Industrial Big Data Cockpit View Recommendation Method Based on Meta-Assisted Learning Framework

Abstract

The present invention relates to an industrial big data cockpit view recommendation method based on a meta-assisted learning framework. The method includes: Step 1, importing the scoring information of various views by industrial big data cockpit users into the cockpit view recommendation system; Step 2, the data processing module of the cockpit view recommendation system reads and processes the scoring information to generate feature vectors of users and views; Step 3, the data sampling module of the cockpit view recommendation system generates a set of sampling data about users and views according to the feature vectors and the scoring information; Step 4, the main model of the cockpit view recommendation system and the auxiliary model constructed based on self-supervised contrast learning calculate the corresponding loss functions L pri and L aux and other steps. The superior effect of the present invention lies in that: through the contribution of the auxiliary model, the problem of data sparsity existing in the traditional industrial big data cockpit view recommendation system is alleviated.

Description

Technical Field

[0001] The present invention relates to the field of recommendation systems, and particularly to a recommendation method for industrial big data cockpit views based on a meta-aided learning framework. Background Art

[0002] With the continuous development of Internet technology, more and more technologies have been applied in the industrial field. Most of the industrial fields have changed from the original manual production control to semi-automatic or fully automatic production modes. While reducing a large amount of labor costs, this production mode has also brought a lot of industrial data reflecting the current production status. Most of these industrial data contain valuable information such as the operating status of machines and the overall production efficiency. Therefore, it is very important for cockpit users to comprehensively understand these data.

[0003] An efficient data visualization technology can help industrial production operators better understand the current status of machines and the overall production situation of the factory, so as to make further instructions and achieve the goal of maximizing the overall profit of the factory. However, different industrial big data cockpit operators have different sensitivities to different data visualization views. Users may be more likely to read data information from some types of visualization views, while it is more difficult to obtain the expressed data status from other types of views. And since the types of data visualization views are very numerous, the views used in the industrial data scenario not only include simple tree diagrams, bar charts, etc., but also include views covering high-dimensional information such as area charts, box plots, etc. Therefore, most industrial big data cockpit users only often use a part of their preferred views and lack the exploration of a large part of other views. This phenomenon to a certain extent limits the functionality of the big data cockpit. The present invention proposes a view recommendation method for industrial big data cockpits to help the visualization technology explore other views that users may have preferences for.

[0004] However, for the industrial big data cockpit view recommendation scenario in real life, there are also some technical challenges to be solved. For the users of industrial data cockpits, most of them have only used a very small part of the views, have a high or low interest in a very small part of the views, and have no historical interaction information with most of the views. Therefore, in the view recommendation system, this will make the data available for model training relatively sparse, which may lead to a relatively poor recommendation effect. Therefore, how to alleviate or even solve the problem of data sparsity faced by the industrial big data cockpit view recommendation method is a huge challenge.

[0005] Similarly, since the training of the overall recommendation model needs to update the corresponding feature vectors through the loss functions of various users and views, some recommendation techniques applied in practice do not consider the personalized information of the user side and the view side, which will make it difficult for the model to complete the personalized recommendation task. Therefore, in the view recommendation system, how to integrate the personalized information from the user and view perspectives into the model has become a major problem.

[0006] Finally, the low model update efficiency and the large time and space costs are also a problem faced by the current industrial big data cockpit view recommendation method. For different users and different views, a feature vector is required for correspondence and the corresponding training is completed. In most scenarios for solving the data sparsity problem, auxiliary tasks need to be introduced to help the main task better learn the representation of the feature vector. Such an update architecture will increase the time and space costs, which is unacceptable for the industrial data cockpit scenario.

[0007] For example, Chinese Patent Application No. CN 202122835104.2 discloses a big data cockpit for a digital information data management platform, including a seat. A bottom plate is installed at the bottom of the seat, a moving component is arranged inside a moving box, and an adjusting component is arranged inside an adjusting box. The adjusting component includes a rotating rod and relates to the field of digital devices. A moving component is arranged at the bottom. By driving the motor in the moving component, the seat can be driven to move left and right for adjustment, which is convenient for the staff to use. And by rotating the threaded rod to drive the moving plate to move, the displacement of the seat caused by external forces during the movement can be prevented. An adjusting component is arranged at the bottom of the seat. By pushing the seat to rotate and driving the shaft rod to rotate fixedly through the bottom plate, the angle adjustment of the seat can be quickly realized. The connection between the bottom plate and the inside of the adjustment box can be increased through the rotating rod, and the stability of the bottom plate can be increased.

[0008] For another example, Chinese Patent Application No. CN107748763A discloses a 3D touch system based on an ELA big data cockpit and a data processing method. The data processing method includes the following steps: S1: The first Logstash server obtains real-time Log data through a web server, and temporarily stores the real-time Log data collected by the web server into a Redis cache module; S2: The real-time Log data temporarily stored in the Redis cache module is transmitted to an ElasticSearch server for storage; S3: The data is sorted by the ElasticSearch server and then transmitted to a database storage module; S4: The database storage module transmits the data to a data analysis and processing module. The present invention provides an effective solution for big data computing and data presentation of services such as user behavior analysis and user profiling, effectively solving the problems of limited storage capacity, low reading efficiency, and rigid report implementation process of traditional data warehouses. The present invention also combines 3D touch technology to create an ELA big data cockpit system that can provide gesture operation services with custom gestures in the air.

