Deep learning-based school selection recommendation optimization method and system

Abstract

The invention discloses a school selection recommendation optimization method and system based on deep learning, and the method comprises the following steps: S1, collecting the personal information of a user, and carrying out the standardization processing; s2, college information is collected, data mapping is carried out, and a college feature vector set is generated; s3, based on the user feature vector set and the college feature vector set, feature matching is carried out by adopting a deep neural collaborative filtering method, and a preliminary college selection recommendation list is generated; s4, on the basis of the preliminary school selection recommendation list, according to the fitness scores of the users and the colleges, a school selection game optimization model based on an entropy regularization game is adopted to optimize strategy distribution of the users and the colleges, and an optimized school selection recommendation scheme is generated; and S5, dynamically updating parameters of the user deep neural collaborative filtering method based on the optimized school selection recommendation scheme. According to the method, deep neural collaborative filtering and entropy regularization game optimization are combined, personalized selection and correction recommendation is realized, and the method has the advantages of accurate matching, stable strategy optimization and adaptive adjustment capability.

Description

Technical Field

[0001] The present invention relates to the technical field of recommendation systems, and particularly to an optimization method and system for school selection recommendation based on deep learning. Background Art

[0002] In the context of current informatization and globalization, school selection decisions have become a crucial link in students' academic planning and career development. Traditional school selection methods mainly rely on manual consultations, college ranking lists, or simple score matching methods. These methods have certain limitations. With the rapid development of Internet technology, various school selection recommendation systems have emerged one after another, attempting to use big data and artificial intelligence technologies to provide more accurate school selection suggestions for users. However, existing school selection recommendation systems still have many deficiencies and are difficult to fully meet the personalized and multi-dimensional needs of users.

[0003] Currently, school selection recommendation systems based on rule matching are relatively common. Such systems usually perform static matching between fixed indicators, such as users' exam scores, GPAs, standardized test scores, etc., and the admission criteria of colleges and universities to generate a recommendation list. Although this method is simple to calculate and easy to understand, due to the lack of consideration of users' interest preferences, career development directions, and the dynamic changes of colleges and universities, the recommended results lack personalization and adaptability, and users often have difficulty making optimal decisions. In addition, this type of recommendation method lacks dynamic learning ability and cannot be optimized according to users' feedback, resulting in low timeliness and accuracy of the recommendation scheme and affecting the user experience.

[0004] The school selection recommendation method based on collaborative filtering is a relatively advanced recommendation strategy. It analyzes the historical school selection behaviors of a large number of users, mines the potential associations between users and colleges and universities, and generates personalized recommendations. However, traditional collaborative filtering methods are mainly divided into user-based collaborative filtering and item-based collaborative filtering. The former relies on the similarity between users, and the latter relies on the similarity between colleges and universities. Both of these methods have problems of data sparsity and cold start. In addition, most traditional collaborative filtering methods use linear models and are difficult to capture the high-order non-linear relationships of user and college features, thus affecting the recommendation effect.

[0005] In recent years, with the development of deep learning technology, recommendation systems based on deep neural networks have gradually emerged. The deep neural collaborative filtering method automatically learns the feature representations of users and institutions through neural networks, and extracts high-dimensional features through non-linear transformations, improving the accuracy and generalization ability of recommendations. However, existing deep learning recommendation methods mainly focus on fitness calculation and fail to effectively combine factors such as the admission strategies of institutions and the actual application decisions of users for optimization, resulting in poor interpretability and matching stability of recommendation results. In addition, existing methods usually lack optimization of the adaptability of the recommendation system and cannot dynamically adjust recommendation strategies as user needs change, affecting the effectiveness of long-term recommendations.

[0006] Another relatively advanced recommendation optimization method is the game optimization method. The school selection decision is essentially a two-way matching problem. Users hope to apply to the best institutions, while institutions hope to recruit the best students. Therefore, introducing a game optimization model can better describe the interaction relationship between users and institutions and optimize the school selection matching strategy. However, traditional game optimization methods usually rely on static payoff functions and fail to fully consider the dynamic adjustment of user and institution strategies, resulting in inflexible recommendation results. In addition, it is difficult for traditional game optimization methods to be effectively combined with deep learning methods, leading to high computational complexity and difficulty in large-scale application in actual school selection recommendation scenarios. Summary of the Invention

[0007] An object of the present invention is to propose a school selection recommendation optimization method based on deep learning. The present invention combines deep neural collaborative filtering and entropy-regularized game optimization, calculates the fitness score between users and institutions through the deep neural collaborative filtering method, and optimizes it using the mean squared error loss function and the contrastive loss function to improve the generalization ability of the recommendation system. Based on the preliminary school selection recommendation list, a school selection game optimization model is constructed, the strategy space and payoff function are defined, and the entropy-regularized game method is used to optimize the strategy distributions of users and institutions, with the ability of adaptive optimization.

