Prediction device, prediction method, and program

Abstract

To provide technology that enables highly interpretable predictions regarding the sensory characteristics of food products. [Solution] A prediction device according to one aspect of the present disclosure includes: a map creation unit that creates a sensory characteristic map, which is multidimensional coordinate information representing the relationship between the sensory characteristics and the samples, based on evaluation results obtained by evaluating the sensory characteristics of a plurality of samples; a model creation unit that creates a prediction model including the sensory characteristic map and a product map, which is multidimensional coordinate information representing the relationship between the samples and the components, based on a plurality of component data representing the concentrations of the components of the plurality of samples; an estimation unit that estimates the range of the coordinate position of the target sample in the sensory characteristic map based on component data of the target sample, which is a sample whose sensory characteristics are to be predicted, and the product map; and a prediction unit that predicts the sensory characteristics of the target sample based on the sensory characteristic map and the range.

Description

[Technical Field]

[0001] This disclosure relates to a prediction device, a prediction method, and a program. [Background technology]

[0002] In recent years, consumer preferences have become more diverse, necessitating an acceleration of new food product development. However, testing to evaluate the sensory characteristics of prototypes is time-consuming and costly, currently posing a bottleneck in the development cycle.

[0003] There are various technologies that can streamline the evaluation of the sensory characteristics of food, and one of them is a technology that uses machine learning to create a model that can predict sensory characteristics from food component data (for example, Non-Patent Documents 1 and 2). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Debnath, T., Prasetyawan, D., & Nakamoto, T. (2021). Predicting Odor Perception of Mixed Scent from Mass Spectrometry. Journal of The Electrochemical Society, 168(117505). [Non-Patent Document 2] Perrot, NM, Roche, A., Tonda, A., Lutton, E., & Thomas Danguin, T. (2023). Predicting odor profile of food from its chemical composition: Towards an approach based on artificial intelligence and flavorists expertise. Mathematical Biosciences and Engineering, 20(12), 20528 20552. [Overview of the Initiative] [Problems that the invention aims to solve]

[0005] However, while models created using machine learning techniques can be expected to have high prediction accuracy, their prediction process is a black box, making it difficult to interpret the prediction results. In particular, in the food industry, it is important to be able to show the basis for the prediction results from the perspective of quality control and safety, so the interpretability of the prediction results is one of the important factors.

[0006] This disclosure is made in view of the above points and aims to provide a technology that can achieve highly interpretable predictions regarding the sensory characteristics of food products. [Means for solving the problem]

[0007] A prediction device according to one aspect of the present disclosure includes: a map creation unit that creates a sensory characteristic map, which is multidimensional coordinate information representing the relationship between the sensory characteristics and the samples, based on evaluation results of evaluating the sensory characteristics of a plurality of samples; a model creation unit that creates a prediction model including the sensory characteristic map and a product map, which is multidimensional coordinate information representing the relationship between the samples and the components, based on a plurality of component data representing the concentrations of the components of the plurality of samples; an estimation unit that estimates the range of the coordinate position of the target sample in the sensory characteristic map based on component data of the target sample, which is a sample whose sensory characteristics are to be predicted, and the product map; and a prediction unit that predicts the sensory characteristics of the target sample based on the sensory characteristic map and the range. [Effects of the Invention]

[0008] This enables highly interpretable predictions regarding the sensory characteristics of food products. [Brief explanation of the drawing]

[0009] [Figure 1] This figure shows an example of how to create a sensory characteristics map. [Figure 2] This figure shows an example of how to create a predictive model. [Figure 3] This figure shows an example of how to estimate the coordinate range of a sample whose sensory characteristics are unknown. [Figure 4] This figure shows an example of how to predict sensory characteristics. [Figure 5] This figure shows an example of the hardware configuration of a prediction device according to one embodiment. [Figure 6] This figure shows an example of the functional configuration of a prediction device according to one embodiment. [Figure 7] A flowchart showing an example of the model creation process according to one embodiment. [Figure 8] A flowchart illustrating an example of a prediction process according to one embodiment. [Modes for carrying out the invention]

[0010] One embodiment of the present invention will be described in detail below with reference to the drawings.

[0011] <Proposed method> Below, we propose a method for predicting the sensory characteristics of beverages and foods from their component data using various analytical techniques, based on a highly interpretable predictive model that can predict sensory characteristics from component data of beverages and foods whose sensory characteristics are unknown. While we assume milk, processed milk, and dairy beverages as examples of beverages and foods below, this is merely one example and is not limited to these. The proposed method can be similarly applied to various beverages and foods other than milk, processed milk, and dairy beverages. Furthermore, in the following, the beverage samples used in the proposed method will be referred to as "samples."

[0012] The proposed method consists of the following four steps, from Step 1 to Step 4.

