CN111564200A - Old people diet feature extraction device and method based on rapid random gradient descent - Google Patents

Abstract

The invention discloses a device and a method for extracting diet characteristics of old people based on rapid random gradient descent, wherein the prediction device comprises a data receiving module, a data processing module and a data processing module, wherein the data receiving module is configured to collect old people-diet scoring data; a data storage module configured to store senior-diet scoring data and senior-diet feature data; the characteristic rapid extraction module is configured to train and extract diet characteristics through known scores of the old people on daily diet and transmit the extracted characteristic data to the characteristic output module; the rapid feature extraction module comprises a data initialization unit, a data division unit and a rapid data training unit. The prediction device and the prediction method provided by the invention execute the data block by block in parallel through the momentum combined with the parallel random gradient descent of the random gradient descent strategy, thereby not only accelerating the convergence speed of the model, but also reducing the training time of the model, and being widely applied to a recommendation platform for providing personalized services.

Description

Old people diet feature extraction device and method based on rapid random gradient descent

Technical Field

The invention relates to the technical field of computer data mining, in particular to a device and a method for extracting diet characteristics of old people based on rapid random gradient descent.

Background

With the development of social science and technology and economy, the population is aging seriously, the burden of the labor age population is increased, the number of the elderly living alone is increased, and the healthy diet management platform for the elderly attracts social attention. The number of the old people is increasing, the nutrition diet collocation types are rich, and the old people or nursing personnel for taking care of the diet of the old people are difficult to select the most suitable diet collocation among a plurality of diet collocations. Therefore, an old-person-diet scoring matrix can be formed by scoring different diets by the old person, the preference degree of the old person to different diets is predicted by utilizing the scoring height, the characteristics of different diets are obtained, and a healthy diet matching package is recommended according to the preference. The daily eating types are more and more, the old people have many diets and do not try, the evaluation is not carried out, the characteristic can be rapidly extracted, and the problem of providing healthy diet matching according to the preference of the old people becomes the concern of each family.

Disclosure of Invention

The invention aims to provide a device and a method for extracting the diet characteristics of old people with rapid random gradient decline, which are used for extracting different diet characteristics of the old people according to the historical evaluation of the old people on daily diet.

In order to solve the above-mentioned technical problem, according to a first aspect of the present invention, there is provided an old people diet feature extraction device based on fast random gradient descent, the device comprising:

the data receiving module is used for acquiring old people-diet scoring data from the old people health management platform and transmitting the acquired data into the data preprocessing module;

the data preprocessing module is used for processing the received data into a data form which can be directly used and transmitting the data into the data storage module;

the data storage module is used for storing the acquired known grading data and storing a hidden feature matrix of the old and the daily diet types;

the characteristic rapid extraction module is used for initializing data, constructing an old people diet characteristic extraction model with rapid random gradient descent and performing data training;

and the characteristic output module outputs the prediction data of the unknown old people-diet scores and the implicit characteristics of the old people and diet.

Further, the rapid feature extraction module comprises a data initialization unit, a model component unit and a rapid data training unit, wherein the data initialization unit is used for initializing relevant parameters related to the old people diet feature extraction process based on rapid random gradient descent; the model construction unit constructs an old people diet characteristic extraction model with rapid random gradient descent, which can be executed in parallel, by using a strategy of combining momentum and random gradient descent; the data rapid training unit is used for training to obtain old people diet characteristic data;

according to the preferable scheme of the old people diet feature extraction device based on the rapid stochastic gradient descent, the data preprocessing module specifically comprises:

the collected data is processed into a triplet form, which may be expressed as P ═ x, y, r, where l denotes geriatric, y denotes diet, and r denotes geriatric x scores diet y. Storing the processed data into a data storage module;

according to the preferable scheme of the old people diet characteristic extraction device based on the rapid random gradient descent, the data storage module comprises a diet grading data storage unit and an old people diet characteristic storage unit.

The diet scoring data storage unit is configured to establish a matrix of | X | rows and | Y | columns as an old people-diet scoring matrix R for storing known scoring data, wherein | X | is a set of old people and | Y | is a set of diet.

The elderly diet feature storage unit is configured to store elderly diet feature data. Storing feature data of old people and diet in a matrix form, and storing the feature data by using U, I two hidden feature matrices, wherein the feature dimensions of the two hidden feature matrices are the same as d.

According to the preferred scheme of the old people diet feature extraction device based on the rapid stochastic gradient descent, the feature rapid extraction module comprises a data initialization unit, a model construction unit and a data rapid training unit.

