TWI865935B - Data processing method and apparatus - Google Patents

Abstract

Data predicting method and apparatus are provided. In the method, distances between a to-be-predicted data and multiple data groups are determined. A first machine learning model corresponding the data group having the shortest distance with the to-be-predicted data is selected from multiple machine learning model. The first machine learning model is used to predict the to-be-predicted data. Those machine learning models are trained by using different data groups, respectively. Accordingly, the predicted result of the module could be improved.

Description

Data prediction method and device

The present invention relates to a data prediction technology, and in particular to a data prediction method and device for machine learning.

Machine learning algorithms can analyze large amounts of data to infer the patterns of these data and make predictions about unknown data. In recent years, machine learning has been widely used in fields such as image recognition, natural language processing, outcome prediction, medical diagnosis, error detection, or speech recognition.

In view of this, the present invention provides a data prediction method and device that can group the predicted data to improve the prediction accuracy.

The data prediction method of the embodiment of the present invention is applicable to machine learning, and the data prediction method includes (but is not limited to) the following steps: determining the distance between the data to be predicted and multiple data groups. Selecting a machine learning model corresponding to the data group with the shortest distance between the data to be predicted from multiple machine learning models. Predicting the data to be predicted using the first machine learning model. Those machine learning models are trained using different data groups respectively.

The data prediction device of the embodiment of the present invention includes (but is not limited to) a memory and a processor. The memory is used to store program code. The processor is coupled to the memory. The processor is configured to load the program code to execute: determine the distance between the data to be predicted and multiple data groups. Select the first machine learning model corresponding to the data group with the shortest distance between the data to be predicted from multiple machine learning models. Use the first machine learning model to predict the data to be predicted. Those machine learning models are trained using different data groups respectively.

Based on the above, according to the data prediction method and device of the embodiment of the present invention, the first machine learning model corresponding to the data group most similar to the data to be predicted is found, and the data to be predicted is predicted accordingly. This helps to improve the accuracy, sensitivity and specificity of machine learning.

In order to make the above features and advantages of the present invention more clearly understood, the following is a detailed description of the embodiments with the accompanying drawings.

10: Electronic devices

11: Memory

12: Processor

15: Sensor

S210~S250, S910~S930: Steps

FIG1 is a block diagram of components of a data prediction device according to an embodiment of the present invention.

Figure 2 is a flow chart of a data prediction method according to an embodiment of the present invention.

Figure 3 is a schematic diagram of the analysis results according to an embodiment of the present invention.

Figure 4 is a principal component distribution diagram according to an embodiment of the present invention.

Figure 5 is a schematic diagram of hierarchical clustering according to an embodiment of the present invention.

Figure 6 is a schematic diagram of the verification results of the first group training according to an embodiment of the present invention.

Figure 7 is a schematic diagram of the verification results of the second group training according to an embodiment of the present invention.

Figure 8 is a schematic diagram of the verification results of multi-group joint training according to an embodiment of the present invention.

Figure 9 is a flow chart of data prediction according to an embodiment of the present invention.

FIG1 is a block diagram of a data prediction device 10 according to an embodiment of the present invention. Referring to FIG1 , the data prediction device 10 includes (but is not limited to) a memory 11 and a processor 12. The data prediction device 10 may be a mobile phone, a tablet computer, a laptop computer, a desktop computer, a voice assistant device, a smart home appliance, a wearable device, a car device, or other electronic devices.

The memory 11 can be any type of fixed or removable random access memory (RAM), read only memory (ROM), flash memory, traditional hard disk drive (HDD), solid-state drive (SSD) or similar components. In one embodiment, the memory 11 is used to store program code, software modules, configurations, data or files (e.g., data, models, or features), and will be described in detail in subsequent embodiments.

