A digital marketing user portrait generation method and system based on deep learning

Abstract

The application discloses a kind of digital marketing user portrait generation method and system based on deep learning, it is related to digital marketing technical field, including, to original fine-grained interaction event flow is standardized, and the output standardized user microcosmic behavior time sequence;Standardized user microcosmic behavior time sequence is input to deep time sequence neural network, and the output user cognitive load quantification score and visual attention focus distribution vector;Based on user cognitive load quantification score and visual attention focus distribution vector, to static long-term interest portrait is weighted and focused fusion, generates situational dynamic user portrait vector;Situational dynamic user portrait vector is input counterfactual explanation engine.The application solves the problem that the marketing communication opportunity and form adaptability caused by lack of real-time psychological state assessment of user in prior art, improves the precision of digital marketing, user experience and communication efficiency.

Description

Technical Field

[0001] This invention relates to the field of digital marketing technology, and in particular to a method and system for generating digital marketing user profiles based on deep learning. Background Technology

[0002] With the development of big data and artificial intelligence, deep learning-based user profiling technology has become crucial for precision marketing. Current mainstream technologies analyze and model users' click, browsing, and purchase history sequences by constructing deep neural networks, such as Transformers or recurrent neural networks. The aim is to uncover stable characteristics such as users' long-term interests and consumption tendencies, ultimately outputting static interest tags or user segments. Some advanced solutions incorporate online learning mechanisms to dynamically adjust interest tags, collectively forming existing solutions with the core paradigm of mining and predicting user interests from behavioral data.

[0003] Existing technologies have limitations, primarily focusing on what users are interested in, rather than effectively perceiving and integrating users' real-time internal cognitive states, such as cognitive load levels and attention focus. This makes it difficult for the system to assess the user's current information reception readiness, resulting in the generated user profiles being unable to directly guide when and in what form marketing communication is optimal. Since the model objectives do not include the assessment of instantaneous psychological load and attention resources, although the system can infer the content theme, it cannot determine the best push timing and the most suitable information complexity. Existing technology-driven marketing lacks adaptability in terms of timing and form, making it difficult to synchronize with the user's current psychological rhythm, thus limiting further optimization of communication efficiency and experience. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based digital marketing user profile generation method to solve the technical problem that existing user profiles cannot quantify users' real-time cognitive state, and thus cannot adaptively optimize the timing and form of marketing communication.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for generating digital marketing user profiles based on deep learning, which includes standardizing the original fine-grained interaction event stream and outputting a standardized time series of user micro-behaviors. The standardized time series of user micro-behaviors is input into a deep temporal neural network, which outputs a quantitative score of user cognitive load and a distribution vector of visual attention focus. Based on the user's cognitive load quantification score and visual attention focus distribution vector, the static long-term interest profile is weighted and focused to generate a contextualized dynamic user profile vector. By inputting contextualized dynamic user profile vectors into the counterfactual interpretation engine, the information entropy density and time phase of preset marketing actions are extrapolated to identify the optimal parameter combination that resonates between the information delivery path and the user's cognitive rhythm. Based on the optimal parameter combination, a personalized communication strategy package is output. Marketing actions are executed based on the resonance parameters in the personalized communication strategy package, and the deep temporal neural network and counterfactual explanation engine are updated based on feedback data.

[0007] As a preferred embodiment of the deep learning-based digital marketing user profile generation method of the present invention, the method includes the following steps: standardizing the original fine-grained interaction event stream to output a standardized time series of user micro-behaviors. Data cleaning is performed on the original fine-grained interaction event stream to obtain the cleaned original fine-grained interaction event stream, and session segmentation is performed to obtain independent session fragments; Perform time alignment on each individual session segment to obtain time-aligned session segments, and perform feature vectorization to transform the event at each time point into a feature vector; Perform sequence standardization on the feature vectors to output a standardized time series of user micro-behaviors.

[0008] As a preferred embodiment of the deep learning-based digital marketing user profile generation method of the present invention, the method includes the following steps: inputting a standardized user micro-behavior time series into a deep temporal neural network, and outputting a user cognitive load quantification score and a visual attention focus distribution vector. Standardized user micro-behavior time series are input into the one-dimensional convolutional layer of a one-dimensional convolutional gated recurrent hybrid neural network to extract local temporal features, resulting in a local temporal feature vector; The local temporal feature vector is input into the gated recurrent unit layer of a one-dimensional convolutional gated recurrent hybrid neural network to capture long-term dependencies and obtain long-term dependency feature vectors. The long-time dependent feature vector is input into the multi-head self-attention mechanism layer of a one-dimensional convolutional gated recurrent hybrid neural network and then subjected to attention weighting across time steps to obtain attention-weighted feature vectors. The attention-weighted feature vector is fed into the cognitive load score output head of a one-dimensional convolutional gated recurrent hybrid neural network, and mental bandwidth assessment and neural information flow compression perception are performed on the attention-weighted feature vector to generate a user cognitive load quantification score. The attention-weighted feature vector is fed into the attention distribution output head of a one-dimensional convolutional gated recurrent hybrid neural network, and visual field reconstruction and spatial attention weight diffusion are performed on the attention-weighted feature vector to generate a visual attention focus distribution vector.

[0009] As a preferred embodiment of the deep learning-based digital marketing user profile generation method of the present invention, the method includes the following steps: Based on the user's cognitive load quantification score and visual attention focus distribution vector, a static long-term interest profile is weighted and focused fused to generate a contextualized dynamic user profile vector. Calculate the dynamic fusion weight coefficient based on the user's cognitive load quantification score; Based on the visual attention focus distribution vector, the content region is filtered from the real-time detected user short-term interest vector to obtain the filtered user short-term interest vector. The static long-term interest profile vector is weighted based on the dynamic fusion weight coefficient to obtain the adjusted static long-term interest profile vector. The adjusted static long-term interest profile vector is merged with the filtered short-term user interest vector to generate a contextualized dynamic user profile vector.

[0010] As a preferred embodiment of the deep learning-based digital marketing user profile generation method of the present invention, the method includes the following steps: inputting a contextualized dynamic user profile vector into a counterfactual interpretation engine to extrapolate the information entropy density and time phase of a preset marketing action, and identifying the optimal parameter combination that resonates between the information transmission path and the user's cognitive rhythm. Input the contextualized dynamic user profile vector into the counterfactual interpretation engine, and construct the information entropy optimization space based on the real-time communication appropriateness label and interest focus in the contextualized dynamic user profile vector; Within the information entropy optimization space, the counterfactual explanation engine takes minimizing cognitive friction and maximizing intention convergence as its dual objectives, and performs collaborative perturbation and virtual simulation on the information entropy density parameter and time phase parameter of each preset marketing action in the preset marketing action template library. The virtual simulation process simulates the expected impact of different combinations of information entropy density parameters and time phase parameters on the cognitive state represented by contextualized dynamic user profile vectors, and evaluates the matching degree between information delivery path and user cognitive rhythm. By comparing the matching degree evaluation results under different parameter combinations, the counterfactual explanation engine identifies the combination of information entropy density parameter and time phase parameter that maximizes the matching degree, and obtains the optimal parameter combination that resonates between the information transmission path and the user's cognitive rhythm.

