Industrial graph construction method and system based on scene type marketing

Abstract

The invention provides an industrial graph construction method and system based on scene type marketing, and relates to the technical field of industrial graph construction, and the method comprises the steps: processing a media text through a BERT pre-training model to generate a text data set, processing scene data through a space-time calibration algorithm to generate a behavior data set, and summarizing to generate an unstructured marketing data set. Generating a multi-modal marketing scene data set in combination with the structured marketing data set; processing the multi-modal data set by adopting a stream-batch integrated framework and an entity disambiguation technology, and constructing a four-dimensional data lake; constructing a three-dimensional feature library and training a marketing scene weight model; a data lake is mapped to a graph database, an industrial chain topological structure is formed, node relation strength is adjusted based on a weight model, node implicit features are learned through a GraphSAGE algorithm, and a scene-sensitive industrial graph topological structure is generated, so that industrial entity relations in different marketing scenes can be reflected in real time, and the marketing efficiency is improved. And the accuracy of the industrial map in marketing decision making is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial map construction, and in particular to a method and system for constructing an industrial map based on scenario-based marketing. Background Art

[0002] Traditional industrial map construction methods rely heavily on structured data such as company registration information and industry reports, and have significant technical bottlenecks: they only contain limited dimensions such as basic company information, data updates lag and are of a single dimension, and cannot accurately reflect real-time supply and demand changes in the industrial chain, and the coordination between human thinking intervention and dynamic updates is poor; a complete scenario-sensitive mechanism has not been established, and there is a lack of deep integration of unstructured data such as social media public opinion and user behavior trajectories, resulting in the existing industrial map being unable to accurately reflect market dynamics and making marketing decisions out of touch with actual market demand; a fixed weight algorithm is used to integrate data, which makes it difficult to adapt to the differentiated needs of different marketing scenarios.

[0003] Therefore, it is necessary to provide an industry map construction method and system based on scenario-based marketing to solve the above technical problems. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides an industrial map construction method and system based on scenario-based marketing, which is used to solve the problems in the existing technology that data updates are delayed and have a single dimension, cannot accurately reflect the real-time supply and demand changes in the industrial chain, and lack the unstructured data fusion capabilities required for scenario-based marketing, resulting in a disconnect between marketing decisions and market demand.

[0005] The present invention provides a method for constructing an industry map based on scenario-based marketing, the method comprising: Social media text is obtained through distributed crawlers and processed by the BERT pre-trained model to generate a marketing text dataset with semantic labels. At the same time, offline scene data is collected through IoT devices and processed by the spatiotemporal calibration algorithm to generate a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. Combined with the structured marketing dataset, a multimodal marketing scenario dataset is generated. The Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology are used to preprocess the multimodal marketing scenario dataset and build a four-dimensional dynamic data lake; Obtain a historical marketing scenario dataset and construct a three-dimensional marketing scenario feature library, and train the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model; The four-dimensional dynamic data lake is mapped to the Neo4j graph database to form an industrial chain topology structure. The node relationship strength in the industrial chain topology structure is adjusted based on the marketing scenario weight model. The graph neural network is introduced and the implicit features of the nodes in the industrial chain topology structure are learned through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure.

[0006] Preferably, social media text is obtained through a distributed crawler, and a marketing text dataset with semantic labels is generated after processing by a BERT pre-training model. At the same time, offline scene data is collected through IoT devices, and a marketing behavior dataset with spatiotemporal labels is generated after processing by a spatiotemporal calibration algorithm. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset, and combined with the structured marketing dataset to generate a multimodal marketing scenario dataset, specifically including: The social media text is acquired in real time by the distributed crawler cluster, and the BERT pre-trained model is used to semantically encode the social media text. The industry chain entities in the social media text are extracted by named entity recognition technology, and the association relationships between the industry chain entities are constructed by using dependency syntax analysis. The industry chain entities and the association relationships are summarized to generate the marketing text dataset with semantic labels; Synchronously using an edge computing gateway to connect to the IoT device to collect the offline scene data, and performing geo-fence matching and timestamp calibration on the offline scene data using the spatiotemporal calibration algorithm, that is, mapping the offline scene data to spatial entity coordinates and marking the collection time granularity of the offline scene data, thereby forming the marketing behavior dataset with spatiotemporal labels; Aggregating the marketing text dataset with semantic tags and the marketing behavior dataset with spatiotemporal tags to obtain the unstructured marketing dataset, and synchronously accessing the ERP system to obtain the structured marketing dataset; The unstructured marketing dataset and the structured marketing dataset are aggregated to form the multimodal marketing scenario dataset.

[0007] Preferably, the Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology are used to preprocess the multimodal marketing scenario dataset and construct a four-dimensional dynamic data lake, specifically including: Using the Apache Flink stream-batch integrated processing framework to perform real-time event processing on the unstructured marketing dataset and simultaneously perform micro-batch processing on the structured marketing dataset; Adopting the knowledge graph entity disambiguation technology, constructing an industry chain entity dictionary based on the knowledge graph, and using a deep semantic matching algorithm to perform entity disambiguation on the unstructured marketing dataset and the structured marketing data based on the industry chain entity dictionary and establish a unified entity identifier; The multimodal marketing scenario dataset is mapped to the four-dimensional dynamic data lake based on a distributed storage architecture.

[0008] Preferably, the process of acquiring a historical marketing scenario dataset and constructing a three-dimensional marketing scenario feature library, and training the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model, specifically includes: Acquire the historical marketing scene dataset, extract temporal marketing features, spatial marketing features, and behavioral marketing features from the historical marketing scene dataset, and construct the three-dimensional marketing scene feature library; The LSTM neural network is used to train the three-dimensional marketing scene feature library, match the marketing weight vectors corresponding to the marketing scene features in the three-dimensional marketing scene feature library, and generate the marketing scene weight model.

[0009] Preferably, the four-dimensional dynamic data lake is mapped to the Neo4j graph database to form an industrial chain topology structure, the node relationship strength in the industrial chain topology structure is adjusted based on the marketing scenario weight model, and the graph neural network is introduced and the implicit features of the nodes in the industrial chain topology structure are learned through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure, which specifically includes: Mapping the four-dimensional dynamic data lake to the Neo4j graph database to construct the industrial chain topology structure, wherein the enterprise ID, product ID, and user ID in the four-dimensional dynamic data lake are mapped as nodes in the industrial chain topology structure, the supply and demand relationship and public opinion association in the four-dimensional dynamic data lake are mapped as edges in the industrial chain topology structure, and the spatiotemporal attributes in the four-dimensional dynamic data lake are used as labels for nodes or edges in the industrial chain topology structure; Adjust edge weights, i.e., node relationship strengths, in real time based on the marketing scenario weight model; The graph neural network is introduced, and the GraphSAGE algorithm is used to learn the embedded representation of the nodes. The features of adjacent nodes are weighted and aggregated in combination with the strength of the node relationships to obtain the implicit features of the nodes. The industrial chain topology is recalculated based on the node relationship strength and the node implicit features, and the key node clusters in the updated industrial chain topology are identified through a community discovery algorithm to generate the scenario-sensitive industrial graph topology.

[0010] Preferably, the real-time adjustment of edge weights based on the marketing scenario weight model specifically includes: Using the FastText text classification model to perform keyword extraction and semantic classification on the current marketing scenario description text in the four-dimensional dynamic data lake, determine the corresponding current marketing scenario type, and extract the corresponding current marketing scenario features; Based on the marketing scenario weight model, the current marketing weight vector corresponding to the current marketing scenario feature, that is, the edge weight, is matched.

