Method and system for generating marketing strategy based on user behavior sequence

Abstract

The application discloses a marketing strategy generation method and system based on user behavior sequence, relates to the technical field of intelligent marketing, and comprises the following steps: obtaining the original behavior sequence of a target user, extracting a behavior object identifier, constructing an object association structure, fusing and encoding time sequence information and the object association structure to generate an enhanced semantic representation, identifying a marketing response intention type based on the enhanced semantic representation and outputting an initial confidence degree, calibrating the confidence degree by calculating the distribution deviation degree of the enhanced semantic representation from a reference semantic representation set, evaluating a conversion potential score in combination with a behavior active feature, identifying a behavior cycle mode based on the time sequence information and predicting a marketing triggering time, and finally using the conversion potential score to determine a resource allocation quota, selecting matched marketing content, and generating a personalized marketing strategy. The application improves the accuracy and effectiveness of the marketing strategy through semantic enhancement representation and multidimensional evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent marketing technology, specifically to a method and system for generating marketing strategies based on user behavior sequences. Background Technology

[0002] With the rapid development of the mobile internet, e-commerce platforms have accumulated massive amounts of user behavior data. This data contains information on user interests, preferences, and consumption needs, providing a data foundation for precision marketing. Existing marketing strategy generation methods mainly rely on static user profile features and historical purchase records to segment users and develop marketing strategies for different user groups. However, this method ignores the temporal dependencies and object associations inherent in user behavior sequences, making it difficult to accurately grasp real-time changes in user intent.

[0003] Current marketing response intent identification methods typically rely on single behavioral features for classification and prediction, failing to fully utilize the contextual semantic information within the behavioral sequence, resulting in low confidence levels in the identification results. Existing methods often employ fixed allocation rules when allocating marketing resources, neglecting differences in user conversion potential, leading to inefficient use of marketing resources. Furthermore, existing technologies primarily depend on simple statistical patterns when predicting marketing trigger timing, failing to effectively incorporate user behavioral cycle characteristics, resulting in inaccurate timing of marketing efforts.

[0004] As consumer behavior becomes increasingly complex and personalized, traditional marketing strategy generation methods are no longer sufficient to meet the demands of precision marketing. How to fully extract deeper information from user behavior sequences, accurately identify users' marketing response intentions, and rationally allocate resources based on users' conversion potential has become a pressing technical challenge. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for generating marketing strategies based on user behavior sequences, aiming to solve at least one of the technical problems existing in the prior art.

[0006] The technical solution of this invention is: a marketing strategy generation method based on user behavior sequences, comprising the following steps:

[0007] Obtain the original behavior sequence of the target user, extract the behavior object identifiers in the original behavior sequence, construct the object association structure based on the co-occurrence relationship between the behavior object identifiers in the original behavior sequence, fuse and encode the temporal information of the original behavior sequence with the object association structure, and generate an enhanced semantic representation.

[0008] Based on enhanced semantic representation, the marketing response intent type of the target user is identified, and an initial confidence level is output;

[0009] Obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, calibrate the initial confidence level based on the distribution deviation, and obtain the calibrated confidence level.

[0010] Behavioral activity features are extracted from the enhanced semantic representation, and the behavioral activity features are comprehensively evaluated with the calibrated confidence to obtain a conversion potential score.

[0011] Based on temporal information in enhanced semantic representation, behavioral cycle patterns are identified, and the prediction weights of behavioral cycle patterns are adjusted according to conversion potential scores to predict marketing triggering opportunities.

[0012] The conversion potential score is used as the allocation weight to determine the amount of marketing resources allocated. Based on the amount of marketing resources allocated, marketing content that matches the type of marketing response intent is selected, and personalized marketing strategies are generated in combination with the timing of marketing triggers.

[0013] Obtain the target user's original behavior sequence, extract the behavior object identifiers from the original behavior sequence, construct an object association structure based on the co-occurrence relationship between behavior object identifiers in the original behavior sequence, and fuse and encode the temporal information of the original behavior sequence with the object association structure to generate an enhanced semantic representation, including:

[0014] The sequence of interactive behaviors of the target user within a preset time window is obtained as the original behavior sequence, and the occurrence time of the behaviors in the original behavior sequence is recorded as time sequence information.

[0015] Hierarchical encoding is performed on the behavior types in the original behavior sequence, and the operational complexity and interaction duration of the behavior types are analyzed to generate behavior weights.

[0016] Extract product identifiers and class identifiers from the original behavior sequence, and combine the product identifiers and class identifiers with behavior weights to form behavior object identifiers;

[0017] The behavior object identifiers are constructed as graph structure nodes, and the connection relationship between the graph structure nodes is established based on the order of appearance of the behavior object identifiers in the original behavior sequence to generate an object association structure;

[0018] The time interval between connected nodes in the object association structure is calculated based on the time sequence information, and the connection relationship is temporally decayed according to the time interval to generate a temporal association structure.

[0019] Message passing is performed on graph structure nodes in the temporal association structure, and the structural features and context information of the graph structure nodes are fused to generate node representations of the graph structure nodes.

[0020] The node representation is fused and encoded with temporal information to generate an enhanced semantic representation.

[0021] Based on enhanced semantic representation, the marketing response intent type of the target user is identified, and the initial confidence level is output, including:

[0022] Obtain the historical marketing response behavior sequence of the target user, extract the response behavior type, response time and response result from the historical marketing response behavior sequence, and construct the response behavior representation according to the response time order;

[0023] Based on the response behavior characterization, the frequency and intensity of target users' responses under different response behavior types are statistically analyzed, and a response behavior preference distribution is generated by combining the response results;

[0024] By fusing response behavior preference distribution with enhanced semantic representation, the real-time intent representation of the target user is extracted, and multi-level intent features are constructed.

[0025] Based on the conversion order of response behavior types in the historical marketing response behavior sequence, the migration probability of response behavior is calculated, and a response behavior conversion matrix is ​​generated;

[0026] The response behavior transformation matrix is ​​combined with multi-level intent features to calculate the degree of matching between the target user and the preset response intent type. The marketing response intent type with the highest degree of matching and its corresponding matching score are output as the initial confidence level.

[0027] Obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, and calibrate the initial confidence level based on the distribution deviation to obtain the calibrated confidence level, which includes:

[0028] Semantic features are extracted from marketing response intent types, a semantic feature combination matrix is ​​constructed, feature decomposition is performed using the semantic feature combination matrix to obtain feature basis vectors, and a reference semantic representation set is constructed based on the linear combination of the feature basis vectors.

[0029] A semantic feature space is constructed based on the feature basis vectors. The enhanced semantic representation is then projected and transformed according to the feature basis vectors to generate the distribution representation of the enhanced semantic representation in the semantic feature space.

[0030] The semantic dispersion is obtained by calculating the degree of dispersion of the distribution representation in each dimension of the semantic feature space, and the semantic clustering is obtained by calculating the degree of offset of the distribution representation relative to the clustering center of the reference semantic representation set. The semantic dispersion and semantic clustering are weighted and fused to generate the distribution deviation.

[0031] A calibration mapping relationship is constructed based on the distribution deviation. The initial confidence level is then transformed nonlinearly through the calibration mapping relationship to generate a calibrated confidence level.

[0032] Behavioral activity features are extracted from the enhanced semantic representation, and these features are comprehensively evaluated with the calibrated confidence level to obtain a conversion potential score, which includes:

[0033] Extract the occurrence time and duration of user behavior from the enhanced semantic representation to construct a sequence of behavior nodes;

[0034] Determine the temporal association between behavior nodes based on the occurrence time, convert the duration into behavior intensity value and assign it to the corresponding behavior node to generate a behavior temporal chain.

[0035] In the behavioral temporal chain, the difference between the time interval between adjacent behavioral nodes and the behavioral intensity value is calculated to generate behavioral transfer features;

[0036] Based on the time interval, periodic features of the behavior nodes are extracted, and intensity change features are extracted based on the change amplitude of the behavior intensity value. These are combined to generate behavior pattern features.

[0037] The behavior intensity value is corrected based on the behavior pattern features to generate a corrected behavior sequence;

[0038] A time decay coefficient is calculated based on the time interval in the modified behavior sequence, and the modified behavior sequence is decayed using the time decay coefficient to generate behavioral activity features.