[0008] A school selection recommendation optimization method and system based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect personal information of users, construct a user feature data set, and perform standardization processing to generate a set of user feature vectors;

[0010] S2. Collect institution information, construct an institution database, and perform data mapping to generate a set of institution feature vectors;

[0011] S3. Based on the user feature vector set and the institution feature vector set, use the deep neural collaborative filtering method for feature matching, and use the mean squared error loss function and the contrastive loss function to optimize the deep neural collaborative filtering method, calculate the fitness score between the user and the institution, and generate a preliminary school selection recommendation list;

[0012] S4. Based on the preliminary school selection recommendation list, construct a school selection game optimization model, define the strategy space and the payoff function, according to the fitness score between the user and the institution, use the school selection game optimization model based on entropy regularization game to optimize the strategy distributions of the user and the institution, and construct a school selection matching matrix to generate an optimized school selection recommendation plan;

[0013] S5. Based on the optimized school selection recommendation plan, dynamically update the parameters of the user's deep neural collaborative filtering method to optimize the adaptability of the deep neural collaborative filtering method.

[0014] Optionally, the personal information includes personal academic information, interest preferences, career goals, and social and economic conditions.

[0015] Optionally, the institution information includes admission requirements, admission standards, professional rankings, academic resources, employment rates, and relevant historical admission data.

[0016] Optionally, the specific content of S3 includes:

[0017] S31. Construct a user-institution interaction matrix based on the user feature vector and the institution feature vector:

[0018] R = {r ij |U′ j ∈U′, C′ j ∈C′};

[0019] Among them, R represents the user-institution interaction matrix, r ij represents the fitness score between user U i and institution C j , ∈ means "belongs to", U′ j and C′ j represent the user feature vector and the institution feature vector respectively, and U′ and C′ represent the user feature vector set and the institution feature vector set respectively;

[0020] S32. Use the deep neural collaborative filtering method for feature matching, perform embedding mapping on the user feature vector and the institution feature vector, and generate embedding vectors E U and E C :

[0021] E U = φ(W U U′ + b U ), E C= φ(W C C′ + b C );

[0022] Among them, E U represents the user embedding vector, E C represents the institution embedding vector, φ represents the non - linear activation function, W U and W C respectively represent the weight matrices of user features and institution features, b U and b C respectively represent the corresponding bias vectors, U′ and C′ respectively represent the set of user feature vectors and the set of institution feature vectors;

[0023] S33. Calculate the matching score between the user embedding vector and the institution embedding vector, using the fitness calculation method based on bilinear interaction:

[0024]

[0025] Among them, s ij represents the initial fitness score between user U i and institution C j , W s represents the interaction weight matrix, E U represents the user embedding vector, E C represents the institution embedding vector, T represents the transpose operation;

[0026] S34. Use a multi - layer perceptron to perform non - linear transformation on the initial fitness score s ij to calculate the final fitness score:

[0027] h1 = σ(W1s ij + b1);

[0028] h2 = σ(W2h1 + b2);

[0029]

[0030] Among them, s ij represents the initial fitness score between user U i and institution C j , σ represents the non - linear activation function, W1, W2 and W3 represent the weight matrices of the multi - layer perceptron, b1, b2 and b3 represent the bias vectors of the multi - layer perceptron, represents the final calculated user - institution fitness score;

[0031] S35. Construct a training data set, and the training data set is composed of the user's past school selection decisions, application records and admission results:

[0032] D = {(U′i , C′ j , r ij ) | i ∈ N, j ∈ M};

[0033] Among them, D represents the training data set, N represents the user set, M represents the institution set, E C represents the institution embedding vector, U′ j and C′ j respectively represent the user feature vector and the institution feature vector, r ij represents the actual feedback score of user U i applying for institution C j :

[0034]

[0035] Among them, r ij represents the actual feedback score of user U i applying for institution C j , λ represents the smoothing factor, used to distinguish the score decay degree of unadmitted users, Rank j represents the comprehensive ranking of institution C j , Threshold represents an adjusted admission threshold;

[0036] S36. Adopt the mean square error loss function and the contrast loss function to jointly optimize the deep neural collaborative filtering method, and construct the comprehensive loss function:

[0037]

[0038] Among them, L represents the comprehensive loss function, |D| represents the total number of training samples, D represents the training data set, r ij represents the actual feedback score of user U i applying for institution C j , represents the final calculated user-institution fitness score, α represents the weight of the contrast loss, δ represents the score difference threshold, max represents taking the maximum value between the two, represents the predicted score of the user for the mis-matched institution;

[0039] S37. Based on the trained deep neural collaborative filtering method, calculate the fitness scores of all users and all institutions, and sort them according to the scores to generate a preliminary school selection recommendation list:

[0040]

[0041] Among them, R′ represents the preliminary school selection recommendation list, sort represents the sorting function, and sorts in descending order according to the fitness score.