The data initialization unit is configured to initialize elderly people and mealsTwo hidden feature matrices U and I; initializing a hidden feature space dimension d; initializing a regularization factor λ₂The method comprises the steps of initializing a node number N of parallel calculation, initializing a data division segment number b, initializing a maximum iteration round number T, initializing an iteration round number control variable T in a training process, initializing a momentum acceleration factor gamma, initializing a model training learning step η, and initializing a convergence termination threshold value a, wherein d determines a feature space dimension of each hidden feature matrix and initializes the hidden feature matrix to be a positive integer, in the hidden feature matrices U and I, U is a hidden feature matrix with | L | row d column and I is a hidden feature matrix with | Y | row d column and initializes the hidden feature matrices with random smaller positive numbers respectively, the maximum training iteration round number T is a variable for controlling an upper limit of training times and initializes to be a larger positive integer, the iteration round number control variable T is initialized to be 1, and a regularization factor lambda is normalized₂The method comprises the steps of determining the number of nodes of parallel calculation of a model, initializing the nodes to be positive integers, determining the number b of data division sections, determining the number b of data sections into which a data set is divided, wherein each data section is provided with b data blocks which are independent of each other and do not share a row or a column, initializing the data sections to be positive integers, determining the acceleration effect of momentum on gradient descent, accelerating the convergence speed of the model, initializing the data sections to be small positive numbers, determining the learning step size of model training by the model training learning step size η, initializing the data sections to be small positive numbers, and judging whether the training is finished or not by the convergence termination threshold value a which is a parameter and is initialized by small positive numbers.

The model construction unit is configured to construct an old people diet characteristic extraction model with rapid random gradient descent, and an objective function RLY of the model is constructed according to a known data set Q in an old people diet scoring matrix R_QThe formula is as follows:

wherein Q represents a known scoring data set of daily diet for the elderly in the elderly-daily diet scoring matrix R; r is_x,yThe meaning is x-to-daily drinking for the agedThe preference degree of the food y is the preference score of the old x on the daily food y; u. of_x′Representing a hidden eigenvalue corresponding to the xth old man in the old man hidden eigenvalue matrix U; i.e. i_y′And (4) representing the implicit characteristic value corresponding to the y-th diet in the daily diet implicit characteristic matrix I.

To prevent overfitting of the model, the generalization performance of the model is enhanced, usually in the objective function RLY_QAdding L₂The regularization term constrains the model, which can be given by the following equation:

wherein λ₂Is a regularization term control parameter, λ₂Represents the measure L₂The limiting effect of the regularization term on the model.

The data rapid training unit is configured to combine the target function constructed by the model construction unit, rapidly train data and extract characteristic values. In order to satisfy U and I for accumulating error RLY on set_QAt a minimum, the cumulative error RLY is computed using a fast converging stochastic gradient descent optimization algorithm_QAnd training the matrixes U and I to obtain the global optimal solution of the matrixes U and I. In the training process, the selection of the hidden characteristics of the old and the daily diet types is adopted to quickly update the direction, the linear combination of the current gradient and the last update gradient of the parameters is realized according to the learning step length and the momentum acceleration factor, and the current update u is calculated_x′And i_y′The formula expression of the gradient value corresponding to the fastest convergence speed is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time. The expression is as follows:

selecting the calculated gradient value with the fastest convergence according to the fast convergence direction, and obtaining a parameter updating rule of an old people diet characteristic extraction model based on fast random gradient decline:

about parameter u_x,kThe update rule of (2) is as follows:

with respect to parameter i_y,kThe update rule of (2) is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time.

The following new rule can accelerate the convergence speed of the model, in order to further accelerate the calculation efficiency of the model, the data is divided into b data segments before the model is updated, b blocks are arranged in each data segment, the blocks are irrelevant, and meanwhile, b calculation nodes are used for respectively and simultaneously training b independent blocks, so that the high-efficiency calculation speed is obtained. And executing a parameter updating rule in each block until all data block training is finished, which represents that the iteration is finished.

Repeating the training process on the known data set until the model is converged, wherein the convergence judgment condition is that the iteration number T reaches the maximum training iteration number T or RLY obtained by calculation of the current iteration and the last iteration is_QThe absolute value of the difference in values is less than the convergence termination threshold a.

According to the preferred embodiment of the device for extracting the dietary characteristics of the elderly based on rapid stochastic gradient descent, the data output module comprises:

the output unit outputs a target loss function RLY_QAnd when the minimum two implicit feature matrixes are reached, the unknown score predicting unit predicts the unknown score in the old man-diet matrix by using the implicit feature matrixes obtained by the output unit.

According to a second aspect of the present invention, there is provided a method for extracting dietary characteristics of elderly people with rapid stochastic gradient descent, comprising the following steps:

s1: and receiving old people diet scoring data of the information acquisition terminal, and sending the old people diet scoring data to an old people diet feature extraction device based on rapid random gradient descent. The old people diet scoring data is real scoring data of diets contained in a nutritional package after the old people select the nutritional package matching on the old people health management platform.