[0013] • Step 1: Create a sensory characteristics map Figure 1 shows an example of how to create a sensory characteristics map. As shown in Figure 1, it is assumed that a cross-tabulation table has been obtained for multiple samples, in which their sensory characteristics have been evaluated using the CATA (Check-All-That-Apply) method, one of the sensory evaluation methods. In the example shown in Figure 1, a cross-tabulation table has been obtained for samples A to D, in which the sensory items representing sensory characteristics, namely "milk flavor," "sweet aroma," "unpleasant aftertaste," "richness," and "clean aftertaste," have been evaluated. However, the CATA method is just one example, and other sensory evaluation methods may be used as long as they can be used to obtain cross-tabulation tables or quantitative data tables for multiple samples that evaluate their sensory characteristics.

[0014] At this time, a modified correspondence analysis method, which is an improved version of correspondence analysis, is used to create information called a sensory characteristic map from the cross-tabulation table. A sensory characteristic map is multidimensional coordinate information that represents the relationship between the sensory items used in the sensory evaluation method and each sample using statistical distance (specifically, a value calculated based on the adjusted standardized residuals used in residual analysis). In the example shown in Figure 1, sensory characteristic map 1100 is shown, which represents the relationship between samples A to D and the sensory items "milk flavor," "sweet aroma," "unpleasant aftertaste," "richness," and "clean aftertaste" using statistical distance.

[0015] For example, when creating a sensory characteristics map from a quantitative data table, a biplot of principal component loadings and principal component scores obtained by methods such as Principal Component Analysis (PCA) or Partial Least Squares (PLS) may be used. As with sensory characteristics map 1100, if multidimensional coordinate information of sensory items and samples can be calculated, other quantitative data analysis methods can also be used.

[0016] Step 2: Create a predictive model Figure 2 shows an example of how to create a predictive model. As shown in Figure 2, information called a product map is created from the component data of the samples used in the sensory evaluation method using principal component analysis. A product map is multidimensional coordinate information that represents the relationship between certain components and each sample. In the example shown in Figure 2, product map 1200 is shown, which represents the relationship between samples A to D and components 1, 3, 6, and 9. Component data refers to data composed of measured values ​​of the concentrations of certain components in the sample. Component data may also be called, for example, "component concentration data".

[0017] Furthermore, as shown in Figure 2, an index (e.g., RV coefficient, contribution rate of each dimension) is used to evaluate the similarity between the sensory characteristics map and the product map, and a combination of components C (hereinafter also referred to as "component combination C") is selected for principal component analysis. The example shown in Figure 2 shows the case where C = {component 1, component 3, component 6, component 9} is selected.

[0018] Furthermore, as shown in Figure 2, the product map is corrected so that the positional relationships between samples in the product map are consistent with the positional relationships between samples in the sensory characteristics map (in other words, to minimize errors in those positional relationships). Then, a transformation matrix T is calculated that associates the component data of each sample with the corrected product map. The example shown in Figure 2 shows the case where the transformation matrix T is calculated after applying a correction called Procrustes rotation correction to the product map.

[0019] This yields a predictive model containing information including the component combination C, the transformation matrix T, the sensory characteristic map, and the product map corresponding to the component combination C. The predictive model also includes, for example, the threshold used in step 4 described later to evaluate the significance of the relationship between "samples with unknown sensory characteristics" and "sensory items," and the average error of the positional relationship of each sample. In the following, the components included in component combination C will also be referred to as "selected components."

[0020] Step 3: Estimating the coordinate range of a sample with unknown sensory characteristics. Figure 3 shows an example of estimating the coordinate range of a sample with unknown sensory characteristics. As shown in Figure 3, it is assumed that the concentrations of the selected components included in component combination C have been obtained from the component data of a sample with unknown sensory characteristics (hereinafter also referred to as the "unknown sample"). The example shown in Figure 3 shows the case where the concentrations of component 1, component 3, component 6, and component 9 have been obtained.

[0021] At this point, as shown in Figure 3, the concentration of the selected component of the unknown sample is transformed into a coordinate position on the product map using the transformation matrix T. Then, the coordinate range on the sensory characteristics map is estimated from that coordinate position. In the example shown in Figure 3, the coordinate range 1300 on the sensory characteristics map 1100 is estimated. Note that various errors exist between the sensory characteristics map and the product map (e.g., errors in the positional relationship of each sample, differences in the contribution rate of the coordinate axes, etc.), so the position coordinates on the product map are estimated as a coordinate range on the sensory characteristics map.

