A large model-based personalized taste modeling and recipe generation method

Abstract

The application discloses a kind of personalized taste modeling and formula generation method based on large model, including obtaining the multimodal input data of user, feature extraction and fusion are carried out to multimodal input data, generate standardized multimodal taste feature vector;Then generate taste semantic component through semantic deconstruction identification, and the intensity of multiple taste semantic components is quantified, aggregated into deconstruction matrix, then the personalized taste image of user is constructed;Initial formula is generated using generative pre-training model, and the initial formula is optimized, and the final formula is output;And adaptively update user personalized taste image;And the final formula generated is verified by ecological simulation, assess its long-term stability and output verification report;A gustatory ecological simulation environment with dynamic evolution, adaptive learning and ecological balance guarantee is constructed, so that the final formula generated is more stable and accurate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and food information processing, and particularly relates to a personalized taste modeling and recipe generation method based on a large model. BACKGROUND

[0002] With the improvement of living standards, people's demand for personalized diet is increasing, and the existing recipe recommendation or recipe generation is mostly based on the user's historical score or simple label. At present, the following defects exist: firstly, it is difficult to process the user's complex and multi-modal taste description, the mode fusion method is only simple splicing, and the internal relationship and credibility difference between different modal information are not fully considered, and the modeling of taste is mostly static, lacking of dynamic interaction between taste components and simulation of user preference evolution, and the generated recipe is lack of innovation and has poor matching accuracy with user's deep preference. SUMMARY

[0003] (1) Technical problem to be solved

[0004] The application discloses a personalized taste modeling and recipe generation method based on a large model, and urgently needs a technology capable of deeply fusing multi-modal information, dynamically and accurately modeling user taste, and generating highly personalized and innovative recipes.

[0005] (2) Technical scheme

[0006] The application discloses a personalized taste modeling and recipe generation method based on a large model, comprising the following steps:

[0007] Step 1: Obtain multi-modal input data of a user, perform feature extraction and fusion on the multi-modal input data, and generate a multi-modal taste feature vector;

[0008] Step 2: Generate a plurality of taste semantic components through semantic deconstruction recognition of the multi-modal taste feature vector, quantify the intensity of each taste semantic component, and aggregate the intensity of each taste semantic component into a deconstruction matrix;

[0009] Step 3: Based on the deconstruction matrix, construct a personalized taste portrait of the user;

[0010] Step 4: Input the personalized taste portrait into a generation model to generate an initial recipe, optimize the initial recipe, and output a final recipe;

[0011] Step 5: According to the feedback data of the user on the final recipe, adaptively update the personalized taste portrait of the user;

[0012] Step 6: Perform ecological simulation verification on the updated personalized taste portrait and final recipe, and output a verification report.

[0013] Further, the step 1 comprises the following steps:

[0014] Step 101: Collecting user information, including text description, image photo and sensory data;

[0015] Step 102: Inputting the collected user information into a multi-modal pre-training model to extract multi-modal features and generate a feature embedding vector In the feature embedding vector , a niche is introduced to generate a fusion embedding vector , and the generation formula of the fusion embedding vector is as follows:

[0016]

[0017] wherein, is the modal type, is the modal weight, is the inter-modal Euclidean distance, is the niche width parameter, is the feature embedding vector of the modal , the text modal is represented by , the image modal is represented by , and the sensory modal is represented by ;

[0018] Step 103: Normalizing the fusion embedding vector to output a standardized multi-modal taste feature vector .

[0019] Further, the step 2 specifically comprises the following steps:

[0020] Step 201: Using a large model to perform semantic deconstruction on the multi-modal taste feature vector to extract a plurality of taste semantic components;

[0021] Step 202: Calculating the strength of the taste semantic components:

[0022] Step 203: Structurally integrating the strengths of the plurality of taste semantic components to generate a deconstruction matrix.

[0023] Further, the step 3 comprises the following steps:

[0024] Step 301: Based on the deconstruction matrix, introducing the niche theory and population dynamic model in ecology, and performing weighted summation on the portrait weight under the corresponding taste dimension to obtain an initial preference value and subtracting a time decay term to obtain a taste dynamic portrait score of the user;

[0025] Step 302: normalizing the taste dynamic image scores of each of the users, and combining the normalized taste dynamic image scores of the users to generate a personalized taste image of the user.

