Intelligent generation system of hotpot seasoning formula based on big data

Abstract

The present application relates to the technical field of computer data processing, in particular to a hot pot base formula intelligent generation system based on big data. The system uniformly encodes historical hot pot base formula, raw material attribute, flavor component, sensory label and processing adaptation information, and constructs a heterogeneous flavor graph containing raw material nodes, flavor component nodes, sensory label nodes and process condition nodes. The embedding features of enhancement, inhibition, substitution and conflict relationships are calculated through multi-relation graph representation learning. According to the target flavor profile, a candidate raw material set and initial proportion are generated, and the graph constraint decoding module is used for compatibility verification, local replacement and proportion reconstruction to output the formula. The present application can convert the nonlinear compatibility relationship between hot pot base raw materials into a computable constraint, reduce the formula deviation caused by aroma masking, spicy imbalance and thick taste covering, and improve the consistency and traceability of the formula generation result and the target flavor profile.

Description

Technical Field

[0001] This invention relates to the field of computer data processing technology, specifically to an intelligent generation system for hot pot base recipes based on big data. Background Technology

[0002] Hot pot base recipes typically consist of multiple ingredients, including oils, chili peppers, Sichuan peppercorns, broad bean paste, spices, savory seasonings, and flavor enhancers. The roles of each ingredient in the recipe are not simply additive. Conventional recipe digitization systems generally record historical recipes as structured data tables, recording ingredient names, addition ratios, processing conditions, recipe categories, and sensory evaluation results. New recipes are then generated using database searches, similar recipe matching, or statistical analysis. In this approach, the system usually first searches for historical recipes with similar labels based on the input target flavor tag. Then, it generates recommended recipes based on the average addition amount of each ingredient in historical recipes, the most frequently occurring ingredient, or manually set ratio boundaries. For situations where the same ingredient has multiple names, place-of-origin modifications, or processing methods, existing systems often use manually maintained thesaurus for merging to reduce duplicate entries. While this approach can complete basic data archiving and recipe similarity recommendations, its core processing object remains the ingredient and ratio fields in the recipe table; the deep compatibility relationships between ingredients are not modeled as independent, computable objects.

[0003] In more advanced intelligent recipe generation technologies, the system further incorporates machine learning models to process historical recipe data. A common approach is to convert each recipe into a feature vector, using the presence or absence of ingredients, their proportions, recipe category, spiciness level, numbing sensation level, and aroma label as input features, and sensory scores or target flavor labels as outputs to train regression, classification, or generative models. After training, the model outputs candidate ingredient combinations and their corresponding proportions based on the input target flavor profile. To avoid generating obviously unreasonable results, the system also incorporates manual rules in the post-processing stage, such as limiting the maximum proportion of certain ingredients, excluding ingredients that are not allowed to appear simultaneously, or requiring candidate recipes to contain specified basic ingredients. This approach adds a data learning process compared to simple retrieval methods, but the model input is mostly flattened numerical features, and the generated results are mainly influenced by the distribution of historical samples. For relationships such as the balance of spiciness and numbing sensation between chili peppers and Sichuan peppercorns, the carrying capacity of oils for fat-soluble aromas, the masking and enhancement between spices, and the complex contribution of fermented broad bean paste to the rich and salty flavors, conventional vector models struggle to retain the source, direction, intensity, and applicable proportion range of these relationships.

[0004] Some existing solutions employ knowledge bases or rule bases to assist in the verification of hot pot base recipes. These solutions typically involve researchers pre-compiling information on raw material categories, prohibited combinations, commonly used substitutes, basic dosage ranges, and flavor label mappings. After the system generates candidate recipes, it compares each recipe against the rule base to determine if it violates prohibited combinations or exceeds the recommended proportions. To improve rule maintenance, some solutions establish a correlation table between raw materials and flavor labels, using the contribution relationship between raw materials and labels as a recommendation criterion. While these methods can eliminate some recipes that clearly do not conform to empirical rules, the rules are often static text or fixed table entries, making it difficult to express how the relationship between the same raw material combination changes across different dosage ranges. For example, a certain spice might enhance the aftertaste at a low proportion but mask the fermented aroma of fermented broad beans at a higher proportion; a certain chili variety might enhance the aroma of chili oil when combined with oil, but might cause a confusion of numbing and spicy layers when the proportion of Sichuan peppercorns is high. Existing rule bases typically only express simple prohibited or limited relationships, making it difficult to incorporate enhancement, inhibition, substitution, and conflict relationships into the generation process.

[0005] The main technical problem with existing big data-based hot pot base recipe generation technologies is that the systems typically focus on ingredient names and proportions, failing to construct a heterogeneous relational structure that integrates ingredients, flavor components, sensory labels, processing conditions, and the enhancing, inhibiting, substituting, and conflicting relationships between ingredients into a model-learning and decoding framework. This results in the recipe generation model's inability to identify the directionality, source reliability, and effective dosage range of ingredient compatibility during the generation process. When the target flavor profile requires multiple ingredients, the model tends to provide candidate combinations based on historical co-occurrence probabilities or average proportions, without assessing potential conflicts such as aroma masking, imbalanced spiciness, insufficient oil phase carrying capacity, or overpowering flavor coverage among the candidate ingredients. The core deficiency is that while candidate recipes may meet the target labels and proportion boundaries in terms of data format, their ingredient combinations lack a structured verification and correction mechanism oriented towards flavor relationships, leading to uncontrollable deviations between the generated results and the target flavor. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent generation system for hot pot base recipes based on big data, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The intelligent generation system for hot pot base recipes based on big data includes a big data recipe sample processing module, a heterogeneous flavor map construction module, a multi-relationship graph representation learning module, a target flavor condition generation module, and a graph constraint decoding module. The big data formula sample processing module uniformly encodes historical hot pot base formulas, raw material properties, flavor components, sensory labels, and processing adaptation information. The heterogeneous flavor profile construction module generates a heterogeneous flavor profile containing raw material nodes, flavor component nodes, sensory label nodes, and process condition nodes based on the unified coding results. The multi-relationship graph representation learning module performs embedding calculations on the enhancing, inhibiting, substitution, and conflicting relationships in the heterogeneous flavor profile; The target flavor condition generation module generates a set of candidate raw materials and an initial proportion based on the target flavor profile; The system that outputs hot pot base recipe after the graph constraint decoding module performs compatibility verification and ratio correction on the candidate raw material set and initial ratio based on the embedding calculation results.

[0008] Preferably, the big data formula sample processing module includes a raw material semantic normalization processing submodule, a flavor attribute vector generation submodule, and a sample credibility labeling submodule; The raw material semantic normalization processing submodule maps synonymous raw material names, place of origin modification names, processing form names, and compound seasoning raw material names to a unified raw material identifier. The flavor attribute vector generation submodule generates raw material flavor attribute vectors according to the dimensions of spiciness, numbing sensation, fat-soluble aroma, volatile aroma, fullness, saltiness and umami, and aftertaste. The sample credibility labeling submodule assigns credibility identifiers to samples based on the source of the formula, the completeness of sensory evaluation, the completeness of raw material proportions, and the completeness of processing conditions, and writes the credibility identifiers into the node attributes or edge attributes of the heterogeneous flavor profile.

