Multi-source data input AI engine user demand analysis method

Abstract

The invention provides an AI engine user demand analysis method for multi-source data input, relates to the field of artificial intelligence, and solves the technical problems that in the prior art, relevance and dynamic insight of user demands are insufficient, and user motivation is difficult to mine. The method comprises the following steps: acquiring multi-modal data from different data sources, and carrying out preprocessing and semantic fusion to generate a multi-modal data stream with unified representation; wherein the data source comprises a public network data source and an Internet of Things data source; based on the multi-modal data flow, through a graph neural network, constructing a dynamic demand graph with demand entities as nodes and association relationships as edges; based on the dynamic demand graph, identifying user demands and user motivations by using graph reasoning and causal discovery algorithms, and generating a structured demand analysis report; and based on real-time monitoring data, updating the dynamic demand map through an incremental learning algorithm, and carrying out continuous tracking of demand evolution.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically a method for analyzing user needs of an AI engine with multi-source data input. Background Technology

[0002] In today's fiercely competitive market environment, accurately understanding user needs has become a core basis for enterprises to innovate products, optimize service experiences, and even formulate market strategies. Traditional user needs analysis methods, such as questionnaires, user interviews, and focus groups, while directly reaching users, generally suffer from limited sample size, strong subjectivity, long implementation cycles, and high costs, making it difficult to achieve large-scale, real-time needs insights. With the popularization of Internet and IoT technologies, the dimensions and total amount of data that enterprises can obtain have increased dramatically, including user reviews on e-commerce platforms, discussions on social media, offline retail scenarios such as customer traffic and transaction records, and a large number of product and scene images. This provides an unprecedented data foundation for needs analysis.

[0003] However, most existing demand analysis solutions based on big data technology still focus on processing text data, such as generating word clouds or trend reports through keyword extraction and sentiment analysis. These methods struggle to construct complete user profiles and behavioral paths. Secondly, the analysis results are often presented as isolated statistical indicators or keyword lists, lacking the ability to uncover the complex relationships between demand points and reveal the underlying logic and user motivations. Furthermore, traditional methods are typically static and lagging, unable to adapt to the rapid and dynamic evolution of market demands and failing to promptly identify emerging needs or shifts in preferences. Specifically, in complex scenarios such as offline retail, existing methods cannot correlate customer behavior at the shelf with their final purchase decision and subsequent online reviews, thus missing crucial opportunities to understand the pain points and pleasure points in shopping behavior. Summary of the Invention

[0004] This application provides a user demand analysis method for AI engines with multi-source data input, which solves the technical problems of insufficient insight into the correlation and dynamism of user demand and difficulty in mining user motivation in existing technologies.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, a method for analyzing user needs of AI engines with multi-source data input is provided, including: Multimodal data from different data sources is acquired, preprocessed, and semantically fused to generate a multimodal data stream with a unified representation; wherein, the data sources include public network data sources and Internet of Things (IoT) data sources; Based on the multimodal data stream, a dynamic demand graph is constructed using a graph neural network, with demand entities as nodes and relationships as edges. Based on the dynamic demand graph, graph reasoning and causal discovery algorithms are used to identify user needs and user motivations, and generate a structured demand analysis report. Based on real-time monitoring data, the dynamic demand map is updated through an incremental learning algorithm to continuously track the evolution of demand.

[0006] Based on the above technical solutions, the AI ​​engine user demand analysis method with multi-source data input provided in this application achieves multi-dimensional characterization of users by integrating heterogeneous data from multiple sources such as public networks and the Internet of Things, laying a solid data foundation for a comprehensive understanding of user needs. Secondly, a dynamic demand graph is constructed using graph neural networks, transforming isolated demand information into a networked knowledge structure with rich relationships. This not only intuitively displays the complex relationships between demands but also leverages the characteristics of graph structures for deep reasoning. Furthermore, by introducing causal discovery algorithms, the method aims to reveal the deep user motivations behind demands, thereby providing more guiding insights for product innovation and optimization. Finally, continuous self-updating is achieved through incremental learning algorithms, enabling the capture of changes in market trends and user preferences, improving the timeliness and foresight of demand analysis.

[0007] Furthermore, the data preprocessing and data fusion include: Text data is processed using natural language processing algorithms and named entity recognition technology, and a five-tuple structured text representation containing product, product function, usage scenario, user attribute, and user evaluation is output. Sensor data is processed using signal processing algorithms, transforming the sensor data stream into an event-based feature sequence that reflects device status or user interaction; wherein, the sensor data includes customer flow heatmap data, POS transaction data, and shelf interaction sensing data collected in offline retail scenarios; The scene graph generation model is used to process image data and output a visual semantic scene graph containing objects, attributes, and relationships between objects. Locality-sensitive hashing algorithm is used to perform semantic deduplication and clustering on the processed data, and data describing the same semantic unit are divided into the same hash bucket to obtain deduplicated multi-source data. The tensor decomposition method is used to perform semantic fusion on the deduplicated multi-source data, mapping text, sensor and image multimodal data to a unified semantic space to generate a multimodal data stream.

[0008] Furthermore, the step of processing sensor data using signal processing algorithms to transform the sensor data stream into an event-based feature sequence reflecting device status or user interaction includes: Wavelet transform is applied to the sensor data for filtering and noise reduction to extract the effective signal components; The isolated forest algorithm is used to detect outliers in the denoised signal and identify data points that deviate from the normal pattern. Based on preset domain knowledge rules or patterns learned through unsupervised algorithms, continuous signal anomaly segments or combinations of specific signal patterns are defined as atomic events. Each atomic event represents a complete product state or user behavior unit. The continuous signal anomaly segments are composed of continuous abnormal data points identified by the isolated forest algorithm. The detected atomic events are sorted by timestamp to obtain an event-based feature sequence.

[0009] Furthermore, the step of processing image data using a scene graph generation model to output a visual semantic scene graph containing objects, attributes, and relationships between objects includes: Use the object detection subnetwork to identify all entity objects in the image and their bounding boxes; The visual attributes of each identified object are predicted through an attribute classification sub-network; wherein, the attribute classification sub-network is a multi-label classification model built on a convolutional neural network; the visual attributes include color, shape, material, and visual state; Using a relation prediction subnetwork, the predicate relationship between each pair of objects is determined based on their spatial location, relative size, and visual context; wherein, the relation prediction subnetwork is a relation classification model built based on a neural network algorithm; The recognition results are organized into a directed graph structure with entity objects as nodes, visual attributes as node labels, and predicate relations as edges to obtain a visual semantic scene graph. The scene graph generation model is an end-to-end or cascaded deep learning model consisting of three sub-networks: object detection, attribute classification, and relationship prediction.

[0010] Furthermore, the locality-sensitive hashing algorithm is used to perform semantic deduplication and clustering on the processed data, dividing data describing the same semantic unit into the same hash bucket to obtain deduplicated multi-source data, including: Structured data from different modalities are mapped into unified semantic vectors through pre-trained encoders; wherein the structured data from different modalities includes structured text representations, event-based feature sequences, and visual semantic scene graphs. A set of hyperplanes is randomly generated for the function family of locality-sensitive hashing algorithms to partition the vector space; For each data semantic vector, a binary hash signature is generated based on the relative position of the semantic vector with each hyperplane; Data points with hash signatures that have the same signature or whose Hamming distance is less than a preset distance threshold are assigned to a hash bucket; Data within the same hash bucket is aggregated, retaining the sample with the latest timestamp or generating an aggregated feature vector to obtain deduplicated multi-source data.

[0011] Furthermore, the process of constructing the dynamic demand map includes: Demand entities and relationships are extracted based on multimodal data streams. Demand entities include products, product functions, usage scenarios, user attributes and user reviews extracted from text, performance indicators mapped from sensor events, and visual objects identified from images. Relationships include causal relationships, temporal connections or semantic connections between entities. A demand graph is constructed using a graph neural network as the backbone network; where nodes in the demand graph represent demand entities, and edges in the demand graph represent the association strength between entities. The association strength is quantified by the frequency of occurrence of the association between entities within a time window and the semantic association strength calculated by the attention mechanism. The multimodal data stream is captured using a sliding window approach. A graph neural network is used to encode the data and historical graph state within each time window to generate node embeddings with temporal memory, thereby obtaining a dynamic demand graph. The historical graph state represents the embedding vector sequence of nodes over multiple consecutive time steps in the past.

[0012] Furthermore, the extraction of demand entities based on multimodal data streams includes: Based on the structured text representation, the target entity and user reviews are extracted as demand entities; Based on the event-based feature sequence, the types of atomic events are mapped to demand entities that characterize performance metrics; Based on the visual semantic scene diagram, entity objects are extracted as requirement entities.