[0039] A time weight matrix is ​​constructed based on the temporal correlation, and the behavioral activity features are weighted and fused to generate an evaluation vector.

[0040] Using the calibrated confidence level as a benchmark, the evaluation vector is normalized and weighted to output a conversion potential score.

[0041] Based on temporal information in enhanced semantic representation, behavioral cycle patterns are identified. The prediction weights of these behavioral cycle patterns are adjusted according to conversion potential scores to predict marketing triggering opportunities, including:

[0042] Extract the occurrence and end times of user behavior from the temporal information in the enhanced semantic representation, calculate the duration interval between the occurrence and end times, determine the temporal attributes of the behavior nodes based on the duration interval, and construct a temporal sequence of behavior containing the temporal attributes.

[0043] The behavior time series is grouped by sliding, and the duration interval difference and time span of the behavior nodes in the group are calculated. The change pattern of behavior intensity is identified based on the duration interval difference, and the periodic pattern of behavior occurrence is identified based on the time span. The change pattern and periodic pattern are combined to generate a behavior feature sequence.

[0044] The conversion potential score is converted into feature mapping coefficients. The feature mapping coefficients are then used to adjust the magnitude of the change pattern and the periodicity of the periodicity to generate a predictive feature sequence.

[0045] Calculate the temporal intensity distribution of the predicted feature sequence, perform mean decomposition on the temporal intensity distribution to obtain a temporal trend line, and determine the peak time of the temporal trend line as the marketing triggering opportunity.

[0046] The conversion potential score is used as the allocation weight to determine the marketing resource allocation amount. Based on the marketing resource allocation amount, marketing content that matches the marketing response intent type is selected. Personalized marketing strategies are generated by combining the marketing trigger timing, including:

[0047] The conversion potential score is normalized to obtain the allocation weight, and the total allocation amount of marketing resources is calculated based on the allocation weight.

[0048] The total allocation of marketing resources is divided according to the priority of marketing response intent types, and a marketing resource allocation amount corresponding to each marketing response intent type is generated.

[0049] Extract intent information and scenario elements from the marketing response intent type, construct the combined features of the intent information and scenario elements, and generate a feature vector of the marketing content;

[0050] Calculate the semantic relevance between the feature vector of the marketing content and the marketing response intent type, and select marketing content based on the semantic relevance and the marketing resource allocation amount;

[0051] The marketing content is arranged and combined according to the marketing resource allocation at the marketing trigger point to generate a personalized marketing strategy.

[0052] This invention provides a marketing strategy generation system based on user behavior sequences, the system comprising:

[0053] The behavior processing unit is used to acquire the original behavior sequence of the target user, extract the behavior object identifiers in the original behavior sequence, construct the object association structure based on the co-occurrence relationship between the behavior object identifiers in the original behavior sequence, fuse and encode the temporal information of the original behavior sequence with the object association structure, and generate an enhanced semantic representation.

[0054] The intent recognition unit is used to identify the marketing response intent type of the target user based on the enhanced semantic representation and output the initial confidence level;

[0055] The confidence calibration unit is used to obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, and calibrate the initial confidence based on the distribution deviation to obtain the calibrated confidence.

[0056] The potential assessment unit is used to extract behavioral activity features from the enhanced semantic representation, and to comprehensively evaluate the behavioral activity features with the calibrated confidence to obtain a conversion potential score.

[0057] The timing prediction unit is used to identify behavioral cycle patterns based on temporal information in the enhanced semantic representation, adjust the prediction weight of the behavioral cycle patterns according to the conversion potential score, and predict the timing of marketing triggers.

[0058] The strategy generation unit is used to determine the amount of marketing resources allocated by using conversion potential scores as allocation weights, select marketing content that matches the marketing response intent type based on the marketing resource allocation amount, and generate personalized marketing strategies by combining marketing trigger timing.

[0059] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0060] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps in any of the aforementioned methods.

[0061] This invention effectively extracts temporal dependencies and object association information from user behavior sequences by fusing temporal information with object association structures to generate enhanced semantic representations, thereby enhancing the expressive power of user behavior features. By introducing a reference semantic representation set to calibrate the initial confidence level, the accuracy of marketing response intent identification is improved. A comprehensive evaluation combining behavioral activity features and calibrated confidence levels achieves precise quantification of user conversion potential. Identifying behavioral cycle patterns based on temporal information and adjusting prediction weights according to conversion potential scores improves the accuracy of marketing trigger timing predictions. Using conversion potential scores to determine resource allocation amounts achieves rational allocation of marketing resources and improves resource utilization efficiency. Through the comprehensive utilization of multi-dimensional information, this invention significantly improves the accuracy and effectiveness of personalized marketing strategies, reduces marketing costs, and increases marketing conversion rates. Attached Figure Description

[0062] Figure 1 A flowchart illustrating a marketing strategy generation method based on user behavior sequences provided in an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of the marketing strategy generation system based on user behavior sequences according to an embodiment of the present invention. Detailed Implementation

[0064] like Figure 1 As shown, Figure 1 A flowchart of a marketing strategy generation method based on user behavior sequences provided in an embodiment of the present invention, the method comprising the following steps:

[0065] Obtain the original behavior sequence of the target user, extract the behavior object identifiers in the original behavior sequence, construct the object association structure based on the co-occurrence relationship between the behavior object identifiers in the original behavior sequence, fuse and encode the temporal information of the original behavior sequence with the object association structure, and generate an enhanced semantic representation.

[0066] Based on enhanced semantic representation, the marketing response intent type of the target user is identified, and an initial confidence level is output;

[0067] Obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, calibrate the initial confidence level based on the distribution deviation, and obtain the calibrated confidence level.

[0068] Behavioral activity features are extracted from the enhanced semantic representation, and the behavioral activity features are comprehensively evaluated with the calibrated confidence to obtain a conversion potential score.

[0069] Based on temporal information in enhanced semantic representation, behavioral cycle patterns are identified, and the prediction weights of behavioral cycle patterns are adjusted according to conversion potential scores to predict marketing triggering opportunities.

[0070] The conversion potential score is used as the allocation weight to determine the amount of marketing resources allocated. Based on the amount of marketing resources allocated, marketing content that matches the type of marketing response intent is selected, and personalized marketing strategies are generated in combination with the timing of marketing triggers.

[0071] Obtain the target user's original behavior sequence, extract the behavior object identifiers from the original behavior sequence, construct an object association structure based on the co-occurrence relationship between behavior object identifiers in the original behavior sequence, and fuse and encode the temporal information of the original behavior sequence with the object association structure to generate an enhanced semantic representation, including:

[0072] The sequence of interactive behaviors of the target user within a preset time window is obtained as the original behavior sequence, and the occurrence time of the behaviors in the original behavior sequence is recorded as time sequence information.

[0073] Hierarchical encoding is performed on the behavior types in the original behavior sequence, and the operational complexity and interaction duration of the behavior types are analyzed to generate behavior weights.

[0074] Extract product identifiers and class identifiers from the original behavior sequence, and combine the product identifiers and class identifiers with behavior weights to form behavior object identifiers;

[0075] The behavior object identifiers are constructed as graph structure nodes, and the connection relationship between the graph structure nodes is established based on the order of appearance of the behavior object identifiers in the original behavior sequence to generate an object association structure;

[0076] The time interval between connected nodes in the object association structure is calculated based on the time sequence information, and the connection relationship is temporally decayed according to the time interval to generate a temporal association structure.

[0077] Message passing is performed on graph structure nodes in the temporal association structure, and the structural features and context information of the graph structure nodes are fused to generate node representations of the graph structure nodes.

[0078] The node representation is fused and encoded with temporal information to generate an enhanced semantic representation.

[0079] Obtain the target user's interaction behavior sequence within a preset time window as the raw behavior sequence. The time window can be set to the past 30 days. Record the specific timestamp of each behavior in the raw behavior sequence, accurate to the second, such as "2025-01-10 08:30:45", as the basic data for subsequent time series information analysis.