[0011] As a preferred embodiment of the deep learning-based digital marketing user profile generation method of the present invention, the method includes the following steps: Based on the optimal parameter combination, a personalized communication strategy package is output. Based on the information entropy density parameter in the optimal parameter combination, the content presentation complexity of the preset marketing action template is adjusted to generate the tuned marketing content. Based on the time phase parameter in the optimal parameter combination, the planned push time of the tuned marketing content is set, and the tuned marketing content and the planned push time are encapsulated to output a personalized communication strategy package.

[0012] As a preferred embodiment of the deep learning-based digital marketing user profile generation method of the present invention, the method includes the following steps: executing marketing actions based on the resonance parameters in the personalized communication strategy package, and updating the deep temporal neural network and counterfactual interpretation engine based on feedback data: Based on the planned push times in the personalized communication strategy package, execute marketing actions for the adjusted marketing content at the corresponding times; Collect user feedback data on the adjusted marketing content, including clicks, browsing time, and conversion behavior; The feedback data is combined with the standardized user micro-behavior time series, contextualized dynamic user profile vectors and optimal parameters used in the process of generating personalized communication strategy packages to form a feedback sample. The deep temporal neural network is incrementally trained using feedback samples to update its parameters. The weights of the dual objective function in the counterfactual explanation engine are adaptively adjusted using feedback samples to update the inference evaluation criteria of the counterfactual explanation engine.

[0013] Secondly, the present invention provides a digital marketing user profile generation system based on deep learning, including a standardization module, which performs standardization processing on the original fine-grained interaction event stream and outputs a standardized time series of user micro-behaviors. The evaluation module inputs standardized time series of user micro-behaviors into a deep temporal neural network and outputs a quantitative score of user cognitive load and a vector of visual attention focus. The fusion module, based on the user's cognitive load quantification score and visual attention focus distribution vector, performs weighted and focused fusion of static long-term interest profiles to generate contextualized dynamic user profile vectors. The deduction module inputs contextualized dynamic user profile vectors into the counterfactual interpretation engine to deduce the information entropy density and time phase of preset marketing actions, and identify the optimal parameter combination that resonates with the information transmission path and the user's cognitive rhythm. Assemble the module and output a personalized communication strategy package based on the optimal parameter combination; The learning module executes marketing actions based on the resonance parameters in the personalized communication strategy package and updates the deep temporal neural network and counterfactual explanation engine based on feedback data.

[0014] The beneficial effects of this invention are as follows: By standardizing the user's fine-grained interactive event stream and simultaneously extracting the user's real-time cognitive load quantification score and visual attention focus distribution vector using a deep temporal neural network, accurate quantitative perception of the user's instantaneous psychological state is achieved. The quantified cognitive state is weighted and focused on the user's static long-term interest profile to generate a dynamic profile that reflects the user's real-time context and communicability. By inputting the dynamic profile into a counterfactual interpretation engine, the information entropy density and time phase of the preset marketing actions are deduced in the information entropy optimization space. The optimal parameter combination that resonates with the user's current cognitive rhythm is identified, thereby generating a directly executable personalized communication strategy. By executing the strategy and continuously updating the model using feedback data, a complete perception-decision-learning process is formed. The invention introduces deep modeling of real-time cognitive load and attention state into the user profile and upgrades the profile from describing what the user is interested in to a dynamic strategy generation tool that can guide when and in what form to communicate optimally. This solves the problem of insufficient adaptability of marketing communication timing and form caused by the lack of assessment of the user's real-time psychological state in existing technologies, and improves the accuracy of digital marketing, user experience, and communication efficiency. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a deep learning-based method for generating user profiles in digital marketing.

[0017] Figure 2 This is a schematic diagram of a deep learning-based digital marketing user profile generation system. Detailed Implementation

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0020] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0021] Reference Figures 1-2 This is one embodiment of the present invention, which provides a method for generating digital marketing user profiles based on deep learning, including the following steps: S1. Standardize the original fine-grained interaction event stream and output a standardized time series of user micro-behaviors.

[0022] S1.1 Perform data cleaning on the original fine-grained interactive event stream to obtain the cleaned original fine-grained interactive event stream, and perform session segmentation to obtain independent session fragments.

[0023] Furthermore, the raw fine-grained interactive event stream is first fed into a filter composed of predefined rules. This filter identifies and discards noisy events generated by web crawlers, automated scripts, or network anomalies based on the IP address range of the event source, pattern matching of user agent strings, and the continuity of the physical time window of the event occurrence, thus obtaining a cleaned raw fine-grained interactive event stream. A state machine based on a session timeout threshold is applied to the cleaned raw fine-grained interactive event stream. The state machine monitors the time interval between consecutive events. When the interval exceeds a preset silence period threshold, it determines that the current session ends and starts a new session, thereby segmenting the continuous event stream into multiple independent session segments that are temporally continuous and behaviorally related to specific topics.

[0024] Specifically, this invention uses the dynamic semantics of the conversation, rather than a static time window, as the segmentation criterion. Traditional methods often employ fixed time slices, disrupting the user's coherent flow of intent. This invention simulates the persistence of human attention, utilizing cognitive silence periods within behavioral intervals as natural boundaries. When a user has been inactive for an extended period, their attention has shifted or the task has been completed. In this case, segmentation ensures that each independent conversation segment represents a relatively complete, self-contained cycle of user task or interest exploration. This behavior-dynamic-based segmentation provides more semantically consistent input units for subsequent deep temporal neural network analysis, making it easier for the model to capture the complete micro-behavioral patterns from intent germination and information exploration to decision-making, rather than fragmented sequences mixed with multiple unrelated intents.

[0025] S1.2 Perform time alignment on each independent session fragment to obtain time-aligned session fragments, and perform feature vectorization to transform the event at each time point into a feature vector.

[0026] Furthermore, a time alignment operation is performed on each independent session segment. The timestamps of all events are re-interpolated using a fixed sampling frequency, ensuring a uniform distribution of all events along the timeline. This eliminates timestamp irregularities caused by network latency or client-side data acquisition jitter, resulting in time-aligned session segments. Subsequently, feature vectorization is performed on the time-aligned session segments. Based on a predefined multi-dimensional event feature dictionary, the event type, associated UI element identifiers, content topic tags, and interaction intensity at each time point are mapped together into a high-dimensional dense feature vector, thus transforming the event at each time point into a feature vector.