[0026] Further, the step 4 comprises the following steps:

[0027] Step 401: mapping the personalized taste image of the user to generate an initial formula through a GPT model;

[0028] Step 402: iteratively optimizing the initial formula to obtain an optimization score, so as to seek a dynamic balance among components in the formula while meeting the taste preference of the user;

[0029] Step 403: when the number of iterations reaches a preset number of iterations, terminating the iterative optimization process, and outputting the formula with the maximum optimization score as a final formula.

[0030] Further, the step 402 comprises the following iterative steps:

[0031] Step 4021: calculating a partial derivative of the initial formula with respect to the personalized taste image of each of the users to obtain a formula sensitivity;

[0032] Step 4022: calculating an Euclidean distance between the taste dynamic image score of the user and a formula value to obtain a taste matching error;

[0033] Step 4023: subtracting the taste matching error multiplied by a value of a penalty parameter from a numerical value 1 to obtain a penalty term;

[0034] Step 4024: multiplying the formula sensitivity, the penalty term and an optimization weight to obtain a contribution value of each of the users, summing the contribution values of all the users to obtain an optimization score.

[0035] Further, the step 5 comprises the following steps:

[0036] Step 501: collecting evaluation data of the user on the final formula, preprocessing and fusing the evaluation data into a feedback vector;

[0037] Step 502: calculating a loss based on a difference between the feedback vector and the current personalized taste image of the user, and re-updating a key weight;

[0038] Step 503: synchronously updating the personalized taste image of the user based on the key weight.

[0039] Further, the step 6 comprises the following steps:

[0040] Step 601: Input the user's personalized taste profile and the final recipe as initial input parameters into the ecological dynamics model for time-series simulation.

[0041] Step 602: Calculate the stability index based on the time-series simulation output of the ecological dynamics model;

[0042] Step 603: Generate a stability verification report based on the calculated stability index.

[0043] Furthermore, based on step 102, an adaptive multimodal fusion framework is introduced to generate an optimized fusion embedding vector:

[0044]

[0045] in, For modal types, For modal weights, Represents text modality, Represents image modality, Representing sensory modalities, For context-adaptive modal weights, Modality Feature embedding vector With reference embedding vector The nucleation distance between them This represents the uncertainty factor of the mode; For context-adaptive niche width parameters, It is a weighting factor for intermodal mutual information.

[0046] Furthermore, based on the adaptive multimodal fusion framework introduced in step 102, a modal competition and symbiosis mechanism is further introduced to generate a secondary optimized fusion embedding vector:

[0047]

[0048] in, For modal types, For modal weights, Represents text modality, Represents image modality, Representing sensory modalities, For context-adaptive modal weights, Modality Feature embedding vector With reference embedding vector The nucleation distance between them For the uncertainty factor of the mode; For context-adaptive niche width parameters, is an inter-modal mutual information weighting factor, is an inter-modal competition-symbiosis coefficient, , are modal types.

[0049] Compared with the prior art, the present application has the beneficial effects that:

[0050] By fusing artificial intelligence large model, multi-modal data processing and biological ecology cross thinking, realizing the dynamic balance of taste formula, analyzing user taste through multi-modal input, deconstructing semantic to generate portrait, optimizing the generation of wine formula, and through adaptive learning iteration, a taste ecological simulation environment with dynamic evolution, adaptive learning and ecological balance guarantee is constructed, so that the final formula generated is more stable and accurate. In the feature extraction and fusion step, the niche theory is introduced, the modal weight, inter-modal Euclidean distance and niche width parameter are used to generate fusion embedding vector, effectively solving the problem of multi-modal data heterogeneity and redundancy, improving the accuracy and robustness of feature representation. At the same time, based on the niche theory and population dynamic model, combined with the time decay term, the user's taste preference score is dynamically calculated, and the normalized personalized taste portrait is generated, so that the model can capture the time sequence change and dynamic evolution of user taste, and through the iterative optimization algorithm, the dynamic balance of the internal components of the formula is realized, ensuring that the formula not only meets the user's preference but also has structural rationality. According to the evaluation data of the user on the final formula, the loss is calculated and the key weight is updated, the user's personalized taste portrait is adjusted in real time, so that the model has the ability of continuous learning and self-adaptation. Through the ecological dynamics model, the updated taste portrait and formula are simulated in time sequence, the stability index is calculated and the verification report is generated, ensuring the feasibility and long-term stability of the formula in the simulation environment. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 is the overall step flowchart of the present application.