[0080] The behavior types in the original behavior sequence are hierarchically coded, including browsing, clicking, favorites, adding to cart, and purchasing. The operational complexity and interaction duration of different behavior types are analyzed to generate corresponding behavior weight values. The weight for browsing is set to 0.2, for clicking 0.4, for favorites 0.6, for adding to cart 0.8, and for purchasing 1.0. Behavior duration is determined by calculating the length of time a user stays on a single product page; browsing behaviors with a stay exceeding 120 seconds have their weight increased to 0.3, indicating a higher user interest in the product.

[0081] Product identifiers and category identifiers are extracted from the original behavior sequence. Product identifiers use unique codes, such as "P10086" representing a specific product; category identifiers use a hierarchical structure, such as "C1001-C100102" representing the "Electronic Products - Mobile Phones" category. The product identifier and category identifier are combined with the behavior weight to form a behavior object identifier. A complete behavior object identifier example is "P10086-C1001-C100102-0.8", indicating that the user added product P10086 from the mobile phone category within the Electronic Products category to their cart.

[0082] If a user browses products P10086, P10087, and P10088 sequentially, then in the corresponding graph structure, node P10086 is connected to P10087, and node P10087 is connected to P10088, forming a directed connection relationship. By constructing this connection relationship, a complete object association structure is generated, capturing the inter-product association information contained in the user's behavior sequence.

[0083] If a user browses product P10086 at 08:30:45 and product P10087 at 08:35:20, the time interval between nodes is 295s. Based on the calculated time interval, a temporal decay process is applied to the connection relationships, using a time decay function to determine the connection strength. Connection strength remains at 1.0 for time intervals within 300s; decreases to 0.8 for time intervals exceeding 300s but within 600s; decreases to 0.5 for time intervals exceeding 600s but within 1800s; and decreases to 0.2 for time intervals exceeding 1800s. This temporal decay process generates a temporal association structure that considers the influence of time factors.

[0084] Each node's initial feature vector contains product attribute information, such as price range, brand, and specifications. The current node's representation is updated by aggregating the feature information of neighboring nodes, and three rounds of message passing are performed iteratively. In each round of message passing, node P10086 receives feature information from its connected node P10087 and performs weighted aggregation based on connection strength, thereby fusing the structural features and contextual information of the graph structure nodes to generate a node representation containing rich semantics.

[0085] For each node, the time-point features of its appearance in the original behavior sequence are extracted, including time-division information such as time of day, weekdays, or rest days. These time-division features are concatenated with the node representation, and then feature transformation is performed through a multilayer perceptron to generate an enhanced representation vector that incorporates temporal semantics. For node P10086, if it is viewed by a user during a weekday morning time slot, its temporal feature is encoded as "weekday-morning," which is then fused with the node representation to generate the final enhanced semantic representation.

[0086] In marketing applications, product recommendations can be based on enhanced semantic representations. If a user has recently viewed high-end smartphones, the system can identify their spending potential and product preferences. It can not only recommend similar high-end phones but also suggest accessories that the user might be interested in, such as phone cases and chargers, based on temporal association structures. By analyzing the differences in user shopping behavior between weekdays and weekends, it can identify users' habit of browsing electronic products on weekend mornings and accurately push promotional information during that time period.

[0087] This invention achieves a deep semantic understanding of user behavior data by integrating the temporal information of user behavior sequences with object association structures, demonstrating significant technical advantages compared to traditional methods. This invention can capture the implicit temporal patterns and product associations within user behavior sequences, accurately depicting the evolution of user interests and improving recommendation accuracy. Through hierarchical behavior encoding and a temporal decay mechanism, it effectively distinguishes the importance of different behaviors, solving the technical problem of inaccurate understanding of user intent.

[0088] Based on enhanced semantic representation, the marketing response intent type of the target user is identified, and the initial confidence level is output, including:

[0089] Obtain the historical marketing response behavior sequence of the target user, extract the response behavior type, response time and response result from the historical marketing response behavior sequence, and construct the response behavior representation according to the response time order;

[0090] Based on the response behavior characterization, the frequency and intensity of target users' responses under different response behavior types are statistically analyzed, and a response behavior preference distribution is generated by combining the response results;

[0091] By fusing response behavior preference distribution with enhanced semantic representation, the real-time intent representation of the target user is extracted, and multi-level intent features are constructed.

[0092] Based on the conversion order of response behavior types in the historical marketing response behavior sequence, the migration probability of response behavior is calculated, and a response behavior conversion matrix is ​​generated;

[0093] The response behavior transformation matrix is ​​combined with multi-level intent features to calculate the degree of matching between the target user and the preset response intent type. The marketing response intent type with the highest degree of matching and its corresponding matching score are output as the initial confidence level.

[0094] We acquire historical marketing response behavior sequences of target users as foundational data, containing all user interaction records from past marketing campaigns. We extract three key information categories from these sequences: response behavior type, response time, and response result. Response behavior types are categorized into six main types: exposure, click, favorite, inquiry, participation, and purchase. Response time is recorded precisely to the second, such as "2025-01-15 14:30:22". Response results include positive and negative responses; positive responses indicate actions that achieve the expected conversion goal, while negative responses indicate actions that do not achieve the expected conversion. We construct response behavior representations according to the chronological order of response time, recording the evolution trajectory of user response behavior.

[0095] Based on the constructed response behavior representation, the frequency and intensity of target users' responses under different response behavior types are statistically analyzed. Response frequency is determined by calculating the number of times each type of response behavior occurs within a specific time window, such as 15 clicks, 3 favorites, and 2 purchases by users in the past 30 days. Response intensity is quantified by analyzing the user's level of engagement with the marketing campaign, including factors such as dwell time and interaction depth. For click behavior, the intensity value is set to 0.6 for dwell time exceeding 60 seconds and 0.3 for dwell time less than 60 seconds; for favorite behavior, the intensity value is uniformly set to 0.8; for purchase behavior, the intensity value is set to 1.0 for amounts above the average price of recommended products and 0.9 for amounts below the average price. A response behavior preference distribution is generated based on the response results, reflecting the user's sensitivity and preference for different marketing stimuli. The weighting coefficient for positive responses is set to 1.2, and the weighting coefficient for negative responses is set to 0.8. These values ​​are combined with response frequency and intensity to calculate the final preference distribution value.

[0096] Enhanced semantic representations originate from the analysis of users' original behavioral sequences, encompassing user interests and behavioral patterns. During feature fusion, a feature concatenation method is used to merge the response behavior preference distribution vector with the enhanced semantic representation vector, forming a fused feature vector. This fused feature vector is processed by a fully connected layer to output a real-time intent representation. Based on this real-time intent representation, multi-level intent features are constructed, including three levels: low-level behavioral intent, mid-level demand intent, and high-level goal intent. Low-level behavioral intent captures users' immediate response behavior tendencies, such as click tendencies and favorites tendencies; mid-level demand intent reflects users' functional needs, such as information acquisition, price comparison, and quality confirmation; high-level goal intent represents users' ultimate consumption purpose, such as essential purchases or enjoyment-oriented consumption.

[0097] Based on the conversion order of response behavior types in the historical marketing response behavior sequence, the migration probability of response behavior is calculated. The migration probability is calculated by statistically analyzing the conversion frequency between behavior types. In historical data, the conversion from impression to click occurred 87 times, from click to favorites occurred 23 times, and from favorites to purchase occurred 18 times, with corresponding migration probabilities of 0.21, 0.26, and 0.78, respectively. These migration probability values ​​constitute a complete response behavior conversion matrix, reflecting the conversion efficiency and correlation strength between each link in the user's marketing response behavior chain. In the response behavior conversion matrix, rows represent the initial behavior type, columns represent the target behavior type, and matrix element values ​​represent the conversion probability from the initial behavior to the target behavior.

[0098] The response behavior transformation matrix is ​​combined with multi-level intent features to calculate the target user's matching degree with preset response intent types. These preset intent types include five main types: browsing intent, inquiry intent, price comparison intent, purchase intent, and sharing intent. During the combination operation, for each preset response intent type, a matching score is calculated by combining multi-level intent features and the response behavior transformation matrix. Specifically, a subset of features related to the current preset response intent type is extracted from the multi-level intent features and multiplied by the conversion probability of the corresponding behavior path in the response behavior transformation matrix to obtain the matching score. For the purchase intent type, purchase-related features from the user's high-level target intent are extracted and combined with the conversion probability path pointing to the purchase behavior in the response behavior transformation matrix to obtain a matching score of 0.82. Similarly, matching scores for other intent types are calculated, such as browsing intent (0.65), inquiry intent (0.43), price comparison intent (0.56), and sharing intent (0.38). The marketing response intent type with the highest matching degree and its corresponding matching score are output as the initial confidence level; in this example, it is purchase intent, with an initial confidence level of 0.82.

