Cross-cultural demand intelligent mining method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of cross-cultural demand intelligent mining, and discloses a cross-cultural demand intelligent mining method and device, an electronic device and a storage medium, wherein the method comprises the following steps: introducing a multi-modal demand analysis engine to process an original data set containing multi-language mixed data, so that the system can simultaneously analyze text semantics and cultural symbols; then a dynamic mapping matrix based on demand nodes and function nodes is constructed by using a graph neural network, the market adaptability and accuracy of scheme generation are enhanced, the optimized target mapping matrix is input into a preset optimization model, and an improved scheme is obtained.The application has the beneficial effect that feasible candidate product improvement schemes can be efficiently generated and screened under the premise of considering multi-dimensional constraints, so that the conversion process from demand to scheme is systematized and automated, the product development cycle is shortened, and the decision cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of cross-cultural demand intelligent mining technology, and in particular to a cross-cultural demand intelligent mining method, apparatus, electronic device and storage medium. Background Technology

[0002] In global product development, users from different cultural backgrounds have different implicit needs, thus requiring the discovery of these implicit requirements. Traditional needs discovery methods mainly rely on structured tools such as questionnaires. However, when faced with massive amounts of multilingual, unstructured data generated by social media and local forums, these methods are inefficient and have limited coverage. They are particularly difficult to identify mixed language expressions and cultural metaphors (such as emojis). Existing technologies typically statically associate needs with product functions, lacking the flexibility to adapt to the dynamic changes in emerging markets. Furthermore, multilingual processing is often fragmented, ignoring language mixing phenomena, leading to a high rate of semantic misinterpretation. Summary of the Invention

[0003] Based on this, it is necessary to address the existing problem of intelligent mining of cross-cultural needs, and propose a method, device, electronic device and storage medium for intelligent mining of cross-cultural needs.

[0004] A cross-cultural needs intelligent mining method, the method comprising:

[0005] User behavior and comment data from the target market region are acquired to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0006] The original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0007] Obtain the corresponding functional feature vector based on the implicit demand feature vector;

[0008] Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix;

[0009] Obtain the correlation between the implicit demand feature vector and the functional feature vector, and convert the correlation into correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0010] The target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement schemes.

[0011] Furthermore, the multimodal demand parsing engine includes a multilingual pre-trained model and a regional cultural symbol library. The step of processing the original dataset through the pre-set multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors includes:

[0012] The first feature vector is generated by parsing the text semantics in the original dataset using the multilingual pre-trained model, and the second feature vector is generated by parsing the non-textual symbols in the original dataset using the regional cultural symbol library.

[0013] The first feature vector and the second feature vector are fused to obtain the implicit demand feature vector.

[0014] Further, the step of obtaining the correlation between the implicit demand feature vector and the functional feature vector, and converting the correlation into a correlation strength input into the preliminary mapping matrix to obtain the target mapping matrix includes:

[0015] Collect market feedback data and engineering test data of historical product solutions to establish the correlation between the implicit demand feature vector and the functional feature vector;

[0016] Based on the market feedback data and engineering test data, the correlation strength of the edges between the demand nodes and the functional nodes is obtained through iterative learning of the graph neural network.

[0017] The correlation strength is input into the preliminary mapping matrix to obtain the target mapping matrix.

[0018] Furthermore, before the step of inputting the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes, the method further includes:

[0019] Obtain the constraints for the target market region;

[0020] The constraints are input into the initial optimization model to obtain the optimization model.

[0021] Furthermore, after the step of inputting the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes, the method further includes:

[0022] Obtain environmental data for the target market area;

[0023] A digital twin simulation system is constructed based on the aforementioned environmental data;

[0024] Each of the candidate product improvement schemes is input into the digital twin simulation system to verify the candidate product improvement schemes and obtain the verification results of each candidate product improvement scheme.

[0025] Based on the verification results, output the performance simulation report and feasibility assessment of each candidate product improvement scheme.

[0026] Furthermore, after the step of outputting the performance simulation report and feasibility assessment of each candidate product improvement scheme based on each verification result, the method further includes:

[0027] Calculate the dimensional values ​​of each of the candidate product improvement schemes in each preset dimension;

[0028] The comprehensive score of each candidate product improvement scheme is obtained by weighted summation of the dimensional values ​​of each candidate product improvement scheme.

[0029] The candidate product improvement proposals are sorted according to the comprehensive scores, and the sorted proposals are sent to designated personnel.

[0030] Further, the step of obtaining the corresponding functional feature vector based on the implicit demand feature vector includes:

[0031] Based on the implicit demand feature vector, search for and match multiple candidate function feature vectors from the preset product function knowledge base;

[0032] Based on historical matching data or real-time engineering test feedback, the functional feature vector with the highest correlation to the implicit requirement feature vector is determined from the multiple candidate functional feature vectors and then output.

[0033] A cross-cultural needs intelligent mining device, the device comprising:

[0034] The first acquisition module is used to acquire user behavior and comment data from the target market area to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0035] The identification module is used to process the original dataset through a preset multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0036] The second acquisition module is used to acquire the corresponding functional feature vector based on the implicit requirement feature vector;

[0037] The module is used to utilize a graph neural network to obtain a preliminary mapping matrix by using the implicit demand feature vector as demand nodes and the functional feature vector as functional nodes.

[0038] The transformation module is used to obtain the correlation between the implicit demand feature vector and the functional feature vector, and to convert the correlation into the correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0039] The generation module is used to input the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes.

[0040] An electronic device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0041] User behavior and comment data from the target market region are acquired to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0042] The original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0043] Obtain the corresponding functional feature vector based on the implicit demand feature vector;

[0044] Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix;

[0045] Obtain the correlation between the implicit demand feature vector and the functional feature vector, and convert the correlation into correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0046] The target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement schemes.

[0047] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps:

[0048] User behavior and comment data from the target market region are acquired to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0049] The original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0050] Obtain the corresponding functional feature vector based on the implicit demand feature vector;

[0051] Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix;

[0052] Obtain the correlation between the implicit demand feature vector and the functional feature vector, and convert the correlation into correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0053] The target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement schemes.

[0054] The beneficial effects of this invention are as follows: By introducing a multimodal demand parsing engine to process the original dataset containing multilingual mixed data, the system can simultaneously parse textual semantics and cultural symbols, improving the coverage and accuracy of implicit demand identification and feature extraction in cross-cultural scenarios. This overcomes the dependence of traditional methods on single language or structured data. Then, a dynamic mapping matrix based on demand nodes and functional nodes is constructed using a graph neural network, so that the relationship between demand and solution is no longer a static rule, but can be continuously optimized through dynamic learning and updating of the relationship strength. This greatly enhances the market adaptability and accuracy of solution generation. The optimized target mapping matrix is ​​input into a preset optimization model, which can efficiently generate and screen feasible candidate product improvement solutions under the premise of considering multidimensional constraints. This systematizes and automates the transformation process from demand to solution, shortens the product development cycle, and reduces decision-making costs. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] in:

[0057] Figure 1 This is a diagram illustrating the application environment of a cross-cultural needs intelligent mining method in one embodiment.

