Method and system for accurate commodity recommendation based on multi-dimensional user portrait

Abstract

The application discloses a kind of based on multi-dimensional user portrait's commodity accurate recommendation method and system, it is related to commodity accurate recommendation field, first, by the orthogonal decoupling projection and capsule instantiation of user attribute and historical behavior sequence, construct long-term intention capsule cluster.Second, for real-time behavior data, the topological structure of commodity knowledge graph is used to capture and encode transient behavior vector, and the coupling degree of transient behavior and each long-term intention capsule is quantified by introducing iterative dynamic routing protocol.The core is, based on the comparison of peak value and dynamic mutation threshold of intention coupling coefficient distribution, mutation detection determination is carried out.When detecting sudden intention, it is directly used as independent capsule instantiation and incorporated into long-term capsule cluster;For conventional behavior, momentum fusion is carried out.Finally, based on the fusion active capsule set, cross attention matching scoring is carried out on candidate commodity, and accurate recommendation considering user stable preference and transient mutation intention is realized.

Description

Technical Field

[0001] This invention belongs to the field of precise product recommendation, specifically relating to a method and system for precise product recommendation based on multi-dimensional user profiles. Background Technology

[0002] With the booming development of e-commerce, precise recommendation systems have become a core driving force for major e-commerce platforms to improve user experience and conversion rates. To gain a deeper understanding of users' consumption habits and potential needs, building a precise product recommendation solution based on multi-dimensional user profiles is particularly important. User attribute data and historical behavior sequences together constitute their multi-dimensional profiles; this long-accumulated data can outline users' stable preferences. Therefore, recommendations based on multi-dimensional user profiles can help platforms achieve personalized product distribution, effectively reducing users' information acquisition costs from a massive product knowledge base, thereby enhancing the platform's overall commercial value and user stickiness.

[0003] However, real-world consumption scenarios are often dynamic, and user intent shifts can be instantaneous. Existing product recommendation solutions based on multi-dimensional user profiles typically rely on attention networks or graph aggregation models. These models generate a comprehensive representation vector of user preferences by weighting or multi-layering massive historical behavior sequences and then matching products. This history-dependent mechanism generally falls into a severe lag trap when dealing with sudden intent shifts. For example, when a user with a long-standing tech geek profile suddenly clicks on a baby product due to gift-giving needs, existing models often fail to respond promptly. The core flaws are: firstly, existing multi-dimensional vectors suffer from severe dilution effects during weighted averaging or pooling aggregation. Because historical preferences accumulate significant weight, isolated new, mutating behavior nodes are instantly overwhelmed in mathematical calculations, leading to a suppression effect of the massive historical profile on immediate conversions; secondly, existing models typically project long-term and short-term features into the same latent space for computation. Long-term stable preferences and short-term sudden intents pull against each other. Due to the lack of effective spatial decoupling, the model inevitably sacrifices small-sample sudden intents to minimize global loss, causing the network to become desensitized to abnormal biases. Because these anomalous behaviors cannot be captured in milliseconds and the suppression of historical profiles cannot be blocked, platforms often lose extremely high instant conversion rates.

[0004] Therefore, how to effectively decouple the long-term and short-term intent spaces, keenly capture transient and sudden behaviors, and overcome the dilution and lag defects brought about by multi-dimensional historical features has become a technical problem that urgently needs to be solved in the field of accurate product recommendation. Summary of the Invention

[0005] This invention application provides a method for accurate product recommendation based on multi-dimensional user profiles.

[0006] The technical solution of this invention is as follows: A method for accurate product recommendation based on multi-dimensional user profiles, comprising: Step 1: Based on a product knowledge graph, orthogonally decoupled projection and capsule instantiation are performed on user attribute data and historical behavior sequences to obtain a long-term intent capsule cluster; Step 2: Through the topological structure of the product knowledge graph, transient micro-behavior capture and characterization encoding are performed on real-time behavior data collected in the current session window to obtain a transient behavior vector; Step 3: Based on an iterative dynamic routing protocol, intent coupling metric is performed on the transient behavior vector and each intent capsule in the long-term intent capsule cluster to obtain an intent coupling coefficient distribution; Step 4: Based on the comparison between the peak value of the intent coupling coefficient distribution and the dynamic mutation threshold, mutation detection and judgment are performed on the transient behavior vector. If a mutation is determined, it is instantiated as an independent capsule and incorporated into the long-term intent capsule cluster; otherwise, it is fused with the momentum of the long-term intent capsule to form a matching capsule, generating a fused activated capsule set; Step 5: Using the candidate product embedding features extracted from the product knowledge graph as the query object, cross-attention matching and scoring are performed on the fused activated capsule set, and the resulting comprehensive matching score is sorted in descending order to generate a product recommendation list.

[0007] This invention application provides another product precision recommendation system based on multi-dimensional user profiles.

[0008] The technical solution of this invention is as follows: A product precision recommendation system based on multi-dimensional user profiles, comprising: a user attribute historical behavior instantiation module, used to perform orthogonal decoupling projection and capsule instantiation on user attribute data and historical behavior sequences based on a product knowledge graph to obtain a long-term intent capsule cluster; a real-time behavior capture and encoding module, used to perform transient micro-behavior capture and representation encoding on real-time behavior data collected within the current session window through the topological structure of the product knowledge graph to obtain a transient behavior vector; and an intent coupling metric module, used to perform intent coupling between the transient behavior vector and each intent capsule in the long-term intent capsule cluster based on an iterative dynamic routing protocol. The system quantizes and obtains the intention coupling coefficient distribution; the mutation detection and judgment module is used to perform mutation detection and judgment on transient behavior vectors based on the comparison between the peak value of the intention coupling coefficient distribution and the dynamic mutation threshold. If a mutation is determined, it is instantiated as an independent capsule and incorporated into the long-term intention capsule cluster; otherwise, it is fused with the momentum of the long-term intention capsule to form a matching capsule, generating a fused activation capsule set; the product recommendation list generation module uses the candidate product embedding features extracted from the product knowledge graph as the query object, performs cross-attention matching and scoring on the fused activation capsule set, and sorts and truncates the results in descending order based on the obtained comprehensive matching score to generate a product recommendation list.

[0009] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention aims to solve the problems of undecoupled long-term and short-term latent spaces in existing recommendation models, and the lag trap of sudden intents being diluted by historical features. First, by orthogonally decoupling and projecting user attributes and historical behavior sequences into capsules and instantiating them, a long-term intent capsule cluster is constructed, fundamentally eliminating the mutual interference between long-term and short-term features in the same latent space. Second, for real-time behavior data, the topological structure of the product knowledge graph is used to capture and encode transient behavior vectors, and an iterative dynamic routing protocol is introduced to quantify the coupling degree between transient behaviors and each long-term intent capsule. The core is to perform mutation detection and judgment based on the comparison between the peak value of the intent coupling coefficient distribution and the dynamic mutation threshold. When a sudden intent is detected, it is directly instantiated as an independent capsule and incorporated into the long-term capsule cluster, thereby completely blocking the suppression effect and mathematical dilution of the large historical profile; for regular behaviors, momentum fusion is performed. Finally, based on this fused activated capsule set, cross-attention matching and scoring are performed on candidate products to achieve accurate recommendations that take into account both stable user preferences and transient mutation intents. Attached Figure Description

[0010] Figure 1 This is a flowchart of a method for accurate product recommendation based on multi-dimensional user profiles according to an embodiment of this application.

[0011] Figure 2 This is a data flow diagram illustrating the precise product recommendation method based on multi-dimensional user profiles according to an embodiment of this application.

[0012] Figure 3 This is a flowchart of step two in the product accurate recommendation method based on multi-dimensional user profiles according to an embodiment of this application.

[0013] Figure 4 This is a flowchart of step five in the product accuracy recommendation method based on multi-dimensional user profiles according to an embodiment of this application.

[0014] Figure 5 This is a block diagram of a product accuracy recommendation system based on multi-dimensional user profiles according to an embodiment of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] In view of the shortcomings of existing technologies, this application proposes a method for accurate product recommendation based on multi-dimensional user profiles. For example... Figure 1and Figure 2 As shown, Figure 1 This is a flowchart of a method for accurate product recommendation based on multi-dimensional user profiles according to an embodiment of this application. Figure 2 This is a data flow diagram illustrating the precise product recommendation method based on multi-dimensional user profiles according to an embodiment of this application.

[0017] In detail, step one: Based on the product knowledge graph, orthogonal decoupled projection and capsule instantiation are performed on user attribute data and historical behavior sequences to obtain long-term intent capsule clusters. It should be understood that in existing data processing and distribution logic, users' long-term stable preferences are often mixed with short-term occasional intentions in the same high-dimensional latent space for representation and calculation. When users generate immediate new needs during specific holidays, promotional activities, or unexpected events, their long-term historical behavior accumulates significant numerical weights and complex connections. These vast and mixed historical behavioral features produce a severe dilution effect during the pooling and aggregation stage of mathematical calculations. This dilution effect directly masks small, isolated, and immediate abrupt behavioral features, leading to severe passivation and lag in the calculation results regarding abrupt biases, causing the output to stubbornly favor long-term preferences. To effectively isolate the mutual interference between long-term and short-term features in the same latent space and to block the suppressive effect of massive historical profile features on sudden, subtle intentions, a process is introduced to orthogonally decouple and project user attribute data and historical behavior sequences based on product knowledge graphs and to instantiate capsules. This process achieves spatial dimension isolation in the initial construction stage of feature vectors, thereby constructing long-term intention capsule clusters with clear boundaries and independent features. This lays a structured spatial foundation for subsequently isolating and keenly capturing transient and abrupt behaviors.

[0018] In one specific embodiment of this application, step one includes: using the topological node structure of the product knowledge graph, performing entity alignment and multilayer perceptron encoding mapping on the product identifiers involved in the user attribute data and historical behavior sequences to obtain the original feature embedding matrix; performing orthogonal constraint-based intention space projection on the original feature embedding matrix to obtain the orthogonal intention representation matrix; and performing nonlinear compression and instantiation of the intention features on the orthogonal intention representation matrix to obtain the long-term intention capsule cluster.