[0011] Preferably, the entity set in the four-dimensional dynamic data lake is defined as ,in represents the i-th entity, i.e., enterprise ID, product ID or user ID, n represents the total number of entities in the four-dimensional dynamic data lake, and the entity set E is mapped to the node set in the industrial chain topology structure G. ,and , and map() represents the mapping function from entity to node; the edge set in the industrial chain topology structure G is defined as ,in Represents the i-th node and the jth node The relationship type between , defines the spatiotemporal attribute set in the four-dimensional dynamic data lake as ,in Represents the time attribute, Represents spatial attributes; The calculation formula of the node implicit feature is as follows: Where, Representation node Hidden features of nodes in the lth layer of the graph neural network; represents the activation function; represents the feature aggregation function of the lth layer; Representation node Hidden features of nodes in the l-1 layer of the graph neural network; Represents the adjusted edge weight, i.e., node and The strength of the node relationships between them; Representation node The set of neighbor nodes of Representation node Hidden features of nodes in the l-1th layer of the graph neural network.

[0012] Preferably, the recalculating the industry chain topology structure according to the node relationship strength and the node implicit features specifically includes: The similarity function is constructed based on the node relationship strength and the node implicit features as follows: Where, Representation node and nodes Similarity function; L represents the last layer of the graph neural network; Representation node Hidden features of nodes in the last layer of the graph neural network; Representation node Hidden features of nodes in the last layer of the graph neural network; Represents node implicit features The L2 norm of Represents node implicit features The L2 norm of Representation node Node implicit features With node Node implicit features The cosine similarity of The empirical coefficient representing the amplification factor of the node relationship strength, It represents the amplification factor of the node relationship strength, ; Setting similarity threshold ,when When retaining the edge , otherwise delete the edge , recalculate the industrial chain topology structure to obtain the updated industrial chain topology structure ; The method of identifying key node clusters in the updated industry chain topology structure by using a community discovery algorithm to generate the scenario-sensitive industry graph topology structure specifically includes: The updated industry chain topology Each node in the network is considered as a community, forming ,in represents the i-th community; The community discovery algorithm is used to calculate the community and neighboring communities The community modularity is and , and calculate the community With neighboring communities The community modularity of the new community after the merger is , and the community modularity gain is ; If the community modularity gain , then merge the communities and neighboring communities , otherwise keep the community and neighboring communities independent; Traverse all adjacent community pairs to obtain the final community set , where p represents the final community set Total number of communities in Calculate the final community set Each community Community importance score ,in , importance scores for all communities Arrange them from large to small, select the communities corresponding to the top M community importance scores to form a key community cluster, that is, the key node cluster, and generate the scenario-sensitive industry map topology structure.

[0013] A method for constructing an industry map based on scenario-based marketing, further comprising: providing a three-dimensional visualization engine based on WebGL to support a user end to perform spatiotemporal slicing of the scenario-sensitive industry map topology structure through geo-fencing and timestamps; The improved PageRank algorithm is used to calculate the importance of nodes in the scenario-sensitive industry graph topology structure, and the DBSCAN clustering algorithm is used to identify key marketing touchpoints based on the importance of the nodes and generate marketing decision recommendations, and the key marketing touchpoints are marked in the form of a heat map; An interactive analysis interface is provided to support the user terminal to click on the key marketing touchpoints and view the marketing decision suggestions.

[0014] An industry map construction system based on scenario-based marketing, the construction system comprising: The data collection module is used to obtain social media text through distributed crawlers, generate a marketing text dataset with semantic labels after processing it with the BERT pre-trained model, and simultaneously collect offline scene data through IoT devices. After processing it with the spatiotemporal calibration algorithm, it generates a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. This is combined with the structured marketing dataset to generate a multimodal marketing scenario dataset. A data processing module, configured to pre-process the multimodal marketing scenario dataset and construct a four-dimensional dynamic data lake using the Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology; A model building module is used to obtain a historical marketing scenario dataset and build a three-dimensional marketing scenario feature library, and train the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model; A graph construction module is used to map the four-dimensional dynamic data lake to the Neo4j graph database to form an industrial chain topology structure, adjust the node relationship strength in the industrial chain topology structure based on the marketing scenario weight model, introduce a graph neural network and learn the implicit features of the nodes in the industrial chain topology structure through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure.

[0015] Compared with related technologies, the method and system for constructing an industry map based on scenario-based marketing provided by the present invention have the following beneficial effects: This paper obtains social media texts through distributed crawlers, generates a marketing text dataset with semantic labels after processing with the BERT pre-training model, and collects offline scene data through IoT devices. After processing with the spatiotemporal calibration algorithm, it generates a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset, which is combined with the structured marketing dataset to generate a multimodal marketing scenario dataset. The Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology pre-process the multimodal marketing scenario dataset and construct a four-dimensional dynamic data lake; obtain the historical marketing scenario dataset and construct a three-dimensional marketing scenario feature library, and train the three-dimensional marketing scenario feature library through the LSTM neural network to obtain the marketing scenario weight model; map the four-dimensional dynamic data lake to the Neo4j graph database to form an industrial chain topology structure, adjust the node relationship strength in the industrial chain topology structure based on the marketing scenario weight model, introduce the graph neural network and learn the implicit features of the nodes in the industrial chain topology structure through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure, so as to construct a scenario-sensitive industrial graph to reflect the relationship between industrial entities in different marketing scenarios in real time, thereby improving the effectiveness and accuracy of the industrial graph in marketing decision-making.

[0016] The present invention generates a multimodal marketing scenario dataset by integrating the semantic data of social media text and the spatiotemporal behavior data collected by IoT devices, combined with the structured data of the ERP system, so that the construction of the industry map is based on a more comprehensive and rich data foundation; the method of the present invention utilizes the Apache Flink stream-batch integrated processing framework to realize real-time processing of unstructured data and micro-batch processing of structured data, ensuring the timeliness and efficiency of data processing; the present invention constructs a three-dimensional marketing scenario feature library and uses LSTM neural network training to obtain a marketing scenario weight model, so that the node relationship strength in the industry chain topology structure can be adjusted in real time according to different marketing scenarios, making the industry map scenario-sensitive and able to more accurately reflect the association relationship of industry entities in different marketing scenarios; the present invention combines graph neural networks and GraphSAGE algorithms to learn node implicit features, optimize the industry chain topology structure through similarity functions and community discovery algorithms, and generate a more accurate scenario-sensitive industry map; in addition, the present invention provides users with intuitive and effective marketing decision support through a three-dimensional visualization engine, node importance calculation, key marketing touchpoint identification and interactive analysis interface, thereby improving the scientificity and accuracy of marketing decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a method for constructing an industry map based on scenario-based marketing provided by an embodiment of the present invention; Figure 2 A system block diagram of an industry graph construction system based on scenario-based marketing provided by an embodiment of the present invention; Figure 3 A schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] like Figure 1 FIG. 1 is a flow chart of a method for constructing an industry graph based on scenario-based marketing according to an embodiment of the present invention. Figure 1The execution subject of the method shown may be a software and / or hardware device. The execution subject of the present application may include but is not limited to at least one of the following: user equipment, network equipment, etc. Among them, user equipment may include but is not limited to computers, smart phones, personal digital assistants (PDAs) and the electronic devices mentioned above. Network equipment may include but is not limited to a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers, wherein cloud computing is a type of distributed computing, a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. It includes steps S1 to S4, as follows: S1: social media text is obtained through distributed crawlers and processed with the BERT pre-trained model to generate a marketing text dataset with semantic labels. At the same time, offline scene data is collected through IoT devices and processed with a spatiotemporal calibration algorithm to generate a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. This is combined with the structured marketing dataset to generate a multimodal marketing scenario dataset. Among them, a distributed crawler is a technology that captures network data in parallel through multiple computer nodes. It uses a distributed architecture to improve data collection efficiency and is used to obtain large-scale social media texts in real time, such as comments, posts, messages, etc., solving the problem of slow crawling speed and easy blocking of single nodes. The BERT pre-trained model refers to a pre-trained language model based on the Transformer architecture. It is trained through a bidirectional masked language model and can capture the deep semantics of the text. In the method of the present invention, it is used to perform named entity recognition, relationship extraction and other processing on social media texts to generate a marketing text dataset with semantic labels. A marketing text dataset with semantic labels refers to a dataset containing clear semantic annotations generated after semantic analysis of social media texts.

