Development support method and development support apparatus

Abstract

This development support method for supporting development of a composition by using a computer comprises: performing data analysis on a first dataset in which a plurality of compositions and a plurality of feature amounts included in the compositions are associated with each other; generating a first feature space by performing dimensionality reduction by a predetermined dimensionality reduction method; setting a target composition for the first feature space; and calculating a plurality of estimated feature amounts that are included in the target composition, by performing an inverse analysis on the target composition on the basis of the position of the target composition in the first feature space and a parameter that is obtained by the dimensionality reduction.

Description

Development support method and development support device

[0001] This disclosure relates to a development support method and a development support device.

[0002] Traditionally, developers have developed new compositions through trial and error, referring to the analysis results of compositions similar to the target composition. Various methods are known for analyzing compositions (for example, Non-Patent Document 1, Patent Document 1, Patent Document 2).

[0003] Japanese Patent Publication No. 2021-76458, International Publication No. 2023 / 181167

[0004] Simone Poggesi et al. , “FUSION OF 2DGC-MS, HPLC-MS AND SENSORY FATA TO ASSIST DECISION-MAKING IN THE MARKETING OF INTERNATIONAL MONOVARIETAL CHARDONNAY AND SAUVIGNON BLANC WINES”, Foods 11 (2022) 3458

[0005] However, conventional methods are difficult to apply when there are few compositions similar to the new composition being developed.

[0006] This disclosure aims to provide a technology that enables the efficient development of new compositions, even if there are few existing compositions similar to the target composition.

[0007] One aspect of the present disclosure provides a development support method for supporting the development of a composition using a computer, which involves data analysis of a first dataset in which a plurality of compositions and a plurality of features contained in the composition are associated, generating a first feature space by compressing the dimensions using a predetermined dimensionality reduction method, setting a target composition in the first feature space, and calculating a plurality of estimated features contained in the target composition by performing an inverse analysis on the target composition based on the position of the target composition in the first feature space and parameters obtained by the dimensionality reduction.

[0008] One aspect of the present disclosure provides a development support device for assisting in the development of a composition, comprising a processor and a memory, wherein the processor cooperates with the memory to perform data analysis on a first dataset in which a plurality of compositions and a plurality of features contained in the composition are associated, and generates a first feature space by compressing the dimensions using a predetermined dimensionality reduction method, sets a target composition in the first feature space, and calculates a plurality of estimated features contained in the target composition by performing an inverse analysis on the target composition based on the position of the target composition in the first feature space and parameters obtained by the dimensionality reduction.

[0009] According to this disclosure, even if there are not many compositions similar to the target new composition, it is possible to efficiently develop a new composition.

[0010] This is a block diagram showing an example configuration of the development support device according to this embodiment. This is a flowchart showing an example of the development support method according to this embodiment. This is a table showing an example of the first dataset according to Example 1. This is a diagram showing the two-dimensional principal component loading matrix L according to Example 1. This is a diagram showing an example of the first map in which each composition and each feature are simultaneously biploted in the first feature space according to Example 1. This is the transpose matrix L of the two-dimensional principal component loading matrix L according to Example 1. ｔThis is a diagram. This is a table showing the estimated feature quantities of each feature included in the target composition T according to Example 1. This is a radar chart showing the estimated feature quantities of each feature included in the target composition T according to Example 1. This is a diagram showing an example of a second map plotting the target composition T in the second feature space according to Example 1. This is a table showing the estimated feature quantities of each feature included in the target composition T according to the first modification of Example 1. This is a diagram showing an example of a second map plotting the target composition T according to the first modification of Example 1. This is a table showing the estimated feature quantities of each feature included in the target composition T according to the second modification of Example 1. This is a diagram showing an example of a second map plotting the target composition T according to the second modification of Example 1. This is a table showing an example of a first dataset according to Example 2. This is a diagram showing a two-dimensional factor loading matrix L according to Example 2. This is a diagram showing an example of a first map simultaneously biplotting each composition and each feature in the first feature space according to Example 2. This is the transpose matrix L of the two-dimensional factor loading matrix L according to Example 2. ｔ This is a figure. This is a table showing the estimated feature quantities of each feature included in the target composition T according to Example 2. This is a bar graph showing the estimated feature quantities of each feature included in the target composition T according to Example 2. This is a figure showing an example of a second map plotting the target composition T in the second feature space according to Example 2. This is a figure showing the two-dimensional canonical weight matrix L according to Example 3. This is a figure showing an example of a first map simultaneously biplotting each composition and each feature in the first feature space according to Example 3. This is the transpose matrix L of the two-dimensional canonical weight matrix L according to Example 3. ｔThis is a diagram. This is a table showing the estimated feature quantities of each feature contained in the target composition T according to Example 3. This is a pie chart showing the estimated feature quantities of each feature contained in the target composition T according to Example 3. This is a diagram showing an example of a second map plotting the target composition T in the second feature space according to Example 3. This is a diagram showing an example of an MFA map of commercially available chocolate according to Example 4. This is a graph showing the estimated values ​​of the raw material compound data of Target 1 according to Example 4. This is a graph showing the estimated values ​​of the subjective evaluation data of Target 1 according to Example 4. This is a diagram showing an example of an MFA map of Japanese citrus according to Example 5. This is a diagram showing an example of an MFA map of Japanese citrus according to Example 6. This is a diagram showing an example of a PCA map of Japanese citrus according to Example 7. This is a diagram showing an example of a 3D MFA map of commercially available chocolate according to Example 8. This is a diagram showing an example of a GCCA map of commercially available chocolate according to Example 9.

[0011] Embodiments of the present disclosure will be described in detail below, with appropriate reference to the drawings. However, descriptions that are unnecessarily detailed may be omitted. For example, detailed descriptions of already well-known matters and redundant descriptions of substantially identical configurations may be omitted. This is to avoid the following description becoming unnecessarily verbose and to facilitate understanding for those skilled in the art. The accompanying drawings and the following description are provided to enable those skilled in the art to fully understand the present disclosure and are not intended to limit the subject matter of the claims. The functions of one configuration shown in this embodiment may be realized by two or more physical configurations, or the functions of two or more configurations may be realized by, for example, one physical configuration.

[0012] (Background to this disclosure) The Blue Ocean Strategy is a strategy for developing new areas that did not previously exist. By using the Blue Ocean Strategy to avoid competition with other companies, companies can more easily achieve product differentiation and corporate growth. The Blue Ocean Strategy is also important in the field of compositions. Products made from similar compositions tend to have similar value to consumers and are prone to price competition. Therefore, by developing products in areas with less competition, companies can gain an advantage in acquiring customers and improving profits.

[0013] On the other hand, a bottleneck when adopting a blue ocean strategy is estimating information about unknown products located in areas where no conventional products exist. Because no similar products exist, it is difficult to perform instrumental analysis or sensory evaluation of existing similar products, as is done in conventional development. Therefore, traditionally, developers imagine the target composition to be newly developed and develop it based on experience and intuition. A composition consists of multiple raw materials. By changing the combination method, content, and manufacturing method of multiple raw materials, a variety of compositions can be developed. For example, when developing a blended fragrance as a composition, in order to realize the target composition set by the customer or developer, the developer combines multiple raw materials at appropriate concentrations to create a composition recipe called a prescription. In this case, if the target composition is a known product as an existing product or natural product, the developer develops the target composition based on instrumental analysis or sensory evaluation of that known product. During this time, the developer may repeatedly perform sensory evaluations of reference prototypes or instrumental analysis of similar products.

[0014] Low-dimensional feature spatial information, created through statistical analysis based on data from instrumental analysis or sensory evaluation, is used to confirm the relative positional relationships between groups of compositions and to set target compositions.

[0015] Non-patent document 1 discloses that a two-dimensional map is created by combining the results of instrumental analysis and sensory evaluation and performing Multiple Factor Analysis (MFA), thereby supporting the setting of marketing targets by understanding the relative relationships between samples. However, the purpose of creating the map is limited to setting strategic target compositions, and there is no mention of compositional information such as which raw materials to combine and at what concentrations to achieve the target compositions.

[0016] Patent Document 1 discloses a method for calculating a correlation coefficient based on the intensity of the fragrance profile of a standard fragrance and the olfactory intensity of compounds contained in that fragrance, in order to facilitate the sharing or transmission of fragrance information. Furthermore, Patent Document 1 discloses a fragrance development support method and apparatus that provides design information by calculating the compounds that contribute to a target fragrance and their olfactory intensity based on the calculated correlation coefficient. Although Patent Document 1 provides information on which compounds to combine and at what concentrations to achieve a target fragrance, the calculation is based on the correlation coefficient. Calculating this correlation coefficient requires a process of varying the gas phase concentration of each fragrance compound and constructing a curve model of gas phase concentration and olfactory intensity through sensory evaluation of the olfactory intensity at that time. To construct a model with high accuracy, repeated sensory evaluation by multiple people is required for each fragrance compound that can be used in fragrance development, which takes an enormous amount of time. In addition, the prediction formula in Patent Document 1 that calculates the fragrance compounds related to the target fragrance profile and their olfactory intensity does not include an interaction term of fragrance compounds, and the prediction accuracy of the fragrance profile of the target fragrance composition is not sufficient. When developing a composition that is expressed by complex features, focusing on only some of its characteristics may result in a composition that meets the intended quality for those specific characteristics, but other characteristics may not meet the intended quality, potentially leading to a composition that is not entirely satisfactory as a whole.

[0017] Patent Document 2 discloses a method for designing tires in which a self-organizing map is used to cluster the resulting product based on certain material information, and the original material information is assigned to the clusters, thereby providing material information assigned to the clusters based on desired material properties. However, the method disclosed in Patent Document 2 cannot provide material information when the target composition does not fall into any existing cluster, i.e., when it is located in a blue ocean area.

[0018] (This Embodiment) Figure 1 is a block diagram showing an example of the configuration of the development support device 1 according to this embodiment.

[0019] Development support device 1 is a device that supports the development of a target composition (hereinafter referred to as the target composition). Development support device 1 comprises, as hardware, a processor 11, memory 12, storage 13, input device 14, display device 15, and communication device 16. Development support device 1 may be a PC (Personal Computer) or a server device. Alternatively, the functions of development support device 1 may be provided as a so-called cloud service.