[0009] Preferably, the heterogeneous flavor map construction module includes a multi-type edge generation submodule, a dosage range association submodule, and a compatibility conflict marking submodule; The multi-type edge generation submodule generates different types of graph edges based on historical co-occurrence frequency, flavor component overlap, sensory label co-contribution, and alternative ingredient labeling data. The dosage range association submodule associates each graph edge with the corresponding raw material ratio range, so that the same raw material combination has independent edge attributes under different addition ratios. The compatibility conflict marking submodule generates conflict edges based on the relationships of aroma masking, spiciness imbalance, insufficient oil phase carrying capacity, and thick flavor coverage, and establishes reference relationships between the conflict edges and the corresponding raw material nodes, flavor component nodes, and sensory label nodes.

[0010] Preferably, the graph constraint decoding module includes a candidate recipe path search submodule, an edge relationship verification submodule, and a ratio reconstruction submodule; The candidate formula path search submodule starts from the sensory tag node corresponding to the target flavor profile and searches for candidate paths that meet the raw material coverage conditions in the heterogeneous flavor map. The edge relationship verification submodule performs combined verification of the strengthening relationship, inhibition relationship, substitution relationship and conflict relationship between raw material nodes in the candidate path; When there are conflicting or invalid edges in the candidate path, the ratio reconstruction submodule reconstructs the ratio of candidate raw materials according to the effective usage range recorded in the edge attributes, and retains the candidate formulas that meet the verification conditions.

[0011] Preferably, the raw material semantic normalization processing submodule further includes a composite raw material disassembly processing submodule and a raw material particle size locking submodule; The composite raw material disassembly and processing submodule disassembles the identifiable basic components contained in broad bean paste, compound spice powder, compound chili powder and seasoning oil into basic raw material identifiers and composite carrier identifiers. The raw material particle size locking submodule determines the mapping particle size among basic raw material level, composite carrier level and finished raw material level based on the completeness of basic component disclosure in historical samples. The flavor attribute vector generation submodule generates basic component vectors, composite carrier vectors, and finished raw material vectors according to the mapping granularity, and establishes hierarchical reference relationships for vectors of different granularities in the heterogeneous flavor map.

[0012] Preferably, the multi-type edge generation submodule further includes a directional edge weight calculation submodule and an edge source fusion submodule; The directional edge weight calculation submodule calculates the degree of flavor enhancement, flavor inhibition, and substitution feasibility of the first raw material on the second raw material, and generates independent edge weights for raw material relationships in opposite directions. The edge source fusion submodule uses historical formula statistical relationships, flavor component similarity relationships, sensory label co-occurrence relationships, and trial blending correction records as edge sources, and sets source identifiers for different edge sources; The dosage range association submodule stores multiple edges of the same raw material pair in layers according to the source identifier, so that the same raw material pair retains independent relationship records under different data sources and different dosage ranges.

[0013] Preferably, the scaling reconstruction submodule further includes a local replacement submodule, a scaling migration submodule, and a constraint write-back submodule; When there is a conflict edge in the candidate path, the local replacement submodule selects a candidate alternative ingredient from the heterogeneous flavor spectrum that has a substitution relationship with the conflicting ingredient and does not trigger a new conflict edge. The proportion migration submodule performs migration calculations on the proportion components of the replaced raw material based on the differences between the replaced raw material and the candidate alternative raw material in the flavor attribute vector. The constraint write-back submodule writes the replaced raw material node, the updated proportion range, and the corresponding edge relationship back to the candidate path, so that subsequent edge relationship verification is performed based on the updated candidate path.

[0014] Preferably, the composite raw material disassembly and processing submodule is further configured with a component missing completion submodule and a disassembly confidence transfer submodule; When the compound raw material contains only some basic component information, the component missing completion submodule generates a set of candidate basic components based on the distribution of basic components of similar compound raw materials, the flavor attribute vector of finished raw materials and the co-occurrence relationship in historical formulas. The disassembly confidence transfer submodule assigns disassembly confidence to the basic component identifiers in the candidate basic component set and transfers the disassembly confidence to the graph edges corresponding to the hierarchical reference relationship; The multi-relationship graph representation learning module reads the decomposition confidence during embedded computation and performs confidence-based representation learning on the edges of the composite material relation graph.

[0015] Preferably, the edge source fusion submodule further includes a conflict evidence decomposition submodule and an edge version management submodule; The conflict evidence decomposition submodule decomposes the edge relationships with opposite sources or opposite directions in the same raw material pair into statistical evidence edges, component evidence edges, sensory evidence edges, and modified evidence edges, and establishes a reference relationship between each evidence edge and the corresponding sample identifier. The edge version management submodule generates a version sequence for the evidence edge according to the sample time, sample credibility identifier, and trial matching correction record; The multi-relation graph representation learning module selects a set of evidence edges according to the version sequence during embedded computation, and performs relation aggregation on multiple versions of evidence edges for the same raw material pair.

[0016] Preferably, the constraint write-back submodule further includes a candidate recipe difference index submodule and a duplicate defect suppression submodule; The candidate recipe difference index submodule writes the candidate path, proportion range, triggered conflict edge, and resolved conflict edge after each replacement into the difference index table. The repetitive defect suppression submodule identifies candidate paths that have the same combination of conflicting edges and overlapping proportions with historical invalid candidate recipes based on the difference index table, and masks the corresponding path segments when the candidate recipe path search submodule searches again. The edge relationship verification submodule performs a combined verification on the candidate paths that are not masked, and associates the verification results with the difference index table.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By uniformly encoding historical hot pot base recipes, raw material attributes, flavor components, sensory labels, and processing adaptation information, and constructing a heterogeneous flavor map containing raw material nodes, flavor component nodes, sensory label nodes, and processing condition nodes, the enhancement, inhibition, substitution, and conflict relationships between raw materials, originally scattered in historical samples, sensory records, and trial mixing correction records, are transformed into graph edge relationships. After embedding the above graph edges using multi-relationship graph representation learning, the candidate raw material set and initial proportions generated by the target flavor conditions no longer solely rely on the co-occurrence probability and average addition amount of historical recipes, but undergo compatibility verification during the graph constraint decoding process. When aroma masking, spiciness imbalance, insufficient oil phase carrying capacity, or heavy flavor coverage occur in the candidate recipes, graph constraint decoding can reconstruct the proportions of candidate raw materials or replace local raw materials based on the effective dosage range, directional edge weight, and substitution relationship in the edge attributes. Thus, the formula generation process has been transformed from a simple numerical prediction to a generation process with relational constraints. Before output, the candidate formula undergoes joint verification of raw material relationships, flavor component relationships, and sensory label relationships, which reduces the situation where the target flavor label is satisfied but the actual compatibility relationship conflicts, and makes the generated result have a more stable mapping relationship with the target flavor profile.

[0018] 2. Through processes such as raw material semantic normalization, compound raw material decomposition, raw material granularity locking, edge source fusion, edge version management, and candidate formula difference indexing, the internal consistency of the data used for formula generation is enhanced. Synonymous raw material names, processing form names, and compound seasoning raw material names are mapped to a unified raw material identifier. Compound raw materials such as broad bean paste, compound spice powder, compound chili powder, and seasoning oil can enter the mapping process at the basic raw material level, compound carrier level, or finished product raw material level, reducing misjudgments of relationships caused by inconsistent data granularity. For relationships of the same raw material with different sources, directions, or periods of correction records, the system decomposes them using statistical evidence edges, component evidence edges, sensory evidence edges, and correction evidence edges, and combines sample credibility and version sequence to participate in relationship aggregation, giving the relationship records in the graph a traceable data foundation. After candidate formulas are replaced, scaled, and constraint written back, the difference index table saves the triggering conflict edges, resolving conflict edges, and scale ranges. Subsequent path searches can shield path segments identical to historical invalid candidate formulas, reducing the repeated generation of the same defective combinations and narrowing the invalid search range during model iteration. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall data processing of a big data-based intelligent generation system for hot pot base recipes. Figure 2 Flowchart for big data formula sample processing and heterogeneous flavor profile construction; Figure 3Flowchart for multi-relation graph representation learning, target flavor condition generation, and graph constraint decoding; Figure 4 A flowchart for the decomposition of composite raw materials, fusion of side sources, and indexing and writing back of candidate formulation differences. Detailed Implementation

