Big data prediction method based on matrix decomposition technology

Abstract

The invention discloses a big data prediction method based on a matrix decomposition technology, and the method comprises the steps: inputting a corresponding time sequence dependence factor matrix into an echo state network, after obtaining a corresponding data matrix, carrying out the matrix decomposition of each data matrix, and carrying out the prediction of the time sequence dependence factor matrix. The method further comprises the steps of determining a data matrix with different data orders of magnitude or dimensions as a target matrix, and performing normalization operation on each data in the target matrix; obtaining a sequence data set, and respectively combining the sequence data of each user in the sequence data set to obtain respectively corresponding data matrixes; performing matrix decomposition on each data matrix to obtain a corresponding time sequence factor matrix and a time sequence dependence factor matrix; and an echo state network is introduced, so that the original capability of capturing time sequence modeling and multi-scale structures in matrix decomposition is enhanced, the modeling capability of complex dynamics is enhanced, and the prediction precision is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of big data prediction methods based on matrix decomposition technology, and specifically to a big data prediction method based on matrix decomposition technology. Background Art

[0002] With the development of network technology, users are lost in the vast amount of information. It is difficult to select the information they really need, which reduces the efficiency of information use. This is the so-called information overload problem. To solve the problem of information overload, the recommendation system came into being. The recommendation system is a process of recommending items of interest to users based on their information needs. This is a personalized recommendation process. The personalized recommendation system is based on the interests of the target users. It does not require users to enter keywords. It actively recommends some things that users may be interested in based on some of their previous behaviors, such as browsing history, shopping cart additions, and purchase history, to help users freely navigate in the ocean of information.

[0003] The existing matrix decomposition prediction methods are all centralized training methods. Specifically, a recommendation system platform is first built to collect the QoS values ​​of all users for all Web services, and then use these data to build a matrix decomposition model. The existing collaborative filtering algorithms can already achieve a high level of recommendation accuracy, but improving the time efficiency on this basis is still a problem to be solved for the existing recommendation algorithms. How to improve the time efficiency of the collaborative filtering algorithm while maintaining a high accuracy is already a big problem. Based on this, a big data prediction method based on matrix decomposition technology is proposed to solve the above problem. Summary of the invention

[0004] The purpose of the present invention is to provide a big data prediction method based on matrix decomposition technology to solve the problems raised in the above background technology.

[0005] By adopting the above technical solution, the big data prediction method based on matrix decomposition technology includes:

[0006] S1: Establish an attribute information neural network with a dimension transformation intermediate layer;

[0007] S2: Collect data and information; initialize the weights of each neural network, and calculate the function output value of each node according to the function;

[0008] S3: classify the data samples and build a decomposition model of the prediction matrix;

[0009] S4: Calculate the predicted output value of the neural network, and compare and analyze the predicted output value of the neural network with the expected output value of the neural network to calculate the network prediction error;

[0010] S5: Modify each network weight and each parameter of the function in the neural network prediction model to achieve the prediction expectation;

[0011] S6: Perform parameter iteration. When the network prediction error reaches a specified smaller value or reaches a specified number of iterations, the iteration ends and the prediction result is obtained.

[0012] According to the above technical solution, in S1, an attribute information neural network with a dimensional transformation intermediate layer is constructed. The network is composed of several layers. There are at least two dimensional transformation layers belonging to the input information between the bottom layer and the top layer. The dimensional transformation of the input information is performed by setting the hyperparameters of the weights and biases of each layer of the network to obtain the dimensional transformation result of the input information in the network. The number of layers required for the network structure varies with the scale of the input information. The network structure is: X is the input information involved in the prediction, and is also the input information of the attribute information neural network; taking the dimensional transformation intermediate layer as two layers as an example, Y is the network output result of the dimensional transformation of X after the two-layer network:

[0013] Y=Relu(X*W 1 +B 1 )*W 2 +B 2 ;

[0014] Where W1 and W2 are the weights of each layer in the intermediate layer of dimensional transformation, B1 and B2 are the biases of each layer in the intermediate layer of dimensional transformation, and Relu is the activation function.

[0015] According to the above technical solution, in S2, the sample data information is input into the constructed network to obtain the corresponding dimension transformation result, collect data and information, integrate the sample data into a data set, and use 80% of the data set as a training set and 20% as a test set.