[0022] Step 4: Predicting sensory characteristics Figure 4 shows an example of how to predict sensory characteristics. As shown in Figure 4, the coordinate range estimated in step 3 above is used to evaluate whether there is a significant positive relationship with each sensory item. The example shown in Figure 4 shows the evaluation of whether there is a significant positive relationship between the coordinate range 1300 and each of the sensory items "milky flavor," "sweet aroma," and "clean aftertaste."

[0023] A more specific evaluation method involves calculating the minimum value of the dot product between all coordinate positions representing the sensory items and the coordinate range estimated in step 3 above. If this minimum value of the dot product is greater than or equal to a threshold, a significant positive relationship is considered to exist; otherwise, no significant positive relationship is considered to exist. This predicts that the unknown sample possesses the sensory characteristics represented by the sensory items that are evaluated as having a significant positive relationship.

[0024] Furthermore, when using a sensory characteristic map obtained from quantitative data tables for prediction, it is sufficient to evaluate whether there is a significant positive relationship using any statistical method. For example, in a sensory characteristic map created using PCA, the dot product of the numerous points included in the estimated coordinate range of the unknown sample and the vector representing the sensory items can be calculated, and the mean of these dot products can be compared with the dot product distribution of the known sample using a Z-score or permutation test to evaluate whether there is a significant positive relationship with each sensory item. Moreover, it is also possible to evaluate the strength and significance of the relationship between the unknown sample and each sensory item by considering the distribution of a quantity obtained by projecting (linearly transforming) a Bayesian prediction, which treats the coordinates of the unknown sample as a probability distribution, in the direction of the vector representing the sensory items, and calculating the Z-score of that distribution.

[0025] The following describes a prediction device 10 that can create a prediction model and predict the sensory characteristics of an unknown sample using the proposed method described above.

[0026] <Example of hardware configuration for prediction device 10> Figure 5 shows an example of the hardware configuration of a prediction device 10 according to one embodiment. As shown in Figure 5, the prediction device 10 according to one embodiment includes an input device 101, a display device 102, an external I / F 103, a communication I / F 104, a RAM (Random Access Memory) 105, a ROM (Read Only Memory) 106, an auxiliary storage device 107, and a processor 108. Each of these hardware components is connected to each other via a bus 109 so as to be able to communicate.

[0027] The input device 101 is, for example, a keyboard, mouse, touch panel, or physical button. The display device 102 is, for example, a display or display panel. The prediction device 10 does not necessarily have to have at least one of the input device 101 and the display device 102.

[0028] External I / F 103 is an interface with external devices such as recording media 103a. Examples of recording media 103a include CD (Compact Disc), DVD (Digital Versatile Disk), SD memory card (Secure Digital memory card), and USB (Universal Serial Bus) memory card.

[0029] The communication interface 104 is an interface for connecting to a communication network. The RAM 105 is a volatile semiconductor memory (storage device) that temporarily holds programs and data. The ROM 106 is a non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. The auxiliary storage device 107 is a non-volatile storage device such as an HDD (Hard Disk Drive), SSD (Solid State Drive), or flash memory. The processor 108 is an arithmetic unit such as a CPU (Central Processing Unit).

[0030] Note that the hardware configuration of the prediction device 10 shown in Figure 5 is just one example, and the hardware configuration of the prediction device 10 is not limited to this. The prediction device 10 may have, for example, multiple auxiliary storage devices 107 or multiple processors 108, or it may not have some of the hardware shown, or it may have various hardware other than the hardware shown.

[0031] <Example of Functional Configuration of Prediction Device 10> Figure 6 shows an example of the functional configuration of a prediction device 10 according to one embodiment. As shown in Figure 6, the prediction device 10 according to one embodiment has a model creation processing unit 201 and a prediction processing unit 202. Each of these units is realized, for example, by processing that one or more programs installed in the prediction device 10 are executed by a processor 108 or the like. The prediction device 10 according to one embodiment also has a storage unit 203. The storage unit 203 is realized, for example, by a storage area such as an auxiliary storage device 107. However, the storage unit 203 may also be realized by a storage area of ​​a storage device (e.g., a storage device of a database server) that is communicably connected to the prediction device 10.

[0032] The model creation processing unit 201 executes a predictive model creation process to create a predictive model. Here, the model creation processing unit 201 includes a sensory characteristic map creation unit 211 and a model creation unit 212. The sensory characteristic map creation unit 211 creates a sensory characteristic map according to the procedure 1 described above. The model creation unit 212 creates a predictive model according to the procedure 2 described above and saves it to the storage unit 203.

[0033] The prediction processing unit 202 performs prediction processing to predict the sensory characteristics of an unknown sample. Here, the prediction processing unit 202 includes a coordinate range estimation unit 221 and a sensory characteristics prediction unit 222. The coordinate range estimation unit 221 estimates the coordinate range of the unknown sample on the sensory characteristics map according to step 3 above. The sensory characteristics prediction unit 222 predicts the sensory characteristics of the unknown sample according to step 4 above.