[0020] Reference Appendix Figure 1 In one embodiment, the big data-based intelligent hot pot base recipe generation system runs on a recipe data processing platform. The system takes historical hot pot base recipes, raw material attribute data, flavor component data, sensory label data, and processing adaptation information as input objects. Figure 1 This section primarily introduces the overall system composition and data input objects. The historical hot pot base recipe includes at least the ingredient names, ingredient proportions, recipe category, sensory evaluation records, and trial-mixing correction records. The ingredient attribute data includes the basic categories, processing forms, and identifiable flavor dimensions of chili peppers, Sichuan peppercorns, broad bean paste, spices, oils, savory seasonings, and flavor-enhancing ingredients. The flavor component data represents the contribution of ingredients to dimensions such as spiciness, numbing sensation, fat-soluble aroma, volatile aroma, richness, savory flavor, and aftertaste. The sensory label data expresses the flavor profile of the target hot pot base. The processing adaptation information represents the compatibility relationships of ingredients during stir-frying, simmering, mixing, and maturation. After receiving the above data, the big data recipe sample processing module performs field alignment, name merging, proportion normalization, and label standardization on data from different sources, enabling different historical samples to be converted into unified computable objects. Ingredient proportions are expressed using relative recipe quality, sensory labels are expressed using a combination of discrete labels and continuous intensity, and process conditions are entered into subsequent graphs as process condition nodes, not described as independent hardware structures.

[0021] Reference Appendix Figure 2 In this embodiment, the big data formula sample processing module establishes a unified raw material identifier for each raw material and converts the raw material attributes into a flavor attribute vector. For chili pepper raw materials, spiciness, red oil aroma, fat-soluble aroma, and volatile aroma can be recorded in the same vector; for Sichuan pepper raw materials, numbing sensation, fresh aroma, bitterness risk, and aftertaste persistence can be recorded; for broad bean paste raw materials, fermented aroma, rich flavor, salty and umami, and oil phase integration characteristics can be recorded. The flavor attribute vector is not a simple preservation of sensory text, but rather a conversion of the sensory contribution records, flavor component annotations, and trial-mixing correction records of the corresponding raw materials in historical formulas into numerical coordinates that can participate in calculations. The normalized value of each dimension can be calculated using the following formula: ; in, Indicates raw materials In terms of flavor Normalized attribute values ​​on Indicates raw materials In terms of flavor The original labeled values ​​on This indicates that all ingredients are in terms of flavor dimension. The minimum label value on, This indicates that all ingredients are in terms of flavor dimension. The maximum labeled value on the scale. Logically, this maps flavor intensity from different sources and scales to the same computational interval. For example, if a chili pepper's original labeled value for the spiciness dimension is 72, and the minimum labeled value for this dimension in the sample database is 20 and the maximum labeled value is 100, then the normalized attribute value for this chili pepper's spiciness dimension is... This value is used for subsequent calculation of graph edge weights and verification of candidate formulation proportions. The raw material unified coding fields and their meanings are shown in Table 1 below.

[0022] Table 1. Raw Material Unified Code Fields and Their Meanings:

[0023] The heterogeneous flavor map construction module generates heterogeneous flavor maps based on unified coding results. The heterogeneous flavor map includes raw material nodes, flavor component nodes, sensory label nodes, and processing condition nodes. Raw material nodes represent optional raw materials in the hot pot base; flavor component nodes represent component characteristics that can be contributed to or influenced by raw materials; sensory label nodes represent target flavor expressions such as numbing, spicy, fragrant, rich, fresh, and sweet aftertaste; and processing condition nodes represent processing adaptation states related to the release and integration of raw materials. Graph edges include co-occurrence relationships, contribution relationships, enhancement relationships, inhibition relationships, substitution relationships, and conflict relationships. Co-occurrence relationships originate from records of co-occurrence of raw materials in historical recipes; contribution relationships originate from correspondence records between raw materials and flavor components or sensory labels; enhancement and inhibition relationships originate from trial blending evaluations and correction records; substitution relationships originate from flavor attribute vector similarity and historical substitution records; and conflict relationships originate from records of compatibility anomalies such as aroma masking, imbalance of numbing and spicy flavors, insufficient oil phase carrying capacity, and coverage of rich flavors.

[0024] When the multi-relationship graph representation learning module performs embedding calculations on the heterogeneous flavor map, it does not compress all relationships into a single adjacency relationship, but calculates them separately according to different relationship types. For any raw material node, its representation vector is determined by its own node attributes, adjacent flavor component nodes, adjacent sensory label nodes, process condition nodes, and different types of graph edges. After receiving the target flavor profile, the target flavor condition generation module converts the target flavor profile into a set of sensory label nodes and an intensity vector, and retrieves a set of candidate raw materials in the graph based on the target flavor profile. After the candidate raw material set is generated, the generation model outputs an initial ratio. The graph constraint decoding module performs compatibility verification on the candidate raw material set and the initial ratio based on the embedding calculation results output by the multi-relationship graph representation learning module. If the raw material combination in the candidate formula triggers a conflict edge, or if the ratio of a certain raw material is not within the effective dosage range of the corresponding enhancement relationship, the graph constraint decoding module performs local replacement or ratio correction of the candidate raw materials, and outputs the hot pot base formula after satisfying the compatibility verification. This embodiment begins to refine the sample processing and pre-graphing data structure, refer to Figure 2 This approach clarifies the sequential relationship between raw material normalization, vector generation, confidence level writing, and graph edge construction, facilitating the subsequent descriptions of Table 1 and edge weight calculation. The advantage of this embodiment is that the formula generation process can simultaneously utilize the proportion information from historical formulas and the relationship information from heterogeneous graphs, enabling candidate formulas to undergo joint verification of raw materials, flavor components, sensory labels, and process conditions before output.

[0025] In one embodiment, the big data formula sample processing module includes a raw material semantic normalization processing submodule, a flavor attribute vector generation submodule, and a sample credibility labeling submodule. The raw material semantic normalization processing submodule normalizes synonymous raw material names, place-of-origin modification names, processing form names, and compound seasoning raw material names. For example, "Erjingtiao chili segments," "Erjingtiao dried chili," and "Erjingtiao chili powder" share the same varietal basis in their raw material names but differ in their processing forms. The system does not directly merge them into identical records but generates unified raw material identifiers and processing form sub-identifiers, preserving both varietal relationships and morphological differences. For raw materials related to numbing sensation, such as "red Sichuan pepper," "green Sichuan pepper," and "Sichuan pepper oil," the system retains category identifiers, source identifiers, and carrier identifiers respectively, avoiding direct confusion between the numbing sensation released from the oil phase carrier and the numbing sensation in solid spices. For compound seasoning raw materials, the system first reads the formula disclosure information; the parts that can be broken down enter the basic raw material layer, while the parts that cannot be broken down remain in the finished product raw material layer.