[0009] For another example, Chinese Patent Application No. 108657186A discloses an interaction method and device for an intelligent cockpit applied to an autonomous vehicle, belonging to the field of autonomous driving technology in intelligent driving. The method includes: when it is detected that an unidentified user is about to enter or is located inside the intelligent cockpit, obtaining the facial information of the user; identifying the user's identity according to the facial information of the user, so that a big data server obtains corresponding user preference information according to the user identity, and finds corresponding information data based on the user preference information; receiving the information data from the big data server; automatically or according to a received play instruction, playing the information data. This application can play content of interest to different users according to their preferences, and the user can obtain the information of interest without manual operation, solving the problem of single human-vehicle interaction mode in the prior art, bringing a good driving experience to the user, and achieving the effect of making the human-vehicle interaction content more rich.

[0010] None of the above patent applications solve the problem of data sparsity faced by the industrial big data cockpit view recommendation method, as well as the problems of how to integrate personalized information from the user and view perspectives into the model, low model update efficiency, and high time and space costs. Summary of the Invention

[0011] In view of the deficiencies and blanks in the prior art, the present invention proposes a recommendation method for industrial big data cockpit views based on a meta-assisted learning framework.

[0012] The recommendation method for industrial big data cockpit views based on the meta-assisted learning framework includes:

[0013] Step 1, import the scoring information of industrial big data cockpit users for various views into the cockpit view recommendation system;

[0014] Step 2, the data processing module of the cockpit view recommendation system reads and processes the scoring information to generate feature vectors of users and views;

[0015] Step 3, the data sampling module of the cockpit view recommendation system generates a set of sampling data about users and views according to the feature vectors and scoring information;

[0016] Step 4, the main model of the cockpit view recommendation system and the auxiliary model constructed based on self-supervised contrast learning calculate the corresponding loss functions L pri and L aux ; and introduce trainable contribution coefficients λ U and λ V , and obtain the objective function L of the overall model through weighted summation;

[0017] Step 5, the parameter update module of the cockpit view recommendation system, through the objective function L of the overall model, adopts a two-layer meta-aided learning update architecture to update the parameters of the main model of the cockpit view recommendation system and the feature vectors of users and views in the inner layer, and update the parameters of the contrast learning auxiliary model of users and views and the contribution coefficient λ U and λ V , and use the implicit gradient method for acceleration;

[0018] Step 6, repeatedly execute Step 3, Step 4, and Step 5 until the parameters of the main model, the parameters of the auxiliary model, the feature vectors of users and views, and the contribution coefficient λ U and λ V no longer change;

[0019] Step 7, the recommendation module of the cockpit view recommendation system calculates the preference values of users for various views through the feature vectors of users and views, and recommends different industrial big data cockpit views for different users.

[0020] Furthermore, in Step 1, the scoring data of industrial big data cockpit users for various views includes multiple scoring triples (u, v, r), where: u represents the ID of different users, with subscripts from 1–N, and N is the number of users; v represents the ID of different views, with subscripts from 1–M, and M is the number of views; r represents the score of user u for different views v, and the value range is 1–5.

[0021] Furthermore, in Step 2, the data processing module of the cockpit view recommendation system reads and processes the scoring information to generate feature vectors of users and views, including:

[0022] Step 201: The data processing module of the cockpit view recommendation system obtains the number of users N and the number of views M according to the ID subscripts of users and views in the scoring information, and generates a sparse scoring matrix R. N*M , and the corresponding positions of the scoring matrix R N*M store the scores of specific users for specific views.

[0023] Step 202: The data processing module of the system obtains the user set U and the view set V. According to the number of users N and the number of views M, N and M trainable feature vectors are generated for the user side and the view Figure 1 side respectively. Through the construction of the mapping, each user and each view are in one-to-one correspondence with these trainable feature vectors.

[0024] Furthermore, in step 3, the data sampling module of the cockpit view recommendation system generates a set of sampling data about users and views according to the feature vectors and the scoring information, specifically including:

[0025] Step 301: The data sampling module preprocesses the scoring information. According to the value of the score of user u for the view, the views with a user score greater than or equal to 4 are defined as the views preferred by user u, that is, the positive sample views of user u Other views are defined as the views not preferred by user u, that is, the negative sample views of user u which can be expressed as the following formula (1):

[0026]

[0027] In the above formula (1): R Uv represents the score of user u for view v, and V represents the view set;

[0028] Step 302: According to the relationships of all users to all views obtained in step 301, calculate the similarities between different users to obtain the similarity matrix U on the user side N*N , and at the same time calculate the similarities between different views to obtain the similarity matrix V on the view side M*M ;

[0029] Step 303: According to the similarity matrix between users obtained in step 302, users with a similarity greater than or equal to 0.3 are selected for each user and regarded as its positive correlated user neighbors. At the same time, the remaining users are regarded as the negative correlated user neighbors of the current user. The same operation is also performed on the view side to select positive and negative correlated neighbors for each view;

[0030] Step 304, the data sampling module generates a set of samples, which consists of seven parts, including the user, the views preferred by the user, the views not preferred by the user, the positive correlated neighbors of the user, the negative correlated neighbors of the user, the positive correlated neighbors of the views preferred by the user, and the negative correlated neighbors of the views preferred by the user.

