Model method for predicting cheese preference degree based on GC-MS combined with machine learning

Abstract

The invention discloses a model method for predicting cheese preference degree based on GC-MS combined with machine learning, and belongs to the technical field of cheese evaluation. The method comprises the following steps: firstly, determining the cheese preference of Chinese consumers by adopting a method for checking all suitable items, determining the nature of volatile flavor compounds by combining headspace-arrow type solid-phase microextraction and gas chromatography-mass spectrometry, and quantifying by adopting an internal standard curve method. Secondly, the correlation between the sensory attributes and the compounds is established through quantitative description analysis and Pearson correlation coefficients, and the characteristic flavor compound of each sensory attribute is obtained; and finally, establishing a cheese consumer preference degree identification method by using a machine learning algorithm, and predicting the Chinese consumer preference degree of unknown cheese by using the model. According to the method, the cheese preference degree of Chinese consumers can be accurately recognized only by means of data obtained through instrument analysis and the classification model, a set of more objective and rapid cheese consumer preference degree prediction method is established, and more accurate prediction is achieved along with the increase of sample libraries.

Description

Technical Field

[0001] The present invention relates to the technical field of dairy product detection, namely, the determination of cheese consumer preference and qualitative and quantitative research on volatile flavor substances, and in particular to a method for establishing a cheese consumer preference prediction model based on GC-MS combined with a machine learning algorithm. Background Art

[0002] Flavor is one of the important factors that determine food quality and consumer appeal. Among the complex volatile substances in food, only a small part can trigger human olfactory senses and have a significant dominant or auxiliary effect on the overall flavor of the food. These compounds that play a dominant role in the flavor of the product are called the characteristic flavor compounds of the food.

[0003] Cheese is a fermented dairy product that is popular among people for its unique flavor and functionality. In terms of functionality, cheese is not only rich in nutrients such as protein, fat, vitamins and minerals, but also contains probiotics and fermentation products that are beneficial to human health, such as improving intestinal health and enhancing immunity. In addition, cheese has a rich and diverse flavor that can meet the taste needs of different people.

[0004] The unique flavor of cheese is an important factor affecting consumer acceptance. The flavor components in cheese include flavor compounds already present in milk, flavor compounds produced during milk fermentation, and flavor compounds produced during cheese maturation. At present, more than 600 different volatile flavor compounds have been identified from cheese, mainly including acids, ketones, alcohols, aldehydes, esters, and sulfur-containing compounds.

[0005] Consumer sensory evaluation is an important tool for evaluating products. It directly reflects consumers' intuitive feelings and preferences for product appearance, taste, touch and smell. This method plays a key role in understanding consumer needs and optimizing product design. In market research, sensory evaluation can help companies collect direct feedback, accurately grasp the advantages and disadvantages of products, and provide data support for product improvement and new product development. These data help identify problems, such as poor taste and unattractive appearance, thereby guiding companies to optimize products and develop new products that better meet consumer preferences.

[0006] By researching a large amount of literature and relevant regulations in GB 25192-2022, GB 5420-2021, RHB 501-2004, and RHB505-2004, the descriptions of the sensory attributes of cheese mainly include fruity flavor, nutty flavor, yogurt flavor, caramel flavor, caramel flavor, milky flavor, milky smell, sourness, sweetness, bitterness, saltiness, umami, astringency, aftertaste, richness, hardness, chewiness and delicateness.

[0007] At present, manual sensory evaluation is the most direct and effective means to determine and test the sensory attributes and quality of cheese. Although manual sensory evaluation data can truly reflect consumers' preferences for cheese, it takes a lot of time and money to recruit consumers. When the sample size is too large, consumers are prone to sensory fatigue and are unwilling to participate in the evaluation, making data collection difficult.

[0008] A Chinese patent document with publication number CN118037365A discloses a method and system for evaluating the preference of spicy food, comprising the following steps: (1) judging the sentiment polarity of the preprocessed review text in combination with the score information, dividing the purchase review data into positive, negative or neutral sentiment, and extracting the keyword sequence of the preprocessed review text; determining the weight of each evaluation index based on the keyword sequence to form the user's weighted coded review data on spicy food; (2) constructing a preference evaluation model; and (3) evaluating the preference of spicy food.