[0026] The flavor attribute vector generation submodule generates raw material flavor attribute vectors based on the unified raw material identifiers. For the performance of the same raw material in different historical formulations, the system does not directly take a simple average, but instead weights it based on sample credibility. The sample credibility labeling submodule assigns a credibility label to each historical sample based on the formulation source, sensory evaluation completeness, raw material proportion completeness, and processing condition completeness. The credibility label can be calculated using the following formula: ; in, Representing historical samples The credibility indicator value, Indicates the credibility of the formula's source. Indicates the completeness of sensory evaluation. Indicates the integrity of the raw material proportions. This indicates the completeness of processing conditions. All variables range from 0 to 1, logically meaning that samples with clear sources, complete evaluations, complete proportions, and clear processing conditions have a higher relationship contribution in subsequent map construction. For example, if a historical sample has a formula source confidence of 0.9, a sensory evaluation completeness of 0.8, a raw material proportion completeness of 1.0, and a processing condition completeness of 0.7, then the confidence indicator value for this historical sample is... This value is written to the node attribute or edge attribute for subsequent edge weight calculation.

[0027] In this embodiment, the sample reliability identifier is not used as a subjective score of the formula's quality, but rather as an identifier of the completeness and computable reliability of the data record. If a formula only contains the names and proportions of the raw materials but lacks sensory evaluation records, the sensory evaluation completeness will be assigned a lower value; if a trial blend record contains clearly defined structured correction records such as "the bitterness is reduced after the proportion of Sichuan peppercorns is reduced" or "the aroma release is more stable after adding oil," then the sensory evaluation completeness and processing condition completeness can be assigned higher values. When the flavor attribute vector generation submodule summarizes multiple sample records of the same raw material, it uses... As a weight, it participates in the calculation. Therefore, the generation of raw material flavor attribute vectors will not be excessively affected by incomplete sample records. The advantage of this embodiment is that the data entering the heterogeneous flavor profile has a unified identifier and a reliable basis, reducing the relationship bias caused by homonyms, homonyms, and missing sample items.

[0028] In one embodiment, the heterogeneous flavor profile construction module includes a multi-type edge generation submodule, a dosage range association submodule, and a compatibility conflict marking submodule. The multi-type edge generation submodule generates different types of edges based on historical co-occurrence frequency, flavor component overlap, sensory label co-contribution, and alternative ingredient labeling data. For co-occurrence relationships between two ingredient nodes, the system reads the number of times the two ingredients co-occur in the same historical formulation and generates co-occurrence edges based on sample confidence. For contribution relationships between an ingredient node and a sensory label node, the system reads the proportion of the ingredient in historical formulations with the corresponding sensory label and sample confidence to generate contribution edges. For reinforcing or inhibiting relationships between two ingredient nodes, the system reads trial formulation correction records and sensory deviation vectors to determine the direction of influence of changes in the proportion of one ingredient on the corresponding flavor contribution of the other ingredient. The system allows multiple relationship edges to exist simultaneously for the same ingredient pair, but each relationship edge has a relationship type, direction, source identifier, and effective dosage range.

[0029] The dosage range association submodule associates each graph edge with the corresponding raw material ratio range, allowing the same raw material combination to have independent edge attributes under different addition ratios. Taking the combination of chili peppers and Sichuan peppercorns as an example, when the chili pepper ratio is in the low range, it can form a numbing and spicy complex relationship with Sichuan peppercorns; when the chili pepper ratio is in the high range, the spiciness may cover the numbness details. Therefore, the same raw material pair may have both enhancing and conflicting edges, the difference being that the effective dosage ranges are different. The compatibility conflict marking submodule generates conflicting edges based on aroma masking, numbing and spicy imbalance, insufficient oil phase carrying capacity, and thick flavor coverage. Conflicting edges do not only indicate that two raw materials cannot appear at the same time, but also that two or more raw materials have an incompatible relationship under a certain ratio range, a certain sensory target, or a certain process condition. Conflicting edges establish reference relationships with the corresponding raw material nodes, flavor component nodes, and sensory label nodes, enabling subsequent decoding to locate the source of the conflict.

[0030] The edge weights of a graph can be calculated using the following formula: ; in, Indicates raw materials Pointing to raw materials At that time, in relation type and dosage range The lower edge rights; This indicates that it contains raw materials. ,raw material And the proportion falls within the dosage range A collection of historical samples; Representing historical samples The credibility indicator value; Representing historical samples For relation types The contribution value of the relationship is used to indicate the reinforcement or substitution support. Positive values ​​indicate reinforcement or conflict support, while negative values ​​indicate inhibition or conflict support. For example, if a raw material pair has three sample records within a target usage range, with confidence levels of 0.8, 0.9, and 0.7, and their contribution values ​​to the reinforcement relationship are 0.6, 0.4, and 0.5 respectively, then the edge weight is... This value is written to the corresponding enhanced edge and used in subsequent graph representation learning and decoding verification. The advantage of this embodiment is that the relationship between the same raw material pair is no longer fixed as a single conclusion, but rather forms a hierarchical graph edge record based on source, direction, and usage range.

[0031] Reference Appendix Figure 3 In one embodiment, the multi-relation graph representation learning module performs node embedding calculations on heterogeneous flavor maps. Since the nodes and edge types in the heterogeneous flavor maps are different, using the same adjacency matrix would mix ingredient co-occurrence, sensory contributions, aroma enhancement, and compatibility conflicts into the same relation space, leading to the loss of relational meaning. To avoid this problem, the multi-relation graph representation learning module configures independent relation transformation matrices for different relation types and configures node update parameters for different node types. For any node... In its first The representation vector after round relation propagation can be calculated using the following formula: ; in, Represents a node In the The representation vector after propagation of the round relation. Represents a nonlinear mapping function. Represents a node Node type, This represents the feature transformation matrix corresponding to the node type. Represents a set of relation types. Indicates by relation type Connect to node The set of adjacent nodes, Indicates adjacent nodes Pointing to node During the dosage range The right of the lower side, Representing relation type The corresponding relation transformation matrix. Its meaning is to map the node's own attributes and adjacency information from different relation sources to the flavor relation space. For example, a chili pepper node has a contributing edge to a red oil aroma node, a reinforcing edge to a Sichuan pepper node, and a conflicting edge to a spice node. These three types of relations are represented by different... When entering the calculation, avoid interpreting contribution, enhancement, and conflict as the same thing.

[0032] In this embodiment, the target flavor condition generation module converts the input target flavor profile into a target vector. The target vector may include dimensions such as spiciness intensity, numbing intensity, body intensity, volatile aroma intensity, and aftertaste intensity. The system retrieves sensory tag nodes related to the target vector in the heterogeneous flavor map, and searches for candidate raw material nodes along contribution edges, enhancement edges, and substitution edges from the sensory tag nodes. After the candidate raw material nodes are combined into a candidate raw material set, the system generates an initial ratio according to historical proportion distribution, node embedding similarity, and category coverage constraints. The initial ratio is not the final formula, but rather a candidate vector entering the map constraint decoding module. If the candidate raw material set lacks an oil-carrying node, and the target flavor contains a high fat-soluble aroma, the map constraint decoding module will supplement the candidate path with raw material nodes that satisfy the oil phase carrying relationship; if a spice node in the candidate raw material set has a body coverage conflict edge with a broad bean paste node, the proportion of that spice will be reduced or replaced. The advantage of this embodiment is that the candidate formula is guided by map relationships during the generation stage and accepts relationship constraints during the decoding stage, making the generated result have a traceable relationship path.