[0052] Figure 2 is the flowchart of step 1 of the present application.

[0053] Figure 3 is the flowchart of step 5 of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] As Figure 1 shown, the application discloses a personalized taste modeling and formula generation method based on a large model, comprising the following steps:

[0056] Step 1: Obtain the multi-modal input data of the user, extract and fuse the features of the multi-modal input data, and generate a multi-modal taste feature vector;

[0057] Step 2: The multi-modal taste feature vector is generated through semantic deconstruction recognition to generate a plurality of taste semantic components, the intensity of each taste semantic component is quantified, and the intensity of each taste semantic component is aggregated into a deconstruction matrix;

[0058] Step 3: Based on the deconstruction matrix, the personalized taste image of the user is constructed;

[0059] Step 4: The personalized taste image is input into a generation model to generate an initial formula, and the initial formula is optimized to output a final formula;

[0060] Step 5: According to the feedback data of the user on the final formula, the user's personalized taste image is updated adaptively;

[0061] Step 6: The updated personalized taste image and final formula are subjected to ecological simulation verification, and a verification report is output.

[0062] Specifically, the step 1 comprises the following steps:

[0063] Step 101: Collect user information, collect all information related to wine taste of the user from various channels, including text description such as "like sweet red wine", image photo: such as wine bottle photo, sensory data: such as tasting score recorded through APP, define the original data set of user information as , wherein is text, is image, is sensory data;

[0064] Step 102: Input the obtained user information into a multi-modal pre-training model (such as a CLIP model) to extract multi-modal features,

[0065] Through the CLIP model, the deep meaning in the text description is understood, the key visual information in the picture is recognized, and the hidden preferences behind the scoring are understood, and a feature embedding vector is generated. The feature embedding vector is input into a niche fusion model to generate a fusion embedding vector, and the generation formula of the fusion embedding vector is as follows:

[0066]

[0067] wherein is the modal type, is the modal weight, is the inter-modal Euclidean distance, is the niche width parameter, is the modal feature embedding vector, represents the text modality, represents the image modality, represents the sensory modality, in fusing these information, the model will intelligently decide whether to believe the user's written description more or the user's score (modal weight) according to the user's historical behavior; at the same time, it also checks whether these information are consistent;

[0068] Step 103: normalizing the niche fusion embedding to the range of [0, 1], outputting the standardized multi-modal taste feature vector ,

[0069] wherein, for text data processing: segmenting user reviews, removing stop words and stemming, using a pre-trained BERT model to extract text embedding vectors ;

[0070] For image data processing: using a pre-trained CLIP model (VisionTransformer variant) to extract embedding vectors of wine bottle label images . Standardize the image (adjust the size to 224x224 pixels, and normalize the pixel value to [0, 1]);

[0071] For sensory data processing: normalize the sensory scores (acidity, sweetness, etc.) to the range of [0, 1] to form a sensory vector , where each dimension corresponds to a sensory attribute; generate embedding vectors from the original input, and output multi-modal taste feature vectors after fusion.

[0072] By introducing niche fusion, intelligent fusion of multi-modal taste data is realized, effectively improving the richness and accuracy of feature expression. By using a dynamic weight distribution mechanism based on user historical preferences, the importance of each modal data in the fusion process can be adjusted. The entire preprocessing process forms a complete technical solution from data collection, feature fusion to standardized output, providing a reliable input basis for personalized taste modeling.

[0073] Preferably, multi-modal fusion is the core step of realizing fine-grained user modeling, and the present scheme optimizes An operator, an adaptive ecological-niche multimodal fusion framework (AENMF) is proposed to generate an optimized fusion embedding vector to improve fusion robustness and personalized modeling accuracy.