S2: for known data, it is stored in the form of a triplet, i.e. P ═ x, y, r, where l denotes geriatric, y denotes diet, and r denotes geriatric x scores diet y.

S3: initializing relevant parameters in the process of extracting the diet characteristics of the old.

S4: and combining the known acquired data and the initialized related parameters to construct an objective loss function.

S5: and updating the hidden characteristics of the old and the diet according to the constructed target loss function, and obtaining two characteristic matrixes when the model reaches convergence.

S6: two feature matrices are output.

According to the preferred embodiment of the method for extracting dietary characteristics of elderly people with fast random gradient descent, step S3 is used for initializing relevant parameters in the process of extracting dietary characteristics of elderly people, and includes:

initializing hidden feature matrixes U and I; initializing a hidden feature space dimension d; initializing a regularization factor λ₂The method comprises the steps of initializing a node number N of parallel calculation, initializing a data division segment number b, initializing a maximum iteration round number T, initializing an iteration round number control variable T in a training process, initializing a momentum acceleration factor gamma, initializing a model training learning step length η, and initializing a convergence termination threshold value a;

wherein:

the hidden feature dimension d determines the spatial dimensions of the feature matrices U and I, and is initialized to a positive integer.

Size of hidden feature matrices U and I: the implicit feature matrix with U being | L | row and d column and the implicit feature matrix with I being | Y | row and d column are initialized by random small positive numbers.

The maximum training iteration round number T is a variable for controlling the upper limit of the training times and is initialized to be a larger positive integer. Initializing an iteration round number control variable t to 1;

regularization factor lambda₂The method is used for measuring the limiting effect of the regularization term on the model and is initialized to be a smaller positive number.

Determining the number of nodes of the model parallel computation by the number N of the nodes of the parallel computation, and initializing the number of the nodes of the model parallel computation into a positive integer;

the data dividing segment number b is used for determining that a data set is divided into b data segments, wherein b data blocks are arranged in each data segment, the b data blocks are independent from each other, do not share rows or columns, and are initialized to be positive integers;

the momentum acceleration factor gamma is used for determining the acceleration effect of momentum on gradient reduction, so that the convergence speed of the model is accelerated and the model is initialized to be a smaller positive number;

the learning step length eta of the model training determines the learning step length of the model training and is initialized to a smaller positive number;

the convergence termination threshold a is a parameter for determining whether or not training is terminated, and is initialized with a very small positive number.

According to the preferred embodiment of the method for extracting the dietary characteristics of the elderly with fast stochastic gradient descent, step S4 combines the known acquired data and the initialized relevant parameters to construct an objective loss function, which includes:

constructing an objective function RLY of the model according to a known data set Q in the old people-diet scoring matrix R_QThe formula is as follows:

wherein Q represents a known scoring data set of daily diet for the elderly in the elderly-daily diet scoring matrix R; r is_x,yThe meaning is that the preference degree of the old people x to the daily diet y is the preference score of the old people x to the daily diet y; u. of_x′Representing a hidden eigenvalue corresponding to the xth old man in the old man hidden eigenvalue matrix U; i.e. i_y′And (4) representing the implicit characteristic value corresponding to the y-th diet in the daily diet implicit characteristic matrix I.

To prevent overfitting of the model, the generalization performance of the model is enhanced, usually in the objective function RLY_QAdding L₂The regularization term constrains the model, which can be given by the following equation:

wherein λ₂Is a regularization term control parameter, λ₂Represents the measure L₂The limiting effect of the regularization term on the model.

According to the preferred scheme of the old people diet characteristic extraction method with the rapid random gradient descent, step S5 is used for solving an objective function extraction characteristic matrix. The method comprises the following steps:

and S51, dividing the data into b data segments, wherein each data segment contains b data blocks and the b data blocks are not related to each other. And selecting one data segment every time, distributing the b data blocks to different computing nodes until all data are distributed and trained, and finishing one iteration.

And S52, different nodes acquire data blocks which are not related to each other, and simultaneously train and optimize the target loss function. In order to satisfy U and I for accumulating error RLY on set_QAt a minimum, the cumulative error RLY is computed using a fast converging stochastic gradient descent optimization algorithm_QAnd training the matrixes U and I to obtain the global optimal solution of the matrixes U and I. In the training process, the selection of the hidden characteristics of the old and the daily diet types is adopted to quickly update the direction, the linear combination of the current gradient and the last update gradient of the parameters is realized according to the learning step length and the momentum acceleration factor, and the current update u is calculated_x′And i_y′The formula expression of the gradient value corresponding to the fastest convergence speed is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time. The expression is as follows:

selecting the calculated gradient value with the fastest convergence according to the fast convergence direction, and obtaining a parameter updating rule of an old people diet characteristic extraction model based on fast random gradient decline:

about parameter u_x,kThe update rule of (2) is as follows:

with respect to parameter i_y,kThe update rule of (2) is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time.

