Intelligent e-commerce behavior event decision-making method and system fused with multi-source perception

Abstract

The invention provides an intelligent e-commerce behavior event decision-making method and system fusing multi-source perception, and the method comprises the steps: obtaining a user behavior data set, a commodity attribute data set and an environment context data set through a preset data collection interface, and carrying out the intention recognition processing of multi-source data; generating an intention feature set containing preliminary intention features and refined intention features, then generating an interactive guidance strategy tree containing node branch weights and strategy execution priorities based on the intention feature set, and performing dynamic path adjustment processing on the interactive guidance strategy tree according to a real-time feedback data set to obtain an optimized strategy tree set; and finally, pushing the optimization strategy tree set to a target interaction interface to activate an interaction guide operation, thereby providing personalized and intelligent interaction guide for the user by fusing multi-source sensing data, deeply understanding the intention of the user and dynamically adjusting a decision strategy, and improving the operation efficiency of an e-commerce platform and the satisfaction of the user.

Description

Technical Field

[0001] This invention relates to the field of e-commerce and internet technology, and more specifically, to an intelligent e-commerce behavior event decision-making method and system that integrates multi-source perception. Background Technology

[0002] In the e-commerce sector, with the continuous expansion of business scale and the increasing diversification of user needs, accurately understanding user intent and providing personalized interactive guidance has become crucial for improving user experience and promoting transaction conversion. However, existing technologies, when making decisions regarding e-commerce behavioral events, often rely solely on a single data source, such as focusing only on user behavior data or product attribute data, while ignoring the significant impact of environmental context data on user decisions. This single-data-source decision-making approach cannot comprehensively and accurately capture the user's true intent, resulting in a lack of targeting and effectiveness in the generated interactive guidance strategies. Furthermore, existing technologies lack detailed analysis of user intent, making it difficult to distinguish between different levels of user intent needs and providing precise interactive guidance. Moreover, in the decision-making process, existing technologies typically employ static decision-making strategies, unable to dynamically adjust based on real-time feedback data. This makes them ill-suited to the complex and ever-changing e-commerce environment, failing to meet the constantly evolving needs of users and limiting the operational efficiency and user satisfaction of e-commerce platforms. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an intelligent e-commerce behavioral event decision-making method integrating multi-source perception, the method comprising:

[0004] The user behavior data set, product attribute data set, and environmental context data set are obtained through a preset data collection interface. The user behavior data set includes the sequence of operation events triggered by the user in the interactive interface and the corresponding page jump path.

[0005] The user behavior data set, the product attribute data set, and the environmental context data set are subjected to intent recognition processing to generate an intent feature set, which includes preliminary intent features and refined intent features.

[0006] An interaction guidance strategy tree is generated based on the intent feature set. The interaction guidance strategy tree includes node branch weights and strategy execution priorities.

[0007] Based on the real-time feedback data set, the interactive guidance strategy tree is dynamically adjusted to obtain an optimized strategy tree set, which includes at least one updated node branch weight and strategy execution priority.

[0008] The optimized strategy tree set is pushed to the target interactive interface to activate the interactive guidance operation.

[0009] In another aspect, embodiments of the present invention also provide an intelligent e-commerce behavior event decision-making system that integrates multi-source perception, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code. The processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-mentioned method.

[0010] Based on the above, this invention, by integrating multi-source perception data, including user behavior data, product attribute data, and environmental context data, can comprehensively and accurately capture various user behaviors and needs in e-commerce scenarios. It performs intent recognition processing on the multi-source data to generate an intent feature set containing preliminary and refined intent features, which helps to deeply understand different levels and details of user intent. The interaction guidance strategy tree generated based on the intent feature set includes node branch weights and strategy execution priorities, making the interaction guidance strategy more scientific and reasonable. It can guide users in an orderly manner according to the importance and urgency of their intent. The interaction guidance strategy tree is dynamically adjusted based on real-time feedback data to obtain an optimized strategy tree set, achieving real-time optimization of decision-making strategies. This allows for rapid adaptation to changes in the e-commerce environment and dynamic adjustments to user needs, effectively improving the accuracy and timeliness of interaction guidance. Finally, the optimized strategy tree set is pushed to the target interaction interface to activate the interaction guidance operation, providing users with a more personalized and intelligent interactive experience and significantly improving user satisfaction and transaction conversion rates on e-commerce platforms. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the execution flow of the intelligent e-commerce behavior event decision-making method that integrates multi-source perception provided in an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of exemplary hardware and software components of the intelligent e-commerce behavior event decision-making system that integrates multi-source perception provided in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an intelligent e-commerce behavior event decision-making method integrating multi-source perception, provided in one embodiment of the present invention. The following is a detailed description of this intelligent e-commerce behavior event decision-making method integrating multi-source perception.

[0014] Step S110: Obtain user behavior data set, product attribute data set and environmental context data set through a preset data acquisition interface, wherein the user behavior data set includes the sequence of operation events triggered by the user on the interactive interface and the corresponding page jump path.

[0015] In the operational architecture of an e-commerce system, a pre-defined data acquisition interface is a crucial data acquisition channel. This interface is a program module developed according to predefined data protocols and interface specifications. It connects to the e-commerce platform's data storage layer, user interaction layer, and device information monitoring layer, thereby enabling the acquisition of data from various sources.

[0016] For user behavior data sets, e-commerce platforms can deploy event listening mechanisms in the background for their user interfaces. When a user interacts with the interface, such as clicking a product image, adding it to the cart, or submitting an order, these actions are encapsulated as operation event objects. Each operation event object contains information such as the operation type (e.g., click, swipe), operation time, operation location, and the product or page element involved. These operation events are arranged in chronological order to form an operation event sequence. Simultaneously, the system can also record user navigation between different pages. During a navigation from page A to page B, the identifiers of page A and page B, as well as the navigation time, can be recorded. Multiple such navigation records are combined to form the page navigation path.

[0017] Product attribute data sets are stored in the e-commerce platform's product database. The product database is a structured data storage system; each product record contains multiple attribute fields, such as product name, brand, material, color, size, functional description, and target audience. The data acquisition interface can extract the attribute information of each product sequentially from the product database based on its unique identifier, and organize this information into a product attribute data set.

[0018] The environmental context data set is related to the user's device. The data acquisition interface can interact with the device's operating system and network services, obtaining device network latency parameters through the device's network management module. These parameters reflect the data transmission latency between the device and the e-commerce server. Simultaneously, it obtains device screen adaptation parameters, including screen resolution and aspect ratio, from the device's display settings module. These device status parameters are then integrated to generate the environmental context data set.

[0019] The preset data collection interface must be developed in accordance with strict data protection regulations and privacy policies. Before data collection, the purpose, scope, and usage of the data collection must be clearly explained to the user, and explicit authorization from the user must be obtained, such as authorization to collect data and authorization to analyze the collected data to facilitate interactive guidance in subsequent embodiments. The data collection process will only be initiated with the user's consent.

[0020] For the acquisition of the user behavior data set, a legal and compliant event listening mechanism is deployed on the interactive interface of the e-commerce platform. When a user operates on the interface, operation events need to be recorded without infringing on the user's privacy. For example, operation events only record the interaction behaviors between the user and commodity or page elements, rather than the user's personal sensitive information such as ID card numbers, bank card numbers, etc. The recording of the operation event sequence and the page jump path also follows the principle of minimum necessity, only collecting information related to the user's shopping intention and behavior.