[0031] Further, in step 4, the main model of the cockpit view recommendation system and the auxiliary model constructed based on self-supervised contrastive learning calculate the corresponding loss function L pri and L aux ; and introduce the trainable contribution coefficients λ U and λ V , and obtain the objective function L of the overall model through weighted summation, including:

[0032] Step 401, the sampling data generated in step 3 is passed to the main model and the auxiliary model. The main model uses the method of BPR-loss to obtain the loss function L pri ; meanwhile, the auxiliary model uses the idea of self-supervised contrastive learning to approximate the lower bound of the mutual information by maximizing the InfoNCE between the user, the view and their corresponding positive and negative sample neighbors, and obtains the loss function L aux corresponding to the auxiliary model, where L aax consists of the auxiliary loss function on the user side and the auxiliary loss function on the view side. The calculation of the objective function is shown in the following formulas (2) to (5):

[0033]

[0034]

[0035]

[0036]

[0037] In the above formulas (2) to (5), u represents a specific user, v p represents the view preferred by user u, v n represents the view not preferred by user u, u + represents the positive correlated neighbor of user u, u - represents the negative correlated neighbor of user u, v + represents the positive correlated neighbor of user v, v - represents the negative correlated neighbor of user v, D pri represents the data sampled and generated by the main model, D aux represents the data sampled and generated by the auxiliary model;

[0038] Step 402: After weighted summation of the loss functions of the main model and the auxiliary model calculated in Step 401, the objective function L of the overall model is obtained. The calculation of the objective function of the overall model is shown in the following formula (6):

[0039]

[0040] In the above formula (6), σ represents the activation function sigmoid, and λ U represents the trainable contribution weight parameter on the user side, and λ V represents the trainable contribution weight parameter on the user side.

[0041] Furthermore, in Step 5, the parameter update module of the cockpit view recommendation system, through the objective function L of the overall model, adopts a two-layer meta-aided learning update architecture to update the parameters of the main model and the feature vectors of the user and the view in the inner layer, and update the parameters of the auxiliary model and the contribution coefficient λ U and λ V in the outer layer, and uses the implicit gradient method for acceleration, including:[[]]

[0042] Step 501: In the inner layer of the parameter update module, use the objective function L of the overall model obtained in Step 4, fix the parameters of the auxiliary model and the contribution coefficient λ U and λ V unchanged, update the parameters of the main model and the feature vectors of the user and the view, and retain the computational graph of the parameters of the auxiliary model and the contribution coefficient λ U and λ V . Here, any optimizer can be used for the update process, and the Adam optimizer is adopted by default. The update steps can be expressed as the following formula (7):

[0043]

[0044] In the above formula (7), θ includes the parameters of the main model and the feature vectors of the user and the view, includes the parameters of the auxiliary model and the contribution coefficient λ U and λ V , α represents the learning rate, represents the first-order partial derivative of the parameter θ; represents the optimal value of θ updated through the inner layer of the meta-aided learning framework while keeping unchanged;

[0045] Step 502: The parameters obtained after the update in Step 501 retain a part of gradients, let the parameter be used as the input of the main model, calculate the loss function again, and obtain And by means of implicit gradient update, a regularization term is added to the objective function, transforming the calculation of the high-order derivative into a quadratic optimization problem. Finally, the meta-gradient is calculated through the conjugate gradient algorithm to update the parameters of the auxiliary model. The update steps can be expressed as the following formula (8):

[0046]

[0047] In the above formula (8), β represents the learning rate, and d represents the calculation of the full derivative.

[0048] Further, in step 6, steps 3, 4, and 5 are repeatedly executed until the parameters of each model no longer change.

[0049] Further, in step 7, the recommendation module of the cockpit view recommendation system calculates the preference values of users for various views through the feature vectors of users and views, and recommends different industrial big data cockpit views for different users, including:[[]]

[0050] Step 701, the recommendation module of the system obtains the preference value matrix P between users and views through the calculation of the inner product based on the trained feature vectors of users and views N*M , where each row in the matrix represents the preference values of a specific user for all M different views;

[0051] Step 702, according to the preference value matrix P of users for views obtained in step 701 N*M , sort according to the magnitude of each row of preference values, and screen out the top 20 views with the highest preference values for a specific user as the industrial big data cockpit views recommended for the specific user.

[0052] Compared with the prior art, the superior effect of the industrial big data cockpit view recommendation method based on the meta - auxiliary learning framework of the present invention lies in:[[]]

[0053] 1. The industrial big data cockpit view recommendation method based on the meta - auxiliary learning framework of the present invention alleviates the problem of data sparsity existing in the traditional industrial big data cockpit view recommendation system through the contribution of the auxiliary model; in view of the fact that the auxiliary model in the present invention uses the idea of self - supervised contrastive learning and is constructed by maximizing the mutual information between users, views and their corresponding positive and negative sample neighbors, which plays a role of data augmentation. Therefore, the overall model is trained not only relying on the source data, but also on the augmented neighbor data, making the perception range of the overall model larger and improving the recommendation effect of the model.

[0054] 2. The industrial big data cockpit view recommendation method based on the meta-assisted learning framework described in the present invention utilizes a two-layer parameter update architecture for meta-assisted learning, overcoming the phenomenon of negative transfer that often occurs in traditional multi-task learning. Through the parameter update architecture of meta-assisted learning, the system separates the parameter updates of the cockpit view recommendation main model and the auxiliary model as a whole. In the first stage, the parameters of the auxiliary model are kept unchanged while the parameters of the main model are updated, and at the same time, the updated parameters of the main model retain a part of the gradients of the auxiliary model. And when updating the parameters of the auxiliary model in the second stage, the parameters of the main model are fixed. Such a parameter update architecture can make each update affect only the parameters of a single task while ensuring the relevance between tasks, which not only guarantees the overall effect of the multi-task model but also avoids the occurrence of the negative transfer phenomenon.