And S53, judging whether the training of all the data blocks is finished, if the data is not completely processed in the current iteration, returning to S51, and acquiring the unassigned data blocks to continue the training. If the training is completed, the process proceeds to S54.

S54, judgmentJudging whether the iteration end condition is met, wherein the convergence judgment condition is that the iteration round number T reaches the maximum training iteration round number T or RLY obtained by the calculation of the current iteration and the last iteration_QThe absolute value of the difference in values is less than the convergence termination threshold a.

The invention discloses an old people diet feature extraction device and method based on rapid random gradient descent, and aims to obtain a feature matrix of old people diet by utilizing a fast-calculated reduction matrix factorization method, and provide healthy and nutritional diet matching service which accords with the favor of the old people for the old people.

Drawings

Fig. 1 is a prediction device for elderly people diet feature extraction based on rapid stochastic gradient descent in the invention.

Fig. 2 is a schematic flow chart of the process for extracting and updating the dietary characteristics of the elderly based on rapid stochastic gradient descent.

Detailed Description

The present invention will be described in further detail with reference to test examples and specific embodiments. It should be understood that the scope of the above-described subject matter is not limited to the following examples, and any techniques implemented based on the disclosure of the present invention are within the scope of the present invention.

Example 1:

referring to fig. 1, there is shown an apparatus for senior diet feature extraction based on fast stochastic gradient descent of the present invention, comprising:

the data receiving module 210 acquires elderly-eating scoring data from the elderly health management platform and transmits the acquired data to the data preprocessing module;

the data preprocessing module 220 is used for processing the received data into a data form which can be directly used and transmitting the data into the data storage module;

the data storage module 230 is used for storing the acquired known grading data and storing a hidden feature matrix of the old and the daily diet types;

the characteristic fast extraction module 240 is used for initializing data, constructing an old people diet characteristic extraction model with fast random gradient descent and performing data training;

and the characteristic output module 250 outputs the prediction data of the unknown old people-diet score and the implicit characteristics of the old people and diet.

Further, the rapid feature extraction module comprises a data initialization unit 241, a model component unit 242 and a data rapid training unit 243, wherein the data initialization unit 241 is used for initializing relevant parameters involved in the elderly diet feature extraction process based on rapid stochastic gradient descent; the model construction unit 242 constructs an old people diet feature extraction model with rapid random gradient descent, which can be executed in parallel, by using a strategy of combining momentum and random gradient descent; the data fast training unit 243 is used for training to obtain old people diet characteristic data;

in a specific embodiment, the data preprocessing module 220 specifically includes:

the collected data is processed into a triplet form, which may be expressed as P ═ x, y, r, where l denotes geriatric, y denotes diet, and r denotes geriatric x scores diet y. Storing the processed data into a data storage module;

in a specific embodiment, the data storage module 230 includes a diet scoring data storage unit 231 and an elderly diet feature storage unit 232.

The diet scoring data storage unit 231 is configured to establish a matrix of | X | rows and | Y | columns as the senior-diet scoring matrix R, and is configured to store known scoring data, where | X | is a set of senior citizens and | Y | is a set of diets.

The elderly diet feature storage unit 232 is used for storing elderly diet feature data. Storing feature data of old people and diet in a matrix form, and storing the feature data by using U, I two hidden feature matrices, wherein the feature dimensions of the two hidden feature matrices are the same as d.

In a specific embodiment, the feature fast extraction module 240 includes a data initialization unit 241, a model construction unit 242, and a data fast training unit 243.

The data initialization unit 241 is used for initializing two hidden feature matrixes U and I of the old and the diet; initializing a hidden feature space dimension d; initializing a regularization factor λ₂The method comprises the steps of initializing a node number N of parallel calculation, initializing a data division segment number b, initializing a maximum iteration round number T, initializing an iteration round number control variable T in a training process, initializing a momentum acceleration factor gamma, initializing a model training learning step η, and initializing a convergence termination threshold value a, wherein d determines a feature space dimension of each hidden feature matrix and initializes the hidden feature matrix to be a positive integer, in the hidden feature matrices U and I, U is a hidden feature matrix with | L | row d column and I is a hidden feature matrix with | Y | row d column and initializes the hidden feature matrices with random smaller positive numbers respectively, the maximum training iteration round number T is a variable for controlling an upper limit of training times and initializes to be a larger positive integer, the iteration round number control variable T is initialized to be 1, and a regularization factor lambda is normalized₂The method comprises the steps of determining the number of nodes of parallel calculation of a model, initializing the nodes to be positive integers, determining the number b of data division sections, determining the number b of data sections into which a data set is divided, wherein each data section is provided with b data blocks which are independent of each other and do not share a row or a column, initializing the data sections to be positive integers, determining the acceleration effect of momentum on gradient descent, accelerating the convergence speed of the model, initializing the data sections to be small positive numbers, determining the learning step size of model training by the model training learning step size η, initializing the data sections to be small positive numbers, and judging whether the training is finished or not by the convergence termination threshold value a which is a parameter and is initialized by small positive numbers.