[0019] The implementation process is as follows: First, acquire the basic data required for processing, mainly including user attribute data, historical behavior sequences, and a product knowledge graph. User attribute data refers to structured user information collected from the business front end and processed through discretization, containing specific discretized features such as age, gender, region, and consumption level. For example, the attribute data of a target user can be specified as a set of tags such as 28 years old, male, East China region, and high consumption level. These tags will be converted into one-hot encoding or initial low-dimensional dense vectors. Historical behavior sequences refer to the list of specific product identifiers and their corresponding timestamps extracted from the behavior log server, containing the user's clicks, favorites, purchases, and other behaviors in the past 90 days. For example, the sequence may contain entries such as product identifier 1045 (mechanical keyboard) and product identifier 2099 (gaming mouse). The product knowledge graph is a pre-built global graph database, which contains all product nodes within the platform and their multi-dimensional attribute graph network structure. The edges in the graph represent the category affiliation, brand association, or matching relationship between products, and the nodes in the graph store detailed multi-dimensional attribute metadata of the products.

[0020] After acquiring the aforementioned data, the historical behavior sequences are traversed, and specific product identifiers are extracted from each sequence. These product identifiers are then used as query keys for entity alignment within the product knowledge graph. The purpose of entity alignment is to accurately locate the topological node corresponding to the historical product within the vast graph structure, and retrieve the corresponding multidimensional attribute metadata along the first- and second-order network structures of that node. This metadata includes information such as the digital category, specific electronic brand, and price range of 500 to 1000 yuan. Next, the retrieved multidimensional attribute metadata is concatenated with the previously acquired user attribute data along the feature dimension. The resulting joint feature vector simultaneously captures the user's profile and the content of their long-term interactive products. To transform this high-dimensional and sparse joint feature into a dense representation suitable for network computation, the concatenated data is input into a multilayer perceptron model for encoding and mapping. This multilayer perceptron mainly consists of an input layer, multiple hidden layers, and an output layer. The neurons in the hidden layers use linear rectified functions as activation functions to capture the non-linear cross-relationships between features. In this multilayer perceptron network, each layer is calculated by multiplying the input vector by a weight matrix and adding a bias vector, then outputting it to the next layer via a linear rectified function. The weight matrix and bias vector in the multilayer perceptron are pre-trained on massive amounts of historical samples using an offline backpropagation algorithm combined with an adaptive moment estimation optimizer. For example, if the feature dimension of user attribute data is 20 dimensions and the feature dimension of product multidimensional attribute metadata is 80 dimensions, they are concatenated to form a 100-dimensional input vector. After processing by a multilayer perceptron containing two hidden layers, it is finally mapped to a uniform 64-dimensional dense vector at the output layer. If the historical behavior sequence contains 50 product interaction records, after entity alignment and one-by-one encoding mapping by the multilayer perceptron, these 50 64-dimensional dense vectors will be stacked in chronological order to obtain a 50×64-dimensional original feature embedding matrix, which serves as the underlying digital carrier representing the user's long-term historical interaction features.

[0021] After obtaining the original feature embedding matrix, it is projected onto the intent space under orthogonal constraints to obtain an orthogonal intent representation matrix. Since the feature dimensions in the original feature embedding matrix may exhibit high collinearity (i.e., the digital preferences and office preferences of the same user overlap in the feature space), direct aggregation could lead to intent confusion. Therefore, the original feature embedding matrix is ​​input into a neural network containing an orthogonal penalty term. The core architecture of this neural network includes an intent space projection layer, which contains a learnable projection weight matrix to abstract and learn multiple specific sets of basis vectors from the original continuous sequence features. Each basis vector represents a potential long-term user intent. To ensure that the extracted intent spaces are independent and do not interfere with each other, orthogonalization constraints based on the Gram-Schmidt principle are strictly enforced during the network's forward propagation and projection calculation. Specifically, this orthogonal constraint requires that the inner product of the projection weight matrix along the column vector dimensions be as small as possible to ensure that the feature correlation between different intent spaces approaches zero. This projection calculation and orthogonal constraint process are expressed through the forward intent mapping formula and an independent orthogonal penalty term loss function, respectively, as follows: in, This represents the orthogonal intention representation matrix output after orthogonal projection processing; This represents the original feature embedding matrix generated in the previous stage; This represents the projection weight matrix in a neural network containing orthogonal penalty terms, used to map the original features to multiple predefined intent dimensions; The normalization exponential function is used to transform the projected values ​​into a probability distribution, so that the sum of each row in the matrix across all intent dimensions is one, thereby representing the degree to which different historical behaviors belong to various potential intents. This represents the orthogonality penalty coefficient, a preset hyperparameter used to balance mapping accuracy and orthogonality constraint strength. In actual configuration, it is preset to 0.01. The transpose of the projection weight matrix; Representative and Identity matrices of the same dimension; The square of the Frobenius norm of the matrix represents the forward intent mapping formula. The formula above is responsible for the core intent attention mapping, calculating the association weights between each product behavior and each potential intent in the sequence and performing a weighted summation with the original features. During the network optimization stage, a structural orthogonal penalty term is introduced through an independent loss function. When there is non-orthogonal overlap between different column vectors of the projection weight matrix (i.e., different intention basis vectors), Off-diagonal elements will deviate significantly from 0. Subtracting the identity matrix and calculating the Frobenius norm at this point will produce a very large penalty value fed back to the network. Through continuous iterative optimization of the network parameters, this penalty term forces the projection weight matrices to evolve towards a mutually perpendicular orthogonal state. If the preset number of intents is set to 5, the above original feature embedding matrix has a dimension of 50×64. After passing through a projection weight matrix with 64×5 dimensions The mapping calculation, combined with the weighted summation of the original features, is finally integrated and transformed into an orthogonal intent representation matrix of dimension 5×64. This matrix contains five mutually orthogonal sub-vectors, which represent five pure long-term intent directions of the user in the past ninety days after removing overlapping features, such as pure digital enthusiast or home life.

[0022] Finally, the orthogonal intent representation matrix is ​​subjected to nonlinear compression and instantiation of intent features to obtain long-term intent capsule clusters. In the neural network architecture, to give the above orthogonal feature vectors a more physically meaningful expression, the design concept of capsule networks is introduced. Ordinary neurons only output scalars, while capsule structures output vectors. The magnitude of the vector represents the probability of the feature's existence, and the direction of the vector instantiates the specific attribute parameters of the feature. To ensure that each sub-vector in the orthogonal intent representation matrix conforms to the probability specification of capsule vectors, a specific squeezing activation function is calculated for each sub-vector in the orthogonal intent representation matrix, thereby achieving nonlinear compression of the magnitude while maintaining the absolute invariance of the vector direction. This squeezing and instantiation process is calculated using the following formula: in, This represents the i-th uncompressed sub-vector extracted from the orthogonal representation matrix; The vector of the i-th instantiated intention capsule, after nonlinear compression, is taken as the central feature vector of the i-th capsule. Represents the input subvector The L2 norm of the vector is the Euclidean length of the vector in multidimensional space. This represents the square of the length. Analyzing this nonlinear compression formula, the right side of the equation contains a product of two parts. The second part... The operation performed is a normalization operation, used to extract the direction information of the input vector, while preserving the distribution of the intention's multi-dimensional attribute parameters after orthogonal decoupling. Part 1 It is the core non-linear scaling factor. When the input vector When the length of the input vector is very short, meaning the long-term intention it represents is extremely weak and noisy, the scaling factor will cause the output value to decay rapidly towards 0 in a quadratic manner, thus suppressing noise. When the length of the input vector is very long, meaning it represents an extremely strong and definite long-term stable preference, the value of the scaling factor will infinitely approach 1, but will never be equal to or exceed 1. For a specific numerical example, if the calculated length of the sub-vector representing the digital geek intention in a user's orthogonal intention representation matrix is ​​2.0, substituting it into the nonlinear scaling factor, we get 4 / 5, or 0.8. At this time, the intention sub-vector will be compressed proportionally, its direction remains unchanged, but the magnitude is strictly limited to 0.8, representing that the probability strength of the user having a digital geek intention is 80%. By traversing the above 5×64 orthogonal intention representation matrix and performing this nonlinear magnitude compression calculation on each of its 5 sub-vectors, the originally infinitely extended feature vector is transformed into 5 capsule vectors with magnitudes between 0 and 1 and independent directions. These five compressed capsule vectors, i.e., the central feature vectors, are encapsulated as a structured object to construct a long-term intent capsule cluster. This long-term intent capsule cluster completely abandons the messy scalar weighted stacking of traditional models, and accurately portrays the user's stable preference background in a structured and instantiated manner.

[0023] In detail, step two involves using the topology of the product knowledge graph to capture and encode transient micro-behaviors in the real-time behavioral data collected within the current session window, resulting in transient behavior vectors. Correspondingly, in real e-commerce interaction scenarios, user intent often exhibits high dynamism and suddenness. When a user generates a new demand that completely deviates from their long-term historical profile due to a specific event within a very short time, this transient micro-intent is usually only reflected in a handful of search or click behaviors within the current active period. Traditional feature construction mechanisms tend to perform indiscriminate global pooling or weighted averaging of users' long and short-term behavior sequences. This leads to such zero-sample-level sudden intents being inevitably diluted and completely submerged by the massive historical feature weights when entering deeper layers of the network, resulting in severe lag in recommendation results. To keenly capture micro-anomalies occurring at the millisecond level and break the mixed suppression effect of traditional models on long and short-term features, a process for capturing and encoding transient micro-behaviors in the real-time behavioral data collected within the current session window is introduced. By isolating pure real-time actions by predefined time boundaries and combining them with the topological structure of the product knowledge graph for in-depth mining of context and causal relationships, highly condensed transient behavior vectors can be extracted independently in an environment that excludes long-term historical noise interference.

[0024] Figure 3 This is a flowchart of step two in the product accuracy recommendation method based on multi-dimensional user profiles according to an embodiment of this application. Figure 3As shown, in a specific embodiment of this application, step two includes: step two-one, based on a preset sliding time window, parsing and removing noisy behaviors from the original log messages in the real-time behavior data, and performing entity alignment and temporal reorganization based on the product knowledge graph to obtain an instant action sequence; step two-two, performing context-aware graph embedding mapping on the instant action sequence to obtain a context graph embedding sequence; step two-three, based on a multi-head self-attention mechanism, performing intra-sequence causal relationship capture and global average pooling aggregation compression on the context graph embedding sequence to obtain a transient behavior vector.