[0016] According to the above technical solution, in S3, the decomposition model of the prediction matrix includes an input layer at the bottom layer, a hidden layer at the middle layer, and an output layer at the top layer, the input layer is composed of a plurality of input nodes, each of which is used to accept and process input variables in the training sample set, the hidden layer is composed of a plurality of hidden nodes, which are used for linear transformation of nonlinear samples, and the hidden layer nodes use wavelet functions as transfer functions; the output layer is composed of a plurality of output nodes, each of which is used to output variable prediction results.

[0017] According to the above technical solution, in S3, the calculation formula of the decomposition model of the prediction matrix is:

[0018] f=clip(U+R_TR*Z. / RN+UA)*(I+IR. / IAN) T ;

[0019] Clip() sets the lower limit of each element of the matrix to 1 and the upper limit to 5; A is the average parameter, B is the deviation parameter, U is the user feature parameter, I is the feature parameter, and Z is the historical feature parameter, all of which are matrices; TR is the prediction result of the training set, the rows represent the training samples, the columns represent the test samples, and each element of the matrix represents the prediction result of big data based on matrix decomposition technology; R_TR is the record matrix of the training set, which sets the result records in the training set scoring matrix TR to 1 and the result records without result records to 0; RN is the module value matrix of each row of the training set result record, each row of which is the square root of the module value of the corresponding row of RTR, which needs to be divided by dot, and the size is the same as R_TR*Z; IAN is the module value matrix of each row of the item attribute, each row of which is the square root of the module value of the corresponding row of the attribute matrix, which needs to be divided by dot, and the size is the same as IA.

[0020] According to the above technical solution, in S4 and S5, the data samples are classified and divided into two parts, one part is a training sample for training the wavelet neural network, and the other part is a test sample for testing the prediction accuracy of the wavelet neural network. According to the network weights in the wavelet neural network prediction model and the function output value of each hidden node in the hidden layer, the wavelet neural network prediction output value is calculated, and the wavelet neural network prediction output value is compared and analyzed with the wavelet neural network expected output value to calculate the network prediction error.

[0021] According to the above technical solution, in S6, when there is an inconsistency in the order of magnitude or dimension of the data in the decomposition model matrix of the prediction matrix, after obtaining the corresponding data matrices and before performing matrix decomposition on each data matrix, it also includes: determining the data matrix with inconsistent order of magnitude or dimension as the target matrix, and performing normalization operation on each data in the target matrix; obtaining a sequence data set, and combining the sequence data of each user in the sequence data set respectively to obtain the corresponding data matrices; performing matrix decomposition on each data matrix to obtain the corresponding time series factor matrix and the timing dependency factor matrix; inputting the corresponding timing dependency factor matrix into the echo state network respectively, and calculating the corresponding next timing dependency factor; and predicting the corresponding next data value by performing matrix operations on the corresponding time series factor matrix and the next timing dependency factor respectively.

[0022] According to the above technical scheme, each data matrix is ​​decomposed separately, including: using non-negative matrix decomposition algorithm to decompose each data matrix separately, and after predicting the corresponding next prediction result, it also includes: obtaining the real data value of each next one; using each real data value and the corresponding prediction value to test the validity of the prediction model used in the prediction process; by using matrix decomposition technology to obtain the corresponding time series factor matrix and time series dependency factor matrix, the time series invariant characteristics and time-varying dependency factors are captured. The introduction of echo state network enhances the original capture ability of matrix decomposition for time series modeling and multi-scale structure, enhances the modeling ability of complex dynamics, and greatly improves the prediction accuracy of big data based on matrix decomposition technology.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. A big data prediction method based on matrix decomposition technology obtains a sequence data set and combines the sequence data of each user in the sequence data set to obtain the corresponding data matrices: performs matrix decomposition on each data matrix to obtain the corresponding time series factor matrix and time series dependency factor matrix, and inputs the corresponding time series dependency factor matrix into the echo state network. After obtaining the corresponding data matrices and before performing matrix decomposition on each data matrix, the method further includes: determining a data matrix with inconsistent order of magnitude or dimension as a target matrix, and normalizing each data in the target matrix; obtaining a sequence data set and combining the sequence data of each user in the sequence data set to obtain the corresponding data matrices; performing matrix decomposition on each data matrix to obtain the corresponding time series factor matrix and time series dependency factor matrix; introducing the echo state network, which enhances the original matrix decomposition ability to capture time series modeling and multi-scale structure, enhances the modeling ability of complex dynamics, and greatly improves the prediction accuracy.