[0055] 3. The industrial big data cockpit view recommendation method based on the meta-assisted learning framework described in the present invention uses an implicit gradient update method when updating the parameters of the auxiliary model, reducing the high costs in time and space. Since meta-assisted learning requires constructing a two-layer update method, it is necessary to retain the computational graph of the parameters of the auxiliary model after updating the parameters of the cockpit view recommendation main model in the first stage. Then, when updating the parameters of the auxiliary model in the second stage, it is necessary to calculate and store the higher-order derivatives, which will result in a large burden in terms of time and space, thereby reducing the efficiency of the model and increasing the cost of the industrial big data cockpit view recommendation system.

[0056] 4. The industrial big data cockpit view recommendation method based on the meta-assisted learning framework described in the present invention avoids this step of calculation through the implicit gradient method and directly obtains the parameter gradients of the auxiliary model through formula derivation to complete the parameter update.

[0057] 5. The industrial big data cockpit view recommendation method based on the meta-assisted learning framework described in the present invention can be loaded into any existing industrial big data cockpit view recommendation model and can significantly improve the recommendation effect. Since the industrial big data cockpit view recommendation method based on the meta-assisted learning framework described in the present invention mainly proposes an auxiliary model framework applied to the industrial big data cockpit view recommendation system, the main contributions are to propose an efficient auxiliary task model based on self-supervised contrast learning and a two-layer parameter optimization architecture. Therefore, it can be loaded into many existing recommendation models and make these recommendation models act as the cockpit view recommendation main model mentioned in the industrial big data cockpit view recommendation method based on the meta-assisted learning framework described in the present invention.

[0058] 6. Since the industrial big data cockpit view recommendation method based on the meta-aided learning framework of the present invention plays a role of data augmentation, expanding the receptive field of the overall model, it can improve the recommendation effect. Moreover, through the double-layer parameter optimization architecture and the implicit gradient update method, the time and space consumption required by the overall system can be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is the overall framework diagram of the industrial big data cockpit view recommendation method based on the meta-aided learning framework of the present invention.

[0060] Figure 2 It is the schematic diagram of the double-layer parameter update framework of the industrial big data cockpit view recommendation method based on the meta-aided learning framework of the present invention.

[0061] Figure 3 It is the schematic diagram of the Top-K recall rate of the model with personalized contribution parameters and the original model.

[0062] Figure 4 It is the schematic diagram of the Top-K normalized discounted cumulative gain of the model with personalized contribution parameters and the original model.

[0063] Figure 5 It is the schematic diagram of the Top-K precision rate of the model with personalized contribution parameters and the original model.

[0064] Figure 6 It is the schematic diagram of the comparison of the computational costs between the implicit gradient descent method and the traditional gradient descent method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the method of the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0066] Combined with Figure 1 , the method of the present invention proposes an industrial big data cockpit view recommendation algorithm based on the meta-aided learning framework, and the method includes:

[0067] Step 1, import the scoring information of various views of the industrial big data cockpit users into the cockpit view recommendation system:

[0068] Through pre-operations such as obtaining and processing the original data, the system administrators obtain the usage habit data of the industrial big data cockpit users. A common form of these data is the scoring data of the industrial big data cockpit users for different views. Each line of the data file is a triple (u, v, r), where u represents the ID of different users, with subscripts from 1 to N, and N is the number of users; v represents the ID of different views, with subscripts from 1 to M, and M is the number of views; r represents the score of user u for view v, and the value range is from 1 to 5. The magnitude of the score value indicates the preference degree of the user for the view. After the system administrators import the scoring data file by setting the data path and other methods, the system automatically applies for a certain amount of memory space to store these scoring data;

[0069] Step 2, the data processing module of the cockpit view recommendation system reads and processes the scoring information to generate feature vectors of users and views:

[0070] Step 201, the data processing module of the system obtains the number of users N and the number of views M according to the ID subscripts of users and views in the scoring information, and generates a sparse scoring matrix R N*M , and the corresponding positions of the matrix store the scores of specific users for specific views. In this step, the data processing module of the system first accesses the memory space storing the scoring data, counts the types of users and views, obtains the number of users N and the number of views M, and then converts the data into a scoring matrix R N*M , and stores it in the corresponding memory space;

[0071] Step 202, the data processing module of the system obtains the user set U and the view set V. According to the number of users N and the number of views M, N and M trainable feature vectors are generated for the user side and the view side respectively, and through the construction of the mapping, each user and each view are in one-to-one correspondence with these trainable feature vectors. In this step, the system sets N + M trainable feature vectors, corresponding to N users and M views respectively, and each feature vector is set to 50 dimensions; Figure 1 Step 3, the data sampling module of the cockpit view recommendation system generates a set of sampling data about users and views according to the feature vectors and the scoring information:

[0072] Step 301, the data sampling module preprocesses the scoring information. According to the value of the score of user u for the view, the views with a user score greater than or equal to 4 are defined as the views preferred by user u, that is, the positive sample views of user u

[0073] Other views are defined as the views not preferred by user u, that is, the negative sample views of user u They are expressed in the following formula in turn: They are expressed in the following formula in turn:

[0074]

[0075]

[0076] In the above formula, R uv represents the score of user u for view v, and V represents the view set;

[0077] Step 302: According to the relationships between all users and all views obtained in Step 301, calculate the similarity between different users to obtain the similarity matrix U on the user side N*N , and at the same time calculate the similarity between different views to obtain the similarity matrix V on the view side M*M :