[0009] However, there is still a lack of objective, fast, effective and accurate methods for identifying cheese consumer preferences. Therefore, developing a reliable and stable method for identifying cheese preference is very important for promoting the development of new products by companies and providing consumers with more choices. Summary of the invention

[0010] In view of the current lack of an accurate and systematic method for evaluating cheese liking, the present invention provides a method for predicting cheese liking based on characteristic flavor compounds, and uses instrumental analysis combined with a machine learning classification model to accurately predict cheese liking, thereby promoting a more objective understanding of cheese liking.

[0011] To achieve the above object, the present invention provides the following technical solutions:

[0012] A method for predicting cheese preference model based on GC-MS combined with machine learning, comprising the following steps:

[0013] (1) A total of 56 types of cheddar cheese were collected from major supermarkets. The CATA method was used to determine the cheese preferences and frequency of sensory attributes of Chinese consumers and to screen sensory descriptors.

[0014] (2) extracting the volatile flavor compounds of the cheese sample in step (1) using HS-SPME-Arrow, and performing qualitative analysis of the volatile compounds in combination with GC-MS;

[0015] (3) screening out characteristic flavor compounds with greater flavor intensity from the volatile flavor substances by combining a GC-O device, and calculating the concentration of the compounds by using an internal standard curve method;

[0016] (4) The correlation between sensory attributes and volatile flavor compounds was quantified through quantitative descriptive analysis (QDA) and Pearson correlation coefficient, the characteristic flavor compounds of each sensory attribute were obtained, and a cheese preference scale was established.

[0017] (5) applying statistics and machine learning algorithms to analyze the data obtained in steps (3) and (4) to determine the identity characteristic flavor compounds of cheeses with different consumer preferences, and at the same time constructing a cheese preference recognition model to predict consumer preference for cheeses with unknown preferences;

[0018] Specifically, in step (1), untrained consumers were recruited to fill out the CATA questionnaire. The sensory attributes of cheese were checked, the frequency was counted, and the cheese preference was evaluated. The questionnaire required participants to check 17 sensory attributes of cheese, including 6 smells, 8 flavors, and 3 mouthfeels. After answering the CATA questions for each evaluation sample, participants were also required to rate the overall liking of the sample using a 0-5 point hedonic scale (0, extremely dislike; 3, neither dislike nor like; 5, extremely like).

[0019] Specifically, in step (2), the method for qualitatively identifying volatile flavor substances by GC-MS is as follows: data processing is performed by MassHunter chemical workstation, the compound to be identified is searched, and its calculated RI value is compared with the RI value of the volatile component in the literature report of DB-WAX capillary column analysis in the NIST 14 spectral library to qualitatively identify the volatile flavor substances;

[0020] More specifically, in step (3), the quantitative method of the characteristic flavor compounds obtained by GC-O sniffing is: prepare a cheese simulation matrix, add 2-methyl-3-heptanone (0.816 μg / μL of dichloromethane solution, used to quantify other compounds) into the cheese simulation matrix as an internal standard, predict and configure the concentration gradient of each test compound based on the semi-quantitative results of the internal standard of each compound in the preliminary experiment, establish a standard curve based on the ratio of the peak area (y) of the test compound at each gradient concentration and the internal standard at a constant concentration and the ratio of their concentrations (x), and quantify the characteristic flavor compounds.

[0021] More specifically, in step (4), a QDA experiment is conducted based on the sensory descriptors selected in step (1). After long-term and rigorous training, the sensory evaluation panel members have reached a consensus on the intensity of cheese sensory attributes, and score them using a 0-5 point sensory attribute intensity evaluation table (0, the attribute is not perceived at all; 3, the attribute intensity is consistent with the reference; 5, the attribute intensity is much higher than the reference).

[0022] More specifically, in step (5), the specific method for establishing the cheese preference prediction model includes: S1 data entry step: collecting raw data on cheese preference and characteristic flavor compounds; S2 data preprocessor establishment step; S3 key characteristic flavor compound filter establishment step; S4 cheese preference determination model establishment step; optionally also including S5 determination result verification step and S6 screening result fine-tuning step.