[0020] IoT (Internet of Things) devices are physical devices with network connectivity, such as Wi-Fi probes, thermal sensors, and RFID tags. They are used to collect offline scene data, such as foot traffic in shopping districts and user movement trajectories. Spatiotemporal calibration algorithms, such as R-tree spatiotemporal indexing, perform timestamp calibration and geospatial matching on data collected by IoT devices. These algorithms can map raw sensor data to specific geofences and time granularities, generating spatiotemporally labeled marketing behavior datasets. A geofence might be the coordinates of a shopping mall, and a time granularity might be 10 minutes per visit. Spatiotemporally labeled marketing behavior datasets refer to offline scene data collected by IoT devices, generated after spatiotemporal calibration and annotated with both temporal and spatial dimensions.

[0021] Unstructured marketing datasets refer to data sets that do not follow fixed data models, including social media text and user behavior trajectories. These data formats are relatively flexible. The method of the present invention converts unstructured data into processable, labeled data through semantic parsing and spatiotemporal alignment. Structured marketing datasets refer to data with a clear data model and format, such as order data in ERP (Enterprise Resource Planning) systems and structured statistical data in industry reports. These are combined with unstructured data to form multimodal marketing scenario datasets that cover the full range of information required for marketing decision-making. Multimodal marketing scenario datasets are comprehensive datasets that integrate multiple data modalities and cover the full range of information in marketing scenarios.

[0022] In practical applications, distributed crawling technology can be used to bulk-collect user-generated text content from social media platforms. This text can then be semantically parsed using the BERT pre-trained model. Natural language processing techniques such as named entity recognition and relationship extraction can then be used to generate structured marketing text datasets with entity labels, relationship annotations, and sentiment. Simultaneously, IoT devices can be used to collect user behavior data from offline scenarios. This data can be synchronized with timestamps and geofences using spatiotemporal calibration algorithms to generate marketing behavior datasets containing both location coordinates and time dimensions. Finally, these two types of unstructured data can be integrated with structured data such as order data from enterprise resource planning systems and industry analysis reports to construct a multimodal marketing scenario dataset encompassing semantic information, spatiotemporal behavior, and business statistics.

[0023] S2, using the Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology to preprocess the multimodal marketing scenario dataset and build a four-dimensional dynamic data lake; It's understood that the Apache Flink stream-batch processing framework refers to an open-source big data processing framework that supports unified processing of both real-time streaming data, such as live stream comments, and batch data, such as daily reports. Knowledge graph entity disambiguation technology solves the entity aliasing problem in multi-source data through a deep semantic matching algorithm. It also builds an entity dictionary based on the knowledge graph, assigning unified identifiers to entities from different data sources, enabling entity alignment across modal data and ensuring accurate data fusion. A four-dimensional dynamic data lake refers to a distributed data storage architecture constructed along four dimensions: entities, relationships, time and space, and attributes. It supports real-time incremental data updates and data lineage tracking.

[0024] It's important to note that the Apache Flink stream-batch processing framework can be used to uniformly clean and standardize real-time live stream comments, device sensor data, and batch-imported daily sales reports. Knowledge graph entity disambiguation technology addresses the semantic consistency of company names and product aliases across different data sources, thereby building a four-dimensional dynamic data lake. This data lake continuously integrates new data through incremental learning, ensuring minimal data update latency.

[0025] S3, obtaining a historical marketing scenario dataset and building a three-dimensional marketing scenario feature library, and training the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model; It should be noted that the historical marketing scenario data set refers to a full-dimensional data set generated in past marketing activities, covering business data, user behavior data and marketing effect data in different marketing scenarios. The three-dimensional marketing scenario feature library refers to a feature set constructed based on historical marketing data, which includes time dimension, space dimension and behavior dimension. It reflects the core attributes of different marketing scenarios through three-dimensional feature vectors, among which time dimension is such as the promotion cycle stage, space dimension is such as the regional consumption power index, behavior dimension is such as user conversion rate, and marketing scenarios are such as new product launches, holiday promotions, regional promotions and other scenarios. LSTM neural network refers to a long short-term memory network, which is used in the method of the present invention to train the three-dimensional feature library, learn the relationship between scene features and weight mapping, and generate a scene-sensitive marketing scenario weight model. The marketing scenario weight model refers to an algorithm model obtained by training based on the historical marketing scenario data set, which is used to dynamically adjust the calculation weights of each data dimension in the industry map according to the current marketing scenario type.

[0026] In practical applications, full data from different scenarios can be collected from historical marketing activities to construct a three-dimensional feature library encompassing time, space, and behavior dimensions. This feature library can then be trained using LSTM to learn the mapping between different scenario features and data weights, generating a marketing scenario weighting model. This model can automatically adjust the calculation weights for dimensions such as supply and demand relationships and competitive dynamics in industry chain analysis based on the real-time scenario type, achieving dynamic adaptation of data weights.

[0027] S4, mapping the four-dimensional dynamic data lake to the Neo4j graph database to form an industrial chain topology structure, adjusting the node relationship strength in the industrial chain topology structure based on the marketing scenario weight model, introducing a graph neural network and learning the implicit features of the nodes in the industrial chain topology structure through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure.

[0028] Neo4j graph databases are NoSQL databases based on graph models. They use nodes, edges, and attributes to represent data and relationships, making them suitable for storing and querying complex relationships, such as those in industrial chains. They can map four-dimensional data lakes into graph structures, supporting efficient graph topology analysis. Node relationship strength refers to the weight of edges in an industrial chain topology, reflecting the closeness of connections between nodes. Graph neural networks (GNNs) are neural networks that operate directly on graph structures. The GraphSAGE algorithm, a classic algorithm in GNNs, addresses the training efficiency challenges of large-scale graph data by sampling neighboring nodes and aggregating their features. Node implicit features are abstract feature vectors containing deep semantic information, generated through multi-layer aggregation learning of node features in the industrial chain topology via GNNs. A scenario-sensitive industrial graph topology structure dynamically adjusts node relationship strength and topological connectivity based on the current marketing scenario. It can adaptively present key nodes and core relationships that are highly relevant to the marketing scenario.

[0029] Furthermore, the entities, relationships, and spatiotemporal attributes in the four-dimensional dynamic data lake can be mapped to a Neo4j graph database to construct an industrial chain topology consisting of nodes and edges. Then, the edge weights in the graph can be calibrated in real time based on a marketing scenario weight model, for example, enhancing the interaction edge weights between stores and users in promotional scenarios. Furthermore, a graph neural network can be introduced, using the GraphSAGE algorithm to perform multi-layer learning on node features. By weightedly aggregating the scenario-sensitive features of neighboring nodes, implicit node representations containing deep semantics are generated, such as store features that integrate regional customer flow and user reviews. Finally, the implicit node features and dynamic edge weights can be combined to recalculate the industrial chain topology through a community discovery algorithm, identifying highly correlated key node clusters, such as core promotional business districts, and generating an industrial graph topology that directly serves the current marketing scenario, achieving an upgrade from data display to decision support.