[0058] Figure 2 A flowchart of a cross-cultural needs intelligent mining method in one embodiment;

[0059] Figure 3 This is a structural block diagram of a cross-cultural demand intelligent mining device in one embodiment;

[0060] Figure 4 This is a structural block diagram of an electronic device in one embodiment. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Figure 1 This is a diagram illustrating the application environment of cross-cultural needs intelligent mining in one embodiment. (Refer to...) Figure 1 This cross-cultural demand intelligent mining method is applied to a cross-cultural demand intelligent mining system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers. The terminal 110 is used to acquire user behavior and comment data from the target market area, and the server 120 is used to generate multiple candidate product improvement solutions.

[0063] like Figure 2 As shown, in one embodiment, a cross-cultural demand intelligent mining method is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to terminals. The cross-cultural demand intelligent mining method specifically includes the following steps:

[0064] S1: Obtain user behavior and comment data from the target market region to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0065] S2: The original dataset is processed by a preset multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors;

[0066] S3: Obtain the corresponding functional feature vector based on the implicit requirement feature vector;

[0067] S4: Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix;

[0068] S5: Obtain the correlation between the implicit demand feature vector and the functional feature vector, and convert the correlation into the correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0069] S6: Input the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes.

[0070] As described in step S1 above, user behavior and comment data from the target market region are acquired to form the raw dataset. The terminal can access local device cache (such as comments from installed applications, chat logs, and product usage logs) through an embedded acquisition module and, in compliance with relevant laws, regulations, and user agreements, obtain user-generated content from the target market region through public or authorized data interfaces. To ensure cross-cultural coverage, samples of relevant languages / dialects and mixed languages ​​from the target market should be actively screened, and the source platform, time, geographic tags, and user anonymization identifiers should be recorded. Data preprocessing includes noise reduction (deduplication, removal of advertising content), segmentation (sentence segmentation of long text), multimedia metadata extraction (emoticons, emojis, image captions), and privacy protection (de-identification, encryption, local storage, or homomorphic encryption before upload). The output is a structured raw dataset (text, symbol sequences, time / geographic metadata, and authorization identifiers). When implementing this on the terminal, bandwidth and storage limitations must be considered, and incremental synchronization, differentiated upload, and local caching strategies should be adopted.

[0071] As described in step S2 above, the original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors. The core of this process is to transform unstructured, multilingual, and multimodal inputs into unified and comparable implicit demand vectors. The multimodal parsing engine deployed on the terminal consists of two main sub-modules: a language semantic sub-module and a cultural symbol sub-module. The language semantic sub-module uses a lightweight multilingual pre-trained model (such as a distilled mBERT or a lightweight multilingual model) to segment the text, encode context, and extract intent slots, identifying explicit demand words (such as "battery life") and potential cues. The cultural symbol sub-module utilizes a local cultural symbol library (emoji mapping table, popular phrase mapping, local slang dictionary) to perform semantic parsing on non-textual symbols and cultural metaphors, mapping symbols to semantic vectors through contrastive learning or similarity retrieval. The vectors output by the two sub-modules are merged through attention fusion or pooling strategies to generate standardized implicit demand feature vectors (including confidence, source weight, and timestamp). To improve terminal efficiency, a hierarchical reasoning approach can be adopted: first, use rules / keywords for rapid screening, and then call deep encoding on suspected implicit demand samples to reduce computational load.

[0072] As described in step S3 above, the corresponding functional feature vector is obtained based on the implicit requirement feature vector. The purpose is to map or retrieve implementable functional descriptions and their feature representations from the requirement vector. The terminal first performs semantic retrieval on the implicit requirement vector in a preset functional knowledge base (local or segmented cached remote library), using vector similarity search (such as nearest neighbor search, ANN) to return several candidate functional entries (functional description, implementation method, associated cost / weight / regulatory tags). Then, combining historical matching data (previous similar requirement function selections and market feedback) or real-time engineering test feedback (locally synchronized test summaries), the candidate functions are scored and ranked. The finally selected functional entries undergo structured parsing (converting text descriptions into functional feature vectors, including material attributes, process parameters, alternative solutions, etc.), and are accompanied by preliminary implementation constraints and confidence levels. The terminal can implement lightweight retrieval and scoring logic; for highly complex retrievals, it can asynchronously request a larger semantic library from the server to obtain an expanded candidate set.

[0073] As described in step S4 above, a preliminary mapping matrix is ​​obtained by using a graph neural network, with the implicit demand feature vector as demand nodes and the functional feature vector as functional nodes. The terminal constructs a graph data structure locally: demand nodes contain implicit demand vectors, time / location labels, and source confidence; functional nodes contain functional feature vectors and implementation constraint metadata. A graph neural network (GNN) model is used to propagate information between nodes and generate node representations and edge features. To adapt to terminal computing resources, a lightweight GNN architecture (such as a simplified version of the graph attention network GAT or a distilled model of the message passing network MPNN) can be adopted, initialized with pre-trained parameters. After one or several rounds of message passing, the node representations are updated, and the similarity or relevance scores between nodes are calculated to generate a preliminary mapping matrix (matrix elements represent the initial matching scores and candidate paths of demand i and function j). In terminal implementation, if the number of candidate nodes is large, candidate filtering (Top-K) can be performed before GNN inference to control the computational load.

[0074] As described in step S5 above, the correlation between the implicit demand feature vector and the functional feature vector is obtained, and the correlation is converted into correlation strength and input into the preliminary mapping matrix to obtain the target mapping matrix. This process is responsible for converting dynamic data feedback into trainable weights, completing the transformation from the preliminary matrix to the target matrix. The correlation information comes from: historical market feedback (sales volume, returns, user ratings), engineering test data (drop tests, damp heat cycling, and other quantitative indicators), and A / B trials or small-batch validation results. After collecting local or synchronous feedback data on the terminal, incremental learning or online learning mechanisms are used, with the feedback signal serving as a supervision term for the loss function. Iterative training / fine-tuning of the GNN is used to update the edge weights (correlation strength). The transformation steps include normalizing feedback from different sources (assigning weights to sources), mapping feedback to the gradient update direction of the edges (e.g., reducing weights for performance degradation and increasing weights for success cases), and regularizing the weights to avoid overfitting or weight explosion. The output is the target mapping matrix, with matrix elements representing the final correlation strength (which may include a confidence interval). This matrix can be used as input for subsequent model optimization. To ensure real-time performance of the terminal, a threshold-triggered local update strategy can be adopted, adjusting only the edges that are significantly affected by feedback.