[0025] 2. A big data prediction method based on matrix decomposition technology obtains the respective dimensional transformation results by inputting information into an attribute information network with a dimensional transformation intermediate layer, and combines it with matrix decomposition to propose a more complete prediction model, which makes the prediction score more accurate; when calculating formulas for different evaluation indicators, a single loss function calculation formula is no longer used; compared with the prior art method of sequentially performing operations on a single element, the model of the present invention obtains the prediction score through matrix operations, and then converts it into modeling of complex dynamics, which greatly improves the prediction accuracy, and selects adaptive moment estimation as the parameter optimization method, so that the efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the process structure of the present invention;

[0027] Figure 2 It is a gradient schematic diagram of the present invention;

[0028] Figure 3 It is a schematic diagram of the calculation of the model of the present invention;

[0029] Figure 4 Schematic diagram of the model structure of different potential vector dimensions of the present invention;

[0030] Figure 5 Schematic diagram of the performance comparison between the deep matrix decomposition of the present invention and the benchmark method. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] Embodiment 1:

[0033] The big data prediction method based on matrix decomposition technology provided by the present invention is preferably implemented as follows: Figures 1 to 3 As shown: A big data prediction method based on matrix decomposition technology is characterized by including:

[0034] S1: Establish an attribute information neural network with a dimensional transformation intermediate layer; construct an attribute information neural network with a dimensional transformation intermediate layer, the network consists of several layers, there are at least two layers of dimensional transformation layers belonging to the input information between the bottom layer and the top layer, the dimensional transformation of the input information is performed by setting the hyperparameters of the weights and biases of each layer of the network, and the dimensional transformation result of the input information in the network is obtained. The number of layers required for the network structure varies with the scale of the input information. The network structure is: X is the input information involved in the prediction, and is also the input information of the attribute information neural network; taking the dimensional transformation intermediate layer as an example with two layers, Y is the network output result of the dimensional transformation of X after the two-layer network:

[0035] Y = Relu(X*W 1 +B 1 )*W 2 +B 2 ;

[0036] Where W1 and W2 are the weights of each layer in the intermediate layer of the dimension transformation, B1 and B2 are the biases of each layer in the intermediate layer of the dimension transformation, and Relu is the activation function;

[0037] S2: Collect data and information; initialize the weights of each neural network, and calculate the function output value of each node according to the function; input the sample data information into the constructed network to obtain the corresponding dimension transformation result, collect data and information, integrate the sample data into a data set, and use 80% of the data set as a training set and 20% as a test set;

[0038] S3: Classify the data samples and construct a decomposition model of the prediction matrix; the decomposition model of the prediction matrix includes an input layer at the bottom, a hidden layer at the middle, and an output layer at the top. The input layer consists of several input nodes, each of which is used to receive and process input variables in the training sample set. The hidden layer consists of several hidden nodes, which are used for linear transformation of nonlinear samples. The hidden layer nodes use wavelet functions as transfer functions. The output layer consists of several output nodes, each of which is used to output variable prediction results.

[0039] The decomposition model calculation formula of the prediction matrix is:

[0040] f=clip(U+R_TR*Z. / RN+UA)*(I+IR. / IAN) T ;

[0041] Clip() sets the lower limit of each element of the matrix to 1 and the upper limit to 5; A is the average parameter, B is the deviation parameter, U is the user feature parameter, I is the feature parameter, and Z is the historical feature parameter, all of which are matrices; TR is the prediction result of the training set, the row represents the training sample, the column represents the test sample, and each element of the matrix represents the prediction result of big data based on matrix decomposition technology; R_TR is the record matrix of the training set, which sets the result record in the training set scoring matrix TR to 1 and the result record to 0; RN is the module value matrix of each row of the training set result record, each row of which is the square root of the module value of the corresponding row of RTR, which needs to be divided by dot, and the size is the same as R_TR*Z; IAN is the module value matrix of each row of the item attribute, each row of which is the square root of the module value of the corresponding row of the attribute matrix, which needs to be divided by dot, and the size is the same as IA;

[0042] S4: Calculate the predicted output value of the neural network, and compare and analyze the predicted output value of the neural network with the expected output value of the neural network to calculate the network prediction error;

