A personalized recommendation method in a robot intelligent aided design system

Abstract

The application provides a personalized recommendation method in a robot intelligent auxiliary design system, belongs to the field of data source construction and analysis, extracts a design project, extracts a design function, a model type, a model parameter, a user operation score and a user search record, labels a supervision label, constructs a training data set, pre-trains a user function recommendation model and a personalized recommendation model, calculates a function score and a personalized recommendation score, combines the function score and the personalized recommendation score, generates a final project function recommendation list, pushes the project function recommendation list to a user, verifies a user operation score correctness, updates a user operation score and a user search record in a reinforcement learning mode, updates a user function recommendation model and a personalized recommendation model, regenerates a project function recommendation list, and repeats until a user operation score correctness convergence value is higher than a preset threshold value, so that the accuracy and flexibility of the personalized recommendation system are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data source construction analysis of personalized recommendation, and particularly relates to a personalized recommendation method in a robot intelligent auxiliary design system. BACKGROUND

[0002] Under the background of rapid development of artificial intelligence and machine learning, personalized recommendation systems have become an important tool in various industries and are widely used in e-commerce, social media and online content platforms. Personalized recommendation systems use users' historical behavior data, preferences and interests to analyze and mine potential patterns and provide customized product or service recommendations for users. In recent years, the introduction of deep learning technology has greatly improved the performance of recommendation systems, but continuous optimization is still needed to meet the increasingly diverse needs of users.

[0003] Traditional personalized recommendation techniques mainly rely on collaborative filtering and content recommendation methods. Collaborative filtering is based on the similarity of user behavior for recommendation, while content recommendation relies on the matching of item features and user interests. These methods often use simple scoring models combined with basic statistics and similarity measures for recommendation. However, these traditional methods often face performance bottlenecks when dealing with high-dimensional data and complex user behavior, and cannot fully mine users' potential needs. At the same time, traditional techniques lack dynamic learning ability and cannot update user models in real time to adapt to changing preferences, resulting in inaccurate and timely recommendation results. SUMMARY

[0004] The technical problem to be solved by the present application is how to improve the accuracy and flexibility of the personalized recommendation system.

[0005] The present application solves the above technical problems by the following technical solution: a personalized recommendation method in a robot intelligent auxiliary design system, comprising:

[0006] Step one, extracting design projects in user historical logs;

[0007] Step two, extracting design functions, model types, model parameters, user operation scores and user search records in the design projects respectively to build a data set;

[0008] Step three, annotating supervision labels based on user operation scores;

[0009] Step four, using the data set and supervision labels to build a training data set, and pre-training user function recommendation models and personalized recommendation models based on the training data set to obtain pre-trained user function recommendation models and personalized recommendation models;

[0010] Step five, using the pre-trained user function recommendation model to calculate the function score of the design project in the user's historical day, and using the pre-trained personalized recommendation model to calculate the personalized recommendation score of the design project in the user's historical day;

[0011] Step six, combining the function score and the personalized recommendation score to generate a final project function recommendation list;

[0012] Step seven, pushing the project function recommendation list to the user;

[0013] Step eight, verifying the correctness of the user operation score, and updating the user operation score and the user search record in a reinforcement learning manner;

[0014] Step nine, updating the user function recommendation model and the personalized recommendation model respectively using the updated user operation score and the user search record;

[0015] Step ten, regenerating the project function recommendation list using the updated user function recommendation model and the personalized recommendation model, repeating steps seven to ten until the convergence value of the user operation score correctness is higher than the preset threshold.

[0016] The present application can improve the accuracy of recommendation by comprehensively analyzing the user's historical log, score and search record, and establishing a more comprehensive user portrait. The present application can effectively capture the relationship between complex user preferences and design functions by labeling supervision tags based on user operation scores and training user function recommendation models and personalized recommendation models based on data extracted from historical logs and supervision tags. The present application can solve the problem that existing methods often cannot effectively mine the user's potential preferences when processing user historical data, resulting in inaccurate and timely recommendation results. The present application uses reinforcement learning to dynamically update the user operation score and the user search record, and continuously adjusts the recommendation strategy according to user feedback, so that the system can adapt to the changing needs of users in real time, improve the accuracy and flexibility of the personalized recommendation system, and significantly improve the user experience and satisfaction.

[0017] Preferably, in step two, the extraction of design functions is based on the basic modules involved in the design project, and the functions implemented by the basic modules are determined by analyzing the basic modules contained in the design project. The determination of model parameters is based on the model parameter type established for each design project and its corresponding design function. The user operation score is obtained by analyzing the model type and model parameters in the design project. The user search record is established by tracking the user's operation behavior in the design process, and records the search keywords and frequency of the user in the use process.

[0018] Preferably, in step three, the supervision tag is a set of design functions in the design project, and is represented as:

[0019] D(i) = {f1, f2,..., fk}, i = 1, 2,..., n k}

[0020] where D(i) is the design function set of the ith design project, f j is a specific design function, j = 1, 2,..., k.