[0078] In this step, first, according to the positive and negative sample relationships between users for samples, generate a one-hot vector of length M for each user u. For each bit of the vector, if the corresponding view belongs to then the corresponding bit of the vector is 1; otherwise, if the corresponding view belongs to then the corresponding bit of the vector is 0; similarly, for the view side, perform the same operation, generate a one-hot vector of length N for each view v. For each bit of the vector, if the corresponding user satisfies R uv ≥4, then the corresponding bit is 1, otherwise it is 0. Then, obtain the similarity matrix U on the user side by calculating the cosine similarity between the one-hot vectors of users and views N*N and the similarity matrix V on the view side M*M ;

[0079] Step 303: According to the similarity matrix between users obtained in Step 302, screen out users with a similarity greater than or equal to 0.3 for each user, regarded as its positive correlation user neighbors, and at the same time regard the remaining other users as the negative correlation user neighbors of the current user. Perform the same operation on the view side to screen out positive and negative correlation neighbors for each view;

[0080] Step 304: The data sampling module generates a set of samples. A set of samples consists of 7 parts, including users, views preferred by users, views not preferred by users, positive correlation neighbors of users, negative correlation neighbors of users, positive correlation neighbors of views preferred by users, and negative correlation neighbors of views preferred by users;

[0081] Step 4: The main model of the cockpit view recommendation system and the auxiliary model constructed based on self-supervised contrast learning calculate the corresponding loss functions L pri and L aux ; and introduce the trainable contribution coefficients λ U and λV , the objective function L of the overall model is obtained by weighted summation:

[0082] Step 401, the sampled data generated in Step 3 is passed to the main model and the auxiliary model. The main model uses the BPR-loss method to calculate the loss function L pri ; meanwhile, the auxiliary model uses the idea of self-supervised contrastive learning to approximate the lower bound of the mutual information by maximizing the InfoNCE between the user, view and their corresponding positive and negative sample neighbors, and calculates the loss function L corresponding to the auxiliary model aux , where L aux consists of the auxiliary loss function on the user side and the auxiliary loss function on the view side . The calculation of the objective function is shown as follows:

[0083]

[0084]

[0085]

[0086]

[0087] In the above formula, u represents a specific user, v p represents the view preferred by user u, v n represents the view not preferred by user u, u + represents the positive neighbor of user u, u - represents the negative neighbor of user u, v + represents the positive neighbor of user v, v - represents the negative neighbor of user v, D pri represents the data sampled and generated by the main model, D aux represents the data sampled and generated by the auxiliary model. In the overall architecture of the present invention, the main model can select any existing industrial big data cockpit view recommendation model, and can complete the change of the main model through some simple model selection operations; if the main model is not specifically set, matrix factorization will be default selected as the main model to complete the subsequent cockpit view recommendation task;

[0088] Step 402, the weighted summation of the loss functions of the main model and the auxiliary model calculated in Step 401 is used to obtain the objective function L of the overall model. The calculation of the objective function of the overall model is shown as follows:

[0089]

[0090] In the above formula, σ represents the activation function sigmiod, λU Denote the contribution weight parameter on the user side that can be trained, λ V Denote the contribution weight parameter on the user side that can be trained;

[0091] The method described in the present invention sets a trainable contribution weight parameter for different users and views. This parameter will be continuously updated adaptively during the parameter update process, and is mapped to the interval (0,1) through the sigmoid activation function, and then weighted and summed to obtain the objective function L of the overall model;

[0092] Step 5, the parameter update module of the cockpit view recommendation system, through the objective function L of the overall model, adopts a two-layer meta-aided learning update architecture. In the inner layer, update the parameters of the cockpit view recommendation main model and the feature vectors of users and views, and in the outer layer, update the parameters of the contrastive learning auxiliary model of users and views and the contribution coefficient λ U and λ V , and use the implicit gradient method for acceleration:

[0093] Step 501, in the inner layer of the parameter update module, use the objective function L of the overall model obtained in Step 4, fix the parameters of the auxiliary model and the contribution coefficient λ U and λ V unchanged, update the parameters of the main model and the feature vectors of users and views, and retain the parameters of the auxiliary model and the contribution coefficient λ U and λ V of the computational graph. Here, any optimizer can be used in the update process, and the Adam optimizer is adopted by default. The update steps are shown as follows:

[0094]

[0095] In the above formula, θ includes the parameters of the main model and the feature vectors of users and views, includes the parameters of the auxiliary model and the contribution coefficient λ U and λ V , α represents the learning rate, represents the first-order partial derivative of the parameter θ; represents the optimal value of θ updated through the inner layer of the meta-aided learning framework while keeping unchanged;

[0096] Combined with Figure 2 , the parameter θ and the parameter are respectively calculated through the main model and the auxiliary model to obtain the overall objective function Then, through any optimizer, if not specified, the method described in the present invention adopts Adam, update the parameter θ and obtain

[0097] Step 502, the parameters obtained after the update in Step 501 keeps a part of the gradient, making the parameter as the input of the main model, calculates the loss function again, and obtains and by the method of implicit gradient update, adds a regularization term to the objective function, transforms the calculation of the higher-order derivative into a quadratic optimization problem, and finally calculates the meta-gradient through the conjugate gradient algorithm to update the parameters of the auxiliary model. The update steps can be expressed as follows:

[0098]

[0099] In the above formula, β represents the learning rate, and d represents the calculation of the total derivative;