[0025] The implementation process is as follows: Real-time behavioral data refers to the user micro-interaction log data stream continuously collected by the business front-end during the current active session through streaming technology. The preset sliding time window is a time boundary parameter used to capture this continuous data stream, and its value is set based on the average lifecycle statistics of sudden intent evolution within the platform. For example, through offline behavioral log analysis, it was found that more than 80% of sudden impulse shopping behaviors occur within a 5-minute intensive interaction period, so the preset sliding time window is specifically set to 5 minutes. During the monitoring of real-time behavioral data, the model strictly captures the original log messages according to this 5-minute sliding window and performs deep parsing to extract key metadata such as search terms, clicked product identifiers, page dwell time, and page scroll depth. Subsequently, based on the product knowledge graph that is completely consistent with step one, entity alignment is performed on the product identifiers extracted from the metadata, that is, the corresponding product topology node is accurately located in the huge graph database. While aligning features, a noisy behavior removal mechanism is introduced because users' real-time operations are often accompanied by accidental touches or passive page scrolling, which, if not filtered, will seriously pollute the purity of transient intents. At this stage, a composite threshold rule based on information entropy and physical operation characteristics is set. For example, interaction logs with a page dwell time of less than 1.5 seconds and a page scroll depth of less than 10% are judged as non-active clicks or invalid accidental touches generated by autoplay. If, within the 5-minute monitoring window, a user sequentially generates interaction messages for product identifier 3001 (baby stroller), product identifier 1045 (mechanical keyboard), and product identifier 3005 (baby diapers), and after analysis it is found that the page dwell time for product identifier 1045 is only 0.8 seconds, then the elimination mechanism is triggered to discard it directly as noise behavior. After rigorous cleaning, the remaining valid interaction behaviors are strictly reorganized in chronological order according to the timestamps to construct an instant action sequence with temporal progression logic. At this time, the length L of this instant action sequence is 2, containing two sequentially arranged interaction nodes for mother and baby products, forming the sequence data basis reflecting abrupt behavior.

[0026] After successfully constructing the instant action sequence, a context-aware graph embedding mapping is performed on the instant action sequence to obtain a context graph embedding sequence. Specifically, each action node in the instant action sequence is traversed, and with that node as the core, network topology addressing is performed in the product knowledge graph to query and obtain the deep attribute features of its first-order and second-order related nodes. For example, for product identifier 3005 in the sequence, not only are its direct attributes (diapers) extracted, but also the first-order related node (mother and baby exclusive brand) and the second-order related node (0-1 year old applicable age group) in the product knowledge graph are extracted. Subsequently, a pre-trained graph representation network model is used to map and transform these complex graph structure data. This graph representation network model adopts a graph convolutional neural network architecture, and its internal network weights are pre-trained iteratively on a massive full-network product topology graph through an unsupervised graph contrastive learning task. This graph convolutional neural network is mainly constructed from a feature input layer, multi-layered cascaded graph convolutional information transfer layers, and a linear aggregation output layer for feature abstraction. In deep graph embedding processing, the initial multidimensional attribute features of the target behavior node and the first-order and second-order association nodes obtained from the product knowledge graph are first fed into the feature input layer to construct an initial node state matrix describing the content of the local subgraph. Subsequently, this initial node state matrix, together with the local adjacency matrix that accurately represents the network connection relationship between products, is input into a multi-layered cascaded graph convolutional information transfer layer for deep computation. Within each graph convolutional information passing layer, the core processing logic strictly follows the graph message passing paradigm: First, a neighbor feature aggregation operation based on topological structure is performed. This involves using the local adjacency matrix as a spatial guide to weight and sum the features of all first- and second-order associated nodes surrounding the central target node according to the edge weights in the graph, thus collecting structured contextual information from the local graph network without loss. Next, an affine transformation of the feature space is performed, merging the aggregated neighbor features with the current features carried by the central target node itself, and multiplying it by the learnable weight matrix configured within the convolutional layer. This achieves a linear projection of features from a low-order space to a high-order graph semantic space. Finally, a nonlinear purification operation is performed, applying nonlinear activation constraints to the projected features using a linear rectified function, thereby outputting the node feature matrix that completes this round of depth update. Through continuous multi-layer forward propagation and iterative graph information aggregation, the local subgraph containing the target node and its associated topological structure is mapped into a highly condensed, high-dimensional, dense feature vector that deeply integrates global topological semantics. To ensure that the mapped vector fully reflects the immediate state at the time the action occurs, it is weighted and corrected by incorporating contextual information from when the action was triggered. Contextual information includes variables such as weather conditions at the current geographical location, the type of device the user is using, and the current marketing campaign identifier.These discrete context variables are converted into fixed-length vectors through a lookup table embedding operation, then concatenated with the previously generated dense vector of nodes, and finally subjected to nonlinear feature fusion and dimensionality reduction projection using a multilayer perceptron. To ensure strict mathematical alignment with the long-term intent capsule generated in step one in terms of feature dimensions, the output layer dimension of the multilayer perceptron is strictly configured to 64 dimensions. After the above mapping and correction operations, each behavior node in the instantaneous action sequence is converted into a 64-dimensional dense vector that integrates deep graph attributes and current environmental variables. These fixed-length vectors are then arranged and superimposed in their original temporal order to generate the context graph embedding sequence.

[0027] Finally, based on the multi-head self-attention mechanism, the context graph embedding sequence is subjected to intra-sequence causal relationship capture and global average pooling aggregation compression to obtain transient behavior vectors. To delve deeper into the profound connections and potential transitional causality among various behaviors within the immediate action sequence, the generated context graph embedding sequence is input into the multi-head self-attention layer. The core architecture of this multi-head self-attention layer includes multiple parallel attention computation heads, each equipped with a set of independent, learnable linear projection weight matrices, used to map the input sequence features into query, key, and value matrices, respectively. These projection weight matrices are obtained through offline pre-training on large-scale online session sequence data using a mask language modeling task to predict the next clicked item. The multi-head self-attention mechanism accurately captures the potential causal relationships between different micro-behaviors by calculating the attention scores between nodes within the sequence. For example, if a user clicks on a stroller first and then on a diaper, these two nodes will generate extremely high bidirectional attention weights in a specific attention head, indicating that they jointly point to a clear and strong sudden parenting intention. After obtaining the attention-weighted sequence features, a global average pooling operation is performed to compress them from a variable-length sequence into a fixed-length single dense vector. The entire causal relationship capture and feature compression process is quantized using the following formula: in, The final output is a transient behavior vector representing the current instantaneous demand. This represents the total number of valid action nodes in the instant action sequence, such as the 2 calculated in the previous example; This represents the 64-dimensional feature vector of the j-th node in the sequence after graph embedding and context correction. This represents the scalar value of the attention weight calculated by the i-th node for the j-th node; and These represent the query vector and key vector generated after linear projection mapping of the i-th node and the j-th node, respectively; Then it represents the key vector of the k-th node in the sequence used to calculate the normalized denominator; The hidden layer dimension represents the vector features; here, it is preset to 64 to maintain consistency with the previous processing. Represents the natural exponential function, used to nonlinearly map attention scores in the real number domain to a positive number space; This represents the layer normalization operation function. The second equation on the right illustrates the core calculation mechanism of the attention weights: first, using the query vector... Transpose of the key vectors of all nodes in the sequence Perform a matrix dot product operation to measure the semantic relevance between the i-th micro-behavioral node and other behavioral nodes in the sequence; then divide the dot product result by a scaling factor. That is, the square root of 64 equals 8, which effectively prevents the dot product value from being too large and causing the subsequent exponential function calculation to fall into the gradient saturation dead zone; finally, by normalizing the exponential function operation, the relevance scores of all nodes are transformed into probability distribution weights whose sum is strictly 1. After accurately obtaining the interaction probability weights between each node, the first equation performs global average pooling and feature aggregation compression: firstly, attention weights are used... With the corresponding original embedded features A weighted summation is performed to calculate the updated context sequence representation for all nodes. Then, the reconstructed features of the entire sequence are averaged. This dynamic weighted summation mechanism is essentially a pooling process with causal attention bias. It significantly amplifies node features with strong mutual corroboration in multidimensional space while effectively suppressing isolated and disordered secondary node features. Finally, a layer normalization operation is applied to the aggregated result obtained from the weighted summation. The feature vectors are smoothed for both mean and variance to prevent covariance shift in the feature values ​​during subsequent transmission in deeper networks. After these rigorous steps, the original... A matrix composed of 64-dimensional vectors was successfully extracted and losslessly compressed into a single, fixed-length 64-dimensional transient behavior vector. This transient behavior vector highly condenses the user's immediate, pure intent within the past 5 minutes, freeing them from the constraints of their historical profile.

[0028] In detail, step three: Based on an iterative dynamic routing protocol, the intent coupling metric is performed on the transient behavior vector and each intent capsule in the long-term intent capsule cluster to obtain the intent coupling coefficient distribution. It is understandable that in the data processing flow of recommendation algorithms, user profiles constructed from long-term historical sequences typically possess multi-dimensional stable features, while real-time collected short-term micro-behaviors often contain dynamic and sudden new demands. Traditional feature fusion schemes typically employ simple multilayer perceptron splicing or static weighted summation mechanisms to crudely fuse these two time-dimensional features directly within the same latent dimensional space. This approach cannot accurately assess the degree of fit between short-term sudden intents and specific long-term intent subspaces, easily leading to weak transient micro-features being submerged and diluted by the large historical weights accumulated over time, thus causing mutual tug-of-war between long-term and short-term features and resulting in a severe lag in recommendation results. To accurately quantify the correlation strength between short-term burst features and multi-dimensional long-term profiles without disrupting the established orthogonal decoupling space, an iterative dynamic routing protocol is introduced. This protocol performs intent coupling metric on each intent capsule within the long-term intent capsule cluster to obtain the intent coupling coefficient distribution. Through a bottom-up, multi-round negotiation and dynamic feedback mechanism, the activation probability of the current transient behavior in different long-term intent subspaces can be accurately measured.

[0029] In one specific embodiment of this application, step three includes: calculating the projection intensity of the transient behavior vector with the central feature vector of each capsule in the long-term intent capsule cluster to obtain the initial spatial alignment score; performing weight iterative updates on the initial spatial alignment score based on routing consistency according to the consensus iteration mechanism of the iterative dynamic routing protocol to obtain the iterative route logarithmic weight; and performing intent coupling degree normalization and distribution mapping on the iterative route logarithmic weight to obtain the intent coupling coefficient distribution.