[0075] As described in step S6 above, the target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement solutions. The correlation results are transformed into executable design / improvement solutions. The preset optimization model can be a multi-constraint optimization engine (at the terminal, it can be a rule engine + lightweight optimizer, and if necessary, it can be linked with a large cloud model). Its inputs include the target mapping matrix, product development constraints (cost, supply chain availability, regulatory requirements), and priority objectives (performance, cost, time window). The optimizer performs a feasibility assessment on each requirement-function combination and generates several levels of candidate solutions (e.g., low-cost solutions, performance-priority solutions, regulatory-priority solutions). Each solution includes a function implementation path, estimated cost, and risk score. At the terminal, preliminary solutions can be quickly generated using heuristic algorithms (simplified versions of greedy and genetic algorithms), and complex cloud optimization (e.g., multi-constraint reasoning based on a large language model) can be initiated for high-priority solutions. The final output candidate solutions will include interpretable explanations (why the function was chosen, the basis for the correlation strength, and the expected improvement range) and simulation or small-sample verification suggestions to facilitate product team decision-making. During implementation, attention should be paid to solution version management and traceability records to facilitate subsequent feedback back to S5 to form a closed loop.

[0076] In one embodiment, the multimodal demand parsing engine includes a multilingual pre-trained model and a regional cultural symbol library. Step S2, which processes the original dataset using the pre-set multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors, includes:

[0077] S201: The first feature vector is generated by parsing the text semantics in the original dataset through the multilingual pre-trained model, and the second feature vector is generated by parsing the non-textual symbols in the original dataset through the regional cultural symbol library.

[0078] S202: The first feature vector and the second feature vector are fused to obtain the implicit demand feature vector.

[0079] As described in step S201 above, the text semantics in the original dataset are parsed by the multilingual pre-trained model to generate a first feature vector, and the non-textual symbols in the original dataset are parsed by the regional cultural symbol library to generate a second feature vector. The goal is to transform multilingual, mixed-language, and unstructured text information into semantic vectors that can be processed by subsequent computation. First, the text is preprocessed, including language recognition (detecting the main language and interspersed words), word segmentation or sub-word segmentation, noise reduction (removing advertisements and meaningless symbols), normalization (slang and spelling correction), and context aggregation (merging multiple statements within the same conversation or time window). Then, the processed text is encoded using a multilingual pre-trained model (such as distilled mBERT, XLM-R, or a custom lightweight variant). When encoding, the model considers contextual dependencies, syntactic structure, sentiment color, and intent slots, and outputs a first feature vector with fixed dimensions. This vector typically contains semantic representations of explicit demand terms (such as "battery life") and implicit cues (pain points inferred from context), carrying meta-information such as confidence level, linguistic labels, and source weights. Simultaneously, it parses non-textual symbols (emojis, emoticons, regional symbols, specific phrases, or image captions): using a pre-built regional cultural symbol library, these non-textual inputs are mapped to semantic labels or vector spaces. The regional cultural symbol library is trained through labeled samples and contrastive learning, capable of recognizing, for example, the camel emoji often implicitly signifies "durability" in a certain region, or the specific emotional connotation of a phrase in a particular place. The parsing process includes symbol standardization (unifying similar emojis and variations), context disambiguation (the same emoji may have different meanings in different contexts), and metadata association (publishing platform, user demographics). The parsing result is encoded into a second feature vector, containing a symbolic semantic vector, parsing confidence level, and cultural domain identifier. Finally, the first and second feature vectors serve as inputs for subsequent fusion, retaining source and confidence information for weighted fusion and tracing. This regional cultural symbol database can be constructed by crawling social media data from the target region, manually annotating it, or automatically associating symbols with text descriptions using cross-modal pre-trained models (such as CLIP). During parsing, the similarity of the input symbols with the symbols in the database is calculated, and the semantic label of the most similar symbol is taken as the parsing result.

[0080] As described in step S202 above, the first feature vector and the second feature vector are fused to obtain the implicit demand feature vector. The goal is to synthesize the text semantic vector and the cultural symbol vector into a unified and robust implicit demand representation for subsequent retrieval and matching. Fusion strategies can be divided into two categories: shallow fusion and deep fusion. Shallow fusion obtains a joint representation by concatenating vectors and then performing linear transformation and nonlinear activation, which is suitable for resource-constrained terminal scenarios. Deep fusion uses cross-modal attention or multimodal Transformer, allowing text vectors and symbol vectors to be queries / keys / values ​​to capture fine-grained cross-modal interactions (such as the situation of enhanced facial expressions or reversed text meaning in the same sentence). When fusing, source confidence and language context should be considered: for texts from low-confidence sources or with uncertain translations, their weight in the fusion should be reduced; for vectors with high confidence in cultural symbol parsing and strong regional relevance, their influence can be increased through gating mechanisms. Commonly used methods include weighted summation, post-concatenation projection, gating, and cross-modal self-attention. To ensure vector numerical stability, each input vector is first normalized and batch statistically corrected. Furthermore, the fusion process should include semantic consistency verification and time window aggregation: pooling (averaging or attention-weighted) multiple vectors from the same user or topic within a certain time period to enhance the capture of persistent implicit needs. At the training level, supervised fine-tuning (using labeled need-function pairs) and contrastive learning (positive / negative sample mining) are used to jointly optimize the fusion module, making the synthesized vectors more conducive to subsequent similarity retrieval and graph modeling. The final output implicit need feature vector is a unified-dimensional vector, accompanied by fusion confidence, timestamp, and source distribution information, for use in S3 and subsequent graph neural network modeling. During terminal implementation, depending on computing power, local lightweight fusion or uploading the original vectors to the server for deep fusion and fine-tuning can be selected.

[0081] In one embodiment, step S5, which involves obtaining the correlation between the implicit demand feature vector and the functional feature vector, and converting the correlation into a correlation strength input into the preliminary mapping matrix to obtain the target mapping matrix, includes:

[0082] S501: Collect market feedback data and engineering test data of historical product solutions to serve as the correlation between the implicit demand feature vector and the functional feature vector;

[0083] S502: Based on the market feedback data and engineering test data, the correlation strength of the edges between the demand node and the functional node is obtained through iterative learning of the graph neural network.

[0084] S503: Input the correlation strength into the preliminary mapping matrix to obtain the target mapping matrix.