[0074] The improved optimized fusion embedding vector provided in the application is as follows:

[0075]

[0076] Wherein, the meanings of various symbols are as follows: : modal embedding vector, represents the text modal, represents the image modal, and represents the sensory modal;

[0077] : Context-adaptive modal weight, dynamically calculated by a context-aware neural network, reflecting the importance of the modal under the current user preference and scene.

[0078] : modal embedding and reference embedding between the kernel distance, using a Gaussian kernel function to adapt to the nonlinear characteristics of high-dimensional space, the formula is:

[0079]

[0080] Wherein is the kernel width parameter, which can be learned by training.

[0081] : Context-adaptive niche width parameter, modal distribution range in the "taste niche", dynamically adjusted by the context-aware network.

[0082] : Modal uncertainty factor, calculated based on the Bayesian probability framework, reflecting the noise level of modal data, the calculation formula is:

[0083]

[0084] Wherein is the variance of the modal embedding, normalized to represent the credibility of the modal, is the variance of the text modal, image modal, and sensory modal embedding, is the modal type;

[0085] : Inter-modal mutual information weighting factor, calculated based on Mutual Information (MI) in information theory, quantifying the semantic complementarity of modal with other modalities, calculated as: where denotes the mutual information between modal embeddings, which can be approximated by a neural network.

[0086] The fused backward vector is further normalized to obtain a multi-modal taste feature vector: .

[0087] To further improve the stability of the formula, another embodiment is proposed, which introduces an inter-modal dynamic interaction term , quantifying the niche overlap and competition relationship between modal and , based on the step 102 introducing an adaptive multi-modal fusion framework, further introducing a modal competition and symbiosis mechanism to generate a secondary optimized fusion embedding vector:

[0088]

[0089] where, is the competition-symbiosis coefficient between modal, are modal types, and the calculation formula is:

[0090]

[0091] The symbols are explained as follows:

[0092] : The niche overlap degree between modal and embeddings, which can be approximated by cosine similarity.

[0093] : The distribution distance between modal and , using kernelized distance measurement.

[0094] : Scaling factor, controlling the sensitivity of the competition-symbiosis coefficient.

[0095] : Threshold parameter, determining the critical point of competition and symbiosis, which can be learned through training.

[0096] The formula calculates the competition-symbiosis coefficient as: The design of the fusion operator gives it stronger biological interpretation and dynamic adaptability, and can dynamically adjust the fusion strategy according to the semantic relationship between modalities.

[0097] Further, to ensure the effectiveness of the operator, the following training and optimization mechanisms are designed:

[0098] A small feedforward neural network (FFN) combined with an attention mechanism is used to dynamically learn and based on user historical preferences and current scene inputs.

[0099] The loss function is designed as: where is the artificially annotated taste similarity triple (such as the user's preference score for food and , is the similarity score function between embeddings, is the marginal parameter, and is a positive truncation function. This loss function ensures the alignment of multi-modal embeddings in the sensory semantic space.

[0100] Uncertainty regularization: by applying a regularization term to prevent a certain modality from dominating the fusion result, the loss function is: where is the regularization coefficient.

[0101] Through the deep optimization of the niche fusion embedding operator , the adaptive niche multi-modal fusion framework (AENMF) is proposed, which improves the modality weight, distance metric, uncertainty calibration, and sensory semantic consistency.

[0102] Specifically, the step 2 includes the following steps:

[0103] Step 201: Use the BERT large model to perform semantic deconstruction on the standardized multi-modal taste feature vector , extract multiple taste semantic components, including but not limited to acidity, sweetness, tannin intensity, fruit intensity, oak flavor intensity, and other basic taste dimensions. This step converts unstructured multi-modal input into a structured set of semantic components.