[0030] The implementation process is as follows: The initial spatial alignment score is obtained by calculating the projection intensity of the transient behavior vector with the central feature vectors of each capsule in the long-term intention capsule cluster. After rigorous processing in the preceding steps, a single, fixed-length 64-dimensional transient behavior vector and a long-term intention capsule cluster consisting of five mutually orthogonal capsules, compressed nonlinearly and all with a dimension of 64, have been obtained. Based on this, each independent capsule unit in the long-term intention capsule cluster is traversed, and the central feature vector corresponding to each capsule is extracted. Since these five capsule vectors have been orthogonally constrained in the previous spatial projection, they represent mutually independent long-term preference dimensions. Subsequently, the captured 64-dimensional transient behavior vector is subjected to a high-dimensional dot product operation with the central feature vectors of these five capsules. The dot product operation geometrically measures the magnitude of the original projection components of the current instantaneous appeal in each long-term intention subspace. In one specific embodiment of this application, the projection intensity of the transient behavior vector is calculated one by one with the center feature vector of each capsule in the long-term intent capsule cluster to obtain the initial spatial alignment score, including: calculating the projection intensity using the following formula, wherein the formula is: in, For transient behavior vectors, For each capsule, the central feature vector is... Let be the i-th component in the initial spatial alignment score. Analysis of the dot product formula reveals that if the micro-intent represented by the transient behavior vector highly matches a specific long-term intent subspace, the angle between them in the high-dimensional space will approach 0, and the dot product result will significantly increase. Conversely, if the two are semantically unrelated or even mutually exclusive, based on the base property of the previous orthogonalization process, the dot product value will approach 0 or even be negative. The results of all five dot product operations are collected and combined according to the original order of the capsules to construct an initial spatial alignment score vector containing five numerical elements. If the current transient behavior vector mainly reflects demands related to maternal and infant products, and the five long-term intent capsules represent digital products, sports and outdoor activities, maternal and infant products, household daily necessities, and book reading, the initial spatial alignment scores obtained after dot product calculations are 0.12, 0.05, 1.85, 0.34, and -0.02, respectively. This score initially reflects the basic correlation strength of the current transient behavior in each intent subspace.

[0031] After obtaining the basic projection intensity, the initial spatial alignment score is updated with weights based on routing consistency through a consensus iteration mechanism of the iterative dynamic routing protocol to obtain the iterative route logarithmic weights.

[0032] Example 1: The routing logarithmic weight matrix is ​​a state matrix used to record and track the dynamic connection strength between transient behaviors and long-term intent capsules. Its dimension is consistent with the number of long-term intent capsules, i.e., it contains 5 scalar elements. During the initialization phase, the initial value of this routing logarithmic weight matrix is ​​directly assigned to the component corresponding to the previously calculated initial spatial alignment score. This initialization method ensures that subsequent iterations are built upon a valid starting point of the basic projection strength. Subsequently, the core consensus iteration mechanism of the iterative dynamic routing protocol is initiated. Unlike the static, one-time feature allocation in traditional feedforward neural networks, this iteration mechanism, through bottom-up multiple negotiations, enables the underlying transient features to dynamically find the best-matching high-level intent capsule. In each specific iteration, the logarithmic weights are non-linearly adjusted based on the context dynamic gain coefficient of the current session. This adjustment and update process follows the following formula: In this formula, This represents the logarithmic weight of the route for the i-th long-term intent capsule after the k-th iteration update; This represents the logarithmic weight of the route obtained after the previous round, i.e., the k-1th iteration; This represents the initial spatial alignment score component obtained from the previous calculation; Represents the context-dynamic gain coefficients that are adaptively generated based on the current session environment. Context-dynamic gain coefficients The acquisition method involves inputting previously acquired discrete contextual variables such as current weather and device type into a lightweight two-layer fully connected network for dimensionality reduction and fusion, and mapping them to scalars between 0 and 1 through a nonlinear activation function. If the current context deviates significantly from the historical normal environment, the value will tend towards 1, thereby amplifying the weight increment of the current transient behavior. Multiple accumulation calculations are performed using a preset number of iterations, for example, three. This continuous accumulation method effectively strengthens intent channels with high consistency; that is, channels with high initial scores expand rapidly after being multiplied by the gain coefficient and accumulated multiple times, while relatively suppressing low-correlation noise channels. If the initial score of the highest-scoring maternal and infant products channel is 1.85, and the contextual dynamic gain coefficient is 0.8, its value becomes 3.33 after the first iteration, 4.81 after the second iteration, and climbs to 6.29 after the third iteration; while the channel with a score of only 0.12 only increases to 0.408 after three iterations. After these three complete consensus iterations, the final output is an iterative routing log weight matrix that clearly represents the intent direction.

[0033] Example 2: Specifically, in the scenario of precise product recommendation based on multi-dimensional user profiles, handling the intersection and update of long-term and short-term intentions is the core hub of the entire algorithm chain. The data processing logic in Example 1 often adopts a linear accumulation method based on a fixed step size when updating intention features. That is, it directly adds the initial spatial alignment score of the current instantaneous behavior in each intention subspace to the historical weights proportionally. Although this approach is extremely simple in algebraic implementation, in the real high-concurrency interaction environment of e-commerce, it directly smooths out several key relationships that determine the success or failure of recommendations. First, this method ignores the semantic mutual exclusion and competition between intention capsules. If a user whose long-term profile stably points to digital geeks suddenly searches for products such as baby strollers and newborn diapers in the current session, this new transient intention not only increases the weight of the maternal and infant dimension capsule, but also inevitably creates a significant squeezing and substitution effect on the original high-weight digital intentions in terms of business logic. If scores are accumulated independently, multiple mutually exclusive intentions are often unreasonably inflated simultaneously, leading to a multi-peak, ambiguous state in the underlying routing weights. This can result in industrial-level false recall disasters, such as mistakenly recommending baby gift sets and high-end graphics cards to the homepage simultaneously during subsequent matching and ranking. Secondly, a uniform, fixed update step size masks significant differences in session context. Adding items to a cart before checkout and aimless scrolling in random browsing have vastly different levels of confidence and urgency in representing intent. A fixed step size lacking context sensitivity causes slow intent switching in high-transaction scenarios, while excessive updates in noisy scenarios lead to frequent profile fluctuations. Finally, this linear accumulation mechanism completely fails to consider the uncertainty of the initial score distribution. When users present highly chaotic browsing sequences, the initial spatially aligned score distribution is usually extremely divergent. Repeatedly accumulating scores using a fixed formula only blindly amplifies noise at the numerical level, failing to achieve stable convergence of true intent. To completely eliminate the dual challenges of inconsistent speed and stability caused by the lack of mutual exclusion control, scene awareness, and entropy constraints, the coarse-grained scorer is upgraded to a routing control mechanism adapted for industrial-grade real-time recommendations. This involves introducing iterative updates of routing consistency weights based on competition suppression and dynamic gain control. This mechanism quantifies and suppresses the relative strength of intent channels and adaptively adjusts the convergence speed according to real-time scene characteristics. This ensures that in complex cross-category jumping and emergency shopping scenarios, new dominant intents quickly emerge while historically conflicting intents are effectively suppressed, resulting in a highly recognizable and pure intent signal.

[0034] In another specific embodiment of this application, based on the consensus iteration mechanism of the iterative dynamic routing protocol, the initial spatial alignment score is iteratively updated with weights based on routing consistency to obtain the iterative route logarithmic weights, including: The initial spatial alignment score is estimated using a cross-capsule competition suppression factor to obtain a competition suppression coefficient matrix. It should be understood that the multiple intent channels triggered in the current session do not coexist in parallel, but rather exhibit a distinct competitive structure with varying strengths. Without pre-estimating which intent channel has overwhelming dominance and mathematically converting it into a suppressive force against other channels, subsequent weight updates cannot proactively reduce the contribution of mutually exclusive channels. In this implementation, the initial spatial alignment score sequence is first obtained by performing a dot product operation between the transient behavior vector and the feature vectors of each long-term intent capsule center in the preceding steps. To transform these absolute values ​​into relative competitive relationships, the initial spatial alignment score is normalized. This normalized alignment probability is precisely calculated using the following formula: The formula symbols are analyzed. This represents the normalized alignment probability calculated for the i-th intention capsule; This represents the i-th score component in the initial spatial alignment score matrix of the input; This represents the j-th score component in the initial spatial alignment score matrix; This represents the total number of intent capsules contained in the long-term intent capsule cluster, which was set to 5 in the previous processing. This formula, through its exponential amplification property, initially differentiates high and low scores and forces the sum to be equal to one. After obtaining the probability distribution of all 5 channels, information entropy theory is further introduced to characterize the dispersion of current session intent. The normalized information entropy is calculated using the following formula: In this formula, Normalized information entropy represents the overall distribution of the initial spatial alignment scores; This is the normalized alignment probability obtained above; The total number of capsules is 5; The normalized denominator used for global scaling of the entropy value makes the final calculated information entropy more accurate. It can stably fall within a preset standard range of 0 to 1, ensuring consistent suppression when facing capsule matrices of different sizes. The closer the entropy value is to 0, the more focused and clear the user's transient intent; the closer the entropy value is to 1, the more chaotic the current behavioral sequence is, lacking a clear focus. Combining the above probability and information entropy, a competition suppression coefficient is finally generated. The competition suppression coefficient is calculated using the following formula: In this formula, This represents the i-th component in the final output competition suppression coefficient matrix; the remaining symbols are completely consistent with the aforementioned meanings. The right-hand side of the multiplication sign, as the core entropy suppression term, plays a crucial regulatory role. When a behavior at a certain instant is highly directional, it leads to information entropy... If the entropy is very low, such as 0.2, the entropy suppression term is as high as 0.8, and the normalized probability of the dominant channel will be largely preserved, thus gaining a highly dominant competitive advantage. Conversely, if the actions are sporadic and scattered, causing the information entropy to soar to 0.9, the entropy suppression term is only 0.1, and the competitive suppression coefficients of all channels will be drastically compressed. Extrapolating with specific values, if the five long-term intention capsules represent digital devices, sports, mother and baby, home, and books, and the current transient behavior is mainly aimed at mother and baby, the calculated initial spatial alignment scores are 0.12, 0.05, 1.85, 0.34, and -0.02, respectively. After exponential normalization calculation, the probability of the mother and baby channel... It will occupy an absolutely high position. Substituting these 5 probabilities into the information entropy formula, due to the extremely steep distribution, the calculated... It is only 0.25. Substituting this into the inhibition coefficient formula, (1-0.25) equals 0.75 as the multiplier, which is the competitive inhibition coefficient corresponding to the mother-infant channel. It will still maintain an extremely high value, such as 0.6, while the suppression coefficient of the digital channel... It then approaches 0.02. Through this series of rigorous algebraic transformations, each intention channel is no longer regarded as an independent accumulation unit, but a dynamic competitive structure is precisely formed that determines which is more like the intention graph and which should be suppressed. This not only significantly improves the recognizability of the dominant intention in subsequent games, but also reduces the intensity of participation of invalid random channels in the iteration from the source.

