CN115019933A - Amblyopia training scheme recommendation method fusing GMF and CDAE - Google Patents

Abstract

A recommendation method for amblyopia training scheme fusing GMF and CDAE belongs to the technical field of computers. It comprises the following steps: 1. acquiring user information data in a weak training platform database; 2. respectively extracting data characteristics of input data through a CDAE model and a GMF model; 3. the fusion layer performs weight fusion on the outputs of the CDAE model and the GMF model; 4. and optimizing the function learning model parameters to minimize the prediction error of the prediction score and the real score, and outputting a user recommendation result. The invention introduces intelligent recommendation into the amblyopia training system, solves the artificial formula by utilizing the neural network algorithm in the field of computers, and liberates medical personnel and technical personnel; the CDAE model and the GMF model are introduced, so that the problem of insufficient recommendation accuracy caused by data sparsity is effectively solved, and the recommendation efficiency is improved.

Description

A recommended method for amblyopia training scheme integrating GMF and CDAE

technical field

The invention belongs to the field of computer technology, relates to neural collaborative filtering and recommendation systems, and in particular relates to a multimedia amblyopia training scheme recommendation method integrating generalized matrix decomposition and the same noise reduction automatic encoder.

Background technique

Amblyopia is a developmental childhood ophthalmopathy, which is a decrease in best-corrected visual acuity in one or both eyes due to abnormal visual experience during visual development, without organic eye disease.

With the development of modern Internet information technology, the treatment methods of amblyopia have also been diversified. Among them, the multimedia amblyopia training system has been widely developed. The multimedia amblyopia training system comprehensively uses the theories and methods of neurobiology, psychophysics and computer vision, utilizes the plasticity of the brain nervous system, combines amblyopia treatment with computer games, and improves visual function through various biological stimulations. The multimedia amblyopia training system adds more interest, and the training method combines the cognitive training content to improve the training effect.

At present, the existing amblyopia training system achieves the purpose of amblyopia treatment by providing users with a series of training items. These training items mostly use animation, graphics and other elements as carriers to stimulate the human eye. In this type of multimedia amblyopia system, the user needs to complete one or more courses of training, and each training needs to complete multiple training items. In recent years, there have been more and more studies on amblyopia training software. As early as 2003, Cecilia et al. developed three multimedia programs to treat children's vision. In 2005, Wang et al. combined computer technology with modern neurophysiology and child psychology to realize a multimedia amblyopia treatment system for children. Lukasz et al developed a computerized therapy system for strabismus amblyopia. Li Xiangjie designed and implemented an adaptive training system for children. Taking into account the parameters in each training item, he proposed a parameter adaptive strategy including difficulty and level. Zou Qinbin designed an online amblyopia training system based on cognitive theory, which combined with cognitive theory to improve patient compliance. Therefore, the multimedia amblyopia training system is widely used now, and the research and improvement of the amblyopia training system are of great significance for the rehabilitation of amblyopia patients.

In these multimedia amblyopia training systems, doctors or professionals usually designate training programs for patients. After these training programs are designated, they will not change until the user completes the course of training. However, the training program specified by humans has shortcomings such as singleness and limitation, which will cause users to not adjust the training program in time during the training process, and repeated training will make the user lose interest in training and cannot achieve the best training. Effect. Currently, these systems only manage the basic information of patients, do not collect data during patient training, and cannot track patient training in a timely manner. Therefore, the formula operation not only requires a lot of manpower, but also cannot achieve a better training effect because the specified training items cannot be adjusted in time during the patient training process.

Recommendation based on neural collaborative filtering has been widely used in recent years. In order to improve the effectiveness of user training programs, reduce unnecessary time loss, and reduce the labor-intensive problem of manual formulations by doctors and professionals, neural network recommendation is introduced to achieve effective recommended training programs.

SUMMARY OF THE INVENTION

Aiming at the limited and single manual designation scheme existing in the amblyopia training system, and poor user training effect, the present invention proposes a method for recommending amblyopia training scheme integrating generalized matrix factorization and cooperative noise reduction autoencoder.