[0023] In one specific embodiment, the specific methods of steps S1 to S6 are as follows:

[0024] The specific method of step S1 is as follows: the collected raw data is the input data of the cheese preference prediction model, and the raw data includes the preference of each cheese determined in step (1); and the quantitative analysis data of characteristic flavor compounds obtained in steps (2) and (3), including 23 compounds: acetic acid, butyric acid, hexanoic acid, octanoic acid, trans-2-nonenal, hexanal, 1-octen-3-one, octanal, phenylacetaldehyde, 2-undecanone, ethyl hexanoate, δ-dodecalactone, furfural, furanone, 3-methylindole, δ-octanolactone, nonanal, 3-hydroxy-2-butanone, γ-dodecalactone, γ-nonalactone, 2-methyl-3-sulfide, 3-methylthiopropionaldehyde and heptanal.

[0025] The specific method of step S2 is as follows: The preprocessor is performed immediately after the step of collecting raw data, and includes two sub-steps:

[0026] 2. The raw data of characteristic flavor compounds are normalized. The normalization formula used here is as follows:

[0027]

[0028] In the above formula, x new represents the normalized data, x represents the data before normalization, and x max is the maximum value of the data, x min is the minimum value of the data;

[0029] 3. Visualize the distribution of the characteristic flavor compound content in each cheese preference. In order to adapt to the huge data scale and facilitate comparison between data sets, a tabular heat map is used here to describe the content distribution. Drawing a tabular heat map requires 2 fields representing positions and 1 numerical field. The field representing the position divides the chart into regular rectangular blocks, and the normalized characteristic flavor compound content data is used to represent the numerical field.

[0030] The specific method of step S3 is as follows: the key characteristic flavor compound filter is connected to the preprocessor, and the normalized characteristic flavor compound content data is used to calculate and complete the screening of the key characteristic flavor compounds of cheese preference. The establishment of the key characteristic flavor compound filter utilizes the normalized characteristic flavor compound data obtained in S2 and the decision tree algorithm in machine learning, wherein the calculation method of the information gain IG used is as follows, where IG(X,Z) represents the information gain contributed by the characteristic flavor compound Z in different cheese preferences X:

[0031] IG(X,Z)=H(X)-H(X|Z)

[0032]

[0033] Then, the influence of each characteristic flavor compound on cheese preference was determined by sorting the information gain, thus completing the feature selection; the screening results of this experiment were: acetic acid, 3-hydroxy-2-butanone, hexanal, dimethyl trisulfide, ethyl hexanoate, 3-methylthiopropionaldehyde and caprylic acid.

[0034] The specific method of step S4 is as follows: the cheese preference determination model adopts a discrimination method based on a logistic regression classification algorithm, and the logistic regression calculation formula used is as follows:

[0035]

[0036] In the above formula, z represents the characteristic flavor compound polynomial with θ as the coefficient, and h is the judgment probability of each category, whose value is between (0, 1). When the model is established, one of the two cheese preferences is taken as the positive class each time, and all other types are regarded as negative classes, and two classifiers are trained. When classifying, the classifier that judges the experimental sample as the positive class is used as the classification result.

[0037] The specific method of step S5 is as follows: 1) The incoming characteristic flavor compound data and its preference evaluation data are randomly divided into two groups with a ratio of 8:2. The former is used to establish the cheese preference judgment model (training data), and the latter is used to judge the rationality of the classifier (test data); 2) The classification accuracy is calculated respectively. If the classification accuracy of the training data and test data used by the classifier is higher than 70%, it is confirmed that the rationality evaluation of the characteristic flavor compound filter and the cheese preference judgement has reached the standard, and the screened characteristic flavor compound and cheese preference judgment model are considered to be valid. If the score is lower than 50%, it is necessary to return to steps S3, S4 and S6 to re-establish the model; among them, the classification accuracy score calculation formula and explanation are as follows:

[0038]

[0039] In the above formula, if the actual situation is a positive example, the predicted result is a positive example, then it is TP, and if the predicted result is a negative example, then it is FN; if the actual situation is a negative example, the predicted result is a positive example, then it is FP, and if the predicted result is a negative example, then it is TN.

[0040] The specific method of step S6 is as follows: compare the key compounds presented in the heat map of the characteristic flavor compound content distribution obtained in S2 with the key compounds screened in S3. If there is a significant deviation between the screening results of S3 and the relationship presented by the characteristic flavor compound content distribution obtained in S2, fine-tune according to experience and then proceed to step S4. After adding the fine-tuning step in this experiment, the screening results are: acetic acid, 3-hydroxy-2-butanone, hexanal, dimethyl trisulfide, ethyl hexanoate and octanoic acid.

[0041] The present invention also provides the establishment method to obtain a cheese preference prediction model.