[0027] Specifically, unstructured interaction events are elevated to temporal feature vectors with a unified semantic space, providing deep learning models with computationally achievable and information-preserving input. Traditional feature engineering often extracts statistical features, losing the sequential nature of behavior and the rich semantics of the original events. The feature vectorization process essentially constructs a behavioral linguistics dictionary, transforming heterogeneous events such as clicking the "add to cart" button for product A and quickly scrolling through the details page of product B into representations in a vector space that can calculate distance and relevance. Temporal alignment ensures that the temporal rhythm information of the behavioral sequence is preserved, eliminating collection noise and enabling subsequent neural networks to accurately perceive changes in the speed and rhythm of user interactions. These micro-rhythms are precisely the key signal sources for inferring cognitive load and attention focus, transforming the raw logs into standardized mathematical objects that retain the original behavioral semantics and micro-temporal dynamics.

[0028] S1.3 Perform sequence standardization on the feature vectors and output a standardized time series of user micro-behaviors.

[0029] Furthermore, sequence standardization is performed on the feature vectors to obtain the mean and standard deviation of all feature vectors in the entire training corpus across each feature dimension. These statistics are then used to perform zero-mean, unit-variance scaling on the real-time generated feature vector sequence dimension by dimension. This operation maps interactive features with different dimensions and value ranges to similar numerical distribution intervals. The entire scaled feature vector sequence is then length-normalized, and sequences of different session lengths are adjusted to the same time step by truncation or padding. The final output is a standardized user micro-behavior time series with uniform dimensions, stable numerical distribution, and direct input into deep temporal neural networks.

[0030] Specifically, sequence-level standardization resolves the fundamental contradiction between the fusion of multi-source heterogeneous behavioral data and the stability of deep model training. Different interaction events have features with different physical meanings and numerical ranges; direct input would lead to the model optimization process being dominated by large numerical features. Zero-mean unit variance scaling ensures that all feature dimensions contribute equally to model parameter updates during gradient descent, enabling the model to learn signals fairly from different types of behaviors such as clicks, scrolling, and dwell. Length normalization addresses the fixed requirement of recurrent neural networks for input sequence length, allowing the model to handle user sessions of arbitrary length. More importantly, standardization enables the model's learned cognitive state assessment capabilities to generalize across users and sessions, because the model learns the mapping relationship between standardized behavioral patterns and cognitive states, rather than specific numerical ranges. This removes obstacles at the data preprocessing level for building a universal and robust real-time cognitive assessment model.

[0031] S2. Input the standardized user micro-behavior time series into a deep temporal neural network, and output the user cognitive load quantification score and visual attention focus distribution vector.

[0032] S2.1 Input the standardized user micro-behavior time series into the one-dimensional convolutional layer of a one-dimensional convolutional gated recurrent hybrid neural network to extract local temporal features and obtain a local temporal feature vector.

[0033] Furthermore, the standardized time series of user micro-behaviors is first input into a one-dimensional convolutional layer of a one-dimensional convolutionally gated recurrent hybrid neural network. This layer is configured with multiple one-dimensional convolutional kernels of different widths, which slide along the temporal dimension to scan and filter feature combinations of adjacent time steps in the input sequence. Each convolutional kernel is responsible for detecting a specific local interaction pattern, such as rapid continuous clicks, slow linear scrolling, or mouse hover sequences with a specific rhythm. Through convolution operations and processing by nonlinear activation functions, the one-dimensional convolutional layer extracts various basic, temporally localized behavioral primitives from the original behavior sequence and concatenates the responses of these primitives along the feature dimension to form a local temporal feature vector that can characterize short-term behavioral patterns.

[0034] Specifically, one-dimensional convolution is used as the front end for perceiving micro-behavioral patterns. Standardized user micro-behavioral time series contain rich, high-frequency local temporal structures, such as the suddenness of clicks and changes in scrolling speed. The one-dimensional convolutional layer acts as a behavioral pattern filter, specifically capturing these short-lived local correlations. For example, a narrow convolutional kernel can identify the impatient pattern of rapid double-clicks, while a slightly wider kernel can identify the focused pattern of slow, reading-style scrolling. Decoupling and extracting these local patterns from the original high-dimensional temporal signal is the cornerstone for subsequent higher-order cognitive state inference. This allows the model to not directly process complex original behavioral points, but first construct an intermediate representation composed of basic behavioral vocabulary, establishing stable and interpretable underlying features for understanding more macro-level behavioral semantics and cognitive states.

[0035] S2.2 Input the local temporal feature vector into the gated recurrent unit layer of the one-dimensional convolutional gated recurrent hybrid neural network to capture long-term dependencies and obtain long-term dependency feature vectors.

[0036] Furthermore, the local temporal feature vectors are then input into the gated recurrent unit layer of a one-dimensional convolutional gated recurrent hybrid neural network. The gated recurrent unit layer processes the local temporal feature vector sequence sequentially, dynamically controlling the retention and forgetting of historical information using internal update and reset mechanisms. The update gate determines how much long-term dependent feature vector information from the past is passed to the current moment, while the reset gate controls how the current input is combined with past states. Through the gating mechanism, the gated recurrent unit layer can span a relatively long time window, integrating and memorizing early local behavioral patterns that influence the current state, thereby capturing the user's evolving intent cues and behavioral context throughout the session, and outputting a high-level behavioral representation that incorporates historical information and long-term dependencies—the long-term dependent feature vector.

[0037] Specifically, the gating mechanism of the gated recurrent unit simulates the dynamic weighting and integration of recent and long-term information in human cognition. A user's cognitive state is not determined by isolated instantaneous actions, but is influenced by a series of preceding actions. The reset gate of the gated recurrent unit layer allows the model to selectively forget earlier behaviors irrelevant to the current context, such as discarding an accidental click; the update gate allows the model to selectively remember long-term behavioral patterns that form the foundation of the current intention, such as continuous exploration of a particular topic. This makes the long-term dependent feature vector not only a historical record of behavior, but also a cognitively filtered summary of behavioral memories most relevant to the current psychological state. The dynamic memory mechanism enables the model to understand the causal chain of behavior and the continuity of intention, providing crucial temporal context for assessing sustained cognitive input and attention shifts.

[0038] S2.3. Input the long-time dependent feature vector into the multi-head self-attention mechanism layer of the one-dimensional convolutional gated recurrent hybrid neural network and perform attention weighting across time steps to obtain the attention-weighted feature vector.

[0039] Furthermore, the long-term dependency feature vector is input into the multi-head self-attention mechanism layer of a one-dimensional convolutionally gated recurrent hybrid neural network. This projects the long-term dependency feature vector into multiple different subspaces through a linear transformation, forming a query vector, key vector, and value vector set. Within each subspace, the similarity between the query vector at the current time step and the key vectors at all time steps in the sequence is obtained, resulting in an attention weight distribution. This distribution is then used to weight and sum the value vectors, yielding the attention-weighted output for that subspace. The outputs from multiple subspaces are concatenated and subjected to another linear transformation, integrating information from different representation subspaces. This allows features from any two time steps in the sequence to directly interact, regardless of their temporal distance. The output is an attention-weighted feature vector that explicitly models long-term dependencies within the sequence and the global context.