[0100] Combined with Figure 2 , after the system is updated to obtain , since when this parameter is obtained by the update, the method of the present invention keeps the computational graph of the parameter , so the parameter contains the gradient of, and since directly using the traditional gradient update method would require the calculation and storage of higher-order derivatives, which would bring a computational and memory burden to the whole. Therefore, the method of the present invention adds a regularization term about the parameter θ and the parameter at, making the overall objective function, through the derivation of the derivative, transform into a quadratic optimization problem, so that the gradient of for the parameter can be implicitly calculated through the conjugate gradient algorithm, and the calculation and storage of higher-order derivatives can be avoided. The derivation formulas are shown as follows in sequence: In the above formula, γ represents the regularization coefficient, d represents the calculation of the total derivative,

[0101]

[0102]

[0103]

[0104]

[0105] In the above formula, γ represents the regularization coefficient, d represents the calculation of the total derivative, represents taking the partial derivative with respect to the parameter θ first and then taking the partial derivative with respect to the parameter . For a matrix multiplication problem of A -1 B, it can be transformed into a quadratic optimization problem that can be solved by the conjugate gradient algorithm, thus avoiding the calculation and storage of higher-order derivatives;

[0106] Step 6: Repeatedly execute Step 3, Step 4, and Step 5 until the parameters of each model no longer change;

[0107] Step 7: The recommendation module of the cockpit view recommendation system calculates the preference values of users for various views through the feature vectors of users and views, and recommends different industrial big data cockpit views for different users:

[0108] The preference value of user u for view v is obtained by the inner product of the feature vectors of user u and view v. The recommendation module calculates the preference values of all users for all views, and then for each user, recommends the top 20 views with the highest preference values.

[0109] In the method of the present invention, during the implementation of the algorithm, the lengths of the feature vectors of users and views are set to 50, the number of samples selected in each sampling process is 5000, the learning rates α and β are 0.001, the number of negative sampling samples is 2, the number of steps of the conjugate gradient algorithm used in the implicit gradient method is 2, and the contribution coefficients λ U and λ V are initially set to 0.1.

[0110] To facilitate the display of the performance of the method of the present invention in the recommendation of industrial big data cockpit views, the method of the present invention has undergone a detailed functional test. Specifically, the test is carried out on two datasets that are relatively similar to the industrial big data cockpit view scenario. The selected datasets are Amazon - Books (recommended for purchasing books, abbreviated as Books) and MovieLens - 20M (recommended movies, abbreviated as ML20m). The specific evaluation indicators are the following three indicators:

[0111] 1. Top - K recall rate (Recall@K, abbreviated as R@K): Measures the ratio of the number of positive sample views among the top K recommendations to the total number of positive sample views in the view library, and measures the recall rate of the recommendation system.

[0112] 2. Top - K normalized discounted cumulative gain (NDCG@K, abbreviated as N@K): Measures the ranking results of the K recommended views and evaluates the accuracy of the ranking.

[0113] 3. Top - K precision rate (Precision@K, abbreviated as P@K): Measures the proportion of the number of positive sample views among the top K recommended views to the total number.

[0114] For all evaluation metrics, the method of the present invention has conducted multiple experiments at K = 5, 10, 15, and 20, making the experimental data more general. At the same time, in order to prove that the framework proposed in the method of the present invention can significantly improve the recommendation effect of the main model, the following four benchmark algorithms are used as the main tasks:

[0115] MF: A matrix factorization-based model that calculates the loss function through Bayesian personalized ranking. This is a classic method for learning pairwise view ranking.

[0116] Multi-VAE: A recommendation model based on the variational autoencoder (VAE) method on the Figure 1 side.

[0117] NGCF: A graph convolutional-based view recommendation model. The view recommendation model uses a bipartite graph to capture user-view interactions and learns high-order information by propagating feature vectors.

[0118] LightGCN: A graph neural network-based view recommendation model. The view recommendation model learns the feature vectors of users and views by performing multi-hop convolutions on the user-view interaction graph and uses the weighted sum of feature vectors at different depths as the final result.

[0119] The method of the present invention has conducted a series of experimental tests, comparing the above four benchmark algorithms with the corresponding algorithms based on the meta-aided learning framework (MAL) proposed by the method of the present invention. Under the condition of ensuring that all other parameters are the same, it explores how much the framework proposed by the method of the present invention can improve the performance of the model recommendation task.

[0120] As shown in Tables 1-3 below, it shows the improvement in the recommendation effect of the model based on the meta-aided learning framework (MAL) proposed by the method of the present invention compared to the corresponding original model. It can be seen from Tables 1 to 3 that compared with the benchmark algorithms, the recommendation effect of the corresponding algorithms with the framework proposed by the method of the present invention is better. The overall Top-K recall rate of the proposed model based on the meta-aided learning framework (MAL) has increased by 3% - 35% compared to the corresponding model, the Top-K normalized discounted cumulative gain has increased by 4% - 30% compared to the corresponding model, and the Top-K precision rate has increased by 3% - 30% compared to the corresponding model.

[0121] Table 1

[0122]

[0123] Table 2

[0124]

[0125] Table 3

[0126]

[0127] Meanwhile, in order to verify that the personalized contribution parameters of users and views proposed by the method of the present invention can effectively help the overall model better learn information, a series of experiments were conducted on the Amazon-Books dataset, and the changes in evaluation metrics were observed by controlling whether to introduce personalized weights. Combining Figure 5 It is found that when a learnable contribution parameter is added for different users and views, the three evaluation metrics are improved when K takes (5, 10, 15, 20), indicating that the learnable contribution parameter provides certain personalized information from users and views, enabling the overall model to better mine the differences between users and views and make recommendation results.