[0035] Based on real-time contextual feature vectors, context-aware adaptive gain modulation is applied to the initial spatial alignment score to obtain a dynamic gain operator. Correspondingly, the same initial spatial alignment score represents different levels of effectiveness and business urgency at different business touchpoints. A precise click at the top of the search results page, a combined checkout on the shopping cart page, and a brief sliding exposure in the event venue's waterfall flow correspond to drastically different decision-making depths in essence. If a uniform, fixed value is continued to be used as the update step size, the online recommendation system will inevitably miss the best opportunity to switch intentions during high-value transaction windows, or experience frequent system oscillations amidst low-value page browsing noise. To endow the underlying routing calculation with context-aware capabilities, real-time collected contextual information and scores are fed together into the entry control modulation structure. First, in the data acquisition stream at the system front end, real-time business environment variables at the time of the current click are extracted, including the specific page level identifier, the time segment, whether a specific marketing activity identifier is currently displayed, the traffic source channel, the cumulative dwell time on the current page, and the continuity of user operations, among other discrete information. This information is then transformed into a dense real-time contextual feature vector through a pre-defined dictionary mapping and one-hot encoding rules. Subsequently, the context feature vector and the initial spatial alignment score from the preceding sequence are input into a pre-constructed neural network gating function. The calculation of this gating function is expressed in a normalized form using the following formula: The mathematical symbols in the formula are broken down in detail. This represents the dynamic gain operator component of the output for the i-th intention capsule channel; This represents a gated activation function with smooth compression characteristics, employing a logistic function to ensure that the output gain operator is strictly limited to the numerical range of 0-1. This represents the dense real-time context feature vector generated by the aforementioned encoding; This represents the real-time context feature mapping weight matrix, which is responsible for reducing the dimensionality of multi-dimensional context features and aligning them to a one-dimensional scalar space. This represents the initial spatial alignment score of the i-th channel; Represents the scalar mapping weights for the initial spatial alignment score; This represents the network bias term set to balance the feature shift. Wherein, , as well as All of these are pre-trained and learned from massive amounts of anonymized historical transaction logs, with the goal of converting them into supervisory signals, through end-to-end backpropagation. Under this gating mechanism, if a target user stays on the shopping cart checkout page for a relatively long period of 60 seconds and then clicks on a recommended product, the encoded context vector... After weight matrix After mapping, a very large positive scalar value is output. This value is added to the initial score, and then subjected to nonlinear compression by a gated activation function, resulting in the output dynamic gain operator. It could be as high as 0.95. This means that effective signals in extremely high conversion chains will be rapidly amplified, prompting the system to confirm the intention switch with a near 100% step size within an extremely short window of a few milliseconds. Conversely, if the user is merely performing a rapid, aimless pull-down operation in the main promotional venue, with each exposure slide lasting only 0.5 seconds, the result after the context vector mapping will be a very small value or even a negative number. After processing by the gating function, the output dynamic gain operator... It may be as low as 0.1. At this point, no matter how high its initial alignment score is, the magnitude of its route update will be actively weakened and forcibly locked at an extremely low level, thereby effectively preventing meaningless image jitter.

[0036] Based on the dynamic gain operator and the competition suppression coefficient matrix, a nonlinear competitive route iterative update is performed on the initial spatial alignment score to obtain the iterative route logarithmic weights. Finally, the previously mined intention mutual exclusion relationships and context-sensitive intelligence are substantially injected into the underlying numerical update loop. By constructing a complex algebraic system encompassing multiple constraints, the linear scoring mode where each channel operates independently, as in traditional mechanisms, is broken. In this implementation phase, the historical logarithmic weights of each channel from the previous round are extracted, and the competition suppression coefficient matrix is ​​used as an explicit penalty interference term, combined with the dynamic gain operator to reconstruct the entire iterative formula. The calculation of the nonlinear competitive route update is performed using the following formula: A deep analysis of the variables in this core iterative formula is conducted. This represents the logarithmic weight of the new iterative route generated for the i-th intent capsule after the k-th complete iteration; This represents the historical logarithmic weights retained at the (k-1)th iteration (or the initial state); This represents the dynamic gain operator, which is precisely generated in the previous step and lies between 0 and 1. This represents the competition suppression coefficient component that belongs to other competing channels (the j-th channel) and is calculated in the preceding steps; This represents the initial spatial alignment score of the corresponding competing channel; the summed term represents a crucial physical concept: the total inhibition penalty jointly imposed on the current i-th channel by all competing intent channels except the i-th channel. Each update is no longer simply adding its own score. Instead, you must first subtract the total suppression penalty from other competitors from your own score. If the competitors in a certain channel are extremely strong (i.e., a certain competitive channel has...), then... and If both are very high, then the difference within the parentheses in the formula will shrink drastically or even become negative. This means that the channel will not only fail to gain a weight increase, but will also be subject to active numerical reduction. After obtaining this difference, it is necessary to use the outer dynamic gain operator. The weights are scaled proportionally according to the urgency of the current business scenario, and then accumulated into the historical weights. For example, in the first round of iterations where k equals 1, the base weight from the previous round is used for the first intent channel representing the digital value. With a score of 0.12, it faces a dominant third mother-infant channel (score 1.85, inhibition coefficient as high as 0.6). When calculating the total inhibition penalty, the pressure exerted by the mother-infant channel alone is as high as 0.6 × 1.85, approximately 1.11. The difference within the parentheses is 0.12 - 1.11 = -0.99. Even on a high-conversion settlement page, such as when the dynamic gain operator is 0.8, the new weight of this digital channel in this round... The result becomes 0.12 + 0.8 × (-0.99), drastically decreasing to -0.672. This achieves precise and slow suppression of historically high-weighted but currently mutually exclusive intentions. Conversely, the dominant third mother-infant channel has a score as high as 1.85, while the other four channels, due to their extremely low probabilities, have negligible suppression coefficients, thus suffering only a tiny fraction of a percent in total suppression penalty. The difference within the parentheses remains almost constant at 1.85. Under the same 0.8 dynamic gain operator amplification, its new weights... The value will rapidly surge from 1.85 to over 3.29. After several non-linear competitive iterations (e.g., three), the initially similar channel weights will quickly become polarized. The updated set of iterative route logarithmic weights will be directly fed into the exponential global normalization layer in subsequent steps. Due to the triple rigorous modulation of mutual exclusion competition, entropy constraint noise reduction, and contextual scene intensity, the downstream signal is no longer a fuzzy preference resulting from linear accumulation and multi-peak confluence, but a highly sharp and clearly defined pure intent signal. This eliminates the redundant computational resource consumption of a large number of invalid, high-weight, noisy channels during the online inference stage, laying the data foundation for subsequent accurate comparison of dynamic mutation thresholds, determination of whether to trigger independent capsule instantiation and traffic splitting, and ultimately locking in the most suitable product sequence from hundreds of millions of candidate products.

[0037] After multiple rounds of dynamic weight reinforcement, the logarithmic weights of the iterative routes are normalized and mapped to their intent coupling coefficient distribution. The 5-element iterative route logarithmic weight matrix output by the consensus iteration mechanism is then input into the global normalization layer. The global normalization layer is a mathematical processing module specifically designed for probability space mapping. It does not contain learnable weight parameters that require complex updates via backpropagation; its core architecture only includes parallel exponential operation units and division allocation units for summation. The global normalization layer transforms real-valued weights of arbitrary ranges into statistically significant probability distributions, while further leveraging exponential amplification to highlight the dominant intent channel. This normalization and mapping process is performed according to the following formula: In this normalization formula, Represents the intent coupling coefficient for the i-th long-term intent capsule in the final output; This represents the logarithmic weight of the route from the input to the i-th iteration of the global normalization layer; This represents the total number of long-term intent capsules, which is 5 in this practical scenario. This represents the iterative route logarithmic weight of the j-th channel out of all 5 channels; Represents the operational unit of exponential functions with the natural constant as the base; This represents a global summation operation on all values ​​after exponential mapping. A deeper analysis of the formula's operational logic reveals that, since the exponential function is monotonically increasing and always positive, the exponential mapping ensures that all iterative logarithmic weights that might have negative values ​​are unconditionally converted to positive numbers greater than 0. Subsequently, by calculating the proportion of each positive component in the total weight sum, the sum of all five output values ​​is forced to be always equal to 1. This division allocation mechanism not only achieves strict constraints on the numerical range but, more importantly, transforms the original abstract weights into relative values ​​with probabilistic significance, thus generating the intention coupling coefficient distribution. If the differences between the five weight values ​​obtained after iterative route updates are significantly widened, the leading weight will experience explosive growth after exponential operation, thereby occupying the vast majority of the probability share in the summation and normalization stage. As mentioned earlier, the weight of the maternal and infant products channel, which increased to 6.29 after three iterations, may have a final intent coupling coefficient as high as 0.89 after the exponential and proportional mapping of the global normalization layer; while the coupling coefficients corresponding to the other four lower weights are forcibly compressed to an extremely low level, with a sum of only 0.11. The final distribution of the output intent coupling coefficients directly and accurately indicates the absolute activation probability of the currently generated instantaneous micro-behavior in each multidimensional long-term intent subspace.