[0033] In one embodiment, the graph constraint decoding module includes a candidate formulation path search submodule, an edge relationship verification submodule, and a proportion reconstruction submodule. The candidate formulation path search submodule starts from the sensory tag node corresponding to the target flavor profile and searches for candidate paths in the heterogeneous flavor graph that satisfy the raw material coverage condition. The candidate path consists of sensory tag nodes, flavor component nodes, raw material nodes, and the graph edges between them, and is not simply a list of raw materials. The edge relationship verification submodule performs combined verification of the strengthening, inhibiting, substituting, and conflicting relationships between raw material nodes in the candidate path. For candidate formulations... and target flavor profile The system can calculate the matching score between candidate formulations and the target flavor profile using the following formula: ; in, Indicates candidate formulations Relative to the target flavor profile The matching score; This represents the flavor prediction vector obtained from the spectrum calculation of the candidate formulation; Represents the target flavor profile vector; This indicates the similarity between two flavor vectors; This indicates the coverage of the candidate formulation to the target sensory tag node; Indicates the conflict penalty value of the candidate formulation; , and This represents the preset calculation weights, used to control the proportions of similarity, coverage, and conflict penalty in the score. For example, a candidate recipe has a similarity of 0.82 to the target flavor profile, a coverage of 0.75, and a conflict penalty value of 0.20. , , When, the matching score is This score is used for candidate recipe ranking and ratio reconstruction triggering decisions.

[0034] Conflict penalty value The cumulative degree of conflict relationships in candidate formulations can be calculated using the following formula: ; in, Indicates candidate formulations The set of raw material pairs that appear simultaneously in the middle, Indicates the type of conflict relationship. Indicates the raw materials in the candidate formulation and raw materials The combined dosage range in which it is located Indicates the weight of the corresponding conflicting edge. This represents the overlap coefficient between the proportions of raw materials and the conflicting dosage ranges in a candidate formulation. For example, the conflicting edge weights of chili peppers and Sichuan peppercorns in a candidate formulation are... The overlap coefficient is 0.50, and the conflict edge weight between a certain spice and fermented bean paste is... If the overlap coefficient is 0.25, then the conflict penalty value is... The proportion reconstruction submodule reads this value and the corresponding conflict edge, and decides to correct the candidate proportion. The advantage of this embodiment is that whether a candidate recipe is output is no longer determined solely by flavor similarity. The compatibility conflict is entered into the generation decoding in a calculable form, reducing candidate results with similar target labels but unstable raw material relationships.

[0035] Reference Appendix Figure 4 In a preferred embodiment, the raw material semantic normalization processing submodule further includes a composite raw material disassembly processing submodule and a raw material particle size locking submodule. (Reference) Figure 4The processing of compound raw materials can be linked with subsequent evidence edge fusion and difference index write-back, making the relationship between progressive embodiments clearer. The compound raw material decomposition and processing submodule is used to process compound raw materials such as broad bean paste, compound spice powder, compound chili powder, and seasoning oil. For "compound spice powder" that appears directly in the formula record, if the historical record discloses that it contains basic ingredients such as star anise, fennel, cinnamon, and cardamom, the system decomposes it into basic raw material identifier and compound carrier identifier; if the historical record only discloses some basic ingredients, the system retains the finished product raw material identifier and establishes a hierarchical reference relationship for identifiable basic ingredients; if the historical record does not disclose basic ingredients, the system does not forcibly infer them as definite ingredients, but instead enters them as finished product raw material nodes into the heterogeneous flavor map. The raw material granularity locking submodule determines the mapping granularity between the basic raw material level, compound carrier level, and finished product raw material level based on the completeness of basic ingredient disclosure in historical samples, avoiding the same compound raw material from participating in mapping repeatedly with inconsistent granularity in different samples. Examples of compound raw material decomposition and mapping granularity processing are shown in Table 2 below.

[0036] Table 2. Examples of particle size processing for composite raw material disassembly and mapping:

[0037] In this embodiment, the flavor attribute vector generation submodule generates basic component vectors, composite carrier vectors, and finished raw material vectors according to the mapping granularity, and establishes hierarchical reference relationships for vectors of different granularities in the heterogeneous flavor map. The basic component vector represents the flavor contribution of decomposable basic raw materials, the composite carrier vector represents the influence of the carrier in the composite raw material on flavor release and carrying capacity, and the finished raw material vector represents the overall flavor performance of composite seasoning raw materials that cannot be completely decomposed. When calculating substitution relationships, the system does not directly compare the basic component vectors with the finished raw material vectors, but instead transforms them to the same granularity space through hierarchical reference relationships. For example, after a compound spice powder is decomposed, the contributions of star anise and fennel in the volatile aroma dimension can be compared, as can the similarity of the finished raw material vectors between the overall compound spice powder and another compound spice powder. During candidate formula path search, if the target flavor requires a finishing aroma supplement, the system can prioritize searching basic raw material level nodes; if historical samples do not disclose basic components, then finished raw material level nodes are searched while retaining confidence limits. The advantage of this embodiment is that the composite raw materials can be modeled in detail when the information is complete, and the overall record can be retained when the information is incomplete, thereby reducing the relationship break caused by the composite seasoning raw materials.

[0038] In a preferred embodiment, the composite raw material disassembly and processing submodule is further configured with a component missing completion submodule and a disassembly confidence transfer submodule. When the composite raw material contains only partial basic component information, the component missing completion submodule generates a set of candidate basic components based on the distribution of basic components of similar composite raw materials, the flavor attribute vector of the finished raw materials, and the co-occurrence relationships in historical formulas. Component missing completion does not directly write the candidate basic components into the determined formula, but instead generates candidate reference relationships with confidence levels. The disassembly confidence level can be calculated using the following formula: ; in, Indicates composite raw materials With candidate basic components The confidence level of the decomposition between them. Indicates the candidate basic components in similar composite raw materials The frequency of occurrence Indicates composite raw materials Finished product raw material vectors and candidate basic components The similarity between the basic component vectors, Indicates candidate basic components With composite raw materials Support for co-occurrence in historical recipes , and This indicates the calculated weight of each component. For example, a compound spice powder and the candidate basic ingredient cardamom have an occurrence frequency of 0.70, a vector similarity of 0.60, and a co-occurrence support of 0.50. , , At that time, the confidence level of disassembly is This value is written to the hierarchical reference edge between the composite raw material node and the candidate basic component node.

[0039] The disassembly confidence transfer submodule transmits the basic component identifiers and corresponding disassembly confidence scores from the candidate basic component set to the graph edges corresponding to the hierarchical reference relationships. The multi-relationship graph representation learning module reads the disassembly confidence scores during embedding computation and performs confidence-based representation learning on the graph edges related to the composite raw materials. For candidate basic components with low confidence scores, the system reduces their contribution to the composite raw material node representation vector during graph propagation; for candidate basic components with high confidence scores, the system allows them to participate in candidate path search and substitution relationship judgment. Therefore, incomplete composite raw materials will not be incorrectly treated as completely unknown nodes, nor will they be treated as definite disassembly results. The advantage of this embodiment is that missing components of the composite raw material participate in the calculation in the form of confidence scores, preserving usable information while limiting the impact of uncertain information on the formulation generation results.

[0040] In one embodiment, the multi-type edge generation submodule further includes a directional edge weight calculation submodule and an edge source fusion submodule. The directional edge weight calculation submodule calculates the flavor enhancement, flavor inhibition, and substitution feasibility of the first ingredient on the second ingredient, and generates independent edge weights for ingredient relationships in opposite directions. In hot pot base formulations, ingredient relationships are directional. For example, a certain oil may carry the aroma of chili oil, but this is not equivalent to chili enhancing the oil to the same degree; a certain spice may mask the aroma of fermented broad bean paste, but this is not equivalent to the masking strength of the broad bean paste on that spice. Therefore, the system will... and The edge source fusion submodule stores the historical formula statistical relationships, flavor component similarity relationships, sensory label co-occurrence relationships, and trial blending correction records as edge sources, and sets source identifiers for different edge sources. The dosage range association submodule stores multiple edges of the same raw material pair hierarchically according to the source identifier, so that the same raw material pair retains independent relationship records under different data sources and different dosage ranges.

