Multi-task learning algorithm design and implementation strategy for intelligent police academy management system

The analysis and design principles of the student management information system for police-oriented institutions of higher education are: 1)

Based on the specification requirements of the national standard and the ministry standard, follow the standard of top-level design. Adopting the analogy method, on the basis of summarizing and learning from the experience of similar information systems, combining the current status and needs of the actual research on the manual system, the analysis, design and archiving are carried out in accordance with the norms of software engineering and the principle of regulatory compliance;

The complex system can adopt the six-layer architecture design based on boundary service, WEB service, portal service, application service, database and storage service, and set up four security fire zones, which provides a safe, reliable and sturdy software and hardware system operation support environment for the application system.

This system is a simple system, adopting the general three-tier B/S architecture as the software and hardware support environment of the application system.

B/S architecture, as shown in Figure 1, is mainly the use of Web technology, combined with the browser’s ability to interpret scripts, with a common browser to achieve the original need for complex special software to achieve powerful functions, and save the development cost and maintenance costs.

Figure 1.

System structure

In the B/S architecture, in addition to the database server, the application program is stored on the Web server in the form of static or dynamic web pages, and the user only needs to enter the corresponding URL in the browser on the client side to run an application program. The user runs an application by entering the URL into the browser on the client. The application on the Web server is called and operates on the database to complete the corresponding data processing work, and finally the result is displayed to the user through the browser. It can be said that in the application system of B/S architecture, the application program is centralized to a certain extent.

In a software system based on the B/S architecture, installation, modification, and maintenance are all handled on the server side. When users use the system, they only need a browser to run all the modules, which really achieves the function of “zero client”, and it is easy to upgrade automatically at runtime.B/S architecture also provides the most realistic and open foundation for the on-line, networking, and unified service of heterogeneous machines, heterogeneous networks, and heterogeneous application services.

Before the emergence of the B/S architecture, the functions of the management information system were mainly internal to the organization, and the B/S-style “zero-client” approach made it easy for the computers of the organization’s suppliers and customers (who might be potential, that is to say, might be unknown in advance) to become clients of the management information system, and then to query organization-related information within a limited range of functions, complete data exchange and processing of various business transactions with the organization, and expand the functional coverage of the organization’s computer application system. In the limited functional scope of the query organization-related information, complete with the organization of the various business transactions of the data exchange and processing work, expanding the organization of the computer application system’s functional coverage, you can make fuller use of a variety of resources on the network, with the application of the workload is also greatly reduced. In addition, the combination of the B/S architecture application system and the Internet also makes it possible to realize some new enterprise computer applications (e.g. e-commerce, cloud computing).

As shown in Figure 1, the system is divided into three layers: presentation layer, business logic layer, and data layer. The performance layer uses the business logic layer, and the business logic layer uses the data layer. The performance layer mainly includes the UI display class to display the interface. The business logic layer includes the entity class and service class, while the data layer includes the mapping class and data control class. The business layer depends on the WEB GUI to show that the data layer relies on ADO.NET technology. The three-tier structure is built on the basis of the .net framework.

The functional architecture design defines the relationship between the main parts and components of the software system. As shown in Figure 2, the system consists of six modules and one sub-system, which are: student variable information management module, difficult student financial aid management module, dormitory management module, leave management module, queue management module and discipline management module. This subsystem is the background management subsystem, which is divided into four modules: user management, log management, system role management, and user level management.

Figure 2.

Total system packet diagram

The basic information management business process mainly includes the registration of new trainees, the division of classes or teams, and the establishment of trainee registration, etc. This business process is based on the analysis of the basic needs of trainee information management and the process formulated with the management personnel of the Academic Affairs Section of the university, and the business process has added a lot of new processing operations, which are described as follows. 1)

Registration of new students

When a new trainee enrolls in a school, he/she needs to register his/her school registration, including the information of his/her class or formation, and his/her school registration number, etc. The school registration number of the new trainee is automatically registered. The registration number of the new trainee is assigned to each new trainee by automatic batch processing, and then the new trainee is automatically assigned to the class or formation according to the order of the registration number, and manual adjustment of the class or formation to which the new trainee belongs is supported.