The processor 12 is coupled to the memory 11. The processor 12 may be a central processing unit (CPU), a graphics processing unit (GPU), or other programmable general-purpose or special-purpose microprocessor, digital signal processor (DSP), programmable controller, field programmable gate array (FPGA), application-specific integrated circuit (ASIC), neural network accelerator or other similar components or a combination of the above components. In one embodiment, the processor 12 is used to execute all or part of the operations of the data prediction device 10, and can load and execute various program codes, software modules, files and data stored in the memory 11. In some embodiments, some operations in the method of the present invention may be implemented by different or the same processor 12.

In one embodiment, the data prediction device 10 further includes a sensor 15. The processor 12 is coupled to the sensor 15. For example, the sensor 15 is connected to the processor 12 via USB, Thunderbolt, Wi-Fi, Bluetooth or other wired or wireless communication technologies. For another example, the data prediction device 10 has a built-in sensor 15. The sensor 15 can be a radar, a microphone, a temperature sensor, a humidity sensor, an image sensor, a motion sensor or other types of sensors. In one embodiment, the sensor 15 is used for sensing to obtain sensing data. In one embodiment, the sensing data is time-dependent data. That is, data recorded with a timing, continuous time or multiple time points. For example, the sensing data is a radar sensing result (e.g., in-phase or quadrature signal), a sound signal, or a continuous image.

In the following, the method described in the embodiment of the present invention will be described with the various devices, components and modules in the data prediction device 10. The various processes of the method can be adjusted according to the implementation situation, but are not limited to this.

FIG2 is a flow chart of a data prediction method according to an embodiment of the present invention. Referring to FIG2, the processor 12 performs dimensionality reduction analysis on multiple feature sets to obtain analysis results (step S201). Specifically, each feature set includes one or more features. The type of feature may vary depending on the type of sensing data of the sensor 15. Taking the IQ signal of the radar as an example, the feature may be a feature related to the variance or waveform between different channels. Another example is the zero-crossing rate (ZCR), pitch, or Mel Frequency Cepstral Coefficients (MFCC) in acoustic features.

In one embodiment, the processor 12 can convert multiple sensing data into those feature sets. For example, the IQ signal is converted into the variance between different channels and waveform-related features. For another example, the sound signal is converted into ZCR, pitch, or MFCC.

處理器12可將表(1)的感測資料透過重新塑形(re-shape)轉換成矩陣形式。例如，300×500的矩陣，且其元素為I或Q資料。 For example, Table (1) shows the IQ sensing data of the radar:

The processor 12 may convert the sensing data in Table (1) into a matrix form by reshaping, for example, a 300×500 matrix whose elements are I or Q data.

In another embodiment, the processor 12 can download or receive sensing data from an external sensor or a feature set generated by an external computing device through a communication transceiver (not shown).

Different feature sets may correspond to sensing data of different subjects or different target objects. For example, the first feature set is converted from sensing data of the first subject, and the second feature set is converted from sensing data of the second subject. Alternatively, different feature sets may correspond to sensing data of the same subject or the same target object but at different times or in different environments. For example, the third feature set corresponds to sensing data of the third subject in the first time period, and the fourth feature set corresponds to sensing data of the third subject in the second time period.

In one embodiment, the processor 12 may mark one or more feature sets. For example, marking events such as hypopnea, wakefulness, or apnea. However, the marking content may still vary depending on the feature type, and the embodiments of the present invention are not limited thereto.

Dimensionality reduction analysis is used to reduce features. That is, each feature is considered as a dimension, and reducing the dimension will also reduce the features. In one embodiment, the dimensionality reduction analysis is principal component analysis (PCA) or principal coordinate analysis (PCoA). For PCA, it uses orthogonal transformation to linearly transform the observed values of a series of possibly related variables (features in this embodiment) to project them into a series of linearly unrelated variable values. These unrelated variables are called principal components. In other words, the most important elements and structures are found from multiple features. Unlike PCA, PCoA is a projection of the distance matrix (recording the difference/distance between two observations) obtained by passing the observations through different distance algorithms. In addition, PCoA finds the most important coordinates in the distance matrix.