The present invention provides the following technical solutions: an amblyopia training scheme recommendation method integrating GMF and CDAE, the recommendation system includes an input layer, a hidden layer, a fusion layer and an output layer; the hidden layer includes a generalized matrix factorization GMF model and a collaborative noise reduction automatic encoder CDAE model, the collaborative noise reduction auto-encoder CDAE adopts a three-layer network and two fully connected layers, which use an offset vector and a tanh activation function;

The recommendation method based on the recommendation system specifically includes the following steps:

Step 1. The input layer obtains the user information data in the database of the vulnerable training platform, and the user information data includes the user's personal basic information, relevant information of amblyopia training items and user-training item interaction information; The data is input into the CDAE model of the collaborative denoising auto-encoder, and another part of the undamaged data is input into the generalized matrix factorization GMF model;

Step 2. The collaborative noise reduction autoencoder CDAE model and the generalized matrix factorization GMF model in the hidden layer respectively perform data feature extraction on the input data;

将其传入输出层用于反馈优化；Step 3. The fusion layer performs weight fusion on the outputs of the collaborative noise reduction auto-encoder CDAE model and the generalized matrix factorization GMF model in the hidden layer to balance the high and low-order features; the predicted user-training item score is obtained through the fusion layer

Pass it into the output layer for feedback optimization;

Step 4. The output layer uses the mean square error loss MSE as the objective optimization function to learn the model parameters, so as to minimize the prediction error between the predicted score and the actual score, and the model parameters are iteratively corrected by the AdamW optimizer; finally, according to the order of the predicted scores, select The set number of TOP-N recommendations is recommended for users.

Further, in the step 1, the data input by the input layer is mapped into a latent vector describing the user and the item in the context of the latent factor model, and then used as the input of the generalized matrix factorization GMF model; The input of the same user latent vector is combined with the damaged user-item interaction data as the input of the CDAE model of the collaborative noise reduction auto-encoder. During the data processing process, the data that is 1 in the user-training item scoring vector is randomly set to 0 to generate Broken data.

Further, in the CDAE model of the collaborative noise reduction auto-encoder, the user-training item data from the input layer is input into the two-layer fully connected layer to extract the high-level interaction information of the user-training item; in the two-layer fully connected layer The activation function is introduced to realize the nonlinear process, as follows:

First, the processed damage vector in the input layer is mapped to the first layer of the fully connected layer; then the feature representation of the first fully connected layer is mapped to the second fully connected layer to reconstruct the input user-item rating vector.

Further, in the generalized matrix factorization GMF model, first multiply the user hidden vector and the corresponding element of the training item hidden vector, then multiply the vector obtained after the multiplication by the weight ^WG and add it, and then input the result into the activation function. Further mapping is done to introduce nonlinear processes.

By adopting the above-mentioned technology, compared with the prior art, the beneficial effects of the present invention are as follows:

1) The present invention introduces intelligent recommendation in the amblyopia training system, utilizes the neural network algorithm in the computer field to solve the artificial formula, and liberates medical personnel and technicians;

2) In the present invention, a collaborative noise reduction auto-encoder is introduced, the data is damaged, and the high-order features of the data are extracted, which can effectively alleviate the problem of insufficient recommendation accuracy caused by the sparsity of the data;

3) In the present invention, the collaborative noise reduction auto-encoder and the generalized matrix decomposition model are combined to extract the high-order nonlinear features and low-order linear features of the data, thereby improving the recommendation efficiency.

Description of drawings

Fig. 1 is the structural representation of the present invention;

Fig. 2 is the structural representation of the present invention;

FIG. 3 is a schematic structural diagram of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments of the description. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

On the contrary, the present invention covers any alternatives, modifications, equivalents and arrangements within the spirit and scope of the present invention as defined by the appended claims. Further, in order to give the public a better understanding of the present invention, some specific details are described in detail in the following detailed description of the present invention. The present invention can be fully understood by those skilled in the art without the description of these detailed parts.