[0034] The memory unit 203 stores the prediction model. The memory unit 203 may also store, for example, intermediate calculation results when creating the prediction model, or intermediate calculation results when predicting the sensory characteristics of an unknown sample.

[0035] <Example of model creation process> Figure 7 is a flowchart showing an example of the model creation process according to one embodiment. Hereinafter, the prediction device 10 is assumed to be provided with a cross-tabulation table in which the sensory characteristics of multiple samples were evaluated using the CATA method, and component data of those multiple samples.

[0036] Here, the cross-tabulation table is a matrix X=(x) of size (N+1)×(M+1). ij Let it be represented as follows: 1≦i≦N is the sample number, 1≦j≦M is the sensory item number, and N≦M is assumed to be the case. Hereafter, the i-th sample will be referred to as "sample i", and the j-th sensory item as "sensory item j". ij (where 1 ≤ i ≤ N, 1 ≤ j ≤ M) represents the number of votes for sensory item j for sample i. Also, x ij (where i=N+1, 1≦j≦M) is the column sum of the jth column (i.e., the sum of the votes for each sample i (1≦i≦N) for sensory item j), x ij (where 1 ≤ i ≤ N, j = M + 1) is the row sum for row i (i.e., the sum of the votes for each sensory item j (1 ≤ j ≤ M) for sample i). Furthermore, x ij (where i=N+1 and j=M+1) represents the total number of votes.

[0037] Furthermore, the component data of sample i (1 ≤ i ≤ N) is a K-dimensional vector Y. i =( y ik Let it be represented as ). 1 ≤ k ≤ K is the component number. Hereafter, the k-th component will be represented as "component k". y ik This represents the concentration of component k in sample i.

[0038] The sensory characteristic map creation unit 211 of the model creation processing unit 201 creates a sensory characteristic map from the given cross-tabulation table X using a modified correspondence analysis method (step S101). The modified correspondence analysis method is a technique that improves upon existing correspondence analysis and can determine adjusted standardized residuals that represent statistically significant relationships.

[0039] Specifically, the sensory characteristic map creation unit 211 creates a sensory characteristic map, for example, according to the following procedures 1-1 to 1-3.

[0040] Procedure 1-1: The sensory characteristic map creation unit 211 calculates the adjusted standardized residual s ij (1 ≤ i ≤ N, 1 ≤ j ≤ M) of each vote count x included in the cross-tabulation table X. ij The adjusted standardized residual s ij is defined as s ij = r ij / √(E ij (1 - m i r (1 - m j c )). Here, r ij is the residual of the vote count x ij , E ij is the expected value of the vote count x ij . Also, m i r a value called the row mass, which is the sum of the vote counts in the i-th row of the cross-tabulation table X / the total vote count (that is, the row total in the i-th row / the total vote count). Similarly, m j c c is a value called the column mass, which is the sum of the vote counts in the j-th column of the cross-tabulation table X / the total vote count (that is, the column total in the j-th column / the total vote count). Note that the definitions of the residual r ij and the expected value E ij are the same as those in known residual analysis. Hereinafter, an N×M matrix with s ij as the (i, j) element is denoted as AdjS = (s ij ). )

[0041] Procedure 1-2: The sensory characteristic map creation unit 211 calculates a threshold value (hereinafter referred to as threshold th) used to evaluate the significance with the sensory items in the above procedure 4. The threshold th may be calculated by any method. For example, when it is assumed that the adjusted standardized residual s ij included in AdjS follows a normal distribution, it is conceivable to set th = 1.96σ. Here, σ is the standard deviation of the adjusted standardized residual s ij . This is because s ijIf ≥ th, it means that sample i and sensory item j have a significant positive relationship (i.e., significance level p < 0.05).

[0042] In this embodiment, the adjusted standardized residual s ij We assumed a normal distribution for the distribution of s, but this is just one example. Adjusted standardized residuals s ij The distribution of is not limited to parametric probability distributions such as the normal distribution; for example, probability distributions obtained by nonparametric methods such as kernel density estimation may also be used. In this case, the threshold th is the adjusted standardized residual s ij Based on the cumulative distribution function of the assumed distribution (e.g., parametric probability distribution, nonparametric distribution such as kernel density distribution, etc.), it is possible to use a value corresponding to the point where the area of ​​the distribution reaches 95% (e.g., the upper 5% point).