[0041] In a preferred embodiment, the edge source fusion submodule further includes a conflict evidence decomposition submodule and an edge version management submodule. The conflict evidence decomposition submodule decomposes edge relationships with opposite sources or opposite directions within the same raw material pair into statistical evidence edges, component evidence edges, sensory evidence edges, and corrected evidence edges, and establishes a reference relationship between each evidence edge and its corresponding sample identifier. Statistical evidence edges are derived from historical formula co-occurrence frequencies, component evidence edges from overlapping or complementary flavor components, sensory evidence edges from shared contributions of sensory labels, and corrected evidence edges from proportion adjustment records after small-scale trial blending. The edge version management submodule generates a version sequence for the evidence edges according to sample time, sample confidence identifier, and trial blending correction records. The multi-relationship graph representation learning module selects the evidence edge set according to the version sequence during embedding computation and performs relationship aggregation on multiple versions of evidence edges for the same raw material pair.

[0042] Relational aggregation can be calculated using the following formula: ; in, This indicates that the same raw material pair is in the relation type and dosage range The aggregate edge weights below, Denotes the set of corresponding evidence edges. Indicates the edge of the evidence Aggregate weights, Indicates the edge of the evidence The edge weights. For example, if the same raw material pair has a statistical evidence edge of 0.30, a sensory evidence edge of 0.55, and a modified evidence edge of 0.70 within a certain usage range, and their aggregation weights are 0.4, 0.8, and 1.0 respectively, then the aggregation edge weights are... The aggregated edge weights participate in subsequent candidate path searches and conflict determination. The advantage of this embodiment is that the raw material relationships can retain evidence from different sources and historical versions, preventing the formation of biased relationships due to a single statistical co-occurrence or a single correction record.

[0043] In one embodiment, the proportion reconstruction submodule further includes a local replacement submodule, a proportion migration submodule, and a constraint write-back submodule. When conflicting edges exist in the candidate path, the local replacement submodule selects candidate substitute ingredients from the heterogeneous flavor profile that have a substitution relationship with the conflicting ingredient and do not trigger new conflicting edges. The substitute candidates are determined not only based on the ingredient category but also need to meet the following conditions: similar flavor attribute vectors, consistent target sensory label coverage, connectable corresponding process condition nodes, and no high-intensity conflicting edges with other ingredients in the candidate formulation. The proportion migration submodule performs migration calculations on the proportion components of the replaced ingredient based on the differences between the replaced ingredient and the candidate substitute ingredient in their flavor attribute vectors. The constraint write-back submodule writes the replaced ingredient nodes, the updated proportion range, and the corresponding edge relationships back to the candidate path, enabling subsequent edge relationship verification to be performed based on the updated candidate path.

[0044] Proportional migration can be calculated using the following formula: ; in, Indicates candidate alternative raw materials The proportion after migration, Indicates the material being replaced The proportion in the original candidate formulation, This represents the set of flavor dimensions involved in the target flavor profile. Indicating flavor dimensions Dimensional weights in the target flavor profile Indicates the material being replaced In terms of flavor Normalized attribute values ​​on Indicates candidate alternative raw materials In terms of flavor The normalized attribute values ​​on the [dimensional values]. For example, if the proportion of the replaced ingredient in the original candidate formulation is 3.0, and the target flavor only considers spiciness and volatile aroma, with weights of 0.6 and 0.4 for the two dimensions respectively; the normalized attribute values ​​of the replaced ingredient on the two dimensions are 0.7 and 0.5 respectively, and the candidate substitute ingredients are 0.8 and 0.4 respectively, then the proportion after migration is [value]. The system uses 2.91 as the initial migration ratio for candidate alternative raw materials, and then the edge relationship verification submodule determines whether it is within the effective usage range.

[0045] After completing local replacement and proportion migration, the constraint write-back submodule updates the candidate paths. The updates include records of removed nodes of the replaced raw material, records of added candidate alternative raw material nodes, records of changes in proportion ranges, resolved conflict edges, retained enhancement edges, and newly added alternative edges. The updated candidate paths are then resubmitted to the edge relationship verification submodule instead of being directly output. If conflict edges still exist after replacement, the system continues to search for the next candidate in the candidate alternative raw material set; if the target sensory label coverage decreases after replacement, the system selects supplementary nodes from compensating raw materials of the same flavor dimension and reconstructs the proportions. The advantage of this embodiment is that the conflict formulation correction process revolves around the spectral relationships and flavor attribute vectors, avoiding treating raw material replacement as a mechanical substitution of raw materials of the same category.

[0046] In a preferred embodiment, the constraint write-back submodule further includes a candidate recipe difference index submodule and a duplicate defect suppression submodule. The candidate recipe difference index submodule writes the candidate path, proportion range, triggered conflict edges, and resolved conflict edges after each replacement into a difference index table. The difference index table records the candidate recipe identifier, original candidate path, replaced candidate path, replaced raw material, candidate substitute raw material, triggered conflict edge combination, resolved conflict edge combination, and proportion overlap range. The duplicate defect suppression submodule identifies candidate paths with the same conflict edge combination and overlapping proportion range as historical invalid candidate recipes based on the difference index table, and masks the corresponding path segments when the candidate recipe path search submodule searches again. The edge relationship verification submodule performs combination verification on the unmasked candidate paths and associates the verification results with the difference index table. The fields and processing meanings of the candidate recipe difference index table are shown in Table 3 below.

[0047] Table 3. Fields and their meanings in the Candidate Recipe Difference Index Table:

[0048] Repetitive defect suppression can be calculated using the following formula: ; in, Indicates the current candidate formulation Compared with historically ineffective candidate formulations The similarity of repeated defects between them This indicates the combination of conflicting edges triggered by the current candidate recipe. This represents the combination of conflicting edges triggered by historically invalid candidate recipes. This indicates the degree of overlap between the two in terms of the relevant raw material ratio range. For example, if the current candidate formulation triggers 3 conflicting edges, and a historical invalid candidate formulation triggers 4 conflicting edges, and they share 2 identical conflicting edges, then the overlap in their relevant ratio ranges is 0.75. Therefore, the similarity of the duplication defect is... When the similarity of repeated defects reaches a preset masking condition, the candidate recipe path search submodule will no longer prioritize the corresponding path segment in subsequent searches. The advantage of this embodiment is that the system can record invalid paths generated during the candidate recipe correction process and reduce the recurrence of identical conflicting combinations in subsequent generation.

[0049] In one embodiment, the target flavor condition generation module and the spectrum constraint decoding module interact in a closed-loop manner. Upon receiving the target flavor profile, the target flavor condition generation module maps the profile to a set of target sensory tag nodes and generates a target flavor vector. The candidate recipe path search submodule, based on the target sensory tag node set, searches for raw material nodes along contribution edges in the heterogeneous flavor spectrum, expands synergistic raw materials that can complement the target flavor along enhancement edges, retains candidate raw materials along substitution edges, and marks potentially incompatible candidate combinations along conflict edges. When the generation model outputs the initial proportion based on the candidate path, it simultaneously reads the raw material proportion range and the effective dosage interval to avoid the initial proportion falling into a clearly invalid range. After completing the edge relationship verification, the edge relationship verification submodule feeds back conflicting edges, missing tag edges, and proportion out-of-bounds edges to the target flavor condition generation module. The target flavor condition generation module adjusts the candidate path search range based on the feedback results, for example, reducing raw material nodes that are weakly correlated with the target flavor but frequently conflict, and increasing raw material nodes directly connected to missing sensory tags.