[0043] S5: Modify each network weight and each function parameter in the neural network prediction model to achieve the prediction expectation; classify the data samples and divide the data samples into two parts, one part is a training sample for training the wavelet neural network, and the other part is a test sample for testing the prediction accuracy of the wavelet neural network. According to each network weight in the wavelet neural network prediction model and the function output value of each hidden node in the hidden layer, calculate the wavelet neural network prediction output value, and compare and analyze the wavelet neural network prediction output value with the wavelet neural network expected output value to calculate the network prediction error;

[0044] S6: perform parameter iteration, and when the network prediction error reaches a specified smaller value or reaches a specified number of iterations, end the iteration and obtain the prediction result; when there is an inconsistency in the order of magnitude or dimension of the data in the decomposition model matrix of the prediction matrix, after obtaining the corresponding data matrices and before performing matrix decomposition on each data matrix, it also includes: determining the data matrix with inconsistent order of magnitude or dimension as the target matrix, and normalizing each data in the target matrix; performing matrix decomposition on each data matrix, including: performing matrix decomposition on each data matrix using a non-negative matrix decomposition algorithm, and after predicting the corresponding next prediction result, it also includes: obtaining each next real data value; using each real data value and the corresponding prediction value to verify the validity of the prediction model used in the prediction process.

[0045] Embodiment 2:

[0046] On the basis of the first embodiment, a preferred embodiment of the big data prediction method based on matrix decomposition technology provided by the present invention is shown in the figure: the model does not introduce other information to assist in describing users and items, and uses the row vector of the rating matrix as user information and the column vector as item information, which are respectively input into two multi-layer neural networks for feature extraction. The general steps are as follows:

[0047]

[0048] Among them, max(x) represents the maximum value of the sequence data in each data matrix, min(x) represents the minimum value of the sequence data in each data matrix, x represents any data value in the data matrix, and represents the data after normalization. Due to the activation function characteristics of the neural network, it is sensitive to data in the interval [-1, 1]. If the input data is in the interval [-1, 1], it will greatly accelerate the training speed and training effect of the attribute information neural network. The non-negative matrix decomposition algorithm is used to decompose each data matrix to obtain the sequence factor matrix and the time series dependency factor matrix corresponding to each user; the non-negative matrix decomposition algorithm NMF (Non-negative Matrix Factorization) is used to decompose each data matrix to obtain the time series factor matrix F and the time series dependency factor matrix X, so that the value of each position corresponding to the matrix obtained by the matrix product of F and X is as small as possible compared with the value of the corresponding position of the original matrix Y. That is:

[0049]

[0050] in, is the Frobenius norm. For example, the Frobenius norm of matrix A is defined as the sum of the squares of the absolute values ​​of the elements of matrix A, followed by a square root operation, that is:

[0051]

[0052] Input each time series dependency factor matrix into the network respectively, calculate the next time series dependency factor; predict the corresponding next data result by performing matrix operations on each time series factor matrix and the next time series dependency factor respectively; obtain the next real data value respectively; obtain the next real data value of each user respectively; use each real data value and the corresponding predicted data value to test the validity of the prediction model used in the prediction process;

[0053] The method provided by the embodiment of the present invention is used to obtain a sequence data set, and the sequence data of each user in the sequence data set are combined to obtain the corresponding data matrices; the matrix decomposition is performed on each data matrix to obtain the corresponding time series factor matrix and time series dependency factor matrix; the corresponding time series dependency factor matrix is ​​input into the echo state network to calculate the corresponding next time series dependency factor; the corresponding next data value is predicted by performing matrix operations on the corresponding time series factor matrix and the next time series dependency factor; the corresponding time series factor matrix and time series dependency factor matrix are obtained by using matrix decomposition technology, and the time series invariant characteristics and dependency factors that change with time are captured. The introduction of the echo state network enhances the original matrix decomposition ability to capture time series modeling and multi-scale structures, enhances the modeling ability of complex dynamics, and greatly improves the prediction accuracy of big data based on matrix decomposition technology.

[0054] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0055] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A big data prediction method based on matrix decomposition technology, characterized in that: include: S1: Establish an attribute information neural network with a dimension transformation intermediate layer; S2: Collect data and information; initialize the weights of each neural network, and calculate the function output value of each node according to the function; S3: classify the data samples and build a decomposition model of the prediction matrix; S4: Calculate the predicted output value of the neural network, and compare and analyze the predicted output value of the neural network with the expected output value of the neural network to calculate the network prediction error; S5: Modify each network weight and each parameter of the function in the neural network prediction model to achieve the prediction expectation; S6: Perform parameter iteration. When the network prediction error reaches a specified smaller value or reaches a specified number of iterations, the iteration ends and the prediction result is obtained.