[0043] Steps 1-3: The sensory characteristic map creation unit 211 performs singular value decomposition (i.e., UDV) on AdjS. τ (expressed in the form of) and converted into coordinates. Specifically, first, the sensory characteristic map creation unit 211 performs singular value decomposition and calculates AdjS = UΣV τ We calculate matrices U, Σ, and V that satisfy the following conditions. U is an N×N orthogonal matrix, Σ is an N×M matrix, and V is an M×M orthogonal matrix. Note that τ is the symbol for transpose. Here, among the components of matrix Σ, the components whose row number and column number match are defined as Σ ii If we consider (1≦i≦N), then the components Σ ii (1≦i≦N) are singular values ​​σ1, σ2, ..., σ N (However, 0≦σ1≦σ2≦···≦σ N ) is composed of the singular values ​​σ1, σ2, ..., σ N If D is an N×N diagonal matrix with the diagonal elements and O is an N×(MN) zero matrix, then Σ=[DO]. Therefore, if V' is an M×N matrix composed of the (i,j) components of V (where 1≦i≦M and 1≦j≦N), then UDV' τ To satisfy =AdjS, the sensory characteristic map creation unit 211 calculates a matrix D such as this. Next, the sensory characteristic map creation unit 211 performs exponential decomposition of matrix D = D1(z / (1+z)) D2 (1 / (1+z)) (where z≠-1 is a parameter, and D1 and D2 are N×N diagonal matrices) are calculated. Then, the sensory characteristic map creation unit 211 sets AdjS=UD1 (z / (1+z)) D2 (1 / (1+z)) V' τ Using the fact that it can be expressed as, UD1 (z / (1+z)) The i-th row (1≦i≦N) is the coordinate of sample i, D2 (1 / (1+z)) V' τ The j-th column is plotted in an N-dimensional coordinate space as the coordinate of sensory item j. This yields a sensory characteristic map consisting of N-dimensional coordinates representing sample i (1 ≤ i ≤ N) and N-dimensional coordinates representing sensory item j (1 ≤ j ≤ M).

[0044] Note D1 (z / (1+z)) These are matrices U and D2. (1 / (1+z)) is matrix V' τ Since it is the scaling factor, the parameter z may also be called the scale parameter. The value of the scale parameter z can be set to any value as appropriate, but for example, sample i and sensory item j (i.e., s) have a significant positive relationship. ij The value to set is the one that minimizes the sum of the distances between the coordinates of sample i and sensory item j) where ≥ th.

[0045] The model creation unit 212 of the model creation processing unit 201 processes the given component data Y i A predictive model is created from (1≦i≦N) and the sensory characteristic map created in step S101 above and stored in the memory unit 203 (step S102).

[0046] Specifically, the model creation unit 212 creates and saves a predictive model, for example, by following steps 2-1 to 2-7.

[0047] Step 2-1: The model creation unit 212 is, K C N For each of the N possible combinations of components, the corresponding N component data Y iPrincipal component analysis is performed on (1≦i≦N). Subsequently, the model creation unit 212 places each component data Y in an N-dimensional space with the first to the Nth principal components calculated as a result of the principal component analysis as axes. i A product map is created by plotting the principal component scores obtained based on each of these (i.e., the principal component scores corresponding to that combination). K C N For each of the N possible combinations of components, a product map is obtained in which the principal component scores of sample i (1 ≤ i ≤ N) are plotted in an N-dimensional space.

[0048] For example, if K=6 and N=4, then 6C4=15. Therefore, for each of the 15 combinations of selecting 4 components from components 1 to 6, we have 4 component data Y. i Principal component analysis is performed on (1≦i≦4). Then, for each of the combinations, the four component data Y i Since four principal components (1st principal component to 4th principal component) can be obtained from (1≦i≦4), the data for each component Y can be placed in a 4-dimensional space with these as axes. i A product map is created by plotting the principal component scores for each (1 ≤ i ≤ 4) group.

[0049] Below, o(1≦o≦ K C N Let's denote the )th product map as "product map o".

[0050] Step 2-2: The model creation unit 212 uses the sensory characteristic map created in step S101 above and each product map o(1≦o≦) created in step 2-1 above. K C N The similarity between each of the two is evaluated. Indicators that can be used to evaluate similarity include, for example, the RV coefficient (Reference 1), the contribution rate of each dimension, and their weighted sum.

[0051] Here, the RV coefficient is an index that represents the strength of the correlation between two matrices, each representing a different multivariate dataset. Below, we have the N×N matrix U created when the sensory characteristics map was created, and the N×N matrix U with the coordinates of each sample i in the product map o as the i-th row. o As such, matrix U and matrix U o The RV coefficient between U,U o ) Note that 0≦RV(U,U o The value is ≤ 1, and the closer the value is to 1, the stronger the correlation.