The analysis result can be the principal component and its proportion, or the principal coordinate and its proportion. The proportion refers to the principal component or the principal coordinate. For example, FIG3 is a schematic diagram of the analysis result according to an embodiment of the present invention. Please refer to FIG3, assuming that the sensing data is data of sleep sensing through a continuous wave (CW) radar, and the corresponding marked verification data is the data produced by a full-night sleep multiple physiological function test (Polysomnography, PSG). The comparison target is sleep events such as hypopnea, wakefulness or apnea events. That is, radar is used to predict sleep events. This embodiment uses data from 32 subjects for analysis. After converting the radar data of the 32 subjects into features and performing PCA/PCoA processing, the principal component structure shown in FIG3 can be obtained. The analysis results include principal components PC1~PC11 and proportions. The principal component PC1 accounts for the highest proportion.

In other embodiments, the dimension reduction analysis may be Linear Discriminant Analysis (LDA), t-Distributed Stochastic Neighbor Embedding (t-SNE), or other dimension reduction. The analysis results include the reduced features or dimensions and their proportions.

Referring to FIG. 2 , the processor 12 may normalize the feature sets according to the analysis results to generate multiple normalized feature sets (step S220). Specifically, normalization is to scale the feature values and make the scaled values fall into a specific interval (e.g., [0,1] or [0,10]). That is, the values of each feature in the feature set are scaled to a specific interval.

In one embodiment, the processor 12 selects one or more first principal components from a plurality of principal components and normalizes the feature set according to the first principal component. For example, the processor 12 sets the maximum and minimum values of the interval and normalizes each principal component to make the benchmark points consistent with each other.

In one embodiment, the first principal component is the principal component with the highest proportion among those principal components. For example, the proportion of principal component PC1 in Figure 3 is much larger than that of other principal components PC2~PC11, so principal component PC1 can be selected for subsequent normalization processing.

In another embodiment, the first principal component is the principal component with the highest proportion or the second highest proportion among the principal components. The difference between the principal component with the highest proportion and the principal component with the second highest proportion is less than a threshold value (e.g., 3, 5, or 10%). For example, if the difference between the principal component with the highest proportion and the principal component with the second highest proportion is within 5%, the principal component with the second highest proportion is considered and selected. If there are other principal components ranked by proportion whose differences with the principal component with the highest proportion are also less than the threshold value, they will also be considered for subsequent regularization.

而表(2)經轉換後的排名為表(3)

In one embodiment, the processor 12 may sort the feature sets by percentile transformation, that is, converting the feature values into rankings. For example, Table (2) is a feature set of features: Table (2)

The ranking of Table (2) after transformation is Table (3)

FIG4 is a principal component distribution diagram according to an embodiment of the present invention. Referring to FIG4, the horizontal axis is the features and is arranged according to the labels, and the vertical axis is the numbers of different subjects. The features of different subjects are not the same. For example, the performance (i.e., the importance, and represented by different gray levels in the figure) of the 17th and 18th features of subjects numbered 5 and 10 are different from those of other subjects.

Referring to FIG. 2 , the processor 12 generates the distance relationship of those normalized feature sets (step S230). Specifically, the distance relationship includes the distance between two of those normalized feature sets. The processor 12 may project the features in the normalized feature sets into the same space and form coordinates, and calculate the distance between the features of different normalized feature sets in the space (i.e., the distance between two coordinates).

In one embodiment, the distance relationship is a distance matrix, and each element in the distance matrix is the distance between features in two normalized feature sets. The distance algorithm can be Euclidean distance, cosine similarity, or KL divergence. For example, the first normalized feature set is [1.5, 2.2], the second normalized feature set is [0.1, 1.6], and the third normalized feature set is [5.7, 4.3]. The distance matrix is [1.52, 4.7, 6.22]. Taking the Euclidean distance as an example, taking the square root of (1.5-0.1)^2+(2.2-1.6)^2 gives 1.52, and the rest are similar.

The distance relationship is not limited to the matrix form. In other embodiments, the distance relationship may also be a comparison table, a mathematical conversion formula, or other relationship that records the distances between different feature sets.