[0104] Step 202: Intensity quantification of the taste semantic components extracted in step 201, calculating the taste component intensity, first, the first step needs to calculate the partial derivative of the taste feature vector with respect to each target taste component, obtaining a semantic component gradient, the numerical value of the semantic component gradient reflects the importance of the taste component in the user's overall taste preference, if the calculated semantic gradient value is larger, it indicates that the user's taste feature vector reacts strongly to the small changes of this component, and this component is regarded as a key feature in the user's taste image; on the contrary, if the semantic gradient value is smaller, it means that the component has a weak influence on the current user's preference, then the second step: for each target taste component, calculate the distance of the target taste component and all other taste components in the semantic space, then sum the inverse of the distance to obtain the ecological energy value of the target component, then, the third step multiplies the semantic gradient obtained in the first step with the ecological energy value obtained in the second step, specifically, with an adjustment coefficient greater than 1 as the base and the ecological energy value as the exponent, an amplification factor is calculated, which significantly amplifies the initial sensitivity of components with high ecological energy values, and the comprehensive sensitivity value after energy weighting amplification is added by 1 and then logarithm is taken, which ensures that the final output intensity is always positive regardless of the input value; at the same time, larger input values have a compression effect, preventing the instability of the model caused by the excessively high intensity value of individual components, thereby ensuring the smoothness and robustness of the output result.

[0105] Step 203: Multiply the intensity value output by the third step by a global scaling coefficient, thereby adjusting the intensity values of all components to a unified, pre-set numerical range, repeat the calculation process of step 202 for each identified taste dimension (sweetness, tannin, etc.). After this standardization processing, the intensity values of all taste components are aggregated into a structured deconstruction matrix, all these dimensions and their calculated intensity scores are combined together, according to this structured matrix, the user's taste model, how many numerical values is the acidity intensity, how many numerical values is the sweetness intensity, how many numerical values is the tannin intensity.

[0106] This step breaks through the limitation of independent calculation of components in traditional models, and captures the potential synergistic or inhibitory effect between taste dimensions by constructing a dynamic associated energy field.

[0107] The semantic gradient is combined with ecological energy, the inherent characteristics of components and the semantic correlation between components are comprehensively considered, the dynamic balance of the semantic components in the formula is realized, the semantic deconstruction process not only reflects the independent characteristics of a single component, but also captures the synergistic effect of mutual enhancement or inhibition between components, and dynamic deconstruction is realized; the quantized taste semantic components and their corresponding intensity values are structured and integrated to generate a deconstruction matrix corresponding to acidity, sweetness, tannin, fruitiness, and oak flavor. The deconstruction matrix represents the semantic composition of user taste preferences and the relative intensity of each component in a machine-readable standard format.

[0108] Specifically, the step 3 includes the following steps:

[0109] Step 301: First, based on the deconstruction matrix, introduce the niche theory and population dynamics model in ecology, obtain the intensity values of all taste components from the previous taste semantic deconstruction, these components include acidity, sweetness, tannin, etc., each taste semantic component has an intensity value representing the user's preference for the component, then assign a weight to each taste component, the weight represents the contribution of the component to the current taste image, the weight is obtained by learning user historical data, for example, if the user often emphasizes acidity, the weight of the acidity component may be higher, next, multiply the intensity value of each taste component by its corresponding weight to obtain a weighted value, finally, add all the weighted values to obtain a total sum to obtain an initial preference value representing the overall influence of the user's taste components on the current image, also consider the evolution of user taste over time, introduce a time decay term, which represents the change in user preference, recent preferences are more important than long-term preferences, when calculating the time decay term, integrate from the initial time (e.g. the time when the user first uses the model) to the current time, the integral function is an exponential decay function, where the decay rate is controlled by a system parameter, the result of the integral is a value representing the cumulative decay effect of historical preferences. Then multiply this integral value by a decay intensity parameter to obtain the final time decay term, the larger the time decay term, the smaller the influence of historical preferences on the current image, finally subtract the time decay term calculated in the second step from the weighted sum calculated in the first step to obtain the user's taste dynamic image score;

[0110] Step 302: normalizing and formatting the taste dynamic image score to generate a standardized user's personalized taste image, inputting the comprehensive input value into a S-shaped function which maps any input value to 0 to 1, representing probability or intensity; taking the negative exponential function of the comprehensive input value plus 1, and then taking the reciprocal, the result is the score of the current taste image, the closer to 1, the more consistent the user is with the image; the closer to 0, the less consistent, repeat the above steps for each possible taste image such as refreshing type, rich type, balanced type, generate a set of image scores, these scores are combined into the user's personalized taste image, through which the user's current preference can be known, such as this user is currently a strong acidity, medium tannin, and rich grape flavor wine lover.