[0038] In detail, step four: Based on the comparison between the peak value of the intent coupling coefficient distribution and the dynamic mutation threshold, mutation detection is performed on the transient behavior vector. If a mutation is detected, it is instantiated as an independent capsule and incorporated into the long-term intent capsule cluster; otherwise, it is fused with the momentum of the long-term intent capsule to form a matching capsule, generating a fused activation capsule set. It is understandable that in complex data interaction scenarios such as e-commerce, recommendation algorithms inevitably face a severe conflict between long-term stable preferences and immediate micro-level demands when processing user behavior preferences. When a user exhibits a rare micro-behavior completely unrelated to their past long-term profile due to specific external stimuli, most existing feature fusion mechanisms often lack clear dynamic boundary judgment criteria. They typically directly perform a simple weighted average or pooling of these extremely weak sudden features with a large historical high-dimensional feature matrix. This indiscriminate global aggregation method mathematically leads to a severe spatial dilution effect, not only completely submerging and suppressing sudden new demands with high conversion value from historical data, but also subjecting the original stable historical profile features to unnecessary noise pollution. To completely block the suppressive effect of massive historical profiles on subtle, sudden intentions, and to avoid the mutual pull and confusion between feature spaces of different time dimensions, a comparison is introduced between the peak value of the intention coupling coefficient distribution and the dynamic mutation threshold. This allows for mutation detection and judgment of transient behavior vectors, followed by controlled traffic diversion. By constructing an isolation and update barrier with rigorous algebraic logic, regular extended behaviors can be smoothly integrated, while anomalous mutation behaviors can be independently instantiated and stripped, thereby ensuring the purity of both long-term and short-term intentions and the immediate and accurate response of the final recommendation results.

[0039] In one specific embodiment of this application, step four includes: extracting the peak value in the distribution of intent coupling coefficients and comparing the peak value with the dynamic mutation threshold using a step function to obtain a mutation state identifier and a target activation index; performing controlled splitting on the long-term intent capsule cluster based on the mutation state identifier, wherein when the mutation state identifier is 0, the transient behavior vector and the capsule corresponding to the target activation index are fused by momentum to obtain a matching capsule, and when the mutation state identifier is 1, it is instantiated as an independent capsule, and the matching capsule or independent capsule is assembled into a reconstructed capsule feature matrix; and performing projection culling and boundary smoothing on the redundant components in the reconstructed capsule feature matrix that are collinear with the other capsules due to the injection of new features to obtain a fused activation capsule set.

[0040] The implementation process is as follows: First, the intent coupling coefficient distribution, precisely quantized from the previous steps, is received. This distribution is a probability array with a strict sum of one, where each numerical component precisely represents the activation response level of the current transient behavior in each independent dimension of the long-term intent capsule cluster. The distribution array is then globally traversed and sorted to precisely extract the maximum response probability value, i.e., the peak value, and the corresponding index of this peak value in the distribution array is recorded simultaneously. This index is the target activation index, which directly points to the historical capsule vector in the long-term intent capsule cluster that best matches the current micro-behavior. The historical capsule vector essentially represents the user's long-term solidified preferences in a specific dimension. After peak extraction, a dynamic mutation threshold is obtained to judge whether the current behavior constitutes an anomalous mutation. The dynamic mutation threshold is dynamically calculated by combining the historical variance of the overall market fluctuations. Specifically, the intention switching frequency of the global user group in different time periods and business scenarios is continuously monitored in the background, and the mean and variance of the intent transition probability for the current session window are calculated and extracted. The mean of this benchmark is subtracted by half of the variance to obtain a highly sensitive dynamic mutation threshold, which adaptively filters out the normal fluctuation noise of the market. After obtaining the peak value and the dynamic mutation threshold, a Herveyd step function is introduced for rigorous logistic algebraic comparison to generate a binary mutation state identifier. The entire detection and comparison process is quantified using the following formula: in, An array representing the distribution of intent coupling coefficients from preceding inputs; Represents the maximum value extraction operator; This represents the peak value of the maximum response probability extracted from the distribution; This represents the dynamic mutation threshold obtained by combining the variance of market volatility mentioned above; This represents the Herveside step function, whose mathematical property is that when the input parameter within the parentheses is greater than 0, the function outputs 1, and when the input parameter is less than or equal to 0, the function outputs 0. The mutation status flag, representing whether an intentional mutation has occurred in the final output representation, takes only two Boolean values: 0 or 1. If a user's transient behavior vector points to a completely new maternal and infant care need, and their historical capsule only contains irrelevant dimensions such as numbers and motion, then the distribution of their intention coupling coefficient will be extremely flat, with the extracted peak value being... It may be as low as 0.25. This is similar to the dynamic mutation threshold calculated for the current time period. If the value is 0.40, substituting both into the formula, 0.40 - 0.25 = 0.15, which is greater than 0. After processing by the step function, the output mutation state indicator is 1, clearly indicating that this is a sudden behavior that deviates from the historical trajectory. Conversely, if the peak value is as high as 0.89, and 0.40 - 0.89 is a negative value, the step function output mutation state indicator is 0, indicating that the current behavior is a reasonable continuation of a long-term historical intention.

[0041] Controlled splitting of long-term intent capsule clusters is performed based on mutation state identifiers. This process utilizes the Boolean value characteristics of mutation state identifiers to isolate and distribute update paths for long-term capsule clusters at the underlying algebraic operation level. When the mutation state identifier is 0, a stable update branch is triggered. At this time, the extracted target activation index is used to extract the corresponding historical capsule vector from the long-term intent capsule cluster as the update base. To preserve the stability of historical preferences while incorporating micro-behavioral features, an exponential moving average algorithm is used to perform momentum-weighted fusion of this historical capsule vector with a 64-dimensional transient behavior vector to generate a matching capsule with enhanced features. When the mutation state identifier is 1, a mutation isolation branch is triggered. At this time, momentum fusion is immediately blocked, and a completely new dimensional space is directly extended outside the existing feature space. The transient behavior vector is fully mapped and filled into the new space, instantiated as a purely independent capsule, completely isolating historical weight dilution. Regardless of whether a matching capsule or an independent capsule is generated, it is ultimately recombined with the remaining capsules to form a reconstructed capsule feature matrix. The algebraic operation process of the two-way splitting and assembly is specified by the following formula: in, This represents the matching capsule vector generated by the stable update branch; Indicates a mutation state identifier; This represents the momentum preservation factor in the exponential moving average algorithm, with a default value of 0.8. The original historical capsule vector extracted from the target activation index; Represents the transient behavior vector; This represents the independent capsule vectors generated by the mutation isolation branches. Seamless decomposition at the computational graph level is achieved by cleverly utilizing mutually exclusive algebraic switches. When it is 0, Forced reset to 0, The linear momentum fusion of 80% historical features and 20% new features takes effect and is executed; when When it is 1, Forced reset to 0, The values ​​of the transient behavior vector are directly copied to generate independent capsules that retain 100% pure mutation features. Subsequently, the updated or newly added capsules are aligned and concatenated with the original capsule clusters to form a reconstructed capsule feature matrix.

[0042] Finally, redundant components collinear with other capsules in the reconstructed capsule feature matrix due to the injection of new features are projected and smoothed to obtain a fused activated capsule set. Although the controlled diversion mechanism achieves feature isolation or fusion on a macroscopic level, in a high-dimensional continuous space, whether it is a matching capsule that changes direction or a completely new independent capsule injected across boundaries, its newly generated features may have a weak linear correlation with other unupdated basic intent capsules in geometric representation. This spatial drift phenomenon leads to redundant components that overlap between different intent dimensions, destroying the independent orthogonality foundation built in the early stage, and thus causing secondary confusion in intent determination in subsequent cross-matching of the deep network. To completely eliminate the hidden spatial coupling, a strict Schmitt orthogonalization rule is introduced for local fine-tuning. Specifically, the update target in the reconstructed matrix is ​​extracted and orthogonally projected onto the vector subspace where each of the other unchanged basic intent capsules is located. The resulting vector is the collinear redundant overlapping component between the two. Subsequently, these overlapping components are precisely subtracted from the original vector of the update target to forcibly separate feature associations and ensure strict orthogonality of each dimension. The projection rejection process is expressed by the following formula: ,in, This represents the newly orthogonalized capsule vector after projection removal. This represents the input update capsule vector that has not yet been fine-tuned; Represents the total number of the remaining benchmark capsules; Represents the vector of the remaining j-th baseline intention capsule; Scalar representing the inner product between two vectors; Represents the square of the magnitude of the reference capsule vector; the fractional part is... Multiplication calculates the exact projection components of the input vector onto the reference vector direction; the summation symbol is used to accumulate and integrate redundant projections across all reference dimensions. Subtracting this set of redundant components from the contaminated input vector outputs an orthogonalized feature vector that remains perpendicular to all other reference dimensions. After precise removal, a nonlinear norm operation is applied to all orthogonally fine-tuned dense vectors for boundary smoothing, ensuring that their magnitudes strictly conform to the physical probability expression specification between 0 and 1. After smoothing constraints are applied, all 64-dimensional high-frequency pure vectors are encapsulated and packaged into a standardized output object, thereby generating a fusion activation capsule set that combines orthogonal independence and intent integrity.

[0043] In detail, step five involves using the candidate product embedding features extracted from the product knowledge graph as the query object, performing cross-attention matching and scoring on the fused activation capsule set, and then sorting and truncating the results in descending order based on the comprehensive matching scores to generate a product recommendation list. In other words, after the preceding deep feature decoupling and mutation isolation processing, a high-quality representational foundation that balances users' long-term stable preferences with immediate micro-level sudden intentions has been constructed. However, existing recommendation decision-making stages typically use simple vector dot products or Euclidean distance to measure the similarity between overall user features and a massive number of products. This globally consistent matching method ignores the fine-grained alignment relationship between product features and the user's multi-dimensional preference space, easily leading to further deviations or compromises in the complex game between long-term and short-term intentions, making it difficult to accurately cover the user's core needs. To fully establish a precise matching and mapping channel between multi-dimensional intent capsules and a vast array of products, a processing step is introduced that uses the embedded features of candidate products extracted from the product knowledge graph as the query object. This step involves cross-attention matching, scoring, and sorting the fused activation capsule set to ensure that each candidate product receives a reasonable evaluation in the most matching intent dimension, thereby ultimately generating a product recommendation list that highly matches the user's current real and diverse needs.