[0021] Preferably, the training data set in step four is represented as:

[0022] T = {F(i), MT(i), MP(i), U(i), R(i), D(i)}

[0023] where T represents the training data set, F(i) represents the design function of the ith design project, MT(i) represents the model type of the ith design project, MP(i) represents the model parameters of the ith design project, U(i) represents the user operation score of the ith design project, R(i) is the user search record of the ith design project, and D(i) is the supervision label of the ith design project.

[0024] Preferably, the user function recommendation model is pre-trained using a deep neural network in step four, and the pre-training process includes:

[0025] 4.1.1, extract design function, model type, model parameter, user operation score, user search record, and input deep neural network N to predict supervision label The calculation formula is:

[0026]

[0027] wherein, is the supervision label predicted by the deep neural network, F(i) represents the design function of the ith design project, MT(i) represents the model type of the ith design project, MP(i) represents the model parameters of the ith design project, U(i) represents the user operation score of the ith design project, R(i) is the user search record of the ith design project, and N is the deep neural network;

[0028] 4.1.2, calculate the loss function, stop training when the loss function is minimum, update all parameters of the deep neural network N using the back propagation algorithm, and obtain the pre-trained user function recommendation model. The calculation formula of the loss function is:

[0029]

[0030] wherein, L represents the loss function of the user function recommendation model pre-training, D j (i) is the actual supervision label, is a supervised label predicted by a deep neural network, and n is the number of supervised labels.

[0031] Preferably, the step four utilizes a deep neural network to pre-train the personalized recommendation model, and the pre-training process comprises:

[0032] 4.2.1, input the model type MT(i) into the personalized recommendation model P, and predict to obtain a personalized score S(i):

[0033] S(i) = P(MT(i))

[0034] wherein S(i) represents the predicted personalized score, and P is the personalized recommendation model;

[0035] 4.2.2, input the user search record R(i) and the user behavior feature B(i) into the personalized recommendation model P to obtain a user item preference sequence P(i):

[0036] P(i) = P(R(i), B(i))

[0037] 4.2.3, calculate a loss function, stop training when the loss function is minimum, update all parameters of the deep neural network N by using a back propagation algorithm, obtain the pre-trained personalized recommendation model, and the calculation formula of the loss function is:

[0038]

[0039] wherein L p represents the loss function of the pre-training of the personalized recommendation model, U k (i) is the actual user operation score, S k (i) is the predicted personalized score, w k is a weight calculated according to the user item preference sequence, and m is the number of scores.

[0040] Preferably, the output of the deep neural network is represented as N(X), X represents input data, the input data X includes design functions, model types, model parameters, user operation scores and user search records, feature extraction is performed through multiple hidden layers, and the output of the tth hidden layer can be represented as:

[0041] N(X) = H t (W t ·H t-1 (W t-1 …H1(W1·X+b1)+b2)+…+b t )

[0042] wherein H t is the output of the tth hidden layer, W t is the weight of the tth hidden layer, and bt is the corresponding bias;

[0043] The calculation formula of the loss function L is defined as:

[0044]

[0045] where D j is the actual supervision label, N(X j ) is the network output, and n is the sample number;

[0046] After calculating the loss function, the gradient with respect to the model parameters is obtained, and the derivative of the loss function with respect to each weight and bias is calculated through the chain rule. The update rule is:

[0047]

[0048] where η is the learning rate, which controls the update step size.

[0049] Preferably, the specific process of step five of calculating the function score comprises:

[0050] 5.1, evaluate the recommended score of each design function in the design project, add and average the recommended scores of all design functions to obtain the function score, and the calculation formula of the recommended score of each design function is:

[0051]

[0052] where S i,n represents the recommended score of the nth design function of the ith design project, w j is the design weight of the design project of each training sample, F j (i) is the design function of the jth training sample in the ith design project;

[0053] 5.2, input the user search record R(i) and the user behavior feature B(i) into the trained personalized recommendation model P to obtain the user project preference sequence P(i) of each design function, and calculate the preference weighted score P of all design projects, and the calculation formula is:

[0054]

[0055] where O(i) is the output of the deep neural network, W i is the preference selection weight of the ith design project;

[0056] 5.3, update the preference selection weight, and the update formula is: W new = W old + α·(R-W old), where R represents the actual preference of the user to the model, and a is a preset update rate.

[0057] Preferably, the process of updating the user operation score and the user search record in step eight by using the reinforcement learning method comprises:

[0058] 8.1. Integrating the user search record, behavior and user operation score into the state of reinforcement learning, selecting the recommended item as the action, calculating the user preference weighted score, defining the reward function and the state value function;

[0059] 8.2. Comparing the result output by the recommendation model with the actual user feedback to determine whether the recommended design function is effective;

[0060] 8.3. Calculating the user feedback and updating the model;

[0061] 8.4. Dynamically adjusting the user behavior and preference.