Please refer to Figure 1-3, a method for recommending amblyopia training scheme integrating GMF and CDAE, which is based on amblyopia training scheme recommendation system, which includes an input layer for data processing, a CDAE model by a collaborative noise reduction autoencoder, a generalized matrix Decompose the hidden layer composed of the GMF model, the fusion layer used for result fusion, and the output layer.

Specifically, the input layer:

(1) Obtain the basic personal information of the user and various visual function inspection information, related information of amblyopia training items and historical training score records in the database of the amblyopia training platform from the backend. User personal basic information includes user ID, name, gender, age, corrected vision, symptom label, etc. Amblyopia training item information includes training item ID, item name, strategy ID, strategy name, category, category ID, etc. Historical training score records include User ID, training item ID, user rating.

(2) The GMF part maps the data of the input layer into latent vectors describing users and items in the context of the latent factor model, and then uses this as input.

(3) The CDAE part takes user latent vectors similar to GMF input combined with broken user-item interaction data as input. When processing the input of the CDAE part, the data that is 1 in the user-item rating vector is randomly set to 0 to generate damaged data, as shown in Figure 2.

The specific formula is as follows:

中y_ui不为0的概率，y_ui为用户对训练项目的评分，

为处理后的破损数据，q为随机概率，σ＝1/(1-q)。where p is the broken user-item interaction data

The probability that y _ui is not 0, y _ui is the user's rating of the training item,

is the damaged data after processing, q is a random probability, σ=1/(1-q).

Specifically, the hidden layer:

(1) CDAE inputs the user-training item data from the input layer to the two-layer fully connected layer to extract the high-order interaction information of the user-training item. An activation function is introduced into the two fully connected layers to realize the nonlinear process.

Specific steps are as follows:

The processed damage vector in the input layer is first mapped to the first layer of the fully connected layer. The formula is as follows:

W ^C represents the weight matrix between the nodes in the input layer and the nodes in the first layer of CDAE, and V _u represents the hidden vector of the user input node. V _u is a user-specific vector, ie for each user, there is a unique vector V _u . b is the offset vector and h( ) is an element-wise mapping function.

The feature representation of the first fully-connected layer is then mapped to the second fully-connected layer to reconstruct the input user-item rating vector. For each predicted user-item rating in the reconstructed vector, the following formula is calculated:

(2) The GMF model does not directly use the linear inner product of the user latent vector and the training item latent vector as the prediction score. First, the user latent vector and the corresponding element of the training item latent vector are multiplied, and then the vector obtained after the multiplication is multiplied by the weight. After W ^G is added, the result is input into the activation function for further mapping to introduce a nonlinear process.

The specific formula is as follows:

P _u and Q _i represent the latent vectors of user U and item I, respectively

Specifically, the fusion layer: the output results of CDAE and the output results of GMF in the hidden layer are weighted with corresponding weights

The fusion formula is:

将传入输出层用于反馈优化。Complete user-item prediction scores obtained through the fusion layer

Use the incoming output layer for feedback optimization.

Specifically, the output layer: The output layer uses the mean square error loss MSE as the objective optimization function to learn the model parameters, so as to minimize the prediction error between the predicted score and the real score.

The specific formula is as follows:

n represents the number of items. The model parameters are iteratively corrected by the AdamW optimizer. Finally, according to the high and low order of the predicted scores, the set TOP-N recommendation number is selected to recommend users. The detailed flow chart of the specific recommendation is shown in Figure 3.

The present invention proposes a method for recommending a multimedia amblyopia training scheme integrating generalized matrix decomposition and cooperative noise reduction auto-encoder. This method effectively utilizes the collaborative noise reduction auto-encoder CDAE and the generalized matrix factorization GMF model, based on a large number of user training data obtained from the training guided by doctors, and uses the user information combined with the user-training item interaction data randomly set to zero with a certain probability as the Inputting and modeling user preferences with implicit feedback data can effectively alleviate the problem that the system is difficult to accurately extract data features due to the sparsity of user training item interaction data, resulting in insufficient recommendation accuracy.