[0085] As described in step S501 above, market feedback data and engineering test data of historical product solutions are collected to establish the correlation between the implicit demand feature vector and the functional feature vector. The key is to construct a multi-source feedback library for determining the effectiveness of demand-function matching. Market feedback data includes, but is not limited to: the sentiment polarity and topic clustering results of user review texts, conversion rates / repurchase rates / return rates after product launch, differences between the control and experimental groups in A / B testing, regional sales changes under channel distribution, and the number of complaints about specific functions in online customer service tickets. Engineering test data includes quantitative physical test results (such as drop counts and breakage probability, water ingress volume under IP protection level, and performance degradation percentage after damp heat cycling), laboratory measurements (current, voltage, thermal resistance, etc.), and reliability indicators from small-batch verification. During the collection process, the data needs to be uniformly labeled and structured: each feedback is associated with a corresponding implicit demand ID and functional ID, and a timestamp, source credibility tag (e.g., high credibility of laboratory data, high social media sentiment noise), and contextual metadata (region, model, firmware version) are added. In addition, data should be cleaned (outliers and duplicate records should be removed), normalized (different units of measurement should be standardized), missing values ​​should be imputed, and privacy should be protected (data should be desensitized, homomorphically encrypted, or stored locally and only aggregated statistics should be uploaded) to ensure that the feedback used for subsequent model training is comparable and interpretable.

[0086] As described in step S502 above, based on the market feedback data and engineering test data, the correlation strength between the edges of the demand nodes and the functional nodes is obtained through iterative learning of the graph neural network. Using the graph neural network (GNN) as the core, the feedback signals are mapped to edge weights (correlation strength) through supervised or semi-supervised iterative learning. First, the structured feedback in S501 is aggregated into training samples according to demand-function pairs: the label corresponding to each pair (demand i, function j) can be a continuous indicator (such as performance improvement, satisfaction score) or a discrete label (such as pass / fail). The GNN architecture can employ graph attention (GAT), GraphSAGE, or message passing network (MPNN), etc., to perform message passing and representation updates through node features (latent demand vector, functional feature vector) and initial edge features (candidate matching score, historical success rate). The training objective includes minimizing the loss (MSE or cross-entropy) between the predicted edge weights and the actual feedback mapping values, while introducing regularization terms (L2, edge sparsity constraints) and time decay terms (decreasing weights of older feedback) to prevent overfitting and response delay. To address feedback sparsity, contrastive learning (same / different sample pairs) can be combined to enhance representation capabilities, or a Bayesian / probabilistic graphical model can be used to provide confidence intervals for edge weights. At the deployment level, online incremental learning is supported: when new market or test data arrives, the GNN performs local fine-tuning to update the edge weights of the affected subgraphs; under data distribution protection or privacy requirements, federated learning can be used to aggregate locally updated gradients into global edge weight updates, ensuring cross-regional knowledge transfer without leaking the original data. The GNN employs a two-layer GraphSAGE aggregation layer, using the market success rate (e.g., sales growth rate) or engineering test pass rate of the "demand node-functional node" edges in historical solutions as supervision signals, and uses mean squared error loss for training to learn the correlation strength.

[0087] As described in step S503 above, the association strength is input into the preliminary mapping matrix to obtain the target mapping matrix, completing the transformation and delivery from model calculation to a usable mapping matrix. Specifically, the association strength of each edge obtained in S502 is used as a correction factor for the corresponding element of the preliminary mapping matrix, and the matrix value is updated according to a predetermined mapping rule (e.g., matrix element = α * preliminary matching score + β * learned edge weight). Normalization processing (e.g., softmax, min-max) should be performed before and after the update to ensure comparability between different rows, and the confidence level or confidence interval of each element is recorded as additional metadata. To reduce the impact of noise, a threshold or Top-K filtering can be set to retain only functional candidates with strengths higher than the threshold or higher rankings; at the same time, a hysteresis mechanism and smoothing filtering are introduced to prevent drastic fluctuations in the matrix due to a single abnormal feedback. After the target mapping matrix is ​​generated, version management is required (marking time, data window, and training model ID), and a change log should be provided for backtracking and auditing. The matrix can be stored on the terminal for rapid offline solution generation, or synchronized to the server for use in more complex optimization models. In terminal-constrained scenarios, only a high-confidence subset of the matrix can be transmitted to save bandwidth. Ultimately, the target mapping matrix, along with the source and confidence information of each element, serves as input to downstream optimization and simulation modules, ensuring that the generated solution is interpretable and verifiable.

[0088] In one embodiment, before step S6 of inputting the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes, the method further includes:

[0089] S511: Obtain the constraints of the target market region;

[0090] S512: Input the constraints into the initial optimization model to obtain the optimization model.

[0091] As described in step S511 above, obtaining the constraints of the target market region aims to comprehensively collect and structure various constraints imposed on product improvement solutions by the target market region before entering the solution optimization stage. This ensures that the subsequently generated candidate solutions are both feasible and compliant with legal, cost, and supply chain realities. Constraints include, but are not limited to: cost constraints (unit cost, target gross margin, R&D budget ceiling); supply chain and capacity constraints (availability of key components, minimum order quantity, production line capacity, delivery cycle); regulatory compliance constraints (safety standards, electromagnetic compatibility, RoHS / REACH, energy efficiency, and certification requirements of the target market); material and process constraints (permitted / prohibited materials, process availability, list of alternative materials); time window constraints (market launch deadline, seasonal demand); user experience and ecosystem compatibility constraints (size, interface compatibility, multi-SIM, and other market preferences); and corporate strategy constraints (brand positioning, differentiation requirements). When implementing on a terminal or server, each type of constraint needs further structuring: quantify quantifiable items (e.g., cost limits expressed numerically), transform regulatory items into verifiable compliance conditions (e.g., maximum lead content ≤ X ppm), map supply constraints to a resource availability matrix (component ID → available quantity → delivery cycle), and record the hard / soft attributes of constraints (hard constraints must be met, soft constraints can be traded off), priority, source of trust, and update timestamp. For constraints with uncertainty (e.g., component availability for the next three months), probability predictions or scenario assumptions must be included. All constraints undergo consistency verification after collection (detecting logical conflicts, unit inconsistencies, etc.), and the structured constraint set is stored in a standard format as the input basis for building the S512 optimization model.