[0040] Specifically, a multi-head self-attention mechanism is introduced to address the bottleneck of traditional recurrent neural networks in capturing ultra-long-range, non-local dependencies. The assessment of user cognitive load and attention often relies on the correlation analysis of non-adjacent key behavioral events across time points. For example, determining whether a user is distracted may require correlating rapid page jumps at the start of a session with prolonged inactivity on a particular page during the middle of the session. The multi-head self-attention mechanism allows the model to explicitly and in parallel calculate the correlation strength between any two behavioral events in a sequence, thereby constructing a global behavioral correlation graph. Different attention heads can learn to focus on different types of relationships; for example, one head might focus on the browsing-purchase decision path correlation, while another might focus on the attention dissipation pattern of rapid scrolling and repeated returns. This capability enables attention-weighted feature vectors to capture the most informative key events in the behavioral sequence that determine cognitive states, along with their complex interactions, providing a global, structured feature representation for subsequent accurate quantitative evaluation.

[0041] S2.4 The attention-weighted feature vector is fed into the cognitive load score output head of a one-dimensional convolutional gated recurrent hybrid neural network, and mental bandwidth assessment and neural information flow compression perception are performed on the attention-weighted feature vector to generate a user cognitive load quantification score.

[0042] Furthermore, the attention-weighted feature vector is fed in parallel into the cognitive load score output head of a one-dimensional convolutionally gated recurrent hybrid neural network. The cognitive load score output head is a small subnetwork that first performs global average pooling on the attention-weighted feature vector in the time dimension, compressing the variable-length sequence into a fixed-dimensional global context vector. This global context vector is then non-linearly transformed through several fully connected layers, with the final layer using a sigmoid activation function to map the high-dimensional features into a continuous scalar value between zero and one. This scalar value is the user's cognitive load quantification score, which comprehensively reflects the temporal disorder, rhythmic abruptness, and dispersion of intent of the behavioral patterns encoded in the attention-weighted feature vector. A higher value indicates a higher inferred real-time cognitive load level for the user.

[0043] Specifically, a single measure of cognitive channel saturation is directly distilled from the behavioral association graph. The concept behind the cognitive load score output head is to compress the perception of neural information flow, compressing the attention-weighted feature vector representing complex global behavioral associations into a scalar representing the degree of mental resource occupancy through pooling and nonlinear transformation. A mapping function is learned that can identify the fingerprints left by high cognitive load in behavior, such as fragmentation of association patterns caused by rapid switching of attention between multiple tasks, interruption of goal-oriented behavioral chains, and disorder of behavioral rhythm. The use of the Sigmoid function ensures that the output has clear probabilistic interpretability and comparability. The mapping from high-dimensional complex behavioral patterns to low-dimensional mental state quantities enables an indirect, continuous, and quantifiable estimation of the user's unobservable internal cognitive state.

[0044] S2.5 The attention-weighted feature vector is fed into the attention distribution output head of a one-dimensional convolutional gated recurrent hybrid neural network, and visual field reconstruction and spatial attention weight diffusion are performed on the attention-weighted feature vector to generate a visual attention focus distribution vector.

[0045] Furthermore, the attention-weighted feature vector is simultaneously fed into the attention distribution output head of a one-dimensional convolutionally gated recurrent hybrid neural network. The attention distribution output head first performs a special spatial dimension mapping on the attention-weighted feature vector, reorganizing the temporal step features onto a two-dimensional grid simulating the visual interface layout based on the interface region identifiers associated with the behavioral events. A lightweight convolutional network processes this spatial feature map, progressively reconstructing a heatmap of the possible distribution of user visual attention through convolution and upsampling operations. Finally, this heatmap is normalized and vectorized to generate a visual attention focus distribution vector. Each dimension of this vector corresponds to a semantic block in the interface, and its value represents the probability weight of the user's attention being allocated to that block, thus quantifying the user's spatial allocation preference for visual attention in the current session.

[0046] Specifically, the spatial attention distribution is reconstructed from temporal behavioral data, achieving visual field reconstruction. Traditional user profiling completely lacks quantification of where users are looking. The attention distribution output head achieves reconstruction through a key assumption: continuous or high-frequency user interactions (such as hovering or micro-scrolling) in specific areas of the interface (such as product main images, price areas, and detail text) are the behavioral manifestations that attract the user's visual attention to that area. This output head learns to deconvolve attention-weighted feature vectors containing spatiotemporal correlations into a two-dimensional attention probability map simulating the visual interface. This process not only utilizes the temporal patterns of behavior but also the positional semantic information attached to the behavior. The generated visual attention focus distribution vector enables the system to understand the specific foothold of the user's interest (whether it is price sensitivity or curiosity about functional details), providing unprecedented fine-grained guidance for subsequent content focus and personalized information presentation, advancing user understanding from the topic level to the interface information architecture level.

[0047] S3. Based on the user's cognitive load quantification score and visual attention focus distribution vector, the static long-term interest profile is weighted and focused to generate a contextualized dynamic user profile vector.

[0048] S3.1 Calculate the dynamic fusion weight coefficient based on the user's cognitive load quantification score.

[0049] Furthermore, the user cognitive load quantification score is input into the dynamic fusion weight coefficient calculation process. The scaling factor and bias term are pre-learned from historical interaction data or set based on business experience. The calculation process first calculates the intermediate term of the expression, which is the difference between one and the user cognitive load quantification score multiplied by the scaling factor, plus the bias term. Minimum and maximum pruning operations are performed on the intermediate result to ensure that the final output value is limited to a closed interval between zero and one. This pruned final calculation result is the dynamic fusion weight coefficient, whose value is negatively correlated with the user cognitive load quantification score, and is used to subsequently adjust the influence strength of the static long-term interest profile vector.

[0050] Specifically, cognitive load, a psychological state, is transformed into a control parameter with clear mathematical meaning for an adjustable fusion strategy. By subtracting the user's cognitive load score, a negative correlation logic is established: low load corresponds to high weight, and high load corresponds to low weight. This aligns with the cognitive science principle that users rely more on habits under high load and are better able to process new information under low load. The scaling factor controls the sensitivity of the negative correlation, allowing the model to adapt to the varying degrees of cognitive load's impact on decision-making in different scenarios. The bias term sets the baseline weight level, ensuring that even under high cognitive load, long-term interests retain a minimum influence, preventing the model from completely deviating from the user's stable preferences due to short-term noise. This ensures that the dynamic fusion weight coefficients always remain within an effective probability range, producing stable and interpretable fusion weights regardless of the learned parameters, making the entire fusion process mathematically rigorous and controllable.