[0020] The processor 11 works in cooperation with the memory 12 to execute a program, thereby realizing the functions of the development support device 1. Details of the functions will be explained as appropriate. The processor 11 may also be read as a CPU (Central Processing Unit), controller, or control unit. Furthermore, the processor 11 may include a GPU (Graphics Processing Unit) and / or an NPU (Neural Processing Unit).

[0021] The memory 12 is composed of a volatile storage medium (e.g., RAM (Random Access Memory)) and / or a non-volatile storage medium (e.g., ROM (Read Only Memory)) and stores programs and data.

[0022] The storage 13 is composed of a non-volatile storage medium and stores programs and data. Examples of the storage 13 include SSDs (Solid State Drives), HDDs (Hard Disk Drives), or flash memory.

[0023] The input device 14 is, for example, a keyboard, mouse, touchpad, and / or microphone, and accepts input from developers, etc.

[0024] The display device 15 is, for example, a liquid crystal display or an organic EL display, and displays images and information.

[0025] The communication device 16 connects to a predetermined communication network (not shown) and controls data communication with other devices (e.g., smartphones, tablet terminals, and / or other server devices). Examples of communication networks include wired LANs (Local Area Networks), wireless LANs, the Internet, and / or mobile communication networks.

[0026] The development support device 1 has the following functions: a data set acquisition unit 21, a data analysis unit 22, a map generation unit 23, a target setting unit 24, an inverse analysis unit 25, and a target composition information display unit 26. Details of the functions will be explained as appropriate.

[0027] <About the Composition> The composition of this disclosure is not particularly limited, but may be an industrial product itself or a raw material for an industrial product. Examples of industrial products include food and beverages, cosmetics, household goods, pharmaceuticals, textile products, petrochemical products, rubber products, general machinery, machine tools, and transportation machinery. Examples of raw materials for industrial products include processed food ingredients, flavors, and fragrances. In this embodiment, the case where the target composition is a flavor or a fragrance will be described. However, the target composition is not limited to these.

[0028] <About the Target Category> The developer uses the development support device 1 to develop a new composition (target composition) belonging to the target category. The target category refers to the group of products to which the composition to be developed belongs. The product may be a composition consisting of multiple features. The features may further consist of multiple features. The target category of this disclosure is not particularly limited, but food and beverages or cosmetics are particularly preferred.

[0029] Examples of food products include dairy products, confectionery, oils and fats products, processed agricultural products, seasonings, soups, processed livestock products, processed seafood products, and other foods. Examples of dairy products include butter, cheese, cheese foods, and other dairy products. Examples of confectionery include frozen desserts, Japanese sweets, Western sweets, desserts, baked goods, bakery items, candies, and gums. Examples of oils and fats products include margarine, coffee whiteners, and chocolates. Examples of processed agricultural products include noodles, processed vegetable protein products, jams, pastes, pickles, canned agricultural products, fruit juices, and processed fruit pulp products. Examples of seasonings include miso, soy sauce, sauces, mayonnaise, and dressings. Examples of soups include powdered soups, retort pouch soups, and canned soups. Examples of processed livestock products include ham, sausages, hamburgers, and canned meats. Examples of processed seafood products include fish ham, fish sausage, processed seafood products, and canned seafood. Examples of other foods include frozen foods, retort foods, instant foods, livestock feed, fish feed, pet feed, and oral care products. Examples of raw materials for these products include agricultural products such as milk, grains, edible oils and fats, fruits and vegetables, livestock products, and seafood.

[0030] Examples of beverages include carbonated drinks, fruit drinks, vegetable drinks, beers, non-alcoholic beers, non-alcoholic beverages, chuhai (a type of Japanese alcoholic beverage), other alcoholic beverages, tea drinks, coffee drinks, functional drinks, cocoa drinks, cacao drinks, sugar-free drinks, sports drinks, nutritional drinks, milk, lactic acid bacteria drinks, milk drinks, and vinegar drinks. Examples of tea drinks include tea leaves and tea leaf extracts used as raw materials for green tea, black tea, and oolong tea. Examples of coffee drinks include coffee extract and instant coffee. The form of beverages is not particularly limited and includes ready-to-drink (RTD) beverages in cans or PET bottles that can be consumed immediately after purchase, powdered beverages that are dissolved in water or hot water, and products that require brewing before consumption, such as tea drinks and coffee drinks. Food and beverages also include food additives.

[0031] Examples of food additives include flavors (synthetic flavors, natural flavors, blended flavors), sweeteners (acesulfame potassium, stevia, erythritol, etc.), acidulants (citric acid, tartaric acid, phosphoric acid, lactic acid, etc.), bittering agents (caffeine, naringin, etc.), seasonings (monosodium glutamate, L-arginine, etc.), colorants (chlorophyllin, grape skin pigment, safflower red pigment, Food Red No. 102, copper chlorophyllin sodium, etc.), preservatives (benzoic acid, sorbic acid, etc.), thickeners and stabilizers (carrageenan, xanthan gum, guar gum, dextran, pullulan, etc.), emulsifiers (quillaja extract, enzyme-hydrolyzed lecithin, glycerin fatty acid ester, sucrose fatty acid ester, polysorbate 60, etc.), and processing aids (sodium bicarbonate, potassium carbonate, magnesium carbonate). Examples of additives usable in food and beverages include: nesium, antioxidants (catechin, quercetin, tea extract, green coffee bean extract, d-α-tocopherol, calcium disodium ethylenediaminetetraacetate, etc.), color fixatives (potassium nitrate, sodium nitrate, etc.), glazing agents (carnauba wax, lanolin, paraffin wax, etc.), bleaching agents (calcium acetate, sodium hyposulfite, etc.), enzymes (lipase, pectinase, polyphenol oxidase, protease, etc.), excipients (dextrin, gum arabic, corn starch, modified starch, etc.), spices (mint, pepper, perilla, garlic, ginger, etc.), inorganic salts (sodium chloride, potassium chloride, etc.), and other flavor enhancers (flavor-enhancing peptides, fruit juice-derived fractions, etc.).

[0032] Examples of cosmetics in this disclosure include, but are not limited to, fine fragrances (such as parfums, extrait de parfums, eau de parfums, parfums de toilettes, eau de toilettes, eau de colognes, body splashes, aftershaves, and body mists containing baby colognes, etc., aqueous alcohol solutions of perfume oils), cosmetics (basic cosmetics, finishing cosmetics, sunscreens, sunscreens, medicated cosmetics), and personal care products (body washes, shampoos, rinses, conditioners, creams, body lotions, moisturizers, shower gels, soaps, hand soaps, toothpastes, etc.). Examples include polishing agents, mouthwashes and other oral care products, hair fixing aids, styling products, body odor deodorants, antiperspirants, etc.), fabric care products (laundry detergents, fabric softeners, bleaches, dryer sheets, scented beads, etc.), home care products (kitchen detergents, bathroom cleaners, floor cleaners, window cleaners, and other cleaning agents), deodorizers (carpet deodorizers, room deodorizers, pet deodorizers, etc.), air care products (candles, aerosols, air fresheners, room sprays, room mists, etc.), bath additives (powdered bath additives, solid effervescent bath additives, bath oils, bubble baths, bath salts, etc.), and repellents.

[0033] Among these, shampoos, laundry detergents, bath salts, dish soaps, hair fixing agents, styling products, body washes, body lotions, colognes, powder foundations, etc., bathroom cleaners, floor cleaners, window cleaners, air fresheners, deodorizers (carpet deodorizers, room deodorizers, etc.), and window cleaners are particularly preferred. Ingredients for cosmetics such as fragrances are also included in the target category.

[0034] While there are no particular limitations on the categories of samples to be included, limiting the categories allows for more detailed estimations. The samples to be included are preferably food and beverage products and cosmetics belonging to the same category, and more preferably their raw materials, such as flavors and fragrances. Furthermore, while there are no particular limitations on the number of samples to be included, it is preferable to select two or more types.

[0035] <Regarding the dataset> A plurality of datasets 40 may be stored in advance in the storage 13. The dataset acquisition unit 21 may acquire a target dataset that matches the target category from the plurality of datasets 40.

[0036] The target dataset is not particularly limited, but preferably represents the characteristics of the composition of the target category. Specifically, examples include instrumental analysis data representing the characteristics of the composition, prescription data, sensory evaluation data, survey data, physiological and psychological data, composition manufacturing parameter data, and the like. The number of types of these data is not particularly limited, but the developer can also select multiple types.

[0037] The instrumental analysis data is data obtained by performing analysis using an analytical instrument. The analytical instrument can be arbitrarily selected according to the purpose. Examples of analytical instruments include mass spectrometers, nuclear magnetic resonance apparatuses, high-performance liquid chromatographs, gas chromatographs, infrared spectrometers, near-infrared spectrometers, ultraviolet-visible spectrometers, X-ray diffractometers, electron microscopes, flow cytometers, thermal analyzers, and the like.

[0038] Particularly preferred are data obtained by chromatography. In the case of gas chromatography, specifically, as the column, non-polar columns (such as DB-5 column, RTX-5 column, SPB-1 column, etc.), polar columns (such as HP-Innowax column, DB-Wax column, SPB-20 column, etc.), columns for isomer separation (such as Chiralpak AD column, Chiralcel OD column), etc. can be mentioned. As the carrier gas, helium, nitrogen, hydrogen, argon, etc. can be mentioned. As the detector, flame ionization detector, mass spectrometer, thermal conductivity detector, electron capture detector, ultraviolet-visible light detector, fluorescence detector, etc. can be mentioned.

[0039] In the case of liquid chromatography, specifically, examples of the mobile phase include water, methanol, acetonitrile, ethanol, hexane, etc. Examples of the stationary phase include silica gel, C18 reverse phase, aminopropyl stationary phase, polymer base, ion exchange resin, etc. Examples of the column include HPLC column, GC column, TLC plate, size exclusion column, ion exchange column, etc. Examples of the detector include UV-Vis detector, fluorescence detector, mass spectrometer, conductivity detector, thermal conductivity detector, etc. Examples of the sample pretreatment method include solid phase extraction, liquid-liquid extraction, filtration, thin layer chromatography, concentration, etc.