[0021] The acquisition of the commodity attribute data set also follows the principle of data legality. The data in the commodity database is legally provided by merchants and has been audited by the platform to ensure the authenticity and compliance of the data. When collecting commodity attribute information, only the data related to the characteristics of the commodity itself is obtained, without involving the commercial secrets or other sensitive information of the merchants.

[0022] The collection of the environmental context data set also pays attention to the protection of user privacy. When interacting with the user device system to obtain the device network latency parameter and the screen adaptation parameter, encryption transmission technology is adopted to prevent the data from being stolen during the transmission process. At the same time, the system does not obtain other sensitive information of the device, such as personal files stored in the device, contact information, etc.

[0023] Step S120: Perform intention recognition processing on the user behavior data set, the commodity attribute data set, and the environmental context data set to generate an intention feature set, where the intention feature set includes preliminary intention features and refined intention features.

[0024] In order to mine the user's intention from the acquired multi-source data, a series of data processing and model analysis are required. The following describes each processing step in detail.

[0025] Step S121: Perform semantic parsing processing on the operation event sequence in the user behavior data set to generate a semantic association feature set, where the semantic association feature set includes event trigger keywords and event association attributes.

[0026] Semantic parsing processing analyzes the text information in the operation event sequence based on natural language processing technology. First, for the text descriptions involved in the operation events, such as commodity names, button texts, etc., use a word segmentation algorithm to split them into single words. Then, remove the stop words (such as words without actual semantics like "of", "is", "in", etc.). Next, determine the词性 of each word through a词性标注 algorithm, and screen out the words with actual semantics such as nouns and verbs as candidate keywords.

[0027] For these candidate keywords, the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm was used to calculate the importance score for each keyword. The TF-IDF algorithm considers both the frequency of the keyword in the current event text and its frequency in the entire event sequence corpus. Keywords with higher scores were identified as event-triggered keywords.

[0028] In addition to basic information such as operation time and location, event-related attributes can also consider the user's historical operation records. By analyzing the user's operational behavior over a period of time, a user's operation history profile can be constructed. For example, if a user has frequently searched for a certain type of product in the past week, then this search behavior record can be used as a related attribute of the current operation event. Integrating the event-triggered keywords and event-related attributes together generates a semantic association feature set.

[0029] Step S122: Perform time-series analysis on the page jump path to generate a path time-series feature set, which includes path dwell time and path node switching frequency.

[0030] Time series analysis is a method for studying page navigation paths. First, based on the navigation time information recorded in the page navigation path log, the user's dwell time on each page is calculated. For a navigation from page A to page B, the user's dwell time on page A is obtained by subtracting the user's exit time from the entry time of page B.

[0031] To analyze the frequency of path node switching, the page navigation path is divided into certain time windows. For example, a 10-minute time window is used to count the number of page jumps by users within each time window. This number of page jumps is the frequency of path node switching within that time window.

[0032] To enable effective comparison and analysis of path dwell time and path node switching frequency in subsequent processing, they need to be normalized. A min-max normalization method can be used to map the values ​​of path dwell time and path node switching frequency to the interval [0, 1]. Combining the normalized path dwell time and path node switching frequency generates the path temporal feature set.

[0033] Step S123: Extract attributes from the product description text in the product attribute data set to generate an attribute keyword vector set, which includes product functional features and product applicable scenario features.

[0034] Attribute extraction involves extracting keywords that represent product features from the product description text and converting them into vector representations. First, the product description text is preprocessed, including removing punctuation and converting to lowercase letters. Then, a named entity recognition algorithm is used to identify the product's functional descriptions and applicable scenarios within the text.

[0035] For the identified product function descriptions and applicable scenario descriptions, the keywords are converted into vectors using a word embedding model. Word embedding models can be Word2Vec, GloVe, etc. Each keyword is converted into a fixed-length vector, representing the keyword's position in the semantic space.

[0036] The keyword vectors corresponding to product functional features and product applicable scenario features are aggregated separately. Average pooling can be used to sum all keyword vectors under the same feature type and calculate the average, resulting in an aggregated vector for that feature type. Combining the aggregated vectors of product functional features and product applicable scenario features generates a set of attribute keyword vectors.

[0037] Step S124: Perform context encoding processing on the device status parameters in the environment context data set to generate a context environment vector set, which includes device network latency parameters and device screen adaptation parameters.

[0038] Context encoding converts device state parameters into vector representations suitable for model processing. For device network latency parameters, the latency is divided into different ranges, such as low latency, medium latency, and high latency. Each range is assigned a unique encoding value, such as 0 for low latency, 1 for medium latency, and 2 for high latency.

[0039] For device screen adaptation parameters, screen resolution and screen ratio are combined and categorized. For example, common combinations of screen resolution and screen ratio are divided into different categories, and a coding value is assigned to each category.

[0040] The encoded values ​​of the device network latency parameters and the device screen adaptation parameters are concatenated to form a vector. To better integrate this vector with other features in subsequent processing, it is standardized to have a mean of 0 and a standard deviation of 1. The standardized vector is the set of context environment vectors.

[0041] Step S125: Input the semantic association feature set, the path temporal feature set, the attribute keyword vector set, and the context environment vector set into a preset intent recognition combination model to perform intent recognition, and obtain the preliminary intent features and the refined intent features, wherein the preliminary intent features are used to represent the user's basic needs knowledge graph, and the refined intent features are used to represent the user's potential needs preferences.

[0042] The pre-defined intent recognition ensemble model is a deep learning model composed of multiple sub-models and layers. The inputs to this intent recognition ensemble model are a set of semantic association features, a set of path temporal features, a set of attribute keyword vectors, and a set of contextual environment vectors. The outputs are preliminary intent features and refined intent features.

[0043] In the input layer of the intent recognition ensemble model, different feature sets are concatenated to form a unified input vector. This input vector contains information from various aspects, including user behavior, product attributes, and environmental context.

[0044] Step S1251: Call the semantic association model to perform context-related encoding on the event-triggered keywords in the semantic association feature set to generate the first intermediate semantic feature set.

[0045] The semantic association model is a neural network model based on an attention mechanism. When processing event-triggered keywords, the semantic association model needs to consider the contextual relationships between keywords. First, the event-triggered keyword vector is input into the embedding layer of the semantic association model for further feature extraction and transformation.

[0046] Then, in the attention layer, the semantic association model calculates the attention weights between each keyword vector and other keyword vectors. These attention weights represent the degree of association between keywords. Each keyword vector is then fused with other keyword vectors through a weighted summation to obtain a keyword vector that includes contextual information.

[0047] By combining all the keyword vectors containing contextual information, the first intermediate semantic feature set is generated.

[0048] Step S1252: Call the time series correlation model to perform weight allocation processing on the path dwell time in the path time series feature set to generate the first intermediate time series feature set.

[0049] Step S12521: Dynamically divide the path node switching frequency in the path time sequence feature set into multiple time sequence window units based on a preset time interval or event number threshold.

[0050] The preset time interval or event number threshold is a parameter determined based on extensive experiments and data analysis. Assume the preset time interval is T and the event number threshold is N. The path node switching frequency data is read sequentially from the path temporal feature set in chronological order.

[0051] When the time interval reaches T or the number of events reaches N, the path node switching frequency data within this time interval is divided into a time-series window unit. For example, starting from time t1, to time t2 (t2-t1=T) or when the number of events reaches N, the path node switching frequency data during this period is used as a time-series window unit. This process continues until all path node switching frequency data has been processed, generating multiple time-series window units.