[0062] Preferably, the process of updating the user function recommendation model and the personalized recommendation model in step nine comprises:

[0063] 9.1. Inputting the user search record, behavior data and score information into the personalized recommendation model;

[0064] 9.2. Calculating the user preference weighted set by the personalized recommendation model, comparing the output of the personalized recommendation model with the actual score, and calculating the loss function, wherein the calculation formula of the loss function L is:

[0065]

[0066] where Y k is the actual user operation score, is the score predicted by the personalized recommendation model, and n is the sample quantity;

[0067] 9.3. Adjusting the parameters of the user function recommendation model by using the gradient descent method, and the update mode of the model parameters is:

[0068]

[0069] where θ represents the model parameters to be updated, is the derivative of the loss function with respect to the parameters, reflecting the influence of the current parameters on the loss;

[0070] 9.4. Updating the score of each design function.

[0071] The application has the advantages that:

[0072] (1) The present application can improve the accuracy of recommendation by comprehensively analyzing the user's historical log, score and search record to establish a more comprehensive user portrait, can label the supervision label based on the user operation score, and can train the user function recommendation model and the personalized recommendation model based on the data extracted from the historical log and the supervision label, which can effectively capture the complex user preference and the relationship between the designed functions, solve the problem that the existing method cannot effectively mine the potential preference of the user when processing the user historical data, leading to the inaccuracy and timeliness of the recommendation result, dynamically update the user operation score and the user search record by using the reinforcement learning method, continuously adjust the recommendation strategy according to the user feedback, so that the system can adapt to the changing needs of the user in real time, improve the accuracy and flexibility of the personalized recommendation system, and significantly improve the user experience and satisfaction.

[0073] (2) The present application takes the set of design functions in the design project as the supervision label, and the user operation score reflects the overall satisfaction of the user to the design project, and the supervision label is characterized by the function characteristics of the design project, and the relationship between the user operation score and the design function is established in this way, forming the training data set in the supervised learning, and the user operation score provides a target value for the supervised learning, and the supervision label is the set of design functions related to the user operation score, and in the training process, the user operation score and the supervision label constitute the input and output together, helping the model to learn how to predict the user's satisfaction according to the different performance of the design function, so as to provide personalized design function suggestions in the subsequent recommendation.

[0074] (3) In the process of training the personalized recommendation model by using the deep neural network, the loss function includes the weight calculated according to the user project preference sequence, and through back propagation and optimization, the model can automatically learn the preference information related to the user project preference sequence.

[0075] (4) The present application adopts deep neural network to train the user function recommendation model and the personalized recommendation model respectively, so that the recommendation system has the ability of self-learning, realizes continuous learning and self-optimization. Reinforcement learning enables the system to learn in each interaction of the user, and through continuous collection of user feedback and behavior data, the system can continuously optimize the model parameters and the recommendation strategy to adapt to the change of user preference. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 A flow chart of a personalized recommendation method in a robot intelligent auxiliary design system provided for an embodiment of the present application;

[0077] Figure 2 A flow chart of training data set construction in a personalized recommendation method in a robot intelligent auxiliary design system provided for an embodiment of the present application;

[0078] Figure 3 A sub-flow chart of a personalized recommendation method in a robot intelligent aided design system provided by an embodiment of the present application is shown in the figure.

[0079] Figure 4 A sub-flow chart of a personalized recommendation method in a robot intelligent aided design system provided by an embodiment of the present application is shown in the figure.

[0080] Figure 5 A sub-flow chart of a personalized recommendation method in a robot intelligent aided design system provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0081] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application are described in detail below with reference to the embodiments and the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0082] As shown in the figure, the present embodiment provides a personalized recommendation method in a robot intelligent aided design system, which includes the following steps: Figure 1

[0083] Step one, extract the design project in the user history log.

[0084] Step two, extract the design function, model type, model parameter, user operation score and user search record in the design project respectively, and construct the data set.

[0085] Among them, the extraction of the design function is based on the basic module involved in the design project. By analyzing the basic module contained in the design project, the function realized by the basic module is determined. For example, the design function of the ith design project can be expressed as F(i) = {p1, p2,..., pn}, where F(i) is the design function of the ith design project, pj is the corresponding model parameter type, and j = 1, 2,..., n.

[0086] The determination of the model parameter depends on the model parameter type established by each design project and its corresponding design function.

[0087] The user operation score of the ith design project can be expressed as:

[0088] U(i) = g(MT(i), MP(i))

[0089] ​wherein U(i) is the user operation score of the ith design project, MT(i) represents the model type of the ith design project, MP(i) represents the model parameter of the ith design project, g is a score function for mapping the model type and the model parameter to the user's satisfaction score.

[0090] The user search record is established by tracking the user's operation behavior in the design process, recording the search keywords and frequency of the user in the use process, and the user search record R(i) of the ith design project can be represented as R(i) = {k1, k2,..., km}, wherein R(i) is the user search record of the ith design project, kj is the user search keyword, and j = 1, 2,..., m.