[0040] Examples of the indicators dealt with in the peak analysis of the chromatogram include peak area, peak height, peak retention time, peak width, peak resolution, etc. The developer can obtain data representing the characteristics of the product by combining these elements according to the target category and / or development purpose.

[0041] The prescription data of the present disclosure are data representing the combination and / or concentration of raw materials in the composition. The form of the prescription data can be arbitrarily selected according to the purpose. Specific examples of the prescription data include data created in advance based on the presence or absence, concentration, amount, relative amount, summary statistic, etc. of the raw materials contained in the target sample. In addition, the developer can select any element (for example, fragrance composition, amount and type of fruit juice raw material, sugar content, amount of alcohol, etc.) from the components of the composition and use the information of any element as it is for the prescription data. Further, when any element is a composition, the developer can further extract any element (fruit juice raw material composition (production area and content ratio, etc.) and / or fragrance composition (representative compound and its content)) from the composition and use it as a data source for the prescription data.

[0042] The sensory evaluation data in this disclosure includes sensory inspection data and sensory test data, and refers to data obtained by evaluating a group of compositions in the target category using human senses (sight, hearing, taste, smell, and touch). Regarding the sensory evaluation method, any method suitable for evaluating the target sample can be arbitrarily selected, and there are no particular restrictions. Examples of sensory evaluation methods include discrimination tests such as the two-point discrimination method and the three-point discrimination method, scoring methods, descriptive tests such as flavor profiling and QDA (Quantitative Descriptive Analysis), rapid methods such as CATA (Check All That Apply), napping, flash profiling, and ultra-flash profiling, TI (Time Intensity) method, TDS (Temporary Dominance of Sensations), TCATA (Temporary Check All That Apply), TDL (Temporary Drivers of Liking) method, or preference-based sensory evaluation. The expertise of the evaluator is not particularly limited. Developers can examine in detail the relative relationships in a sample's composition by selecting data obtained from descriptive tests conducted by experts in the target category. Furthermore, by having general consumers conduct evaluations, developers can directly examine the perceived value of the product from a consumer's perspective.

[0043] The survey data disclosed herein refers to data obtained using methods such as questionnaire surveys, interview surveys, focus group surveys, observational surveys, experimental surveys, purchase data surveys, segmentation analysis surveys, word-of-mouth surveys, diary surveys, and online panel surveys. Sensory evaluation is not necessarily required to be conducted simultaneously when acquiring survey data. Regarding the survey method, it is possible to arbitrarily select a method suitable for the target sample. Conducting surveys targeting consumers in the target category allows for a more detailed examination.

[0044] The psychological data disclosed herein is not particularly limited. Examples of psychological data include behavioral data, physiological measurement data, psychological test data, self-report data, social network data, life history data, and cognitive data. Psychological data that is particularly preferred includes physiological measurement data, behavioral data, psychological test data, self-report data, and cognitive data. Examples of physiological measurement data include data obtained by measuring indicators such as central nervous system activity, autonomic nervous system activity, and immune and endocrine system activity. Further examples of physiological measurement data include heart rate, heart rate variability, skin electrical activity, electroencephalogram, blood pressure, respiratory rate, electromyogram, blood oxygen saturation, blood glucose level, biomarkers such as stress hormones, blood lactate concentration, pupillary response, cerebral blood flow, and blood antibody levels. Examples of behavioral data include behavioral observation data, behavioral recording data, behavioral field experiment data, social interaction data, and reaction time data. Examples of psychological test data include scores from standardized psychological tests such as POMS (Profile of Mood States), PANAS (Positive and Negative Affect Schedule), STAI (State-Trait Anxiety Inventory), MMPI (Minnesota Multiphasic Personality Inventory), Big Five personality tests, WAIS (Wetchler Adult Intelligence Scale), TAT (Thematic Apperception Test), and BDI (Beck Depression Inventory). Examples of self-report data include questionnaire data on emotions, thoughts, and behaviors, psychological scale data, diary data, journal data, interview data, emotion records, and self-report data on health status. Examples of cognitive data include cognitive ability test data, attention test data, cognitive style assessment data, memory test data, cognitive bias test data, and problem-solving ability data. Further examples of cognitive data include scores from the N-back test, Stroop test, sustained attention test, free recall test, memory span test, bias measurement survey, and logical reasoning test.

[0045] The manufacturing parameter data for a composition is not particularly limited and refers to all parameters that affect the quality of the composition. Examples of data for food composition manufacturing include heating temperature, heating time, manufacturing batch size, stirring speed, extraction temperature, extraction solvent composition, concentration temperature, concentration method, concentration time, drying time, drying temperature, sterilization method, sterilization temperature, and sterilization time.

[0046] In this disclosure, it is preferable to use a dataset to be analyzed (for example, the first dataset described later) that includes at least one type of data selected from sensory evaluation data and survey data, and at least one type of data selected from instrumental analysis data, prescription data, and physiological measurement data.

[0047] <Dimensionality Reduction Method> The data analysis unit 22 performs dimensionality reduction on the target dataset acquired by the dataset acquisition unit 21 using a predetermined dimensionality reduction method.

[0048] In this disclosure, the data analysis method for creating a feature space after dimensionality reduction refers to a data analysis method that compresses a high-dimensional space into a lower-dimensional space, and a data analysis method that allows for the matching of the relationship between a group of compositions and their features. Even with a complex data structure, by reducing it to a lower dimension using a new comprehensive index, the relative position of the compositions and the features related to those compositions can be confirmed. There are no particular restrictions on the data analysis methods that can be used, and any method can be selected according to the purpose. For example, data analysis methods include PCA (Principal Component Analysis), MFA (Multiple Factor Analysis), CCA (Canonical Correlation Analysis), GCCA (Generalized Canonical Correlation Analysis), ICA (Independent Component Analysis), Autoencoder, Kernel Principal Component Analysis, and NMF (Nonnegative Matrix Factorization). Of these data analysis methods, PCA, MFA, CCA, or GCCA are particularly preferred. Furthermore, before implementing these data analysis methods, preprocessing such as normalization, standardization, logarithmization, Box-Cox transformation, or Yeo-Johnson transformation may be performed on the numerical values. After creating the feature space after dimensionality reduction, only the necessary features may be selected, and the feature space may be created again.

[0049] <Setting and recommending target compositions> The map generation unit 23 visualizes the analysis results (map 50 (see Figure 5, etc.)) that have been dimensionally compressed and analyzed by the data analysis unit 22.

[0050] The dimensional configuration of map 50 is not particularly limited, but two, three, or four dimensions are preferred, and two or three dimensions are even more preferred. The visualization method of map 50 is not particularly limited, but a biplot in two or three-dimensional space (a method of visualizing the relationship between observed values ​​and variables at once) is preferred. Furthermore, the map generation unit 23 may visualize the map in a manner that allows for the distinction of colors and shapes based on subcategories of the composition group or subcategories of feature quantities.

[0051] The target setting unit 24 searches for areas in the feature space that are estimated to be blue oceans based on the relative positions and / or feature relationships of compositions in the map 50 (e.g., the first map 51A) generated by the map generation unit 23, and visualizes these areas using circles or different colors (e.g., the dotted ellipse 101 in Figure 5), thereby enabling the setting of target compositions. Furthermore, the map generation unit 23 visualizes sales, preferences, and / or cluster information together with the map 50 (first map 51A), allowing developers to strategically set target compositions.

[0052] In this way, by visualizing as Map 50, developers can understand the relative positions of multiple compositions, which are represented by multidimensional and complex features, from a macro perspective, and appropriately set target compositions for development. Conventionally, development proceeds by focusing on some of the features of such a complex composition, and even if a composition of the intended quality can be created for those features, the quality of other features may not be as intended. In other words, conventionally, it has been difficult to create a composition of sufficiently satisfactory quality as a whole.

[0053] Furthermore, the development support device 1 according to this embodiment may quantify the relative positional relationship of the compositions in the feature space after dimensionality reduction, the relationship between the feature quantities, and the target composition.

[0054] Furthermore, in this embodiment, the developer may input various conditions for setting the target composition (for example, quantifying whether it is a blue ocean market), and the development support device 1 may output design information of a target composition that matches the conditions as a recommendation. Examples of conditions for setting the target composition include the coordinate point of the target composition where the sum of the distances from the positions of each composition in the feature space is maximized, or the coordinate point where the objective variable has the optimal value. However, the conditions for setting the target composition are not limited to these. In this case, optimization methods can be applied to search for the coordinate point of the target composition. Examples of optimization methods include gradient descent, Newton's method, particle swarm optimization, genetic algorithms, simulated annealing, Bayesian optimization, constrained optimization, integer programming, local search algorithms, and field-dependent optimization. However, the optimization methods are not limited to these.

[0055] <Inverse Analysis> Inverse analysis means estimating the input from the output, the opposite of ordinary causality where the output is derived from the input. In contrast to the relationship between the relative position and features of a sample obtained by a dimensionality reduction method (output) from a dataset of composition information (input), the inverse analysis unit 25 estimates the dataset of composition information (input) from the relationship between the relative position and features of a sample (output). The inverse analysis method is a method that utilizes parameters obtained during the dimensionality reduction process, and in particular, it is preferable to perform inverse analysis by calculating the transpose matrix of the loading matrix of each dimension obtained when transforming to a lower dimension in the analysis method of the dimensionality reduction method, and taking the product with the position information on the feature space after dimensionality reduction, thereby estimating the original dimensional dataset (input) related to composition information from the relative position of the target composition on the feature space after dimensionality reduction (output). However, there are no particular restrictions on the inverse analysis method, and any method can be selected. In this disclosure, the dataset may consist of data composed of multiple rows or data composed of a single row.

[0056] <Visualization of Target Composition Information> The target composition information display unit 26 visualizes the estimated values ​​regarding the target composition information obtained by the inverse analysis unit 25 in a format useful to the composition developer. Examples of methods for visualizing the target composition information include tables, bar graphs, line graphs, pie charts, scatter plots, bubble charts, radar charts, heat maps, etc. However, the methods for visualizing the target composition information are not limited to these. Furthermore, two or more of these methods for visualizing the target composition information may be selected. It is preferable that the visualization of the target composition information be modified in a format useful to the composition developer.