[0052] Step S12522: For each time window unit, extract the dwell time distribution parameters and node switching interval parameters within the time window unit.

[0053] For each time-series window unit, the path dwell time data is collected. The mean, variance, and other distribution parameters of the path dwell time are calculated. The mean represents the average time a user spends on the page within that time-series window unit, and the variance represents the dispersion of the dwell time.

[0054] Simultaneously, the node switching interval parameter is calculated. The node switching interval refers to the time interval between two adjacent page transitions. The mean and standard deviation of all node switching intervals within this time-series window unit are calculated. The mean represents the average node switching interval time, and the standard deviation represents the fluctuation of the node switching interval.

[0055] Step S12523: Call the time-series correlation model to perform joint weight calculation on the dwell time distribution parameters and the node switching interval parameters to generate local weight values ​​for each time-series window unit.

[0056] The temporal correlation model is a linear regression-based model. By using the dwell time distribution parameter and the node switching interval parameter as input features, the temporal correlation model learns the relationship between these features and user intent.

[0057] During the model training phase, a large amount of historical data is used for training, and the model's weight parameters are adjusted. In the prediction phase, the dwell time distribution parameters and node switching interval parameters of the current time-series window unit are input into the time-series correlation model. The time-series correlation model calculates a local weight value based on the learned weight parameters. This local weight value represents the importance of the current time-series window unit within the entire path's time-series feature set.

[0058] Step S12524: Perform weighted aggregation processing on the path dwell time within the time window unit according to the local weight value to obtain the first intermediate time series feature set.

[0059] For each time series window unit, the path dwell time within it is multiplied by the corresponding local weight value. Then, the weighted path dwell times from all time series window units are concatenated. This concatenation can be done sequentially according to the order of the time series window units to form the first intermediate time series feature set.

[0060] Step S1253: Call the attribute association model to perform scene mapping processing on the product functional features in the attribute keyword vector set to generate the first intermediate attribute feature set.

[0061] The attribute association model is a knowledge graph-based model. The knowledge graph contains the relationships between product features and applicable scenarios. For example, for the product feature "waterproof," the knowledge graph would record applicable scenarios such as "outdoor sports" and "traveling in rainy weather."

[0062] The product functional feature vectors from the attribute keyword vector set are input into the attribute association model. The attribute association model searches the knowledge graph for applicable scenarios related to each product functional feature. For each product functional feature, the vectors of applicable scenarios are weighted and summed according to the strength of their association with the applicable scenarios.

[0063] By combining the weighted sums of the vectors corresponding to all product functional features, the first set of intermediate attribute features is generated.

[0064] Step S1254: Call the context association model to perform latency impact assessment on the device network latency parameters in the context environment vector set, and generate the first intermediate context feature set.

[0065] Step S12541: Perform delay fluctuation interval division processing on the network delay parameters of the device to generate multiple delay interval units.

[0066] Based on the range of values ​​for the device's network latency parameters, the latency is divided into multiple latency intervals. For example, the latency range can be divided into three intervals: [0, D1), [D1, D2), and [D2, +∞), where D1 and D2 are thresholds determined based on actual conditions.

[0067] The frequency and duration of network latency parameters for statistical devices within each latency interval are recorded. The latency data within each interval is considered a single latency interval unit.

[0068] Step S12542: For each of the delay interval units, extract the delay fluctuation amplitude parameter and the delay duration parameter.

[0069] For each delay interval cell, calculate the delay fluctuation amplitude parameter. The delay fluctuation amplitude refers to the difference between the maximum and minimum delay values ​​within that delay interval cell. Simultaneously, record the delay duration of that delay interval cell, i.e., the time from entering the delay interval to leaving it.

[0070] Step S12543: Call the context association model to quantify the influence of the delay fluctuation amplitude parameter and the delay duration parameter, and generate an influence score value for each delay interval unit.

[0071] The context association model is a decision tree-based model. Using delay fluctuation amplitude and delay duration parameters as input features, the context association model evaluates each delay interval unit according to predefined decision rules.

[0072] During the model training phase, a large amount of historical data is used to train the decision tree and determine the node splitting rules and scoring criteria. In the prediction phase, the latency fluctuation amplitude and latency duration parameters of the current latency interval unit are input into the context association model. The context association model calculates an impact score based on the decision rules. This impact score represents the degree of influence of the current latency interval unit on the user's operation.

[0073] Step S12544: Prioritize the delay interval units according to the impact score to generate a delay priority sequence.

[0074] The impact scores of all delay interval units are compared, and the delay interval units are sorted in descending order. The sorted sequence of delay interval units is the delay priority sequence.

[0075] Step S12545: Perform feature recombination processing on the device network delay parameters based on the delay priority sequence to obtain the first intermediate context feature set.

[0076] The device network latency parameters are rearranged according to the latency priority sequence. Latency parameters corresponding to higher priority latency intervals are placed first, and lower priority parameters are placed later. The rearranged device network latency parameters are then encoded and concatenated to form the first intermediate context feature set.

[0077] Step S1255: Input the first intermediate semantic feature set, the first intermediate temporal feature set, the first intermediate attribute feature set and the first intermediate context feature set into the feature fusion layer. After mapping the dimensions of each feature set to a preset dimension through the linear projection layer, perform cross attention calculation processing to generate a cross attention weight set.

[0078] The linear projection layer of the feature fusion layer is a fully connected layer. Its function is to map the dimensions of the first intermediate semantic feature set, the first intermediate temporal feature set, the first intermediate attribute feature set, and the first intermediate context feature set to a preset dimension. The preset dimension is a fixed value determined based on the model design and experiments.

[0079] For each feature set, the linear projection layer transforms the vectors of its original dimensions into vectors of a preset dimension through matrix multiplication. The transformed vectors are dimensionally consistent, facilitating subsequent cross-attention computation.

[0080] Cross-attention computation is a method based on the attention mechanism. When calculating cross-attention weights, the relationships between different feature sets are considered. For a vector in each feature set, its attention score is calculated relative to vectors in other feature sets. The attention score represents the degree of correlation between two vectors.

[0081] The attention scores are converted into attention weights using the softmax function. Combining all the attention weights together generates a set of cross-attention weights.

[0082] Step S1256: After weighted fusion of the first intermediate semantic feature set, the first intermediate temporal feature set, the first intermediate attribute feature set, and the first intermediate context feature set according to the cross-attention weight set, the results are input into the intent knowledge construction layer for demand knowledge graph mapping to obtain the preliminary intent features.

[0083] Based on the cross-attention weight set, each vector in the first intermediate semantic feature set, the first intermediate temporal feature set, the first intermediate attribute feature set, and the first intermediate context feature set is weighted. The weighted vectors are then concatenated to form a fused feature vector.

[0084] The intent knowledge construction layer is a graph neural network-based model that maps the fused feature vectors to a pre-built demand knowledge graph. This demand knowledge graph contains various products, demand categories, and the relationships between them.

[0085] Through propagation and aggregation operations in a graph neural network, the fused feature vectors are matched and associated with nodes and edges in the demand knowledge graph. The resulting knowledge graph representation related to user needs is the preliminary intent feature.

[0086] Step S1257: Input the preliminary intent features into the intent refinement layer for potential demand mining processing to obtain the refined intent features.