[0092] As described in step S512 above, the constraints are input into the initial optimization model to obtain the optimization model. The structured constraint set is formalized into a computable optimization model and integrated with the objective mapping matrix and optimization objective to obtain an initial optimization model that can be used to solve candidate solutions. Implementation includes the following key steps: First, mathematical modeling of constraints—converting hard constraints into equality / inequality constraints (e.g., cost ≤ C, component usage ≤ inventory), and modeling soft constraints as penalty terms or weighted objectives (e.g., preference terms as weights of the objective function); applying logical constraints or verifiable rule sets to regulatory items (if a violation is detected, the solution is deemed infeasible); and using a resource constraint matrix for the supply chain and capacity, introducing a time dimension (multi-period planning). Second, determining the optimization objective (single or multiple objectives). Common objectives include minimizing cost, maximizing performance improvement, minimizing risk, or a weighted sum that trades off between multiple objectives. For multi-objective problems, Pareto front calculation, weighted sum, or hierarchical optimization strategies are used. Then, select an appropriate solution method: for discrete and integer quantization problems, integer / mixed-integer programming (MILP) or heuristic algorithms (genetic algorithms, simulated annealing) can be used; for continuous nonlinear problems, gradient methods or evolutionary strategies can be used; on resource-constrained terminals, lightweight heuristic algorithms can be prioritized, and complex solutions can be offloaded to the cloud. Furthermore, associate the correlation strength in the target mapping matrix with the constraint model: combine requirements and functions with high correlation strength in the model and set lower penalties or higher priorities to ensure that optimization tends to select high-confidence matching terms. For soft constraints and uncertainties, use robust optimization or scenario optimization methods (solve different constraint scenarios and compare results). After the model is built, a feasibility pre-check should be performed (quickly solve small-scale samples to verify consistency), and model metadata (constraint set version, weight settings, solver type, timestamp) should be saved for subsequent auditing and online updates. Finally, if a large language model is integrated as an interpretation layer, the mathematical model can be translated into readable rules and priority descriptions, facilitating understanding and adjustment by product decision-makers. Specifically, the elements of the target mapping matrix can be set. , representing the strength of the association between requirement i and function j. The cost of implementing function j. The optimization objective is to maximize the total association strength. The constraint is the total cost. ,and ∈{0,1} indicates whether to select function j, B is the budget, and a set of Pareto optimal candidate solutions is obtained by using a greedy algorithm or an integer programming solver.

[0093] In one embodiment, after step S6 of inputting the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes, the method further includes:

[0094] S701: Obtain environmental data for the target market area;

[0095] S702: Construct a digital twin simulation system based on the environmental data;

[0096] S703: Input each of the candidate product improvement schemes into the digital twin simulation system to verify the candidate product improvement schemes and obtain the verification results of each candidate product improvement scheme;

[0097] S704: Based on the verification results, output the performance simulation report and feasibility assessment of each candidate product improvement scheme.

[0098] As described in step S701 above, environmental data for the target market area is acquired. The purpose is to collect and structure real environmental parameters of the target market in order to subsequently build a digital twin simulation system that reflects the usage scenarios in that area. Environmental data includes, but is not limited to, meteorological data (temperature, relative humidity, rainfall, wind speed, air pressure), atmospheric composition (salt spray, conductivity, dust / particulate matter concentration, sandstorm frequency), altitude and solar radiation intensity, seasonal fluctuations, characteristics of power and communication infrastructure (power supply stability, charging voltage fluctuations, mobile network coverage / bandwidth / latency), user habits and typical operating conditions (continuous use duration, charging frequency, carrying method), and the physical exposure conditions of the product in the area (indoor / outdoor usage ratio, coastal / inland, etc.). For example, for mobile phone drop test simulation, the digital twin system includes a finite element model of the shell based on CAD files, with the material properties set as polycarbonate (elastic modulus 2.4 GPa), and the boundary conditions set as impacting a marble floor from a height of 1.5 meters at a specific angle. The simulation output is a maximum stress cloud map of the shell; if the stress exceeds the material yield strength, the solution is deemed infeasible. Data sources may include public meteorological databases, third-party environmental monitoring platforms, local sensor networks (customer deployments or test sites), market research data, and historical after-sales records. During data collection, time-series synchronization, missing value completion, outlier detection, and smoothing are required. Data should be stratified and labeled according to geographic grids or user groups (urban / rural, coastal / inland, different climate zones). To support uncertainty analysis in simulations, the statistical distribution of various environmental parameters (mean, variance, frequency of extreme values) should be estimated, and a set of scenarios (typical scenarios, extreme scenarios, probabilistic scenarios) should be constructed. The final output is a structured environmental dataset and several preset scenarios.

[0099] As described in step S702 above, a digital twin simulation system is constructed based on the environmental data. The environmental parameters obtained in S701 are integrated with the product's physical, material, and functional models to construct a digital twin system capable of replicating the behavior of the real object in software. First, a virtual product model needs to be established, including geometric modeling, a material property library (elastic modulus, coefficient of thermal expansion, electrical conductivity, corrosion resistance, etc.), structural connection relationships, and functional models of internal components (batteries, circuits, speakers, etc.). Then, the environmental data is mapped to simulation boundary conditions, such as temperature and humidity cycle curves, salt spray concentration time series, impact / vibration spectra, continuous drop distribution, or communication packet loss rate models. At the simulation level, multiphysics joint solution (thermal-mechanical-humidity-electrical coupling), finite element analysis (FEA), multibody dynamics, fluid dynamics (for water ingress / dustproof modeling), and network simulation targeting software or communication performance can be employed. The implementation platform can use Unity3D for scene-driven and rapid prototyping verification, combined with dedicated solvers (ANSYS, COMSOL) to achieve high-precision engineering simulation. It should also support scenario parameterization, parallel batch simulation, and Monte Carlo uncertainty assessment to obtain robust performance predictions under different environmental scenarios. After the model is built, baseline validation (comparing it with known measured data) should be performed to calibrate the model parameters and confidence levels.

[0100] As described in step S703 above, each candidate product improvement scheme is input into the digital twin simulation system to verify the candidate product improvement scheme and obtain the verification results of each candidate product improvement scheme. The candidate scheme generated in S6 is converted into a simulable virtual configuration and the verification process is executed. First, each candidate scheme is virtualized: the component parameters in the digital twin are replaced or modified (such as material replacement, structural reinforcement, coating thickness, battery capacity, speaker grille design, etc.), and the corresponding production and assembly tolerances, supply chain substitute information and actual implementation constraints are injected. Then, the simulation is run one by one in the preset scenario set, including normal scenarios and extreme scenarios, and key performance indicators (such as the probability of damage caused by drop, functional failure rate in humid environments, thermal performance degradation, battery life estimation, signal attenuation and sound quality changes, etc.) are recorded. To improve reliability, batch parallel simulation and random disturbance (Monte Carlo) are used to evaluate the robustness of the scheme, and the average performance, variance and confidence interval are calculated. Failure modes and impact paths should also be collected during the simulation process for subsequent interpretability analysis. The output verification results include quantitative indicator time series, failure rate distribution, key failure scenarios and corresponding root cause descriptions, and generate reproducible simulation logs and model version information for each solution to facilitate auditing and result verification.