[0091] For example, if a design project involves a sensor and an actuator, the corresponding design function can include data acquisition and action control, the model parameter can be the sensitivity of the sensor and the response time of the actuator, and the user operation score can be 5 points, indicating the user's satisfaction with the design project, and the user search record can show that the user frequently searches for sensor types and actuator selection. By integrating these information, a data set is constructed for subsequent model training and optimization.

[0092] Step three, labeling the supervised label based on the user operation score.

[0093] The supervised label is a set of design functions in the design project, which can be represented as:

[0094] D(i) = {f1, f2,..., fk} k}

[0095] wherein D(i) is the set of design functions of the ith design project, f j is a specific design function, and j = 1, 2,..., k.

[0096] For example, if the model type of a design project is "linear regression" and the model parameters are "learning rate 0.01, iteration number 1000", the user operation score may be 4.5, indicating that the user's satisfaction with the design project is relatively high. At this time, the supervision label D(i) may include "prediction accuracy", "model stability" and "computing efficiency", which constitute the functional characteristics of the design project. The user operation score U(i) reflects the overall satisfaction of the user with the design project, and the supervision label D(i) represents the specific performance of the design project through its functional characteristics. In this way, the relationship between the user operation score U(i) and the design function is established, forming the training data set in supervised learning. The user operation score U(i) provides a target value for supervised learning, and the supervision label D(i) is a set of design functions related to the user operation score U(i). In the training process, the user operation score U(i) and the supervision label D(i) together constitute the input and output, helping the model to learn how to predict user satisfaction according to the different performance of the design function, so as to provide personalized design function recommendations in subsequent recommendations. The user operation score and its corresponding supervision label will be used for the training and optimization of subsequent models to ensure that the recommendation system can provide personalized design function recommendations based on the user's historical interaction behavior.

[0097] Step four, using the data set constructed in step two and the supervision label obtained in step three to construct a training data set, and pre-training the user function recommendation model and the personalized recommendation model based on the training data set to obtain the pre-trained user function recommendation model and the personalized recommendation model.

[0098] As shown in Figure 2 , the data set is constructed according to the design function, model type, model parameter, user operation score and user search record of the design project extracted in step two, and the training data set can be represented as:

[0099] T = {F(i), MT(i), MP(i), U(i), R(i), D(i)}

[0100] Where T represents the training data set, F(i) represents the design function of the i-th design project, MT(i) represents the model type of the i-th design project, MP(i) represents the model parameter of the i-th design project, U(i) represents the user operation score of the i-th design project, R(i) is the user search record of the i-th design project, and D(i) is the supervision label of the i-th design project.

[0101] As shown in Figure 3 , the user function recommendation model and the personalized recommendation model are pre-trained using deep neural networks in step four, wherein the process of pre-training the user function recommendation model based on the training data set T includes:

[0102] 4.1.1, extract design function, model type, model parameter, user operation score, user search record, and input deep neural network N to predict supervision label The calculation formula is:

[0103]

[0104] wherein, is the supervision label predicted by the deep neural network, F(i) represents the design function of the i-th design project, MT(i) represents the model type of the i-th design project, MP(i) represents the model parameter of the i-th design project, U(i) represents the user operation score of the i-th design project, R(i) is the user search record of the i-th design project, and N is the deep neural network.

[0105] 4.1.2, calculate the loss function, stop training when the loss function is minimum, update all parameters of the deep neural network N using the back propagation algorithm, and obtain the pre-trained user function recommendation model. The loss function is represented by mean square error, and the calculation formula is:

[0106]

[0107] wherein, L represents the loss function of the pre-training of the user function recommendation model, D j (i) is the actual supervision label, is the supervision label predicted by the deep neural network, and n is the number of supervision labels.

[0108] The process of pre-training the personalized recommendation model based on the training data set T includes:

[0109] 4.2.1, input the model type MT(i) into the personalized recommendation model P to predict the personalized score S(i):

[0110] S(i)=P(MT(i))

[0111] wherein, S(i) represents the predicted personalized score, and P is the personalized recommendation model.

[0112] 4.2.2, input the user search record R(i) and the user behavior feature B(i) into the personalized recommendation model P to obtain the user project preference sequence P(i):

[0113] P(i)=P(R(i),B(i))

[0114] 4.2.3, calculate the loss function, stop training when the loss function is minimum, update all parameters of the deep neural network N using the back propagation algorithm, and obtain the pre-trained personalized recommendation model. The calculation formula of the loss function is:

[0115]

[0116] where L p represents the loss function of the pre-training of the personalized recommendation model, U k (i) is the actual user operation score, S k (i) is the predicted personalized score, w k is the weight calculated according to the user item preference sequence, and m is the number of scores. In the recommendation system, the user's preference is represented by the input of the model (for example, the user's embedding vector). These vectors are learned through the training process, so the loss function itself does not directly contain the user item preference sequence P(i), but through back propagation and optimization, the model will automatically learn the preference information related to the user item preference sequence P(i).

[0117] By minimizing the loss function and performing back propagation, the optimized parameters of the personalized recommendation model are finally obtained, and the pre-trained personalized recommendation model is obtained based on the optimized parameters.