[0057] <Data Update> In this disclosure, data update means, without any particular limitations, when the number of samples belonging to the target category increases, performing analysis processing again using a dimensionality reduction method based on the increased data, and updating the output regarding composition information obtained by inverse analysis of the analysis results. By performing data updates, it is possible to provide estimated values ​​of information that take marketability into consideration. Furthermore, by accumulating updated information, it is possible to provide estimated values ​​of design information that take into account the time-series changes in the market.

[0058] <Development Support Method> Figure 2 is a flowchart showing an example of a development support method according to this embodiment. The development support device 1 may execute the processes related to the development support method shown below.

[0059] The dataset acquisition unit 21 acquires at least one dataset belonging to the target category (hereinafter referred to as the first dataset) from the multiple datasets 40 stored in the storage 13 (S11).

[0060] The data analysis unit 22 performs data analysis and dimensionality reduction on the first dataset using a predetermined dimensionality reduction method, and calculates the first feature space after dimensionality reduction (S12). This flowchart describes an example in which PCA (Principle Component Analysis) is used as the data analysis method, but other data analysis methods may be used.

[0061] The data analysis unit 22 calculates the principal component score of each composition included in the first dataset using PCA (S13).

[0062] The data analysis unit 22 calculates the principal component loadings of each feature included in the first dataset using PCA (S14).

[0063] The map generation unit 23 plots each composition in the first feature space after dimensionality reduction calculated in step S12 based on the principal component scores calculated in step S13, and plots each feature based on the principal component loadings calculated in step S14, thereby generating a first map 51A (see Figure 5) (S15). The map generation unit 23 may display the generated first map 51A on the display device 15.

[0064] The target setting unit 24 sets the target composition (target composition T) at a predetermined position on the first map 51A (S16). For example, the developer looks at the first map 51A and sets the target composition T in an area where there are not many other compositions (for example, the area of ​​the dotted ellipse 101). Alternatively, the target setting unit 24 searches the first map 51A for an area where there are not many other compositions and automatically sets the target composition T in that area (for example, the area of ​​the dotted ellipse 101). This is because an area where there are not many other compositions has a high probability of being a blue ocean with little competition. Note that the position where the target composition T is set is not limited to this and may be any position.

[0065] The inverse analysis unit 25 performs an inverse analysis on the set target composition T and calculates the estimated feature quantities of each feature contained in the target composition T (S17). Details of the inverse analysis method will be described later.

[0066] The dataset acquisition unit 21 generates a second dataset (S18) by adding the target composition T, which includes estimated feature quantities for each feature calculated in step S17, to the first dataset.

[0067] The data analysis unit 22 performs data analysis and dimensionality reduction on the second dataset using a predetermined dimensionality reduction method (e.g., PCA), and calculates the second feature space after dimensionality reduction (S19).

[0068] The data analysis unit 22 calculates the principal component score of each composition (including the target composition T) included in the second dataset using PCA (S20).

[0069] The data analysis unit 22 calculates the principal component loadings of each feature included in the second dataset using PCA (S21).

[0070] The map generation unit 23 plots each composition (including the target composition T) based on the principal component score calculated in step S20, and plots each feature based on the principal component loading calculated in step S21, in the second feature space after dimensionality reduction calculated in step S19, thereby generating a second map 51B (see Figure 9) (S22). The map generation unit 23 may display the generated second map 51B on the display device 15. The map generation unit 23 may also display the first map 51A, on which the target composition T set in step S16 is plotted, and the second map 51B, on which the target composition T calculated in step S20 is plotted, in a manner that allows for comparison. For example, the map generation unit 23 may display the first map 51A and the second map 51B simultaneously. Alternatively, the map generation unit 23 may superimpose and plot the target composition T set in step S16 on the second map 51B on which the target composition T calculated in step S20 is plotted.

[0071] The developer confirms the position of the target composition T in the first map 51A and the second map 51B (S23). If there is no significant discrepancy in the position of the target composition T in the first map 51A and the second map 51B (for example, if the distance of the discrepancy is less than a predetermined threshold), the developer can determine that the accuracy of the inverse analysis in step S17 is sufficient. Therefore, the developer can efficiently create a target composition T of the desired quality by composing the target composition T based on the estimated feature quantities of each feature of the target composition T obtained in the inverse analysis in step S17, saving the effort of trial and error.

[0072] The above-mentioned process will be explained below with specific examples. However, the following examples are not limiting to this embodiment.

[0073] (Example 1) Example 1 describes the case where PCA is used as the dimensionality reduction method.

[0074] <Preparation of the first dataset> Figure 3 is a table showing an example of the first dataset 41A according to Example 1.

[0075] In the table of the first dataset 41A shown in Figure 3, the numbers in the column labeled "Composition" indicate the identification number of the composition, and the numbers in columns a, b, c, d, and e of the column labeled "Features" indicate the feature quantities of each of the five types of features contained in the composition. Each feature may also be a component contained in the composition.

[0076] For example, in the first dataset 41A shown in Figure 3, the row for composition "1" indicates that in composition "1", the feature quantity of feature "a" is "5.86", the feature quantity of feature "b" is "3.18", the feature quantity of feature "c" is "4.17", the feature quantity of feature "d" is "4.25", and the feature quantity of feature "e" is "5.85". Feature quantities may be relative values ​​between features.

[0077] In step S11 shown in Figure 2, the dataset acquisition unit 21 acquires the first dataset 41A shown in Figure 3.

[0078] <Dimensionality Reduction and Parameter Calculation> Figure 4 shows the two-dimensional principal component loading matrix L according to Example 1.

[0079] In step S12 shown in Figure 2, the data analysis unit 22 analyzes the first dataset 41A (see Figure 3) acquired by the dataset acquisition unit 21 using PCA, a dimensionality reduction method, to obtain a first feature space that has been dimensionally reduced to two dimensions. In other words, the data analysis unit 22 analyzes the first dataset 41A using PCA and obtains the first principal component and the second principal component of the dimensionally reduced first feature space.

[0080] In step S13 shown in Figure 2, the data analysis unit 22 obtains principal component scores for the first principal component and the second principal component of each composition (1 to 20) in the first dataset 41A. The principal component scores for the first principal component and the second principal component of each composition can be represented by a two-dimensional principal component score matrix Z.

[0081] In step S14 shown in Figure 2, the data analysis unit 22 obtains principal component loadings for the first principal component and the second principal component of each feature (a to e) of the first dataset 41A. The principal component loadings (PC1, PC2) for the first principal component and the second principal component of each feature can be represented by a two-dimensional principal component loading matrix L, as illustrated in Figure 4.

[0082] <Display of the First Map> Figure 5 is a diagram showing an example of the first map 51A in which each composition and each feature are simultaneously biploted in the first feature space according to Example 1.

[0083] The map generation unit 23 generates a first map 51A with the first principal component on the horizontal axis and the second principal component on the vertical axis. The map generation unit 23 may also display the proportion of the first principal component on the horizontal axis and the proportion of the second principal component on the vertical axis in the first map 51A.

[0084] In step S15 shown in Figure 2, the map generation unit 23 identifies the position of each composition on the first map 51A based on the principal component score matrix Z of each composition and plots it on the first map 51A. In the first map 51A shown in Figure 5, the circles indicate the position of each composition, and the numbers near the circles correspond to the composition identification numbers (1 to 20) in the first dataset 41A shown in Figure 3.

[0085] In step S15 shown in Figure 2, the map generation unit 23 identifies the position of each feature (a to e) on the first map 51A based on the principal component loading matrix L of each feature, and plots it on the first map 51A. In the first map 51A shown in Figure 5, the plus signs indicate the position of each feature (a to e), and the letters near the plus signs correspond to the type of feature (a to e).

[0086] The map generation unit 23 may display the generated first map 51A on the display device 15, as shown in Figure 5.

[0087] <Setting the Target Composition> The developer may look at the first map 51A shown in Figure 5 and set the target composition T in an area where there are few compositions (for example, the area of ​​the dotted ellipse 101). This is because an area with few compositions is likely to be a blue ocean with little competition. For example, in step S16 shown in Figure 2, the developer inputs the target composition T to the coordinate point p(1,1) of the first map 51A shown in Figure 5 through the input device 14, and the target setting unit 24 sets the target composition T to the input coordinate point p.

[0088] <Inverse Analysis> Figure 6 shows the transpose matrix L of the two-dimensional principal component loading matrix L according to Example 1. ｔ This figure shows the estimated feature quantities of each feature contained in the target composition T according to Example 1. Figure 8 is a radar chart showing the estimated feature quantities of each feature contained in the target composition T according to Example 1.

[0089] In step S17 shown in Figure 2, the inverse analysis unit 25 performs an inverse analysis of the coordinate point p(1,1) set in the first map 51A by, for example, the following process, and calculates the estimated feature quantities of each feature of the target composition T.

[0090] First, the inverse analysis unit 25 converts the two-dimensional principal component loading matrix L shown in Figure 4 into the transpose matrix L shown in Figure 6. ｔ Next, the inverse analysis unit 25 calculates the transpose matrix L using the following formula. ｔ Using this method, inverse analysis is performed on the coordinate point p(1,1) to calculate the estimated feature quantities for each feature of the target composition T, as shown in Figure 7.

[0091] T = pL ｔ

[0092] Furthermore, if standardization is performed during dimensionality reduction, the inverse analysis unit 25 may adjust the estimated feature quantities of the target composition T calculated using the above formula using the feature quantities and standard deviations in the first dataset 41A shown in Figure 3.

[0093] Furthermore, the target composition information display unit 26 may display the estimated feature quantities of each feature of the target composition T in a radar chart as shown in Figure 8.

[0094] <Display of the second map> Figure 9 is a diagram showing an example of the second map 52A in which the target composition T is plotted in the second feature space according to Example 1.

[0095] In step S18 shown in Figure 2, the dataset acquisition unit 21 generates a second dataset (not shown) by adding the target composition T estimated by the inverse analysis unit 25 to the first dataset 41A shown in Figure 3.

[0096] Then, the development support device 1 analyzes the second dataset using PCA by the processes shown in steps S19 to S22 of Figure 2, and generates and displays a second map 52A corresponding to the second feature space, as shown in Figure 9. The map generation unit 23 may display the first map 51A and the second map 52A simultaneously on the display device 15. As shown in Figure 9, the target composition T is plotted on the second map 52A.