Please refer to FIG. 2 , the processor 12 clusters those feature sets according to the distance relationship to generate multiple data groups (step S240). Specifically, each data group includes one or more feature sets. The similarity between different feature sets can be known from the distance relationship. Clustering is to assign multiple feature sets with higher similarity to the same data group. The clustering method can be K-Means Clustering, Hierarchical Clustering or Fuzzy Clustering.

For example, FIG5 is a schematic diagram of clustering according to a hierarchical clustering method according to an embodiment of the present invention. Referring to FIG5, each feature set number corresponds to a feature set (e.g., a subject). The processor 12 uses the hierarchical clustering method to cluster the feature sets with the closest distance into one of those data groups. The feature sets can be split into two groups between numbers 28 and 16 (e.g., feature sets corresponding to subjects with numbers 16 and 28). It should be noted that the closest refers to the result of comparison with the distance threshold. If the distance between the two feature sets is less than the distance threshold, they are regarded as the closest two feature sets; otherwise, they are regarded as two non-close feature sets.

In one embodiment, the processor 12 may determine the number of groups of the data groups, determine the clustering distance according to the number of groups, and cluster the feature sets according to the clustering distance. Taking FIG. 5 as an example, if the number of groups is 2, the clustering distance is 60. The distances between the feature sets numbered 5, 12, 11, 27, 19, 23, 30, 3, 28 (for example, the feature sets corresponding to the subject numbers 5, 12, 11, 27, 19, 23, 30, 3, 28) are within 60, so they are all assigned to the same data group. If the number of groups is 3, the clustering distance is 50. The distances between the feature sets numbered 16, 24, 10, 15, 29 are within 50, so they are all assigned to the same data group.

Please refer to FIG. 2 , the processor 12 uses those data groups to train multiple machine learning models (step S250). Specifically, after the clustering results are obtained, each data group can be trained separately from other data groups. Those machine learning models are trained using different data groups. The processor 12 can use the feature sets corresponding to each data group (that is, the feature sets converted from the sensor data) or the unconverted sensor data to train the corresponding machine learning models. For example, the feature set of the first data group is used to train the first machine learning model, and the feature set of the second data group is used to train the second machine learning model. The first data group will not be used to train the second machine learning model. In addition, the machine learning algorithm can be deep learning, decision tree, recurrent neural network (RNN) or other algorithms.

The following verification results can prove that the clustering training of the embodiment of the present invention is helpful for the training of machine learning.

FIG6 is a schematic diagram of the verification results of the first group training separately according to an embodiment of the present invention, FIG7 is a schematic diagram of the verification results of the second group training separately according to an embodiment of the present invention, and FIG8 is a schematic diagram of the verification results of the multi-group joint training according to an embodiment of the present invention. Please refer to FIG6, FIG7 and FIG8, accuracy is the correctness of multiple prediction results obtained by each machine learning model and the actual results. Sensitivity is the proportion of samples judged as positive in actual positive samples. Specificity is the proportion of samples judged as negative in actual negative samples.

FIG6 is a verification result of training using the feature sets numbered 5, 12, 11, 27, 19, 23, 30, 3, 28 in FIG5 (e.g., feature sets corresponding to subject numbers 5, 12, 11, 27, 19, 23, 30, 3, 28 or original sensing data). FIG7 is a verification result of training using other feature sets in FIG5 (e.g., feature sets corresponding to other subjects other than subject numbers 5, 12, 11, 27, 19, 23, 30, 3, 28 or original sensing data). FIG8 is a verification result of training using all feature sets in FIG5 (e.g., feature sets corresponding to all subjects or original sensing data). Cluster training (corresponding to Figure 6 and Figure 7) is superior to joint training (corresponding to Figure 8) in accuracy, sensitivity and specificity. Taking accuracy as an example, under joint training, the accuracy shown in Figure 8 converges to about 0.7. Under cluster training, the accuracy shown in Figure 6 and Figure 7 can converge to above 0.7. Even the accuracy of Figure 6 for the same data group can converge to about 0.9.