[0044] Figure 4 This is a flowchart of step five in the product accuracy recommendation method based on multi-dimensional user profiles according to an embodiment of this application. Figure 4 As shown, in a specific embodiment of this application, step five includes: step five-one, retrieving nodes and their first-order neighborhood attribute features of the current target user's candidate product set from the product knowledge graph, and performing deep structured embedding extraction and dimension alignment to obtain candidate product embedding features; step five-two, using the candidate product embedding features as query terms, performing multi-capsule cross-attention aggregation scoring on the fusion activation capsule set to obtain a comprehensive matching score; step five-three, globally sorting the values ​​of each candidate product in the comprehensive matching score in descending order, and extracting the top-ranked product entity sequence according to the preset recommendation list capacity, and encapsulating it after business filtering to generate a product recommendation list.

[0045] The implementation process is as follows: First, a candidate product set is obtained based on a preset recall strategy. The preset recall strategy is a set of rules used to initially filter out absolutely irrelevant products and define the matching range from a global product pool of hundreds of millions. This strategy integrates the results of multiple recall channels. For example, it uses collaborative filtering to recall products purchased by similar users, combines inverted index retrieval to recall products related to the user's current search terms, and supplements high-conversion products for the current period with trending product recall. To balance the processing efficiency of subsequent deep learning and the diversity and coverage of the recommendation list, a quota parameter is set to truncate the total number of products recalled from multiple channels. For example, the top 1000 potential products are extracted as the candidate product set for the current target user. After obtaining these 1000 candidate products, their unique identifiers are used as query keys, and they are again integrated into the product knowledge graph consistent with the previous steps. In the graph database, not only are the core nodes of these 1000 candidate products precisely located, but their first-order neighborhood attribute features are also extracted outwards. First-order neighborhood attribute features encompass all topological entities directly connected to the product node, such as its exact category tree node, associated brand matrix nodes, and complementary product nodes. Subsequently, a graph convolutional neural network (GCNN) is used to deeply structure and extract the rich neighborhood attributes of these 1000 candidate product nodes. This GCNN architecture includes an input layer, multiple graph convolutional message passing layers, and a linear output layer for dimensionality reduction and alignment. In the message passing layer, the core processing logic is a graph message passing mechanism: the initial features of the central product node and the features of all its first-order neighbor nodes are weighted and aggregated according to edge weights, then multiplied by a learnable weight matrix built into the model, and subjected to a non-linear transformation using a linear rectified activation function. The weight bias parameters within the network are pre-obtained offline using graph contrastive learning self-supervised training on the entire product graph. Through this layer-by-layer aggregation, each candidate product not only includes its own business attributes but also deeply integrates its global structured contextual associations within the graph topology network. To enable seamless dot product operations with the fusion activation capsule set from the preceding output in the same high-dimensional mathematical space, the output layer parameters of the graph convolutional neural network are strictly set to enforce dimension alignment, ensuring that each extracted product feature vector is precisely reduced in dimensionality and fixed to a dense 64-dimensional feature vector. These 1000 depthwise extracted and precisely aligned 64-dimensional vectors are then stacked sequentially in a matrix row format, ultimately constructing a 1000×64 candidate product embedding feature matrix.

[0046] After the structured extraction of product features, the candidate product embedding features are used as query terms. A multi-capsule cross-attention aggregation scoring method is applied to the fused activation capsule set to obtain a comprehensive matching score. The fused activation capsule set generated in the previous processing may contain five orthogonally smoothed long-term intent capsules and one new capsule independently instantiated due to a sudden intent. That is, the set contains a total of six 64-dimensional standardized capsule vectors, each representing an independent subspace of the user's intent. At this stage, a multi-capsule cross-attention model is constructed. This model no longer uses the crude method of compressing all user features into a single global vector. Instead, it treats each product feature vector in the candidate product embedding feature matrix as an independent query term, and considers the six intent capsules in the fused activation capsule set as a whole as key and value terms. For a specific candidate product query term, the model first calculates the affinity weights of the product to different intent dimensions. Specifically, the 64-dimensional query vector of the candidate product is multiplied by the six intent capsule vectors. The dot product result reflects the degree of projection matching between the product feature and the corresponding intent direction in geometric space. Next, a weighted inner product operation is performed using a specific formula. By dynamically aggregating the response values ​​of all intent capsules, a comprehensive score for the candidate product targeting the current user's multidimensional and complex intents is calculated. This calculation process is quantified using the following formula: in, This represents the overall matching score for the j-th candidate item. This represents the total number of independent intent capsules in the fusion activation capsule set, which is 6 in this case. The 64-dimensional embedded feature vector represents the query item, i.e., the j-th candidate product; The representative value is the vector of the i-th intention capsule; The result of the basic inner product operation between the product vector and the capsule vector; The temperature regulation hyperparameter is a pre-set empirical value of 0.1, used to control the smoothness of the probability distribution. The score consists of cross-attention weights based on a normalized exponential function, which are then multiplied by the corresponding base inner product value. Finally, the total score is calculated by summing the results. For a candidate product belonging to the maternal and infant product category, such as a pack of baby diapers, when it is dot-producted with the first five capsules representing historical intentions such as numbers and sports, the dot product value is extremely low or even negative due to the orthogonal separation of the feature spaces. However, when it is dot-producted with the sixth parenting intention capsule, which was just instantiated due to a mutation, it produces a very high positive value. This value is then factored by the temperature coefficient. Dividing by 0.1 achieves a tenfold amplification effect. After exponential mapping and global normalization, the attention weight assigned to the maternal and infant capsule will approach 100%. The final weighted summation result is almost entirely determined by the product's performance in the mutation intent dimension, resulting in an extremely high overall matching score. By traversing all 1000 candidate products and using this cross-attention mechanism to perform multi-dimensional matching and scoring aggregation one by one, the final output is a sequence of overall matching scores containing 1000 one-dimensional scalar values.

[0047] Finally, the values ​​of each candidate product in the comprehensive matching score are globally sorted in descending order. After obtaining the one-dimensional sequence containing 1000 scores, each specific value is first precisely bound and mapped to its corresponding candidate product's unique identifier in key-value pairs at the data processing level, ensuring that the attributes of the product entities are not misplaced or disordered during the sorting process. Subsequently, a heap sort or quick sort algorithm with extremely low time complexity is used to globally and strictly sort these 1000 unique identifier pairs bound to scores in descending order, thereby generating a complete sequence stream with scores decreasing from high to low. After sorting, a partial subset of products needs to be extracted for front-end page rendering. At this point, a preset recommendation list capacity parameter is introduced. This parameter is determined based on the window display capacity limit of a single page refresh on a mobile application device and the fatigue threshold of normal user scrolling. To balance server bandwidth and user visual experience, this value is strictly set to 50. The model strictly follows this preset recommendation list capacity of 50, extracting the top 50 product entity sequences starting from the top of the descending sort. However, products with extremely high scores calculated solely by algorithms may have flaws or conflicts in actual business operations. Therefore, these 50 product sequences are reviewed and optimized by multiple filtering operations on the business side. Business filtering operations include: an inventory status verification operator, which removes products that, while highly matched, are currently out of stock in the warehouse; a historical purchase deduplication operator, which compares the user's order history over the past seven days to forcibly remove duplicate category products that have recently been purchased, avoiding wasted display resources and user frustration; and a category dispersion operator to ensure that the products in the list are not overly concentrated on a single brand, maintaining reasonable exposure diversity. After this series of rigorous business and logical filtering and the addition of high-scoring products, the information of these 50 products, along with corresponding image links and copywriting materials, are finally packaged and encapsulated. The resulting object is the final product recommendation list, which is sent to the front end via an API. This recommendation list, thanks to its advantages of intention decoupling, mutation isolation, and cross-matching, can not only prominently and accurately showcase the mother and baby products that users urgently need due to sudden demands, but also smoothly accommodate their long-term needs for digital products in auxiliary positions.

[0048] In summary, the product recommendation method based on multi-dimensional user profiles, as described in this application, aims to address the problems of undecoupled long-term and short-term latent spaces in existing recommendation models, and the lag trap of sudden intents being diluted by historical features. First, by orthogonally decoupling and projecting user attributes and historical behavior sequences into capsules, a long-term intent capsule cluster is constructed, fundamentally eliminating the mutual interference between long-term and short-term features in the same latent space. Second, for real-time behavior data, transient behavior vectors are captured and encoded using the topological structure of the product knowledge graph, and an iterative dynamic routing protocol is introduced to quantify the coupling degree between transient behaviors and each long-term intent capsule. The core is to perform mutation detection based on the comparison between the peak value of the intent coupling coefficient distribution and the dynamic mutation threshold. When a sudden intent is detected, it is directly instantiated as an independent capsule and incorporated into the long-term capsule cluster, thereby completely blocking the suppression effect and mathematical dilution of the massive historical profile; for regular behaviors, momentum fusion is performed. Finally, based on this fused activated capsule set, cross-attention matching and scoring are performed on candidate products to achieve accurate recommendations that take into account both stable user preferences and transient mutation intents.

[0049] Figure 5 This is a block diagram of a product accuracy recommendation system based on multi-dimensional user profiles according to an embodiment of this application. Figure 5 As shown, the product precision recommendation system 100 based on multi-dimensional user profiles according to an embodiment of this application includes: a user attribute historical behavior instantiation module 110, used to perform orthogonal decoupling projection and capsule instantiation on user attribute data and historical behavior sequences based on a product knowledge graph to obtain a long-term intent capsule cluster; a real-time behavior capture and encoding module 120, used to perform transient micro-behavior capture and representation encoding on real-time behavior data collected in the current session window through the topological structure of the product knowledge graph to obtain a transient behavior vector; and an intent coupling metric module 130, used to perform iterative dynamic routing protocol on the transient behavior vector and each intent capsule in the long-term intent capsule cluster. The intent coupling metric is quantified to obtain the intent coupling coefficient distribution; the mutation detection and judgment module 140 is used to perform mutation detection and judgment on transient behavior vectors based on the comparison between the peak value of the intent coupling coefficient distribution and the dynamic mutation threshold. If a mutation is determined, it is instantiated into an independent capsule and incorporated into the long-term intent capsule cluster. Otherwise, it is fused with the momentum of the long-term intent capsule to form a matching capsule, generating a fused activation capsule set; the product recommendation list generation module 150 is used to use the candidate product embedding features extracted from the product knowledge graph as the query object, perform cross-attention matching and scoring on the fused activation capsule set, and sort and truncate it in descending order according to the obtained comprehensive matching score to generate a product recommendation list.