[0052] Furthermore, principal component analysis reveals the matrix U o The contribution rate c to the N principal components (i.e., the N dimensions of product map o) based on the combination of components corresponding to product map o. o1 ,···,c oN Similarly, the singular values ​​σ1, σ2, ..., σ obtained by singular value decomposition can be calculated. N Using this, the contribution rates c1,···,c to the N dimensions of the sensory characteristic map from matrix U. N Therefore, the matrix U can be calculated. o SEVRI is an index that represents the closeness between the contribution rate for each dimension calculated from the matrix U and the contribution rate for each dimension calculated from the matrix U. o (Similarity of Explained Variance Ration Index)=1 / ((c o1 -c1) 2 +···+(c oN -c N ) 2 ) is defined as SEVRI o A larger value indicates that the contribution rates of each dimension are closer together.

[0053] Step 2-3: The model creation unit 212 selects component combination C using the evaluation results from Step 2-2 above. For example, the model creation unit 212 selects RV(U,U o ) From the combinations where the value is greater than or equal to a predetermined value, SEVRI oThe combination with the largest [value] may be selected as the component combination C. However, this is just an example, and the component combination C may be selected by other methods. For example, the model creation unit 212 may select, as the component combination C, the combination with the largest weighted sum of RV(U,U o ) and SEVRI o .

[0054] Hereinafter, the product map o corresponding to the component combination C selected in the above procedure 2-3 (that is, the product map o with N principal components based on the components included in the component combination C as each dimension) will be referred to as the "selected product map".

[0055] Procedure 2-4: The model creation unit 212 corrects the selected product map so as to minimize the error in the positional relationship between the sensory characteristic map and the selected product map. Specifically, the model creation unit 212 performs Procrustes rotation correction on the selected product map so as to minimize the error between the same samples, using the coordinates of each sample on the sensory characteristic map and the coordinates of each sample on the selected product map. In Procrustes rotation correction, inversion, rotation, translation, reduction, expansion, etc. are performed on the polygon with each sample on the selected product map as vertices so that the polygon with each sample on the sensory characteristic map as vertices and the polygon with each sample on the selected product map as vertices coincide as much as possible.

[0056] Hereinafter, the selected product map after Procrustes rotation correction will be referred to as the "corrected selected product map". Also, the average of the errors between the coordinates of each sample on the sensory characteristic map and the coordinates of each sample on the corrected selected product map is defined as Δ (hereinafter, also referred to as the average error Δ). As the error, for example, RMSE (Root Mean Squared Error) or the like can be used.

[0057] Procedure 2-5: The model creation unit 212 uses the component data Y of each sample i (1 ≤ i ≤ N)[[ID=2)]] iCalculate a conversion matrix T for converting the concentration of the selected components (i.e., the components included in the component combination C) to the coordinates of the same sample on the corrected selection product map. This is for each component data Y i If the N×N matrix with the concentrations of the selected components for each row is W, and the N×N matrix with the coordinates of sample i on the corrected selection product map for each row is W', then the N×N matrix T that satisfies WT = W' can be obtained as the conversion matrix T.

[0058] Procedure 2-6: The model creation unit 212 creates a prediction model using the information including the sensory characteristic map, the corrected selection product map, the component combination C, the conversion matrix T, the threshold th, and the average error Δ. In addition to these, the prediction model may also include, for example, the contribution rate c n of the n-th (1≦n≦N) dimension of the sensory characteristic map, and the contribution rate c' n of the n-th (1≦n≦N) dimension of the corrected selection product map, and the error δ n between them. Hereinafter, this error δ n will be referred to as the "contribution rate error δ n ".

[0059] Procedure 2-7: The model creation unit 212 stores the prediction model created in the above Procedure 2-6 in the storage unit 203. Thereby, a highly interpretable prediction model is obtained.

[0060] <Example of prediction processing> FIG. 8 is a flowchart showing an example of the prediction processing according to an embodiment. Hereinafter, it is assumed that the component data Y = (y k )(1≦k≦K) of an unknown sample is given. 1≦k≦K is the component number, and y k is the concentration of component k in the unknown sample. However, the component data Y of the unknown sample does not necessarily have to be composed of the concentrations of K components, and it is sufficient if it includes the concentrations of N selected components. Hereinafter, for simplicity, among the concentrations y k of the component data Y of the unknown sample, y1, ···, y N are taken as the concentrations of the selected components.

[0061] The coordinate range estimation unit 221 of the prediction processing unit 202 estimates the coordinate range of the unknown sample on the sensory characteristics map (step S201).

[0062] Specifically, the coordinate range estimation unit 221 estimates the coordinate range of an unknown sample by, for example, following the steps 3-1 to 3-3 below.

[0063] Step 3-1: The coordinate range estimation unit 221 uses the transformation matrix T to determine the concentrations of selected components y1,···,y included in the component data of the unknown sample. N The coordinates are converted to coordinates on the corrected selected product map. That is, the coordinate range estimation unit 221 calculates Y'=(y1,···,y N )T determines the concentrations of the selected components y1,···,y N Convert the corrected coordinates to Y' on the selected product map.