[0087] For example, step S1257-1: The knowledge graph parsing module in the intent refinement layer performs entity node traversal processing on the user basic needs knowledge graph in the preliminary intent features, and extracts the entity association path set in the user basic needs knowledge graph.

[0088] The knowledge graph parsing module starts from the root node of the user's basic needs knowledge graph and visits each entity node sequentially according to the graph traversal algorithm (such as depth-first search or breadth-first search). During the visit, it records the path from the root node to each entity node; these paths are the entity association paths.

[0089] For each entity association path, the node identifiers and edge types on the path are recorded. Combining all entity association paths together yields the entity association path set.

[0090] Step S1257-2: Call the multi-hop inference module in the intent refinement layer, and perform path extension processing on the entity association path set based on the preset maximum hop count threshold and loop detection rules to generate an extended path set containing indirect association nodes.

[0091] The multi-hop inference module expands the set of entity association paths. The preset maximum hop threshold is a pre-defined parameter that indicates the maximum number of steps a path can extend.

[0092] Starting from the end node of each entity's associated path, the path is extended based on the edge relationships in the knowledge graph. During the extension process, loop detection rules are used to avoid loops in the path. If the number of hops in the extended path does not exceed the maximum hop count threshold and does not form a loop, it is added to the extended path set.

[0093] Step S1257-3: Use the context embedding module in the intent refinement layer to perform context semantic encoding on each node in the extended path set to generate a node context vector set.

[0094] The context embedding module is primarily used to perform contextual semantic encoding on each node in the extended path set, thereby capturing the semantic information of the node within the entire knowledge graph context. This process considers the node's neighboring nodes and the types of relationships between them. First, for each node in the extended path set, its neighboring node information is extracted; neighboring nodes are other nodes directly connected to that node via edges. Simultaneously, the types of edges between nodes are recorded; different edge types represent different semantic relationships.

[0095] Next, the context embedding module transforms the information of nodes and their neighboring nodes into vector representations. To achieve this, pre-trained word embedding models, trained on large-scale text data, are used to convert the entity names represented by nodes into semantically informative vectors. The types of edges between nodes are also encoded and converted into corresponding vectors.

[0096] After obtaining the vector representations of nodes, adjacent nodes, and edge types, the context embedding module fuses these vectors using an aggregation method. Common aggregation methods include weighted summation, which integrates the vector information of different adjacent nodes and edge types by assigning different weights to them, forming the node's context vector. This process is performed on each node in the extended path set, ultimately generating a set of node context vectors.

[0097] Step S1257-4: The attention aggregation module in the intent refinement layer performs dynamic weight allocation processing on the node context vector set to generate a node attention weight distribution set.

[0098] The attention aggregation module dynamically assigns weights to the set of node context vectors. Since different nodes within the expanded path set may have varying importance in uncovering potential user needs, it's necessary to assign appropriate weights to each node. This attention aggregation module determines the weights based on the similarity and correlation between the node context vectors.

[0099] Specifically, for each vector in the node context vector set, the attention aggregation module calculates a similarity score between it and other vectors. Similarity scores can be calculated in various ways, such as cosine similarity. By calculating similarity scores, the degree of connection between nodes can be understood.

[0100] Then, based on these similarity scores, an attention mechanism is used to generate attention weights for each node. The attention mechanism normalizes the similarity scores so that the sum of the attention weights for all nodes is 1. In this way, each node has a corresponding attention weight, and these attention weights form a set of node attention weight distributions.

[0101] Step S1257-5: Based on the node attention weight distribution set, perform preference intensity quantization on the nodes in the extended path set to generate a node preference intensity sequence.

[0102] Based on the node attention weight distribution set, the preference intensity of nodes in the extended path set can be quantified. The attention weight of each node reflects its importance in the entire extended path set, and this importance is related to the user's preference intensity for the entity represented by that node.

[0103] Specifically, each weight value in the node attention weight distribution set is used as a measure of the preference intensity of the corresponding node. These weight values ​​can be arranged according to the order of the nodes in the extended path set to form a node preference intensity sequence, which intuitively shows the degree of user preference for the entities represented by each node in the extended path set.

[0104] Step S1257-6: Input the node preference intensity sequence into the preference mapping module in the intent refinement layer for multi-dimensional preference projection processing to generate the user's potential demand preferences to constitute the refined intent features.

[0105] The preference mapping module performs multidimensional preference projection on the node preference intensity sequence. Internally, this module has a predefined multidimensional preference space containing multiple different preference dimensions, such as price preference, feature preference, and brand preference.

[0106] When the node preference intensity sequence is input into the preference mapping module, the module projects the node preference intensity onto various dimensions of the multidimensional preference space based on the attribute information of the entity represented by the node. For example, if a node represents a smartphone, and its attribute information includes price, camera function, brand, etc., then the preference intensity of that node will be projected onto the corresponding positions of the price preference dimension, function preference dimension, brand preference dimension, etc.

[0107] Through this multidimensional preference projection processing, the node preference intensity sequence can be transformed into a representation of the user's preference degree in different preference dimensions. These representations combined together constitute the user's potential demand preferences, which is to refine the intent features.

[0108] In the above embodiments, the method for identifying combined models of intent includes the following steps:

[0109] Step S210: Obtain a training sample set through a preset multimodal data acquisition interface. The training sample set includes labeled user behavior sequence samples, labeled product attribute samples, and labeled context environment samples. The user behavior sequence samples include intent category labels and refined preference labels for operation event sequences.

[0110] Step S220: Semantic vector encoding is performed on the event triggering keywords in the user behavior sequence sample to generate a semantic training vector set, and time-series normalization is performed on the path dwell time in the operation event sequence to generate a time-series training vector set.

[0111] Step S230: Perform attribute vector mapping processing on the product function description in the product attribute annotation sample to generate an attribute training vector set, and perform context normalization processing on the device network delay parameter in the context environment annotation sample to generate a context training vector set.

[0112] Step S240: Input the semantic training vector set, the temporal training vector set, the attribute training vector set, and the context training vector set into the initialized intent recognition combination model to generate a joint training feature vector set, and input it into the intent knowledge construction layer for multi-label classification training to output preliminary intent prediction features and refined intent prediction features.

[0113] Step S250: Calculate a first loss function value based on the preliminary intent prediction features and the intent category label, and calculate a second loss function value based on the refined intent prediction features and the refined preference label. Optimize the parameters of the semantic association model, temporal association model, attribute association model, and context association model in the intent recognition combination model using the gradient backpropagation algorithm until the first loss function value and the second loss function value satisfy a preset convergence condition.

[0114] Step S130: Generate an interaction guidance strategy tree based on the intent feature set. The interaction guidance strategy tree includes node branch weights and strategy execution priorities.

[0115] After obtaining the intent feature set, an interaction guidance strategy tree needs to be generated based on it. This interaction guidance strategy tree can be used to guide the user's interaction behavior. The specific steps are as follows:

[0116] Step S131: Perform demand category mapping processing on the preliminary intent features to generate a demand category identifier set.

[0117] The initial intent features are presented in the form of a user basic needs knowledge graph. The needs category mapping process is the process of extracting user needs category information from the knowledge graph. It analyzes the entity nodes and relationships in the knowledge graph to identify different needs categories.

[0118] For example, if a knowledge graph contains nodes and relationships related to clothing, electronic products, and food, the user's needs can be determined by mining and classifying this information. Each need category is assigned a unique identifier, and these identifiers are then aggregated to generate a set of need category identifiers.