[0097] In step S23 shown in Figure 2, the developer compares the position of the target composition T (coordinate point p) set on the first map 51A with the position of the target composition T plotted on the second map 52A to confirm whether the two positions are similar. If the two positions are similar, the developer can determine that the accuracy of the estimated feature quantities for each feature of the target composition T calculated by inverse analysis is sufficient. Therefore, by creating the target composition T according to the estimated feature quantities for each feature (a to e) shown in Figure 7, the developer can efficiently create a target composition T with the intended quality, which differs in quality from the other 20 compositions (1 to 20) (see Figure 3) belonging to the same category, while saving the effort of trial and error.

[0098] <First Modification of Example 1: Recommendation of Target Composition Using the Maximum Value of Principal Component Score> Above, we described a case in which the developer sets the target composition T by looking at the first map 51A. Below, we will describe a case in which the development support device 1 automatically recommends the target composition T.

[0099] Figure 10 is a table showing the estimated feature quantities of each feature contained in the target composition T according to the first modification of Example 1. Figure 11 is a diagram showing an example of the second map 52B plotted with the target composition T according to the first modification of Example 1.

[0100] The target setting unit 24 automatically sets the maximum absolute value of the principal component scores for each dimension in the first map as the recommended target composition T. In other words, the target setting unit 24 sets the recommended target composition T in a blue ocean area with little competition. For example, the target setting unit 24 sets the recommended target composition T at coordinate point p(2.81, -2.91). In this case, the map generation unit 23 does not need to display the first map on the display device 15.

[0101] The inverse analysis unit 25 performs inverse analysis on the coordinate point p(2.81, -2.91) as described above, and calculates the estimated feature quantities of each feature of the target composition T as shown in Figure 10.

[0102] The development support device 1 generates and displays a second map 52B as shown in Figure 11, using the same process as described above, as shown in steps S18 to S22 of Figure 2, for the target composition T shown in Figure 10. As shown in Figure 11, the recommended target composition T is plotted on the second map 52B.

[0103] Developers can refer to the second map 52B shown in Figure 11 to confirm the relative position of the recommended target composition T in relation to other compositions. Furthermore, by creating the target composition T according to the estimated feature quantities of each feature (a to e) shown in Figure 10, developers can efficiently create a target composition T with the intended quality, which differs in quality from the 20 other compositions (1 to 20) (see Figure 3) belonging to the same category, while saving the effort of trial and error.

[0104] <Second Modification of Example 1: Recommendation of Target Composition Using Optimization> In the above, we described a case in which the developer sets the target composition T by looking at the first map 51A. Below, we will describe a second modification in which the development support device 1 automatically recommends the target composition T.

[0105] Figure 12 is a table showing the estimated feature quantities of each feature contained in the target composition T according to the second modification of Example 1. Figure 13 is a diagram showing an example of the second map 52C plotted with the target composition T according to the second modification of Example 1.

[0106] The target setting unit 24 searches for a coordinate where the sum of the distances from all principal component scores is maximized, within the range of the minimum and maximum values ​​for each dimension of all principal component scores in the first map, using an optimization method, and automatically sets the searched coordinate as the recommended target composition T. For example, the target setting unit 24 sets the coordinate point p(-0.51, 0.05) as the recommended target composition T. In this case, the map generation unit 23 does not need to display the first map on the display device 15.

[0107] The inverse analysis unit 25 performs inverse analysis on the coordinate point p(-0.51, 0.05) as described above, and calculates the estimated feature quantities of each feature of the target composition T as shown in Figure 12.

[0108] The development support device 1 generates and displays a second map 52C as shown in Figure 13, using the same process as described above, as shown in steps S18 to S22 of Figure 2, for the target composition T shown in Figure 12. As shown in Figure 13, the recommended target composition T is plotted on the second map 52C.

[0109] Developers can refer to the second map 52C shown in Figure 13 to confirm the relative position of the recommended target composition T in relation to other compositions. Furthermore, by creating the target composition T according to the estimated feature quantities of each feature (a to e) shown in Figure 12, developers can efficiently create a target composition T with the intended quality, which differs in quality from the 20 other compositions (1 to 20) belonging to the same category (see Figure 3), while saving the effort of trial and error.

[0110] (Example 2) Example 2 describes the case in which MFA is used as the dimensionality reduction method.

[0111] <Preparation of Dataset> Figure 14 is a table showing an example of the first dataset 41B according to Example 2.

[0112] The first dataset 41B shown in Figure 14 consists of 20 compositions (1-20), each possessing five feature quantities (a-e) and four feature quantities (A-D). Features (a-e) and features (A-D) may be from different fields. The interpretation of the first dataset 41B is the same as in Figure 3, so an explanation is omitted here.

[0113] In step S11 shown in Figure 2, the dataset acquisition unit 21 acquires the first dataset 41B shown in Figure 14.

[0114] <Dimensionality Reduction and Parameter Calculation> Figure 15 shows the two-dimensional factor loading matrix L according to Example 2.

[0115] In step S12 shown in Figure 2, the data analysis unit 22 analyzes the first dataset 41B (see Figure 14) acquired by the dataset acquisition unit 21 using the dimensionality reduction method MFA to obtain a first feature space that has been dimensionally reduced to two dimensions. In other words, the data analysis unit 22 analyzes the first dataset 41B using MFA to obtain the first and second factors in the dimensionally reduced first feature space.

[0116] In step S13 shown in Figure 2, the data analysis unit 22 obtains factor scores for the first and second factors of each composition (1 to 20) in the first dataset 41B. The factor scores for the first and second factors of each composition can be represented by a two-dimensional factor score matrix Z.

[0117] In step S14 shown in Figure 2, the data analysis unit 22 obtains the factor loadings for the first and second factors for each feature (a to e, A to D) of the first dataset 41B. The factor loadings (Dim1, Dim2) for the first and second factors of each feature can be represented by a two-dimensional factor loading matrix L, as illustrated in Figure 15.

[0118] <Display of the first map> Figure 16 is a diagram showing an example of the first map 51D in which each composition and each feature are simultaneously biploted in the first feature space according to Example 2.

[0119] The map generation unit 23 generates a first map 51D with the first factor on the horizontal axis and the second factor on the vertical axis. The map generation unit 23 may also display the proportion of the first factor on the horizontal axis and the proportion of the second factor on the vertical axis in the first map 51D.

[0120] In step S15 shown in Figure 2, the map generation unit 23 identifies the position of each composition on the first map 51D based on the factor score matrix Z of each composition and plots it on the first map 51D. In the first map 51D shown in Figure 16, the circles indicate the position of each composition, and the numbers near the circles correspond to the identification numbers (1 to 20) of the compositions in the first dataset 41B shown in Figure 14.

[0121] In step S15 shown in Figure 2, the map generation unit 23 identifies the position of each feature (a to e, A to D) on the first map 51D based on the factor loading matrix L of each feature, and plots it on the first map 51D. In the first map 51D shown in Figure 16, the plus signs indicate the position of each feature (a to e), and the lowercase letters near the plus signs correspond to the type of feature (a to e). Also, in the first map 51D shown in Figure 16, the diamond shapes indicate the position of each feature (A to D), and the uppercase letters near the diamond shapes correspond to the type of feature (A to D).

[0122] The map generation unit 23 may display the generated first map 51D on the display device 15, as shown in Figure 16.

[0123] <Setting the Target Composition> The developer may look at the first map 51D shown in Figure 16 and set the target composition T in an area where there are few existing products. This is because an area with few existing products is likely to be a blue ocean with little competition. For example, in step S16 shown in Figure 2, the developer inputs the target composition T to the coordinate point p (0.5, -0.5) of the first map 51D shown in Figure 16 via the input device 14, and the target setting unit 24 sets the target composition T at the input coordinate point p.

[0124] <Inverse Analysis> Figure 17 shows the transpose matrix L of the two-dimensional factor loading matrix L according to Example 2. ｔ This figure shows the estimated feature quantities of each feature contained in the target composition T according to Example 2. Figure 19 is a bar graph showing the estimated feature quantities of each feature contained in the target composition T according to Example 2.

[0125] In step S17 shown in FIG. 2, the inverse analysis unit 25 performs inverse analysis on the coordinate point p(0.5, -0.5) set in the first map 51D, for example, by the following process, and calculates the estimated feature amounts of each feature of the target composition T.

[0126] First, the inverse analysis unit 25 calculates the transposed matrix L shown in FIG. 17 from the two-dimensional factor loading matrix L shown in FIG. 15. ｔ Next, the inverse analysis unit 25 performs inverse analysis on the coordinate point p(0.5, -0.5) using the transposed matrix L by the following equation, and calculates the estimated feature amounts of each feature of the target composition T as shown in FIG. 18. ｔ T = pL

[0127] ｔ

[0128] When normalization is performed during dimensional compression, the inverse analysis unit 25 may adjust the estimated feature amounts of the target composition T calculated by the above equation using each feature amount and the standard deviation in the first data set 41B shown in FIG. 14.

[0129] Further, the target composition information display unit 26 may display the estimated feature amounts of each feature of the target composition T in a bar graph as shown in FIG. 19.

[0130] <Display of the second map> FIG. 20 is a diagram showing an example of a second map 52D in which the target composition T is plotted in the second feature space according to Example 2.

[0131] The data set acquisition unit 21 generates a second data set (not shown) obtained by adding the target composition T estimated by the inverse analysis unit 25 to the first data set 41B shown in FIG. 14 in step S18 shown in FIG. 2.

[0132] Then, the development support device 1 analyzes the second data set using MFA by the processes shown in steps S19 to S22 of FIG. 2, and generates and displays a second map 52D as shown in FIG. 20 corresponding to the first feature space. The map generation unit 23 may display the first map 51D and the second map 52D on the display device 15 at the same time. As shown in FIG. 20, the target composition T is plotted on the second map 52D.

[0133] In step S23 shown in Figure 2, the developer compares the position of the target composition T (coordinate point p) set on the first map 51D with the position of the target composition T plotted on the second map 52D to confirm whether the two positions are similar. If the two positions are similar, the developer can determine that the accuracy of the estimated feature quantities for each feature of the target composition T calculated by inverse analysis is sufficient. Therefore, by creating the target composition T according to the estimated feature quantities for each feature (a to e, A to D) shown in Figure 18, the developer can efficiently create a target composition T with the intended quality, which differs in quality from the other 20 compositions (1 to 20) (see Figure 14) belonging to the same category, while saving the effort of trial and error.