[0099] This invention achieves accurate identification of users' marketing response intentions through in-depth analysis of their historical marketing response behavior sequences, significantly improving the targeting and effectiveness of marketing strategies. This invention not only considers users' static preference characteristics but also integrates the dynamic evolution patterns inherent in the behavioral sequences, accurately capturing the temporal changes in user intentions. By constructing a multi-level intention feature and response behavior transformation matrix, it comprehensively portrays users' decision-making paths and response tendencies in marketing scenarios, solving the problem of the one-sidedness of traditional methods in understanding user intentions.

[0100] Obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, and calibrate the initial confidence level based on the distribution deviation to obtain the calibrated confidence level, which includes:

[0101] Semantic features are extracted from marketing response intent types, a semantic feature combination matrix is ​​constructed, feature decomposition is performed using the semantic feature combination matrix to obtain feature basis vectors, and a reference semantic representation set is constructed based on the linear combination of the feature basis vectors.

[0102] A semantic feature space is constructed based on the feature basis vectors. The enhanced semantic representation is then projected and transformed according to the feature basis vectors to generate the distribution representation of the enhanced semantic representation in the semantic feature space.

[0103] The semantic dispersion is obtained by calculating the degree of dispersion of the distribution representation in each dimension of the semantic feature space, and the semantic clustering is obtained by calculating the degree of offset of the distribution representation relative to the clustering center of the reference semantic representation set. The semantic dispersion and semantic clustering are weighted and fused to generate the distribution deviation.

[0104] A calibration mapping relationship is constructed based on the distribution deviation. The initial confidence level is then transformed nonlinearly through the calibration mapping relationship to generate a calibrated confidence level.

[0105] Semantic features are extracted from marketing response intent types, including browsing intent, inquiry intent, price comparison intent, purchase intent, and sharing intent. For each intent type, key semantic features are extracted. The semantic features of browsing intent include information acquisition, short dwell time, and frequent redirects; the semantic features of inquiry intent include asking questions, paying attention to details, and frequent interaction; the semantic features of price comparison intent include price focus, comparing multiple products, and repeated viewing; the semantic features of purchase intent include specification confirmation, inventory check, and payment preference; and the semantic features of sharing intent include social attributes, content copying, and forwarding. These extracted semantic features are organized according to intent type to construct a semantic feature combination matrix. The rows of this matrix represent different intent types, the columns represent different semantic features, and the matrix element values ​​represent the correlation strength between a specific intent type and a specific semantic feature. The association strength between browsing intent and information acquisition features is 0.9, the association strength with short dwell time features is 0.8, and the association strength with multiple jumps features is 0.7; the association strength between purchase intent and specification confirmation features is 0.8, the association strength with inventory query features is 0.7, and the association strength with payment preference features is 0.9.

[0106] Feature decomposition is performed using a semantic feature combination matrix, employing singular value decomposition (SVD) to decompose the semantic feature combination matrix into a product of three sub-matrices. The left matrix contains basis vectors for intent types, the right matrix contains basis vectors for semantic features, and the middle diagonal matrix contains singular values. The first five column vectors are extracted from the right matrix as feature basis vectors, forming an orthogonal basis for the semantic feature space. Feature basis vector 1 primarily expresses the price sensitivity dimension, feature basis vector 2 primarily expresses the interaction depth dimension, feature basis vector 3 primarily expresses the dwell time dimension, feature basis vector 4 primarily expresses the social attribute dimension, and feature basis vector 5 primarily expresses the decision stage dimension. A reference semantic representation set is constructed based on the linear combination of feature basis vectors. For each intent type, feature basis vectors are combined with different weights to generate a reference semantic representation for that intent type. The reference semantic representation of purchase intent has a weight of 0.4 in the decision stage dimension, 0.25 in the price sensitivity dimension, 0.2 in the interaction depth dimension, 0.1 in the dwell time dimension, and 0.05 in the social attribute dimension.

[0107] A semantic feature space is constructed based on the feature basis vectors. This space is a five-dimensional vector space, with each dimension corresponding to a feature basis vector. The enhanced semantic representation is then projected onto the feature basis vectors, transforming the original enhanced semantic representation into the newly constructed semantic feature space. Specifically, the inner product of the enhanced semantic representation and each feature basis vector is calculated to obtain the projection value of the enhanced semantic representation in each feature basis vector direction. The projection value of the enhanced semantic representation for the target user is 0.35 in the price sensitivity dimension, 0.25 in the interaction depth dimension, 0.15 in the dwell time dimension, 0.1 in the social attribute dimension, and 0.15 in the decision-making stage dimension. These projection values ​​constitute the distribution representation of the enhanced semantic representation in the semantic feature space, reflecting the distribution of user behavior across each semantic dimension.

[0108] The semantic dispersion is calculated by determining the degree of dispersion of the distributed representation across each dimension of the semantic feature space. This is done by calculating the deviation of the projected values ​​of each dimension from the uniform distribution. In a uniform distribution, the projected value for each dimension is 0.2. In the actual distributed representation, the deviations are 0.15 for price sensitivity, 0.05 for interaction depth, 0.05 for dwell time, 0.1 for social attributes, and 0.05 for decision-making stage. The average of these deviations yields a semantic dispersion of 0.08. The semantic clustering is then calculated by determining the offset of the distributed representation from the cluster center of the reference semantic representation set. This is done by calculating the Euclidean distance between the distributed representation and the reference semantic representation for each intent type, selecting the minimum distance as the semantic clustering index. The distance between the distributed representation and the reference semantic representation for browsing intent is 0.4, for consultation intent is 0.35, for price comparison intent is 0.25, for purchase intent is 0.18, and for sharing intent is 0.45. Therefore, the semantic clustering is 0.18. The semantic dispersion and semantic clustering are weighted and fused to generate the distribution deviation, with the weights set to 0.4 and 0.6 respectively. The calculated distribution deviation is 0.14.

[0109] A calibration mapping relationship is constructed based on the distribution deviation. The calibration mapping relationship adopts a piecewise function form, and the specific expression is as follows:

[0110] ;

[0111] Where d represents the distribution deviation, and f(d) is the corresponding calibration factor.

[0112] The target user distribution deviation is 0.14, corresponding to a calibration factor of 0.96. The initial confidence level is 0.82, and the calibrated confidence level is 0.82 multiplied by 0.96, resulting in 0.79. The calibrated confidence level more accurately reflects the user's actual marketing response intent, eliminating the error caused by the abnormal distribution of semantic representation in the initial confidence level.

[0113] This invention achieves precise calibration of the initial confidence level by constructing a reference semantic representation set and calculating the distribution deviation between the enhanced semantic representation and the reference semantic representation, demonstrating significant technical effectiveness. The semantic feature space constructed through feature decomposition technology can characterize the semantic features of users' marketing response intent from multiple dimensions, improving the accuracy of intent recognition. The introduction of distribution deviation solves the judgment bias problem that may arise from relying solely on the initial confidence level, making the intent recognition results more robust. The nonlinear design of the calibration mapping relationship adapts to calibration needs under different degrees of deviation, effectively suppressing the interference of abnormal samples.

[0114] Behavioral activity features are extracted from the enhanced semantic representation, and these features are comprehensively evaluated with the calibrated confidence level to obtain a conversion potential score, which includes:

[0115] Extract the occurrence time and duration of user behavior from the enhanced semantic representation to construct a sequence of behavior nodes;

[0116] Determine the temporal association between behavior nodes based on the occurrence time, convert the duration into behavior intensity value and assign it to the corresponding behavior node to generate a behavior temporal chain.

[0117] In the behavioral temporal chain, the difference between the time interval between adjacent behavioral nodes and the behavioral intensity value is calculated to generate behavioral transfer features;

[0118] Based on the time interval, periodic features of the behavior nodes are extracted, and intensity change features are extracted based on the change amplitude of the behavior intensity value. These are combined to generate behavior pattern features.