The model construction unit 242 is used for constructing an old people diet feature extraction model with fast random gradient descent, and constructing a target function RLY of the model according to a known data set Q in an old people-diet scoring matrix R_QThe formula is as follows:

wherein Q represents a known scoring data set of daily diet for the elderly in the elderly-daily diet scoring matrix R; r is_x,yThe meaning is that the preference degree of the old people x to the daily diet y is the preference score of the old people x to the daily diet y; u. of_x′Representing hidden characteristic moments of the agedHidden characteristic values corresponding to the x-th old man in the array U; i.e. i_y′And (4) representing the implicit characteristic value corresponding to the y-th diet in the daily diet implicit characteristic matrix I.

To prevent overfitting of the model, the generalization performance of the model is enhanced, usually in the objective function RLY_QAdding L₂Regularization constrains the model, which can be given by the following equation:

wherein λ₂Is a regularization control parameter, λ₂Represents the measure L₂Regularization limiting the effect on the model.

The data fast training unit 243 is configured to fast train data and extract feature values in combination with the objective function constructed by the model construction unit. In order to satisfy U and I for accumulating error RLY on set_QAt a minimum, the cumulative error RLY is computed using a fast converging stochastic gradient descent optimization algorithm_QAnd training the matrixes U and I to obtain the global optimal solution of the matrixes U and I. In the training process, the selection of the hidden characteristics of the old and the daily diet types is adopted to quickly update the direction, the linear combination of the current gradient and the last update gradient of the parameters is realized according to the learning step length and the momentum acceleration factor, and the current update u is calculated_x′And i_y′The formula expression of the gradient value corresponding to the fastest convergence speed is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e. the kth implicit of geriatric x and diet yAnd (5) characterizing.

And

respectively, the current gradient value of the t time. The expression is as follows:

selecting the calculated gradient value with the fastest convergence according to the fast convergence direction, and obtaining a parameter updating rule of an old people diet characteristic extraction model based on fast random gradient decline:

about parameter u_x,kThe update rule of (2) is as follows:

with respect to parameter i_y,kThe update rule of (2) is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time.

The following new rule can accelerate the convergence speed of the model, in order to further accelerate the calculation efficiency of the model, the data is divided into b data segments before the model is updated, b blocks are arranged in each data segment, the blocks are irrelevant, and meanwhile, b calculation nodes are used for respectively and simultaneously training b independent blocks, so that the high-efficiency calculation speed is obtained. And executing a parameter updating rule in each block until all the data blocks are trained to end, which represents the end of the iteration.

Repeating the training process on the known data set until the model is converged, wherein the convergence judgment condition is that the iteration number T reaches the maximum training iteration number T or RLY obtained by calculation of the current iteration and the last iteration is_QThe absolute value of the difference in values is less than the convergence termination threshold a.

In a specific embodiment, the data output module 250 specifically includes:

output target loss function RLY_QAnd when the minimum two implicit feature matrixes are reached, the unknown score predicting unit predicts the unknown score in the old man-diet matrix by using the implicit feature matrixes obtained by the output unit.

Example 2:

referring to fig. 2, fig. 2 shows the method for extracting dietary characteristics of the elderly based on rapid stochastic gradient descent according to the present invention, wherein the prediction method comprises the following steps:

s1: and receiving old people diet scoring data of the information acquisition terminal, and sending the old people diet scoring data to an old people diet feature extraction device based on rapid random gradient descent. The old people diet scoring data is real scoring data of diets contained in a nutritional package after the old people select the nutritional package matching on the old people health management platform.

S2: for known data, it is stored in the form of a triplet, i.e. P ═ x, y, r, where l denotes geriatric, y denotes diet, and r denotes geriatric x scores diet y.

S3: initializing relevant parameters in the process of extracting the diet characteristics of the old.

S4: and combining the known acquired data and the initialized related parameters to construct an objective loss function.

S5: and updating the hidden characteristics of the old and the diet according to the constructed target loss function, and obtaining two characteristic matrixes when the model reaches convergence.

S6: two feature matrices are output.

In a specific embodiment, step S3 is used to initialize relevant parameters in the process of extracting dietary features of the elderly, including:

initializing hidden feature matrixes U and I; initializing a hidden feature space dimension d; initializing a regularization factor λ₂The method comprises the steps of initializing a node number N of parallel calculation, initializing a data division segment number b, initializing a maximum iteration round number T, initializing an iteration round number control variable T in a training process, initializing a momentum acceleration factor gamma, initializing a model training learning step length η, and initializing a convergence termination threshold value a;

wherein:

the hidden feature dimension d determines the spatial dimensions of the feature matrices U and I, and is initialized to a positive integer.