[0013] Furthermore, the associations include causal relationships, temporal connections, and semantic associations obtained through co-occurrence analysis, semantic model prediction, and cross-modal temporal rule mining, specifically: The co-occurrence analysis refers to statistically analyzing the frequency of different entities appearing together within the same data unit. For entity pairs whose frequency exceeds a preset threshold, an undirected co-occurrence edge is established to form a preliminary co-occurrence association. The semantic model prediction refers to inputting entity pairs and their contextual features into a fine-tuned relation classification model to predict the specific semantic relationship type between entity pairs and obtain the semantic association between entity pairs; the contextual features represent the information of the entity's adjacent nodes in the dynamic demand graph. The cross-modal temporal rule mining refers to aligning the sequence of timestamped event units extracted from multimodal data on a unified time axis, and using sequence pattern mining algorithms or causal discovery algorithms to identify the order of events and obtain the temporal or causal relationships between events.

[0014] Furthermore, the association strength is quantified by the frequency of occurrence of the association relationship between entities within a time window and the semantic association strength calculated by the attention mechanism, including: For any association within a given time window, calculate the normalized value of the frequency of the association to obtain the frequency intensity; By using the attention mechanism in the graph attention network, the attention coefficient between the head entity node and the tail entity node is calculated to obtain the semantic association strength between the two entities; wherein, the head entity represents the starting node in the relation triple and the tail entity represents the ending node in the relation triple. The association strength is obtained by weighted summation of the frequency intensity and the semantic association strength.

[0015] Furthermore, the encoding of data and historical graph states within each time window using a graph neural network includes: For the graph at the current time t, the attention weights of each target node and its first-order neighbor nodes are calculated using the multi-head graph attention mechanism in the graph neural network, and the features of the neighbor nodes are aggregated in a weighted manner to obtain the spatial aggregated features of each target node at time t; wherein, the graph at the current time t is composed of entities and relations extracted within the current time window; The spatial aggregation features of the target node at time t are combined with the final node embedding vectors of the target node over the past k consecutive time steps to form a feature sequence in chronological order. The feature sequence is input into a gated loop unit, which outputs a new embedding vector of the target node at time t that includes the historical graph state. Once all nodes have been updated, a dynamic demand graph encoding temporal dependencies is obtained.

[0016] Furthermore, the process of identifying user needs and user motivations based on the dynamic demand graph, using graph reasoning and causal discovery algorithms, includes: Hierarchical clustering or community detection algorithms are used to cluster demand entity nodes into multiple demand clusters based on the embedding vectors of nodes in the dynamic demand graph, and a generalized label is generated for each demand cluster through a large language model. On a graph subgraph containing entities including user attributes, usage scenarios, user reviews, and product functions, a causal discovery algorithm is used to search for conditional independence relationships between variables, and a causal directed acyclic graph structure representing the causal relationships between entities is output. Based on the causal directed acyclic graph structure, starting from the root nodes of the two types of entities, "user attributes" and "usage scenarios", the causal paths pointing to "user reviews" and "product functions" are traversed. By combining the generalized tags of the aforementioned demand clusters, the causal path is summarized into a description of user motivation using a large language model.

[0017] Furthermore, updating the dynamic demand map using an incremental learning algorithm includes: Using statistical testing methods, the differences between newly collected data and historical data in feature distribution or model prediction results are compared. When the test statistic exceeds the preset threshold, it is determined to be a case of inverted concept drift, and the newly collected data is stored in the buffer pool. When the buffer pool data reaches a preset size, an incremental update process is initiated; the incremental update process is as follows: New entities and relationships are extracted based on the newly collected data and added to the existing dynamic demand graph as new nodes or edges; Using an online graph representation learning algorithm, the embedding vectors in the dynamic demand graph are iteratively updated based on new entities and relationships; The edge association strength is recalculated within the sliding time window to update the edge weights.

[0018] Compared with the prior art, the beneficial effects of this application are: This application utilizes a multimodal data processing and fusion framework to transform heterogeneous data such as text, sensor signals, and images into a data stream with unified semantic representation, laying a high-quality data foundation for in-depth demand analysis. Then, by constructing a dynamic demand graph with a graph neural network as its core, it networks demand entities and their relationships, continuously updating them using time-series analysis. This solves the technical problems of how to characterize complex relationships between demands, quantify the strength of these relationships, and track their dynamic changes over time. Specifically, by combining co-occurrence analysis, semantic prediction, and cross-modal time-series rule mining, it not only establishes static relationships between entities but also captures their dynamic causal and temporal connections. The embedding of graph attention networks and time-series analysis ensures that the calculation of relationship strength simultaneously considers statistical significance and semantic importance, and can continuously track changes in relationships through a sliding window mechanism. This dynamic, weighted graph structure can more profoundly and accurately reflect the true picture of user needs and their evolutionary trends. Ultimately, this application achieves a leap from massive data to actionable insights through the synergistic application of graph reasoning, causal discovery, and large language models. Hierarchical clustering and community discovery algorithms summarize scattered needs into representative clusters, while causal discovery algorithms deeply analyze the driving relationships between user attributes, usage scenarios, product functions, and user reviews to uncover user motivations. The incremental learning mechanism ensures that the entire analysis system can continuously adapt to market changes and possess self-evolution capabilities, thereby providing enterprises with structured requirements analysis reports that are forward-looking, accurate, and interpretable. Attached Figure Description

[0019] 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.

[0020] Figure 1 A flowchart illustrating a user demand analysis method for an AI engine with multi-source data input, provided in an embodiment of this application; Figure 2 A flowchart illustrating another method for analyzing user needs of an AI engine with multi-source data input, provided in an embodiment of this application; Figure 3 A flowchart illustrating another method for analyzing user needs of an AI engine with multi-source data input, provided in an embodiment of this application; Figure 4 A flowchart illustrating another method for analyzing user needs of an AI engine with multi-source data input, provided in an embodiment of this application; Figure 5 This is a flowchart illustrating another method for analyzing user needs of an AI engine with multi-source data input, provided in an embodiment of this application. Detailed Implementation

[0021] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0022] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] To address the technical problems in existing technologies, such as insufficient insight into the correlation and dynamism of user needs, difficulty in uncovering user motivations, inadequate fusion of multi-source heterogeneous data, delayed analysis results, and lack of continuous tracking of demand evolution, this application provides a user demand analysis method for an AI engine with multi-source data input. This method includes: Multimodal data from different data sources is acquired, preprocessed, and semantically fused to generate a unified multimodal data stream; the data sources include public network data sources and closed IoT data sources. Based on multimodal data flow, a dynamic demand graph is constructed using graph neural networks, with demand entities as nodes and relationships as edges. Based on dynamic demand graphs, graph reasoning and causal discovery algorithms are used to identify user needs and user motivations, and generate structured demand analysis reports. Based on real-time monitoring data, the dynamic demand map is updated through incremental learning algorithms to continuously track the evolution of demand.

[0024] Based on this, this application can characterize user needs from multiple dimensions, accurately capture the correlation and dynamic changes of needs, and deeply explore the core motivations of users, while ensuring the timeliness and forward-looking nature of the analysis results, providing comprehensive and reliable support for enterprise decision-making.

[0025] like Figure 1 As shown in the embodiment of this application, a user demand analysis method for an AI engine with multi-source data input is provided, including: S1. Acquire multimodal data from different data sources, perform preprocessing and semantic fusion, and generate a multimodal data stream with a unified representation.

[0026] Multimodal data refers to data types encompassing various information carriers, primarily including text data, sensor data, and image data, and may also include audio data. Text data is commonly found in e-commerce platform user reviews, social media discussions, and product inquiry records, directly conveying users' attitudes towards products, their experiences, and their specific needs, serving as a fundamental source of information for demand analysis. Sensor data originates from closed IoT data sources, such as customer flow data, POS transaction data, and shelf interaction data in offline retail scenarios, and equipment operating status data in industrial scenarios, objectively recording user behavior trajectories, product usage scenarios, and equipment interaction. Image data includes real-life product photos, usage scenario images, and shelf display images, intuitively presenting the product's appearance, usage environment, and user interaction with the product, providing visualized demand clues. Audio data, such as user voice evaluations and customer service call recordings, can capture users' emotions and needs expressed beyond text.

[0027] In some implementations, common preprocessing methods include: text data is refined using techniques such as word segmentation, stop word removal, spell correction, and part-of-speech tagging; sensor data quality is ensured through filtering and denoising, data standardization, missing value imputation, and outlier removal; image data is enhanced by size normalization, pixel standardization, and image enhancement to improve feature recognition; and audio data is transformed into analyzable information through speech-to-text conversion, noise reduction, and feature extraction. Common semantic fusion methods include feature concatenation fusion, weighted summation fusion, deep neural network fusion, and matrix factorization fusion. Among these, deep neural network fusion automatically learns the deep relationships between different modalities by constructing a multi-layered network structure, while matrix factorization fusion maps multimodal data to the same semantic space through mathematical transformations.