[0042] The present invention further provides a method for predicting cheese liking based on characteristic flavor compounds, which comprises the following steps: extracting key characteristic flavor compounds from a cheese sample to be tested using a solid phase microextraction method to obtain raw data; inputting the raw data into the cheese liking prediction model, and outputting a judgment result of like or dislike.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] At present, there are relatively few methods for identifying cheese preference in China. Generally, the statistical method of Partial Least Squares Regression (PLSR) is used to establish the correlation between cheese volatile compounds and cheese sensory attributes, but it is not possible to predict the preference of unknown cheese. The present invention uses CATA combined with QDA to determine the cheese preference, and then conducts qualitative and quantitative analysis on the characteristic flavor compounds in the cheese, establishes a database based on this, and obtains a cheese preference classification model based on machine learning. On this basis, the present method can accurately predict the cheese preference, laying a foundation for further rapid identification of cheese consumer preference and related detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flow chart of the steps of the model method for predicting cheese preference based on GC-MS combined with machine learning.

[0046] Figure 2 Schematic diagram of the cheese consumer preference survey questionnaire

[0047] Figure 3 Schematic diagram of the correlation between cheese liking and flavor compounds

[0048] Figure 4 This is the process of establishing the cheese preference prediction model of the present invention.

[0049] Figure 5 The present invention is a process for using the cheese preference determination model.

[0050] Figure 6 Schematic diagram of data allocation when establishing a cheese preference determination model.

[0051] Figure 7 From left to right are the distribution heat maps of characteristic flavor compound contents of cheeses with different degrees of preference output by the data preprocessor of the present invention.

[0052] Figure 8 It is the output result of the key characteristic flavor compound filter of the present invention.

[0053] Fig. 9 This is the evaluation result of the cheese preference prediction model of the present invention. DETAILED DESCRIPTION

[0054] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0055] Embodiment 1

[0056] The method for establishing a cheese consumer preference prediction model based on GC-MS combined with a machine learning algorithm comprises the following steps:

[0057] 1. Cheese sample selection and sensory descriptor determination

[0058] We selected 56 commercially available cheese samples from different dairy brands in different regions of the world. First, we planned to recruit 200 consumers from different regions to complete the CATA questionnaire ( Figure 2 ). Before the experiment began, each participant signed an informed consent form. The recruited consumer evaluators were introduced to the process required for the experiment and trained on basic knowledge such as sensory evaluation. After answering the CATA questions for each evaluation sample, participants were also required to rate the overall liking of the sample using a hedonic scale of 0 to 5. Each cheese sample was placed in a 30 ml tasting cup. The samples were randomly coded with 3 digits and presented to the participants. Salt-free biscuits and bottled water were also provided for oral cleaning. CATA data were expressed as the frequency of attributes, and Cochran's Q test was performed; the correlation between the frequency of attributes and the preference of cheese samples was correlated through correspondence analysis (CA). The data obtained by CATA were correlated with the degree of preference, and the sensory attributes related to the degree of preference (correlation greater than 0.5) were obtained through Pearson correlation analysis as the sensory descriptors of QDA.

[0059] The cheese aroma attributes were scored by 10 professional sensory assessors (6 females and 4 males, average age 25 years) using the QDA method: sensory assessors were conducted in a standard sensory laboratory, and the cheese samples were placed at room temperature (24±2℃) for 30min. Next, about 10g of cheese was placed in a transparent tasting cup, the mouth of the cup was covered with a lid to minimize aroma loss, marked with a three-digit code, and presented in a random order. Panelists were asked to rate the intensity of each aroma attribute using a 5-point scale (0, the attribute was not perceived at all; 3, the attribute intensity was consistent with the reference; 5, the attribute intensity was much higher than the reference). The sensory attribute scoring table is shown in Table 1. Unsalted crackers and purified bottled water were provided as taste bud cleansers. To prevent fatigue, each panelist rested for 5min after evaluating four samples. Each cheese sample was evaluated twice.