[0134] (Example 3) Example 3 describes the case where GCCA is used as the dimensionality reduction method. In Example 3, the first dataset 41B shown in Figure 14 will be used for the explanation.

[0135] <Preparation of Dataset> In step S11 shown in Figure 2, the dataset acquisition unit 21 acquires the first dataset 41B shown in Figure 14.

[0136] <Dimensionality Reduction and Parameter Calculation> Figure 21 shows the two-dimensional canonical weight matrix L according to Example 3.

[0137] In step S12 shown in Figure 2, the data analysis unit 22 analyzes the first dataset 41B (see Figure 14) acquired by the dataset acquisition unit 21 using the dimensionality reduction method GCCA to obtain a first feature space that has been dimensionally reduced to two dimensions. In other words, the data analysis unit 22 analyzes the first dataset using GCCA to obtain the first canonical variables and the second canonical variables in the dimensionally reduced first feature space.

[0138] In step S13 shown in Figure 2, the data analysis unit 22 obtains canonical variable scores for the first and second canonical variables of each composition (1 to 20) of the first dataset 41B. The canonical variable scores for the first and second canonical variables of each composition can be represented by a two-dimensional canonical variable score matrix Z.

[0139] In step S14 shown in Figure 2, the data analysis unit 22 obtains the canonical weights and canonical loadings for the first and second canonical variables of each feature (a to e, A to D) of the first dataset 41B. The canonical weights (Dim1, Dim2) for the first and second canonical variables of each feature can be represented by a two-dimensional canonical weight matrix L, as illustrated in Figure 21. The canonical loadings for the first and second canonical variables of each feature can be represented by a two-dimensional canonical loading matrix E.

[0140] <Display of the first map> Figure 22 is a diagram showing an example of the first map 51E in which each composition and each feature are simultaneously biploted in the first feature space according to Example 3.

[0141] The map generation unit 23 generates a first map 51E with the first canonical variable on the horizontal axis and the second canonical variable on the vertical axis. The map generation unit 23 may also display the proportion of the first canonical variable on the horizontal axis and the proportion of the second canonical variable on the vertical axis in the first map 51E.

[0142] In step S15 shown in Figure 2, the map generation unit 23 identifies the position of each composition on the first map 51E based on the canonical variable score matrix Z of each composition and plots it on the first map 51E. In the first map 51E shown in Figure 22, the circles indicate the position of each composition, and the numbers near the circles correspond to the composition identification numbers (1 to 20) in the first dataset 41B shown in Figure 14.

[0143] In step S15 shown in Figure 2, the map generation unit 23 identifies the position of each feature (a to e, A to D) on the first map 51E based on the canonical loading matrix E of each feature, and plots it on the first map 51E. In the first map 51E shown in Figure 22, the plus signs indicate the position of each feature (a to e), and the lowercase letters near the plus signs correspond to the type of feature (a to e). Also, in the first map 51E shown in Figure 22, the diamond shapes indicate the position of each feature (A to D), and the uppercase letters near the diamond shapes correspond to the type of feature (A to D).

[0144] The map generation unit 23 may display the generated first map 51E on the display device 15, as shown in Figure 22.

[0145] <Setting the Target Composition> The developer may set the target composition T in an area where there are few products, by looking at the first map 51E shown in Figure 22. This is because an area with few products is likely to be a blue ocean with little competition. For example, in step S16 shown in Figure 2, the developer sets the target composition T at coordinate point p(-0.5, 0.5) on the first map 51E shown in Figure 22 via the input device 14.

[0146] <Inverse Analysis> Figure 23 shows the transpose matrix L of the two-dimensional canonical weight matrix L according to Example 3. ｔ This figure shows the estimated feature quantities of each feature contained in the target composition T according to Example 3. Figure 25 is a pie chart showing the estimated feature quantities of each feature contained in the target composition T according to Example 3.

[0147] In step S17 shown in Figure 2, the inverse analysis unit 25 performs an inverse analysis of the coordinate point p(-0.5, 0.5) set in the first map 51E by, for example, the following process, and calculates the estimated feature quantities of each feature of the target composition T.

[0148] First, the inverse analysis unit 25 converts the two-dimensional canonical weight matrix L shown in Figure 21 to the transpose matrix L shown in Figure 23. ｔ Next, the inverse analysis unit 25 calculates the transpose matrix L using the following formula. ｔ Using this method, inverse analysis is performed on the coordinate point p(-0.5, 0.5) to calculate the estimated feature quantities for each feature of the target composition T, as shown in Figure 24.

[0149] T = pL ｔ

[0150] Furthermore, if standardization is performed during dimensionality reduction, the inverse analysis unit 25 may adjust the estimated feature quantities of the target composition T calculated using the above formula using the feature quantities and standard deviations in the first dataset 41B shown in Figure 14.

[0151] Furthermore, the target composition information display unit 26 may display the estimated feature quantities of each feature of the target composition T in a pie chart as shown in Figure 25.

[0152] <Display of the second map> Figure 26 is a diagram showing an example of the second map 52E in which the target composition T is plotted in the second feature space according to Example 3.

[0153] In step S18 shown in Figure 2, the dataset acquisition unit 21 generates a second dataset (not shown) by adding the target composition T estimated by the inverse analysis unit 25 to the first dataset 41B shown in Figure 14.

[0154] Then, the development support device 1 analyzes the second dataset using GCCA by the process shown in steps S19 to S22 of Figure 2, and generates and displays a second map 52E corresponding to the second feature space, as shown in Figure 26. The map generation unit 23 may display the first map 51E and the second map 52E simultaneously on the display device 15. As shown in Figure 26, the target composition T is plotted on the second map 52E. As shown in Figure 26, the target composition T is plotted as a star shape on the second map 52E.

[0155] In step S23 shown in Figure 2, the developer compares the position of the target composition T (coordinate point p) set on the first map 51E with the position of the target composition T plotted on the second map 52E to confirm whether the two positions are similar. If the two positions are similar, the developer can determine that the accuracy of the estimated feature quantities for each feature of the target composition T calculated by inverse analysis is sufficient. Therefore, by creating the target composition T according to the estimated feature quantities for each feature (a to e, A to D) shown in Figure 21, the developer can efficiently create a target composition T of the intended quality, which differs in quality from the other 20 compositions (1 to 20) (see Figure 14) belonging to the same category, while saving the effort of trial and error.

[0156] (Example 4) (Provision of design information for an unknown target of commercially available chocolate) Figure 27 is a diagram showing an example of an MFA map of commercially available chocolate according to Example 4. Figure 28 is a graph showing estimated values ​​of raw material compound data for Target 1 according to Example 4. Figure 29 is a graph showing estimated values ​​of subjective evaluation data for Target 1 according to Example 4. Next, referring to Figures 27 to 29, the effect of the development support device 1 according to this embodiment providing design information for an unknown target (target composition) of commercially available chocolate will be explained as Example 4.

[0157] <Preparation of Dataset> (Instrumental Analysis Data) Aromas were recovered from 17 commercially available chocolates using the FEDHS (Full Evaporation Dynamic Headspace) method and subjected to GC-MS analysis. Peak area values ​​for 68 raw material compounds that are important as design information for the fragrance composition were selected from the peaks obtained by GC-MS analysis.

[0158] (Sensory Evaluation Data) Seventeen commercially available chocolates, similar to those used for the raw material compound data, were used as samples. The samples were transferred to containers labeled with a three-digit code and evaluated in a blind manner. The sensory evaluation panel consisted of 10 experts, including flavorists and applicators, who possess knowledge of chocolate flavor. For each sample, the panel evaluated 29 flavor characteristics that are important as design information for the flavor composition, using a five-point scoring system. For each sample, the average values ​​of the 10 panelists for each flavor characteristic item were calculated and selected as data.

[0159] <Dimensionality Reduction and Parameter Calculation> The dataset acquisition unit 21 combined GC-MS peak area value data for 68 components and average value data for 29 flavor characteristic items for 17 selected commercially available chocolates to create a single dataset 40. The data analysis unit 22 analyzed this dataset 40 using the MFA method.

[0160] <Display of the first map> The map generation unit 23 created the MFA map 100 shown in Figure 27 by visualizing the factor scores of the first and second axes for the 17 samples obtained as a result of the analysis, and the loadings of the first and second axes of each feature multiplied by four, as a two-dimensional biplot (a generalized two-variable scatter plot).

[0161] <Setting the target composition> The target setting unit 24 set three arbitrary points on the MFA map 100 as Target 1, 2, and 3 as unknown targets, as shown in Figure 27.

[0162] <Inverse Analysis> The inverse analysis unit 25 performed an inverse analysis on three unknown target points to calculate estimated values ​​for raw material compound data related to the design information of the fragrance composition and estimated values ​​for sensory evaluation data. The target composition information display unit 26 then output a bar graph 110 (see Figure 28) showing the estimated values ​​for raw material compound data and a radar chart 120 (see Figure 29) showing the estimated values ​​for sensory evaluation data based on these calculation results.

[0163] <Confirmation of Development Usefulness> Instead of displaying the second map, the flavor list confirmed whether the estimated values ​​obtained through inverse analysis were useful for development. The flavor list created fragrance composition prescriptions based on experience and intuition for three arbitrary points (Target 1, 2, 3) on the MFA map 100, without considering the above estimated values.

[0164] When the coordinates on the MFA map 100 are denoted as (Dim1, Dim2), the three arbitrary points for which inverse analysis was performed were, as shown in Figure 27, Target1 (-1.58, -2.50), Target2 (-3.00, -1.00), and Target3 (-0.55, -2.50). The dotted ellipses drawn on the MFA map 100 in Figure 27 are for convenience to make the positions of the Targets easier to see. Subsequently, the flavor list compared the estimated values ​​calculated using inverse analysis for Target1, 2, and 3 with the prescriptions created by the flavor list. As a result of the comparison, the flavor list concluded that "the calculated estimated values ​​and the flavors of the created prescriptions have similar directions" and "the calculated estimated values ​​can serve as a basis for creating prescriptions." This clearly demonstrates that providing estimated target design information based on inverse analysis supports the development of fragrance compositions.

[0165] The map creation step and inverse analysis step described above were performed using R (version 4.1.2).