[0119] The behavior intensity value is corrected based on the behavior pattern features to generate a corrected behavior sequence;

[0120] A time decay coefficient is calculated based on the time interval in the modified behavior sequence, and the modified behavior sequence is decayed using the time decay coefficient to generate behavioral activity features.

[0121] A time weight matrix is ​​constructed based on the temporal correlation, and the behavioral activity features are weighted and fused to generate an evaluation vector.

[0122] Using the calibrated confidence level as a benchmark, the evaluation vector is normalized and weighted to output a conversion potential score.

[0123] The time of occurrence and duration of user behaviors are extracted from the enhanced semantic representation to construct a sequence of behavior nodes. The enhanced semantic representation contains temporal information of the user's historical behaviors. By parsing this representation, the specific time of occurrence of each behavior is extracted, such as: on January 15, 2025, at 14:30:22, the user clicked on the marketing push notification; on January 15, 2025, at 14:35:45, the user browsed the product details; and on January 15, 2025, at 14:42:10, the user added the product to the shopping cart. The duration of each behavior is also extracted, such as: clicking on the marketing push notification lasted 15 seconds, browsing the product details lasted 320 seconds, and adding the product to the shopping cart lasted 60 seconds. Using the time points as indexes, the behaviors are arranged in chronological order to form a sequence of behavior nodes.

[0124] Temporal associations are determined by the temporal order between adjacent behavior nodes, forming directed connections. The conversion from duration to behavior intensity value uses a segmented mapping method. For browsing behaviors, the intensity value is 0.2 for durations between 0-30 seconds; 0.5 for durations between 31-120 seconds; 0.7 for durations between 121-300 seconds; and 0.9 for durations exceeding 300 seconds. For interactive behaviors, the intensity value is 0.4 for durations between 0-60 seconds; 0.7 for durations between 61-180 seconds; and 0.9 for durations exceeding 180 seconds. For conversion behaviors, the intensity value is 0.8 for durations between 0-120 seconds; and 1.0 for durations exceeding 120 seconds.

[0125] The time interval is obtained by calculating the difference in the time points of adjacent behavior nodes. For example, the time interval between clicking a marketing push and browsing product details is 323 seconds, and the time interval between browsing product details and adding to cart is 385 seconds. The difference in behavior intensity values ​​is obtained by calculating the difference in behavior intensity values ​​between adjacent behavior nodes. For example, the difference in behavior intensity value between browsing product details and clicking a marketing push is 0.7, and the difference in behavior intensity value between adding to cart and browsing product details is -0.1. Combining the time interval and the difference in behavior intensity values ​​constitutes the behavior transfer feature.

[0126] Periodic features are extracted by analyzing the temporal patterns of behavior occurrences. The average time interval of all behavior nodes is calculated as the baseline period, which is 354 seconds. The deviation of each time interval from the baseline period is used as the periodic feature; the deviation from click to browse is -31 seconds, and the deviation from browse to add to cart is 31 seconds. Intensity variation features are extracted by analyzing the variation patterns of behavior intensity values. The average absolute value of the difference in behavior intensity values ​​is calculated as the variation baseline, which is 0.4. The ratio of the difference in each behavior intensity value to the variation baseline is used as the intensity variation feature; the intensity variation feature from click to browse is 1.75, and the intensity variation feature from browse to add to cart is -0.25.

[0127] The behavior intensity value is adjusted based on behavioral pattern characteristics, taking into account the influence of behavioral periodicity and intensity variation characteristics. For periodicity characteristics, the behavior intensity value is increased when the actual time interval is less than the baseline period, and decreased when the actual time interval is greater than the baseline period. The adjustment magnitude is proportional to the degree of deviation; for every 100 seconds increase in deviation, the behavior intensity value is adjusted by 0.1. For intensity variation characteristics, the behavior intensity value is increased when the intensity variation amplitude is large and positive, and decreased when the intensity variation amplitude is large and negative. The adjustment magnitude is proportional to the absolute value of the intensity variation characteristic; for every 1.0 increase in the absolute value of the intensity variation characteristic, the behavior intensity value is adjusted by 0.05.

[0128] The time decay coefficient is calculated based on the time intervals in the modified behavior sequence. This coefficient reflects the impact of the timeliness of the behavior on activity level and is calculated using an exponential decay model. The half-life is set to 720 seconds, meaning that the activity impact of a behavior halves every 720 seconds. The current time is 14:45:00 on January 15, 2025. The times elapsed since each behavior occurred are 878 seconds, 258 seconds, and 173 seconds, respectively, with corresponding time decay coefficients of 0.43, 0.80, and 0.86. Multiplying the modified behavior intensity value by the corresponding time decay coefficient yields the activity characteristics considering timeliness. The activity characteristic values ​​for the three behavior nodes are 0.10, 0.76, and 0.65, respectively.

[0129] A time-weighted matrix is ​​constructed based on temporal correlation. This matrix reflects the importance of behaviors within different time windows, with recent behaviors having a higher weight than distant behaviors. Behaviors are divided into three time windows based on their distance from the current time: 0-300s is the recent window (weight 0.6); 301-600s is the intermediate window (weight 0.3); and over 600s is the distant window (weight 0.1). The three behavior nodes fall into the distant, recent, and recent windows respectively, with corresponding time weights of 0.1, 0.6, and 0.6. The activity features of the behaviors are multiplied by their corresponding time weights and summed to obtain a weighted fusion evaluation vector with a value of 0.85.

[0130] Using the calibrated confidence level as a baseline, the evaluation vector was normalized and weighted to output a conversion potential score. With a calibrated confidence level of 0.79, the evaluation vector of 0.85 was normalized by dividing it by the maximum possible value of 1.0 (based on the premise that behavioral activity characteristics have been standardized to the [0, 1] interval and the sum of time weights is 1), resulting in 0.85. The weights of the calibrated confidence level and the evaluation vector in the conversion potential score were set to 0.6 and 0.4, respectively, resulting in a final conversion potential score of 0.81. This score comprehensively reflects the user's intent confidence and behavioral activity, providing data support for subsequent marketing strategy development.

[0131] This invention achieves precise quantification of user conversion potential by extracting behavioral activity features from enhanced semantic representations and comprehensively evaluating them with calibrated confidence scores. This invention not only considers the temporal characteristics and intensity changes of user behavior but also introduces a time decay mechanism, enabling accurate capture of the dynamic trends in user interests. By constructing behavioral temporal chains and behavioral pattern features, it comprehensively characterizes the consistency of user behavior and conversion paths in the marketing environment. The introduction of a time weight matrix solves the problem of traditional methods neglecting the influence of time windows, making the evaluation of activity features more reasonable.

[0132] Based on temporal information in enhanced semantic representation, behavioral cycle patterns are identified. The prediction weights of these behavioral cycle patterns are adjusted according to conversion potential scores to predict marketing triggering opportunities, including:

[0133] Extract the occurrence and end times of user behavior from the temporal information in the enhanced semantic representation, calculate the duration interval between the occurrence and end times, determine the temporal attributes of the behavior nodes based on the duration interval, and construct a temporal sequence of behavior containing the temporal attributes.

[0134] The behavior time series is grouped by sliding, and the duration interval difference and time span of the behavior nodes in the group are calculated. The change pattern of behavior intensity is identified based on the duration interval difference, and the periodic pattern of behavior occurrence is identified based on the time span. The change pattern and periodic pattern are combined to generate a behavior feature sequence.

[0135] The conversion potential score is converted into feature mapping coefficients. The feature mapping coefficients are then used to adjust the magnitude of the change pattern and the periodicity of the periodicity to generate a predictive feature sequence.

[0136] Calculate the temporal intensity distribution of the predicted feature sequence, perform mean decomposition on the temporal intensity distribution to obtain a temporal trend line, and determine the peak time of the temporal trend line as the marketing triggering opportunity.