[0028] It should be noted that the multimodal data collected and processed in this application strictly comply with the provisions on personal privacy protection in relevant laws and regulations such as the "Personal Information Protection Law of the People's Republic of China" and the "Cybersecurity Law of the People's Republic of China." All data has undergone de-identification and anonymization preprocessing and does not contain any information that can directly or indirectly identify a natural person, such as name, mobile phone number, ID card number, home address, bank account information, or other private data. The data collection process strictly follows the principle of minimum necessity, collecting only information directly related to user needs analysis, such as publicly available product reviews, de-identified customer flow statistics, and scene images without personal portraits, without collecting any additional irrelevant private content. Furthermore, the data processed in this application is only used for technical research and business optimization purposes such as user needs insight, product optimization, and service improvement, and will never be used for any unauthorized commercial promotion, targeted marketing, data selling, or other acts that infringe on personal rights. In addition, the data storage and processing process has adopted security protection measures such as encrypted transmission and access control to further prevent the risk of data leakage and ensure that the entire data processing process is compliant and secure, without infringing on the privacy rights and other legitimate rights of any natural person.

[0029] S2. Based on multimodal data flow, a dynamic demand graph is constructed using graph neural networks, with demand entities as nodes and relationships as edges.

[0030] Among them, dynamic demand graph is a structured knowledge representation form that uses demand-related entities as nodes and various relationships between entities as edges. It can dynamically update the number of nodes, the type of edges, and the strength of relationships as time goes by and new data is added. It is used to intuitively present the complex relationships between demand entities and reflect the evolution of these relationships in real time.

[0031] In some implementations, common methods for constructing dynamic demand graphs include the following steps: The first step is to extract demand entities from the unified representation of the multimodal data stream, including but not limited to products, product functions, usage scenarios, user attributes, user reviews, performance indicators, visual objects, device status, and interaction events. The second step is to explore the relationships between entities, including but not limited to causal relationships, semantic relationships, temporal relationships, co-occurrence relationships, attribute relationships, demand-satisfaction relationships, defect-feedback relationships, etc. The third step is to construct an initial graph using graph neural networks. Commonly used graph neural networks include graph convolutional networks and graph sampling generation networks such as GraphSAGE. These networks encode entities and relationships to generate node embeddings with semantic information. The fourth step is to introduce a time dimension to make the graph dynamic. By using a sliding time window and a periodic update mechanism, new entities are added, new or adjusted relationships are added, and the relationship strength of edges is updated based on new data flows, forming a dynamically evolving demand graph.

[0032] S3. Based on dynamic demand graphs, using graph reasoning and causal discovery algorithms, identify user needs and user motivations, and generate structured demand analysis reports.

[0033] Graph reasoning, based on the structured information of nodes and edges in a dynamic demand graph, uses logical deduction, path search, and feature propagation to uncover potential relationships or hidden information between nodes, extracting implicit demand connections from the explicit graph structure. Causal discovery algorithms are a class of algorithms capable of analyzing dependencies between variables in data, distinguishing between correlation and causal relationships, and identifying causal transmission paths between variables.

[0034] In some implementations, graph reasoning algorithms, including pathfinding algorithms, belief propagation algorithms, and community detection algorithms, are used to mine potential relationships between entities in a dynamic demand graph and identify the hierarchical structure and dependencies of demands. Commonly used causal discovery algorithms include PC algorithms and Bayesian network algorithms, which analyze the causal relationships between entities to clarify the deep motivations behind user demands. Then, the identified demands and motivations are classified and organized according to dimensions such as demand category, motivation level, related product characteristics, and optimization direction. Finally, natural language generation technology is used to transform the organized demands, motivations, and analysis conclusions into a clear and detailed structured analysis report, including a demand overview, core motivation analysis, and action recommendations.

[0035] For example, a path search algorithm is used to discover potential related paths in a dynamic demand graph, namely "young users - campus scenarios - lightweight design - high cost-effectiveness". The entities under this path are clustered into "core demand clusters of campus users" using a community discovery algorithm. The demand cluster is analyzed using a Bayesian network algorithm to determine that the usage environment limitations of "campus scenarios" and the purchasing power of "young users" are the core motivations driving the demand for "lightweight design" and "high cost-effectiveness". The analysis results are organized according to the structure of "demand category - core motivation - product optimization suggestions" to generate a structured report, which clearly puts forward specific suggestions for optimizing product portability and controlling cost pricing for the campus user group.

[0036] S4. Based on real-time monitoring data, the dynamic demand map is updated through incremental learning algorithms to continuously track the evolution of demand.

[0037] Incremental learning algorithms are a type of algorithm that can use newly collected real-time data to gradually update the parameters and structure of a model or system without retraining the entire model. Its core advantage lies in efficiently integrating new information while retaining effective knowledge from historical data, ensuring that the system can continuously optimize as data accumulates and adapt to dynamic changes in requirements.

[0038] In some implementations, common methods for continuous tracking of requirement evolution include: The first step is to continuously collect multimodal real-time monitoring data, including user feedback, transaction records, and product usage data, through methods such as web crawling, real-time uploading from IoT sensors, and synchronization with platform data interfaces. The second step involves using statistical tests and distribution comparisons to analyze the differences between the newly collected data and historical data in terms of characteristic distribution, frequency of occurrence of demand entities, and strength of correlation, in order to determine whether there are signs of demand changes. The third step involves using incremental learning algorithms such as incremental graph neural networks and online learning algorithms to integrate new entities and new relationships extracted from new data into the existing graph, update the node embedding vectors and the relationship strength of edges, and delete outdated entities and invalid relationships. The fourth step is to set demand evolution indicators, such as the rate of change in the popularity of core demand entities, the proportion of new / disappearing relationships, and the degree of change in demand cluster structure. When the indicators exceed the preset threshold, it is determined that the demand has evolved significantly. The evolution trajectory is recorded, including the types of new demands, changes in demand intensity, and changes in motivation, and is fed back to the demand analysis report in real time.

[0039] It should be noted that during incremental learning, the weight of new data and historical data must be balanced to avoid the excessive dominance of new data leading to the loss of useful historical information.

[0040] For example, an incremental graph sampling generation network is used as the incremental learning algorithm. User reviews, purchase records, and shelf interaction data are collected in real time through e-commerce platform interfaces and offline store sensors. The feature distribution of new data is compared with historical data every day. When the frequency of occurrence of entities related to "foldable screen function" increases by 50% compared with the historical average, the incremental update process is initiated to add "foldable screen function" as a new entity to the graph and update its association with "high-end users" and "business scenarios". A significant evolution threshold is set when the degree of change in the demand cluster structure exceeds 20%. When the "high-end user demand cluster" undergoes a 30% structural change due to the addition of "foldable screen function", it is determined to be a significant evolution of demand. The evolution trajectory is tracked and recorded, and the structured analysis report is updated.

[0041] Based on the above technical solutions, this application provides an AI engine user demand analysis method with multi-source data input. By integrating multi-source, multi-modal data from public networks and closed IoT, it achieves a comprehensive and multi-perspective characterization of user needs. It utilizes graph neural networks to construct a dynamic demand graph, transforming demand entities and relationships into a structured knowledge network that intuitively presents the complex relationships between demands and their temporal evolution. Combining graph reasoning and causal discovery algorithms, it uncovers the core motivations behind demands, providing enterprises with more guiding decision-making support. Through incremental learning algorithms, it achieves continuous self-optimization of the system, ensuring that demand analysis keeps pace with market changes and shifts in user preferences, balancing timeliness and foresight.

[0042] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 2 As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Perform targeted preprocessing on the multi-source heterogeneous multimodal data to generate structured characterization data corresponding to each modality.

[0043] In some implementations, the preprocessing operations for different types of data are as follows: 1. Text Data Preprocessing: Utilizing natural language processing algorithms and named entity recognition technology, a series of operations including word segmentation, part-of-speech tagging, stop word filtering, and named entity extraction are employed to construct a five-tuple structured text representation encompassing product, product features, usage scenarios, user attributes, and user reviews. The specific process is as follows: First, the text data is segmented using a word segmentation algorithm that combines dictionary and statistics, such as the enhanced version of jieba word segmentation, to break the text into semantic units. Subsequently, the part-of-speech tagging model (such as Hidden Markov Model HMM) is used to tag the part of speech of each semantic unit, and core part-of-speech units such as nouns, verbs, and adjectives are selected. Then, named entity recognition models, such as the joint model based on bidirectional long short-term memory network BiLSTM and conditional random field CRF, are used to identify entities such as product name, function description, scene keywords, user profile tags, and evaluation words. Finally, entities are combined according to a fixed structure of [product, product function, usage scenario, user attribute, user evaluation] to form a quintuple representation.