[0051] The expression for the dynamic fusion weight coefficient is: ; in, For dynamic fusion of weighting coefficients, Scaling factor Quantify the cognitive load of users into a score. This is a bias term.

[0052] S3.2. Based on the visual attention focus distribution vector, the content region is filtered from the real-time detected user short-term interest vector to obtain the filtered user short-term interest vector.

[0053] Furthermore, the visual attention focus distribution vector and the real-time detected user short-term interest vector are jointly input into the content region filtering process. Each dimension of the visual attention focus distribution vector corresponds to a specific semantic region in the user interface or content layout. The content region filtering process treats the value of each dimension in the visual attention focus distribution vector as the probability of the corresponding user's attention to that semantic region. The filtering process sets the dimensions in the visual attention focus distribution vector below a preset attention threshold to zero, retaining only those dimensions above the threshold that represent areas of significant user attention. Using this binary attention mask vector, the real-time detected user short-term interest vector is multiplied element-wise. In the real-time detected user short-term interest vector, short-term interest components associated with areas not of user attention are suppressed to zero, while short-term interest components matching areas of high user attention are retained, thus obtaining the filtered user short-term interest vector.

[0054] Specifically, it achieves precise spatial filtering from visual attention to semantic interest, establishing a location-semantic mapping and gating mechanism. The visual attention focus distribution vector is essentially a cognitive eye-tracking heatmap, indicating the entry point for user visual information intake. The content area filtering process assumes that only interest signals originating from areas within the user's field of vision possess high immediacy and genuine intent. For example, on a product details page, if the user's visual focus remains on areas with negative reviews, the resulting short-term interest stemming from quality concerns is far more valuable for decision-making than interest generated by highlighting areas where the user's gaze has not lingered. Through threshold filtering and masking multiplication, this process achieves attention gating, allowing only short-term interest signals verified through the attention channel to enter subsequent fusion. This effectively filters out interest noise caused by non-active attention such as autoplay and pop-up ads, ensuring that the short-term interests fused into the dynamic profile are products of the user's active cognitive input, thus improving the purity and credibility of the immediate intent components in the dynamic profile.

[0055] S3.3. The static long-term interest profile vector is weighted based on the dynamic fusion weight coefficient to obtain the adjusted static long-term interest profile vector.

[0056] Furthermore, the dynamic fusion weight coefficients and the static long-term interest profile vector are jointly input into the weighting process. The static long-term interest profile vector is a high-dimensional vector retrieved from the user's historical behavior database, representing the user's stable preference patterns across time and scenarios. The weighting process uses the dynamic fusion weight coefficients as a global scalar multiplier, performing element-wise scalar multiplication with the static long-term interest profile vector. This operation is equivalent to scaling each interest dimension component in the static long-term interest profile vector by the same proportion. The scaling ratio is determined by the dynamic fusion weight coefficients. When the dynamic fusion weight coefficients are high, the static long-term interest profile vector is largely preserved; when the dynamic fusion weight coefficients are low, the amplitude of the static long-term interest profile vector is correspondingly weakened, resulting in an adjusted static long-term interest profile vector whose vector direction remains unchanged, but whose overall amplitude (i.e., its contribution to the final fusion result) has been adaptively adjusted according to the user's real-time cognitive load level.

[0057] Specifically, global gain control based on cognitive state was implemented for long-term interests, demonstrating a deep understanding of how user decision-making patterns are constrained by cognitive resources. The static long-term interest profile vector represents a relatively stable preference structure formed after careful consideration by the user. The weighting process does not modify this structure itself, but rather adjusts the volume of this knowledge base in the current decision-making process by dynamically fusing weight coefficients. When the user's cognitive load is low and their mind is clear, the volume is increased, allowing their rich long-term experience and stable preferences to fully participate in the judgment of the current situation. When the user's cognitive load is high and their mental bandwidth is insufficient, the volume is decreased, because in this state, the user is more likely to rely on heuristic judgments or be dominated by immediate stimuli, and the guiding role of their long-term preferences is relatively weakened. The global scaling method is mathematically efficient and semantically intuitive, avoiding the complexity of setting weights for each interest dimension individually. At the same time, it grasps the essence of how cognitive load affects decision-making: it affects the individual's overall ability to call upon and rely on their internal stable knowledge system, rather than selectively forgetting certain specific interests. This makes the adjusted static long-term interest profile vector a context-sensitive long-term knowledge representation.

[0058] S3.4. Merge the adjusted static long-term interest profile vector with the filtered short-term user interest vector to generate a contextualized dynamic user profile vector.

[0059] Furthermore, the adjusted static long-term interest profile vector is added element-wise to the filtered short-term user interest vector. Before addition, a fixed short-term interest gain coefficient can be multiplied by the filtered short-term user interest vector according to a specific strategy to balance the potential difference in magnitude between long-term and short-term signals. The vector addition operation linearly superimposes the long-term stable preferences and the immediate intent from the current context in the same high-dimensional semantic space. The resulting vector is the contextualized dynamic user profile vector, which simultaneously encodes the user's steady-state interest profile and the instantaneous interest fluctuations after cognitive state and attention filtering correction, and naturally contains the user's real-time communication suitability information, because the weight coefficients that generate it are themselves derived from cognitive load assessment.

[0060] Specifically, linear superposition in vector space is used to achieve cognitive algebraic fusion of long-term memory and working memory. Essentially, this involves constructing a unified, dynamic user representation containing information across multiple time scales. Under the same semantic coordinate system, the decayed long-term preference vector and the purified short-term intention vector are synthesized. This allows the two signals to coexist, compete, and reinforce each other. For example, if the long-term preference vector has a high value in the high-end digital dimension and the short-term intention vector has a high value in the camera lens dimension, their addition further enhances the value in the camera lens dimension, generating a precise dynamic profile of the user currently focusing on the camera lens in high-end digital devices. If a signal in the short-term signal slightly contradicts the long-term preference but has been verified through attention, additive fusion can retain this contradiction, which may reflect the user's temporary comparative considerations or interest exploration. The generated contextualized dynamic user profile vector is thus a mathematical embodiment of the current user. It is neither a mechanical repetition of the past nor an overreaction to the moment, but rather a dynamic equilibrium achieved between the long-term self and the instantaneous context under cognitive state modulation. It is the most ideal and complete snapshot of the cognitive state for subsequent strategy deduction.

[0061] S4. Input the contextualized dynamic user profile vector into the counterfactual interpretation engine to extrapolate the information entropy density and time phase of the preset marketing actions, and identify the optimal parameter combination that resonates with the information transmission path and the user's cognitive rhythm.

[0062] S4.1 Input the contextualized dynamic user profile vector into the counterfactual interpretation engine, and construct the information entropy optimization space based on the real-time communication suitability label and interest focus in the contextualized dynamic user profile vector.