[0050] Here, those skilled in the art will understand that the specific operations of each step in the above-described product precision recommendation system based on multi-dimensional user profiles have been referenced above. Figures 1 to 4 The method for accurate product recommendation based on multi-dimensional user profiles has been described in detail, and therefore, its repeated description will be omitted.

[0051] Those skilled in the art will understand that the steps, measures, and schemes in the various operations, methods, and processes discussed in this application can be alternated, modified, combined, or deleted; furthermore, other steps, measures, and schemes in the various operations, methods, and processes discussed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted; furthermore, the steps, measures, and schemes in the prior art that are similar to those disclosed in this application can also be alternated, modified, rearranged, decomposed, combined, or deleted. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above-described embodiments are merely illustrative of several implementation methods of this disclosure, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent for the embodiments of this disclosure. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this disclosure, and these all fall within the protection scope of the embodiments of this disclosure. Therefore, the protection scope of the embodiments of this disclosure should be determined by the appended claims. As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and detail can be made without departing from the spirit and scope of the present invention as defined in the appended claims.

[0052] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for accurate product recommendation based on multi-dimensional user profiles, characterized in that, include: Step 1: Based on the product knowledge graph, orthogonal decoupled projection and capsule instantiation are performed on user attribute data and historical behavior sequences to obtain long-term intent capsule clusters. This includes: using the topological node structure of the product knowledge graph, entity alignment and multilayer perceptron encoding mapping are performed on the product identifiers involved in the user attribute data and historical behavior sequences to obtain the original feature embedding matrix; the original feature embedding matrix is ​​then projected into the intent space under orthogonal constraints to obtain the orthogonal intent representation matrix. The original feature embedding matrix is ​​input into a neural network containing an orthogonal penalty term. The neural network includes an intent space projection layer, which internally contains a learnable projection weight matrix. The projection calculation and orthogonal constraint processes are expressed through the forward intent mapping formula and an independent orthogonal penalty term loss function, respectively, as detailed below: in, This represents the orthogonal intention representation matrix output after orthogonal projection processing; This represents the original feature embedding matrix generated in the previous stage; This represents the projected weight matrix in a neural network containing orthogonal penalty terms; Represents the normalized exponential function; Represents the orthogonal penalty coefficient; The transpose of the projection weight matrix; Representative and Identity matrices of the same dimension; The square of the Frobenius norm of the matrix is ​​used to represent the intention feature; nonlinear compression and instantiation of the intention feature of the orthogonal intention representation matrix are performed to obtain a long-term intention capsule cluster; Step 2: Through the topological structure of the product knowledge graph, the real-time behavioral data collected in the current session window is subjected to transient micro-behavior capture and representation encoding to obtain transient behavior vectors; Step 3: Based on the iterative dynamic routing protocol, the intent coupling metric is performed on the transient behavior vector and each intent capsule in the long-term intent capsule cluster to obtain the intent coupling coefficient distribution. This includes: calculating the projection intensity of the transient behavior vector with the central feature vector of each capsule in the long-term intent capsule cluster to obtain the initial spatial alignment score; based on the consensus iteration mechanism of the iterative dynamic routing protocol, the initial spatial alignment score is iteratively updated with routing consistency-based weights to obtain the iterative route logarithmic weights. The routing consistency-based weight iterative update, through multiple bottom-up negotiations, enables the lower-level transient features to dynamically find the best-matching higher-level intent capsule. In each iteration, the logarithmic weights are non-linearly corrected according to the context dynamic gain coefficient of the current session; the iterative route logarithmic weights are normalized and mapped with intent coupling degree to obtain the intent coupling coefficient distribution. Step 4: Based on the comparison between the peak value of the intention coupling coefficient distribution and the dynamic mutation threshold, the transient behavior vector is subjected to mutation detection and judgment. If it is determined to be a mutation, it is instantiated as an independent capsule and incorporated into the long-term intention capsule cluster. Otherwise, it is fused with the momentum of the long-term intention capsule to form a matching capsule, generating a fused activation capsule set. Step 5: Using the candidate product embedding features extracted from the product knowledge graph as the query object, perform cross-attention matching and scoring on the fusion activation capsule set, and sort and truncate the results in descending order based on the obtained comprehensive matching scores to generate a product recommendation list.

2. The method for accurate product recommendation based on multi-dimensional user profiles according to claim 1, characterized in that, Step two includes: Based on a preset sliding time window, the original log messages in the real-time behavior data are parsed and noisy behaviors are removed. Entity alignment and temporal reorganization are performed based on the product knowledge graph to obtain the real-time action sequence. Context-aware graph embedding mapping is performed on the instantaneous action sequence to obtain the context graph embedding sequence; Based on the multi-head self-attention mechanism, the context graph embedding sequence is subjected to intra-sequence causal relationship capture and global average pooling aggregation compression to obtain transient behavior vectors.

3. The method for accurate product recommendation based on multi-dimensional user profiles according to claim 1, characterized in that, The transient behavior vector is projected onto the central feature vector of each capsule in the long-term intent capsule cluster to obtain the initial spatial alignment score. This includes calculating the projection intensity using the following formula: in, For transient behavior vectors, For each capsule, the central feature vector is... Let i be the i-th component in the initial spatial alignment score.

4. The method for accurate product recommendation based on multi-dimensional user profiles according to claim 1, characterized in that, Step four includes: Extract the peak value from the distribution of the intention coupling coefficient, and compare the peak value with the dynamic mutation threshold using a step function to obtain the mutation state identifier and the target activation index; Controlled diversion of long-term intent capsule clusters is performed based on mutation state identifiers. When the mutation state identifier is 0, the transient behavior vector is fused with the capsule corresponding to the target activation index to obtain a matching capsule. When the mutation state identifier is 1, it is instantiated as an independent capsule, and the matching capsule or independent capsule is assembled into a reconstructed capsule feature matrix. The redundant components that are collinear with other capsules in the reconstructed capsule feature matrix due to the injection of new features are projected and smoothed to obtain the fused activated capsule set.

5. The method for accurate product recommendation based on multi-dimensional user profiles according to claim 1, characterized in that, Step five includes: Retrieve nodes and their first-order neighborhood attribute features from the current target user's candidate product set in the product knowledge graph, and perform deep structured embedding extraction and dimension alignment to obtain candidate product embedding features; Using the embedded features of candidate products as query terms, a multi-capsule cross-attention aggregation score is performed on the fused activation capsule set to obtain a comprehensive matching score; The values ​​of each candidate product in the comprehensive matching score are sorted globally in descending order, and the top-ranked product entity sequence is extracted according to the preset recommendation list capacity. After business filtering, the sequence is packaged to generate a product recommendation list.

6. The method for accurate product recommendation based on multi-dimensional user profiles according to claim 1, characterized in that, The consensus iteration mechanism based on the iterative dynamic routing protocol updates the initial spatial alignment score with weights based on routing consistency to obtain the iterative route logarithmic weights, including: The initial spatial alignment score is used to estimate the cross-capsule competition inhibition factor to obtain the competition inhibition coefficient matrix; Based on the real-time context feature vector, the initial spatial alignment score is subjected to context-aware adaptive gain modulation to obtain a dynamic gain operator; Based on the dynamic gain operator and the competition suppression coefficient matrix, the initial spatial alignment score is updated nonlinearly through competitive routing iteration to obtain the iterative route log weight.

7. A product precision recommendation system based on multi-dimensional user profiles, characterized in that, include: The user attribute historical behavior instantiation module is used to perform orthogonal decoupled projection and capsule instantiation on user attribute data and historical behavior sequences based on a product knowledge graph to obtain long-term intent capsule clusters. This includes: using the topological node structure of the product knowledge graph, performing entity alignment and multilayer perceptron encoding mapping on the product identifiers involved in the user attribute data and historical behavior sequences to obtain the original feature embedding matrix; and performing orthogonal constraint-based intent space projection on the original feature embedding matrix to obtain an orthogonal intent representation matrix. Specifically, the original feature embedding matrix is ​​input into a neural network containing an orthogonal penalty term. The neural network includes an intent space projection layer with a learnable projection weight matrix. The projection calculation and orthogonal constraint processes are expressed through a forward intent mapping formula and an independent orthogonal penalty term loss function, respectively, as follows: in, This represents the orthogonal intention representation matrix output after orthogonal projection processing; This represents the original feature embedding matrix generated in the previous stage; This represents the projected weight matrix in a neural network containing orthogonal penalty terms; Represents the normalized exponential function; Represents the orthogonal penalty coefficient; The transpose of the projection weight matrix; Representative and Identity matrices of the same dimension; The square of the Frobenius norm of the matrix is ​​used to represent the intention feature; nonlinear compression and instantiation of the intention feature of the orthogonal intention representation matrix are performed to obtain a long-term intention capsule cluster; The real-time behavior capture and encoding module is used to capture and encode transient micro-behaviors in real-time behavior data collected within the current session window through the topological structure of the product knowledge graph to obtain transient behavior vectors. The intent coupling metric module, based on an iterative dynamic routing protocol, performs intent coupling metric on transient behavior vectors and each intent capsule in a long-term intent capsule cluster to obtain the intent coupling coefficient distribution. This includes: calculating the projection intensity of the transient behavior vector with the central feature vector of each capsule in the long-term intent capsule cluster to obtain an initial spatial alignment score; using a consensus iteration mechanism based on the iterative dynamic routing protocol, iteratively updating the initial spatial alignment score with routing consistency-based weights to obtain iterative route logarithmic weights. The routing consistency-based weight iteration update, through multiple bottom-up negotiations, allows the lower-level transient features to dynamically find the best-matching higher-level intent capsule. In each iteration, the logarithmic weights are non-linearly corrected based on the current session's context dynamic gain coefficient; and the iterative route logarithmic weights are normalized and mapped to obtain the intent coupling coefficient distribution. The mutation detection and determination module is used to detect and determine the mutation of transient behavior vectors based on the comparison between the peak value of the intention coupling coefficient distribution and the dynamic mutation threshold. If it is determined to be a mutation, it is instantiated as an independent capsule and incorporated into the long-term intention capsule cluster. Otherwise, it is fused with the momentum of the long-term intention capsule to form a matching capsule and generate a fused activation capsule set. The product recommendation list generation module uses the embedded features of candidate products extracted from the product knowledge graph as the query object, performs cross-attention matching and scoring on the fusion activation capsule set, and sorts and truncates the results in descending order based on the obtained comprehensive matching score to generate a product recommendation list.

Images