[0060] Table 1 Sensory attribute rating table of cheese samples Gender: Age: Date:

[0061] Sample No. Fruity Nutty flavor Milky flavor Milk smell sour Caramel flavor Likeability C1 C2 C3 C4 C5 ……

[0062] 2. Determination of characteristic flavor compounds in cheese samples

[0063] (1) Extraction of volatile flavor compounds in yogurt samples using HS-SPME-Arrow

[0064] Weigh 6 g of cheese and transfer it into a 40 mL headspace vial with a polytetrafluoroethylene-lined silicone rubber septum, then add 1 μL of 2-methyl-3-heptanone with a mass concentration of 0.816 mg / mL as an internal standard. The cheese was placed in a 50 °C water bath for equilibration for 30 min. Then the SPME-Arrow extraction head was inserted into the headspace vial, and the extraction fiber (120 μm × 20 mm DVB / CWR / PDMS, Agilent, USA) was pushed out to adsorb in the headspace 5 mm above the sample surface for 30 min, the extraction fiber was retracted, and the extraction head was pulled out for GC-O-MS analysis.

[0065] (2) GC-O-MS analysis conditions

[0066] The gas phase conditions were as follows: a DB-WAX capillary column (60m×0.25mm, 0.25μm) was used, helium was used as the carrier gas, and the carrier gas flow rate was 1mL / min. The column temperature program was as follows: the initial column temperature was 40℃, maintained for 5min, and finally increased to 230℃ at a constant rate of 3℃ / min. The splitless mode was used, and the injection port temperature was 250℃. The extraction head was placed in front of the injection port for 5min.

[0067] The mass spectrometry conditions were as follows: electron ionization source, ionization energy of 70 eV, ion source temperature of 230 °C, interface temperature of 250 °C, quadrupole temperature of 150 °C, detector temperature of 280 °C, full scan scanning mode, and mass range of 25 to 350 m / z.

[0068] (3) GC-O-MS qualitative and quantitative analysis of characteristic flavor compounds

[0069] Qualitative analysis: Data processing was completed by MassHunter Chemical Workstation, and the compounds were searched through the NIST14 mass spectrometry database. The characteristic flavor compounds in yogurt were qualitatively analyzed by comparing the retention index (RI) of the volatile compound in the literature reported in DB-WAX capillary column analysis with the actual RI value, comparing the peak time of the standard under the same conditions, and olfactory measurement. Calculation of actual RI: The peak time of normal alkanes C7~C30 and cheese under the same chromatographic conditions was used to calculate the actual RI value of unknown compound i according to the formula (t n <t i <t n+1 ).

[0070]

[0071] Where: n—number of carbon atoms; t n —Retention time of normal alkanes with carbon number n; t n+1 —peak time of normal alkanes with carbon number n+1; t i —Peak time of unknown compound i.

[0072] Quantitative method: 23 characteristic flavor compounds were screened by sniffing. Each sample was sniffed by four sensory evaluators. The same sensory description was given at the same sniffing time for more than three times, which was a valid result. Finally, 23 characteristic flavor compounds with the greatest influence on cheese aroma were determined (Table 2), and then these compounds were accurately quantified. Specific steps: First, a cheese simulation system was established. 6g of cheese simulation system and mixed standard were transferred into a 40mL headspace vial with a polytetrafluoroethylene-lined silicone rubber septum, and then 1μL of 2-methyl-3-heptanone with a mass concentration of 0.816mg / mL was added as an internal standard. The extraction conditions and GC-MS conditions were the same as above. According to the semi-quantitative results of the internal standard of each compound in the preliminary experiment, the concentration gradient of each test compound was predicted and configured. The peak area ratio (y) of the test compound at each gradient concentration and the internal standard at a constant concentration and their concentration ratio (x) were used to establish a standard curve (Table 3).

[0073] Table 223 characteristic flavor compounds

[0074]

[0075]

[0076] Note: a is the odor perceived at the sniffing port during GC-O

[0077] Table 3 Standard curves of characteristic flavor compounds

[0078]

[0079]

[0080] The precise quantitative results of characteristic flavor compounds in yogurt identified by SPME-Arrow-GC-O-MS are shown in Table 4 , and a total of 23 characteristic flavor compounds were identified from all cheese samples.

[0081] Table 4 Concentrations of characteristic flavor compounds in cheese

[0082]

[0083]

[0084]

[0085] 3. Establishment of correlation between sensory attributes and flavor compounds

[0086] The cheese preference scale was established by performing Pearson correlation analysis between the QDA results and the volatile characteristic flavor compounds.

[0087] 4. Establishment of cheese preference model

[0088] Based on the characteristic flavor compounds in the measured cheese and the corresponding preference, a cheese preference prediction model was established by combining statistics and machine learning methods. Figure 4 As shown, the establishment of a cheese preference prediction model includes collecting raw data on cheese preference and characteristic flavor compounds, establishing a data preprocessor, establishing a key characteristic flavor compound filter, establishing a preference prediction model, and verifying the judgment results.