In addition to training optimization, the embodiment of the present invention can optimize model prediction. FIG. 9 is a flow chart of data prediction according to an embodiment of the present invention. Referring to FIG. 9, the processor 12 can determine the distance between the data to be predicted and those data groups (step S910). Specifically, the processor 12 can first obtain the data to be predicted. The data to be predicted can refer to the description of the aforementioned sensing data, which will not be repeated here. As required, the processor 12 converts the data to be predicted into a feature set to be predicted. The description of feature conversion can refer to the conversion of the aforementioned sensing data to the feature set, which will not be repeated here. Then, the processor 12 determines the distance between the feature set to be predicted and multiple data groups.

For example, the representative values (e.g., mean, median, or other statistical values) of the first data group are [8.16, 9.8, 3.7, 15.54, 2.74, 4.04, 16.82, 4.56, 21., 11.88, 12.78, 11.1, 9.54, 7.22, 7.24, 18.34, 17.04, 4.24, 20., 12.1, 13.16], and the representative values of the second data group are [4.61, 6.4 2,9.95,5.7,4.,6.61,2.85,10.28,21,15.85,14.66,12.047,8.28,10.38,9.95,18.85,16.42,3.57,20,13.33,16.09], and the feature set to be predicted is [10,13,6,16,2,3,17,5,21,9,15,12,8,7,4,19,18,1,20,11,14]. Taking the Euclidean distance as an example, the distance between the feature set to be predicted and the first data group is 7.855, and the distance between the feature set to be predicted and the second data group is 23.495.

The processor 12 can select a first machine learning model corresponding to the data group with the shortest distance between the data to be predicted from multiple machine learning models (step S920), and use the first machine learning model to predict the data to be predicted (step S930). For example, 7.855 is less than 23.495, so the data group with the shortest distance between the feature set to be predicted is the first data group. The processor 12 can load the first machine learning model of the first data group, and input the data to be predicted into the loaded first machine learning model to predict the result. The predicted data takes the sensing result of the radar as an example, and the predicted result can be a sleep event. However, the predicted result can still be changed according to actual needs.

It should be noted that in one embodiment, in response to the distance between multiple data groups and the data to be predicted being less than the lower limit of the distance or greater than the upper limit of the distance, the machine learning models of these data groups may all be selected to predict the results of the data to be predicted. In another embodiment, in response to the distance between multiple data groups and the data to be predicted being the same or less than a preset value, the processor 12 may load the machine learning model trained by multiple data groups for prediction.

In summary, in the data prediction method and device of the embodiment of the present invention, the feature set is normalized according to the result of dimension reduction and further grouped. Then, different machine learning models are trained using different data groups. In addition, machine learning models corresponding to data groups with similar distances are selected for prediction. In this way, the effect of training and prediction can be improved.

Although the present invention has been disclosed as above by the embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of the attached patent application.

S910~S920: Steps

Claims (20)