[0044] 2. Sensor Data Preprocessing: Signal processing algorithms are used to transform continuous sensor data streams into event-based feature sequences reflecting device status or user interactions. Specific steps include: First, wavelet transform is applied to the raw sensor data to filter and remove noise, retaining the effective signal components that reflect the true state. Then, the Isolation Forest algorithm is used to identify outliers in the denoised signal, identifying data points that deviate from the normal pattern and defining continuous outlier data points as signal outlier segments. The Isolation Forest algorithm constructs multiple isolated trees and calculates the path length h(x) for each data point; the shorter the path length, the more likely the data point is to be an outlier. The outlier determination formula is: Where s(x,n) is the outlier score of data point x, n is the number of samples, and c(n) is the average path length; when s(x,n)>θ (θ is a preset threshold, usually 0.8), the data point is determined to be an outlier. Next, based on patterns learned from pre-defined domain knowledge rules or unsupervised algorithms (such as K-means clustering), continuous signal anomalies or combinations of specific signal patterns are defined as atomic events. An atomic event must represent a complete product state or user behavior unit, such as "shelf dwell time exceeds 30 seconds," "POS transaction successful," or "equipment operating temperature exceeds standard." Finally, the detected atomic events are sorted in ascending order by timestamp to form an ordered sequence of event-based features, which fully records the temporal evolution of user interaction or device state.

[0045] 3. Image Data Preprocessing: Image data is processed using a scene graph generation model to output a visual semantic scene graph containing objects, attributes, and relationships between objects. The scene graph generation model is an end-to-end or cascaded deep learning model, consisting of three sub-networks: a. Object Detection Subnetwork: Employs a series of region-based convolutional neural network (R-CNN) models (such as Faster R-CNN). It extracts image features through a feature extraction backbone network to generate candidate regions, and then uses a classifier and bounding box regressor to identify all entity objects in the image and their corresponding bounding box coordinates. Entity objects include, but are not limited to, product bodies, shelves, display stands, shopping tools, de-identified users, environmental furnishings, price tags, equipment components, and usage scenario-related carriers. b. Attribute Classification Sub-network: A multi-label classification model is constructed based on a convolutional neural network. Taking the entity object region features output by the object detection sub-network as input, it identifies the visual attributes of each object, including color, shape, material, and visual state. This highlights the differentiated features and actual state of the entity object, providing detailed support for subsequent semantic association analysis. For example, it matches user aesthetic preferences based on product color and material, identifies usage-related needs based on product visual state, and mines suitability requirements based on scene environment attributes (such as lighting and space size). c. Relationship Prediction Subnetwork: Based on neural network algorithm, a relationship classification model is constructed. The input is the region features of two objects, the relative position features of the bounding boxes, and the visual context features. It predicts the predicate relationship between each pair of objects, such as "the user is holding the product" or "the product is placed on the shelf". Finally, the recognition results are organized into a directed graph structure with entity objects as nodes, visual attributes as node labels, and predicate relationships as edges, resulting in a visual semantic scene graph.

[0046] It should be noted that the purpose of collecting sensor data is to objectively capture the behavioral trajectories of users in real-world scenarios, the product usage status, and quantitative information during interactions. For example, using heatmap data of customer flow in offline retail scenarios to identify popular product areas, using POS transaction data to link user purchasing behavior with product types, and using shelf interaction sensor data to record the duration of user attention and actions for specific products, we can provide objective and quantifiable behavioral evidence for demand entity extraction and relationship mining, making up for the subjectivity of pure text data and the non-quantitative defects of image data, and achieving complementary verification of multi-dimensional data. The role of visual semantic scene graphs is to transform the visual information in images into structured semantic knowledge. They can intuitively reflect the distribution pattern of entities in the actual scene, the specific characteristics of each object (such as product appearance and scene environmental conditions), and the interaction logic between objects, such as the user-product usage relationship and the product-scene adaptation relationship. At the same time, they can realize cross-modal semantic alignment of image data, text data, and sensor data, providing intuitive and resolvable support for comprehensively depicting user demand scenarios and mining contextual implicit needs.

[0047] S102. Perform semantic deduplication and clustering on the preprocessed structured representation data of each modality to eliminate redundant information and aggregate data of the same semantic unit.

[0048] Among them, structured data of different modalities may contain redundant information describing the same semantic unit. For example, multiple user comments may express the same need for "clear photos". It is necessary to filter core data through semantic deduplication and clustering operations to reduce the amount of subsequent fusion calculations.

[0049] In some implementations, locality-sensitive hashing (LSH) algorithms are used to perform semantic deduplication and clustering, including: The first step is semantic vector mapping: structured data of different modalities are mapped into semantic vectors of a unified dimension through pre-trained encoders. For structured text representations, a pre-trained language model (such as BERT) can be used as an encoder, taking a five-tuple text as input and outputting the vector of the last layer [CLS] token as the semantic vector; for event-based feature sequences, a gated recurrent unit (GRU) can be used as an encoder to transform ordered atomic events into semantic vectors of a fixed dimension; for visual semantic scene graphs, a graph neural network can be used as an encoder to aggregate node features and edge features, outputting a global semantic vector of the graph.

[0050] The second step is hyperplane generation: A set of hyperplanes is randomly generated for the family of functions in the Locality Sensitive Hashing (LSH) algorithm to partition the vector space. Let the semantic vector dimension be d, and generate k hyperplanes. k is typically 16-64, adjusted according to the semantic discriminative requirements. Each hyperplane is defined by a normal vector. Define, where , Let represent the d-dimensional Euclidean space, which is the set of all ordered vectors containing d real components.

[0051] The third step is hash signature generation: For each data semantic vector v, calculate its relationship with each hyperplane normal vector. The inner product is used to generate binary hash bits based on the inner product sign. Concatenating k hash bits yields a binary hash signature h(v) of length k, as shown in the formula: ; ;in, For the i-th bit hash value, Let the semantic vector v and the hyperplane normal vector be... The inner product; Step 4, Hash Bucket Allocation: Calculate two hash signatures. and Hamming distance between When the Hamming distance is below a preset threshold When two pieces of data are determined to be semantically similar, they are assigned to the same hash bucket; where, Usually, k / 4 is taken, such as when k=32. =8.

[0052] Step 5, Data Aggregation: Aggregate data within the same hash bucket using one of two methods: If the data is time-series related (e.g., user comments from different times), retain the sample with the latest timestamp; if the data is non-time-series, calculate the mean of all semantic vectors within the bucket as the aggregation feature vector. Aggregation results in deduplicated multi-source data, eliminating semantic redundancy.

[0053] S103. Perform semantic fusion on the deduplicated multi-source data, map different modal data to a unified semantic space, and generate a multimodal data stream with a unified representation.

[0054] Even after deduplication, the multi-source data still belong to different modalities and have different semantic spaces. Semantic fusion is needed to break down modal barriers and achieve information complementarity and enhancement.

[0055] In some implementations, tensor decomposition is used for semantic fusion, which includes the following steps: First, the semantic vectors (or aggregated feature vectors) of the three modalities after deduplication are constructed into a third-order tensor. Where I is the number of samples in the text modal data, J is the number of samples in the sensor modal data, K is the number of samples in the image modal data, and tensor elements Let represent the semantic association strength between the i-th text sample, the j-th sensor sample, and the k-th image sample. The semantic association strength is obtained by calculating the cosine similarity product of the semantic vectors of the three samples, as shown in the formula: .

[0056] Then, the PARAFAC tensor decomposition method (parallel factor analysis) is used to decompose the third-order tensor X into the product of three factor matrices and a core tensor, as shown in the formula: Where R is the decomposition rank, representing the dimension of the unified semantic space, which is usually 128 to 512 and needs to be determined through cross-validation; This is the factor matrix of the text modalities, where each row corresponds to a vector of a text sample in a unified semantic space; This is the factor matrix of sensor modes, where each row corresponds to a vector of a sensor sample in a unified semantic space; This is a factor matrix of image modalities, where each row corresponds to a vector of an image sample in a unified semantic space; To decompose the error term, it must satisfy the following condition: , To preset the error threshold, it is usually taken as... ; , , These are the elements of the three factor matrices.

[0057] By using tensor decomposition, the high-dimensional features of the three modalities are mapped to a low-dimensional unified semantic space, thereby achieving cross-modal semantic alignment.