[0137] The enhanced semantic representation includes detailed temporal information of user behavior, from which the start and end times of each behavior are extracted. Taking an e-commerce marketing scenario as an example, a user starts browsing recommended products at 9:30:15 on January 10, 2025, and ends browsing at 9:35:45; starts viewing product details at 14:22:30 on January 10, 2025, and ends viewing at 14:28:15; adds products to the shopping cart at 20:15:00 on January 10, 2025, and ends the operation at 20:16:45; initiates payment at 10:10:25 on January 11, 2025, and completes payment at 10:12:10. The duration intervals for each behavior are calculated to be 330s, 345s, 105s, and 105s, respectively. Based on the duration intervals, the temporal attributes of the behavior nodes are determined, including duration, time period, and behavior density. The duration is directly taken from the duration interval. The occurrence time is divided into four categories: morning, noon, afternoon, and evening, based on the start time of the behavior. Behavior density is defined as the number of behavior operations completed per unit time. A behavior time sequence is constructed, and each node includes behavior type, occurrence time, end time, duration interval, and derived time attributes.

[0138] A sliding window method was used to group behavior sequences, with a window size of 48 hours and a sliding step of 6 hours, ensuring that each window contained sufficient behavior samples and that adjacent windows had some overlap. For each behavior node within a group, the duration interval difference between adjacent behaviors was calculated. For example, the second behavior's duration interval was 15 seconds longer than the first, the third behavior's duration interval was 240 seconds shorter than the second, and the fourth behavior's duration interval was the same as the third. These differences reflected the changing trend of user behavior intensity. The time span between adjacent behaviors within a group was also calculated. For example, the interval between the first and second behaviors was 4 hours, 52 minutes, and 15 seconds; the interval between the second and third behaviors was 5 hours, 46 minutes, and 45 seconds; and the interval between the third and fourth behaviors was 13 hours, 53 minutes, and 40 seconds. Behavior intensity change patterns were identified based on the duration interval difference. A positive difference exceeding 60 seconds was considered an increase in intensity, a negative difference exceeding 60 seconds was considered a decrease in intensity, and a difference within ±60 seconds was considered stable intensity. Based on the time span identification of periodic patterns, the average value of all time spans was calculated to be 8 hours, 10 minutes, and 53 seconds, which was used as the baseline period. The deviations of each time span from the baseline period were -3 hours, 18 minutes, and 38 seconds, -2 hours, 24 minutes, and 08 seconds, and 5 hours, 42 minutes, and 47 seconds, respectively. The intensity change pattern was combined with the periodic pattern to generate a behavioral feature sequence.

[0139] The conversion potential score reflects the strength of a user's conversion intention, ranging from 0 to 1. In this example, the user's conversion potential score is 0.78. The conversion potential score is converted into a feature mapping coefficient using a piecewise function mapping method: when the score is in the range of 0-0.3, the mapping coefficient is 0.5; when the score is in the range of 0.31-0.6, the mapping coefficient is 0.8; when the score is in the range of 0.61-0.85, the mapping coefficient is 1.2; and when the score is in the range of 0.86-1, the mapping coefficient is 1.5. Based on the score of 0.78, the feature mapping coefficient is determined to be 1.2. The amplitude of the change pattern is adjusted using the feature mapping coefficient by multiplying the original duration interval difference by the feature mapping coefficient. The adjusted differences are 18s, -288s, and 0s, respectively. The periodic pattern is adjusted by scaling the period deviation according to the feature mapping coefficient, making the prediction period for high-potential users more accurate. The adjusted period deviations are -3 hours 58 minutes 22 seconds, -2 hours 52 minutes 58 seconds, and 6 hours 51 minutes 22 seconds, respectively. By combining the adjusted change patterns and periodic regularities, a predictive feature sequence is generated.

[0140] The temporal intensity distribution reflects the changing pattern of user behavior intensity over time, calculated by aggregating the temporal attributes and adjusted change patterns in the predicted feature sequence. On a 24-hour timeline, time slots are divided at the hourly granularity, and the intensity values ​​of historical user behaviors are assigned to the corresponding time slots. The intensity value is determined comprehensively based on the behavior type and duration interval: the base value for browsing behavior is 0.3, for viewing product details is 0.6, for adding to the shopping cart is 0.8, and for payment behavior is 1.0, adjusted according to the proportion of the duration interval relative to the average duration of that type of behavior. This results in an intensity value distribution for 24 time slots, such as 0.35 for 9:00-10:00, 0.8 for 10:00-11:00, 0.65 for 14:00-15:00, and 0.85 for 20:00-21:00, etc. The temporal intensity distribution is then decomposed by mean. First, the average intensity for the entire day is calculated (0.45 in this example). The intensity value for each time slot is then subtracted from the average to obtain the intensity fluctuation value. A time-series trend line is generated by smoothing fluctuation values ​​using a three-point weighted moving average. Peak times on the trend line represent the times when user behavior intensity reaches its peak, and these moments are suitable for triggering marketing campaigns. In this embodiment, the peak times of the trend line occur at 10:30, 15:15, and 20:45. Considering the high conversion potential score of users, the moment with the highest intensity, 20:45, is selected as the priority marketing trigger time.

[0141] This invention identifies behavioral cycle patterns based on temporal information in enhanced semantic representation and adjusts the prediction weights of these patterns according to conversion potential scores to accurately predict marketing trigger times. This invention can extract patterns and periodic characteristics of behavioral intensity changes from temporal data of user behavior and make targeted adjustments based on the user's conversion intention strength, avoiding the inefficiency caused by fixed-time marketing triggers in traditional methods. Through sliding grouping analysis, it captures the dynamic changes of user behavior patterns over time, improving the adaptability of predictions. The introduction of a feature mapping mechanism transforms the abstract conversion potential score into specific prediction parameter adjustment coefficients, enabling differentiated processing for users with different conversion intentions.

[0142] The conversion potential score is used as the allocation weight to determine the marketing resource allocation amount. Based on the marketing resource allocation amount, marketing content that matches the marketing response intent type is selected. Personalized marketing strategies are generated by combining the marketing trigger timing, including:

[0143] The conversion potential score is normalized to obtain the allocation weight, and the total allocation amount of marketing resources is calculated based on the allocation weight.

[0144] The total allocation of marketing resources is divided according to the priority of marketing response intent types, and a marketing resource allocation amount corresponding to each marketing response intent type is generated.

[0145] Extract intent information and scenario elements from the marketing response intent type, construct the combined features of the intent information and scenario elements, and generate a feature vector of the marketing content;

[0146] Calculate the semantic relevance between the feature vector of the marketing content and the marketing response intent type, and select marketing content based on the semantic relevance and the marketing resource allocation amount;

[0147] The marketing content is arranged and combined according to the marketing resource allocation at the marketing trigger point to generate a personalized marketing strategy.

[0148] Conversion potential score reflects the likelihood of a user converting into a purchase under marketing incentives, ranging from 0 to 1. For example, a user's conversion potential score is 0.75. Normalization uses a non-linear mapping method to convert the conversion potential score into allocation weights. Considering that users with high conversion potential have a higher marginal benefit to marketing resources, a quadratic function mapping is used. When the conversion potential score is 0.75, the calculated allocation weight is 0.84. Given the limited total marketing resources available to the company, assuming a maximum allocation of 1000 units per user, the actual allocation amount is determined by the allocation weight. For a user with an allocation weight of 0.84, the calculated total allocation amount is 840 units. The allocation weight is fine-tuned considering the user's historical response. For example, if the user's average response rate to marketing campaigns over the past 30 days is 0.62, the adjusted allocation weight is 0.82, corresponding to a total allocation amount of 820 units.

[0149] Semantic analysis identified the user's marketing response intent types, including price-sensitive, quality-seeking, new product-trying, and limited-time offer-seeking. For this user, analysis of their historical behavior sequence revealed the following distribution of marketing response intent types: price-sensitive (45%), quality-seeking (30%), new product-trying (15%), and limited-time offer-seeking (10%). Based on the priority and percentage of intent types, total resources were allocated as follows: price-sensitive users received 45% of the 820 units (369 units); quality-seeking users received 30% (246 units); new product-trying users received 15% (123 units); and limited-time offer-seeking users received 10% (82 units). The allocation ratio is adjusted to take into account the conversion efficiency of each intent type. For example, the average conversion rate of price-sensitive marketing is 0.25, quality-seeking marketing is 0.20, new product trial marketing is 0.15, and limited-time flash sale marketing is 0.30. The original allocation ratio is then adjusted by weighting based on the conversion rate, and the adjusted allocation amounts are 378 units, 201 units, 90 units, and 151 units, respectively.