2. The big data prediction method based on matrix decomposition technology according to claim 1, characterized in that: In S1, an attribute information neural network with a dimensionality transformation intermediate layer is constructed. The network is composed of several layers. There are at least two dimensionality transformation layers belonging to the input information between the bottom layer and the top layer. The dimensionality transformation of the input information is performed by setting the hyperparameters of the weights and biases of each layer of the network to obtain the dimensionality transformation result of the input information in the network. The number of layers required for the network structure varies with the scale of the input information. The network structure is: X is the input information involved in the prediction, and is also the input information of the attribute information neural network; taking the dimensionality transformation intermediate layer as an example with two layers, Y is the network output result of the dimensionality transformation of X through the two-layer network: Y = Re lu (X * W1 + B1) * W2 + B2; Where W1 and W2 are the weights of each layer in the intermediate layer of dimensional transformation, B1 and B2 are the biases of each layer in the intermediate layer of dimensional transformation, and Relu is the activation function.

3. The big data prediction method based on matrix decomposition technology according to claim 1, characterized in that: In S2, the sample data information is input into the constructed network to obtain the corresponding dimension transformation result, collect data and information, integrate the sample data into a data set, and use 80% of the data set as a training set and 20% as a test set.

4. The big data prediction method based on matrix decomposition technology according to claim 1, characterized in that: In S3, the decomposition model of the prediction matrix includes an input layer at the bottom layer, a hidden layer at the middle layer, and an output layer at the top layer, wherein the input layer is composed of a plurality of input nodes, each of which is used to receive and process input variables in a training sample set, the hidden layer is composed of a plurality of hidden nodes, which are used for linear transformation of nonlinear samples, and the hidden layer nodes use wavelet functions as transfer functions; the output layer is composed of a plurality of output nodes, each of which is used to output variable prediction results.

5. The big data prediction method based on matrix decomposition technology according to claim 1, characterized in that: In S3, the calculation formula of the decomposition model of the prediction matrix is: f=clip(U+R_TR*Z. / RN+UA)*(I+IR. / IAN) T ; Clip() sets the lower limit of each element of the matrix to 1 and the upper limit to 5; among them, A is the average parameter, B is the deviation parameter, U is the user feature parameter, I is the feature parameter, and Z is the historical feature parameter, all of which are matrices; TR is the prediction result of the training set, the rows represent the training samples, the columns represent the test samples, and each element of the matrix represents the prediction result of big data based on matrix decomposition technology; RTR is the record matrix of the training set, which sets the result records in the training set scoring matrix TR to 1 and the result records without result records to 0; RN is the module value matrix of each row of the training set result record, each row of which is the square root of the module value of the corresponding row of RTR, which needs to be divided by dot, and the size is the same as RTR*Z; IAN is the module value matrix of each row of item attributes, each row of which is the square root of the module value of the corresponding row of the attribute matrix, which needs to be divided by dot, and the size is the same as IA.

6. The big data prediction method based on matrix decomposition technology according to claim 1, characterized in that: In S4 and S5, the data samples are classified into two parts, one part is a training sample for training the wavelet neural network, and the other part is a test sample for testing the prediction accuracy of the wavelet neural network. The wavelet neural network prediction output value is calculated according to each network weight in the wavelet neural network prediction model and the function output value of each hidden node in the hidden layer, and the wavelet neural network prediction output value is compared and analyzed with the wavelet neural network expected output value to calculate the network prediction error.

7. The big data prediction method based on matrix decomposition technology according to claim 1, characterized in that: In the S6, when there is inconsistency in the order of magnitude or dimension of the data in the decomposition model matrix of the prediction matrix, after obtaining the corresponding data matrices and before performing matrix decomposition on each data matrix, it also includes: determining the data matrix with inconsistent order of magnitude or dimension of the data as the target matrix, and performing normalization operation on each data in the target matrix.

8. The big data prediction method based on matrix decomposition technology according to claim 7, characterized in that: Performing matrix decomposition on each data matrix respectively, including: performing matrix decomposition on each data matrix respectively using a non-negative matrix decomposition algorithm, and after predicting the corresponding next prediction result, further including: obtaining each next real data value respectively; and using each real data value and the corresponding prediction value respectively to test the validity of the prediction model used in the prediction process.

Images