[0064] Step 3-2: The coordinate range estimation unit 221 maps the coordinates Y' obtained in Step 3-1 onto the sensory characteristics map. Since the polygons with the coordinates of each sample as vertices on the corrected selected product map and the polygons with the coordinates of each sample as vertices on the sensory characteristics map are similar and have little error, the coordinate range estimation unit 221 can simply map the coordinates Y' directly onto the sensory characteristics map. Hereafter, the coordinates obtained by mapping the coordinates Y' onto the sensory characteristics map will also be represented by Y'.

[0065] Step 3-3: The coordinate range estimation unit 221 estimates the range of the unknown sample as the range of a circle or ellipse region (more precisely, an N-dimensional circle or ellipse region) centered on coordinate Y' on the sensory characteristic map. Specifically, the coordinate range estimation unit 221 estimates the range of the unknown sample as, for example, a circle region with radius Δ centered on coordinate Y'. However, in addition to this, the coordinate range estimation unit 221 also estimates, for example, a region with dimension n (1 ≤ n ≤ N) centered on coordinate Y' and axis length Δ × (100 + δ n The elliptical region defined as ) / 100 may be estimated as the coordinate range of the unknown sample.

[0066] The sensory characteristic prediction unit 222 of the prediction processing unit 202 predicts the sensory characteristics of an unknown sample using the coordinate range estimated in step S202 and the coordinates of the sensory item j (1≦j≦M) on the sensory characteristic map (step S202).

[0067] Specifically, the sensory characteristic prediction unit 222 predicts the sensory characteristics of an unknown sample by, for example, following the procedures 4-1 to 4-3 below.

[0068] Step 4-1: The sensory characteristic prediction unit 222 calculates the minimum value of the dot product between the coordinate of each sensory item j (1 ≤ j ≤ M) on the sensory characteristic map and the coordinates included in the coordinate range estimated in step S202 above. Hereinafter, the minimum value of the dot product between the coordinate of sensory item j and the coordinates included in the coordinate range is calculated as e j Let's assume that.

[0069] Step 4-2: The sensory characteristic prediction unit 222 performs the following for each sensory item j (1 ≤ j ≤ M): j Determine whether or not ≥ th. This is e j If ≥ th, then regardless of which coordinate the true coordinate of the unknown sample is within that coordinate range, the dot product of the true coordinate of the unknown sample and the coordinate of sensory item j will be greater than or equal to the threshold th.

[0070] Step 4-3: The sensory characteristic prediction unit 222 is e j If there is a sensory item j that is determined to be ≥ th, the unknown sample is predicted to possess the sensory characteristics represented by that sensory item j. This allows for the estimation of the sensory characteristics of the unknown sample.

[0071] <Summary> As described above, the prediction device 10 according to one embodiment creates a prediction model that predicts sensory characteristics from a certain number of selected components using a highly interpretable method such as a modified correspondence analysis, which is an improved method of correspondence analysis. As a result, the prediction device 10 according to one embodiment can predict the sensory characteristics of a sample whose sensory characteristics are unknown with high interpretability and accuracy. Therefore, by using the prediction device 10 according to one embodiment, it becomes possible to accelerate, for example, the development of new food products.

[0072] The present invention is not limited to the embodiments specifically disclosed above, and various modifications, changes, and combinations with known technologies are possible as long as they do not deviate from the spirit described in the claims.

[0073] [References] Reference 1: Robert, P. and Escaufier, Y. (1976). A unifying tool for linear multivariate statistical methods: the RV-coefficient. Appl. Statist, 25, 257-65. [Explanation of symbols]

[0074] 10 Prediction device 101 Input Device 102 Display device 103 External I / F 103a Recording medium 104 Communication I / F 105 RAM 106 ROM 107 Auxiliary storage 108 processors 109 Bus 201 Model Creation Processing Unit 202 Prediction Processing Unit 203 Storage section 211 Sensory Characteristics Map Creation Department 212 Model Creation Department 221 Coordinate Range Estimation Unit 222 Sensory Characteristics Prediction Unit