[0042] Optionally, S4 specifically includes:

[0043] S41. Based on the preliminary school selection recommendation list R′, construct a school selection game optimization model, define the user set N and the school set M, define the school selection strategy of user U i as A i , and the admission strategy of school C j as B j , and construct the strategy space:

[0044] A i ={a i1 ,a i2 ,...,a ik}, B j ={b j1 ,b j2 ,...,b jl};

[0045] Among them, A i represents the school selection strategy of user U i , describes the set of schools to be selected by user U i , B j represents the admission strategy of school C j , describes the set of students to be selected by school C j , a ik represents the probability that user U i selects school C k , b jl represents the probability that school C j admits user U i ;

[0046] S42. Use the preliminary school selection recommendation list R′ as the preliminary recommendation strategy of the user, define the preliminary selection probability of the user, and define the initial admission strategy probability of the school based on historical admission data:

[0047]

[0048] Among them, P0(A) represents the preliminary selection probability of the user, represents the fitness score of user U i for school C j , e represents the natural exponential function, represents the fitness score of user U i for school C k , K represents the total number of schools, Q0(B) represents the initial admission strategy probability of the school, ψ ij and ψ ik represent the school historical admission data;

[0049] S43. Define user U iThe revenue function U i (A, B) and institution C j The revenue function U j (A, B), and combine the entropy regularization term to constrain the stability of strategy selection:

[0050]

[0051] Among them, U i (A, B) represents the revenue function of user U i The revenue function U j (A, B) represents the revenue function of institution C j The revenue function, α1, α2, and α3 represent adjustment parameters, Represents user U i And institution C j The fitness score between them, E j Represents institution C j The graduate employment rate of, C i Represents user U i The application cost of, μ C Represents user U i The average value of the application cost of, λ1, λ2, and λ3 represent adjustment parameters, Ql i Represents the academic quality of the user, D i Represents institution C j The diversity of the student source of, μ D Represents institution C j The average value of the diversity of the student source of, R i Represents the research ability of the user, β1 and γ1 represent adjustment parameters, τ represents the entropy regularization coefficient, P(A) and Q(B) respectively represent the strategy probability distributions of the user and the institution, H(P(A)) and H(Q(B)) respectively represent the entropy regularization terms of the user and the institution:

[0052] H(P(A)) = -∑P(A)log2P(A), H(Q(B)) = -∑Q(B)log2Q(B);

[0053] Among them, H(P(A)) represents the entropy regularization term of the user, H(Q(B)) represents the entropy regularization term of the institution, P(A) and Q(B) respectively represent the strategy probability distributions of the user and the institution;

[0054] S44. Construct the entropy regularization game solution process, define the user as the follower of the game, the institution as the leader of the game, and construct the optimization objective:

[0055]

[0056] Among them, P * (A) represents the optimal school selection strategy probability distribution obtained by the user after optimization; Q* (B) represents the probability distribution of the optimal admission strategy obtained by the institution after optimization, P(A) represents the current strategy probability distribution of the user, P0(A) represents the preliminary selection probability of the user, obtained from the preliminary school selection recommendation list R′, η represents the learning rate, controlling the strategy update step size, Q(B) represents the current strategy probability distribution of the institution, Q0(B) represents the initial admission strategy probability of the institution, argmin represents the variable value when taking the maximum value, D KL represents the Kullback-Leibler divergence, U i (A,B) represents the benefit function of user U i ; U j (A,B) represents the benefit function of institution C j ; N represents the set of users, and M represents the set of institutions;

[0057] S45. Calculate the optimization parameters of the school selection strategy based on the entropy regularization gradient optimization method, and update the strategy distributions of users and institutions using the entropy regularization gradient descent method:

[0058]

[0059] Among them, P t+1 (A) and Q t+1 (B) represent the user and institution strategy distributions in the (t + 1)-th round respectively, P t (A) and Q t (B) represent the user and institution strategy distributions in the t-th round respectively, η represents the learning rate, controlling the strategy update step size, represents the gradient of the user benefit function, represents the gradient of the institution benefit function, τ represents the entropy regularization coefficient, and represent the gradients of the entropy regularization terms, ensuring that the strategy does not change violently;

[0060] S46. Based on the optimized user strategy P * (A) and institution strategy Q * (B), calculate the matching rules, construct the school selection matching matrix, and generate the optimized school selection recommendation plan:

[0061]

[0062] Among them, M ij represents the school selection matching matrix, and θ1 and θ2 represent the thresholds for the user to select an institution and the institution to admit a user respectively.

[0063] A school selection recommendation optimization system based on deep learning according to an embodiment of the present invention includes:

[0064] A user data processing module for collecting user information and generating user feature vectors;

[0065] A college data processing module for collecting college information and generating college feature vectors;

[0066] A deep neural collaborative filtering recommendation module for calculating the matching degree score between users and colleges and generating a preliminary college selection recommendation list;

[0067] A college selection game optimization module for optimizing the strategy distributions of users and colleges and generating an optimized college selection recommendation plan;

[0068] An adaptive optimization module for updating the parameters of the deep neural collaborative filtering method based on the optimized college selection recommendation plan to improve adaptability.

[0069] The beneficial effects of the present invention are as follows:

[0070] First, compared with the static recommendation method based on rule matching, the present invention can deeply mine multi-dimensional features of users such as academic background, interest preferences, and career goals, and combine data such as the enrollment criteria and historical admission situations of colleges, and automatically learn the matching relationship between users and colleges by using the deep neural collaborative filtering method, improving the accuracy and personalization degree of recommendations.

[0071] Second, the present invention uses the deep neural collaborative filtering method to model the matching relationship between users and colleges, utilizes the non-linear transformation ability of the neural network to learn the high-dimensional features of users and colleges, and mines deep matching patterns. By constructing a user-college interaction matrix, performing embedding mapping on the user feature vector and the college feature vector, adopting a matching degree calculation method based on bilinear interaction, and using a multi-layer perceptron to perform non-linear transformation on the initial matching degree score, the accuracy of the matching degree calculation is improved. In addition, the present invention uses the mean square error loss function and the contrastive loss function for optimization, enabling the model to better distinguish correctly matched and wrongly matched colleges, improving the accuracy and generalization ability of the recommendation results, thereby generating a high-quality preliminary college selection recommendation list to provide input data for subsequent game optimization.