[0111] By introducing a dynamic integral operation with a time decay factor, the taste image can accurately capture and adapt to the user's taste drift and change over time, improving the timeliness and prediction accuracy of the model. The taste semantic components are integrated in a niche quantization manner, so that the finally generated taste image not only contains intensity information, but also contains dynamic correlation between components, and the image information is more rich and stereoscopic.

[0112] Specifically, the step 4 comprises the following steps:

[0113] Step 401: inputting the user's personalized taste image into a GPT model variant based on a Transformer architecture to map and generate an initial formula through the GPT model variant;

[0114] Step 402: iteratively optimizing the initial formula to obtain an optimization score to seek a dynamic balance between components in the formula while meeting the user's taste preference, comprising the following steps:

[0115] Step 4021: calculating the partial derivative of the initial formula to each of the user's personalized taste images to obtain the formula sensitivity;

[0116] Step 4022: calculating the Euclidean distance between the user's taste dynamic image score and the formula value to obtain the taste matching error;

[0117] Step 4023: subtracting the value of the taste matching error from 1 multiplied by the value of the penalty parameter to obtain the penalty term;

[0118] Step 4024: multiply the formula sensitivity, penalty term and optimization weight to obtain the contribution value of each user, sum the contribution values of all the users to obtain an optimization score, so as to seek dynamic balance among components in the formula while meeting the user taste preference; the square root function form is adopted to prevent the overall flavor imbalance caused by excessive prominence of any formula component, and the design makes the optimization process not only pursue the user portrait, but also forcibly pursue the stability of the internal structure of the formula.

[0119] Step 403: iterative solution is performed through gradient descent optimization algorithm, when the iterative formula comprehensive score converges to the maximum value or reaches the preset iteration number, the optimization process is terminated, and the formula with the maximum optimization score is output as the final formula.

[0120] By introducing the ecological imbalance penalty term based on the square root function, the balance concept in ecology is quantified as a calculable optimization target, which effectively avoids the problem that some flavors or components in the generated formula are too prominent, and ensures the overall coordination and realizability of the formula.

[0121] Specifically, the step 5 comprises the following steps:

[0122] Step 501: collect the evaluation data of the user on the final formula, pre-process and fuse the evaluation data into a standardized feedback vector; for example, after tasting the recommended wine, give it a score (such as 1-5 stars), or write a simple evaluation such as "not bad, but a little sweeter would be better", translate the user's evaluation into a feedback signal, and this vector accurately captures the user's satisfaction with each dimension of the wine such as acidity, sweetness, etc. in numerical form;

[0123] Step 502: calculate the loss based on the difference between the feedback vector and the current user personalized taste portrait and re-update the key weight, calculate the "cosine similarity" between the two vectors, if the directions of the two vectors are completely consistent, the similarity is 1, which means the prediction is completely accurate; if the directions of the two vectors are completely opposite, the similarity is -1, which means the formula prediction is completely wrong, by using the inverse cosine function, the similarity value is converted into an "angle"; when the prediction is completely consistent: the similarity is 1, and the corresponding angle is 0 degrees, which means "no deviation"; when the prediction is completely opposite: the similarity is -1, and the corresponding angle is 180 degrees, which means "completely opposite, great deviation"; and for example, the similarity is 0.5, which may correspond to an angle of 60 degrees, which means there is a moderate deviation; by converting the similarity into an angle, the deviation is more delicate and stable.

[0124] Step 503: Based on the key weight calculated in step 502, apply the calculated adjustment to the internal weight parameters, fine-tune the personalized taste image through the user's new feedback, and make it more in line with the user's current preferences, and synchronize the update of the user's personalized taste image; for example: the user's previous personalized taste image shows "sweetness preference medium", but the user this time feedbacks "sweeter", that is, increase the "sweetness" weight in the user's personalized taste image.

[0125] An end-to-end adaptive learning link from user feedback to parameters and user image is established, forming a complete "generation-feedback-learning-optimization" closed loop.