[0150] Intent information includes core elements such as the user's purchase motivation, price expectations, and quality requirements. For example, a user's price-sensitive intent might be expressed as "expecting a price 20% lower than the market average," while a quality-seeking intent might be expressed as "paying attention to product materials and craftsmanship details." Contextual elements include environmental factors such as the user's usage scenario, purchase timing, and social influence. For example, a user's primary usage scenario might be "daily commuting," and their purchase timing might be "end-of-season discounts." Combining intent information and contextual elements constructs combined features. For example, "price-sensitive + daily commuting" indicates that the user expects products used in their daily commuting scenario to have a price advantage. These combined features are quantified into feature vectors, with each dimension corresponding to the strength of a particular feature. For example, the strength of "price-sensitive + daily commuting" is 0.85, and the strength of "quality-seeking + formal occasions" is 0.72. Corresponding feature vectors are also generated for all possible marketing content. For example, the value of the "price-sensitive" dimension in the feature vector for discount coupons is 0.90, and the value of the "quality-seeking" dimension in the feature vector for quality assurance is 0.95.

[0151] The semantic relevance of marketing content feature vectors to marketing response intent types is calculated. The cosine similarity method is used to measure the degree of matching between marketing content features and user intent features. For price-sensitive intents, the relevance of various marketing content types is calculated as follows: 0.92 for full-reduction offers, 0.90 for discount coupons, 0.85 for flash sales, 0.78 for discount packages, and 0.75 for member-exclusive offers. For quality-seeking intents, the relevance is as follows: 0.95 for quality assurance, 0.88 for recommended products, 0.84 for positive reviews, 0.82 for detailed parameters, and 0.76 for comparative evaluation. A semantic relevance threshold of 0.75 is set, and marketing content with relevance higher than the threshold is selected as candidates. Based on the resource allocation quota for each intent type, marketing content is selected according to relevance ranking and specific resources are allocated. Price-sensitive users are allocated 378 units of resources, with 151 units allocated to the most relevant discount offers, 132 units to discount coupons, and 95 units to limited-time flash sales. Quality-conscious users are allocated 201 units of resources, with 105 units allocated to quality assurance and 96 units to recommended premium products.

[0152] Personalized marketing strategies are generated by arranging and combining marketing content at the designated marketing trigger times according to the allocated marketing resources. The marketing trigger times, identified in the previous steps, are peak user activity periods, such as 9:00 AM, 3:30 PM, and 9:00 PM. Based on user behavior analysis, user receptiveness to different marketing content varies across different time periods: users tend to browse information in the morning, compare and choose during the afternoon, and make purchase decisions in the evening. To address these time-specific characteristics, the marketing content is arranged chronologically: at 9:00 AM, quality assurance content is pushed out, allocating 105 units of resources to showcase product quality features and user reviews in text and image format; at 3:30 PM, discounts and coupons are pushed out, allocating 283 units of resources to showcase price advantages in the form of a discount details page; at 9:00 PM, limited-time flash sales and featured product recommendations are pushed out, allocating 191 units of resources to stimulate user decision-making through countdowns and recommendation lists. Each marketing content session includes direct conversion entry points, such as "Buy Now" and "Claim Coupon" action buttons, facilitating quick user response. Maintain visual consistency and informational coherence in marketing content across all time periods to avoid user confusion. Based on historical user interaction data, select preferred delivery channels such as push notifications, emails, or in-app news feeds to ensure effective reach of marketing content.

[0153] This invention determines the allocation of marketing resources by using conversion potential scores as weights, selects matching marketing content based on marketing response intent types, and generates personalized marketing strategies by combining marketing trigger timing, thus achieving precise allocation and efficient utilization of marketing resources. It employs nonlinear mapping technology to transform abstract user potential indicators into specific resource allocation weights, solving the problem of coarse-grained marketing resource allocation in traditional methods. Through intent type identification and priority division, it achieves reasonable allocation of marketing resources in different marketing directions, avoiding resource waste. The introduction of a semantic relevance calculation mechanism ensures that the selected marketing content highly matches user intent, improving marketing effectiveness. The content arrangement strategy based on time-period characteristics fully considers users' psychological states and decision-making tendencies at different times, making marketing outreach more targeted.

[0154] like Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of a marketing strategy generation system based on user behavior sequences according to an embodiment of the present invention. The system includes:

[0155] The behavior processing unit is used to acquire the original behavior sequence of the target user, extract the behavior object identifiers in the original behavior sequence, construct the object association structure based on the co-occurrence relationship between the behavior object identifiers in the original behavior sequence, fuse and encode the temporal information of the original behavior sequence with the object association structure, and generate an enhanced semantic representation.

[0156] The intent recognition unit is used to identify the marketing response intent type of the target user based on the enhanced semantic representation and output the initial confidence level;

[0157] The confidence calibration unit is used to obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, and calibrate the initial confidence based on the distribution deviation to obtain the calibrated confidence.

[0158] The potential assessment unit is used to extract behavioral activity features from the enhanced semantic representation, and to comprehensively evaluate the behavioral activity features with the calibrated confidence to obtain a conversion potential score.

[0159] The timing prediction unit is used to identify behavioral cycle patterns based on temporal information in the enhanced semantic representation, adjust the prediction weight of the behavioral cycle patterns according to the conversion potential score, and predict the timing of marketing triggers.

[0160] The strategy generation unit is used to determine the amount of marketing resources allocated by using conversion potential scores as allocation weights, select marketing content that matches the marketing response intent type based on the marketing resource allocation amount, and generate personalized marketing strategies by combining marketing trigger timing.

[0161] One technical solution provided in this embodiment of the invention is an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0162] One technical solution provided in this embodiment of the invention is a computer-readable storage medium storing a computer program, wherein the processor executes the computer program to implement the steps in any of the aforementioned methods.

[0163] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A marketing strategy generation method based on user behavior sequences, characterized in that, Includes the following steps: Obtain the original behavior sequence of the target user, extract the behavior object identifiers in the original behavior sequence, construct the object association structure based on the co-occurrence relationship between the behavior object identifiers in the original behavior sequence, fuse and encode the temporal information of the original behavior sequence with the object association structure, and generate an enhanced semantic representation. Based on enhanced semantic representation, the marketing response intent type of the target user is identified, and an initial confidence level is output; Obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, calibrate the initial confidence based on the distribution deviation, and obtain the calibrated confidence. Behavioral activity features are extracted from the enhanced semantic representation, and the behavioral activity features are comprehensively evaluated with the calibrated confidence to obtain a conversion potential score. Based on temporal information in enhanced semantic representation, behavioral cycle patterns are identified, and the prediction weights of behavioral cycle patterns are adjusted according to conversion potential scores to predict marketing triggering opportunities. The conversion potential score is used as the allocation weight to determine the amount of marketing resources allocated. Based on the amount of marketing resources allocated, marketing content that matches the type of marketing response intent is selected, and personalized marketing strategies are generated in combination with the timing of marketing triggers. Behavioral activity features are extracted from the enhanced semantic representation, and these features are comprehensively evaluated with the calibrated confidence level to obtain a conversion potential score, which includes: Extract the occurrence time and duration of user behavior from the enhanced semantic representation to construct a sequence of behavior nodes; Determine the temporal association between behavior nodes based on the occurrence time, convert the duration into behavior intensity value and assign it to the corresponding behavior node to generate a behavior temporal chain. In the behavioral temporal chain, the difference between the time interval between adjacent behavioral nodes and the behavioral intensity value is calculated to generate behavioral transfer features; Based on the time interval, periodic features of the behavior nodes are extracted, and intensity change features are extracted based on the change amplitude of the behavior intensity value. These are combined to generate behavior pattern features. The behavior intensity value is corrected based on the behavior pattern features to generate a corrected behavior sequence; A time decay coefficient is calculated based on the time interval in the modified behavior sequence, and the modified behavior sequence is decayed using the time decay coefficient to generate behavioral activity features. A time weight matrix is ​​constructed based on the temporal correlation, and the behavioral activity features are weighted and fused to generate an evaluation vector. Using the calibrated confidence level as a benchmark, the evaluation vector is normalized and weighted to output a conversion potential score.