It first learns the implicit high- and low-order interaction features between users and training items through a collaborative denoising autoencoder and generalized matrix factorization, respectively, and then performs weighted fusion before the output layer. It learns the deep interaction relationship between users and training items more comprehensively, and further improves the recommendation performance. In the present invention, the matrix of users and training items is randomly set to zero to damage the data with corresponding probability values, the damaged data after processing is input into the CDAE model for training, and the users and training items without damage are input into the GMF model for training. Finally, the output results of the two models are weighted and fused with a certain weight to balance the high and low-order features. Combined with the preference information, the prediction score is obtained, and the user's visual training scheme is recommended according to the prediction score. Therefore, the problem of consuming a lot of human resources due to the manual configuration scheme is simplified, intelligent recommendation and user-friendly use are realized, and the training effect is improved.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (4)

1. a kind of amblyopia training scheme recommendation method of fusion GMF and CDAE, it is characterized in that: recommendation system comprises input layer, hidden layer, fusion layer and output layer; Hidden layer comprises generalized matrix decomposition GMF model and collaborative noise reduction automatic encoder CDAE model, the collaborative noise reduction auto-encoder CDAE adopts a three-layer network and two fully connected layers, which use an offset vector and a tanh activation function; The recommendation method based on the recommendation system specifically includes the following steps: Step 1. The input layer obtains the user information data in the database of the vulnerable training platform, and the user information data includes the user's personal basic information, relevant information of amblyopia training items and user-training item interaction information; The data is input into the CDAE model of the collaborative denoising auto-encoder, and another part of the undamaged data is input into the generalized matrix factorization GMF model; Step 2. The collaborative noise reduction autoencoder CDAE model and the generalized matrix factorization GMF model in the hidden layer respectively perform data feature extraction on the input data; Step 3. The fusion layer performs weight fusion on the outputs of the collaborative noise reduction auto-encoder CDAE model and the generalized matrix factorization GMF model in the hidden layer to balance the high and low-order features; the predicted user-training item score is obtained through the fusion layer

Pass it into the output layer for feedback optimization; Step 4. The output layer uses the mean square error loss MSE as the objective optimization function to learn the model parameters, so as to minimize the prediction error between the predicted score and the actual score, and the model parameters are iteratively corrected by the AdamW optimizer; finally, according to the order of the predicted scores, select The set number of TOP-N recommendations is recommended for users. 2. the amblyopia training scheme recommendation method of a kind of fusion GMF and CDAE according to claim 1, is characterized in that in described step 1, the data of input layer input is mapped to describe user and item in latent factor model context. The hidden vector is then used as the input of the generalized matrix factorization GMF model; the same user hidden vector as the generalized matrix factorization GMF model input is combined with the broken user-item interaction data as the input of the collaborative noise reduction autoencoder CDAE model, In the process of data processing, the data that is 1 in the user-training item rating vector is randomly set to 0 to generate damaged data. 3. the amblyopia training scheme recommendation method of a kind of fusion GMF and CDAE according to claim 2, it is characterized in that in described collaborative noise reduction autoencoder CDAE model, the user-training item data from input layer is input into two. The fully connected layer extracts the high-order interaction information of the user-training item; the activation function is introduced into the two fully connected layers to realize the nonlinear process, as follows: First, the processed damage vector in the input layer is mapped to the first layer of the fully connected layer; then the feature representation of the first fully connected layer is mapped to the second fully connected layer to reconstruct the input user-item rating vector. 4. the amblyopia training scheme recommendation method of a kind of fusion GMF and CDAE according to claim 3, it is characterized in that in described generalized matrix factorization GMF model, at first user hidden vector and training item hidden vector corresponding element are multiplied, then The vector obtained after multiplication is multiplied by the weight ^WG and then added, and then the result is input into the activation function for further mapping to introduce a nonlinear process.

Images