[0118] Through pre-training, both the user function recommendation model and the personalized recommendation model are effectively trained, providing a basis for subsequent personalized recommendation. For example, if a user gives a high score to the design function of "prediction accuracy" in multiple design projects, the model will adjust the weight of this function accordingly during training, thereby improving the accuracy of the recommendation and user satisfaction.

[0119] In the process of pre-training the user function recommendation model and the personalized recommendation model using deep neural networks, gradient method is used to update all parameters in the deep neural network to minimize the loss function.

[0120] The output of the deep neural network is represented as N(X), and X represents the input data, which includes design functions, model types, model parameters, user operation scores, and user search records. Feature extraction is performed through multiple hidden layers, and the output of the tth hidden layer can be represented as:

[0121] N(X)=H t (W t ·H t-1 (W t-1 …H1(W1·X+b1)+b2)+…+b t )

[0122] where H tis the output of the t-th hidden layer, W t is the weight of the t-th hidden layer, b t is the corresponding bias, through this hierarchical structure, the model can capture the complex relationship in the input data.

[0123] To optimize the performance of the model, define the loss function L, the purpose is to measure the difference between the model output and the actual supervised label, usually use mean square error to represent:

[0124]

[0125] where D j is the actual supervised label, N(X j ) is the network output, n is the number of samples. After calculating the loss function, the gradient about the model parameters can be obtained, the derivative of the loss function to each weight and bias is calculated through the chain rule, through the update rule:

[0126]

[0127] where η is the learning rate, control the update step, using the gradient of the loss function, the model can adjust the weight to better fit the training data set. For example, suppose the model predicts the value is 6 in a certain iteration, while the actual label is 4, at this time, the loss function will calculate a larger error, reflecting the inaccuracy of the model's prediction, through the calculation of the gradient of the loss function to update the model parameters, the model will strive to reduce this error, so as to improve the accuracy of the next prediction. Through multiple iterations, the network can gradually learn the relationship between user preferences and design functions, making personalized recommendations more accurate and effective.

[0128] Step five, using the pre-trained user function recommendation model to calculate the function score of the design project in the user history day, using the pre-trained personalized recommendation model to calculate the personalized recommendation score of the design project in the user history day.

[0129] The core of calculating the function score is to evaluate the recommendation score of each design function in the design project, refer to Figure 3 , the specific process of calculating the function score includes:

[0130] 5.1, evaluate the recommendation score of each design function in the design project, add up the recommendation scores of all design functions and get the function score, the calculation formula of the recommendation score of each design function is:

[0131]

[0132] where S i,n represents the recommendation score of the n-th design function of the i-th design project, w jis the design weight of the design project of each training sample, F j (i) is the design function of the jth training sample in the ith design project, and this weighted summation process enables the user function recommendation model to comprehensively consider the importance of multiple design functions, thereby generating a comprehensive score.

[0133] 5.2, input the user search record R(i) and the user behavior feature B(i) into the trained personalized recommendation model P to obtain the user project preference sequence P(i) of each design function, and calculate the preference weighted score P of all design projects, the calculation formula is:

[0134]

[0135] wherein O(i) is the output of the deep neural network, the output O(i) is obtained by processing the model type and model parameters through the deep neural network, the output O(i) represents the prediction result based on the user historical data and the current input, W i is the preference selection weight of the ith design project, and the formula shows how the preference selection weight is combined with the model output to generate a final preference weighted score, reflecting the user's preference for a specific design function.

[0136] 5.3, update the preference selection weight, and in the process of updating the weight, the user's preference for the model will affect the final recommendation result. The update formula can be simplified as:

[0137] W new = W old + α·(R-W old )

[0138] wherein W new represents the updated model weight, i.e. the new recommendation strategy or preference adjusted according to the user feedback, W old represents the model weight before updating, i.e. the recommendation strategy or preference of the model before the current time, W old = W i , R represents the actual preference of the user for the model, and α is the preset update rate. This mechanism ensures that in each iteration, the model can continuously adjust the recommendation strategy based on user feedback, gradually optimize the weight, and make it better reflect the real preference of the user.

[0139] For example, if a user scores a function as 5 in a design project, and the function obtains a higher weight in the previous model, the updated weight will make the function obtain a higher score in future recommendations. This dynamic adjustment process allows the model to more accurately match user needs, and through multiple iterations, ultimately achieves more personalized recommendation results and enhances user experience.

[0140] In step five, the function score is calculated by the user function recommendation model, the preference weighted score of the design project is calculated by the personalized recommendation model, and the personalized score of the design project is calculated based on the function score and the preference weighted score of the design project. The function score is obtained by comprehensively calculating the recommendation score of each design function. Specifically, first, the score of each function in each design project needs to be calculated, which reflects the degree of recommendation of the function. Then, the function scores are averaged to obtain the overall project recommendation score, i.e., the function score. This process ensures that the model can consider the influence of all functions on user decision-making, thereby providing a comprehensive evaluation.