Claims

1. A first creation unit creates a sensory characteristic map, which is multidimensional coordinate information representing the relationship between the sensory characteristics and the sample, based on evaluation results obtained by evaluating the sensory characteristics of multiple samples. A second creation unit creates a product map, which is multidimensional coordinate information representing the relationship between the sample and the component, based on a plurality of component data representing the concentrations of the components of the plurality of samples and the sensory characteristics map. An estimation unit estimates the range of the coordinate position of the target sample in the sensory characteristics map by setting a predetermined circular or elliptical region centered on the coordinate position obtained by mapping the coordinate position of the target sample in the product map to the sensory characteristics map, based on the component data of the target sample which is the sample to be predicted for sensory characteristics and the product map, A prediction unit that predicts the sensory characteristics of the target sample based on the sensory characteristics map and the range, It has, The second creation unit described above is: Let K be the number of components included in the component data, and N be the number of samples in the sensory characteristic map and the dimension of the sensory characteristic map. For each of the N combinations of components (K C N), principal component analysis is performed on the multiple component data to create K C N product map candidates. A prediction device that defines the product map as the product map obtained by correcting the product map candidate that is most similar to the sensory characteristic map among the KCN product map candidates, so as to minimize the error in the positional relationship with the sensory characteristic map.

2. The second creation unit described above is: The prediction device according to claim 1, wherein the product map is obtained by applying Procrustes rotation correction to the product map candidate that is most similar to the sensory characteristic map among the KCN product map candidates, in order to minimize the error in the positional relationship with the sensory characteristic map.

3. The estimation unit, The prediction device according to claim 1 or 2, wherein a predetermined circular or elliptical region is set centered on the coordinate position obtained by mapping the coordinate position represented by the concentration of a component included in a combination of N components corresponding to the product map, among the concentrations included in the component data of the target sample, onto the sensory characteristics map.

4. The prediction unit, The prediction device according to claim 3, which predicts the sensory characteristics of the target sample based on the dot product of coordinates included in the range and coordinates representing the sensory characteristics on the sensory characteristics map.

5. The prediction unit, The prediction device according to claim 4, wherein if the minimum value of the dot product is greater than or equal to a predetermined threshold, the device predicts that the target sample has the aforementioned sensory characteristics.

6. The first creation unit described above is: Based on the evaluation results representing the quantitative data table, a sensory characteristic map is created from a biplot of principal component loadings and principal component scores obtained by principal component analysis or partial least squares method. The prediction unit, The prediction device according to claim 1, which predicts the sensory characteristics of the target sample using a predetermined statistical method.

7. A first creation procedure for creating a sensory characteristic map, which is multidimensional coordinate information representing the relationship between the sensory characteristics and the sample, based on evaluation results obtained by evaluating the sensory characteristics of multiple samples, A second creation procedure for creating a product map, which is multidimensional coordinate information representing the relationship between the sample and the component, based on multiple component data representing the concentrations of the components of the multiple samples and the sensory characteristics map. An estimation procedure for estimating the range of the coordinate positions of the target sample in the sensory characteristics map by setting a predetermined circular or elliptical region centered on the coordinate position obtained by mapping the coordinate position of the target sample in the product map to the sensory characteristics map, based on the component data of the target sample which is the sample to be predicted for sensory characteristics and the product map, A prediction procedure for predicting the sensory characteristics of the target sample based on the sensory characteristics map and the range, The computer executes this, The second creation procedure described above is: Let K be the number of components included in the component data, and N be the number of samples in the sensory characteristic map and the dimension of the sensory characteristic map. For each of the N combinations of components (K C N), principal component analysis is performed on the multiple component data to create K C N product map candidates. A prediction method in which, among the KCN product map candidates mentioned above, the product map that is most similar to the sensory characteristic map is corrected to minimize the error in the positional relationship with the sensory characteristic map, and this corrected product map is defined as the product map.

8. A first creation procedure for creating a sensory characteristic map, which is multidimensional coordinate information representing the relationship between the sensory characteristics and the sample, based on evaluation results obtained by evaluating the sensory characteristics of multiple samples, A second creation procedure for creating a product map, which is multidimensional coordinate information representing the relationship between the sample and the component, based on multiple component data representing the concentrations of the components of the multiple samples and the sensory characteristics map. An estimation procedure for estimating the range of the coordinate positions of the target sample in the sensory characteristics map by setting a predetermined circular or elliptical region centered on the coordinate position obtained by mapping the coordinate position of the target sample in the product map to the sensory characteristics map, based on the component data of the target sample which is the sample to be predicted for sensory characteristics and the product map, A prediction procedure for predicting the sensory characteristics of the target sample based on the sensory characteristics map and the range, Have the computer run it, The second creation procedure described above is: Let K be the number of components included in the component data, and N be the number of samples in the sensory characteristic map and the dimension of the sensory characteristic map. For each of the N combinations of components (K C N), principal component analysis is performed on the multiple component data to create K C N product map candidates. A program that defines the product map as the product map obtained by correcting the product map candidate that is most similar to the sensory characteristic map among the KCN product map candidates mentioned above, so as to minimize the error in the positional relationship between the sensory characteristic map and the product map candidate that is corrected to minimize the error in the positional relationship between the sensory characteristic map and the product map candidate.

Images