[0072] Secondly, the present invention uses the entropy-regularized game optimization method to model the user's school selection strategy and the school's admission strategy, and introduces an entropy regularization term in the optimization process to make the strategy optimization process smoother, avoiding the problem that the traditional game optimization method converges to a local optimum during the search process. The school selection game optimization model constructed based on the preliminary school selection recommendation list enables the user's school selection decision to not only consider the fitness score but also incorporate the school's enrollment strategy and the user's actual application behavior, thus making the recommendation result more in line with the user's actual needs. Through strategy optimization, the present invention can establish an equilibrium relationship among the user's needs, the school's admission probability, and the stability of the recommendation system, such that the recommendation scheme not only improves the user's application success rate but also optimizes the quality and diversity of the school's student sources, achieving an optimized effect of two-way matching.

[0073] Finally, the present invention has an adaptive optimization ability and can dynamically adjust the parameters of the deep neural collaborative filtering model based on the optimized school selection recommendation scheme, enabling the recommendation system to automatically optimize the recommendation result when the user's needs change or the school's enrollment policy is adjusted. Compared with the traditional collaborative filtering method that only relies on static historical data, the present invention continuously learns the user's feedback data, enabling the recommendation system to continuously update the user profile and improve the effectiveness and sustainability of long-term recommendations. This adaptive optimization mechanism not only improves the accuracy of the recommendation result but also enhances the scalability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0075] Figure 1 is a flowchart of a school selection recommendation optimization method based on deep learning proposed by the present invention;

[0076] Figure 2 is a schematic structural diagram of a school selection recommendation optimization method based on deep learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0077] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, and thus only showing the components related to the present invention.

[0078] Refer to Figure 1 and Figure 2 , a school selection recommendation optimization method and system based on deep learning, including the following steps:

[0079] S1. Collect the user's personal information, construct a user feature data set, and perform standardization processing to generate a set of user feature vectors;

[0080] S2. Collect college information, construct a college database, and perform data mapping to generate a set of college feature vectors;

[0081] S3. Based on the set of user feature vectors and the set of college feature vectors, use the deep neural collaborative filtering method for feature matching, and use the mean squared error loss function and the contrastive loss function to optimize the deep neural collaborative filtering method, calculate the fitness score between the user and the college, and generate a preliminary college selection recommendation list;

[0082] S4. Based on the preliminary college selection recommendation list, construct a college selection game optimization model, define the strategy space and the revenue function, according to the fitness score between the user and the college, use the college selection game optimization model based on entropy regularization game to optimize the strategy distributions of the user and the college, and construct a college selection matching matrix to generate an optimized college selection recommendation plan;

[0083] S5. Based on the optimized college selection recommendation plan, dynamically update the parameters of the user's deep neural collaborative filtering method to optimize the adaptability of the deep neural collaborative filtering method.

[0084] In this embodiment, the personal information includes personal academic information, interest preferences, career goals, and social and economic conditions.

[0085] In this embodiment, the college information includes enrollment requirements, admission standards, professional rankings, academic resources, employment rates, and relevant historical admission data.

[0086] In this embodiment, the specific content of S3 includes:

[0087] S31. Construct a user-college interaction matrix based on the user feature vector and the college feature vector:

[0088] R = {r ij |U′ j ∈U′, C′ j ∈C′};

[0089] Among them, R represents the user-college interaction matrix, r ij represents the fitness score between the user U i and the college C j ∈ means "belongs to", U′ j and C′ j represent the user feature vector and the college feature vector respectively, and U′ and C′ represent the set of user feature vectors and the set of college feature vectors respectively;

[0090] S32. Use the deep neural collaborative filtering method for feature matching, perform embedding mapping on the user feature vector and the college feature vector to generate the embedding vectors E U and EC :

[0091] E U = φ(W U U′ + b U ), E C = φ(W C C′ + b C );

[0092] Wherein, E U represents the user embedding vector, E C represents the institution embedding vector, φ represents the non-linear activation function, W U and W C respectively represent the weight matrices of user features and institution features, b U and b C respectively represent the corresponding bias vectors, U′ and C′ respectively represent the set of user feature vectors and the set of institution feature vectors;

[0093] S33. Calculate the matching score between the user embedding vector and the institution embedding vector, and adopt an adaptation degree calculation method based on bilinear interaction:

[0094]

[0095] Wherein, s ij represents the initial adaptation degree score between user U i and institution C j , W s represents the interaction weight matrix, E U represents the user embedding vector, E C represents the institution embedding vector, and T represents the transpose operation;

[0096] S34. Use a multi-layer perceptron to perform non-linear transformation on the initial adaptation degree score s ij to calculate the final adaptation degree score:

[0097] h1 = σ(W1s ij + b1);

[0098] h2 = σ(W2h1 + b2);

[0099]