[0126] Specifically, the step 6 includes the following steps:

[0127] Step 601: input the user's personalized taste image and the final formula as initial input parameters into the ecological dynamics model for time series simulation running; and set a future time range to be simulated,

[0128] Step 602: in the initialization of the ecological simulation verification environment, perform evolution simulation of a specified time span, such as simulating the situation for 100 days or 1 year in the future, continuously calculate the natural changes that may occur to the user's taste through the ecological dynamic model, and also simulate the dynamic balance of the interaction between the components in the formula, for example, advance the simulation clock every day as a time point, and calculate the "instantaneous satisfaction score" of the formula at each time point, and obtain the score at each time point: at each time point of the simulation running, calculate an "instantaneous satisfaction score". This score reflects the matching degree of the formula and the user's taste image at that particular time. An "exponential decay" rule is used to assign weights: the closer to the present, the higher the weight of the score. This means that more attention is paid to whether the user likes the formula in the near future; the farther away from the present, the score weight will decrease exponentially over time; the decrease speed of this weight is controlled by a decay constant; multiply the instantaneous satisfaction score of each time point by the decay weight of the corresponding time point. Then, all these weighted scores during the simulation period are added up. Finally, in order to obtain a standardized index, this cumulative sum is divided by the total length of the simulation, to obtain a stability index, the higher the index, the more stable and satisfactory the formula performs during the entire simulation period, especially in the critical near-term stage.

[0129] Step 603: Based on the calculated stability index, combined with the dynamic fluctuation data recorded in the simulation process, a structured stability verification report is generated. According to the calculated "stability index", a qualitative conclusion is given. For example: the formula is highly consistent with the current preferences of the user and performs stably in long-term simulation, and is strongly recommended; the formula basically meets the requirements, but there are fluctuations in long-term simulation, and is recommended for use; the formula has short-term or long-term risks, and the stability is insufficient, and is recommended to be re-optimized; this report not only contains the final stability conclusion, but also may contain the change curve of the key indicators (such as user satisfaction) in the simulation process, and the analysis of potential risk points; ensure that the generated formula is not only good to use now, but also reliable and stable in the future, so as to greatly improve the use experience and trust.

[0130] The present application significantly improves the multi-modal data processing and semantic deconstruction ability through niche fusion embedding and ecological energy driven component quantization; further enhances the dynamic adaptability and stability of the final formula obtained by the method.

[0131] The innovation of the present application is that by using the population dynamics model in ecology, the taste preference is regarded as a "population", the "niche" concept is introduced to conceptualize user preference, the cross-disciplinary integration is realized, the population dynamic change is introduced into the artificial intelligence driven taste modeling and formula generation method, and a taste ecological simulation environment with dynamic evolution, self-adaptive learning and ecological balance protection is constructed, so that the generated final formula is more stable and accurate.

[0132] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand.