[0119] Step S132: Perform preference dimension decomposition on the refined intent features to generate a preference dimension identifier set.

[0120] Refined intent features represent users' potential needs and preferences. Preference dimension decomposition is the process of breaking down these potential needs and preferences into different preference dimensions. Since the refined intent features have already been projected into a multidimensional preference space, decomposition can be performed based on the dimensional information of this multidimensional preference space.

[0121] For example, the multidimensional preference space covers dimensions such as price, function, and brand. The information on the degree of preference of the detailed intention features on these dimensions is extracted, and a unique identifier is assigned to each preference dimension. By integrating these identifiers together, we obtain the preference dimension identifier set.

[0122] Step S133: Perform node initialization processing based on the demand category identifier set and the preference dimension identifier set to generate an initial strategy node set.

[0123] Step S1331: Map each requirement category identifier in the requirement category identifier set to the corresponding node type identifier to generate a node type identifier set.

[0124] For each demand category identifier in the demand category identifier set, it is converted into a corresponding node type identifier according to pre-defined mapping rules. These mapping rules are formulated based on the characteristics and needs of e-commerce business. For example, the clothing demand category identifier is mapped to the "clothing" node type identifier, and the electronics demand category identifier is mapped to the "electronics" node type identifier, and so on. Combining all the mapped node type identifiers generates the node type identifier set.

[0125] Step S1332: Perform sub-dimension decomposition on each preference dimension identifier in the preference dimension identifier set to generate a sub-dimension identifier set.

[0126] Each preference dimension identifier in the preference dimension identifier set may contain more detailed sub-dimension information. For example, the price preference dimension may include sub-dimensions of different price ranges; the feature preference dimension may cover sub-dimensions of different functional characteristics.

[0127] Each preference dimension identifier is decomposed into sub-dimensions to identify the sub-dimension information. A unique identifier is assigned to each sub-dimension. By summing up all the sub-dimension identifiers, a set of sub-dimension identifiers is generated.

[0128] Step S1333: Combine and match each node type identifier in the node type identifier set with at least one sub-dimension identifier in the sub-dimension identifier set, and filter them according to the preset requirement category and preference dimension correlation rules to generate a candidate strategy node set.

[0129] Each node type identifier in the node type identifier set is combined with a sub-dimension identifier in the sub-dimension identifier set. For example, the "clothing" node type identifier can be combined with the price sub-dimension identifier, style sub-dimension identifier, etc.

[0130] After the combinations are completed, they are filtered according to preset rules regarding the correlation between demand categories and preference dimensions. These rules are based on a large amount of user data and business experience and are used to determine which combinations are reasonable and relevant to user needs. For example, if there is no obvious correlation between a certain demand category and a certain preference dimension, then the corresponding combination will be filtered out. After filtering, the resulting combinations are called candidate strategy nodes. Integrating these candidate strategy nodes generates a set of candidate strategy nodes.

[0131] Step S1334: Perform validity verification on each candidate policy node in the candidate policy node set according to the preset node validity verification rules, filter out invalid policy nodes, and obtain the initial policy node set.

[0132] Predefined node validity verification rules are used to ensure the validity and rationality of candidate strategy nodes. These rules may include whether the node's attribute values ​​are within a reasonable range, and whether the association between the node and other nodes conforms to business logic.

[0133] For each candidate policy node in the candidate policy node set, these rules are applied for validation. If a candidate policy node does not conform to the rules—for example, if its attribute value exceeds the normal range or its association with other nodes contradicts the rules—then the node will be considered an invalid policy node and filtered out. After filtering, the remaining candidate policy nodes constitute the initial policy node set.

[0134] Step S134: Perform hierarchical partitioning on the initial strategy node set based on the category priority in the demand category identifier set to generate a strategy hierarchy structure.

[0135] Step S1341: Perform priority sorting on each requirement category identifier in the requirement category identifier set to generate a category priority sequence.

[0136] Based on the characteristics of e-commerce operations and the importance of user needs, each demand category identifier in the demand category identifier set is assigned a priority. Priority assignment can be based on various factors, such as the prevalence and urgency of the demand.

[0137] These priorities are sorted, and the requirement category identifiers are arranged into a category priority sequence according to the order of priority from high to low.

[0138] Step S1342: Based on the category priority sequence and the preset node hierarchy constraints, assign each strategy node in the initial strategy node set to the corresponding hierarchy position, and perform hierarchy reallocation processing on isolated nodes to generate hierarchy allocation results.

[0139] The preset node hierarchy constraints define the hierarchy range that different types of policy nodes should occupy in the policy tree. Based on the category priority sequence, each policy node in the initial policy node set is assigned to the corresponding hierarchy position according to the priority of its corresponding requirement category.

[0140] During the allocation process, some isolated nodes may appear, that is, nodes that have no obvious connection with other nodes. For these isolated nodes, hierarchical reallocation processing will be performed based on their attributes and associations to ensure they can be reasonably integrated into the hierarchical structure of the policy tree. After allocation and reallocation processing, the hierarchical position information of each policy node is obtained, and this information constitutes the hierarchical allocation result.

[0141] Step S1343: Based on the hierarchical allocation result, perform parent-child relationship binding processing on the strategy nodes in the initial strategy node set to generate a parent-child relationship connection graph.

[0142] Based on the hierarchical allocation results, the parent-child relationships between policy nodes in the initial policy node set are determined. Policy nodes at higher levels are designated as parent nodes, and policy nodes at lower levels that are logically related to their parent nodes are designated as child nodes.

[0143] These parent-child relationships are bound together and represented graphically to generate a parent-child relationship connection graph. This graph visually illustrates the hierarchical relationships and connections between strategy nodes.

[0144] Step S1344: Perform structural integrity verification on the parent-child relationship connection graph according to the preset hierarchical structure verification rules. If there are unconnected isolated nodes, reallocate the hierarchical positions based on the category priority sequence until the strategy hierarchical structure that meets the hierarchical depth constraints is generated.

[0145] The pre-defined hierarchical structure verification rules are used to check the structural integrity of the parent-child relationship connection graph. These rules include checking for the existence of unconnected isolated nodes and whether the hierarchy depth meets the requirements.

[0146] The parent-child relationship connection graph is validated. If any isolated nodes are found, their hierarchical positions are reassigned according to the category priority sequence. After reassignment, structural integrity is validated again until the generated strategy hierarchy structure meets the hierarchy depth constraints.

[0147] Step S135: Based on the preference dimension identifier set, perform branch weight allocation processing on the strategy hierarchy to obtain the node branch weight and the strategy execution priority, and generate the interaction guidance strategy tree.

[0148] Step S1351: Extract the dimension priority parameter and dimension association strength parameter from the set of preference dimension identifiers.

[0149] The preference dimension identifier set contains relevant parameters for each preference dimension. Among them, the dimension priority parameter reflects the importance of the preference dimension in the user's needs, and the dimension association strength parameter indicates the degree of association between the preference dimension and other dimensions.

[0150] Step S1352: The dimension priority parameter and the dimension association strength parameter are subjected to nonlinear transformation processing through a preset nonlinear activation function to generate a preliminary weight set, and the preliminary weight set is normalized by a probability distribution normalization method to generate a normalized weight set.

[0151] The preset non-linear activation function can be a common one such as the Sigmoid function or the ReLU function. Its function is to perform non-linear transformation on the dimension priority parameter and the dimension correlation strength parameter to enhance the expressive power of the data.