Figure 3.

Student information management service flow chart

The basic information management subsystem is a series of business processes that must be completed after the enrollment of new students, and it is a series of business processes that must be carried out for each new student, while the school registration management subsystem does not have exactly the same business processes for all the students, and this subsystem mainly deals with various school registration businesses of the students, such as registration, rewards and punishments, suspension and withdrawal of students, graduation information and employment information, and so on. Data. The main business flow diagram of the student registration management subsystem is shown in Figure 4 and described as follows.

Figure 4.

Student status management system business flow

Multi-task learning (MTL) is a learning strategy that forces algorithms to learn multiple tasks simultaneously. In particular, the learning performance of the main task can be improved by learning the auxiliary tasks and will also reduce label annotations by learning the shared knowledge between different tasks.MTL networks are usually categorized as either hard-parameter sharing networks or soft-parameter sharing networks. In hard parameter sharing (also known as multi-head), the network outputs separately by using separate task layers on top of the shared encoder.

Whereas in soft parameter sharing all parameters are specific to different tasks, but all networks have the ability to handle cross-task learning. The hard parameter sharing approach was chosen to keep the model simple and minimize the risk of overfitting. The structure of the MTL based on hard parameter sharing for neural networks is shown in Fig. 5 and consists of three layers: the input layer, the shared layer and the task-related layer. The shared layer is where we learn the shared representations of all tasks, while the task-related layer is where we learn the specific representations of each task. The key idea of multi-task learning is to find the relationship between tasks, which can provide useful information for the primary and secondary tasks to achieve better learning results.

Figure 5.

MTL structure

It can be seen that the hidden layers of the network are shared by multiple tasks, while the task-related layers perform different task learning. Sharing several hidden layers at the bottom of the network, different tasks learn common low-level features from labeled samples related to different tasks, which can effectively utilize the labeled sample information and also reduce the risk of data overfitting. It should be emphasized that in this learning paradigm, auxiliary tasks can help improve the learning performance of the main task, i.e: (1)LearningAlg(Main Task||Related tasks)>LearningAlg(Main Task)

However, it is also noted that different learning algorithms may achieve different learning results for the same task. Typically, most machine learning methods employ single-task learning, which focuses on accomplishing a task with a specific model. It is possible to deal with complex problems by decomposing them into different independent tasks solved by a specific model, but ignoring the intrinsic correlation between the various tasks. In contrast, MTL can utilize a single model to solve complex problems by learning tasks in parallel. In general, the training cost of multiple models is significantly higher than that of a single model. Therefore, MTL models are more acceptable in practice than single-task models, unless the results of MTL are significantly worse. Moreover, MTL usually shows better performance when different tasks can be considered as related or similar. The main reason is that MTL can obtain correlation information between different tasks. By exploiting the useful information contained in multiple related tasks, the generalization performance of all tasks can be improved.

In the initial stage when multi-task learning was proposed, the most important motivation was to cope with the problem of data sparsity, because each individual task has limited labeled data by itself, and the amount of respective labeled data is not enough to support the monologue to train accurately, on the contrary, MTL learns the knowledge of all the tasks at the same time with the purpose of enhancing the labeling of the data, which can lead to a more accurate answer for each task. So multi-task learning can actually make better use of the existing knowledge and thus reduce the required labeling cost. When the era of big data arrived in areas such as computer vision and natural language processing, it was discovered that deep MTL models could achieve better performance than their single-task counterparts and reduce the risk of overfitting in each task.

MTL has attracted a lot of attention in the field of artificial intelligence and machine learning in the last few decades. Many MTL models have been designed and many MTL applications in other fields have been developed by virtue of their flexibility and efficiency, but there are few applications in the chemical industry, and this chapter applies the strategy of multi-task learning to the classification of faults in chemical processes.