[0050] In this embodiment, the closed-loop interaction does not rely on manual temporary rules, but rather operates based on nodes, edges, edge weights, dosage ranges, and version records in the heterogeneous flavor map. If a candidate recipe is insufficient in terms of numbing sensation, the system searches for Sichuan peppercorns, Sichuan peppercorns, or oil-phase numbing sensation carriers connected to the numbing sensation tag nodes, and determines whether these ingredients trigger conflicting edges with the chili peppers, broad beans, and spices in the current candidate recipe. If a candidate recipe is too high in terms of rich flavor, the system searches for the raw material nodes and edge sources that cause rich flavor coverage, identifying whether the proportion of broad beans is too high, the proportion of salty and umami seasonings is too high, or whether the compound spice powder and broad beans jointly cause the coverage relationship. Then, the proportion reconstruction submodule performs local processing. Each verification result is written to the difference index table, so that the next round of candidate path search can utilize existing verification records. The advantage of this embodiment is that generation, verification, correction, and write-back form a continuous data processing link. The candidate recipe is output after multiple rounds of relational constraints, reducing the dependence of a single generation result on the historical average proportion.

[0051] In one embodiment, the system uses a structured representation for the input of the target flavor profile. The target flavor profile is not a single text label such as "Chongqing flavor" or "chili oil flavor," but rather consists of sensory label nodes and intensity values. For example, the target flavor profile may include dimensions such as spiciness, numbing sensation, fat-soluble aroma, volatile aroma, body, and aftertaste, each corresponding to a target intensity value. The system writes these target intensity values ​​into the target flavor vector and compares it with the flavor prediction vector obtained from the graph calculation of the candidate recipe. The flavor prediction vector is calculated jointly by the candidate ingredient flavor attribute vector, ingredient proportions, graph enhancement edges, and inhibition edges. If multiple ingredients in a candidate recipe contribute to the same sensory label, the system performs relational correction on the contributions, rather than simply adding them together. For example, if two spices both contribute to volatile aroma, but there is a masking conflict edge between them, the system deducts the prediction contribution of the corresponding dimension; if a certain oil node and the chili oil aroma node have a carrying enhancement edge, the system increases the prediction contribution of the corresponding fat-soluble aroma dimension.

[0052] In this embodiment, the flavor prediction vector of the candidate formulation can be generated jointly by the weighted raw material vector and the relationship correction term. The relationship correction term reads the aggregated edge weights of enhancing edges, inhibiting edges, and conflicting edges, and calls them according to the dosage range of the candidate ratio. If the ratio of a certain raw material shifts, the system recalculates the flavor prediction vector of the candidate formulation and recalculates the matching score with the target flavor profile. For sensory tag nodes in the target flavor profile that are not covered by the candidate formulation, the target flavor condition generation module retrieves the candidate raw materials connected to the tag node from the graph and determines its edge relationship with the existing candidate raw materials. Thus, the target flavor input, candidate raw material generation, and candidate ratio verification are all based on the same graph relationship structure. The advantage of this embodiment is that the target flavor is no longer processed as an isolated tag, and the system can jointly determine the candidate formulation based on multi-dimensional flavor intensity and graph relationships.

[0053] In one embodiment, the system performs structured parsing on natural language sensory descriptions in historical recipes and trial blending correction records. Sensory descriptions may include text such as "numbing sensation suppresses spicy aroma," "aroma is prominent in the initial stage but lacks aftertaste," "fermented bean aroma is covered by spices," and "red oil aroma is insufficient." The system maps the above text to sensory tag nodes, defect types, associated raw materials, and correction directions. If the trial blending record includes the proportions before and after correction, the system establishes a correlation between the proportion change and the sensory change, forming a correction evidence edge. If the trial blending record only contains defect descriptions without a clear correction proportion, the system uses it as a sensory evidence edge in relation fusion, but does not directly generate dosage ranges. The results obtained from the structured parsing enter the edge source fusion submodule and, together with historical recipe statistical relationships and flavor component similarity relationships, form multi-source evidence. For descriptions that cannot be mapped to existing sensory tag nodes, the system writes them into the pending confirmation tag pool and does not directly participate in candidate recipe generation.

[0054] In this embodiment, sensory description parsing employs a combination of term mapping and contextual relationship recognition. Term mapping is used to identify sensory objects such as "numbing sensation," "spicy aroma," "rich flavor," and "sweet aftertaste," while contextual relationship recognition is used to identify relational directions such as "suppression," "insufficiency," "coverage," "enhancement," and "weakening." If the text states "increasing the proportion of Sichuan peppercorns makes the numbing sensation more prominent, but the aftertaste is suppressed," the system establishes an enhancing evidence edge between the Sichuan peppercorn node and the numbing sensation tag node, and establishes an inhibition or conflict evidence edge between the Sichuan peppercorn node and the aftertaste tag node. If the text states "adding oil makes the aroma of red oil more fully released," the system establishes a carrying enhancing evidence edge between the oil node and the red oil aroma node. All of the above evidence edges retain the sample identifier, correction record identifier, and credibility identifier. The advantage of this embodiment is that defect descriptions and correction experiences that were originally difficult to calculate directly in historical trial blending can be converted into graph relationships and incorporated into subsequent candidate formula generation and constraint decoding.

[0055] In one embodiment, the overall working principle of the system is as follows: The big data formula sample processing module processes multi-source hot pot base formula data into unified identifiers, flavor attribute vectors, and sample credibility identifiers; the heterogeneous flavor graph construction module generates multiple types of nodes and edges based on the above processing results, and associates edge relationships with source, direction, and dosage range; the multi-relationship graph representation learning module calculates the representation vector of nodes in the flavor relationship space based on the transformation matrix of different relationship types; the target flavor condition generation module maps the target flavor profile into a set of sensory label nodes and a target flavor vector, and generates a set of candidate raw materials and initial proportions along the graph relationships; the graph constraint decoding module verifies the strengthening, inhibiting, substituting, and conflicting relationships in the candidate paths, and performs local replacement, proportion migration, and constraint write-back on candidate raw materials that trigger conflicts; the candidate formula difference index submodule records the correction process, and the repetition defect suppression submodule masks path segments that are the same as or highly overlap with historical invalid candidate formulas. The above processing links all revolve around the data objects and algorithmic relationships in the hot pot base formula, and do not involve any modification to the structure of the hot pot base production equipment.

[0056] In this embodiment, the hot pot base recipe output by the system includes a set of candidate ingredients, the proportion of each ingredient, the target flavor matching score, records of triggered conflict edges, records of resolved conflict edges, and recipe relationship paths. The recipe relationship paths represent the graph connections between candidate ingredients and target sensory labels. For example, a chili pepper node is connected to a fat-soluble aroma label via a red oil aroma contribution edge, an oil node is connected to the same label via a carrying enhancement edge, and a Sichuan pepper node is connected to a numbing sensation label via a numbing sensation contribution edge. If the candidate recipe undergoes partial replacement, the output records the ingredients before replacement, the ingredients after replacement, and the basis for proportion migration. This output method provides data path support for the recipe generation results, facilitating subsequent sample write-back and model training. The advantage of this embodiment is that the system covers the complete technical chain from sample processing, heterogeneous graph construction, relationship learning, target generation, constraint decoding to difference write-back, providing an implementable data processing method to address the difficulty in calculating the compatibility relationships of ingredients in hot pot base recipes.

