized career recommendation tasks, the CTR@10 is an important metric for evaluating the cold-start capability of a recommendation system and reflects its effectiveness in addressing cold-start challenges.

As shown in Fig. 14, the HRA-TS algorithm (orange bars) consistently outperformed the DQN algorithm (green bars) in terms of the cold-start click-through rate (CTR@10) across all recommendation sizes. Specifically, HRA-TS achieved a 6.09% improvement in CTR@10 compared with the DQN when recommending 10 positions (\(p <0.01\)). As the number of recommended positions increased, the CTR@10 value of HRA-TS increased steadily and remained relatively high, demonstrating its superior ability to recommend relevant positions under cold-start conditions.

In contrast, DQN also demonstrated incremental improvement in CTR@10 as the number of recommended positions increased; however, its performance remained consistently below that of HRA-TS, indicating limited adaptability in cold-start scenarios. Similarly, A3C (purple bars) and PPO (yellow bars) outperformed DQN, with A3C achieving a CTR@10 value of 0.711 at 100 recommendations and PPO reaching 0.7095 under the same setting. However, both models remained inferior to HRA-TS, reaching 0.7421 at 100 recommendations.

Overall, the HRA-TS algorithm demonstrated superior cold-start performance, providing more relevant job recommendations than DQN, A3C, and PPO.

Federated learning is a distributed learning paradigm that allows multiple parties to participate in model training, while retaining the privacy of their local data. In federated learning, accuracy loss typically refers to the performance gap between the global model and locally trained models, and is commonly used to evaluate the impact of distributed training on model effectiveness. Federated learning was adopted to enable collaborative knowledge sharing across multiple nodes or devices without centralizing raw data, thereby preserving data privacy while improving model performance. To assess the effectiveness of federated learning, it is necessary to analyze the accuracy loss to understand the gap between local models and the global model and to evaluate the algorithm’s performance in a distributed environment.

As presented in Table 3, the accuracy loss of the HRA-TS algorithm under the federated learning setting was consistently lower than that of the DQN algorithm. Specifically, the average accuracy loss of HRA-TS was 2.9%, whereas that of the DQN was 5.2%, demonstrating the advantage of HRA-TS in preserving model accuracy during distributed training. This indicates that HRA-TS can better maintain the global model performance in a federated learning environment, thereby achieving more efficient model training and stronger overall performance while ensuring data privacy.