[0100] Wherein, s ij represents the initial adaptation degree score between user U i and institution C j , σ represents the non-linear activation function, W1, W2 and W3 represent the weight matrices of the multi-layer perceptron, b1, b2 and b3 represent the bias vectors of the multi-layer perceptron, represents the final calculated adaptation degree score between the user and the institution;

[0101] S35. Construct a training dataset, which is composed of the user's past school selection decisions, application records, and admission results:

[0102] D = {(U′ i , C′ j , r ij )|i ∈ N, j ∈ M};

[0103] Among them, D represents the training dataset, N represents the user set, M represents the institution set, E C represents the institution embedding vector, U′ j and C′ j represent the user feature vector and the institution feature vector respectively, r ij represents the actual feedback score of user U i applying to institution C j :

[0104]

[0105] Among them, r ij represents the actual feedback score of user U i applying to institution C j , λ represents the smoothing factor, used to distinguish the degree of score attenuation of unadmitted users, Rank j represents the comprehensive ranking of institution C j , Threshold represents an adjusted admission threshold;

[0106] S36. Adopt the mean squared error loss function and the contrast loss function to jointly optimize the deep neural collaborative filtering method, and construct a comprehensive loss function:

[0107]

[0108] Among them, L represents the comprehensive loss function, |D| represents the total number of training samples, D represents the training dataset, r ij represents the actual feedback score of user U i applying to institution C j , represents the finally calculated user-institution fitness score, α represents the weight of the contrast loss, δ represents the score difference threshold, max represents taking the maximum value between the two, represents the predicted score of the user for the mis-matched institution;

[0109] S37. Based on the trained deep neural collaborative filtering method, calculate the fitness scores of all users and all institutions, and sort them by score to generate a preliminary school selection recommendation list:

[0110]

[0111] Among them, R′ represents the preliminary school selection recommendation list, and sort represents the sorting function, which sorts in descending order according to the fitness score.

[0112] In this embodiment, the S4 specifically includes:

[0113] S41. Based on the preliminary school selection recommendation list R′, construct a school selection game optimization model, define the user set N and the school set M, define the school selection strategy of user U i as A i , and the admission strategy of school C j as B j , and construct the strategy space:

[0114] A i ={a i1 , a i2 ,..., a ik}, B j ={b j1 , b j2 ,..., b jl};

[0115] Among them, A i represents the school selection strategy of user U i , describes the set of schools to be selected by user U i , B j represents the admission strategy of school C j , describes the set of students to be selected by school C j , a ik represents the probability that user U i selects school C k , b jl represents the probability that school C j admits user U i ;

[0116] S42. Use the preliminary school selection recommendation list R′ as the preliminary recommendation strategy of the user, define the preliminary selection probability of the user, and define the initial admission strategy probability of the school based on historical admission data:

[0117]

[0118] Among them, P0(A) represents the preliminary selection probability of the user, represents the fitness score of user U i for school C j , e represents the natural exponential function, represents the fitness score of user U i for school C k ​​​​​​​​​​​​​​​​The fitness score, K represents the total number of institutions, Q0(B) represents the initial admission strategy probability of the institution, ψ ij and ψ ik represent the historical admission data of the institution;

[0119] S43. Define the revenue function U i of user U i (A, B0 and the revenue function U j of institution C j (A, B0, and combine the entropy regularization term to constrain the stability of strategy selection:

[0120]

[0121] where U i (A, B0 represents the revenue function of user U i , U j (A, B) represents the revenue function of institution C j , α1, α2, and α3 represent adjustment parameters, represents the fitness score between user U i and institution C j , E j represents the employment rate of graduates of institution C j , C i represents the application cost of user U i , μ C represents the mean of the application cost of user U i , λ1, λ2, and λ3 represent adjustment parameters, Ql i represents the academic quality of the user, D i represents the diversity of the student source of institution C j , μ D represents the mean of the diversity of the student source of institution C j , R i represents the scientific research ability of the user, β1 and γ1 represent adjustment parameters, τ represents the entropy regularization coefficient, P(A0 and Q(B0 respectively represent the strategy probability distributions of the user and the institution, and H(P(A)) and H(Q(B)) respectively represent the entropy regularization terms of the user and the institution:

[0122] H(P(A)) = -∑P(A)log2P(A), H(Q(B)) = -∑Q(B)log2Q(B);

[0123] where H(P(A)) represents the entropy regularization term of the user, H(Q(B)) represents the entropy regularization term of the institution, and P(A) and Q(B) respectively represent the strategy probability distributions of the user and the institution;

[0124] S44. Construct the entropy-regularized game solution process, define users as the followers of the game and institutions as the leaders of the game, and construct the optimization objective:

[0125]

[0126] Among them, P * (A) represents the probability distribution of the optimal school selection strategy obtained by users after optimization; Q * (B) represents the probability distribution of the optimal admission strategy obtained by institutions after optimization, P(A) represents the current strategy probability distribution of users, P0(A) represents the preliminary selection probability of users, which is obtained from the preliminary school selection recommendation list R′, η represents the learning rate, controlling the strategy update step size, Q(B) represents the current strategy probability distribution of institutions, Q0(B) represents the initial admission strategy probability of institutions, argmin represents the variable value when taking the maximum value, D KL represents the Kullback-Leibler divergence, U i (A, B) represents the revenue function of user U i , U j (A, B) represents the revenue function of institution C j , N represents the set of users, and M represents the set of institutions;