[0141] The user function model aims to extract the actual needs and function preferences of the user, while the personalized recommendation model is optimized based on the user's historical behavior and scores. The combination of the two can more accurately reflect the user's interests and needs. In actual application, assuming that a design project contains three design functions with recommendation scores of 4, 5, and 3, we add these scores and take the average to obtain the project recommendation score of 4.

[0142] This scoring mechanism provides users with an intuitive reference, helping them make choices among numerous design projects. With continuous iteration and optimization of the model, the final recommendation results will better meet the personalized needs of users, thereby improving the accuracy and effectiveness of the recommendation system.

[0143] Step six, combine the function score and the personalized recommendation score to generate the final project function recommendation list.

[0144] Step seven, push the project function recommendation list to the user.

[0145] Step eight, verify the correctness of the user operation score and update the user operation score and user search record using reinforcement learning, as shown in Figure 4 The specific process of updating the user operation score and user search record using reinforcement learning includes the following:

[0146] 8.1, integrate the user search record, behavior, and user operation score into the state of reinforcement learning, which provides the current context information of the user for the model. The action of reinforcement learning is to select the recommended project based on these states, and the user preference weighted score is derived therefrom, reflecting the relevance of the recommendation results.

[0147] Within this framework, a reward function and a state value function can be defined. The reward function represents the user's feedback on the recommendation results, while the state value function represents the long-term rewards that can be obtained by taking specific actions in the current state. This structure enables the model to gradually learn how to choose the optimal recommendation action in a given state, thereby improving user satisfaction.

[0148] 8.2. Comparing the results output by the recommendation model with the actual user feedback to determine whether the recommended design features are effective. This comparison provides the basis for subsequent feedback loops, enabling the model to accurately identify which recommendations are successful and which need improvement. For each design project, it is necessary to ensure that the recommended design features truly meet the needs of the user.

[0149] 8.3. The process of calculating user feedback involves the estimation of optimal solutions. The degree of matching between the user's actual preferences and the recommended features will affect the calculation of rewards, thereby influencing the direction of subsequent model learning. This feedback process ensures continuous optimization of the model.

[0150] 8.4. The step of updating user behavior and preferences is based on the difference between the recommended results and actual values. By introducing user update weights and update rates, user behavior and preferences can be dynamically adjusted. For example, when the model's output recommendation matches the user's actual preferences, the weight of that item can be increased to make it appear more frequently in subsequent recommendations.

[0151] In this way, the model adapts to the changing needs of users, making the recommendation system more personalized and accurate. As shown in the flowchart for reinforcement learning optimization of user ratings and recommendation models, if a design project has received high ratings in past recommendations, the model will increase the recommended score for that project, giving it higher priority in future recommendations. Figure 5

[0152] In summary, this optimization process, through continuous feedback loops and dynamic adjustments, enables the recommendation system to effectively learn user preferences, thereby improving the accuracy of recommendations and user satisfaction. Through this reinforcement learning approach, the system not only adapts to users' immediate needs but also extracts potential patterns from historical data, further enhancing the quality of recommendations.

[0153] Step 9. Update the user function recommendation model and the personalized recommendation model using the updated user operation scores and user search records, respectively. This includes the following processes:

[0154] ​9.1, Input user search records, behavior data, and rating information into the personalized recommendation model, which provides valuable information for the personalized recommendation model to better understand the user's interest points. For ease of processing, the user's input information is converted into a vector representation, which is initially a zero vector and then updated by input data.

[0155] 9.2, The personalized recommendation model calculates a user preference weighting set, which quantifies the user's preference for different design project design functions. The accuracy of the personalized recommendation model can be evaluated by comparing the user's historical ratings with the output of the current personalized recommendation model. This process involves using the loss function of the deep neural network to minimize the difference between the model's predicted values and the user's actual operation ratings. The formula for calculating the loss function is:

[0156]

[0157] where Y k is the actual rating of the kth design project, is the predicted rating of the kth design project, and n is the sample size. By calculating the mean square error of these ratings, the personalized recommendation model can accurately understand its prediction ability.

[0158] 9.3, Adjust the parameters of the user function recommendation model using gradient descent. In the user function recommendation model update process, similar processing is performed for each design project and each design function. As shown in Figure 5 , the flowchart of the user function recommendation model update process and parameter adjustment. For each model type and model parameter, first output the user preference weighting set and compare it with the actual rating to get the corresponding error. Then use the gradient descent method to adjust the model parameters, which ensures that the user function recommendation model can be continuously optimized. The same scoring and updating operations are performed for each design function to ensure that all functions are reasonably recommended.

[0159] If the user gives a high rating in a certain design project, the model will record this preference. In subsequent updates, the preference weighting set for that project will be improved, so that in future recommendations, the project is more likely to be prioritized for the user. This dynamic adjustment mechanism not only improves the accuracy of the model's recommendations, but also enhances user satisfaction, ultimately achieving a personalized recommendation experience. Through continuous iteration and updating, the recommendation system can better adapt to the changing needs of users and provide more accurate services.