[0152] The dimension priority parameter and dimension association strength parameter are input into a nonlinear activation function for processing, resulting in a preliminary weight set. To ensure that these weights can be used appropriately in subsequent calculations, a probability distribution normalization method is used to normalize the preliminary weight set so that the sum of all weights is 1, thus obtaining a normalized weight set.

[0153] Step S1353: Update the node connection relationships in the strategy hierarchy according to the normalized weight set to obtain the node branch weights.

[0154] The weights in the normalized weight set are assigned to the node connections in the strategy hierarchy. Each node connection corresponds to a weight value, which represents the importance of that connection in guiding user interaction.

[0155] The weights of these node connections are updated based on the normalized weight set, and the updated weights are the node branch weights.

[0156] Step S1354: Based on the node branch weights, prioritize the execution order of nodes in the strategy hierarchy to generate the strategy execution priority.

[0157] The execution order of nodes in the strategy hierarchy is prioritized based on the weight of their branch connections. Nodes connected by branches with larger weights have higher execution priority.

[0158] Sort the nodes in descending order of priority; the resulting sorting order is the policy execution priority. Applying the node branch weights and policy execution priorities to the policy hierarchy generates the interactive guidance policy tree.

[0159] Step S140: Perform dynamic path adjustment processing on the interactive guidance strategy tree based on the real-time feedback data set to obtain an optimized strategy tree set, wherein the optimized strategy tree set includes at least one updated node branch weight and strategy execution priority.

[0160] To enable the interactive guidance strategy tree to better adapt to users' real-time behavior and feedback, it needs to be dynamically adjusted based on the real-time feedback data set. The specific steps are as follows:

[0161] Step S141: Collect user feedback data set in real time from the target interactive interface. The user feedback data set includes strategy response time parameters and strategy execution effect parameters.

[0162] Deploy a data acquisition module on the target interactive interface to collect user feedback data in real time. The strategy response time parameter refers to the time interval from when the system displays the interactive guidance strategy to the user to when the user responds; it reflects the user's response speed to the strategy.

[0163] The strategy execution effect parameters are used to measure the effectiveness of the interactive guidance strategy, such as whether the user performed the operation according to the strategy and whether the result of the operation met expectations. Combining these collected parameters forms a user feedback data set.

[0164] Step S142: Perform response delay evaluation processing on the strategy response duration parameter to generate a delay impact score set, and perform effect quantification processing on the strategy execution effect parameter to generate an effect score set.

[0165] For the strategy response time parameter, response latency will be evaluated based on preset evaluation criteria. The evaluation criteria can be formulated based on a large amount of historical data and business experience, such as dividing the response time into different intervals, with each interval corresponding to a latency impact score.

[0166] Each strategy response time parameter is evaluated to obtain a corresponding delay impact score. Combining these scores generates a delay impact score set.

[0167] For strategy execution performance parameters, a quantitative method is used to convert them into specific scores. For example, if a user follows the strategy and achieves the expected results, a higher score is given; if the user does not follow the strategy or the results are unsatisfactory, a lower score is given. Combining the quantitative scores of all strategy execution performance parameters generates a performance score set.

[0168] Step S143: The delay impact scores in the delay impact score set and the effect scores in the effect score set are processed by a preset standardization method and then weighted to generate a delay weight set and an effect weight set.

[0169] The preset standardization method can be Z-score standardization, etc., which standardizes the score values ​​in the delay impact score set and the effect score set so that they have the same dimensions and distribution range.

[0170] After standardization, these scores undergo weight transformation. Weight transformation can assign different weights based on business needs and importance. For example, different weight coefficients can be assigned to delay impact scores and effect scores respectively. Multiplying the score values ​​by the corresponding weight coefficients yields delay weight sets and effect weight sets.

[0171] Step S144: Perform joint analysis and processing on the delay weight set and the effect weight set to generate a node optimization priority sequence.

[0172] Joint analysis takes into account the weight information in both the delay weight set and the effect weight set. Methods such as weighted summation can be used to combine the delay weights and effect weights to obtain the comprehensive weight for each node.

[0173] The nodes are sorted according to their overall weight, arranged in descending order. The resulting sorting order is the node optimization priority sequence. This node optimization priority sequence indicates the priority of each node in the interaction guidance strategy tree.

[0174] Step S145: Mark the target nodes in the interaction guidance strategy tree according to the node optimization priority sequence to generate a set of node identifiers to be optimized, and generate a set of node optimization instructions based on the set of node identifiers to be optimized.

[0175] Based on the node optimization priority sequence, target nodes in the interaction guidance strategy tree are marked. Nodes with higher priority are marked as nodes to be optimized. Combining the identifiers of these nodes to be optimized generates a set of identifiers for nodes to be optimized.

[0176] Based on the set of node identifiers to be optimized, a set of node optimization instructions is generated. The node optimization instructions contain specific optimization operations for each node to be optimized, such as adjusting node branch weights and updating strategy execution priorities.

[0177] Step S146: Dynamically adjust the node branch weights in the interaction guidance strategy tree according to the node optimization instruction set, and update the strategy execution priority to obtain the optimization strategy tree set.

[0178] Based on the set of node optimization instructions, the branch weights of nodes in the interactive guidance strategy tree are dynamically adjusted. Operations such as increasing, decreasing, or reallocating the branch weights of corresponding nodes are performed according to the requirements of the instructions.

[0179] After adjusting the node branch weights, the policy execution priority is updated based on the new node branch weights. The updated policy tree set is the optimized policy tree set, which can better adapt to the user's real-time feedback and behavior.

[0180] Step S150: Push the optimized strategy tree set to the target interactive interface to activate the interactive guidance operation.

[0181] The optimized strategy tree set is applied to the target user interface to guide user interaction. The specific steps are as follows:

[0182] Step S151: Perform interface adaptation processing on the strategy execution priorities in the optimization strategy tree set to generate an interface layout parameter set.

[0183] Interface adaptation involves transforming the execution priorities of strategies in the optimization strategy tree into layout parameters suitable for the target user interface. Factors such as the target interface's size, resolution, and layout style need to be considered.

[0184] Based on the strategy execution priority, determine the display position, size, color, and other layout parameters of each interactive guidance strategy on the interface. Combining these layout parameters generates the interface layout parameter set.

[0185] Step S152: Dynamically render the target interactive interface based on the interface layout parameter set to generate an updated interactive interface.

[0186] The target interactive interface is the visual interface through which users interact with the e-commerce system. Its rendering depends on the front-end rendering engine and the set of interface layout parameters. The rendering engine is a software module specifically designed to transform layout parameters into visual interface elements, following specific rendering rules and algorithms.

[0187] During dynamic rendering, the rendering engine reads every parameter in the interface layout parameter set. For the display position parameter of interface elements, the rendering engine places the elements accurately in the specified positions according to the coordinate system of the target interactive interface. For example, if the layout parameters specify that a tooltip for a certain interactive guidance strategy should be displayed in the upper right corner of the interface, the rendering engine will calculate the specific coordinates of that position in the interface coordinate system and render the tooltip to that position.

[0188] Regarding the size parameters of UI elements, the rendering engine adjusts the element size according to the resolution and layout ratio of the target interactive interface. If the layout parameters require that the width of a button be a certain proportion of the interface width, the rendering engine will calculate the specific width of the button based on the actual width of the current interface and then render it.