[0127] S45. Calculate the optimization parameters of the school selection strategy based on the entropy-regularized gradient optimization method, and update the strategy distributions of users and institutions using the entropy-regularized gradient descent method:

[0128]

[0129] Among them, P t+1 (A) and Q t+1 (B) represent the strategy distributions of users and institutions in the (t + 1)-th round respectively, P t (A) and Q t (B) represent the strategy distributions of users and institutions in the t-th round respectively, η represents the learning rate, controlling the strategy update step size, represents the gradient of the user revenue function, represents the gradient of the institution revenue function, τ represents the entropy regularization coefficient, and represent the gradients of the entropy regularization terms, ensuring that the strategy does not change violently;

[0130] S46. Based on the optimized user strategy P * (A) and institution strategy Q * (B), calculate the matching rules, construct the school selection matching matrix, and generate the optimized school selection recommendation plan:

[0131]

[0132] Among them, M ij represents the school selection matching matrix, and θ1 and θ2 respectively represent the thresholds for users to select schools and for schools to admit users.

[0133] An optimized school selection recommendation system based on deep learning, comprising:

[0134] A user data processing module, configured to collect user information and generate user feature vectors;

[0135] A school data processing module, configured to collect school information and generate school feature vectors;

[0136] A deep neural collaborative filtering recommendation module, configured to calculate the fitness scores between users and schools and generate a preliminary school selection recommendation list;

[0137] A school selection game optimization module, configured to optimize the strategy distributions of users and schools and generate an optimized school selection recommendation plan;

[0138] An adaptive optimization module, configured to update the parameters of the deep neural collaborative filtering method based on the optimized school selection recommendation plan to improve adaptability.

[0139] Embodiment 1:

[0140] To verify the feasibility of the present invention in implementation, the present invention is applied to a certain study abroad application platform. The platform provides school selection suggestions for users applying to universities and assists in completing the school application. The main problems currently faced by the platform include: the recommendation system is based on static rule matching and is difficult to adapt to the personalized needs of users; the collaborative filtering method is limited by data sparsity, resulting in insufficient recommendation accuracy; the matching between users and schools lacks an optimization mechanism, resulting in poor stability of the recommendation results. To address these problems, the present invention realizes more accurate and intelligent school selection recommendation optimization through a method combining deep neural collaborative filtering and entropy regularization game optimization.

[0141] In practical applications, the system first collects the academic information of users, including standardized test scores, research experiences, internship backgrounds, intended majors, interest preferences, and career goals. At the same time, it obtains information such as the admission requirements, academic rankings, academic resources, admission data, and employment situations of schools. It calculates the fitness scores between users and schools through a deep neural collaborative filtering model and preliminarily screens out a suitable school recommendation list. Then, based on the preliminary school selection recommendation list, the system constructs a school selection game optimization model, considers the school selection strategies of users and the admission strategies of schools, and optimizes the school selection matching matrix through the entropy regularization method, and finally generates an optimized school selection recommendation plan.

[0142] During the experiment, 8,000 users were selected from a data set of 9,853 users as the training set, and the remaining 1,853 users were used as the test set. This simulated the school selection decisions of users with different backgrounds in a real application environment. The list of schools recommended by the system was matched with the list of schools ultimately applied for by the users to evaluate the recommendation accuracy.

[0143] Table 1 Experimental data comparison table

[0144] Indicator Traditional collaborative filtering Matrix factorization Method of the present invention Accuracy of school selection recommendation 71.2％ 75.8％ 88.7％ Application success rate 63.5％ 67.2％ 79.5％ Recommendation calculation time (seconds) 2.31 2.05 1.25 User satisfaction 74.8％ 78.1％ 89.3％ Proportion of the top five recommended being adopted 57.1％ 62.8％ 83.7％ Variance of admission rate and user matching degree 0.217 0.194 0.165

[0145] Experimental data show that the method of the present invention is superior to traditional collaborative filtering and matrix decomposition methods in multiple key indicators, and improves the accuracy of school recommendation, application success rate and user satisfaction. In terms of school recommendation accuracy, the accuracy of traditional collaborative filtering methods is only 71.2%, and the matrix decomposition method is slightly improved to 75.8%. The method of the present invention combines deep neural collaborative filtering with entropy regularized game optimization, and the accuracy reaches 88.7%, which is at least 17.5% higher than the traditional method, indicating that this method can more accurately capture the matching relationship between users and colleges.

[0146] In terms of application success rate, the success rate of traditional methods is between 63.5% and 67.2%, while the method of the present invention increases it to 79.5%, indicating that the optimized recommendation system not only improves the matching accuracy, but also can effectively predict the user's admission probability, so that the recommended colleges and universities are more in line with the user's actual application situation. In addition, user satisfaction with the recommendation system has also been significantly improved. The user satisfaction of the traditional collaborative filtering method is 74.8%, the matrix decomposition method is 78.1%, and the method of the present invention reaches 89.3%, indicating that users have a higher degree of recognition of the recommendation results, and the recommendation system can better meet personalized needs.