[0063] Furthermore, the contextualized dynamic user profile vector is input into the counterfactual interpretation engine. The engine parses the real-time communication suitability label attached to the contextualized dynamic user profile vector, which categorizes the user's state into categories such as unsuitable for communication, suitable for brief communication, or suitable for in-depth communication. Simultaneously, the engine extracts the principal components or high-amplitude dimensions of the contextualized dynamic user profile vector as the current user's focus of interest. Based on the user's cognitive resource boundaries defined by the real-time communication suitability label and the semantic direction pointed to by the focus of interest, the counterfactual interpretation engine dynamically constructs an information entropy optimization space. This space is a multi-dimensional parameter space, with coordinate axes defined by information entropy density parameters and time phase parameters. The boundaries and structure of the space are constrained by the real-time communication suitability label; for example, a high-load state corresponds to a low-value exploration range of the information entropy density parameter, ensuring that the inference is conducted within the user's cognitively tolerable boundaries.

[0064] Specifically, the counterfactual explanation engine defines the space based on state, and then, according to the real-time communication suitability labels carried by the contextualized dynamic user profile vector, it a priori and adaptively shrinks and shapes the feasible range of information entropy density and time phase. For example, when the label is "inappropriate for communication," the information entropy optimization space may completely shut down the immediate push area in the time phase dimension and restrict the high-information-volume area in the information entropy density dimension, forcing the inference engine to find solutions in the direction of delayed communication and simplified information. This is equivalent to installing a cognitive safety barrier for the strategy search process, ensuring that all considered virtual interventions are based on the ethics and effectiveness of respecting the user's current cognitive capacity, avoiding ineffective strategies that generate cognitive overload or harassing communication, and allowing subsequent inferences to focus on potentially efficient and humane solution areas from the beginning.

[0065] S4.2 Within the information entropy optimization space, the counterfactual explanation engine takes minimizing cognitive friction and maximizing intention convergence as its dual objectives, and performs collaborative perturbation and virtual simulation on the information entropy density parameter and time phase parameter of each preset marketing action in the preset marketing action template library.

[0066] Furthermore, within the constructed information entropy optimization space, the counterfactual explanation engine initiates an optimization process with the dual objectives of minimizing cognitive friction and maximizing intent convergence. A library of preset marketing action templates stores various basic marketing action templates, each associated with a set of adjustable initial values ​​for information entropy density and time phase parameters. For each preset marketing action template, the engine applies a series of small, structured cooperative perturbations to its information entropy density and time phase parameters, such as simultaneously increasing information entropy density and delaying the time phase, or decreasing information entropy density and advancing the time phase, generating various parameter combination variations. For each perturbated parameter combination, the counterfactual explanation engine performs an internal, virtual simulation without real user participation, simulating the expected sequence of impacts that the marketing action defined by that set of parameters might have on the cognitive state represented by the current contextualized dynamic user profile vector when executed.

[0067] Specifically, this involves collaborative perturbation and virtual simulation. Information entropy density and temporal phase are not independent variables, but rather coupled factors that jointly determine the intensity of information stimuli. High-density information coupled with the wrong timing (such as late at night) may trigger high cognitive friction; low-density information coupled with perfect timing may also miss opportunities to deepen intent. Collaborative perturbation simulates the joint exploration of these two key factors under dual-objective driving forces, seeking the perfect balance. Minimizing cognitive friction and maximizing intent convergence constitute the gold standard for evaluating strategies; the former ensures a smooth user experience, while the latter ensures the achievement of business goals. Virtual simulation is the core, enabling the counterfactual explanation engine to rapidly test and evaluate various if-then scenarios in an internal simulation environment without disturbing real users.

[0068] S4.3 The virtual simulation process simulates the expected impact of different combinations of information entropy density parameters and time phase parameters on the cognitive state represented by contextualized dynamic user profile vectors, and evaluates the matching degree between information transmission paths and user cognitive rhythms.

[0069] Furthermore, the virtual simulation process utilizes a lightweight effect prediction model pre-trained on historical feedback data. This model receives the current contextualized dynamic user profile vector and a specific combination of information entropy density and time phase parameters as input. Internally, the effect prediction model simulates the possible evolution path of the user's cognitive state after the marketing action defined by this parameter combination is applied, specifically predicting its expected impact on the user's cognitive load, attention focus, and core interest dimensions. Based on the predicted state changes, the evaluation process obtains a matching score. This matching score comprehensively quantifies the synchronization between the information delivery path and the user's cognitive rhythm, specifically reflected as a function combination of multiple intermediate indicators such as predicted information absorption efficiency, intent reinforcement, and cognitive resource utilization. A higher score indicates that the information delivery under this parameter combination is expected to integrate more smoothly into the user's current cognitive process, evoking more positive resonance rather than resistance.

[0070] Specifically, communication is modeled as a synchronization problem between information flow and cognitive flow. The effect prediction model serves as the implementation vehicle; it learns from historical data to summarize the micro-dynamics of how a user's state evolves under specific cognitive conditions and when faced with information of specific complexity and timing. Virtual simulation involves running this micro-dynamic model one step forward. The measure is whether the way information flow is injected aligns with the current capacity and direction of cognitive flow. For example, for users with high cognitive load and scattered attention, a push notification with low information entropy density, delayed until their load decreases, may have a high matching degree because it avoids current cognitive congestion. The matching degree function concretizes the two abstract goals of minimizing cognitive friction and maximizing intention convergence into calculable indicators, such as the predicted increase in cognitive load (negative), the improvement in attention focus, and the change in the dot product with the core interest vector, providing a unified, psychologically utility-oriented value scale for strategy comparison.

[0071] S4.4 By comparing the matching degree evaluation results under the parameter combination, the counterfactual explanation engine identifies the combination of information entropy density parameter and time phase parameter that achieves the maximum matching degree, and obtains the optimal parameter combination that resonates with the information transmission path and the user's cognitive rhythm.

[0072] Furthermore, the virtual simulation process calculates a matching degree evaluation result for each template in the pre-set marketing action template library based on multiple parameter combinations. The counterfactual interpretation engine collects and compares all these matching degree evaluation results corresponding to all parameter combinations. The comparison process typically includes sorting all matching degree scores or identifying the specific combination that achieves the highest matching degree score among all evaluated combinations. Once this parameter combination with the highest matching degree is identified, the counterfactual interpretation engine determines that this combination is the optimal parameter combination within the current information entropy optimization space, resonating with the user's cognitive rhythm and the information delivery path found for the current contextualized dynamic user profile vector. This optimal parameter combination explicitly includes the optimal information entropy density parameter and the optimal time phase parameter, which together define the core characteristics of the communication strategy that theoretically interacts most effectively and smoothly with the user's current psychological state.