[0089] (1) The collected raw data are the input data of the cheese preference prediction model, including the preference of each cheese determined in Table 5; and the characteristic flavor compound quantitative analysis data obtained in Table 4, including 23 compounds: acetic acid, 2-methyl-propionic acid, butyric acid, 3-methyl-butyric acid, valeric acid, caproic acid, heptanoic acid, nonanoic acid, limonene, decanal, benzaldehyde, lauryl aldehyde, 2,3-butanedione, 2-heptanone, 2-nonanone, 2-undecanone, furanone, soy sauce ketone, ethyl butyrate, ethyl caproate, ethyl octanoate, ethyl decanoate, δ-caprolactone, ethyl laurate, δ-nonalactone, δ-decanoic acid, γ-dodecalactone and δ-dodecalactone.

[0090] Table 5 Cheese aroma attribute scores and cheese preference

[0091]

[0092]

[0093] (2) The establishment of the data preprocessor is carried out immediately after the step of collecting raw data, and includes two sub-steps:

[0094] First, the raw data of characteristic flavor compounds are normalized. The normalization formula used here is as follows:

[0095]

[0096] In the above formula, x new represents the normalized data, x represents the data before normalization, and x max is the maximum value of the data, x min is the minimum value of the data;

[0097] The second is to visualize the distribution of the characteristic flavor compound content in each cheese preference. In order to adapt to the huge data scale and facilitate comparison between data sets, a tabular heat map is used here to describe the content distribution. Drawing a tabular heat map requires two fields representing the position and one numerical field. The field representing the position divides the chart into regular rectangular blocks, and the normalized characteristic flavor compound content data is used to represent the numerical field.

[0098] (3) The key characteristic flavor compound filter is connected to the preprocessor, and uses the normalized characteristic flavor compound content data to calculate and complete the screening of key characteristic flavor compounds for cheese preference. The establishment of the key characteristic flavor compound filter utilizes the normalized characteristic flavor compound data obtained in S2 and the decision tree algorithm in machine learning, wherein the calculation method of the information gain IG used is as follows, where IG(X,Z) represents the information gain contributed by the characteristic flavor compound Z in different cheese preference X:

[0099] IG(X,Z)=H(X)-H(X|Z)

[0100]

[0101] Then, the influence of each characteristic flavor compound on cheese preference was determined by sorting the information gain, thus completing the feature selection; the screening results of this experiment were: acetic acid, 3-hydroxy-2-butanone, hexanal, dimethyl trisulfide, ethyl hexanoate and caprylic acid.

[0102] (4) The cheese preference determination model adopts a discrimination method based on a logistic regression classification algorithm, and the logistic regression calculation formula used is as follows:

[0103]

[0104] In the above formula, z represents the characteristic flavor compound polynomial with θ as the coefficient, and h is the judgment probability of each category, whose value is between (0, 1). When the model is established, one of the two cheese preferences is taken as the positive class each time, and all other types are regarded as negative classes, and two classifiers are trained. When classifying, the classifier that judges the experimental sample as the positive class is used as the classification result.

[0105] (5) The specific method for verifying the judgment result is as follows: 1) The input characteristic flavor compound data and its preference evaluation data are randomly divided into two groups with a ratio of 8:2. The former is used to establish the cheese preference judgment model (training data), and the latter is used to judge the rationality of the classifier (test data); 2) The classification accuracy is calculated respectively. If the classification accuracy of the training data and test data used by the classifier is higher than 70%, it is confirmed that the rationality evaluation of the characteristic flavor compound filter and the cheese preference judgement has reached the standard, and the screened characteristic flavor compound and cheese preference judgment model are considered to be valid. If the score is lower than 50%, it is necessary to return to S3, S4 and S6 steps to re-establish the model; among them, the classification accuracy score calculation formula and explanation are as follows:

[0106]

[0107] In the above formula, if the actual situation is a positive example, the predicted result is a positive example, then it is TP, and if the predicted result is a negative example, then it is FN; if the actual situation is a negative example, the predicted result is a positive example, then it is FP, and if the predicted result is a negative example, then it is TN.