[0058] Next, the three factor matrices A, B, and C are concatenated according to the semantic relationships of the samples. For text, sensor, and image samples with strong semantic relationships (cosine similarity ≥ 0.7), their corresponding unified semantic space vectors are concatenated into a fusion vector; for single-modal samples without strong relationships, their unified semantic space vectors are directly retained. This ultimately forms a multimodal data stream with unified dimensions and consistent semantics, containing the core information of all modalities and eliminating modal heterogeneity.

[0059] Based on the above technical solution, firstly, differentiated preprocessing methods are designed for the characteristics of text, sensor, and image data to ensure the effective extraction of core features; secondly, semantic redundancy is eliminated through locality-sensitive hashing (LSH) to improve data purity; finally, tensor decomposition is used to achieve cross-modal semantic alignment, generating a unified representation of multimodal data streams, breaking down semantic barriers between modalities, achieving information complementarity and enhancement, and enabling subsequent analysis to be based on a more complete and consistent information view. Step S1 ensures the targeted and accurate nature of data processing and achieves efficient fusion of multi-source heterogeneous data, providing high-quality and high-completeness data source support for the subsequent construction of dynamic demand maps, effectively solving the technical problems of insufficient multimodal data fusion and information redundancy interfering with analysis in traditional methods.

[0060] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 3 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Extract the required entities from the multimodal data stream and mine the relationships between the entities.

[0061] Among them, demand entities refer to the core units that can represent information related to user needs and are the basic nodes for building demand graphs; relationships refer to the logical connections between demand entities and are the edges that connect the nodes.

[0062] In some implementations, the specific operations for extracting demand entities and mining relationships include: 1. Extraction of Requirement Entities: a. Extraction based on structured text representation: Directly extract five types of entities, namely "product", "product function", "usage scenario", "user attribute" and "user evaluation", from the five-tuple structured text representation generated by S1; b. Event-based feature sequence extraction: Map the atomic event types generated by S1 to demand entities representing performance indicators. First, establish a mapping rule base between atomic events and performance indicators, such as mapping "shelf stay exceeds 30 seconds" to "product attractiveness", "POS transaction success rate" to "purchase intention", and "equipment operating temperature exceeds the standard" to "heat dissipation performance defect". Then, traverse the event-based feature sequence and convert each atomic event into a corresponding performance indicator entity according to the mapping rules. c. Visual semantic scene graph extraction: Extract all entity objects as requirement entities from the visual semantic scene graph generated by S1. For example, extract the entity {user, laptop} from the scene graph of “user-handheld-laptop”.

[0063] 2. Relationship mining: a. Co-occurrence analysis to uncover co-occurrence associations: This involves statistically analyzing the frequency of co-occurrence of different entities within the same data unit (e.g., the same user comment, sensor data within the same time window, or the same image scene) to obtain the co-occurrence frequency. The formula is: ;in, For entities and co-occurrence frequency, The number of times the two appear together. For entities Total number of times it appears alone For entities Total number of times it appears alone. At that time, an undirected co-occurrence relation edge is established between the two entities; The preset frequency threshold is usually set to 0.3 to 0.5.

[0064] b. Semantic model predicts semantic associations: entities are paired The relationship classification model, fine-tuned by taking the contextual features as input, predicts the specific semantic relationship type between the two, such as "demand-satisfaction", "attribute-association", "defect-complaint", etc., as shown in the formula: ;in, For entity pairs The probability of belonging to relation r in context C. For feature splicing operations, and As classifier weights and biases, the BERT model outputs fused semantic features of entity pairs and their context; contextual features refer to information about adjacent entities in the multimodal data stream, such as entity... Nearby "use case" entities, Nearby "user reviews" entities, etc. When the probability is highest... ( When the probability threshold (usually 0.6) is used, the semantic relationship is determined to be an association relationship between entity pairs.

[0065] c. Cross-modal temporal rule mining for temporal / causal relationships: First, align all timestamped event unit sequences on a unified timeline, including text comment timestamps, sensor event timestamps, image acquisition timestamps, etc. Then, use sequence pattern mining algorithms, such as the PrefixSpan algorithm, to identify the chronological order of events and obtain temporal relationships, or use causal discovery algorithms, such as the PC algorithm, to analyze the conditional independence relationships between events and obtain causal relationships. For example, the PrefixSpan algorithm can be used to discover the temporal relationship of "user views product details page → adds to cart → submits order," while the PC algorithm can be used to discover the causal relationship of "small battery capacity → short battery life → negative user reviews."

[0066] For example, in an offline smart home experience scenario, entities {air conditioner, cooling function, living room scene, family of three, loud noise} are extracted from the text quintuple. Entities {cooling efficiency, noise level, user comfort} are mapped from the sensor event sequence "living room temperature 32℃ → air conditioner starts cooling → temperature drops to 26℃ → user presses mute button". Entities {air conditioner, living room, user} are extracted from the image scene graph. Co-occurrence analysis is used to calculate the co-occurrence frequency F=0.42 between "living room scene" and "cooling function", establishing a co-occurrence association. Entity pairs (noise level, user negative review) and contextual features are input into a relation classification model to predict the semantic relationship "defect-complaint". The causal association "high noise level → user presses mute button → user negative review" is mined using a PC algorithm, ultimately yielding material containing 10 entities and 8 association relationships.

[0067] S202. Calculate the correlation strength and construct the initial demand map with graph neural network as the core.

[0068] Among them, the association strength is an indicator that quantifies the closeness of the relationship between entities, and assigns weights to the edges of the demand graph so that the graph can reflect the differences in the importance of the relationship; the initial demand graph is a static graph built based on the entities and relationships within the current time window, and is the basic framework of the dynamic demand graph.

[0069] In some implementations, the specific operations for calculating correlation strength and constructing the initial demand map are as follows: First, based on frequency intensity With semantic association strength The correlation strength is obtained by weighted summation. The formula is: ;in, This is the weighting coefficient, with a value range of [0,1], and is usually taken as 0.5; The frequency intensity is determined by: the co-occurrence frequency obtained in S201. After normalization, the frequency intensity is obtained, as shown in the formula: , The maximum co-occurrence frequency of all entity pairs. The minimum co-occurrence frequency of all entity pairs. The semantic association strength is determined by calculating the head entity through the attention mechanism of a graph attention network (GAT). With tail entity The attention coefficient between the two elements is used as a measure of the semantic association strength. The formula for the attention coefficient is: , For entities For entities The attention coefficient, i.e., the semantic association strength. , Let be the learnable parameter vector for the attention mechanism, and W be the linear transformation weight matrix. , , Entities , , The original feature vector, For entities The set of neighboring entities, For activation function, This is a feature splicing operation.

[0070] Next, we construct the initial requirements graph: Using the demand entities extracted from S201 as nodes, the mined relationships as edges, and the calculated relationship strength... Assign edge weights to construct the initial framework of the undirected weighted graph; Then, a Graph Convolutional Network (GCN) is selected as the backbone network to semantically encode the initial framework and generate initial entity embedding vectors. Through the neighborhood feature weighted aggregation and inter-layer feature propagation mechanism of the GCN, combined with the difference in association strength quantified by the edge weights, differential information fusion is performed. The entity embedding vector can aggregate the features and association strength information of itself and its neighboring entities, enabling the nodes to have semantic representation capabilities, and finally forming the initial demand graph.

[0071] S203. Using a sliding window and timing coding mechanism, the initial demand map is updated to a dynamic demand map.

[0072] The sliding window refers to capturing continuous segments of a multimodal data stream at fixed time intervals to capture changes in demand information at different time stages. The time-series coding mechanism, by fusing historical graph states with current window data, gives the graph the ability to remember time, reflecting the evolution of demand. In some implementations, the specific operations for setting up the sliding window and generating the dynamic demand graph include: First, set the length T and step size s of the sliding window. The length T is adjusted according to the required rate of change, and can be set to 7 days, 30 days, etc. The step size s is usually s = T / 2 or s = T; a smaller step size means more frequent updates. Extract the multimodal data stream in chronological order, with each window corresponding to a dataset of a time segment, denoted as . t is the window number, t=1,2,...,n.

[0073] Then, time-series coding and graph updating are performed: For the graph corresponding to the current window t (by window t) The extracted entities, relationships, and relationship strengths constitute the spatial aggregation features of the node at time t. Through the multi-head attention mechanism of the graph attention network, the attention weights of each target node and its first-order neighbor nodes are calculated, and the features of the neighbor nodes are aggregated in a weighted manner to obtain the spatial aggregation features of the node at time t. Node Spatial aggregation characteristics at time t The final node embedding vectors of the past k consecutive time steps (k is the number of memory steps, usually 3 to 5) , ... Concatenate them in chronological order to form a temporal feature sequence. ; Then the time series feature sequence The input gated recurrent unit (GRU) learns temporal dependencies, adaptively fuses historical node embeddings with current spatial aggregation features, and outputs nodes. The new embedding vector at time t This allows the new embedding vector to contain temporal memory information, reflecting the temporal evolution of node features.