A data prediction method, applicable to machine learning, comprises: grouping a plurality of feature sets; performing a dimensionality reduction analysis on the feature sets to obtain an analysis result, wherein the dimensionality reduction analysis is principal component analysis (PCA) or principal coordinates analysis (PCoA), and the analysis result comprises the proportion of a plurality of principal components; normalizing the feature sets according to the analysis result to generate a plurality of normalized feature sets, and generating a plurality of data groups, wherein each of the normalized feature sets comprises at least one feature, and each of the data groups comprises at least one of the normalized feature sets; training a plurality of machine learning models respectively using the data groups; determining a data to be predicted The distance between the data groups; determining the data group with the shortest distance to the data to be predicted as a first data group, wherein a first machine learning model has been trained using the first data group; selecting the first machine learning model from the machine learning models; and predicting the data to be predicted using the first machine learning model, wherein the machine learning models are trained using different data groups respectively. The data prediction method as described in claim 1 further includes: generating a distance relationship among the normalized feature sets, wherein the distance relationship includes the distance between two of the normalized feature sets; and grouping the feature sets according to the distance relationship to generate the data groups. The data prediction method as described in claim 2, wherein the step of normalizing the feature sets based on the analysis results includes: selecting a first principal component from the principal components; and normalizing the feature sets based on the first principal component. The data prediction method as described in claim 3, wherein the first principal component is the principal component with the highest proportion among the principal components. The data prediction method as described in claim 3, wherein the first principal component is the principal component with the highest proportion or the second highest proportion among the principal components, and the difference between the principal component with the highest proportion and the principal component with the second highest proportion is less than a threshold value. A data prediction method as described in claim 2, wherein the distance relationship is a distance matrix, and each element in the distance matrix is the distance between two features in the normalized feature set. The data prediction method as described in claim 2, wherein the step of clustering the feature sets according to the distance relationship includes: clustering the feature sets with the closest distance into one of the data groups using a hierarchical clustering method. The data prediction method as described in claim 7 further includes: determining the number of groups of the data groups; determining a clustering distance based on the number of groups; and clustering the feature sets based on the clustering distance. The data prediction method as described in claim 2 further includes: converting a plurality of sensing data into the feature sets, wherein the sensing data are time-dependent data; and using the feature sets or the sensing data corresponding to each data group to train the corresponding machine learning model. The data prediction method as described in claim 9, wherein each of the sensing data is a sensing result of a radar. A data prediction device includes: a memory storing a program code; and a processor coupled to the memory and configured to load the program code to execute: grouping a plurality of feature sets; performing a dimensionality reduction analysis on the feature sets to obtain an analysis result, wherein the dimensionality reduction analysis is principal component analysis (PCA) or principal coordinate analysis (PCA). The method comprises the following steps: performing a PCoA analysis, wherein the analysis result comprises the proportion of a plurality of principal components; normalizing the feature sets according to the analysis result to generate a plurality of normalized feature sets, and generating a plurality of data groups, wherein each of the normalized feature sets comprises at least one feature, and each of the data groups comprises at least one of the normalized feature sets; training a plurality of machine learning models using the data groups respectively; determining a pair of data to be predicted and a pair of data to be predicted; The distance between the data groups; determining that the data group with the shortest distance to the data to be predicted is a first data group, wherein a first machine learning model has been trained using the first data group; selecting the first machine learning model from the machine learning models; and predicting the data to be predicted using the first machine learning model, wherein the machine learning models are trained using different data groups respectively. The data prediction device as described in claim 11, wherein the processor further executes: generating a distance relationship of the normalized feature sets, wherein the distance relationship includes the distance between two of the normalized feature sets; and clustering the feature sets according to the distance relationship to generate the data groups. The data prediction device as described in claim 12, wherein the processor further performs: selecting a first principal component from the principal components; and normalizing the feature sets according to the first principal component. A data prediction device as described in claim 13, wherein the first principal component is the principal component with the highest proportion among the principal components. The data prediction device as described in claim 13, wherein the first principal component is the principal component with the highest proportion or the second highest proportion among the principal components, and the difference between the principal component with the highest proportion and the principal component with the second highest proportion is less than a threshold value. A data prediction device as described in claim 12, wherein the distance relationship is a distance matrix, and each element in the distance matrix is the distance between two features in the normalized feature set. The data prediction device as described in claim 12, wherein the processor further performs: using a hierarchical clustering method to cluster the feature sets with the closest distance into one of the data groups. The data prediction device as described in claim 17, wherein the processor further performs: determining the number of groups of the data groups; determining a clustering distance based on the number of groups; and clustering the feature sets based on the clustering distance. The data prediction device as described in claim 18, wherein the processor further performs: converting a plurality of sensing data into the feature sets, wherein the sensing data are time-dependent data; and using the feature sets or the sensing data corresponding to each data group to train the corresponding machine learning model. A data prediction device as described in claim 19, wherein each of the sensing data is a sensing result of a radar.

Images