Size of hidden feature matrices U and I: the implicit feature matrix with U being | L | row and d column and the implicit feature matrix with I being | Y | row and d column are initialized by random small positive numbers.

The maximum training iteration round number T is a variable for controlling the upper limit of the training times and is initialized to be a larger positive integer. Initializing an iteration round number control variable t to 1;

regularization factor lambda₂The method is used for measuring the limiting effect of regularization on the model and is initialized to be a small positive number. Determining the number of nodes of the model parallel computation by the number N of the nodes of the parallel computation, and initializing the number of the nodes of the model parallel computation into a positive integer;

the data dividing segment number b is used for determining that a data set is divided into b data segments, wherein b data blocks are arranged in each data segment, and the b data blocks are independent from each other, namely, the data segments do not share rows or columns and are initialized to be positive integers;

the momentum acceleration factor gamma is used for determining the acceleration effect of momentum on gradient reduction, so that the convergence speed of the model is accelerated and the model is initialized to be a smaller positive number;

the learning step length eta of the model training determines the learning step length of the model training and is initialized to a smaller positive number;

the convergence termination threshold a is a parameter for determining whether or not training is terminated, and is initialized with a very small positive number.

In a specific embodiment, step S4 combines the known acquired data and initialized relevant parameters to construct an objective loss function, which includes:

constructing an objective function RLY of the model according to a known data set Q in the old people-diet scoring matrix R_QThe formula is as follows:

wherein Q represents a known scoring data set of daily diet for the elderly in the elderly-daily diet scoring matrix R; r is_x,yThe meaning is that the preference degree of the old people x to the daily diet y is the preference score of the old people x to the daily diet y; u. of_x′Representing a hidden eigenvalue corresponding to the xth old man in the old man hidden eigenvalue matrix U; i.e. i_y′And (4) representing the implicit characteristic value corresponding to the y-th diet in the daily diet implicit characteristic matrix I.

To prevent overfitting of the model, the generalization performance of the model is enhanced, usually in the objective function RLY_QAdding L₂Regularization constrains the model, which can be given by the following equation:

wherein λ₂Is a regularization control parameter, λ₂Represents the measure L₂Regularization limiting the effect on the model.

In a specific embodiment, step S5 is implemented to solve the objective function extraction feature matrix, and includes:

and S51, dividing the data into b data segments, wherein each data segment contains b data blocks and the b data blocks are not related to each other. And selecting one data segment every time, distributing the b data blocks to different computing nodes until all data are distributed and trained, and finishing one iteration.

And S52, different nodes acquire data blocks which are not related to each other, and simultaneously train and optimize the target loss function. In order to satisfy U and I for accumulating error RLY on set_QAt a minimum, the cumulative error RLY is computed using a fast converging stochastic gradient descent optimization algorithm_QAnd training the matrixes U and I to obtain the global optimal solution of the matrixes U and I. In the training process, the selection of the hidden characteristics of the old and the daily diet types is adopted to quickly update the direction, the linear combination of the current gradient and the last update gradient of the parameters is realized according to the learning step length and the momentum acceleration factor, and the current update u is calculated_x′And i_y′The formula expression of the gradient value corresponding to the fastest convergence speed is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time. The expression is as follows:

selecting the calculated gradient value with the fastest convergence according to the direction of the rapid convergence, and obtaining a parameter updating rule of an old diet characteristic extraction model based on rapid random gradient decline:

about parameter u_x,kThe update rule of (2) is as follows:

with respect to parameter i_y,kThe update rule of (2) is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time.

And S53, judging whether the training of all the data blocks is finished, if the data is not completely processed in the current iteration, returning to S51, and acquiring the unassigned data blocks to continue the training. If the training is completed, the process proceeds to S54.

S54, judging whether the iteration end condition is satisfied, wherein the convergence judgment condition is that the iteration round number T reaches the maximum training iteration round number T or RLY obtained by the calculation of the current iteration and the last iteration_QThe absolute value of the difference in values is less than the convergence termination threshold a.

While embodiments of the invention have been shown and described, it will be understood by those of ordinary skill in the art that: various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims (7)

1. The utility model provides an old man's diet feature extraction element based on quick random gradient descends, its characterized in that includes:

the data receiving module is used for acquiring old people-diet scoring data from the old people health management platform and transmitting the acquired data into the data preprocessing module;

the data preprocessing module is used for processing the received data into a data form which can be directly used and transmitting the data into the data storage module;

the data storage module is used for storing the acquired known grading data and storing a hidden feature matrix of the old and the daily diet types;

the characteristic rapid extraction module is used for initializing data, constructing an old people diet characteristic extraction model with rapid random gradient descent and performing data training;

and the characteristic output module outputs the prediction data of the unknown old people-diet scores and the implicit characteristics of the old people and diet.