Claims

1. A big data-based intelligent generation system for hot pot base recipes, characterized in that: It includes a big data formula sample processing module, a heterogeneous flavor map construction module, a multi-relationship graph representation learning module, a target flavor condition generation module, and a map constraint decoding module; The big data formula sample processing module uniformly encodes historical hot pot base formulas, raw material properties, flavor components, sensory labels, and processing adaptation information. The heterogeneous flavor profile construction module generates a heterogeneous flavor profile containing raw material nodes, flavor component nodes, sensory label nodes, and process condition nodes based on the unified coding results. The multi-relationship graph representation learning module performs embedding calculations on the enhancing, inhibiting, substitution, and conflicting relationships in the heterogeneous flavor profile; The target flavor condition generation module generates a set of candidate raw materials and an initial proportion based on the target flavor profile; The graph constraint decoding module outputs a hot pot base recipe after performing compatibility verification and ratio correction on the candidate raw material set and initial ratio based on the embedded calculation results. The graph constraint decoding module includes a candidate recipe path search submodule, an edge relationship verification submodule, and a ratio reconstruction submodule. When there are conflicting or invalid edges in the candidate path, the ratio reconstruction submodule reconstructs the ratio of candidate raw materials according to the effective usage range recorded in the edge attributes, and retains the candidate recipes that meet the verification conditions. The heterogeneous flavor map construction module includes a multi-type edge generation submodule, a dosage range association submodule, and a compatibility conflict marking submodule; The multi-type edge generation submodule generates different types of graph edges based on historical co-occurrence frequency, flavor component overlap, sensory label co-contribution, and alternative ingredient labeling data. The dosage range association submodule associates each graph edge with the corresponding raw material ratio range, so that the same raw material combination has independent edge attributes under different addition ratios. The compatibility conflict marking submodule generates conflict edges based on the relationship between aroma masking, spiciness imbalance, insufficient oil phase carrying capacity and thick flavor coverage, and establishes reference relationships between the conflict edges and the corresponding raw material nodes, flavor component nodes and sensory label nodes. The multi-type edge generation submodule further includes a directional edge weight calculation submodule and an edge source fusion submodule; The directional edge weight calculation submodule calculates the degree of flavor enhancement, flavor inhibition, and substitution feasibility of the first raw material on the second raw material, and generates independent edge weights for raw material relationships in opposite directions. The edge source fusion submodule uses historical formula statistical relationships, flavor component similarity relationships, sensory label co-occurrence relationships, and trial blending correction records as edge sources, and sets source identifiers for different edge sources; The dosage range association submodule stores multiple edges of the same raw material pair hierarchically according to the source identifier, so that the same raw material pair retains independent relationship records under different data sources and different dosage ranges; The scaling reconstruction submodule further includes a local replacement submodule, a scaling migration submodule, and a constraint writeback submodule; When there is a conflict edge in the candidate path, the local replacement submodule selects a candidate alternative ingredient from the heterogeneous flavor spectrum that has a substitution relationship with the conflicting ingredient and does not trigger a new conflict edge. The proportion migration submodule performs migration calculations on the proportion components of the replaced raw material based on the differences between the replaced raw material and the candidate alternative raw material in the flavor attribute vector. The constraint write-back submodule writes the replaced raw material node, the updated proportion range, and the corresponding edge relationship back to the candidate path, so that subsequent edge relationship verification is performed based on the updated candidate path.

2. The intelligent generation system for hot pot base recipe based on big data as described in claim 1, characterized in that, The big data formula sample processing module includes a raw material semantic normalization processing submodule, a flavor attribute vector generation submodule, and a sample credibility labeling submodule; The raw material semantic normalization processing submodule maps synonymous raw material names, place of origin modification names, processing form names, and compound seasoning raw material names to a unified raw material identifier. The flavor attribute vector generation submodule generates raw material flavor attribute vectors according to the dimensions of spiciness, numbing sensation, fat-soluble aroma, volatile aroma, fullness, saltiness and umami, and aftertaste. The sample credibility labeling submodule assigns credibility identifiers to samples based on the source of the formula, the completeness of sensory evaluation, the completeness of raw material proportions, and the completeness of processing conditions, and writes the credibility identifiers into the node attributes or edge attributes of the heterogeneous flavor profile.

3. The intelligent generation system for hot pot base recipe based on big data according to claim 2, characterized in that, The candidate formula path search submodule starts from the sensory tag node corresponding to the target flavor profile and searches for candidate paths that meet the raw material coverage conditions in the heterogeneous flavor map. The edge relationship verification submodule performs combined verification of the strengthening relationship, inhibition relationship, substitution relationship and conflict relationship between raw material nodes in the candidate path.

4. The intelligent generation system for hot pot base recipe based on big data according to claim 3, characterized in that, The raw material semantic normalization processing submodule further includes a composite raw material disassembly processing submodule and a raw material particle size locking submodule; The composite raw material disassembly and processing submodule disassembles the identifiable basic components contained in broad bean paste, compound spice powder, compound chili powder and seasoning oil into basic raw material identifiers and composite carrier identifiers. The raw material particle size locking submodule determines the mapping particle size among basic raw material level, composite carrier level and finished raw material level based on the completeness of basic component disclosure in historical samples. The flavor attribute vector generation submodule generates basic component vectors, composite carrier vectors, and finished raw material vectors according to the mapping granularity, and establishes hierarchical reference relationships for vectors of different granularities in the heterogeneous flavor map.

5. The intelligent generation system for hot pot base recipe based on big data according to claim 4, characterized in that, The composite raw material disassembly and processing submodule is further configured with a component missing completion submodule and a disassembly confidence transfer submodule; When the compound raw material contains only some basic component information, the component missing completion submodule generates a set of candidate basic components based on the distribution of basic components of similar compound raw materials, the flavor attribute vector of finished raw materials and the co-occurrence relationship in historical formulas. The disassembly confidence transfer submodule assigns disassembly confidence to the basic component identifiers in the candidate basic component set and transfers the disassembly confidence to the graph edges corresponding to the hierarchical reference relationship; The multi-relationship graph representation learning module reads the decomposition confidence during embedded computation and performs confidence-based representation learning on the edges of the composite material relation graph.

6. The intelligent generation system for hot pot base recipe based on big data according to claim 5, characterized in that, The edge source fusion submodule further includes a conflict evidence decomposition submodule and an edge version management submodule; The conflict evidence decomposition submodule decomposes the edge relationships with opposite sources or opposite directions in the same raw material pair into statistical evidence edges, component evidence edges, sensory evidence edges, and modified evidence edges, and establishes a reference relationship between each evidence edge and the corresponding sample identifier. The edge version management submodule generates a version sequence for the evidence edge according to the sample time, sample credibility identifier, and trial matching correction record; The multi-relation graph representation learning module selects a set of evidence edges according to the version sequence during embedded computation, and performs relation aggregation on multiple versions of evidence edges for the same raw material pair.

7. The intelligent generation system for hot pot base recipe based on big data according to claim 6, characterized in that, The constraint writeback submodule further includes a candidate recipe difference index submodule and a duplicate defect suppression submodule; The candidate recipe difference index submodule writes the candidate path, proportion range, triggered conflict edge, and resolved conflict edge after each replacement into the difference index table. The repetitive defect suppression submodule identifies candidate paths that have the same combination of conflicting edges and overlapping proportions with historical invalid candidate recipes based on the difference index table, and masks the corresponding path segments when the candidate recipe path search submodule searches again. The edge relationship verification submodule performs a combined verification on the candidate paths that are not masked, and associates the verification results with the difference index table.

Images