[0101] As described in step S704 above, based on the verification results, output performance simulation reports and feasibility assessments for each candidate product improvement scheme. Organize the simulation outputs into a structured evaluation report to support decision-makers' comparison and selection. The report should include a scheme description, input parameters and constraints, the digital twin version used and model calibration information, a list of environmental scenarios used, key performance indicators (KPIs) and their statistical summaries (mean, variance, 95% confidence interval), sub-item performance in different scenarios, main failure modes and probabilities, simulation estimates of cost and supply chain impact, compliance check results (whether regulatory thresholds are met), and recommendations for further verification (e.g., the type and sample size of physical tests to be completed). Furthermore, a feasibility score based on preset weights (comprehensively considering performance, cost, risk, and compliance) should be provided, and candidate schemes should be ranked by priority; sensitivity analysis results should be provided for key uncertainties, explaining which parameter changes would significantly alter the scheme ranking. The report also needs to include recommended actions (such as immediate trial production, small-batch verification, or further material testing) and corresponding time / cost estimates, and link the report results back to the S5 and S6 modules to update the confidence level of the mapping matrix and the constraints of the optimization model to form a closed-loop continuous optimization process.

[0102] In one embodiment, after step S704 of outputting performance simulation reports and feasibility assessments of each candidate product improvement scheme based on each verification result, the method further includes:

[0103] S7051: Calculate the dimension value of each of the candidate product improvement schemes in each preset dimension;

[0104] S7052: The comprehensive score of each candidate product improvement scheme is obtained by weighted summation of the dimension values ​​of each candidate product improvement scheme.

[0105] S7053: Sort the candidate product improvement schemes according to the comprehensive scores and send the sorted schemes to the designated personnel.

[0106] As described in step S7051 above, the dimensional values ​​of each candidate product improvement scheme in each preset dimension are calculated. Based on the performance simulation report and feasibility assessment output in S704, each candidate scheme is quantified into comparable dimensional values ​​on the predefined evaluation dimensions. The preset dimensions typically include, but are not limited to: functional performance (such as drop resistance probability, improved battery life), cost (manufacturing cost, unit cost change), supply chain availability (key component availability score), regulatory compliance (whether certification thresholds are met, violation risk score), manufacturability (process complexity, capacity fit score), user experience (satisfaction prediction in simulation or user testing), risk and reliability (failure rate, repair cost estimation), and time to market (estimated cycle from design to production). For each dimension, clear quantification rules need to be defined: simulation indicators are directly mapped to numerical values ​​(e.g., drop damage rate → 1 - damage rate), costs and time are converted into a unified currency or time unit, regulations are represented by binary or graded scores, and qualitative items are converted into scores using expert scoring or probability distributions based on model predictions. During execution, the values ​​of different dimensions should be normalized (e.g., linearly scaled to the 0–1 range or using Z-scores), and the confidence level and source (simulation model ID, data time window, sample size) of each dimension value should be recorded for subsequent weighted synthesis and sensitivity analysis.

[0107] As described in step S7052 above, the comprehensive score of each candidate product improvement scheme is obtained by weighted summation based on the dimensional values ​​of each scheme. The standardized dimensional values ​​obtained in S7051 are then combined into a single comprehensive evaluation score by weights that are preset or dynamically determined. Weights can come from various mechanisms: fixed strategies (e.g., weight allocation determined by corporate strategy), expert scoring or stakeholder voting (determined jointly by product, engineering, legal, and procurement departments), or through historical data and machine learning methods (learning optimal weights based on the success rate of past schemes). A commonly used mathematical implementation is the weighted sum model: Comprehensive Score = Σ (w_k * v_k), where w_k is the weight of the k-th dimension and v_k is the normalized dimensional value. For multiple or conflicting objectives, AHP, TOPSIS, or Pareto ranking can be used for supplementary judgment; for soft constraints, penalty terms are used (e.g., deducting high scores for violating legal rules); for uncertainties in the input dimensions, Monte Carlo simulation or interval analysis is used to generate confidence intervals for the comprehensive score, or robust / scenario weighting is used to obtain stable scores under different constraint scenarios. The final result should output the point estimate score, confidence interval, weight distribution details, and contribution rate of each dimension to the total score for each candidate solution, so that decision-makers can understand the causal relationship of the score formation and make sensitivity judgments.

[0108] As described in step S7053 above, the candidate product improvement schemes are sorted according to the comprehensive score, and the sorted schemes are sent to designated personnel responsible for integrating the final decision information and conveying it to relevant personnel in the decision-making chain. First, the candidate schemes are sorted in descending order of comprehensive score, and then divided into several levels or queues according to business rules (e.g., Top-Tier schemes can directly enter small-batch trial production, Mid-Tier schemes require further testing, and Low-Tier schemes are not considered for the time being). The output should include key supporting information for each scheme: comprehensive score, scores and weights for each dimension, key simulation indicators, main risk items and confidence intervals, recommended actions (trial production / prototype testing / material substitution, etc.), and change logs (model version, data time window). The delivery method can be diversified: pushed to designated personnel and teams through a product decision dashboard (interactive dashboard), or automatically generated structured reports sent to engineering, procurement, legal, and senior approval personnel via email or API. The report should include a drill-down proof chain (tracing from simulation results to input data and model ID) and provide feedback entry points to allow the recipient to adjust weights or trigger the next round of simulation / testing. Finally, the distribution process is audited and recorded (recipient, time, version), and the receiver feedback loop is connected to the S5 / S6 module to update the mapping matrix and optimize the model, forming a closed-loop iteration.

[0109] In one embodiment, step S3, which involves obtaining the corresponding functional feature vector based on the implicit demand feature vector, includes:

[0110] S301: Based on the implicit requirement feature vector, search for and match multiple candidate function feature vectors from the preset product function knowledge base;

[0111] S302: Based on historical matching data or real-time engineering test feedback, determine the functional feature vector with the highest correlation to the implicit requirement feature vector from the multiple candidate functional feature vectors and output it.

[0112] As described in step S301 above, based on the implicit requirement feature vector, multiple candidate function feature vectors are searched from a preset product function knowledge base. The aim is to perform a preliminary match between the extracted implicit requirement vectors and existing function knowledge to generate a candidate set for subsequent filtering. First, the preset product function knowledge base should be stored in structured entries. Each entry includes a function description text, a function feature vector (encoded through the same vectorization process), implementation constraints (materials, cost range, process requirements), historical adaptation tags, and regional adaptation metadata. During implementation, the implicit requirement feature vector is used as the query vector. A vector retrieval engine (such as Faiss-based ANN retrieval, nearest neighbor search, or a combination of inverted index and semantic retrieval) is used to retrieve several Top-K candidate function entries from the knowledge base. The retrieval can employ a multi-stage strategy: the first stage uses low-latency approximate retrieval to quickly obtain the candidate set; the second stage uses more precise similarity metrics (cosine similarity, Euclidean distance, or reordering based on interactive attention) and rule filtering (such as excluding functions that conflict with target market regulations) on the candidate set to refine the results. During the retrieval process, the similarity score, source tag, and knowledge base version number of each candidate option are recorded, along with preliminary implementation feasibility metadata, for downstream scoring and tracing. For scenarios with limited terminal resources, a subset of locally cached functions can be retrieved first, with an expanded candidate set asynchronously requested from the server if necessary. The key to this step is ensuring that the knowledge base representation and the demand vector are in the same semantic space, and maintaining retrieval efficiency and candidate diversity to prevent the early elimination of potential innovative solutions.