[0030] In the specific implementation process, social media text is obtained through distributed crawlers, and after processing with the BERT pre-training model, a marketing text dataset with semantic labels is generated. At the same time, offline scene data is collected through IoT devices, and after processing with the spatiotemporal calibration algorithm, a marketing behavior dataset with spatiotemporal labels is generated. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. Combined with the structured marketing dataset, a multimodal marketing scenario dataset is generated, which specifically includes: The social media text is acquired in real time by the distributed crawler cluster, and the BERT pre-trained model is used to semantically encode the social media text. The industry chain entities in the social media text are extracted by named entity recognition technology, and the association relationships between the industry chain entities are constructed by using dependency syntax analysis. The industry chain entities and the association relationships are summarized to generate the marketing text dataset with semantic labels; Synchronously using an edge computing gateway to connect to the IoT device to collect the offline scene data, and performing geo-fence matching and timestamp calibration on the offline scene data using the spatiotemporal calibration algorithm, that is, mapping the offline scene data to spatial entity coordinates and marking the collection time granularity of the offline scene data, thereby forming the marketing behavior dataset with spatiotemporal labels; Aggregating the marketing text dataset with semantic tags and the marketing behavior dataset with spatiotemporal tags to obtain the unstructured marketing dataset, and synchronously accessing the ERP system to obtain the structured marketing dataset; The unstructured marketing dataset and the structured marketing dataset are aggregated to form the multimodal marketing scenario dataset.

[0031] Among them, a distributed crawler cluster architecture can be adopted to capture user-generated content from social media platforms in real time through a multi-node parallel computing strategy, forming a large-scale text data collection network. The original text is deeply semantically encoded with the help of the BERT pre-training model, and named entity recognition technology is used to accurately extract the industry chain entities contained in the text, covering semantic units such as corporate entities, product types, and market behaviors. Through dependency syntax analysis technology, the grammatical dependency relationships and semantic logical associations between entities are mined, and structured association networks such as supply relationships, competitive relationships, and consumer behaviors are constructed. Unstructured text is converted into a semantically annotated dataset containing entity labels and relationship links, providing semantic support for the text dimension of the industry map.

[0032] Furthermore, an IoT device access hub can be built through an edge computing gateway to achieve unified management and data collection of offline scene sensing devices, including terminal devices such as wireless probes, thermal sensors, and radio frequency identification tags. Then, a spatiotemporal calibration algorithm can be used to perform dual processing on the collected raw sensor data: in the spatial dimension, sensor data is mapped to specific physical coordinate entities such as shopping mall areas and store locations through geofence matching technology; in the temporal dimension, the data is timestamped and standardized to unify the collection time granularity, such as sampling at fixed time intervals. Ultimately, a marketing behavior dataset with precise spatial location and time stamps can be generated, achieving a digital representation of user behavior and providing the industry map with spatiotemporal dynamic information of offline scenarios.

[0033] Furthermore, semantically parsed marketing text datasets can be heterogeneously integrated with spatiotemporally aligned marketing behavior datasets to form unstructured marketing datasets. This integration process involves data cleaning, format standardization, and metadata annotation, addressing the format consistency issues faced by unstructured data from different sources. Entity alignment technology is used to associate semantic entities in text data with physical entities in spatiotemporal data. For example, by matching store names mentioned in social media with customer traffic data collected by IoT devices, a cross-modal entity association chain is formed, enhancing the integration value of unstructured data.

[0034] Finally, business data from the ERP (Enterprise Resource Planning) system, including typical structured data such as order records, inventory information, and sales reports, can be synchronized and deeply integrated with unstructured marketing datasets. Multi-source data mapping technology is used in the integration process to establish correspondences between business entities in the structured data and semantic entities and spatiotemporal entities in the unstructured data, constructing a four-dimensional data model of entities, relationships, spatiotemporal space, and attributes. The multimodal marketing scenario dataset formed through the integration of this model encompasses multidimensional information such as textual semantics, physical spatiotemporal space, and business statistics, breaking through the dimensional limitations of traditional single-structured data and providing a comprehensive data foundation for scenario-based industry maps.

[0035] The Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology are used to preprocess the multimodal marketing scenario dataset and build a four-dimensional dynamic data lake, specifically including: Using the Apache Flink stream-batch integrated processing framework to perform real-time event processing on the unstructured marketing dataset and simultaneously perform micro-batch processing on the structured marketing dataset; Adopting the knowledge graph entity disambiguation technology, constructing an industry chain entity dictionary based on the knowledge graph, and using a deep semantic matching algorithm to perform entity disambiguation on the unstructured marketing dataset and the structured marketing data based on the industry chain entity dictionary and establish a unified entity identifier; The multimodal marketing scenario dataset is mapped to the four-dimensional dynamic data lake based on a distributed storage architecture.

[0036] Based on the Apache Flink stream-batch processing framework, a real-time event-driven processing model can be used for unstructured marketing data. For real-time data streams such as social media sentiment and IoT device sensor data, continuous processing based on event time semantics ensures data timeliness. For structured marketing data, micro-batch processing can be implemented, dividing calculations into sub-batches at fixed time intervals, thereby balancing processing efficiency and resource consumption. This architecture achieves a unified processing paradigm for data of different modalities, solving the problem of collaborative computing of real-time and batch data.

[0037] Then, based on the knowledge graph, an industry chain entity dictionary can be constructed, integrating entity representations from multiple data sources to form a standardized entity knowledge base. Furthermore, deep semantic matching algorithms can be used to semantically align semantic entities in unstructured data with business entities in structured data. Semantic entities, such as company aliases mentioned in social media text, and business entities, such as registered company names in ERP systems, can be used. Furthermore, a globally unique unified entity identifier can be assigned to each entity, establishing entity association links across modal data to ensure consistency and traceability across data from different sources.

[0038] Furthermore, based on a distributed storage architecture, the pre-processed multimodal marketing scenario dataset can be mapped to a dynamic data lake containing four dimensions: entity, relationship, time, and space. The entity dimension integrates various entities in the industry chain, the relationship dimension depicts the business connections between entities, the time dimension records the timing of data generation, and the space dimension annotates physical location attributes. Through the four-dimensional dynamic data lake, centralized data storage and management can be achieved, ensuring incremental data updates and version control, thereby providing standardized, real-time data support for scenario-based industry maps, significantly improving data integration efficiency and analysis quality.

[0039] The method of obtaining a historical marketing scenario dataset and constructing a three-dimensional marketing scenario feature library, and training the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model specifically includes: Acquire the historical marketing scene dataset, extract temporal marketing features, spatial marketing features, and behavioral marketing features from the historical marketing scene dataset, and construct the three-dimensional marketing scene feature library; The LSTM neural network is used to train the three-dimensional marketing scene feature library, match the marketing weight vectors corresponding to the marketing scene features in the three-dimensional marketing scene feature library, and generate the marketing scene weight model.