[0166] (Example 5) (Provision of design information for an unknown target of Japanese citrus) Figure 30 is a diagram showing an example of an MFA map 200 of Japanese citrus according to Example 5. Next, referring to Figure 30, the effect of the development support device 1 according to this embodiment providing design information for an unknown target (target composition) of Japanese citrus will be explained as Example 5.

[0167] <Preparation of Dataset> (Instrumental Analysis Data) Fourteen Japanese citrus fruits were used as samples and subjected to GC-FID and GC-MS analysis. Peak area values ​​were selected for 47 raw material compounds that are important as design information for fragrance compositions.

[0168] (Sensory Evaluation Data) Fourteen Japanese citrus fruits, similar to those used for instrumental analysis, were used as samples. The samples were transferred to containers labeled with a three-digit code and evaluated in a blinded manner. The sensory evaluation panel consisted of 12 flavorists with knowledge of citrus sensory evaluation. For each sample, the panel evaluated 10 sensory evaluation characteristics that are important as design information for the fragrance composition, using a five-point scoring system. For each sample, the average values ​​of the 12 panelists for each sensory evaluation characteristic were calculated and selected as data.

[0169] <Dimensionality Reduction and Parameter Calculation> The dataset acquisition unit 21 combined peak area value data of 47 components related to selected Japanese citrus fruits with average value data of 10 sensory evaluation feature items to create a single dataset 40. The data analysis unit 22 performed preprocessing on this dataset 40 and then analyzed it using the MFA method.

[0170] <Display of the first map> The map generation unit 23 created the MFA map 200 shown in Figure 30 by visualizing the factor scores of the first and second axes for the 17 samples obtained as a result of the analysis, and the loadings of the first and second axes of each feature multiplied by 10, as a two-dimensional biplot.

[0171] <Setting the Target Composition> The target setting unit 24 set an arbitrary point (-2.00, -0.75) on the MFA map 200 as the Target, as shown in Figure 30, as an unknown target. The dotted ellipse drawn on the MFA map 200 in Figure 30 is for convenience to make the position of the Target easier to see.

[0172] <Inverse Analysis> The inverse analysis unit 25 calculated estimated values ​​of raw material compound data and sensory evaluation data related to the design information of the fragrance composition for the set Target by performing an inverse analysis. The target composition information display unit 26 output the calculated estimated values ​​of raw material compound data and sensory evaluation data on a number line (not shown).

[0173] <Confirmation of Development Usefulness> Based on the output of estimated values ​​using inverse analysis, the flavor list was found to "serve as a starting point for formula creation." This revealed that providing estimated target design information based on inverse analysis supports the development of fragrance compositions, even for samples in different categories.

[0174] (Example 6) (Provision of design information for an unknown target of Japanese citrus) Figure 31 is a diagram showing an example of an MFA map 300 of Japanese citrus according to Example 6. Next, referring to Figure 31, the effect of the development support device 1 according to this embodiment providing design information for an unknown target (target composition) of Japanese citrus will be explained as Example 6.

[0175] <Preparation of Dataset> (Instrumental Analysis Data) In Example 6, the same data as in Example 5 was used.

[0176] (Consumer Survey Data) The same 14 Japanese citrus fruits used in the instrumental analysis data were sampled, and 676 consumers were surveyed. The panel selected all fruits that matched the image of 21 image items, such as "relaxing" and "energizing." Although the types of fruits used in the survey included non-Japanese citrus fruits, only the data from the 14 Japanese citrus fruits selected for this study was chosen and used for analysis. The selection frequency of each image item was calculated for the 14 Japanese citrus fruit samples.

[0177] <Dimensionality Reduction and Parameter Calculation> The dataset acquisition unit 21 combined the peak area value data of 47 components and the selection frequency of 21 items related to Japanese citrus fruits selected in the same manner as in Example 5 to create a single dataset 40. The data analysis unit 22 analyzed this dataset 40 using the MFA method.

[0178] <Display of the first map> The map generation unit 23 created the MFA map 300 shown in Figure 31 by visualizing the factor scores of the first and second axes for the 14 sample points obtained as a result of the analysis, and the loadings of the first and second axes of each feature multiplied by 10, as a two-dimensional biplot.

[0179] <Setting the Target Composition> The target setting unit 24 set an arbitrary point (0.50, -0.50) on the MFA map 300 as the Target, as shown in Figure 31, as an unknown target. The dotted ellipse drawn on the MFA map 300 in Figure 31 is for convenience to make the position of the Target easier to see.

[0180] <Inverse Analysis and Verification of Development Usability> The inverse analysis unit 25 performed an inverse analysis on the set unknown target to calculate estimated values ​​of raw material compound data related to the design information of the fragrance composition and estimated values ​​of consumer survey data. The calculated estimated values ​​were then used in the design of the target fragrance composition.

[0181] (Example 7) (Provision of design information for an unknown target of Japanese citrus) Figure 32 is a diagram showing an example of a PCA map 400 of Japanese citrus according to Example 7. Next, referring to Figure 32, the effect of the development support device 1 according to this embodiment providing design information for an unknown target (target composition) of Japanese citrus will be explained as Example 7.

[0182] <Preparation of Dataset> (Instrumental Analysis Data) In Example 7, the same data as in Examples 5 and 6 was used.

[0183] <Dimensionality Reduction and Parameter Calculation> The data analysis unit 22 analyzed only the raw material compound data using the PCA method.

[0184] <Display of the first map> The map generation unit 23 created the PCA map 400 shown in Figure 32 by visualizing the first and second principal component scores for the 14 samples obtained as a result of the analysis, and the principal component loadings of the first and second axes of each feature multiplied by 10, as a two-dimensional biplot.

[0185] <Setting the Target Composition> The target setting unit 24 set three arbitrary points on the PCA map 400 as Target 1, 2, and 3 as unknown targets, as shown in Figure 32. The dotted ellipses drawn on the PCA map 400 in Figure 32 are drawn for convenience to make the positions of the targets easier to see.

[0186] <Inverse Analysis and Verification of Development Usefulness> The inverse analysis unit 25 calculated estimated values ​​of raw material compound data related to the design information of the fragrance composition for the set Targets 1, 2, and 3 through inverse analysis. The target composition information display unit 26 output the calculated estimated values ​​on a number line (not shown). The three arbitrary points for performing the inverse analysis were designated as (Dim1, Dim2) on the PCA map, with Target1 being (-2.00, 4.00), Target2 being (8.00, 6.00), and Target3 being (-3.00, -0.10). After carefully examining the outputted estimated values, the Flavor List concluded that it is "useful as a starting point for creating prescriptions."

[0187] (Example 8) (Provision of design information for an unknown target of commercially available chocolate) Figure 33 is a diagram showing an example of a 3D MFA map 500 of commercially available chocolate according to Example 8. Next, referring to Figure 33, the effect of the development support device 1 according to this embodiment providing design information for an unknown target (target composition) of commercially available chocolate will be explained as Example 8.

[0188] <Preparation of Dataset> (Instrumental Analysis Data) In Example 8, the same data as in Example 4 was used.

[0189] (Sensory evaluation data) In Example 8, the same data as in Example 4 was used.

[0190] <Dimensionality Reduction and Parameter Calculation> The dataset acquisition unit 21 combined GC-MS peak area value data for 68 components of 17 selected commercially available chocolates with average value data for 29 sensory evaluation feature items to create a single dataset 40. The data analysis unit 22 analyzed this dataset 40 using the MFA method.

[0191] <Display of the first map> The map generation unit 23 created the three-dimensional MFA map 500 shown in Figure 33 by visualizing the factor scores of the first, second, and third axes for the 17 sample points obtained as a result of the analysis, and the values ​​obtained by multiplying the loadings of the first, second, and third axes of each feature by four, using a three-dimensional plot.

[0192] <Setting the Target Composition> The target setting unit 24 set an arbitrary point (-3.00, -1.00, 3.00) on the 3D MFA map 500 as the Target, as shown in Figure 33, as an unknown target. The dotted ellipse drawn on the 3D MFA map 500 in Figure 33 is drawn for convenience to make the position of the Target easier to see.

[0193] <Inverse Analysis and Verification of Development Usefulness> The inverse analysis unit 25 calculated estimated values ​​of raw material compound data and sensory evaluation data related to the design information of the fragrance composition for the set Target by performing an inverse analysis. The target composition information display unit 26 output a bar graph (not shown) showing the estimated values ​​of the raw material compound data and a radar chart (not shown) showing the estimated values ​​of the sensory evaluation data. After carefully examining the output results, the Flavor List was found to be "useful as a starting point for creating prescriptions."

[0194] (Example 9) (Provision of design information for an unknown target of commercially available chocolate) Figure 34 is a diagram showing an example of a GCCA map 600 of commercially available chocolate according to Example 9. Next, referring to Figure 34, the effect of the development support device 1 according to this embodiment providing design information for an unknown target (target composition) of commercially available chocolate will be explained as Example 9.

[0195] <Preparation of Dataset> (Instrumental Analysis Data) In Example 9, the same data as in Example 4 was used.

[0196] (Sensory evaluation data) In Example 9, the same data as in Example 4 was used.

[0197] <Dimensionality Reduction and Parameter Calculation> The dataset acquisition unit 21 combined GC-MS peak area value data for 68 components of 17 selected commercially available chocolates with average value data for 29 sensory evaluation feature items to create a single dataset 40. The data analysis unit 22 analyzed this dataset 40 using the GCCA method.

[0198] <Display of the first map> The map generation unit 23 created the GCCA map 600 shown in Figure 34 by visualizing the canonical weights of the first and second axes for the 17 sample points obtained as a result of the analysis, and the canonical loadings of the first and second axes for each feature, as a two-dimensional biplot.

[0199] <Setting the Target Composition> Next, the target setting unit 24 set an arbitrary point (-2.00, 3.00) on the GCCA map 600 as the Target, as shown in Figure 34, as an unknown target. The dotted ellipse drawn on the GCCA map 600 in Figure 34 is for convenience to make the position of the Target easier to see.