[0113] As described in step S302 above, based on historical matching data or real-time engineering test feedback, the functional feature vector with the highest correlation to the implicit requirement feature vector is determined from the multiple candidate functional feature vectors and output. This is responsible for scoring and ultimately selecting the candidate function set obtained in S301 based on evidence. First, historical matching data (e.g., historical performance of this requirement type and function combination, user satisfaction, return rate, A / B test results) and real-time engineering test feedback (prototype test results, reliability indicators, material compatibility tests) are collected and correlated. For each candidate function, a comprehensive scoring model is constructed: the semantic similarity score is weighted with indicators such as historical success rate, engineering test performance, and regulatory and supply chain constraint adaptability according to predefined weights or weights obtained through training to obtain the final correlation score. The score can be normalized and a confidence interval can be output. For candidates with sparse data, Bayesian smoothing or model-based inference is used to compensate for their confidence. If scores are close or conflicting, decision-making strategies can be introduced, such as retaining multiple tied options for the next step (Top-N), or triggering manual review / LLM to generate explanations to assist in the selection. Furthermore, the implementation should support an online learning mechanism: when new market feedback or test data arrives, the weights and biases of the scoring model should be automatically adjusted, so that the selection results converge to a higher market fit over time. The final output feature vector should not only contain the vector ontology, but also carry the selection rationale, confidence level, association history, and implementation constraints for subsequent graph construction and model optimization.

[0114] Reference Figure 3 The present invention also provides a cross-cultural demand intelligent mining device, the device comprising:

[0115] The first acquisition module 902 is used to acquire user behavior and comment data in the target market area to form an original dataset; wherein, the original dataset includes comment data in at least two languages;

[0116] The identification module 904 is used to process the original dataset through a preset multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0117] The second acquisition module 906 is used to acquire the corresponding functional feature vector based on the implicit requirement feature vector;

[0118] Module 908 is used to utilize a graph neural network to obtain a preliminary mapping matrix by using the implicit demand feature vector as demand nodes and the functional feature vector as functional nodes.

[0119] The transformation module 910 is used to obtain the correlation between the implicit demand feature vector and the functional feature vector, and to convert the correlation into the correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0120] The generation module 912 is used to input the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes.

[0121] In one embodiment, the identification module 904 includes:

[0122] The parsing submodule is used to parse the text semantics in the original dataset using the multilingual pre-trained model to generate a first feature vector, and to parse the non-text symbols in the original dataset using the regional cultural symbol library to generate a second feature vector;

[0123] The fusion submodule is used to fuse the first feature vector and the second feature vector to obtain the implicit demand feature vector.

[0124] In one embodiment, the conversion module 910 includes:

[0125] The collection submodule is used to collect market feedback data and engineering test data of historical product solutions, so as to establish the correlation between the implicit demand feature vector and the functional feature vector.

[0126] The learning submodule is used to obtain the correlation strength of the edges between the demand nodes and the functional nodes through iterative learning of the graph neural network based on the market feedback data and engineering test data.

[0127] The input submodule is used to input the correlation strength into the preliminary mapping matrix to obtain the target mapping matrix.

[0128] In one embodiment, the cross-cultural needs intelligent mining device further includes:

[0129] A constraint acquisition module is used to acquire the constraints of the target market area;

[0130] The optimization model acquisition module is used to input the constraints into the initial optimization model to obtain the optimization model.

[0131] In one embodiment, the cross-cultural needs intelligent mining device further includes:

[0132] An environmental data acquisition module is used to acquire environmental data of the target market area;

[0133] A construction module is used to build a digital twin simulation system based on the environmental data;

[0134] The verification module is used to input each of the candidate product improvement schemes into the digital twin simulation system to verify the candidate product improvement schemes and obtain the verification results of each candidate product improvement scheme.

[0135] The output module is used to output performance simulation reports and feasibility assessments of each candidate product improvement scheme based on the verification results.

[0136] In one embodiment, the cross-cultural needs intelligent mining device further includes:

[0137] The calculation module is used to calculate the dimensional values ​​of each of the candidate product improvement schemes in each preset dimension;

[0138] The weighted summation module is used to perform weighted summation based on the dimension values ​​of each candidate product improvement scheme to obtain a comprehensive score for each candidate product improvement scheme.

[0139] The sending module is used to sort the candidate product improvement schemes according to the comprehensive score and send the sorted schemes to designated personnel.

[0140] In one embodiment, the second acquisition module 906 includes:

[0141] The search submodule is used to search for multiple candidate function feature vectors from a preset product function knowledge base based on the implicit requirement feature vectors.

[0142] The determination submodule is used to determine, based on historical matching data or real-time engineering test feedback, the functional feature vector with the highest correlation to the implicit requirement feature vector from the multiple candidate functional feature vectors and output it.

[0143] Figure 4 An internal structural diagram of an electronic device in one embodiment is shown. This electronic device can specifically be a terminal or a server, and more specifically, a computer device. Figure 4 As shown, the electronic device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a cross-cultural needs intelligent mining method. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the cross-cultural needs intelligent mining method. Those skilled in the art will understand that... Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0144] In one embodiment, an electronic device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps:

[0145] User behavior and comment data from the target market region are acquired to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0146] The original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0147] Obtain the corresponding functional feature vector based on the implicit demand feature vector;

[0148] Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix;

[0149] Obtain the correlation between the implicit demand feature vector and the functional feature vector, and convert the correlation into correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0150] The target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement schemes.

[0151] By introducing a multimodal demand parsing engine to process the original dataset containing multilingual mixed data, the system can simultaneously parse textual semantics and cultural symbols, improving the coverage and accuracy of implicit demand identification and feature extraction in cross-cultural scenarios. This overcomes the dependence of traditional methods on single-language or structured data. Then, a dynamic mapping matrix based on demand nodes and functional nodes is constructed using graph neural networks. This makes the relationship between demand and solution no longer a static rule, but can be continuously optimized through dynamic learning and updating of the relationship strength. This greatly enhances the market adaptability and accuracy of solution generation. The optimized target mapping matrix is ​​input into a preset optimization model, which can efficiently generate and screen feasible candidate product improvement solutions under the premise of considering multidimensional constraints. This systematizes and automates the transformation process from demand to solution, shortens the product development cycle, and reduces decision-making costs.