[0073] Specifically, by conducting exhaustive or heuristic searches within a constrained information entropy optimization space, data and the model jointly determine what is optimal. When the matching degree reaches its maximum, the sweet spot between information complexity and push timing is found under a specific cognitive state. At this point, the information stimulus can be effectively processed by the user's cognitive system and transformed into positive intent to the greatest extent possible, i.e., resonance is generated. Resonance is an emergent property at the system level; it is not necessarily better to have higher information density or earlier timing, but rather a precise combination of both based on the current cognitive state. The process of identifying the optimal parameter combination is the ultimate manifestation of the counterfactual explanation engine's value. It condenses complex user psychological understanding, communication strategies, and causal effect predictions into a set of executable, personalized operational parameters, providing a deterministic, data-driven decision-making basis for the leap from understanding users to optimal action, realizing a paradigm shift in marketing automation from matching content to matching communication patterns.

[0074] S5. Based on the optimal parameter combination, output a personalized communication strategy package.

[0075] S5.1 Based on the information entropy density parameter in the optimal parameter combination, adjust the content presentation complexity of the preset marketing action template to generate the tuned marketing content.

[0076] Furthermore, the information entropy density parameter in the optimal parameter combination is used in the input content presentation complexity adjustment process. Each template in the preset marketing action template library is associated with a basic content version and a content complexity gradient mapping table. This mapping table defines the specific content presentation rule set corresponding to different information entropy density parameter values, such as rules for text length, number of information levels, richness of visual elements, and complexity of interactive controls. The adjustment process queries the corresponding rule set based on the input information entropy density parameter value and applies these rules to automatically reconstruct the basic content of the preset marketing action template. For example, a high information entropy density parameter may trigger the generation of long-form graphic content containing detailed feature comparisons, technical parameters, and multi-angle evaluations; while a low information entropy density parameter may trigger the generation of short copy or dynamic posters that only highlight the core selling points and calls to action. After this rule-driven content transformation, a tuned marketing content is generated that matches the target cognitive load level in terms of information volume, structure, and presentation format.

[0077] Specifically, through a predefined content grammar mapping table, it transforms into actionable, multimodal content presentation rules, achieving a precise translation of communication strategies from psychological parameters to physical content, driven by parameter-driven dynamic content generation. The information entropy density parameter represents the dosage requirements at the cognitive level, while the adjustment process is the automatic formulation of the reagent. The content complexity gradient mapping table acts as a recipe manual; it doesn't create new content but rather structurally combines, reduces, simplifies, and strengthens basic content materials (text, images, video clips, interactive components). For example, in the same product promotion, high-density parameters trigger a deep mode, automatically inserting links to technical white papers, third-party certification icons, and user case videos; low-density parameters trigger a quick-view mode, retaining only the main product image, a one-sentence value proposition, and a large purchase button, ensuring that content output is not random but strictly follows the guidance of cognitive load theory—providing rich information to support in-depth decision-making when user mental bandwidth is ample, and providing minimal information to reduce cognitive friction when bandwidth is limited. This makes the form of communication truly serve the purpose of adapting to cognitive states, realizing a new dimension of content personalization: adapting to the complexity of real-time cognitive abilities, rather than just adapting to topics based on interests.

[0078] S5.2 Based on the time phase parameter in the optimal parameter combination, set the planned push time point of the tuned marketing content, encapsulate the tuned marketing content and the planned push time point, and output a personalized communication strategy package.

[0079] Furthermore, the time phase parameter in the optimal parameter combination, along with the tuned marketing content generated in the previous step, is input into the plan setting and packaging process. The time phase parameter is a time offset relative to the current moment, or a future time point identifier predicted based on user behavior patterns. The plan setting process parses the time phase parameter to calculate the specific planned push time point for the tuned marketing content. This time point may be immediate, delayed by a specific duration, or the predicted next low cognitive load period for users (such as after work). The packaging process binds the tuned marketing content to this planned push time point and may attach metadata such as channel identifiers and user identifiers required for execution, packaging them together into a structured data object. This complete data object is the personalized communication strategy package, which contains all executable instructions on what to push and when to push, which can be directly read by downstream marketing execution channels to trigger corresponding delivery actions.

[0080] Specifically, the key temporal dimension of the communication strategy, time phase, is atomically bound and encapsulated with the tuned marketing content to create a self-executable strategy capsule with complete spatiotemporal attributes—the personalized communication strategy package. This lies in the spatiotemporal encapsulation of the strategy and the instructive nature of its execution. The time phase parameter is not an isolated delayed command, but rather an optimal interaction timing closely coupled with content complexity, derived from the deduction of the user's cognitive rhythm. The encapsulation process integrates content and timing into an indivisible strategy unit, ensuring that the optimal information entropy density is delivered at the optimal moment. As a standardized instruction set, the personalized communication strategy package achieves clear decoupling and an efficient interface between strategy generation and execution. The execution unit only needs to execute the content within the package according to the time points within it, without needing to understand the complex cognitive deduction logic behind it. Encapsulation makes the entire invention highly engineering-practical and scalable, enabling flexible integration with various existing marketing automation platforms, push notification services, or advertising systems. It represents a new type of marketing automation arsenal, which is not loaded with generic content materials, but with highly personalized communication solutions that have been jointly optimized by cognition, content and timing, realizing the evolution from pushing content to delivering a complete communication context.

[0081] S6. Execute marketing actions based on the resonance parameters in the personalized communication strategy package, and update the deep temporal neural network and counterfactual explanation engine based on feedback data.

[0082] S6.1 Based on the planned push time points in the personalized communication strategy package, execute the marketing actions of the adjusted marketing content at the corresponding time points.

[0083] Furthermore, the personalized communication strategy package is sent to a marketing action distributor with scheduled execution capabilities. The distributor parses the personalized communication strategy package, extracts the encapsulated time metadata (i.e., the planned push time), and enters a waiting state until that time arrives. When the planned push time is triggered, the distributor, based on the channel and user identifiers carried in the personalized communication strategy package, reads the tuned marketing content from the package and delivers it to the target user's terminal through the corresponding push channel, such as an app push server, SMS gateway, or advertising exchange platform, completing a full, time-triggered marketing action execution.

[0084] Specifically, the abstract, cognitively derived optimal time phase is rigorously converted into precise timing control in the physical world. This ensures that the theoretical strategy of cognitive rhythm resonance is faithfully reproduced in real-world interactions, strictly adhering to the cognitive dimension conclusion of the time phase parameter. This imbues the entire invention with temporal authenticity, meaning that the simulation in the virtual deduction... Real-time push notifications are precisely replicated in reality as... The physical basis for timely delivery, verification, and optimization of the effectiveness of the entire cognitive reasoning model.

[0085] S6.2 Collect user feedback data on the adjusted marketing content, including clicks, browsing time, and conversion behavior.