2. The method according to claim 1, characterized in that, Obtain the target user's original behavior sequence, extract the behavior object identifiers from the original behavior sequence, construct an object association structure based on the co-occurrence relationship between behavior object identifiers in the original behavior sequence, and fuse and encode the temporal information of the original behavior sequence with the object association structure to generate an enhanced semantic representation, including: The sequence of interactive behaviors of the target user within a preset time window is obtained as the original behavior sequence, and the occurrence time of the behaviors in the original behavior sequence is recorded as time sequence information. Hierarchical encoding is performed on the behavior types in the original behavior sequence, and the operational complexity and interaction duration of the behavior types are analyzed to generate behavior weights. Extract product identifiers and class identifiers from the original behavior sequence, and combine the product identifiers and class identifiers with behavior weights to form behavior object identifiers; The behavior object identifiers are constructed as graph structure nodes, and the connection relationship between the graph structure nodes is established based on the order of appearance of the behavior object identifiers in the original behavior sequence to generate an object association structure; The time interval between connected nodes in the object association structure is calculated based on the time sequence information, and the connection relationship is temporally decayed according to the time interval to generate a temporal association structure. Message passing is performed on graph structure nodes in the temporal association structure, and the structural features and context information of the graph structure nodes are fused to generate node representations of the graph structure nodes. The node representation is fused and encoded with temporal information to generate an enhanced semantic representation.

3. The method according to claim 1, characterized in that, Based on enhanced semantic representation, the marketing response intent type of the target user is identified, and the initial confidence level is output, including: Obtain the historical marketing response behavior sequence of the target user, extract the response behavior type, response time and response result from the historical marketing response behavior sequence, and construct the response behavior representation according to the response time order; Based on the response behavior characterization, the frequency and intensity of target users' responses under different response behavior types are statistically analyzed, and a response behavior preference distribution is generated by combining the response results; By fusing response behavior preference distribution with enhanced semantic representation, the real-time intent representation of the target user is extracted, and multi-level intent features are constructed. Based on the conversion order of response behavior types in the historical marketing response behavior sequence, the migration probability of response behavior is calculated, and a response behavior conversion matrix is ​​generated; The response behavior transformation matrix is ​​combined with multi-level intent features to calculate the degree of matching between the target user and the preset response intent type. The marketing response intent type with the highest degree of matching and its corresponding matching score are output as the initial confidence level.

4. The method according to claim 1, characterized in that, Obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, and calibrate the initial confidence level based on the distribution deviation to obtain the calibrated confidence level, which includes: Semantic features are extracted from marketing response intent types, a semantic feature combination matrix is ​​constructed, feature decomposition is performed using the semantic feature combination matrix to obtain feature basis vectors, and a reference semantic representation set is constructed based on the linear combination of the feature basis vectors. A semantic feature space is constructed based on the feature basis vectors. The enhanced semantic representation is then projected and transformed according to the feature basis vectors to generate the distribution representation of the enhanced semantic representation in the semantic feature space. The semantic dispersion is obtained by calculating the degree of dispersion of the distribution representation in each dimension of the semantic feature space, and the semantic clustering is obtained by calculating the degree of offset of the distribution representation relative to the clustering center of the reference semantic representation set. The semantic dispersion and semantic clustering are weighted and fused to generate the distribution deviation. A calibration mapping relationship is constructed based on the distribution deviation. The initial confidence level is then transformed nonlinearly through the calibration mapping relationship to generate a calibrated confidence level.

5. The method according to claim 1, characterized in that, Based on temporal information in enhanced semantic representation, behavioral cycle patterns are identified. The prediction weights of these behavioral cycle patterns are adjusted according to conversion potential scores to predict marketing triggering opportunities, including: Extract the occurrence and end times of user behavior from the temporal information in the enhanced semantic representation, calculate the duration interval between the occurrence and end times, determine the temporal attributes of the behavior nodes based on the duration interval, and construct a temporal sequence of behavior containing the temporal attributes. The behavior time series is grouped by sliding, and the duration interval difference and time span of the behavior nodes in the group are calculated. The change pattern of behavior intensity is identified based on the duration interval difference, and the periodic pattern of behavior occurrence is identified based on the time span. The change pattern and periodic pattern are combined to generate a behavior feature sequence. The conversion potential score is converted into feature mapping coefficients. The feature mapping coefficients are then used to adjust the magnitude of the change pattern and the periodicity of the periodicity to generate a predictive feature sequence. Calculate the temporal intensity distribution of the predicted feature sequence, perform mean decomposition on the temporal intensity distribution to obtain a temporal trend line, and determine the peak time of the temporal trend line as the marketing triggering opportunity.

6. The method according to claim 1, characterized in that, The conversion potential score is used as the allocation weight to determine the marketing resource allocation amount. Based on the marketing resource allocation amount, marketing content that matches the marketing response intent type is selected. Personalized marketing strategies are generated by combining the marketing trigger timing, including: The conversion potential score is normalized to obtain the allocation weight, and the total allocation amount of marketing resources is calculated based on the allocation weight. The total allocation of marketing resources is divided according to the priority of marketing response intent types, and a marketing resource allocation amount corresponding to each marketing response intent type is generated. Extract intent information and scenario elements from the marketing response intent type, construct the combined features of the intent information and scenario elements, and generate a feature vector of the marketing content; Calculate the semantic relevance between the feature vector of the marketing content and the marketing response intent type, and select marketing content based on the semantic relevance and the marketing resource allocation amount; The marketing content is arranged and combined according to the marketing resource allocation at the marketing trigger point to generate a personalized marketing strategy.

7. A marketing strategy generation system based on user behavior sequences, used to implement the method described in any one of claims 1-6, characterized in that, The system includes: The behavior processing unit is used to acquire the original behavior sequence of the target user, extract the behavior object identifiers in the original behavior sequence, construct the object association structure based on the co-occurrence relationship between the behavior object identifiers in the original behavior sequence, fuse and encode the temporal information of the original behavior sequence with the object association structure, and generate an enhanced semantic representation. The intent recognition unit is used to identify the marketing response intent type of the target user based on the enhanced semantic representation and output the initial confidence level; The confidence calibration unit is used to obtain the reference semantic representation set corresponding to the marketing response intent type, calculate the distribution deviation between the enhanced semantic representation and the reference semantic representation set, and calibrate the initial confidence based on the distribution deviation to obtain the calibrated confidence. The potential assessment unit is used to extract behavioral activity features from the enhanced semantic representation, and to comprehensively evaluate these behavioral activity features with the calibrated confidence level to obtain a conversion potential score, which specifically includes: Behavioral activity features are extracted from the enhanced semantic representation, and these features are comprehensively evaluated with the calibrated confidence level to obtain a conversion potential score, which includes: Extract the occurrence time and duration of user behavior from the enhanced semantic representation to construct a sequence of behavior nodes; Determine the temporal association between behavior nodes based on the occurrence time, convert the duration into behavior intensity value and assign it to the corresponding behavior node to generate a behavior temporal chain. In the behavioral temporal chain, the difference between the time interval between adjacent behavioral nodes and the behavioral intensity value is calculated to generate behavioral transfer features; Based on the time interval, periodic features of the behavior nodes are extracted, and intensity change features are extracted based on the change amplitude of the behavior intensity value. These are combined to generate behavior pattern features. The behavior intensity value is corrected based on the behavior pattern features to generate a corrected behavior sequence; A time decay coefficient is calculated based on the time interval in the modified behavior sequence, and the modified behavior sequence is decayed using the time decay coefficient to generate behavioral activity features. A time weight matrix is ​​constructed based on the temporal correlation, and the behavioral activity features are weighted and fused to generate an evaluation vector. Using the calibrated confidence level as the baseline, the evaluation vector is normalized and weighted to output a conversion potential score. The timing prediction unit is used to identify behavioral cycle patterns based on temporal information in the enhanced semantic representation, adjust the prediction weight of the behavioral cycle patterns according to the conversion potential score, and predict the timing of marketing triggers. The strategy generation unit is used to determine the amount of marketing resources allocated by using conversion potential scores as allocation weights, select marketing content that matches the marketing response intent type based on the marketing resource allocation amount, and generate personalized marketing strategies by combining marketing trigger timing.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 6.

Images