[0133] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be realized in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for personalized flavor modeling and recipe generation based on a large model, characterized in that, Includes the following steps; Step 1: Obtain the user's multimodal input data, perform feature extraction and fusion on the multimodal input data, and generate a multimodal taste feature vector; Step 2: Generate multiple flavor semantic components from the multimodal flavor feature vector through semantic deconstruction recognition, quantify the strength of each flavor semantic component, and aggregate the strength of each flavor semantic component into a deconstruction matrix; Step 2 specifically includes the following steps: Step 201: Use a large model to semantically deconstruct the multimodal taste feature vector and extract multiple taste semantic components; Step 202: Calculate the strength of the flavor semantic component; calculate the partial derivative of the multimodal flavor feature vector with respect to each flavor semantic component to obtain the semantic gradient; calculate the distance of the flavor semantic component to all other flavor semantic components in the semantic space, and calculate the ecological energy value based on the sum of the inverse distances; multiply the semantic gradient with the ecological energy value, and perform exponential amplification, logarithmic transformation and global scaling to obtain the strength value of the flavor semantic component. Step 203: Normalize and structurally integrate the intensity values ​​of each flavor semantic component to generate a standardized deconstruction matrix; Step 3: Based on the deconstruction matrix, construct a personalized taste profile for the user; Step 4: Input the personalized taste profile into the generation model to generate an initial recipe, optimize the initial recipe, and output the final recipe; Step 5: Adaptively update the user's personalized taste profile based on user feedback data regarding the final recipe; Step 6: Perform ecological simulation verification on the updated personalized flavor profile and final formula, and output a verification report; Step 1 includes the following steps: Step 101: Collect and obtain user information, including text descriptions, images, and sensory data; Step 102: Input the acquired user information into the multimodal pre-trained model to extract multimodal features and generate feature embedding vectors. In the feature embedding vector Introducing niches to generate fused embedding vectors The fused embedding vector The formula for generating it is as follows: ； in, For modal types, For modal weights, The Euclidean distance between modalities. This is a niche width parameter. For the feature embedding vector of the modality, Represents text modality, Represents image modality, Indicates sensory modality; Step 103: merging the embedded vector Normalization, outputting a standardized multimodal taste feature vector. ; Based on step 102, an adaptive multimodal fusion framework is introduced to generate an optimized fusion embedding vector: ； in, For modal types, For modal weights, Represents text modality, Represents image modality, Representing sensory modalities, For context-adaptive modal weights, : Feature embedding vector of a modality With reference embedding vector The nucleation distance between them This represents the uncertainty factor of the mode; For context-adaptive niche width parameters, The intermodal mutual information weighting factor; Based on the adaptive multimodal fusion framework introduced in step 102, a second optimization fusion embedding vector is generated by fusing modal competition and symbiosis mechanisms. ； in, For modal weights, Represents text modality, Represents image modality, Representing sensory modalities, For context-adaptive modal weights, : Feature embedding vector of a modality With reference embedding vector The nucleation distance between them This represents the uncertainty factor of the mode; For context-adaptive niche width parameters, As a weighting factor for intermodal mutual information, The intermodal competition-co-occurrence coefficient. , All are modal types.

2. The method for personalized flavor modeling and recipe generation based on a large model according to claim 1, characterized in that, Step 3 includes the following specific steps: Step 301: Based on the deconstruction matrix, introduce the niche theory and population dynamics model in ecology, and perform a weighted summation of the strength of each taste semantic component and the profile weight under the corresponding taste dimension to obtain the initial preference value and subtract a time decay term to obtain the user's taste dynamic profile score. Step 302: Normalize the dynamic taste profile scores of each user, and combine the normalized dynamic taste profile scores of the users to generate personalized taste profiles for the users.

3. The method for personalized flavor modeling and recipe generation based on a large model according to claim 2, characterized in that, Step 4 includes the following steps: Step 401: The user's personalized taste profile is mapped to generate an initial recipe using the GPT model; Step 402: Iteratively optimize the initial formula to obtain an optimization score, in order to seek to achieve a dynamic balance between the components within the formula while satisfying user taste preferences; Step 403: When the number of iterations reaches the preset number of iterations, terminate the iterative optimization process and output the recipe with the maximum optimization score as the final recipe.

4. The personalized flavor modeling and recipe generation method based on a large model according to claim 3, characterized in that, The specific iterative steps of step 402 are as follows: Step 4021: Calculate the partial derivative of the initial formula with respect to each user's personalized taste profile to obtain the formula sensitivity; Step 4022: Calculate the Euclidean distance between the user's dynamic taste profile score and the recipe value to obtain the taste matching error; Step 4023: Subtract the value of the flavor matching error multiplied by the penalty parameter from the value 1 to obtain the penalty term; Step 4024: Multiply the formula sensitivity, penalty term and optimization weight to obtain the contribution value of each user, and sum the contribution values ​​of all users to obtain the optimization score.

5. The personalized flavor modeling and recipe generation method based on a large model according to claim 4, characterized in that, Step 5 includes the following steps: Step 501: Collect user evaluation data on the final formula, preprocess the evaluation data and fuse it into a feedback vector; Step 502: Calculate the loss based on the difference between the feedback vector and the current user personalized taste profile, and update the key weights again; Step 503: Based on the key weights, the user's personalized taste profile is updated synchronously.

6. The method for personalized flavor modeling and recipe generation based on a large model according to claim 5, characterized in that, Step 6 includes the following steps: Step 601: Input the user's personalized taste profile and the final recipe as initial input parameters into the ecological dynamics model for time-series simulation. Step 602: Calculate the stability index based on the time-series simulation output of the ecological dynamics model; Step 603: Generate a stability verification report based on the calculated stability index.

Images