[0147] In terms of computational efficiency, the recommended calculation time of the method of the present invention is 1.25 seconds, which is optimized by 45.9% compared with 2.31 seconds of traditional collaborative filtering and 2.05 seconds of matrix decomposition. This shows that while improving the accuracy, the method still maintains a high computational efficiency and can meet the real-time school selection needs of large-scale users. In addition, the adoption rate of users for the top five recommended colleges and universities has also increased significantly. The adoption rate of the traditional method is 57.1%, while the adoption rate of the method of the present invention reaches 83.7%, an increase of 26.6%. This proves that the optimized recommendation scheme can better meet the user's school selection intentions and reduce the number of invalid recommendations.

[0148] The method of the present invention also performs well in terms of the stability of the admission rate of institutions and the matching degree of users. The matching variance of the traditional collaborative filtering method is 0.217, and that of the matrix factorization method is 0.194. However, the method of the present invention is reduced to 0.165, indicating that the rationality of users' application for institutions is optimized, the quality of the student source of institutions is more stable, and the recommendation system has achieved remarkable results in the two-way matching optimization.

[0149] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered within the protection scope of the present invention.

Claims

1. A school selection recommendation optimization method based on deep learning, characterized in that: The steps include: S1. Collect user personal information, build a user feature data set, and perform standardization to generate a user feature vector set; S2. Collect information about institutions, build an institution database, and perform data mapping to generate a set of institution feature vectors; S3. Based on the user feature vector set and the school feature vector set, a deep neural collaborative filtering method is used for feature matching, and the mean square error loss function and contrast loss function are used to optimize the deep neural collaborative filtering method, calculate the fitness score between the user and the school, and generate a preliminary school selection recommendation list; S4. Based on the preliminary school selection recommendation list, a school selection game optimization model is constructed, the strategy space and benefit function are defined, and according to the fitness score between users and schools, the school selection game optimization model based on entropy regularization game is used to optimize the strategy distribution of users and schools, and a school selection matching matrix is ​​constructed to generate an optimized school selection recommendation plan; S5. Based on the optimized school recommendation scheme, dynamically update the parameters of the user's deep neural collaborative filtering method to optimize the adaptability of the deep neural collaborative filtering method.

2. The deep learning-based school selection recommendation optimization method according to claim 1, characterized in that: The personal information includes personal academic information, interest preferences, career goals and socioeconomic conditions.

3. The deep learning-based school selection recommendation optimization method according to claim 1, characterized in that: The institution information includes enrollment requirements, admission standards, major rankings, academic resources, employment rates and relevant historical admission data.

4. The deep learning-based school selection recommendation optimization method according to claim 1, characterized in that: The S3 specifically includes: S31. Construct a user-school interaction matrix based on the user feature vector and the school feature vector: R={r ij |U′ j ∈U′,C′ j ∈C′}; Among them, R represents the user-institution interaction matrix, r ij Represents user U i With Institution C j The fitness score between them, ∈ means "belongs to", U′ j and C′ j denote the user feature vector and the institution feature vector respectively, U′ and C′ denote the user feature vector set and the institution feature vector set respectively; S32, use the deep neural collaborative filtering method to perform feature matching, embed and map the user feature vector and the school feature vector, and generate an embedding vector E U and E C : E U =φ(W U U′+b U ),E C =φ(W C C′+b C ); Among them, E U represents the user embedding vector, E C represents the school embedding vector, φ represents the nonlinear activation function, W U and W C Represents the weight matrix of user characteristics and school characteristics, b U and b C They represent the corresponding bias vectors, U′ and C′ represent the user feature vector set and the institution feature vector set respectively; S33. Calculate the matching score between the user embedding vector and the school embedding vector, using a fitness calculation method based on bilinear interaction: Among them, s ij Represents user U i With Institution C j The initial fitness score between s represents the interaction weight matrix, E U represents the user embedding vector, E C represents the embedding vector of the institution, and T represents the transposition operation; S34, use multi-layer perceptron to score the initial fitness s ij Perform nonlinear transformation and calculate the final fitness score: h1=σ(W1s ij +b1); h2=σ(W2h1+b2); Among them, s ij Represents user U i With Institution C j The initial fitness score between , σ represents the nonlinear activation function, W1, W2 and W3 represent the weight matrix of the multilayer perceptron, b1, b2 and b3 represent the bias vector of the multilayer perceptron, Indicates the final calculated user-school compatibility score; S35. Construct a training data set, which is composed of the user's past school selection decisions, application records, and admission results: D={(U′ i ,C′ j ,r ij )|i∈N,j∈M}; Among them, D represents the training data set, N represents the user set, M represents the school set, and E C represents the embedding vector of the institution, U′ j and C′ j Represent the user feature vector and the institution feature vector respectively, r ij Represents user U i Apply to school C j Actual feedback rating: Among them, r ij Represents user U i Apply to school C j The actual feedback score of , λ represents the smoothing factor, which is used to distinguish the degree of score decay of unaccepted users, Rank j Indicates institution C j Comprehensive ranking, Threshold represents an adjusted admission threshold; S36. Use the mean square error loss function and the contrast loss function to jointly optimize the deep neural collaborative filtering method and construct a comprehensive loss function: Among them, L represents the comprehensive loss function, |D| represents the total number of training samples, D represents the training data set, and r ij Represents user U i Apply to school C j Actual feedback rating of represents the final calculated user-school fit score, α represents the weight of the contrast loss, δ represents the score difference threshold, and max represents the maximum value between the two. represents the user's predicted rating for the wrongly matched institutions; S37. Based on the trained deep neural collaborative filtering method, the compatibility scores of all users and all colleges are calculated, and the scores are sorted to generate a preliminary list of recommended schools: Here, R′ represents the preliminary list of recommended schools, and sort represents the sorting function, which sorts the schools from high to low according to the fitness score.