[0074] Once the embedding vectors of all nodes have been updated, the association strength update results within the current window are combined to obtain a dynamic demand graph that encodes temporal dependencies. As the sliding window moves, the above steps are repeated to achieve the continuous evolution of the dynamic demand graph.

[0075] Based on the above technical solution, step S2 extracts the demand entities and relationships in S201, comprehensively mining the core demand elements in multimodal data and constructing the basic nodes and edge structure of the dynamic demand graph. Through the association strength quantification calculation and initial graph construction in S202, edge weights are determined by integrating frequency statistics and semantic association information, resulting in an initial demand graph with reasonable weight configuration. Through the sliding window mechanism and gated recurrent unit (GRU) temporal encoding in S203, the initial graph is injected with temporal memory and evolution capabilities, achieving an upgrade from a static graph to a dynamic graph. This solves the technical problem of traditional demand graphs being statically lagging and unable to capture temporal changes in demand. The constructed dynamic demand graph not only improves the semantic representation level with the strong encoding capability of graph neural networks but also accurately depicts the complex relationships and evolutionary patterns between demand entities, providing structured and dynamic knowledge support for subsequent deep identification of user needs and motivations.

[0076] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Cluster the demand entities in the dynamic demand graph to form demand clusters, and generate demand cluster summary labels through a large language model.

[0077] The dynamic requirement graph contains numerous scattered requirement entities, such as product functions, usage scenarios, and user reviews extracted in S201. Directly analyzing individual entities makes it difficult to extract the core requirement direction. The core objective of clustering is to aggregate semantically related entities that point to the same core requirement into requirement clusters, achieving structured classification of requirements. The summary label, on the other hand, provides a refined description of the core demands of each requirement cluster, offering clear anchor points for subsequent motivational analysis. This step must be carried out based on the embedding vectors of nodes in the dynamic requirement graph generated in S2 to ensure the semantic consistency of the clustering results.

[0078] In some implementations, the specific operations of S301 include: First, clustering algorithms are used to group and aggregate demand entities in the dynamic demand graph. These algorithms include hierarchical clustering and the Louvain community detection algorithm. Hierarchical clustering employs a bottom-up aggregation strategy, initially treating each demand entity as a separate cluster, then iteratively merging the clusters with the highest semantic similarity until a stopping condition is met; this approach is suitable for small-scale demand graph data. The Louvain algorithm, on the other hand, automatically divides demand clusters by maximizing the modularity index, enabling efficient processing of large-scale demand graph data and adapting to the need for rapid clustering of massive amounts of entities.

[0079] Next, semantic similarity or dissimilarity between entities is calculated to provide a basis for clustering. This is based on the node embedding vectors generated by S2. (i is the entity number). The cosine similarity formula can be used to calculate the degree of semantic association between entities. The closer the value is to 1, the more similar the entities are semantically. Alternatively, the Euclidean distance formula can be used to quantify the differences between entities. The smaller the distance, the more similar the entities are.

[0080] Then, the clustering stopping condition is set to complete entity clustering and generate demand cluster summary labels. Hierarchical clustering can terminate iteration by setting the maximum number of clusters or the minimum similarity threshold; the Louvain algorithm stops at maximizing modularity, and the formula for calculating modularity Q is: The value range is [-1, 1]. Generally, a clustering effect is considered good when the modularity is ≥ 0.3; where m is the total number of edges in the dynamic demand graph. The edge weight between entities i and j (i.e., the association strength calculated by S202). Let i be the degree of entity i. For the cluster to which entity i belongs, For indicator functions, when The value is 1 if the condition is met, otherwise it is 0. After clustering, the entity names, relationship descriptions, and semantic information corresponding to the node embedding vectors in each requirement cluster are used as input to construct prompt words with clear label formats and length requirements. For example, the instruction statement "Based on the entity list {air conditioner, cooling function, living room scene, rapid cooling} and the relationship {living room scene-requirement-cooling function, cooling function-core demand-rapid cooling}, extract one core requirement label of no more than 10 characters" is input into a large language model (such as GPT-3.5, Tongyi Qianwen, etc.) to obtain concise and accurate requirement cluster summary labels.

[0081] For example, based on the dynamic demand map of smart home appliances generated by S2, the embedding vectors of 20 demand entities are extracted and clustered using the Louvain algorithm (modularity Q=0.42) to obtain 3 demand clusters: Cluster 1 contains entities {air conditioner, cooling function, living room scene, rapid cooling, temperature adjustment}, Cluster 2 contains entities {air conditioner, noise level, sleep scene, silent mode, negative user reviews}, and Cluster 3 contains entities {air conditioner, energy saving mode, family of three, electricity cost saving, positive user reviews}; and the tags generated by the large language model are "living room rapid cooling demand", "sleep silent demand", and "energy saving demand".

[0082] S302. Construct a causal directed acyclic graph to explore the causal path from user attributes and usage scenarios to user evaluations and product functions.

[0083] The causal directed acyclic graph (DAG) is a structured model that depicts the causal dependencies between entities. The nodes are the four core entities extracted in S201: "user attributes," "usage scenarios," "user reviews," and "product functions," while the edges represent the causal relationships between entities. The goal of this step is to reveal the deeper logic behind "why users have this need" by looking beyond the surface connections between entities.

[0084] In some implementations, the specific operation of S302 is as follows: First, select subgraphs from the dynamic demand graph that contain only four types of entities: "user attributes", "usage scenarios", "user reviews" and "product functions". Remove irrelevant entities and redundant edges to ensure that the subgraphs focus on the core elements of the demand-driven logic.

[0085] Then, causal discovery algorithms (such as PC algorithm and Bayesian network structure learning algorithm) are used to search for conditional independence relationships between variables. Taking the PC algorithm as an example: a. Initialization: Using four types of entities as nodes, with undirected edges between any two nodes, construct a complete undirected graph; b. Conditional Independence Test: Based on the association strength data of entities in the subgraph, for each pair of nodes (X,Y), under different condition sets Z (subsets of other entities), test whether conditional independence is satisfied. Mutual information is used as the indicator for the conditional independence test, and the formula is: Where H(X|Z) is the conditional entropy of X given Z, and H(X|Y,Z) is the conditional entropy of X given Y and Z; the formula for calculating entropy is: ; ; When mutual information ( When the preset threshold is 0.01, X and Y are determined to be conditionally independent given Z, and the undirected edge between the two nodes is deleted. c. Edge Orientation: Orient the remaining undirected edges using a V-structure (such as...). Identify the causal direction, avoid cyclic structures, and finally obtain a causal directed acyclic graph.

[0086] Finally, starting from the root node of the causal directed acyclic graph, traverse all directed paths pointing to the leaf nodes. Each path is a complete demand-driven logic chain. The root node consists of "user attributes" and "usage scenarios" entities, which serve only as causes, while the leaf nodes consist of "user evaluations" and "product functions" entities, which serve only as results.

[0087] S303. Combining requirement cluster tags and causal paths, summarize user motivations and generate a structured requirement analysis report.

[0088] Among these, user motivation is the core driving force of user needs. It needs to be extracted through the "cause-effect" logic revealed by the causal path, combined with the core demands of the demand cluster tags. The structured demand analysis report organizes the analysis results of demand clusters, causal paths, and user motivations into a fixed framework, providing intuitive and actionable references for enterprise decision-making. This step must be closely linked to the demand cluster tags in S301 and the causal path in S302 to ensure the logical consistency of motivation summarization and the practicality of the report.

[0089] In some implementations, summarizing user motivations includes the following steps: First, associate the paths with tags. Match each causal path with the corresponding demand cluster tag. For example, the causal path "family of three → living room scene → cooling function → positive user reviews" corresponds to the demand cluster tag "living room rapid cooling demand". Then, starting from the root node of the causal path, and combining the intermediate nodes (product features) and leaf nodes (user reviews) in the path, construct motivation description prompts and input them into the large language model to generate motivation. Example prompt: "Based on the causal path {family of three → living room scenario → cooling function → positive user reviews} and the demand cluster tag {rapid cooling demand in the living room}, extract the user's core motivation (explaining the user's identity, usage scenario, core demand, and the result after the demand is met), no more than 50 words." The motivations extracted from multiple causal paths corresponding to the same demand cluster are integrated, and duplicate expressions are removed to form the core motivation of the demand cluster. The importance of the motivation is evaluated based on the number of entities or the sum of the correlation strength within the demand cluster, and the motivations of different demand clusters are ranked by importance.