[0040] This involves collecting the full amount of data from historical marketing campaigns, encompassing business data, user behavior data, and market feedback data from various marketing scenarios, to form a historical marketing scenario dataset. This dataset can then be systematically extracted from three core dimensional features: temporal features such as the marketing campaign cycle and holiday attributes can be extracted from the temporal dimension; geographic attributes such as regional consumer power index and business district heat distribution can be mined from the spatial dimension; and behavioral features such as user conversion rate, interaction frequency, and consumer preferences can be analyzed from the behavioral dimension. These multi-dimensional features are structured and integrated to construct a three-dimensional marketing scenario feature library encompassing temporal, spatial, and behavioral elements, providing standardized data samples for subsequent model training.

[0041] Furthermore, the LSTM neural network can be used as the core training algorithm, leveraging its unique gating mechanism to effectively process long-sequence dependencies in historical marketing data and avoid the vanishing gradient problem of traditional recurrent neural networks. A three-dimensional marketing scenario feature library is input into the LSTM neural network as training samples. The network parameters are optimized using a backpropagation algorithm to learn the mapping relationship between different marketing scenario features and data weights. During model training, a corresponding marketing weight vector is matched to each scenario feature combination. This vector contains weight coefficients for dimensions such as supply and demand relationships, competitive landscape, and public opinion heat, and is used to quantify the importance of each data dimension in a specific marketing scenario.

[0042] After training with an LSTM neural network, a decision model, known as a marketing scenario weight model, can be formed that automatically identifies marketing scenario features and outputs corresponding weight vectors. This model exhibits scenario-adaptive capabilities, rapidly activating corresponding weighting strategies based on the real-time marketing scenario type. This model dynamically adjusts the computational weights of various data dimensions in industry graph analysis, such as increasing the weight of geofence data in regional promotion scenarios and strengthening the weight of public opinion data in new product launches. This allows marketing decisions to be precisely adapted to market dynamics. This model overcomes the limitations of traditional fixed-weight algorithms and significantly enhances the scenario sensitivity and decision effectiveness of industry graph analysis.

[0043] The four-dimensional dynamic data lake is mapped to the Neo4j graph database to form an industrial chain topology structure. The node relationship strength in the industrial chain topology structure is adjusted based on the marketing scenario weight model. The graph neural network is introduced and the implicit features of the nodes in the industrial chain topology structure are learned through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure. Specifically, the following steps are involved: Mapping the four-dimensional dynamic data lake to the Neo4j graph database to construct the industrial chain topology structure, wherein the enterprise ID, product ID, and user ID in the four-dimensional dynamic data lake are mapped as nodes in the industrial chain topology structure, the supply and demand relationship and public opinion association in the four-dimensional dynamic data lake are mapped as edges in the industrial chain topology structure, and the spatiotemporal attributes in the four-dimensional dynamic data lake are used as labels for nodes or edges in the industrial chain topology structure; Adjust edge weights, i.e., node relationship strengths, in real time based on the marketing scenario weight model; The graph neural network is introduced, and the GraphSAGE algorithm is used to learn the embedded representation of the nodes. The features of adjacent nodes are weighted and aggregated in combination with the strength of the node relationships to obtain the implicit features of the nodes. The industrial chain topology is recalculated based on the node relationship strength and the node implicit features, and the key node clusters in the updated industrial chain topology are identified through a community discovery algorithm to generate the scenario-sensitive industrial graph topology.

[0044] In practical applications, the multimodal data in the four-dimensional dynamic data lake can be mapped to the Neo4j graph database to construct a visual industrial chain topology. In specific implementation, the enterprise identity, product identity, and user identity in the four-dimensional dynamic data lake are used as the basic nodes of the industrial chain topology structure, representing the subject, object, and actor in the industrial chain respectively; the business connections such as supply and demand relationships and public opinion associations in the four-dimensional dynamic data lake are mapped as edges of the industrial chain topology structure to depict the business links and semantic associations between entities; at the same time, the time attributes and spatial attributes in the industrial chain topology structure are used as node or edge labels to embed the spatiotemporal dimension in the industrial chain topology structure. This mapping process constructs a three-dimensional graph model containing entity nodes, relationship edges, and spatiotemporal attributes, providing a structured visualization carrier for industrial chain analysis.

[0045] Then, based on the marketing scenario weight model, the edge weights in the industry chain topology can be calibrated in real time. This model is trained based on historical marketing data and can automatically generate corresponding weight vectors based on the current marketing scenario type. During the weight adjustment process, the weight values ​​of specific types of edges are dynamically increased or decreased based on different scenario characteristics, such as time period, geographical scope, user behavior, etc. For example, in a holiday promotion scenario, the weight of the interaction edge between store nodes and user nodes is strengthened, and the weight of the cross-regional supply chain edge is weakened, so that the edge weight distribution of the industry chain topology can accurately reflect the business correlation strength under the requirements of the marketing scenario, and achieve preliminary adaptation of the topology structure to the marketing scenario.

[0046] Furthermore, graph neural networks can be introduced, and the GraphSAGE algorithm can be used to learn the embedded representations of nodes in the topological structure of the industrial chain. The algorithm iterates through multiple layers of neural networks, performs weighted aggregation of the features of adjacent nodes based on the strength of node relationships, and at the same time integrates the initial features of the nodes themselves, such as enterprise scale and product evaluation, to generate node implicit features containing deep semantic information. Specifically, in each layer of the network, the algorithm weights the features of neighboring nodes according to edge weights, extracts common features through aggregation operations such as pooling and mean calculation, and then generates higher-level abstract representations through nonlinear transformations. This process enables node features to capture complex associations in the industrial chain. For example, the implicit features of store nodes can integrate multi-dimensional information such as their geographical location, surrounding customer flow, and user evaluations to form a scene-sensitive node semantic representation.

[0047] Finally, the topology of the industrial chain can be recalculated and optimized based on the adjusted edge weights and the learned implicit features of the nodes. Specifically, the node implicit features and edge weights are first fused through the similarity function to measure the semantic association strength and scene adaptation between nodes, so as to reconstruct the node connection relationship. Secondly, the community discovery algorithm is used to modularize the new topology, and the node communities with high cohesion and low coupling are identified by maximizing the modularity index. Finally, based on the weighted degree and average similarity of the nodes in the community, the importance score of each community is calculated, and the communities with high scores are selected as key node clusters to form a scene-sensitive industry map topology structure. This topology can highlight the core business links and high-value touchpoints in the current marketing scenario, such as promotion core business districts, high public opinion product clusters, etc., and provide visual and accurate guidance for marketing decisions.

[0048] The real-time adjustment of edge weights based on the marketing scenario weight model specifically includes: Using the FastText text classification model to perform keyword extraction and semantic classification on the current marketing scenario description text in the four-dimensional dynamic data lake, determine the corresponding current marketing scenario type, and extract the corresponding current marketing scenario features; Based on the marketing scenario weight model, the current marketing weight vector corresponding to the current marketing scenario feature, that is, the edge weight, is matched.

[0049] The FastText text classification model can be used to perform semantic analysis on the current marketing scenario description text in the four-dimensional dynamic data lake. This model automatically identifies the current marketing scenario type by extracting keywords and semantically encoding the scenario description text, matching it with predefined scenario type labels. Based on the determined scenario type, it further extracts scenario features in dimensions such as time, space, and behavior from the text to form a structured feature vector of the current marketing scenario.

[0050] The extracted features of the current marketing scenario can then be input into the marketing scenario weight model. Through matrix operations and nonlinear transformations, the corresponding current marketing weight vector is matched. This weight vector includes weight coefficients for multiple data dimensions, such as supply and demand relationships, competitive landscape, and geographic distribution, and is used to quantify the importance of each dimension in the current scenario. For example, in a regional promotion scenario, the marketing scenario weight model will automatically increase the weight coefficient of the geographic distribution dimension and reduce the weight of irrelevant dimensions, so that the weight vector can accurately reflect the scenario requirements.