5. The deep learning-based school selection recommendation optimization method according to claim 1, characterized in that: The S4 specifically includes: S41. Based on the preliminary school selection recommendation list R′, a school selection game optimization model is constructed, and the user set N and the school set M are defined. i The school selection strategy is A i , College C j The admission strategy is B j , construct the strategy space: A i ={a i1 ,a i2 ,...,a ik },B j ={b j1 ,b j2 ,...,b jl }; Among them, A i Represents user U i The school selection strategy of user U i The set of institutions to be selected, B j Indicates institution C j Describe the admission strategy of institution C j The set of students to be selected, a ik Represents user U i Select School C k The probability of b jl Indicates institution C j Admission User U i probability; S42. Use the preliminary school selection recommendation list R′ as the user's preliminary recommendation strategy, define the user's preliminary selection probability, and define the initial admission strategy probability of the school based on historical admission data: Among them, P0(A) represents the user's initial selection probability, Represents user U i For College C j The fitness score of , e represents the natural exponential function, Represents user U i For College C k The fitness score of the school, K represents the total number of colleges, Q0(B) represents the probability of the initial admission strategy of the college, ψ ij and ψ ik Indicates the historical admission data of the institution; S43. Define user U i The profit function U i (A,B) and institution C j The profit function U j (A,B), and combined with the entropy regularization term to constrain the stability of strategy selection: Among them, U i (A,B) represents user U i The profit function, U j (A,B) represents institution C j The profit function of α1, α2 and α3 represent the adjustment parameters. Represents user U i With Institution C j The fitness score between j Indicates institution C j The employment rate of graduates is C i Represents user U i The application cost, μ C Represents user U i The average of the application costs, λ1, λ2 and λ3 represent adjustment parameters, Ql i Indicates the academic quality of the user, D i Indicates institution C j Diversity of students, μ D Indicates institution C j The mean diversity of the students, R i represents the user's scientific research ability, β1 and γ1 represent adjustment parameters, τ represents the entropy regularization coefficient, P(A) and Q(B) represent the strategy probability distribution of users and institutions respectively, and H(P(A)) and H(Q(B)) represent the entropy regularization terms of users and institutions respectively: H(P(A))=-∑P(A)log2P(A), H(Q(B))=-∑Q(B)log2Q(B); Among them, H(P(A)) represents the entropy regularization term of the user, H(Q(B)) represents the entropy regularization term of the institution, P(A) and Q(B) represent the strategy probability distribution of the user and the institution respectively; S44. Construct the entropy regularized game solving process, define users as followers of the game, institutions as leaders of the game, and construct the optimization goal: Among them, P * (A) represents the probability distribution of the optimal optimization strategy obtained by the user after optimization; Q * (B) represents the probability distribution of the optimal admission strategy of the school after optimization, P(A) represents the probability distribution of the user's current strategy, P0(A) represents the user's initial selection probability, which is obtained from the initial school recommendation list R′, η represents the learning rate, and controls the strategy update step size, Q(B) represents the probability distribution of the school's current strategy, Q0(B) represents the school's initial admission strategy probability, argmin represents the variable value when the maximum value is taken, and D KL represents the Kullback-Leibler divergence, U i (A,B) represents user U i The profit function, U j (A,B) represents institution C j The profit function, N represents the user set, M represents the college set; S45. Calculate the optimization parameters of the school selection strategy based on the entropy regularized gradient optimization method, and use the entropy regularized gradient descent method to update the strategy distribution of users and schools: Among them, P t+1 (A) and Q t+1 (B) represents the distribution of user and school strategies in round t+1, P t (A) and Q t (B) represents the distribution of user and school strategies in round t, η represents the learning rate, and controls the strategy update step size. represents the gradient of the user's revenue function, represents the gradient of the school's profit function, τ represents the entropy regularization coefficient, and Represents the gradient of the entropy regularization term, ensuring that the strategy does not change drastically; S46, based on the optimized user strategy P * (A) and institutional strategy Q * (B) Calculate the matching rules, construct the school selection matching matrix, and generate the optimized school selection recommendation plan: Among them, M ij represents the school selection matching matrix, θ1 and θ2 represent the thresholds for users to select schools and schools to admit users, respectively.

6. A school selection recommendation optimization system based on deep learning, executing the school selection recommendation optimization method based on deep learning according to any one of claims 1 to 6, characterized in that: include: User data processing module, used to collect user information and generate user feature vectors; The school data processing module is used to collect school information and generate school feature vectors; The deep neural collaborative filtering recommendation module is used to calculate the compatibility score between users and colleges and generate a preliminary list of recommended schools; The school selection game optimization module is used to optimize the strategy distribution of users and colleges and generate optimized school selection recommendation plans; The adaptive optimization module is used to update the parameters of the deep neural collaborative filtering method based on the optimized school selection recommendation scheme to improve its adaptability.

Images