[0090] In some implementations, generating a structured requirements analysis report includes the following steps: First, design the report framework, which includes a requirements overview (number of requirement clusters, distribution of core requirements), requirements cluster details (entity composition, tags, and correlation strength statistics for each cluster), user motivation analysis (core motivations ranked by importance, and causal path support), and action recommendations (propose directions for product optimization and service improvement based on core motivations). Then, fill in the clustering results of S301, the causal paths of S302, and the motivational summarization results of S303 by module. Charts can be used to assist in the explanation and improve readability.

[0091] Finally, the report content was polished using a large language model to ensure concise language and clear logic.

[0092] For example, let's take insights into the user needs of new energy vehicles as an example: First, obtain text data, including but not limited to: collecting discussion posts from car forums, car owner groups, and social media such as "Winter range is drastically reduced, advertised as 500km, but can only run 300km in reality" and "Autonomous driving is inaccurate in recognizing lane lines in rainy weather"; Acquire sensor data, including but not limited to: anonymized actual driving range, battery charging and discharging efficiency, ambient temperature, autonomous driving system activation status, system intervention records, etc. collected from the vehicle terminal; Acquire image / video data, including but not limited to: user-uploaded dashboard photos, screenshots of the autonomous driving visualization interface, and dashcam video clips; Subsequently, data fusion was performed to associate textual complaints about "poor winter battery life" with sensor data showing "battery efficiency drops by 30% when ambient temperature is below 5℃" and visual information from the dashboard image showing "sudden drop in driving range," forming a unified multimodal data stream.

[0093] During the dynamic demand graph construction phase, entities such as winter range, nominal range, battery efficiency, ambient temperature, autonomous driving, lane line recognition, rainy weather, and user anxiety are extracted from the data stream. The "impact" relationship between ambient temperature and battery efficiency, the causal relationship that low ambient temperature leads to decreased battery efficiency and thus causes the actual range to be lower than the nominal range, and the "correlation" relationship between autonomous driving and lane line recognition performance in rainy weather scenarios are also explored. After constructing the initial graph with these entities and relationships, winter range is a key node connecting multiple causes (ambient temperature, battery efficiency) and results (user anxiety). The graph will be dynamically updated with the seasons—the weight of the winter range node is lower in summer, but as more related discussions and sensor data increase in autumn, the popularity of this node continues to rise, and the weight of the associated edges with nodes such as battery technology and thermal management system also strengthens. The system predicts that "range" will become the core concern in winter.

[0094] In the demand and motivation identification phase, clustering was used to categorize nodes such as winter range, battery efficiency, and thermal management systems into the "range reliability" demand cluster, while nodes such as autonomous driving, lane recognition, and severe weather were categorized into the "intelligent driving safety" demand cluster. Causal path analysis revealed that users' core demand for "range reliability" is not absolute long range, but rather the accuracy and stability of range prediction under various operating conditions such as low temperatures. The underlying motivation is to eliminate range anxiety and ensure travel reliability. On the other hand, the core demand behind "intelligent driving safety" is the stable performance of the system in extreme scenarios and clear notification of its capability boundaries. The motivation is to build trust in the technology. The resulting structured report proposes optimization of the battery thermal management system, development of a more accurate dynamic range prediction model, and product improvement and service communication suggestions that clearly inform users of the capability boundaries of the intelligent driving system.

[0095] Based on the above technical solutions, S3 achieves the structured classification of scattered needs through demand clustering and label generation in S301, solving the problem of fragmented needs being difficult to extract; through the construction of causal directed acyclic graphs and path mining in S302, it reveals the driving logic behind the needs; through the motivation induction and structured report generation in S303, it transforms the technical analysis results into decision-making basis that enterprises can directly apply, making full use of the semantic association and temporal features of the graph, ensuring both the accuracy of user demand identification and the depth and interpretability of user motivation mining.

[0096] In one possible implementation of the embodiments of this application, combined with Figure 1 ,like Figure 5 As shown, the above S4 specifically includes the following S401 to S403: S401. Detect concept drift using statistical testing methods to identify differences between newly collected data and historical data.

[0097] Concept drift refers to the significant changes in the characteristic distribution or demand patterns of user demand-related data over time, such as the emergence of new demands or the weakening of existing demands. Failure to capture these changes in a timely manner can lead to a disconnect between the dynamic demand map and actual demand. The goal of this step is to determine whether a map update is necessary by quantitatively comparing the differences between new and historical data, which is a prerequisite for ensuring the timeliness of demand evolution tracking.

[0098] In some implementations, step S401 is performed as follows: The newly collected data is transformed into a multimodal data stream with a unified representation according to the preprocessing flow S1, and its feature vectors are extracted and denoted as the dataset. The set of feature vectors is Extract a set of feature vectors of the same period scale from historical data, denoted as the dataset. The set of feature vectors is And the sample sizes are matched; Selecting statistical test methods and calculating test statistics includes: If a characteristic distribution is used for difference testing, such as the Kolmogorov-Smirnov test, then the formula for the test statistic is: ;in, This means taking the maximum value among all x values. Let be the cumulative distribution function of the new data feature vectors. KS is the cumulative distribution function of the feature vector of historical data; the value of KS ranges from [0,1], and the larger the value, the more significant the distribution difference.

[0099] If we use model prediction results for difference testing, such as using KL divergence (Kullback-Leibler Divergence) to compare the differences in the distribution of demand identification results based on new data and historical data (the demand identification results are the distribution of demand cluster labels output by S3), then the formula for the test statistic is: ;in, For the demand cluster in the new data The probability of its occurrence, i.e. , For the new data belonging to the cluster The number of samples, This represents the total number of samples in the new data. Demand clusters in historical data The probability of occurrence is given by m, where m is the total number of demand clusters. The KL divergence ranges from [0, +∞), where a value of 0 indicates a completely uniform distribution, and a larger value indicates more significant differences in demand patterns.

[0100] When KS value ≥ or KL divergence ≥ If a concept drift is detected, the newly collected data is stored in the buffer pool; otherwise, if no significant drift is detected, the new data is not included in the update process. , This is the preset threshold for the test statistic.

[0101] S402. Trigger incremental update process based on buffer pool data size to ensure a balance between update efficiency and data validity.

[0102] The buffer pool is a container that temporarily stores newly collected data that has been detected as having a concept drift. Its purpose is to avoid the waste of computing resources caused by frequent updates and to ensure the accuracy of updates by accumulating a certain amount of data.

[0103] In some implementations, step S402 is performed as follows: The buffer pool data is managed using a first-in, first-out (FIFO) strategy. When new data is stored, if the buffer pool data exceeds the maximum capacity, the oldest expired data is automatically deleted. Preset buffer pool trigger thresholds, such as setting a data volume threshold. Typically, this is 500-1000 records, or a data storage duration threshold. (Usually 3-7 days). When the amount of data in the buffer pool is ≥ or data storage duration ≥ If the threshold is not reached, the incremental update process is triggered; if the threshold is not reached, new data continues to be accumulated. After the update is triggered, the complete preprocessing flow of S1 is performed on all data in the buffer pool to generate a multimodal data stream with a unified representation, ensuring compatibility with historical data.

[0104] S403. Perform incremental update operation to update the nodes, edges and embedding vectors of the dynamic demand graph to achieve demand evolution tracking.

[0105] In some implementations, step S403 is performed as follows: First, for the valid data preprocessed in S402, the method in S201 is used to extract new required entities and new relationships. New entities need to undergo naming consistency verification with entities in the historical graph to avoid duplicate nodes; new relationships need to pass the initial screening of relationship strength in S201 to ensure validity.

[0106] Then, the new entities are added as new nodes and the new relationships are added as new edges to the existing dynamic demand graph; the relationship strength between the original nodes is recalculated based on the new data and historical sliding window data (k windows before the current window, where k is the number of memory steps of S203); Next, an online graph representation learning algorithm (Incremental Graph Convolutional Network - GCN) is used to iteratively update the embedding vectors of all nodes based on the updated graph structure, using the following formula: ;in, Let e_i be the updated embedding vector for node e_i. For nodes The embedding vector before the update The weight coefficients of the historical embedding vector. This is the weighting coefficient for new information, typically taken as 0.3. For nodes The updated set of neighbor nodes, For neighboring nodes The updated feature vector, where b is the bias vector. It is the ReLU activation function; Finally, the changes in this update are recorded, including the number of new entities / edges, node pairs whose association strength changes by more than 0.2, and new requirement clusters, forming a requirement evolution trajectory log.

[0107] Based on the above technical solutions, S4 first captures pattern changes in demand data through statistical testing to avoid invalid and missed updates; then, it balances update efficiency and resource costs through a buffer pool mechanism; finally, through an incremental update strategy, it incorporates new information while retaining historical time sequence memory, which not only ensures the real-time performance and accuracy of the map, but also avoids the resource waste of full updates. The generated dynamic demand map can completely track the entire life cycle of demand from its inception to its evolution, providing dynamic knowledge support for the continuous updating of S3 structured reports, and solving the technical defects of traditional demand analysis methods that are static, lagging, and unable to adapt to the rapid evolution of demand.