[0189] Color parameters are also an important part of layout parameters. The rendering engine will fill the interface elements with the corresponding color based on the specified color value. For example, setting a specific color for the title text of an interactive guidance strategy can make the title more eye-catching.

[0190] In addition to basic position, size, and color parameters, the set of interface layout parameters may also include animation effect parameters. If the layout parameters require a certain interface element to have a fade-in / fade-out animation effect, the rendering engine will implement this animation through a series of frame changes. In each frame, the element's opacity is gradually changed, thereby achieving the fade-in / fade-out visual effect.

[0191] During the rendering process, the rendering engine continuously updates the state of the interface elements based on the set of interface layout parameters until the entire target interactive interface is fully rendered. This dynamic rendering process generates an updated interactive interface that can appropriately display various interactive guidance strategies based on the strategy execution priority of the optimization strategy tree set.

[0192] Step S153: Load the interactive guidance control corresponding to the optimization strategy tree set in the updated interactive interface, and activate the real-time response function of the interactive guidance control to execute the interactive guidance operation.

[0193] The updated user interface provides a basic visual environment for loading interactive guide controls. Interactive guide controls are the specific UI elements that implement interactive guide operations, and they correspond to the various strategies in the optimization strategy tree set.

[0194] First, the type and number of interactive guidance controls to be loaded need to be determined based on the optimization strategy tree set. Different strategies may correspond to different types of controls. For example, a strategy for recommending products may correspond to product display card controls; a strategy for guiding users to perform a certain operation may correspond to operation prompt button controls.

[0195] After determining the type and quantity of controls, a suitable control template will be selected from a pre-designed control library. The control library contains various types of interactive guidance control templates, each with its own specific style and function. For example, a product display card control template might include a display area for product images, names, prices, and other information, as well as interactive elements such as a purchase button.

[0196] Next, the selected control templates are loaded into the updated user interface. During loading, the controls are precisely placed in their corresponding positions on the screen based on the strategy execution priority in the optimization strategy tree set and the set of interface layout parameters. For example, controls corresponding to higher-priority strategies are placed in prominent positions on the screen so that users can easily notice them.

[0197] After loading the interactive guide controls, their real-time response capability needs to be activated. Real-time response means that the controls can respond promptly to user actions and execute corresponding interactive guide actions based on those actions. To achieve this, a corresponding event handler function can be bound to each control.

[0198] For example, a purchase button in a product display card control can be bound to a purchase event handler function. When the user clicks the button, the event handler function will be triggered, executing the corresponding purchase process, such as redirecting to the product details page or adding the product to the shopping cart.

[0199] For the operation prompt button control, a prompt event handler function will be bound. When the user clicks the button, an operation prompt message will pop up, guiding the user to complete the corresponding operation.

[0200] By activating the real-time response function of the interactive guidance control, users can complete various operations based on the guidance of the optimized strategy tree set when interacting with the updated interactive interface, thereby realizing the execution of interactive guidance operations.

[0201] Figure 2 The diagram illustrates exemplary hardware and software components of an intelligent e-commerce behavior event decision-making system 100 that integrates multi-source perception and implements the inventive concept, according to some embodiments of the present invention. For example, a processor 120 can be used in the intelligent e-commerce behavior event decision-making system 100 that integrates multi-source perception and performs the functions described in the present invention.

[0202] The intelligent e-commerce behavior event decision-making system 100, which integrates multi-source perception, can be a general-purpose server or a special-purpose server; both can be used to implement the intelligent e-commerce behavior event decision-making method integrating multi-source perception of the present invention. Although only one server is shown in this invention, for convenience, the functions described in this invention can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0203] For example, the intelligent e-commerce behavior event decision-making system 100 integrating multi-source sensing may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the intelligent e-commerce behavior event decision-making system 100 integrating multi-source sensing may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present invention can be implemented according to these program instructions. The intelligent e-commerce behavior event decision-making system 100 integrating multi-source sensing also includes an I / O interface 150 between the computer and other input / output devices.

[0204] For ease of explanation, only one processor is described in the intelligent e-commerce behavior event decision-making system 100 that integrates multi-source perception. However, it should be noted that the intelligent e-commerce behavior event decision-making system 100 that integrates multi-source perception may also include multiple processors. Therefore, the steps executed by one processor described in this invention may also be executed jointly or individually by multiple processors. For example, if the processor of the intelligent e-commerce behavior event decision-making system 100 that integrates multi-source perception executes steps A and B, it should be understood that steps A and B may also be executed jointly by two different processors or individually by one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0205] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned intelligent e-commerce behavior event decision-making method integrating multi-source perception is realized.

[0206] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for intelligent e-commerce behavioral event decision-making that integrates multi-source sensing, characterized in that, The method includes: The user behavior data set, product attribute data set, and environmental context data set are obtained through a preset data collection interface. The user behavior data set includes the sequence of operation events triggered by the user in the interactive interface and the corresponding page jump path. The user behavior data set, the product attribute data set, and the environmental context data set are subjected to intent recognition processing to generate an intent feature set, which includes preliminary intent features and refined intent features. An interaction guidance strategy tree is generated based on the intent feature set. The interaction guidance strategy tree includes node branch weights and strategy execution priorities. Based on the real-time feedback data set, the interactive guidance strategy tree is dynamically adjusted to obtain an optimized strategy tree set, which includes at least one updated node branch weight and strategy execution priority. The optimized strategy tree set is pushed to the target interactive interface to activate the interactive guidance operation.

2. The intelligent e-commerce behavior event decision-making method integrating multi-source perception as described in claim 1, characterized in that, The step of performing intent recognition processing on the user behavior data set, the product attribute data set, and the environmental context data set to generate an intent feature set includes: The operation event sequence in the user behavior data set is subjected to semantic parsing processing to generate a semantic association feature set, which includes event triggering keywords and event association attributes. The page jump path is subjected to time sequence analysis to generate a path time sequence feature set, which includes path dwell time and path node switching frequency. The product description text in the product attribute data set is subjected to attribute extraction processing to generate an attribute keyword vector set, which includes product functional features and product applicable scenario features. The device status parameters in the environmental context data set are subjected to context encoding to generate a context environment vector set, which includes device network latency parameters and device screen adaptation parameters. The semantic association feature set, the path temporal feature set, the attribute keyword vector set, and the context environment vector set are input into a preset intent recognition combination model for intent recognition to obtain the preliminary intent features and the refined intent features. The preliminary intent features are used to represent the user's basic needs knowledge graph, and the refined intent features are used to represent the user's potential needs and preferences.

3. The intelligent e-commerce behavior event decision-making method integrating multi-source perception according to claim 2, characterized in that, The intent recognition combination model includes a semantic association model, a temporal association model, an attribute association model, a contextual association model, a feature fusion layer, an intent knowledge construction layer, and an intent refinement layer. The step of inputting the semantic association feature set, the path temporal feature set, the attribute keyword vector set, and the contextual environment vector set into the preset intent recognition combination model for intent recognition, to obtain the preliminary intent features and the refined intent features, includes: The semantic association model is invoked to perform context-associative encoding on the event-triggered keywords in the semantic association feature set, thereby generating a first intermediate semantic feature set. The temporal correlation model is invoked to perform weight allocation processing on the path dwell time in the path temporal feature set, thereby generating a first intermediate temporal feature set; The attribute association model is invoked to perform scene mapping processing on the product functional features in the attribute keyword vector set, generating a first intermediate attribute feature set; The context association model is invoked to perform latency impact assessment on the device network latency parameters in the context environment vector set, generating a first intermediate context feature set. The first intermediate semantic feature set, the first intermediate temporal feature set, the first intermediate attribute feature set, and the first intermediate context feature set are input into the feature fusion layer. After the dimensions of each feature set are uniformly mapped to a preset dimension through the linear projection layer, cross attention calculation is performed to generate a cross attention weight set. After weighted fusion of the first intermediate semantic feature set, the first intermediate temporal feature set, the first intermediate attribute feature set, and the first intermediate context feature set according to the cross-attention weight set, the results are input into the intent knowledge construction layer for demand knowledge graph mapping to obtain the preliminary intent features. The preliminary intent features are input into the intent refinement layer for potential demand mining to obtain the refined intent features.