[0128] Meanwhile, in order to verify that the proposed double-layer meta-assisted learning update architecture can effectively reduce the computational amount of parameters during the update process, a series of experiments were conducted on the method of the present invention. When using the gradient descent method to update the inner-loop parameter θ in the double-layer update architecture, for different numbers of gradient descent steps, the time required to update the outer-loop parameter was statistically counted Combining Figure 6 It can be seen that when the number of update steps takes (1, 5, 10, 15, 20), updating the inner-loop parameter θ using the traditional gradient descent method requires more computational amount than using the implicit gradient method, thus consuming more time. Moreover, with the increase of the number of update steps in the traditional gradient descent method, the growth rate of the required time is also higher, that is, the gradient of the curve is higher. This phenomenon also shows that when using the traditional gradient descent method to update the inner-loop parameter θ, it will make the update of the outer-loop parameter more time-consuming, which also means that more time is needed to calculate the higher-order derivatives, while the implicit gradient method proposed by the method of the present invention can effectively avoid this calculation.

[0129] By observing the evaluation metrics and running time of the method of the present invention in experiments on various datasets, the following rules can be observed:

[0130] 1. Benefiting from the auxiliary model based on self-supervised contrast learning constructed by the method of the present invention, after combining this auxiliary model, the recommendation effect of the overall model has been greatly improved compared with that of the corresponding single model. This phenomenon indicates that the auxiliary model designed by the method of the present invention by constructing the mutual information between the user side and the view side in the neighborhood covers some implicit information that is difficult to learn in the main model, alleviating the problem of data sparsity to a certain extent and enabling the overall model to achieve better results.

[0131] 2. Thanks to the personalized contribution parameters proposed by the method of the present invention for different users and views, the overall model can learn the personalized information of users and views, so as to better recommend to users.

[0132] 3. The double-layer meta-assisted learning framework proposed by the method of the present invention can not only effectively combine the auxiliary model to improve the effect of the overall model, but also lead to a double-layer parameter update framework, effectively reducing the update time in the parameter update process of the outer loop and avoiding the calculation of high-order derivatives.

[0133] The present invention is not limited to the above embodiments. Without departing from the essence of the present invention, any deformation, improvement, or replacement that can be thought of by those skilled in the art falls within the protection scope of the present invention.

Claims

1. An industrial big data cockpit view recommendation method based on a meta-assisted learning framework, characterized in that including: Step 1: Import the scoring information of various views by the users of the industrial big data cockpit into the cockpit view recommendation system; Step 2: The data processing module of the cockpit view recommendation system reads and processes the scoring information to generate feature vectors of users and views; Step 3: The data sampling module of the cockpit view recommendation system generates a set of sampling data about users and views according to the feature vectors and scoring information; Step 4, the main model of the cockpit view recommendation system and the auxiliary model constructed based on self-supervised contrastive learning calculate the corresponding loss functions L pri and L aux ; and introduce the trainable contribution coefficients λ U and λ V , and obtain the objective function L of the overall model through weighted summation; the main model of the cockpit view recommendation system and the auxiliary model constructed based on self-supervised contrastive learning calculate the corresponding loss functions L pri and L aux ; and introduce the trainable contribution coefficients λ U and λ V , and obtain the objective function L of the overall model through weighted summation, including: Step 401: The sampling data generated in Step 3 is passed to the main model and the auxiliary model. The main model uses the BPR-loss method to calculate the loss function L pri ; meanwhile, the auxiliary model uses the idea of self-supervised contrastive learning to approximate the lower bound of the mutual information by maximizing the InfoNCE between the user, view and their corresponding positive and negative sample neighbors, and calculates the loss function L aux corresponding to the auxiliary model, where L aux consists of the auxiliary loss function on the user side and the auxiliary loss function on the view side. The calculation of the objective function is shown in the following formulas (2) to (5): In the above formulas (2) to (5), u represents a specific user, and v p represents the view preferred by user u, and v n represents the view not preferred by user u, and u + represents the positive correlation neighbor of user u, and u - represents the negative correlation neighbor of user u, and v + represents the positive correlation neighbor of user v, and v - represents the negative correlation neighbor of user v, and D pri represents the data sampled and generated for the main model, and D aux represents the data sampled and generated for the auxiliary model; Step 402: After the weighted summation of the loss functions of the main model and the auxiliary model calculated in Step 401, the objective function L of the overall model is obtained. Among them, the calculation of the objective function of the overall model is shown in the following formula (6): In the above formula (6), σ represents the activation function sigmiod, and λ U represents the trainable contribution weight parameter on the user side, and λ V represents the trainable contribution weight parameter on the user side; Step 5, the parameter update module of the cockpit view recommendation system, through the objective function L of the overall model, adopts a two-layer meta-aided learning update architecture to update the parameters of the main model of the cockpit view recommendation system and the feature vectors of users and views in the inner layer, and updates the parameters of the contrastive learning aided models of users and views and the contribution coefficient λ in the outer layer U and λ V , and uses the implicit gradient method for acceleration; Step 6, repeatedly execute Step 3, Step 4, and Step 5 until the parameters of the main model, the parameters of the auxiliary model, the feature vectors of the user and the view, and the contribution coefficients λ U and λ V no longer change; Step 7: The recommendation module of the cockpit view recommendation system calculates the preference values of users for various views through the feature vectors of users and views, and recommends different industrial big data cockpit views for different users.