[0051] Finally, the matched marketing weight vector can be applied to the edge weight adjustment of the industry chain graph. For each edge in the graph, the weight value of the edge is dynamically adjusted according to the data dimension involved and the corresponding coefficient in the weight vector. Specifically, for edges that are highly relevant to the current scenario, their weight values ​​are strengthened according to the weight coefficient of the corresponding dimension; for edges that are not relevant to the scenario, the weight values ​​are maintained or weakened. Through this mechanism, the edge weight distribution of the industry chain graph can adapt to changes in marketing scenarios in real time, so that the graph topology structure more accurately reflects the strength of business associations in the current scenario, providing dynamic data support for subsequent graph analysis and marketing decisions.

[0052] Define the entity set in the four-dimensional dynamic data lake as ,in Represents the i-th entity, i.e., enterprise ID, product ID, or user ID, n represents the total number of entities in the four-dimensional dynamic data lake, and maps the entity set E to the node set in the industry chain topology structure G. ,and , and map() represents the mapping function from entity to node; the edge set in the industrial chain topology structure G is defined as ,in Represents the i-th node and the jth node The relationship type between , defines the spatiotemporal attribute set in the four-dimensional dynamic data lake as ,in Represents the time attribute, Represents spatial attributes; The calculation formula of the node implicit feature is as follows: Where, Representation node Hidden features of nodes in the lth layer of the graph neural network; represents the activation function; represents the feature aggregation function of the lth layer; Representation node Hidden features of nodes in the l-1 layer of the graph neural network; Represents the adjusted edge weight, i.e., node and The strength of the node relationships between them; Representation node The set of neighbor nodes of Representation node Hidden features of nodes in the l-1th layer of the graph neural network.

[0053] The recalculating the industry chain topology structure according to the node relationship strength and the node implicit features specifically includes: The similarity function is constructed based on the node relationship strength and the node implicit features as follows: Where, Representation node and nodes Similarity function; L represents the last layer of the graph neural network; Representation node Hidden features of nodes in the last layer of the graph neural network; Representation node Hidden features of nodes in the last layer of the graph neural network; Represents node implicit features The L2 norm of Represents node implicit features The L2 norm of Representation node Node implicit features With node Node implicit features The cosine similarity of The empirical coefficient representing the amplification factor of the node relationship strength, It represents the amplification factor of the node relationship strength, ; Setting similarity threshold ,when When retaining the edge , otherwise delete the edge , recalculate the industrial chain topology structure to obtain the updated industrial chain topology structure ; The method of identifying key node clusters in the updated industry chain topology structure by using a community discovery algorithm to generate the scenario-sensitive industry graph topology structure specifically includes: The updated industry chain topology Each node in the network is considered as a community, forming ,in represents the i-th community; The community discovery algorithm is used to calculate the community and neighboring communities The community modularity is and , and calculate the community With neighboring communities The community modularity of the new community after the merger is , and the community modularity gain is ; If the community modularity gain , then merge the communities and neighboring communities , otherwise keep the community and neighboring communities independent; Traverse all adjacent community pairs to obtain the final community set , where p represents the final community set Total number of communities in Calculate the final community set Each community Community importance score ,in , importance scores for all communities Arrange them from large to small, select the communities corresponding to the top M community importance scores to form a key community cluster, that is, the key node cluster, and generate the scenario-sensitive industry map topology structure.

[0054] At the data modeling level, the entities, relationships, and spatiotemporal attributes in the four-dimensional dynamic data lake can be mapped to nodes, edges, and labels of the graph database to build an industrial chain topology structure that includes entities such as enterprises, products, and users. This enables a unified representation of multimodal data using a graph model, solves the problem of traditional data fragmentation, and visualizes the relationships in the industrial chain.

[0055] At the feature learning level, the GraphSAGE algorithm of the graph neural network can be used to perform weighted aggregation on node features, and the implicit features of the nodes can be calculated in combination with the edge weights after scenario adjustment, so as to capture the deep semantic associations of nodes in the industrial chain. This is more in line with the needs of marketing scenarios than traditional feature extraction methods and significantly improves the feature representation capabilities.

[0056] At the topology optimization level, the industry chain topology is reconstructed by integrating similarity functions between implicit node features and edge weights. This automatically filters out scenario-irrelevant connections while retaining highly relevant edges, dynamically adapting the graph topology to the current marketing scenario. Combined with community discovery algorithms to identify key node clusters, this allows for the precise location of high-value marketing touchpoints and the generation of a scenario-sensitive graph. Field tests have shown a significant improvement in touchpoint identification accuracy, effectively supporting the targeted allocation of resources for scenario-based marketing.

[0057] A method for constructing an industry map based on scenario-based marketing, further comprising: providing a three-dimensional visualization engine based on WebGL to support a user end to perform spatiotemporal slicing of the scenario-sensitive industry map topology structure through geo-fencing and timestamps; The improved PageRank algorithm is used to calculate the importance of nodes in the scenario-sensitive industry graph topology structure, and the DBSCAN clustering algorithm is used to identify key marketing touchpoints based on the importance of the nodes and generate marketing decision recommendations, and the key marketing touchpoints are marked in the form of a heat map; An interactive analysis interface is provided to support the user terminal to click on the key marketing touchpoints and view the marketing decision suggestions.

[0058] First, a three-dimensional visualization engine can be developed based on WebGL technology. This engine has powerful graphics rendering capabilities and supports users to define spatial scope through geo-fencing tools, filter specific time periods with timestamp controls, and perform spatiotemporal slice analysis on the topological structure of the industrial map. This allows for dynamic filtering of industrial chain data, allowing users to focus on changes in business relationships within specific areas and time periods, such as viewing the customer flow and sales links of a certain business district during holiday promotions.

[0059] Secondly, we can use a modified PageRank algorithm to assess the importance of nodes in the graph. We can also use the DBSCAN density clustering algorithm to cluster graph nodes based on their importance scores, automatically identifying high-density node areas as key marketing touchpoints, such as high-potential promotional business districts or popular product clusters. These touchpoints are intuitively annotated on the graph in the form of a heat map, with the color depth reflecting the degree of node clustering and commercial value.

[0060] Finally, an interactive analysis interface can be set up, allowing users to click on key marketing touchpoints in the visualization interface to instantly access marketing decision recommendations generated based on the graph topology and clustering results. These recommendations cover resource allocation strategies, competitive landscape analysis, and other content, completing a closed loop from data visualization to intelligent decision-making, significantly enhancing the application value and decision-making efficiency of industry graphs in scenario-based marketing.

[0061] like Figure 2 FIG. 1 is a system block diagram of an industry graph construction system based on scenario-based marketing provided by an embodiment of the present invention. The construction system includes: The data collection module is used to obtain social media text through distributed crawlers, generate a marketing text dataset with semantic labels after processing it with the BERT pre-trained model, and simultaneously collect offline scene data through IoT devices. After processing it with the spatiotemporal calibration algorithm, it generates a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. This is combined with the structured marketing dataset to generate a multimodal marketing scenario dataset. A data processing module, configured to pre-process the multimodal marketing scenario dataset and construct a four-dimensional dynamic data lake using the Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology; A model building module is used to obtain a historical marketing scenario dataset and build a three-dimensional marketing scenario feature library, and train the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model; A graph construction module is used to map the four-dimensional dynamic data lake to the Neo4j graph database to form an industrial chain topology structure, adjust the node relationship strength in the industrial chain topology structure based on the marketing scenario weight model, introduce a graph neural network and learn the implicit features of the nodes in the industrial chain topology structure through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure.