[0200] <Inverse Analysis and Verification of Development Usefulness> The inverse analysis unit 25 used the inverse analysis formula described above to calculate estimated values ​​of raw material compound data and subjective evaluation data related to the design information of the fragrance composition. The target composition information display unit 26 output a bar graph (not shown) showing the estimated values ​​of the raw material compound data and a radar chart (not shown) showing the estimated values ​​of the subjective evaluation data. After carefully examining the output, the flavor list was found to be "useful as a starting point for creating prescriptions."

[0201] (Summary of this disclosure) The contents of this disclosure above can be expressed as follows:

[0202] <Note 1> A development support method for supporting the development of a composition using a computer (e.g., development support device 1) according to one aspect of the present disclosure involves data analysis of a first dataset (41A, 41B) in which a plurality of compositions and a plurality of feature quantities contained in the compositions are associated, and a first feature space is generated by dimensionality reduction using a predetermined dimensionality reduction method. A target composition (T) is set in the first feature space, and a plurality of estimated feature quantities contained in the target composition is calculated by performing an inverse analysis on the target composition based on the position of the target composition in the first feature space and the parameters obtained by dimensionality reduction. As a result, a plurality of estimated feature quantities contained in the target composition set in the first feature space can be calculated. Therefore, the developer can save the effort of trial and error and efficiently create a target composition by creating a composition based on the calculated estimated feature quantities.

[0203] <Note 2> In the development support method described in Note 1, a first map (51A, 51D, 51E) showing the relative positional relationship of a plurality of compositions in the first feature space is displayed on a predetermined display device (15). This allows the developer to confirm the relative positional relationship of the plurality of compositions by looking at the displayed first map.

[0204] <Note 3> In the development support method described in Note 2, a position or region (e.g., dotted ellipse 101) in which the setting of the target composition in the first feature space is proposed is identified based on the relative positional relationship of the plurality of compositions in the first feature space, and the identified position or region in which the setting of the target composition is proposed is displayed on the first map. This allows the developer to visually confirm, for example, a region in which there are few other compositions (i.e., there is little competition).

[0205] <Note 4> In the development support method described in Note 2 or 3, the user (e.g., a developer) sets the target composition on the first map. This allows the user to set the target composition at any position on the first map.

[0206] <Note 5> In the development support method described in any one of Notes 1 to 4, the target composition and the plurality of estimated features are associated with each other and added to the first dataset to generate a second dataset. The second dataset is then analyzed and dimensionality is reduced using the dimensionality reduction method to generate a second feature space. A second map (52A, 52B, 52C, 52D, 52E) showing the relative positional relationship of the target composition in the second feature space is displayed on a predetermined display device. This allows the developer to view the displayed second map and confirm the position of the target composition related to the calculated estimated features in the second map.

[0207] <Note 6> In the development support method described in any one of Notes 1 to 4, a second dataset is generated by adding the target composition and the plurality of estimated features associated with it to the first dataset, the second dataset is data-analyzed, and a second feature space is generated by dimensionality reduction using the dimensionality reduction method, the target composition set for the first feature space is placed on a first map showing the relative positional relationship of the plurality of compositions in the first feature space, the target composition in the second feature space is placed on a second map showing the relative positional relationship of the plurality of compositions in the second feature space, and the positional relationship between the target composition in the first map and the target composition in the second map is displayed on a predetermined display device in a manner that allows comparison. This allows the developer to visually confirm the discrepancy between the position of the target composition in the first map and the position of the target composition in the second map. Based on this discrepancy, the developer can then confirm the accuracy of the estimated features of the calculated target composition.

[0208] <Note 7> In the development support method described in any one of Notes 1 to 6, the predetermined dimensionality reduction method is at least one of Multiple Factor Analysis (MFA), Principle Component Analysis (PCA), Generalized Canonical Correlation Analysis (GCCA), and Canonical Correlation Analysis (CCA). This enables data analysis and dimensionality reduction of the first dataset and / or the second dataset to generate a feature space.

[0209] <Note 8> In the development support method described in any one of Notes 1 to 7, the features in the first dataset relate to at least one of the following: information on the component analysis results of the composition, information on the formulation of the composition, information on the sensory evaluation results of the composition, information on the consumer survey results of the composition, information on the psychological survey results of the composition, and manufacturing information of the composition. This makes it possible to generate a feature space based on various perspectives on the composition. Therefore, it is possible to calculate estimated features of the target composition based on various perspectives.

[0210] <Note 9> In the development support method described in any one of Notes 1 to 8, the number of dimensions of the first feature space is either 2D, 3D, or 4D. This makes it possible to generate and display a 2D, 3D, or 4D first map.

[0211] <Note 10> In the development support method described in any one of Notes 1 to 9, the composition and the target composition contain at least a fragrance. This makes it possible to support the creation of a new target composition containing at least a fragrance.

[0212] <Note 11> In the development support method described in Note 10, at least two of the plurality of features in the first dataset relate to information on the component analysis results of the composition and information on the sensory evaluation results of the composition. This makes it possible to calculate estimated features related to the component analysis results and sensory evaluation results of the target composition.

[0213] <Note 12> In the development support method described in Note 10, at least two of the plurality of features in the first dataset relate to information on the formulation of the composition and information on the sensory evaluation results of the composition. This makes it possible to calculate estimated features related to the formulation and sensory evaluation results of the target composition.

[0214] <Note 13> A development support device (1) for supporting the development of a composition according to one aspect of the present disclosure comprises a processor (11) and a memory (12). The processor, in cooperation with the memory, performs data analysis on a first dataset in which a plurality of compositions and a plurality of feature quantities contained in the composition are associated, and generates a first feature space by compressing the dimensions using a predetermined dimensionality reduction method. A target composition is set in the first feature space, and a plurality of estimated feature quantities contained in the target composition is calculated by performing an inverse analysis on the target composition based on the position of the target composition in the first feature space and the parameters obtained by the dimensionality reduction. This makes it possible to calculate a plurality of estimated feature quantities contained in the target composition set in the first feature space. Therefore, by creating a composition based on the calculated estimated feature quantities, the developer can save the effort of trial and error and efficiently create a target composition.

[0215] While embodiments have been described above with reference to the attached drawings, this disclosure is not limited to such examples. It is clear to those skilled in the art that various modifications, alterations, substitutions, additions, deletions, and equivalents can be conceived within the scope of the claims, and these are also understood to fall within the technical scope of this disclosure. Furthermore, the components of the embodiments described above can be combined in any way without departing from the spirit of the invention.

[0216] This application is based on Japanese Patent Application No. 2025-017756 filed on February 5, 2025, and its contents are incorporated herein by reference.

[0217] The technology disclosed herein is useful for developing new compositions or products.

[0218] 1 Development support device 11 Processor 12 Memory 13 Storage 14 Input device 15 Display device 16 Communication device 21 Data set acquisition unit 22 Data analysis unit 23 Map generation unit 24 Target setting unit 25 Inverse analysis unit 26 Target composition information display unit 40 Data set 41A, 41B First data set 50 Map 51A, 51D, 51E First map 52A, 52B, 52C, 52D, 52E Second map T Target composition

Claims

1. A development support method for assisting the development of a composition using a computer, comprising: data analysis of a first dataset in which a composition and a plurality of feature quantities contained in the composition are associated; dimensionality reduction is performed using a predetermined dimensionality reduction method to generate a first feature space; a target composition is set for the first feature space; and a plurality of estimated feature quantities contained in the target composition is calculated by performing an inverse analysis on the target composition based on the position of the target composition in the first feature space and parameters obtained by the dimensionality reduction.

2. The development support method according to claim 1, wherein a first map showing the relative positional relationship of a plurality of compositions in the first feature space is displayed on a predetermined display device.

3. The development support method according to claim 2, comprising: identifying a position or region in the first feature space that proposes setting the target composition based on the relative positional relationship of a plurality of compositions in the first feature space; and displaying the identified position or region that proposes setting the target composition on the first map.

4. The development support method according to claim 2 or 3, wherein the user sets the target composition on the first map.

5. The development support method according to claim 1, comprising: adding the correspondence between the target composition and the plurality of estimated features to the first dataset to generate a second dataset; performing data analysis on the second dataset and reducing its dimensionality using the dimensionality reduction method to generate a second feature space; and displaying a second map showing the relative positional relationship of the target composition in the second feature space on a predetermined display device.

6. A development support method according to claim 1, comprising: adding the target composition and the plurality of estimated features associated with it to the first dataset to generate a second dataset; performing data analysis on the second dataset and compressing its dimensions using the dimensionality reduction method to generate a second feature space; placing the target composition set for the first feature space on a first map showing the relative positional relationships of the plurality of compositions in the first feature space; placing the target composition in the second feature space on a second map showing the relative positional relationships of the plurality of compositions in the second feature space; and displaying the positional relationship between the target composition in the first map and the target composition in the second map on a predetermined display device in a manner that allows for comparison.

7. The development support method according to claim 1, wherein the predetermined dimensionality reduction method is at least one of Multiple Factor Analysis (MFA), Principle Component Analysis (PCA), Generalized Canonical Correlation Analysis (GCCA), and Canonical Correlation Analysis (CCA).

8. The development support method according to claim 1, wherein the feature quantities in the first dataset relate to at least one of the following: information on the component analysis results of the composition, information on the formulation of the composition, information on the sensory evaluation results of the composition, information on the results of a consumer survey of the composition, information on the results of a psychological survey of the composition, and information on the manufacturing of the composition.

9. The development support method according to claim 1, wherein the number of dimensions of the first feature space is one of two, three, or four dimensions.

10. The development support method according to claim 1, wherein the composition and the target composition each contain at least a fragrance.

11. The development support method according to claim 10, wherein at least two of the plurality of features in the first dataset relate to information on the results of component analysis of the composition and information on the results of sensory evaluation of the composition.

12. The development support method according to claim 10, wherein at least two of the plurality of features in the first dataset relate to information on the formulation of the composition and information on the results of sensory evaluation of the composition.

13. A development support device for assisting the development of a composition, comprising a processor and memory, wherein the processor, in cooperation with the memory, performs data analysis on a first dataset in which a plurality of compositions and a plurality of feature quantities contained in the composition are associated, and generates a first feature space by compressing the dimensions using a predetermined dimensionality reduction method, sets a target composition in the first feature space, and calculates a plurality of estimated feature quantities contained in the target composition by performing an inverse analysis on the target composition based on the position of the target composition in the first feature space and parameters obtained by the dimensionality reduction.

Images