2. The industrial big data cockpit view recommendation method based on the meta-aided learning framework according to claim 1, wherein In Step 1, when importing the scoring information of various views by the users of the industrial big data cockpit into the cockpit view recommendation system, the scoring data includes multiple scoring triples u, v, r, where: u represents the ID of different users, with subscripts from 1 to N, and N is the number of users; v represents the ID of different views, with subscripts from 1 to M, and M is the number of views; r represents the score of user u for different views v, and the value range is 1 - 5.

3. The industrial big data cockpit view recommendation method based on the meta-assisted learning framework according to claim 1, wherein In Step 2, when the data processing module of the cockpit view recommendation system reads and processes the scoring information to generate feature vectors of users and views, it includes: Step 201, the data processing module of the cockpit view recommendation system obtains the number of users N and the number of views M according to the ID subscripts of the users and views in the scoring information, and generates a sparse scoring matrix R N*M , the scoring matrix R N*M stores the scores of specific users for specific views at the corresponding positions; Step 202: The data processing module of the system obtains the user set U and the view set V. According to the number of users N and the number of views M, N and M trainable feature vectors are generated for the user side and the view side respectively, and through the construction of a mapping, each user and each view are in one-to-one correspondence with these trainable feature vectors.

4. The industrial big data cockpit view recommendation method based on the meta-assisted learning framework according to claim 1, wherein In Step 3, when the data sampling module of the cockpit view recommendation system generates a set of sampling data about users and views according to the feature vectors and scoring information, it specifically includes: Step 301, the data sampling module preprocesses the scoring information. According to the value of the user u's view score, the views with a user score greater than or equal to 4 are defined as the views preferred by the user u, that is, the positive sample views of the user u Other views are defined as the views not preferred by the user u, that is, the negative sample views of the user u It can be expressed as the following formula (1): In the above formula (1): R uv represents the score given by user u to view v, and V represents the set of views; Step 302: Calculate the similarity between different users based on the relationships between all users and all views obtained in Step 301, and obtain the user-side similarity matrix U N*N , and at the same time calculate the similarity between different views to obtain the view-side similarity matrix V M*M ; Step 303: According to the similarity matrix between users obtained in Step 302, users with a similarity greater than or equal to 0.3 are selected for each user and regarded as its positive correlation user neighbors, while the remaining users are regarded as the negative correlation user neighbors of the current user. The same operation is also performed on the view side to select positive and negative correlation neighbors for each view; Step 304: The data sampling module generates a set of samples, which consists of 7 parts, including users, views preferred by users, views not preferred by users, positive correlation neighbors of users, negative correlation neighbors of users, positive correlation neighbors of views preferred by users, and negative correlation neighbors of views preferred by users.

5. The method for recommending industrial big data cockpit views based on the meta-assisted learning framework according to claim 1, characterized in that In step 5, the parameter update module of the cockpit view recommendation system adopts a two-layer meta-aided learning update architecture through the objective function L of the overall model. The parameters of the main model, as well as the feature vectors of users and views, are updated in the inner layer, and the parameters of the auxiliary model and the contribution coefficient λ are updated in the outer layer. U and λ V , and the implicit gradient method is used for acceleration, including: Step 501, inside the inner layer of the parameter update module, use the objective function L of the overall model obtained in Step 4, fix the auxiliary model parameters and the contribution coefficient λ U and λ V remaining unchanged, update the parameters of the main model and the feature vectors of the users and views, and retain the parameters of the auxiliary model and the contribution coefficient λ U and λ V computational graph. Here, any optimizer can be used for the update process, and the Adam optimizer is used by default. The update steps can be expressed as the following equation (7): In the above formula (7), θ includes the parameters of the main model, as well as the feature vectors of the user and the view. It includes the parameters of the auxiliary model and the contribution coefficient λ U and λ V , α represents the learning rate, denotes the first-order partial derivative of the parameter θ; denotes the optimal value after θ passes through the inner layer of the meta-auxiliary learning framework and remains unchanged. Step 502, the parameters obtained after the update in Step 501 keeps a part of the gradient, and uses the parameter as the input of the main model to calculate the loss function again, obtaining and by means of the implicit gradient update method, adding a regularization term to the objective function, transforming the calculation of the higher-order derivative into a quadratic optimization problem, and finally calculating the meta-gradient through the conjugate gradient algorithm to realize the update of the parameters of the auxiliary model. The update step can be expressed as the following formula (8): In the above formula (8), β represents the learning rate, and d represents the full derivative.

6. The method for recommending views of an industrial big data cockpit based on a meta-assisted learning framework according to claim 1, wherein In Step 6, repeatedly execute Step 3, Step 4, and Step 5 until the parameters of each model no longer change.

7. The method for recommending views of an industrial big data cockpit based on a meta-assisted learning framework according to claim 1, wherein In Step 7, when the recommendation module of the cockpit view recommendation system calculates the preference values of users for various views through the feature vectors of users and views, and recommends different industrial big data cockpit views for different users, it includes: Step 701: The recommendation module of the system calculates the preference value matrix P between the user and the view through the inner product of the well-trained feature vectors of the user and the view. N*M Each row in the matrix represents the preference values of a specific user for all M different views. Step 702: According to the user's view preference value matrix P obtained in Step 701 N*M , sort according to the magnitude of the preference values in each row, and screen out the top 20 views with the highest preference values for a specific user as the industrial big data cockpit views recommended for the specific user.

Images