The multi-task learning algorithm is enabled to learn both the primary task classification and the secondary task simultaneously. The two outputs can be considered as y^fc and y^de respectively in the final task layer, denoted as (2)yfc=1∑m=1Meznjc[ ez1jc ez2jc ... ezNjc] (3)yde=[ z1de z2de ... zDde]

where M is the number of category labels for the primary task; D is the dimension of the input data; zi=wiTh+bi is a linear sum of weighting vectors and deviations; and {zmk}m=1M and {zdkc}d=1D represent the final linear projections of the primary and secondary tasks, respectively.

The designed objective function (L_MLCD) is specified as (4)LMLCD=β1LFC(y^k;ωsh,ωfc)+β2LDE(y^de;ωsh,ωde)

where L_FC and L_DE are the loss functions for the main and auxiliary tasks; ω_sh, ω_fc and ω_de represent the training parameters for the shared, main and auxiliary task layers, respectively; and β₁, β₂ is the weighted hyperparameters for each task, respectively. The parameters are updated as follows: (5)ωshi+1=ωshi−λ( β∂LFC∂ωshi+β2 ∂LDE∂ωshi) $$\omega _{sh}^{i + 1} = \omega _{sh}^i - \lambda (\>\beta {{\partial {L_{FC}}} \over {\partial \omega _{sh}^i}} + {\beta _2}\>{{\partial {L_{DE}}} \over {\partial \omega _{sh}^i}})$$ (6)ωfct+1=ωfct−λ∂LFC∂ωfct (7)ωfdt+1=ωfdt−λ∂LFD∂ωfdt

where t is the number of iterations during training; λ is the learning rate; and β and λ are key hyperparameters, where λ adjusts the extent of training while β manages the relative importance between tasks. In general, hyperparameters are specified empirically; however, this rarely helps to achieve desired algorithmic performance. For example, λ too large can lead to non-convergence of the overall training, while β too small can neglect learning of the main tasks. Here, a Bayesian optimization approach is used to select key hyperparameters λ by using an agent model and Bayesian updating to achieve a solution. Its general step is to search for every possible combination and randomly select the top few. Then, based on the performance of these hyperparameters, the next best possible value is determined. Thus, the selection of each hyperparameter depends on previous attempts. The next set of hyperparameters is selected based on a historical evaluation of performance until the best combination is found or the maximum number of tests is reached. It provides superior results compared to traditional methods such as random and grid search.

Cognitive diagnosis, as one of the important research contents of educational data mining, is mainly to explore the potential cognitive state of students based on their answer records and related learning behaviors, such as the degree of knowledge mastery, etc. Accurate results of student assessment are the basis of many education-related researches and applications, such as test question recommendation, early warning of dropping out, etc. Therefore, research on cognitive diagnosis of students is attracting attention both from researchers and the general public. Therefore, research related to student cognitive diagnosis has attracted the attention of researchers and the public.

In recent years, researchers have proposed several cognitive diagnostic models for student assessment. Figure 6 shows an example of using a cognitive diagnostic model for student assessment. As can be seen in Figure 6, most of the existing static cognitive diagnostic models for student assessment target the information from one exam. Usually, the assessment is performed based on the student’s answer record to obtain the student’s cognitive state and the corresponding test parameters (test difficulty, differentiation, etc.). Some experts in the field of pedagogy have pointed out that there is no direct comparison between the scores obtained by students taking different exams. For example, if Student A and Student B take the same test, and Student A scores higher than Student B, Student A is considered to be more capable than Student B. However, if students A and B do not take the same test, even if student A scores higher than student B, student A cannot be considered to be more capable than student B, because there may be differences in the difficulty of the questions themselves, resulting in no comparability between the scores of the two students. Similar to the problem that students’ scores are not directly comparable between different exams, the existing cognitive diagnostic model, when applied to real-world scenarios, also suffers from the problem that there is no comparability between the student assessment results obtained by diagnosing the cognitive diagnostic model when applied to two independent exams if the two students have taken different exams. In most practical teaching scenarios, students in different classes and schools usually use different test questions in their daily practice and examinations, but when taking a unified selection test (e.g., secondary school examination, college entrance examination, etc.), candidates need to be selected through their ranking order of performance according to the selection criteria for candidates in a specific ranking interval or score band. Therefore, it is necessary to diagnose an accurate and comparable cognitive state of students based on independent exercises and examinations, so as to have a more comprehensive and objective understanding of them.