[0108] (5) Fine-tune the screening results of the data preprocessor and the key characteristic flavor compound filter. The specific method is as follows: compare the key compounds presented in the characteristic flavor compound content distribution heat map obtained in the data preprocessor with the key characteristic flavor compounds screened out by the key characteristic flavor compound filter. If there is a significant deviation between the screening results of the key characteristic flavor compound filter and the relationship presented by the characteristic flavor compound content distribution obtained by the data preprocessor, fine-tune based on experience and then establish the cheese preference prediction model. After adding the fine-tuning step in this experiment, the screening results are: acetic acid, 3-hydroxy-2-butanone, hexanal, dimethyl trisulfide, ethyl hexanoate and octanoic acid.

[0109] (6) The judgment result is the output result of the cheese preference prediction model, and the judgment result is expressed as two types: "like" and "dislike" of the consumer.

Claims

1. A method for predicting cheese preference based on GC-MS combined with machine learning, characterized in that: The following steps are involved: (1) Different brands of commercially available cheddar cheese were selected and the check-all-that-apply (CATA) method was used to determine Chinese consumers' cheese preferences and overall liking. (2) Headspace-arrow solid phase microextraction combined with gas chromatography-mass spectrometry was used to qualitatively and quantitatively analyze the extracted volatile flavor compounds. (3) The correlation between sensory attributes and volatile flavor compounds was quantified through quantitative descriptive analysis and Pearson correlation coefficient, the characteristic flavor compounds of each sensory attribute were obtained, and the correlation between cheese sensory attributes and preference was established. (4) Applying statistics and machine learning algorithms to analyze the data obtained in steps (1), (2) and (3) to determine the characteristic flavor compounds of cheeses with different preferences, and at the same time constructing a cheese preference recognition model and a cheese preference scale to predict the preference of unknown cheeses.

2. The method for establishing a cheese preference prediction model based on GC-MS combined with a machine learning algorithm according to claim 1, characterized in that: It is characterized in that In step (1), the number of cheese samples selected is no less than 50, for example, 56 types of cheddar cheese of different brands or maturity. The sample size is large, and the obtained data is comprehensive and representative. The sensory attributes in the CATA questionnaire are determined by cheese sensory evaluation experts who have been trained in the laboratory for a long time and 50 consumers of different ages, genders, and provinces. The obtained data is comprehensive and representative.

3. The method for predicting cheese preference based on GC-MS combined with machine learning according to claim 1, characterized in that: In step (2), the solid phase microextraction uses an SPME-Arrow extraction head, and the coating material is 120 μm×20 mm DVB / CWR / PDMS; the method for qualitatively identifying volatile flavor substances by GC-MS is as follows: data processing is completed by MassHunter chemical workstation, the compounds to be identified are retrieved, and their calculated RI values ​​are compared with the RI values ​​of the volatile components in the literature reported in DB-WAX capillary column analysis in the NIST 14 spectral library to qualitatively identify the volatile flavor substances; The quantitative method of characteristic flavor compounds is as follows: prepare a cheese simulation matrix, add 2-methyl-3-heptanone (0.816 μg / μL n-hexane solution, used to quantify other compounds) into the cheese simulation matrix as an internal standard, predict and configure the concentration gradient of each test compound based on the semi-quantitative results of the internal standard of each compound in the preliminary experiment, establish a standard curve based on the ratio of the peak area of ​​the test compound at each gradient concentration and the internal standard at a constant concentration (y) and the ratio of their concentrations (x), and quantify the characteristic flavor compounds.

4. The method for predicting cheese preference based on GC-MS combined with machine learning according to claim 1, characterized in that: In step (3), after long-term and rigorous training, the sensory evaluation panel members have reached a consensus on the intensity of cheese aroma attributes; the correlation between cheese sensory attributes and preference is established by performing Pearson correlation analysis between QDA results and volatile characteristic flavor compounds.

5. The method for predicting cheese preference based on GC-MS combined with machine learning according to claim 1, characterized in that: In step (4), the specific method for establishing a cheese preference prediction model includes: S1 data entry step: collecting raw data on cheese preference and characteristic flavor compounds; S2 data preprocessor establishment step; S3 characteristic flavor compound filter establishment step; S4 cheese preference determination model establishment step; optionally also including S5 determination result verification step and S6 screening result fine-tuning step.