[0160] 9.4、Updating the score of each design function, adjusting the parameters through gradient descent is a key step in updating each design function. Specifically, first define the score as the target value of each design project, denoted as S n Then, for the update rate of the jth design function, denoted as n, adjust it by calculating the derivative of the loss function with respect to the score. This derivative reflects the difference between the current model output and the actual user preference. If the derivative is positive, it means the current score is too low; if it is negative, it means the score is too high.

[0161] In this process, user behavior data is input into the personalized recommendation model to obtain a user preference weighted set. This set integrates user search records and behavior data to form a feedback information for a specific design function. Then, the objective function is solved to optimize the model to reduce errors.

[0162] Through gradient descent, the objective function can be expressed in the form of a loss function. The derivative of the loss function will guide us on how to adjust the model parameters. Specifically, assuming the loss function is L, the parameters can be updated using the following formula:

[0163]

[0164] where θ represents the model parameters to be updated, is the derivative of the loss function with respect to the parameters, reflecting the current parameter's impact on the loss. In this way, the parameters can be effectively updated in the direction of reducing the loss, thereby improving the prediction accuracy of the model.

[0165] Further, the update process includes inputting user behavior data k times and combining the preset initial value to optimize the objective function. For each input, we calculate the loss function and update the parameters according to the derivative to ensure that the model gradually approaches the optimal state.

[0166] For example, if a user's score for a certain design function is significantly higher than the model's predicted score, the derivative of the loss function will be a large negative value, indicating that the score of that design function needs to be significantly increased. By applying gradient descent, the model will adjust the parameters so that in the next prediction, the score of that function will be increased, better meeting the user's actual preferences. After multiple iterations, the final model will be able to accurately reflect the user's true preferences for each design function.

[0167] Step ten, regenerate the project function recommendation list using the updated user function recommendation model and personalized recommendation model, repeat steps seven to ten until the user operation score accuracy convergence value is higher than the preset threshold.

[0168] The present application can deeply understand the user's preferences and needs by extracting the user's historical logs, user operation scores and user search records, thereby generating more accurate design function recommendations and improving the accuracy of personalized recommendations. This data-driven approach enables the recommendation system to identify the specific needs of users in different design projects, thereby reducing the time and effort wasted by users in the selection process.

[0169] By continuously updating user operation scores and feedback, the recommendation system can adjust the recommended content in real time to better meet the actual needs of users. The introduction of reinforcement learning algorithms enables the system to learn from each user interaction, thereby optimizing the recommendation strategy, improving user experience, and enhancing user trust and satisfaction with the system.

[0170] The present application can help users quickly find the most suitable design solution and improve design efficiency. By generating a function score for the design project and a recommended list, users can more clearly see the strengths and weaknesses of different projects, thereby making better decisions and promoting the optimization of design projects and the design process.

[0171] The use of deep learning and reinforcement learning algorithms enables the recommendation system to have the ability to learn and optimize itself. By continuously collecting user feedback and behavior data, the system can continuously optimize model parameters and recommendation strategies to adapt to changes in user preferences. This dynamic adaptability ensures that the system can always provide high-quality recommendations and improve overall intelligence.

[0172] By analyzing user preferences for different design functions, the system can recommend diverse design options, encourage users to explore new design ideas, and promote innovation. This function not only meets the current needs of users, but also provides them with more possibilities, supporting them to make more extensive attempts in the design process and improving the diversity and innovation of design functions.

[0173] The present application not only improves the effectiveness of personalized recommendations, but also significantly improves user experience, design efficiency and system intelligence, ultimately achieving higher levels of user satisfaction and system value.

[0174] The above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art should understand that they can modify the technical solutions described in the foregoing examples, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A personalized recommendation method within a robot-assisted intelligent design system, characterized by: include: S1. Extract design projects from user history logs; S2. Extract the design functions, model types, model parameters, user operation scores, and user search records in the design projects to construct a data set; S3, label supervision labels based on user operation scores; S4. Use the data set and the supervision label to construct a training data set, and pre-train the user function recommendation model and the personalized recommendation model based on the training data set to obtain the pre-trained user function recommendation model and the personalized recommendation model; wherein the process of pre-training the user function recommendation model includes: extracting design functions, model types, model parameters, user operation scores, user search records, and inputting them into the deep neural network Predicting supervised labels : 、 、 、 、 Respectively i Design function, model type, model parameters, user operation score, and user search history of each design project; The process of pre-training the personalized recommendation model includes: Input to personalized recommendation model , predict and get personalized score : ; User search history and user behavior characteristics Enter the personalized recommendation model , get the user item preference sequence : ; S5. Calculate function scores for the user's design projects in history using the pre-trained user function recommendation model, and calculate personalized recommendation scores for the user's design projects in history using the pre-trained personalized recommendation model. S6. Combine the function score and the personalized recommendation score to generate a final project function recommendation list; S7. Push a list of project function recommendations to the user; S8. Verify the correctness of the user operation score and update the user operation score and user search history using reinforcement learning; S9, using the updated user operation score and user search history to update the user function recommendation model and the personalized recommendation model respectively; S10. Regenerate the project function recommendation list using the updated user function recommendation model and personalized recommendation model, and repeat steps S7 to S10 until the convergence value of the user operation score accuracy rate is higher than the preset threshold.

2. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: In step S2, the design function is extracted based on the basic modules involved in the design project. By analyzing the basic modules included in the design project, the functions implemented by the basic modules are determined; the model parameters are determined based on the model parameter type established for each design project and the corresponding design function; Obtain user operation scores by analyzing the model types and model parameters in the design project; User search records are established by tracking users' operational behaviors in the design process, recording the users' search keywords and frequency during use.

3. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: The supervision label in step S3 is the set of design features in the design project, expressed as: in, It is i A collection of design features for a design project. It is a specific design feature. .

4. A personalized recommendation method in a robot-assisted intelligent design system according to claim 3, characterized in that: The training data set in step S4 is represented as: in, represents the training dataset, Indicates the i The design function of a design project, Indicates the i Model type for each design project, Indicates the i Model parameters for each design project, Indicates the i User operation ratings of design projects, It is i User search records for design projects, It is i Supervisory label for a design project.

5. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: In step S4, the user function recommendation model is pre-trained using a deep neural network, and the loss function is calculated. When the loss function is minimized, the training is stopped and the deep neural network is trained using the back propagation algorithm. N All parameters are updated to obtain the pre-trained user function recommendation model. The loss function is calculated as follows: in, represents the loss function of the pre-training of the user feature recommendation model, is the actual supervision label, is the supervised label predicted by the deep neural network, n is the number of supervision labels.

6. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: In step S4, the personalized recommendation model is pre-trained using a deep neural network, and the loss function is calculated. When the loss function is minimized, the training is stopped and the deep neural network is trained using the back propagation algorithm. N All parameters are updated to obtain the pre-trained personalized recommendation model. The loss function is calculated as follows: in, represents the loss function of the personalized recommendation model pre-training, is the actual user operation score, is the predicted personalized score, is the weight calculated based on the user-item preference sequence, m is the number of ratings.

7. A personalized recommendation method in a robot-assisted intelligent design system according to claim 5 or 6, characterized in that: The output of the deep neural network is represented as , Indicates input data, input data Including design functions, model types, model parameters, user operation scores and user search records, feature extraction is performed through multiple hidden layers, t The output of the hidden layer can be expressed as: in, It is t The output of the hidden layer, It is t The weights of the hidden layers, is the corresponding bias; Define the loss function L The calculation formula is: in, is the actual supervision label, is the network output, n is the sample size; After calculating the loss function, we get the gradient of the model parameters. We use the chain rule to calculate the derivative of the loss function with respect to each weight and bias. The update rule is: in, η is the learning rate, which controls the update step size.

8. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: The specific process of calculating the function score in step S5 includes: 5.

1. Evaluate the recommendation score of each design function in the design project. Add up the recommendation scores of all design functions and take the average to get the function score. The calculation formula for the recommendation score of each design function is: in, Indicates the i The first design project n Recommendation score of design features, is the design weight of the design item for each training sample, It is i Design project No. j The design function of training samples; 5.

2. User search history and user behavior characteristics Input and trained personalized recommendation model , get the user item preference sequence for each design function , calculate the preference weighted scores of all design items , the calculation formula is: in, is the output of the deep neural network, It is i The preference selection weight of each design project; 5.

3. Update the preference selection weights. The update formula is: ,in, represents the updated model weights, represents the model weight before updating, R represents the user's actual preference for the model, α is the preset update rate.

9. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: The process of updating the user operation score and user search history using reinforcement learning in step S8 includes: 8.

1. Integrate user search history, behavior, and user action scores into reinforcement learning states, select recommended items as actions, calculate user preference weighted scores, and define reward functions and state value functions. 8.

2. Compare the output of the recommendation model with actual user feedback to determine whether the recommended design features are effective; 8.

3. Calculate user feedback and update the model; 8.

4. Dynamically adjust user behavior and preferences.

10. A personalized recommendation method in a robot-assisted intelligent design system according to claim 1, characterized in that: The process of updating the user function recommendation model and the personalized recommendation model in step S9 includes: 9.

1. Input user search history, behavior data, and rating information into the personalized recommendation model; 9.

2. The personalized recommendation model calculates the user preference weighted set, compares the output of the personalized recommendation model with the actual score, and calculates the loss function. The calculation formula is: in, Score actual user actions, is the score predicted by the personalized recommendation model, n is the sample size; 9.

3. Use the gradient descent method to adjust the parameters of the user function recommendation model. The update method of the model parameters is: in, θ represents the model parameters to be updated, It is the derivative of the loss function with respect to the parameters, reflecting the impact of the current parameters on the loss; 9.

4. Update the score of each design feature.

Images