Figure 6.

Example of student assessment

Assume that there is an exam set E = {E₁, E₂, ⋯, E_T}, where each exam E_t(t = 1, 2, ⋯, T) contains a student set Ut={Ut1,Ut2,⋯,UtU} and a test question set Vt={Vt1,Vt2,⋯,VtV} , and consider the student diagnosis for each exam E_t as a separate task t(t = 1, 2, ⋯, T). It is worth noting that the students and test questions contained in each separate task are non-intersecting with each other. For each task t, there is a matrix of student answer records Y_t, where Y_tuv is the student U_tu answers on the test questions V_tv. In common scoring methods and cognitive diagnostic models, M Y_tuv is usually scored as 1 when a student U_tu correctly answers question V_tv, and M Y_tuv is scored as 0 when a student U_tu incorrectly answers question V_tv. Thus, the student answer record matrix is usually a binary matrix of 0s and 1s. In addition to the student answer record matrix, the question face information of the corresponding test question is needed as supplementary information connecting each individual student assessment task. For each task t, a test question feature matrix F_t is generated based on the question face information of the test questions, where the question face feature of each test question V_tv is represented by a row vector F_tv, and the feature vectors of all the test questions in task t form a test question feature matrix: F_t = (F_t1, F_t2, ⋯, F_tV).

Based on the above description of the research problem and data, the multi-task student assessment problem can be formally defined as follows:

Given the set of exams E = {E₁, E₂, ⋯, E_T}, as well as the set of students U_t corresponding to each exam E_t, the set of test questions V_t, the matrix of student answer records Y_t, and the corresponding matrix of test question facet features F_t, the main objectives of multi-task student assessment are to (1) conduct student assessment of T independent exams simultaneously, so as to obtain accurate and comparable student cognitive states and corresponding test question features (e.g., difficulty, differentiation, etc.); and (2) Prediction of student responses based on the student cognitive states and test question parameters obtained from the assessments.

The first experiment is to verify the effectiveness of the MTL algorithm for the optimization objective of minimizing distributional differences in domain differences. The distribution of knowledge states in the source and target domains before and after hard parameter sharing are compared. Figure 7 shows the visualization results of the MTL algorithm for dimensionality reduction, before migration on the left and after migration on the right. There is an interesting conclusion about the M → P task. Before minimizing the knowledge discovery between the source and target domains, the knowledge state distributions of the different domains are inconsistent due to their distance from each other. However, after sharing through hard parameters, the distributions of the two domains completely overlap. This indicates that hard parameter sharing is an effective method to realize domain adaption, and hard parameter sharing can be used effectively for domain adaption.

Figure 7.

The data distribution minimizes the visualization results

In order to fully validate the effectiveness of the proposed MTL algorithm, this paper compares several common methods. All of these methods are capable of tracking knowledge and applying it to task transfer to some extent. Specifically, these methods include:

Original: It refers to the scenario where no migration operation is performed. Its experimental results are based on the best performance of the following baselines, which are trained on the source domain and directly applied to the target domain. Replacing only the output layer solves the problem of different output dimensions in different domains without any migration process.

DKT: It is an important knowledge tracking model that applies recurrent neural networks to model the learning process of students in order to estimate their mastery level.

GKT: It is a knowledge tracking method based on graph neural networks, which uses only prerequisite relationships to construct knowledge edge structures.

DKVMN: It is able to utilize the relationships between basic concepts to directly output the learner’s proficiency level for each knowledge point.

SKVMN: It combines the strengths of recurrent neural network modeling and memory network capabilities. It is better equipped to detect the relationship between potential knowledge points and topics, and to track students’ knowledge status.