[0108] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0109] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and modifications.

Claims

1. A method for analyzing user needs of an AI engine with multi-source data input, characterized in that, include: Multimodal data from different data sources is acquired, preprocessed, and semantically fused to generate a multimodal data stream with a unified representation; wherein, the data sources include public network data sources and Internet of Things (IoT) data sources; Based on the multimodal data stream, a dynamic demand graph is constructed using a graph neural network, with demand entities as nodes and relationships as edges. Based on the dynamic demand map, graph reasoning and causal discovery algorithms are used to identify user needs and user motivations, and generate a structured demand analysis report. Based on real-time monitoring data, the dynamic demand map is updated through an incremental learning algorithm to continuously track the evolution of demand.

2. The method for analyzing user needs of an AI engine with multi-source data input according to claim 1, characterized in that, The preprocessing and semantic fusion include: Natural language processing algorithms and named entity recognition technology are used to process text data and output a five-tuple structured text representation containing product, product function, usage scenario, user attribute, and user evaluation. Sensor data is processed using signal processing algorithms, transforming the sensor data stream into an event-based feature sequence that reflects device status or user interaction; wherein, the sensor data includes customer flow heatmap data, POS transaction data, and shelf interaction sensing data collected in offline retail scenarios; The scene graph generation model is used to process image data and output a visual semantic scene graph containing objects, attributes, and relationships between objects. Locality-sensitive hashing algorithm is used to perform semantic deduplication and clustering on the processed data, and data describing the same semantic unit are divided into the same hash bucket to obtain deduplicated multi-source data. The tensor decomposition method is used to perform semantic fusion on the deduplicated multi-source data, mapping text, sensor and image multimodal data to a unified semantic space to generate a multimodal data stream.

3. The method for analyzing user needs of an AI engine with multi-source data input according to claim 2, characterized in that, The process of using signal processing algorithms to process sensor data and transform the sensor data stream into an event-based feature sequence reflecting device status or user interaction includes: Wavelet transform is applied to the sensor data for filtering and noise reduction to extract the effective signal components; The isolated forest algorithm is used to detect outliers in the denoised signal and identify data points that deviate from the normal pattern. Based on preset domain knowledge rules or patterns learned through unsupervised algorithms, continuous signal anomaly segments or combinations of specific signal patterns are defined as atomic events. Each atomic event represents a complete product state or user behavior unit. The continuous signal anomaly segments are composed of continuous abnormal data points identified by the isolated forest algorithm. The detected atomic events are sorted by timestamp to obtain an event-based feature sequence.

4. The method for analyzing user needs of an AI engine with multi-source data input according to claim 2, characterized in that, The process of using a scene graph generation model to process image data and output a visual semantic scene graph containing objects, attributes, and relationships between objects includes: Use the object detection subnetwork to identify all entity objects in the image and their bounding boxes; The visual attributes of each identified object are predicted through an attribute classification sub-network; wherein, the attribute classification sub-network is a multi-label classification model built on a convolutional neural network; the visual attributes include color, shape, material, and visual state; Using a relation prediction subnetwork, the predicate relationship between each pair of objects is determined based on their spatial location, relative size, and visual context; wherein, the relation prediction subnetwork is a relation classification model built based on a neural network algorithm; The recognition results are organized into a directed graph structure with entity objects as nodes, visual attributes as node labels, and predicate relations as edges to obtain a visual semantic scene graph. The scene graph generation model is an end-to-end or cascaded deep learning model consisting of three sub-networks: object detection, attribute classification, and relationship prediction.

5. The method for analyzing user needs of an AI engine with multi-source data input according to claim 2, characterized in that, The locality-sensitive hashing algorithm is used to perform semantic deduplication and clustering on the processed data, dividing data describing the same semantic unit into the same hash bucket to obtain deduplicated multi-source data, including: Structured data from different modalities are mapped into unified semantic vectors through pre-trained encoders; wherein the structured data from different modalities includes structured text representations, event-based feature sequences, and visual semantic scene graphs. A set of hyperplanes is randomly generated for the function family of locality-sensitive hashing algorithms to partition the vector space; For each data semantic vector, a binary hash signature is generated based on the relative position of the semantic vector with each hyperplane; Data points with hash signatures that are identical or whose Hamming distance is less than a preset distance threshold are assigned to a hash bucket; Data within the same hash bucket is aggregated, retaining the sample with the latest timestamp or generating an aggregated feature vector to obtain deduplicated multi-source data.

6. The method for analyzing user needs of an AI engine with multi-source data input according to claim 1, characterized in that, The process of constructing the dynamic demand map includes: Demand entities and relationships are extracted based on multimodal data streams. Demand entities include products, product functions, usage scenarios, user attributes and user reviews extracted from text, performance indicators mapped from sensor events, and visual objects identified from images. Relationships include causal relationships, temporal connections or semantic connections between entities. A demand graph is constructed using a graph neural network as the backbone network; where nodes in the demand graph represent demand entities, and edges in the demand graph represent the association strength between entities. The association strength is quantified by the frequency of occurrence of the association between entities within a time window and the semantic association strength calculated by the attention mechanism. The multimodal data stream is captured using a sliding window approach. A graph neural network is used to encode the data and historical graph state within each time window to generate node embeddings with temporal memory, thereby obtaining a dynamic demand graph. The historical graph state represents the embedding vector sequence of nodes over multiple consecutive time steps in the past.

7. The method for analyzing user needs of an AI engine with multi-source data input according to claim 6, characterized in that, The association strength is quantified by the frequency of occurrence of the association between entities within a time window and the semantic association strength calculated by the attention mechanism, including: For any association within a given time window, calculate the normalized value of the frequency of the association to obtain the frequency intensity; By using the attention mechanism in the graph attention network, the attention coefficient between the head entity node and the tail entity node is calculated to obtain the semantic association strength between the two entities; wherein, the head entity represents the starting node in the relation triple and the tail entity represents the ending node in the relation triple. The association strength is obtained by weighted summation of the frequency intensity and the semantic association strength.

8. The method for analyzing user needs of an AI engine with multi-source data input according to claim 6, characterized in that, The method of encoding data and historical graph states within each time window using a graph neural network includes: For the graph at the current time t, the attention weights of each target node and its first-order neighbor nodes are calculated using the multi-head graph attention mechanism in the graph neural network, and the features of the neighbor nodes are aggregated in a weighted manner to obtain the spatial aggregated features of each target node at time t; wherein, the graph at the current time t is composed of entities and relations extracted within the current time window; The spatial aggregation features of the target node at time t are combined with the final node embedding vectors of the target node over the past k consecutive time steps to form a feature sequence in chronological order. The feature sequence is input into a gated loop unit, which outputs a new embedding vector of the target node at time t that includes the historical graph state. Once all nodes have been updated, a dynamic demand graph encoding temporal dependencies is obtained.

9. The method for analyzing user needs of an AI engine with multi-source data input according to claim 1, characterized in that, The process of identifying user needs and user motivations based on the dynamic demand graph, using graph reasoning and causal discovery algorithms, includes: Hierarchical clustering or community detection algorithms are used to cluster demand entity nodes into multiple demand clusters based on the embedding vectors of nodes in the dynamic demand graph, and a generalized label is generated for each demand cluster through a large language model. On a graph subgraph containing entities including user attributes, usage scenarios, user reviews, and product functions, a causal discovery algorithm is used to search for conditional independence relationships between variables, and a causal directed acyclic graph structure representing the causal relationships between entities is output. Based on the aforementioned causal directed acyclic graph structure, starting from the root nodes of the two types of entities, "user attributes" and "usage scenarios", traverse the causal paths pointing to "user reviews" and "product functions". By combining the generalized tags of the aforementioned demand clusters, the causal path is summarized into a description of user motivation using a large language model.

10. The method for analyzing user needs of an AI engine with multi-source data input according to claim 1, characterized in that, The step of updating the dynamic demand map using an incremental learning algorithm includes: Using statistical testing methods, the differences between newly collected data and historical data in feature distribution or model prediction results are compared. When the test statistic exceeds the preset threshold, it is determined to be a case of inverted concept drift, and the newly collected data is stored in the buffer pool. When the buffer pool data reaches a preset size, an incremental update process is initiated; the incremental update process is as follows: New entities and relationships are extracted based on the newly collected data and added to the existing dynamic demand graph as new nodes or edges; Using an online graph representation learning algorithm, the embedding vectors in the dynamic demand graph are iteratively updated based on new entities and relationships; The edge association strength is recalculated within the sliding time window to update the edge weights.

Images