2. The elderly diet feature extraction device based on fast stochastic gradient descent according to claim 1, wherein the data storage module comprises a diet score data storage unit and an elderly diet feature storage unit,

the diet scoring data storage unit is configured to: and establishing a matrix of | X | rows and | Y | columns as an old people-diet scoring matrix R for storing known scoring data, wherein | X | is a set of old people and | Y | is a set of diet.

The elderly diet feature storage unit is configured to: storing the diet characteristic data of the old. Storing feature data of old people and diet in a matrix form, and storing the feature data by using U, I two hidden feature matrices, wherein the feature dimensions of the two hidden feature matrices are the same as d.

3. The old people diet feature extraction device based on rapid stochastic gradient descent according to claim 2, wherein the feature rapid extraction module comprises a data initialization unit, a model construction unit and a data rapid training unit,

the data initialization unit is configured to: initializing two implicit feature matrixes U and I of the old and diet; initializing a hidden feature space dimension d; initializing a regularization factor λ₂The method comprises the steps of initializing the number N of nodes calculated in parallel, initializing the number b of data division sections, initializing the maximum iteration round number T, initializing an iteration round number control variable T in the training process, initializing a momentum acceleration factor gamma, initializing a model training learning step η, initializing a convergence termination threshold value a, wherein d determines the feature space dimension of each hidden feature matrix, in the hidden feature matrices U and I, U is a hidden feature matrix with L and Y rows and d columns, I is a hidden feature matrix with Y and d columns, the maximum training iteration round number T is a variable for controlling the upper limit of the training times, the iteration round number control variable T and a regularization factor lambda₂The method comprises the steps of measuring the limiting effect of a regularization item on a model, determining the number of nodes of the parallel calculation of the model by the number N of nodes of the parallel calculation of the regularization item, determining the number b of data division sections, determining that a data set is divided into b data sections, wherein each data section is provided with b data blocks which are independent of each other and do not share a row or a column, determining the acceleration effect of momentum on gradient reduction to accelerate the convergence speed of the model, determining the learning step size of model training by the learning step size η of the model training, and judging whether the training is finished or not by the convergence termination threshold value a.

The model building unit is configured to: constructing an old people diet characteristic extraction model with rapid random gradient descent, and constructing a target function RLY of the model according to a known data set Q in an old people-diet scoring matrix R_QThe formula is as follows:

wherein Q represents a known scoring data set of daily diet for the elderly in the elderly-daily diet scoring matrix R; r is_x,yThe meaning is that the x-to-daily diet y preference degree of the old is the x-to-daily diet y preference degree of the oldPreference score for diet y; u. of_x′Representing a hidden eigenvalue corresponding to the xth old man in the old man hidden eigenvalue matrix U; i.e. i_y′And (4) representing the implicit characteristic value corresponding to the y-th diet in the daily diet implicit characteristic matrix I.

To prevent overfitting of the model, the generalization performance of the model is enhanced, usually in the objective function RLY_QAdding L₂The regularization term constrains the model, which can be given by the following equation:

wherein λ₂Is a regularization term control parameter, λ₂Represents the measure L₂The limiting effect of the regularization term on the model.

The data fast training unit is configured to: and combining the target function constructed by the model construction unit, quickly training data and extracting characteristic values. In order to satisfy U and I for accumulating error RLY on set_QAt a minimum, the cumulative error RLY is computed using a fast converging stochastic gradient descent optimization algorithm_QAnd training the matrixes U and I to obtain the global optimal solution of the matrixes U and I. In the training process, the selection of the hidden characteristics of the old and the daily diet types is adopted to quickly update the direction, the linear combination of the current gradient and the last update gradient of the parameters is realized according to the learning step length and the momentum acceleration factor, and the current update u is calculated_x′And i_y′The formula expression of the gradient value corresponding to the fastest convergence speed is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time. The expression is as follows:

selecting the calculated gradient value with the fastest convergence according to the fast convergence direction, and obtaining a parameter updating rule of an old people diet characteristic extraction model based on fast random gradient decline:

about parameter u_x,kThe update rule of (2) is as follows:

with respect to parameter i_y,kThe update rule of (2) is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time.

The following new rule can accelerate the convergence speed of the model, in order to further accelerate the calculation efficiency of the model, the data is divided into b data segments before the model is updated, b blocks are arranged in each data segment, the blocks are irrelevant, and meanwhile, b calculation nodes are used for respectively and simultaneously training b independent blocks, so that the high-efficiency calculation speed is obtained. And executing a parameter updating rule in each block until all data block training is finished, which represents that the iteration is finished.