The above methods are not migratable due to the inconsistency of output dimensions between different domains. I add fine-tuning operations to make them migratable. Among them, MTL algorithm and its variants, i.e., DKT+F, GKT+F, DKVMN+F, and SKVMN+F, are the models after adding fine-tuning operations to DKT, GKT, DKVMN, and SKVMN, respectively; MTL algorithm is the MTL algorithm that includes only the fine-tuning process; MTL-Q is the MTL algorithm that includes only the process of minimizing the difference in distributions and the fine-tuning process; MTL-M is the MTL algorithm that the MTL algorithm that includes only the problem selection and fine-tuning processes, and MTL algorithm is the complete algorithm we propose.

In order to examine the performance of the methods for inter-domain knowledge and semantic migration, we selected the datasets of four schools of thought as four independent domains and named them A, B, C, and D. As a result, there are 12 migration tasks (e.g., A → B).The experimental results of AUC are shown in Table 2.

It is easy to see that among all the methods, MTL performs best in the inter-school migration task. This proves that the multitasking algorithm has superior migration capabilities. However, there is more to be explained in the table. First, comparing DKT+F, SKVMN+F, GKT+F and DKVMN+F with Original, it is found that fine-tuning using a small amount of target data helps to improve the knowledge tracking performance. Second, MTL slightly outperforms DKT+F because DKT+F and MTL-QM are structurally very similar, whereas MTL-QM draws on topic text word vectors. Third, MTL outperforms all variants, which is consistent with the intuition that the three migration stages are effective.

The student skill mastery and time weighting function T(u,j) are introduced to solve the student similarity, and the effect of different values of T(u,j) on MAE is experimentally determined. In view of the dynamic tracking of student answering behavior data in Math1, Math1 was chosen as the dataset for the experiments in this subsection to confirm the optimal value of T(u,j).

As can be seen from Figure 8, for the three sets of student information contained in Math1, the trend of the MAE curves is roughly similar. With the increase of the time weight parameter, the value of MAE firstly decreases very quickly, and then the decrease begins to slow down, the value of MAE gradually slows down when the parameter takes the value of 160 days, and then gradually tends to stabilize after taking the value of 200, and finally roughly stabilizes at about 0.65. Then this paper will use the time-weighted parameter T(u, i) instead of the above time value. The final result will be very close.

Figure 8.

Influence of different time weight parameters on MAE value

In order to verify the effectiveness of the algorithm, the algorithm probability matrix factorization (PMF), the traditional collaborative filtering algorithm (CF), and the MTL algorithm of this paper are compared. The comparison of accuracy between each algorithm is shown in Table 3.

As can be seen from Table 3, MTL, the personalized exercise recommendation algorithm based on improved cognitive diagnostic model proposed in this paper, has better results in terms of accuracy. On the FrcSub dataset, the accuracy of this paper’s MTL algorithm is 8.88% and 27.05% higher than the PMF model and CF algorithm, respectively. The main reason is that the MTL algorithm proposed in this paper is based on the cognitive diagnostic model for recommendation, and the algorithm focuses on describing the skill mastery of the students and takes it as one of the important influencing factors for similar students’ solutions, which can more accurately reflect which skill the students mastered better, which skill they mastered more weakly, or which skill they did not master at all. Moreover, the MTL algorithm takes into account the characteristics of students’ skill mastery over time in the recommendation process, and depicts the influence of students’ answer data in different periods on the recommendation effect. The PMF algorithm and the traditional CF algorithm make recommendations at the level of exercises, ignoring the degree of students’ skill mastery, which leads to lower accuracy of these two methods. And the traditional CF algorithm does not introduce the two features of students’ potential response scores on exercises and the factor of students’ skill mastery over time into the calculation formula when solving for similar students, resulting in the accuracy of the traditional CF algorithm being much lower than that of the MTL algorithm. For the Math1 and Math2 datasets, which have a larger amount of data, the accuracy of the MTL algorithm is higher than the PMF model and CF algorithm by 5.53% and 25.72%, and 7.87% and 24.27%, respectively, which once again validates the effectiveness of the MTL algorithm.