[0062] Figure 2 The apparatus of the embodiment shown can be used to perform Figure 1 The implementation principles and technical effects of the steps in the method embodiment shown are similar and will not be repeated here.

[0063] An electronic device includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the steps of a method for constructing an industry map based on scenario-based marketing as described in any one of the above items.

[0064] like Figure 3 FIG. 1 is a schematic diagram of the hardware structure of an electronic device provided by an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32 and a computer program; The memory 32 is used to store the computer program, which may also be a flash memory. The computer program is, for example, an application program or a functional module for implementing the above method.

[0065] The processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant description in the above method embodiment.

[0066] Optionally, the memory 32 may be independent or integrated with the processor 31 .

[0067] When the memory 32 is a device independent of the processor 31, the device may further include: The bus 33 is used to connect the memory 32 and the processor 31 .

[0068] A readable storage medium having a computer program stored therein, wherein the computer program, when executed by a processor, is used to implement the steps of a method for constructing an industry map based on scenario-based marketing as described in any one of the above.

[0069] The readable storage medium may be a computer storage medium or a communication medium. Communication media include any medium that facilitates the transfer of computer programs from one location to another. Computer storage media may be any available medium that can be accessed by a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium may also be an integral part of the processor. The processor and the readable storage medium may be located in an application-specific integrated circuit (ASIC). In addition, the ASIC may be located in a user device. Of course, the processor and the readable storage medium may also exist as discrete components in a communication device. The readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0070] The present invention also provides a program product, which includes execution instructions stored in a readable storage medium. At least one processor of a device can read the execution instructions from the readable storage medium, and at least one processor executes the execution instructions so that the device implements the methods provided in the various embodiments described above.

[0071] In the embodiments of the above-mentioned devices, it should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), etc. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly executed by a hardware processor or by a combination of hardware and software modules within the processor.

[0072] Through the introduction of the above embodiments, the present invention uses an industry map construction method and system based on scenario-based marketing, obtains social media text through distributed crawlers, generates a marketing text dataset with semantic labels after processing by the BERT pre-training model, and collects offline scene data through IoT devices. After processing by the spatiotemporal calibration algorithm, a marketing behavior dataset with spatiotemporal labels is generated. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset, and combined with the structured marketing dataset to generate a multimodal marketing scenario dataset; the Apache Flink stream-batch integrated processing framework and the knowledge graph entity disambiguation technology are used to preprocess and construct the multimodal marketing scenario dataset. Build a four-dimensional dynamic data lake; obtain historical marketing scenario data sets and construct a three-dimensional marketing scenario feature library, and obtain a marketing scenario weight model by training the three-dimensional marketing scenario feature library through the LSTM neural network; map the four-dimensional dynamic data lake to the Neo4j graph database to form an industrial chain topology structure, adjust the node relationship strength in the industrial chain topology structure based on the marketing scenario weight model, introduce a graph neural network and learn the implicit features of the nodes in the industrial chain topology structure through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure, so as to construct a scenario-sensitive industrial graph to reflect the relationship between industrial entities in different marketing scenarios in real time, thereby improving the effectiveness and accuracy of the industrial graph in marketing decision-making.

[0073] The present invention generates a multimodal marketing scenario dataset by integrating the semantic data of social media text and the spatiotemporal behavior data collected by IoT devices, combined with the structured data of the ERP system, so that the construction of the industry map is based on a more comprehensive and rich data foundation; the method of the present invention utilizes the Apache Flink stream-batch integrated processing framework to realize real-time processing of unstructured data and micro-batch processing of structured data, ensuring the timeliness and efficiency of data processing; the present invention constructs a three-dimensional marketing scenario feature library and uses LSTM neural network training to obtain a marketing scenario weight model, so that the node relationship strength in the industry chain topology structure can be adjusted in real time according to different marketing scenarios, making the industry map scenario-sensitive and able to more accurately reflect the association relationship of industry entities in different marketing scenarios; the present invention combines graph neural networks and GraphSAGE algorithms to learn node implicit features, optimize the industry chain topology structure through similarity functions and community discovery algorithms, and generate a more accurate scenario-sensitive industry map; in addition, the present invention provides users with intuitive and effective marketing decision support through a three-dimensional visualization engine, node importance calculation, key marketing touchpoint identification and interactive analysis interface, thereby improving the scientificity and accuracy of marketing decisions.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing an industry map based on scenario-based marketing, characterized in that: The construction method comprises: Social media text is obtained through distributed crawlers and processed by the BERT pre-trained model to generate a marketing text dataset with semantic labels. At the same time, offline scene data is collected through IoT devices and processed by the spatiotemporal calibration algorithm to generate a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. Combined with the structured marketing dataset, a multimodal marketing scenario dataset is generated. The Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology are used to preprocess the multimodal marketing scenario dataset and build a four-dimensional dynamic data lake; Obtain a historical marketing scenario dataset and construct a three-dimensional marketing scenario feature library, and train the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model; The four-dimensional dynamic data lake is mapped to the Neo4j graph database to form an industrial chain topology structure. The node relationship strength in the industrial chain topology structure is adjusted based on the marketing scenario weight model. The graph neural network is introduced and the implicit features of the nodes in the industrial chain topology structure are learned through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure.

2. The method for constructing an industry map based on scenario marketing according to claim 1, characterized in that: The method involves acquiring social media text through a distributed crawler and processing it with a BERT pre-trained model to generate a marketing text dataset with semantic labels. Meanwhile, offline scene data is collected through IoT devices and processed with a spatiotemporal calibration algorithm to generate a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. This dataset is then combined with the structured marketing dataset to generate a multimodal marketing scenario dataset, specifically including: The social media text is acquired in real time by the distributed crawler cluster, and the BERT pre-trained model is used to semantically encode the social media text. The industry chain entities in the social media text are extracted by named entity recognition technology, and the association relationships between the industry chain entities are constructed by using dependency syntax analysis. The industry chain entities and the association relationships are summarized to generate the marketing text dataset with semantic labels; Synchronously using an edge computing gateway to connect to the IoT device to collect the offline scene data, and performing geo-fence matching and timestamp calibration on the offline scene data using the spatiotemporal calibration algorithm, that is, mapping the offline scene data to spatial entity coordinates and marking the collection time granularity of the offline scene data, thereby forming the marketing behavior dataset with spatiotemporal labels; Aggregating the marketing text dataset with semantic tags and the marketing behavior dataset with spatiotemporal tags to obtain the unstructured marketing dataset, and synchronously accessing the ERP system to obtain the structured marketing dataset; The unstructured marketing dataset and the structured marketing dataset are aggregated to form the multimodal marketing scenario dataset.

3. The method for constructing an industry map based on scenario marketing according to claim 1, characterized in that: The Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology are used to preprocess the multimodal marketing scenario dataset and build a four-dimensional dynamic data lake, specifically including: Using the Apache Flink stream-batch integrated processing framework to perform real-time event processing on the unstructured marketing dataset and simultaneously perform micro-batch processing on the structured marketing dataset; Adopting the knowledge graph entity disambiguation technology, constructing an industry chain entity dictionary based on the knowledge graph, and using a deep semantic matching algorithm to perform entity disambiguation on the unstructured marketing dataset and the structured marketing data based on the industry chain entity dictionary and establish a unified entity identifier; The multimodal marketing scenario dataset is mapped to the four-dimensional dynamic data lake based on a distributed storage architecture.