6. The method for predicting cheese preference based on GC-MS combined with machine learning according to claim 5, characterized in that: The specific method of step S1 is as follows: the collected raw data is the input data of the cheese preference prediction model, and the raw data includes the preference of each cheese determined in step (1); and the quantitative analysis data of characteristic flavor compounds obtained in step (2), including 23 compounds: acetic acid, butyric acid, hexanoic acid, octanoic acid, trans-2-nonenal, hexanal, 1-octen-3-one, octanal, phenylacetaldehyde, 2-undecanone, ethyl hexanoate, δ-dodecalactone, furfural, furanone, 3-methylindole, δ-octanolactone, nonanal, 3-hydroxy-2-butanone, γ-dodecalactone, γ-nonalactone, 2-methyl-3-sulfide, 3-methylthiopropionaldehyde and heptanal. The specific method of step S2 is as follows: The preprocessor is performed immediately after the step of collecting raw data, and includes two sub-steps:

1. The raw data of characteristic flavor compounds are normalized. The normalization formula used here is as follows: In the above formula, x new represents the normalized data, x represents the data before normalization, and x max is the maximum value of the data, x min is the minimum value of the data; The specific method of step S3 is as follows: the key characteristic flavor compound filter is connected to the preprocessor, and the normalized characteristic flavor compound content data is used to calculate and complete the screening of the key characteristic flavor compounds of cheese preference. The establishment of the key characteristic flavor compound filter utilizes the normalized characteristic flavor compound data obtained in S2 and the decision tree algorithm in machine learning, wherein the calculation method of the information gain IG used is as follows, where IG(X,Z) represents the information gain contributed by the characteristic flavor compound Z in different cheese preferences X: IG(X,Z)=H(X)-H(X|Z) Then, the influence of each characteristic flavor compound on cheese preference is determined by ranking by information gain, thereby completing feature selection; preferably, the characteristic flavor compounds are acetic acid, 3-hydroxy-2-butanone, hexanal, dimethyl trisulfide, ethyl hexanoate, 3-methylthiopropionaldehyde and octanoic acid. The specific method of step S4 is as follows: the cheese preference determination model adopts a discrimination method based on a logistic regression classification algorithm, and the logistic regression calculation formula used is as follows: In the above formula, z represents the characteristic flavor compound polynomial with θ as the coefficient, and h is the judgment probability of each category, whose value is between (0, 1). When the model is established, one of the two cheese preferences is taken as the positive class each time, and all other types are regarded as negative classes, and two classifiers are trained. When classifying, the classifier that judges the experimental sample as the positive class is used as the classification result. The specific method of step S5 is as follows: 1) The incoming characteristic flavor compound data and its preference evaluation data are randomly divided into two groups with a ratio of 8:

2. The former is used to establish the cheese preference judgment model (training data), and the latter is used to judge the rationality of the classifier (test data); 2) The classification accuracy is calculated respectively. If the classification accuracy of the training data and test data used by the classifier is higher than 70%, it is confirmed that the rationality evaluation of the characteristic flavor compound filter and the cheese preference judgement has reached the standard, and the screened characteristic flavor compound and cheese preference judgment model are considered to be valid. If the score is lower than 50%, it is necessary to return to steps S3, S4 and S6 to re-establish the model; among them, the classification accuracy score calculation formula and explanation are as follows: In the above formula, if the actual situation is a positive example, the predicted result is a positive example, then it is TP, and if the predicted result is a negative example, then it is FN; if the actual situation is a negative example, the predicted result is a positive example, then it is FP, and if the predicted result is a negative example, then it is TN. The specific method of step S6 is as follows: compare the key compounds presented in the heat map of the characteristic flavor compound content distribution obtained in S2 with the key compounds screened in S3. If there is a significant deviation between the screening results of S3 and the relationship presented by the characteristic flavor compound content distribution obtained in S2, fine-tune according to experience and then proceed to step S4. Preferably, the screening results are: acetic acid, 3-hydroxy-2-butanone, hexanal, dimethyl trisulfide, ethyl hexanoate and octanoic acid.

7. A cheese preference prediction model is obtained according to the method for establishing a cheese preference prediction model based on GC-MS combined with a machine learning algorithm according to any one of claims 1 to 5.

8. The method for predicting cheese preference based on GC-MS combined with machine learning according to claim 1, characterized in that: The cheese sample to be tested for preference is extracted with characteristic flavor compounds using GC-MS technology to obtain raw data; the raw data is input into the cheese preference prediction model as described in claim 7, and the output result is "like" or "dislike".

Images