[0152] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps:

[0153] User behavior and comment data from the target market region are acquired to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages;

[0154] The original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors.

[0155] Obtain the corresponding functional feature vector based on the implicit demand feature vector;

[0156] Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix;

[0157] Obtain the correlation between the implicit demand feature vector and the functional feature vector, and convert the correlation into correlation strength and input it into the preliminary mapping matrix to obtain the target mapping matrix;

[0158] The target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement schemes.

[0159] By introducing a multimodal demand parsing engine to process the original dataset containing multilingual mixed data, the system can simultaneously parse textual semantics and cultural symbols, improving the coverage and accuracy of implicit demand identification and feature extraction in cross-cultural scenarios. This overcomes the dependence of traditional methods on single-language or structured data. Then, a dynamic mapping matrix based on demand nodes and functional nodes is constructed using graph neural networks. This makes the relationship between demand and solution no longer a static rule, but can be continuously optimized through dynamic learning and updating of the relationship strength. This greatly enhances the market adaptability and accuracy of solution generation. The optimized target mapping matrix is ​​input into a preset optimization model, which can efficiently generate and screen feasible candidate product improvement solutions under the premise of considering multidimensional constraints. This systematizes and automates the transformation process from demand to solution, shortens the product development cycle, and reduces decision-making costs.

[0160] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0161] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0162] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A cross-cultural needs intelligent mining method, characterized in that, The method includes: User behavior and comment data from the target market region are acquired to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages; The original dataset is processed by a pre-defined multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors. The corresponding functional feature vector is obtained based on the implicit demand feature vector; wherein, vector similarity search is performed in a preset functional knowledge base based on the implicit demand feature vector to obtain the corresponding functional feature vector. Using a graph neural network, the implicit demand feature vector is used as a demand node, and the functional feature vector is used as a functional node to obtain a preliminary mapping matrix; The process involves: obtaining the correlation between the implicit demand feature vector and the functional feature vector, and converting the correlation into a correlation strength, which is then input into the preliminary mapping matrix to obtain the target mapping matrix. Specifically, this involves: collecting market feedback data and engineering test data from historical product solutions to serve as the correlation between the implicit demand feature vector and the functional feature vector; based on the market feedback data and engineering test data, obtaining the correlation strength between the edges connecting the demand nodes and the functional nodes through iterative learning of the graph neural network; and inputting the correlation strength into the preliminary mapping matrix to obtain the target mapping matrix. The target mapping matrix is ​​input into a preset optimization model to generate multiple candidate product improvement schemes. The preset optimization model is a multi-constraint optimization engine, whose inputs include the target mapping matrix, product development constraints, and priority targets. Each candidate product improvement scheme generated by the engine includes a functional implementation path, estimated cost, and risk score.

2. The cross-cultural needs intelligent mining method according to claim 1, characterized in that, in, The multimodal demand parsing engine includes a multilingual pre-trained model and a regional cultural symbol library. The step of processing the original dataset through the pre-set multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors includes: The first feature vector is generated by parsing the text semantics in the original dataset using the multilingual pre-trained model, and the second feature vector is generated by parsing the non-textual symbols in the original dataset using the regional cultural symbol library. The first feature vector and the second feature vector are fused to obtain the implicit demand feature vector.

3. The cross-cultural needs intelligent mining method according to claim 1, characterized in that, After the step of inputting the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes, the method further includes: Obtain environmental data for the target market area; A digital twin simulation system is constructed based on the aforementioned environmental data; Each of the candidate product improvement schemes is input into the digital twin simulation system to verify the candidate product improvement schemes and obtain the verification results of each candidate product improvement scheme. Based on the verification results, output the performance simulation report and feasibility assessment of each candidate product improvement scheme.

4. The cross-cultural needs intelligent mining method according to claim 3, characterized in that, After the step of outputting the performance simulation report and feasibility assessment of each candidate product improvement scheme based on each verification result, the method further includes: Calculate the dimensional values ​​of each of the candidate product improvement schemes in each preset dimension; The comprehensive score of each candidate product improvement scheme is obtained by weighted summation of the dimensional values ​​of each candidate product improvement scheme. The candidate product improvement proposals are sorted according to the comprehensive scores, and the sorted proposals are sent to designated personnel.

5. The cross-cultural needs intelligent mining method according to claim 1, characterized in that, The step of obtaining the corresponding functional feature vector based on the implicit demand feature vector includes: Based on the implicit demand feature vector, search for and match multiple candidate function feature vectors from the preset product function knowledge base; Based on historical matching data or real-time engineering test feedback, the functional feature vector with the highest correlation to the implicit requirement feature vector is determined from the multiple candidate functional feature vectors and then output.

6. A cross-cultural demand intelligent mining device, characterized in that, The device includes: The first acquisition module is used to acquire user behavior and comment data from the target market area to form a raw dataset; wherein, the raw dataset includes comment data in at least two languages; The identification module is used to process the original dataset through a preset multimodal demand parsing engine to identify and extract cross-cultural implicit demand feature vectors. The second acquisition module is used to acquire the corresponding functional feature vector based on the implicit demand feature vector; wherein, a vector similarity search is performed in a preset functional knowledge base based on the implicit demand feature vector to acquire the corresponding functional feature vector. The module is used to utilize a graph neural network to obtain a preliminary mapping matrix by using the implicit demand feature vector as demand nodes and the functional feature vector as functional nodes. The transformation module is used to obtain the correlation between the implicit demand feature vector and the functional feature vector, and to convert the correlation into a correlation strength, which is then input into the preliminary mapping matrix to obtain the target mapping matrix. Specifically, it involves: collecting market feedback data and engineering test data from historical product solutions as the correlation between the implicit demand feature vector and the functional feature vector; based on the market feedback data and engineering test data, obtaining the correlation strength of the edges between the demand nodes and the functional nodes through iterative learning of the graph neural network; and inputting the correlation strength into the preliminary mapping matrix to obtain the target mapping matrix. The generation module is used to input the target mapping matrix into a preset optimization model to generate multiple candidate product improvement schemes. The preset optimization model is a multi-constraint optimization engine, whose inputs include the target mapping matrix, product development constraints, and priority targets. Each candidate product improvement scheme generated by the engine includes a functional implementation path, estimated cost, and risk score.

7. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, causes the processor to perform the steps of the cross-cultural needs intelligent mining method as described in any one of claims 1 to 5.

8. An electronic device, characterized in that, The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the cross-cultural needs intelligent mining method as described in any one of claims 1 to 5.

Images