4. The method for constructing an industry map based on scenario marketing according to claim 1, characterized in that: The method of obtaining a historical marketing scenario dataset and constructing a three-dimensional marketing scenario feature library, and training the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model specifically includes: Acquire the historical marketing scene dataset, extract temporal marketing features, spatial marketing features, and behavioral marketing features from the historical marketing scene dataset, and construct the three-dimensional marketing scene feature library; The LSTM neural network is used to train the three-dimensional marketing scene feature library, match the marketing weight vectors corresponding to the marketing scene features in the three-dimensional marketing scene feature library, and generate the marketing scene weight model.

5. The method for constructing an industry map based on scenario marketing according to claim 1, characterized in that: The four-dimensional dynamic data lake is mapped to the Neo4j graph database to form an industrial chain topology structure. The node relationship strength in the industrial chain topology structure is adjusted based on the marketing scenario weight model. The graph neural network is introduced and the implicit features of the nodes in the industrial chain topology structure are learned through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure. Specifically, the following steps are involved: Mapping the four-dimensional dynamic data lake to the Neo4j graph database to construct the industrial chain topology structure, wherein the enterprise ID, product ID, and user ID in the four-dimensional dynamic data lake are mapped as nodes in the industrial chain topology structure, the supply and demand relationship and public opinion association in the four-dimensional dynamic data lake are mapped as edges in the industrial chain topology structure, and the spatiotemporal attributes in the four-dimensional dynamic data lake are used as labels for nodes or edges in the industrial chain topology structure; Adjust edge weights, i.e., node relationship strengths, in real time based on the marketing scenario weight model; The graph neural network is introduced, and the GraphSAGE algorithm is used to learn the embedded representation of the nodes. The features of adjacent nodes are weighted and aggregated in combination with the strength of the node relationships to obtain the implicit features of the nodes. The industrial chain topology is recalculated based on the node relationship strength and the node implicit features, and the key node clusters in the updated industrial chain topology are identified through a community discovery algorithm to generate the scenario-sensitive industrial graph topology.

6. The method for constructing an industry map based on scenario-based marketing according to claim 5, characterized in that: The real-time adjustment of edge weights based on the marketing scenario weight model specifically includes: Using the FastText text classification model to perform keyword extraction and semantic classification on the current marketing scenario description text in the four-dimensional dynamic data lake, determine the corresponding current marketing scenario type, and extract the corresponding current marketing scenario features; Based on the marketing scenario weight model, the current marketing weight vector corresponding to the current marketing scenario feature, that is, the edge weight, is matched.

7. The method for constructing an industry map based on scenario marketing according to claim 5, characterized in that: Define the entity set in the four-dimensional dynamic data lake as ,in Represents the i-th entity, i.e., enterprise ID, product ID, or user ID, n represents the total number of entities in the four-dimensional dynamic data lake, and maps the entity set E to the node set in the industry chain topology structure G. ,and , and map() represents the mapping function from entity to node; the edge set in the industrial chain topology structure G is defined as ,in Represents the i-th node and the jth node The relationship type between , defines the spatiotemporal attribute set in the four-dimensional dynamic data lake as ,in Represents the time attribute, Represents spatial attributes; The calculation formula of the node implicit feature is as follows: Where, Representation node Hidden features of nodes in the lth layer of the graph neural network; represents the activation function; represents the feature aggregation function of the lth layer; Representation node Hidden features of nodes in the l-1 layer of the graph neural network; Represents the adjusted edge weight, i.e., node and The strength of the node relationships between them; Representation node The set of neighbor nodes of Representation node Hidden features of nodes in the l-1th layer of the graph neural network.

8. The method for constructing an industry map based on scenario-based marketing according to claim 7, characterized in that: The recalculating the industry chain topology structure according to the node relationship strength and the node implicit features specifically includes: The similarity function is constructed based on the node relationship strength and the node implicit features as follows: Where, Representation node and nodes Similarity function; L represents the last layer of the graph neural network; Representation node Hidden features of nodes in the last layer of the graph neural network; Representation node Hidden features of nodes in the last layer of the graph neural network; Represents node implicit features The L2 norm of Represents node implicit features The L2 norm of Representation node Node implicit features With node Node implicit features The cosine similarity of The empirical coefficient representing the amplification factor of the node relationship strength, It represents the amplification factor of the node relationship strength, ; Setting similarity threshold ,when When retaining the edge , otherwise delete the edge , recalculate the industrial chain topology structure to obtain the updated industrial chain topology structure ; The method of identifying key node clusters in the updated industry chain topology structure by using a community discovery algorithm to generate the scenario-sensitive industry graph topology structure specifically includes: The updated industry chain topology Each node in the network is considered as a community, forming ,in represents the i-th community; The community discovery algorithm is used to calculate the community and neighboring communities The community modularity is and , and calculate the community With neighboring communities The community modularity of the new community after the merger is , and the community modularity gain is ; If the community modularity gain , then merge the communities and neighboring communities , otherwise keep the community and neighboring communities independent; Traverse all adjacent community pairs to obtain the final community set , where p represents the final community set Total number of communities in Calculate the final community set Each community Community importance score ,in , importance scores for all communities Arrange them from large to small, select the communities corresponding to the top M community importance scores to form a key community cluster, that is, the key node cluster, and generate the scenario-sensitive industry map topology structure.

9. The method for constructing an industry map based on scenario marketing according to claim 1, characterized in that: It also includes setting up a WebGL-based 3D visualization engine to support the user end to perform spatiotemporal slicing of the scene-sensitive industry map topology structure through geo-fencing and timestamps; The improved PageRank algorithm is used to calculate the importance of nodes in the scenario-sensitive industry graph topology structure, and the DBSCAN clustering algorithm is used to identify key marketing touchpoints based on the importance of the nodes and generate marketing decision recommendations, and the key marketing touchpoints are marked in the form of a heat map; An interactive analysis interface is provided to support the user terminal to click on the key marketing touchpoints and view the marketing decision suggestions.

10. An industry map construction system based on scenario-based marketing, applied to an industry map construction method based on scenario-based marketing as described in any one of claims 1 to 9, characterized in that: The construction system includes: The data collection module is used to obtain social media text through distributed crawlers, generate a marketing text dataset with semantic labels after processing it with the BERT pre-trained model, and simultaneously collect offline scene data through IoT devices. After processing it with the spatiotemporal calibration algorithm, it generates a marketing behavior dataset with spatiotemporal labels. The marketing text dataset with semantic labels and the marketing behavior dataset with spatiotemporal labels are aggregated to obtain an unstructured marketing dataset. This is combined with the structured marketing dataset to generate a multimodal marketing scenario dataset. A data processing module, configured to pre-process the multimodal marketing scenario dataset and construct a four-dimensional dynamic data lake using the Apache Flink stream-batch integrated processing framework and knowledge graph entity disambiguation technology; A model building module is used to obtain a historical marketing scenario dataset and build a three-dimensional marketing scenario feature library, and train the three-dimensional marketing scenario feature library through an LSTM neural network to obtain a marketing scenario weight model; A graph construction module is used to map the four-dimensional dynamic data lake to the Neo4j graph database to form an industrial chain topology structure, adjust the node relationship strength in the industrial chain topology structure based on the marketing scenario weight model, introduce a graph neural network and learn the implicit features of the nodes in the industrial chain topology structure through the GraphSAGE algorithm to generate a scenario-sensitive industrial graph topology structure.

Images