4. The intelligent e-commerce behavior event decision-making method integrating multi-source perception according to claim 3, characterized in that, The invocation of the time-series correlation model performs weight allocation processing on the path dwell time in the path time-series feature set to generate a first intermediate time-series feature set, including: The path node switching frequency in the path time sequence feature set is dynamically divided into multiple time sequence window units based on a preset time interval or event number threshold. For each of the time-series window units, extract the dwell time distribution parameters and node switching interval parameters within the time-series window unit; The time-series correlation model is invoked to perform joint weight calculation on the dwell time distribution parameters and the node switching interval parameters to generate local weight values ​​for each time-series window unit; The path dwell time within the time window unit is weighted and aggregated according to the local weight value to obtain the first intermediate time series feature set. Furthermore, the invocation of the context association model performs latency impact assessment processing on the device network latency parameters in the context environment vector set to generate a first intermediate context feature set, including: The network latency parameters of the device are divided into latency fluctuation intervals to generate multiple latency interval units; For each of the aforementioned delay interval units, extract the delay fluctuation amplitude parameter and the delay duration parameter; The context association model is invoked to quantify the impact of the delay fluctuation amplitude parameter and the delay duration parameter, and an impact score value is generated for each delay interval unit. Based on the impact score, the delay interval units are prioritized to generate a delay priority sequence; Based on the delay priority sequence, the network delay parameters of the device are reorganized to obtain the first intermediate context feature set.

5. The intelligent e-commerce behavioral event decision-making method integrating multi-source perception according to claim 1, characterized in that, The step of generating an interaction guidance strategy tree based on the intent feature set includes: The initial intent features are subjected to demand category mapping processing to generate a demand category identifier set; The refined intent features are decomposed into preference dimensions to generate a set of preference dimension identifiers; Based on the set of demand category identifiers and the set of preference dimension identifiers, node initialization processing is performed to generate an initial set of strategy nodes; Based on the category priority in the set of requirement category identifiers, the initial strategy node set is hierarchically divided to generate a strategy hierarchy structure. Based on the set of preference dimension identifiers, the branch weights of the strategy hierarchy are assigned to obtain the node branch weights and the strategy execution priorities, and the interaction guidance strategy tree is generated.

6. The intelligent e-commerce behavior event decision-making method integrating multi-source perception according to claim 5, characterized in that, The step of initializing nodes based on the demand category identifier set and the preference dimension identifier set to generate an initial strategy node set includes: Map each demand category identifier in the demand category identifier set to a corresponding node type identifier to generate a node type identifier set; Each preference dimension identifier in the preference dimension identifier set is subjected to sub-dimension decomposition to generate a sub-dimension identifier set; Each node type identifier in the node type identifier set is combined and matched with at least one sub-dimension identifier in the sub-dimension identifier set, and then filtered according to preset requirements category and preference dimension correlation rules to generate a candidate strategy node set. The validity of each candidate strategy node in the candidate strategy node set is verified according to the preset node validity verification rules, and invalid strategy nodes are filtered out to obtain the initial strategy node set. And, the step of performing hierarchical partitioning of the initial strategy node set based on the category priority in the demand category identifier set to generate a strategy hierarchy structure includes: Each requirement category identifier in the requirement category identifier set is sorted by priority to generate a category priority sequence; Based on the category priority sequence and the preset node hierarchy constraints, each strategy node in the initial strategy node set is assigned to the corresponding hierarchy position, and isolated nodes are reassigned hierarchically to generate a hierarchy allocation result. Based on the hierarchical allocation result, the strategy nodes in the initial strategy node set are subjected to parent-child relationship binding processing to generate a parent-child relationship connection graph. The parent-child relationship connection graph is subjected to structural integrity verification processing according to the preset hierarchical structure verification rules. If there are unconnected isolated nodes, the hierarchical positions are reassigned based on the category priority sequence until the strategy hierarchical structure that meets the hierarchical depth constraints is generated.

7. The intelligent e-commerce behavior event decision-making method integrating multi-source perception according to claim 5, characterized in that, The step of performing branch weight allocation processing on the strategy hierarchy based on the preference dimension identifier set to obtain the node branch weights and the strategy execution priorities includes: Extract the dimension priority parameter and dimension association strength parameter from the preference dimension identifier set; The dimension priority parameter and the dimension association strength parameter are subjected to nonlinear transformation processing through a preset nonlinear activation function to generate a preliminary weight set, and the preliminary weight set is normalized by a probability distribution normalization method to generate a normalized weight set. The node connection relationships in the strategy hierarchy are updated according to the normalized weight set to obtain the node branch weights. Based on the node branch weights, the execution order of nodes in the strategy hierarchy is prioritized to generate the strategy execution priority.

8. The intelligent e-commerce behavioral event decision-making method integrating multi-source perception according to claim 1, characterized in that, The step of dynamically adjusting the interactive guidance strategy tree based on the real-time feedback data set to obtain an optimized strategy tree set includes: The user feedback data set is collected in real time from the target interactive interface. The user feedback data set includes strategy response time parameters and strategy execution effect parameters. The strategy response time parameter is evaluated for response latency to generate a latency impact score set, and the strategy execution effect parameter is quantified to generate an effect score set. The delay impact scores in the delay impact score set and the effect scores in the effect score set are processed by a preset standardization method and then weighted to generate a delay weight set and an effect weight set. The delay weight set and the effect weight set are jointly analyzed and processed to generate a node optimization priority sequence; The target nodes in the interaction guidance strategy tree are marked according to the node optimization priority sequence to generate a set of node identifiers to be optimized, and a set of node optimization instructions is generated based on the set of node identifiers to be optimized. The node branch weights in the interaction guidance strategy tree are dynamically adjusted according to the node optimization instruction set, and the strategy execution priority is updated to obtain the optimization strategy tree set.

9. The intelligent e-commerce behavioral event decision-making method integrating multi-source perception according to claim 1, characterized in that, The step of pushing the optimized strategy tree set to the target interactive interface to activate the interactive guidance operation includes: The execution priorities of the optimization strategy tree set are processed for interface adaptation, and a set of interface layout parameters is generated. Based on the set of interface layout parameters, the target interactive interface is dynamically rendered to generate an updated interactive interface. The updated interactive interface loads the interactive guidance control corresponding to the optimized strategy tree set, and activates the real-time response function of the interactive guidance control to execute the interactive guidance operation.

10. A smart e-commerce behavior event decision-making system integrating multi-source sensing, characterized in that, The system includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the intelligent e-commerce behavior event decision-making method integrating multi-source perception as described in any one of claims 1-9.

Images