Repeating the training process on the known data set until the model is converged, wherein the convergence judgment condition is that the iteration number T reaches the maximum training iteration number T or RLY obtained by calculation of the current iteration and the last iteration is_QThe absolute value of the difference in values is less than the convergence termination threshold a.

4. A method for extracting dietary features of old people with fast random gradient descent according to any one of claims 1 to 3, comprising the following steps:

s1: receiving old people diet scoring data of an information acquisition terminal, and establishing an old people-diet preference matrix;

s2: storing known data in a form of a triple, namely P ═ x, y, r, wherein l represents geriatric, y represents diet, and r represents geriatric x scores diet y;

s3: initializing relevant parameters in the process of extracting the diet characteristics of the old;

s4: constructing a target loss function by combining known acquired data and initialized related parameters;

s5: updating hidden characteristics of old people and diet according to the constructed target loss function, and obtaining two characteristic matrixes when the model reaches convergence;

s6: two feature matrices are output.

5. The method for elder diet feature extraction with fast stochastic gradient descent according to claim 4, wherein the step S3 comprises:

initializing hidden feature matrixes U and I; initializing a hidden feature space dimension d; initializing a regularization factor λ₂The method comprises the steps of initializing the number N of nodes in parallel calculation, initializing the number b of data division sections, initializing the maximum iteration round number T, initializing an iteration round number control variable T in the training process, initializing a momentum acceleration factor gamma, initializing a model training learning step η and initializing a convergence termination threshold a.

6. The method for elder diet feature extraction with fast stochastic gradient descent according to claim 5, wherein the step S4 comprises:

constructing an objective function RLY of the model according to a known data set Q in the old people-diet scoring matrix R_QThe formula is as follows:

wherein Q represents a known scoring data set of daily diet for the elderly in the elderly-daily diet scoring matrix R; r is_x,yThe meaning is that the preference degree of the old people x to the daily diet y is the preference score of the old people x to the daily diet y; u. of_x′Representing a hidden eigenvalue corresponding to the xth old man in the old man hidden eigenvalue matrix U; i.e. i_y′And (4) representing the implicit characteristic value corresponding to the y-th diet in the daily diet implicit characteristic matrix I.

To prevent overfitting of the model, the generalization performance of the model is enhanced, usually in the objective function RLY_QAdding L₂The regularization term constrains the model, which can be given by the following equation:

wherein λ₂Is a regularization term control parameter, λ₂Represents the measure L₂The limiting effect of the regularization term on the model.

7. The method for elder diet feature extraction with fast stochastic gradient descent according to claim 6, wherein the step S5 comprises:

and S51, dividing the data into b data segments, wherein each data segment contains b data blocks and the b data blocks are not related to each other. And selecting one data segment every time, distributing the b data blocks to different computing nodes until all data are distributed and trained, and finishing one iteration.

And S52, different nodes acquire data blocks which are not related to each other, and simultaneously train and optimize the target loss function. In order to satisfy U and I for accumulating error RLY on set_QAt a minimum, the cumulative error RLY is computed using a fast converging stochastic gradient descent optimization algorithm_QAnd training the matrixes U and I to obtain the global optimal solution of the matrixes U and I. In the training process, the selection of the hidden characteristics of the old and the daily diet types is adopted to quickly update the direction, the linear combination of the current gradient and the last update gradient of the parameters is realized according to the learning step length and the momentum acceleration factor, and the current update u is calculated_x′And i_y′The formula expression of the gradient value corresponding to the fastest convergence speed is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time. The expression is as follows:

selecting the calculated gradient value with the fastest convergence according to the fast convergence direction, and obtaining a parameter updating rule of an old people diet characteristic extraction model based on fast random gradient decline:

about parameter u_x,kThe update rule of (2) is as follows:

with respect to parameter i_y,kThe update rule of (2) is as follows:

where gamma is the momentum acceleration factor, η is the parameter training step length, parameter w^(t-1)And k^(t-1)Are respectively a parameter u_x,kAnd i_y,kThe update gradient size at the t-th update is respectively w^(t-1)And k^(t-1)And the corresponding decision parameter consists of two parts of an updated gradient and a current gradient accumulated value at t-1 times. Wherein u is_x,kAnd i_y,kAre each u_x′And i_y′The kth element of the vector, i.e., the kth implicit feature of geriatric x and diet y.

And

respectively, the current gradient value of the t time.

And S53, judging whether the training of all the data blocks is finished, if the data is not completely processed in the current iteration, returning to S51, and acquiring the unassigned data blocks to continue the training. If the training is completed, the process proceeds to S54.

S54, judging whether the iteration end condition is satisfied, wherein the convergence judgment condition is that the iteration round number T reaches the maximum training iteration round number T or RLY obtained by the calculation of the current iteration and the last iteration_QThe absolute value of the difference in values is less than the convergence termination threshold a.

Images