[0086] Furthermore, after the tuned marketing content is pushed out and executed, log collection components related to user terminal and server interactions are activated, asynchronously capturing and recording subsequent user interaction events in response to the push. The feedback data collection process monitors user behavior within a specific time window after receiving the tuned marketing content; these behaviors are categorized into the main components of the feedback data. Click behavior records whether and after how long the user triggered interactive elements in the content. Browsing duration records the total duration the user remained on the landing page guided by the marketing content. Conversion behavior records whether the user completed the preset core conversion goals, such as placing an order, submitting registration, or using key functions. These discrete event sequences, along with their timestamps and session identifiers, are aggregated, cleaned, and structured into a set of multi-dimensional feedback data strongly correlated with this marketing action, used to quantitatively evaluate the actual effectiveness of this personalized communication.

[0087] Specifically, the selection and structuring of feedback data closely serve to assess the achievement of the dual objectives of minimizing cognitive friction and maximizing intention convergence. Clicking behavior is the initial signal for capturing attention, browsing time indirectly reflects the matching degree between content and cognitive load (a high matching degree indicates a willingness to invest more cognitive resources), and conversion behavior is the ultimate evidence of intention convergence. By combining these three, feedback data can outline the complete micro-conversion funnel from information exposure and cognitive processing to decision-making action. Multi-dimensional feedback allows the model to learn, for example, that high browsing time but low conversion may indicate appropriate information complexity but insufficient intention guidance; low browsing time but high conversion may indicate excellent timing but the information could be further deepened. It also allows the model to learn which communication patterns are effective under which cognitive states, thereby achieving refined calibration of the cognitive state assessment model and strategy inference criteria.

[0088] S6.3 Combine the feedback data with the standardized user micro-behavior time series, contextualized dynamic user profile vectors, and optimal parameters used in the process of generating personalized communication strategy packages to form a feedback sample.

[0089] Furthermore, the standardized user micro-behavior time series, contextualized dynamic user profile vectors, and optimal parameter combinations used in generating personalized communication strategy packages are retrieved from temporary caches or process logs. These three sets of data record the complete historical inputs and decision outputs that triggered this marketing decision. The assembly process associates and binds this historical data with newly collected feedback data representing the outcome of this action, according to a pre-defined structured format. The standardized user micro-behavior time series, contextualized dynamic user profile vectors, optimal parameter combinations, and feedback data are sequentially arranged and encapsulated into a logically coherent data unit. This complete data unit includes the pre-decision state and decision basis, decision output, and decision result.

[0090] Specifically, this involves the mirroring and replaying of the entire decision-making process. Traditional machine learning models typically use isolated input-output pairs for training, losing the context in which decisions are made. Assembled feedback samples faithfully record the parameterized decisions made under different cognitive states and intent profiles, ultimately leading to specific user feedback. This is equivalent to creating a complete and traceable experimental archive for each decision. Standardized user micro-behavior time series represent the raw observations, contextualized dynamic user profile vectors represent internal cognitive state estimates, optimal parameter combinations represent decision actions, and feedback data represents environmental rewards. This sample structure enables end-to-end optimization in subsequent incremental training and allows for more refined ablation learning. For example, deep temporal neural networks can learn to more accurately predict cognitive states that lead to positive feedback from behavioral sequences; counterfactual explanation engines can learn to adjust the weights of their bi-objective functions so that the parameter combinations they derive, which achieve high matching scores, produce better feedback data in actual execution. The sample structure is crucial for enabling models to infer the merits of cognitive modeling and decision criteria from execution results, and is the core of transforming a business interaction into a high-quality learning experiment.

[0091] S6.4. Use feedback samples to incrementally train the deep temporal neural network, update the parameters of the deep temporal neural network, use feedback samples to adaptively adjust the weights of the dual objective function in the counterfactual explanation engine, and update the inference evaluation criteria of the counterfactual explanation engine.

[0092] Furthermore, feedback samples are input into the model update pipeline. For updating the deep temporal neural network, the standardized user micro-behavior time series in the feedback samples is used as input, while the cognitive utility label obtained by mapping the feedback data in the feedback samples through a reward function is used as part of the supervision signal to incrementally train the deep temporal neural network. The training process uses the mini-batch gradient descent algorithm to calculate the loss between the cognitive state predicted by the current network and the cognitive state that should lead to a better result, inferred from the actual feedback, and backpropagates to update the parameters of the deep temporal neural network. For updating the counterfactual explanation engine, the contextualized dynamic user profile vector in the feedback samples, the optimal parameter combination, and the actual feedback data are used together to calculate the deviation between the expected utility and the actual utility of the policy under the current dual-objective function weight configuration. Based on this deviation, the gradient method or metaheuristic algorithm is used to adaptively adjust the weights of the dual-objective function defined in the counterfactual explanation engine, thereby updating the inference evaluation criteria used inside the counterfactual explanation engine to evaluate the matching degree, making its inference of similar future scenarios more consistent with historically verified successful patterns.

[0093] Specifically, bidirectional optimization and criterion evolution target and optimize the core components of the model responsible for understanding and decision-making. The incremental training of deep temporal neural networks essentially teaches the perceptron to extract cognitive state features from behavior that are more predictive of which communication patterns will succeed—an evolution of feature representation. The weight adjustment of the counterfactual explanation engine represents the evolution of the decision-maker's internal evaluation criteria. It learns how to rebalance the importance of minimizing cognitive friction versus maximizing intention convergence in specific scenarios. For example, historical data might show that in certain situations, slightly increasing cognitive friction to provide more detailed information can significantly improve long-term conversion rates. In this case, the weights of the dual objective function will adaptively tilt towards maximizing intention convergence. This co-evolutionary mechanism transforms the entire solution from a static, deployment-and-determined model into a learning system that continuously optimizes its mental theories and communication skills by learning from every interaction with real users. It directly transforms the achievement of business goals (feedback) into the energy driving the evolution of cognitive and decision-making models, giving the invention a core vitality that maintains high adaptability and effectiveness in dynamic markets and changing user behavior.

[0094] This embodiment also provides a deep learning-based digital marketing user profile generation system, including: a standardization module, which performs standardization processing on the original fine-grained interaction event stream and outputs a standardized time series of user micro-behaviors; The evaluation module inputs standardized time series of user micro-behaviors into a deep temporal neural network and outputs a quantitative score of user cognitive load and a vector of visual attention focus. The fusion module, based on the user's cognitive load quantification score and visual attention focus distribution vector, performs weighted and focused fusion of static long-term interest profiles to generate contextualized dynamic user profile vectors. The deduction module inputs contextualized dynamic user profile vectors into the counterfactual interpretation engine to deduce the information entropy density and time phase of preset marketing actions, and identify the optimal parameter combination that resonates with the information transmission path and the user's cognitive rhythm. Assemble the module and output a personalized communication strategy package based on the optimal parameter combination; The learning module executes marketing actions based on the resonance parameters in the personalized communication strategy package and updates the deep temporal neural network and counterfactual explanation engine based on feedback data.

[0095] This embodiment also provides a computer device applicable to the deep learning-based digital marketing user profile generation method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the deep learning-based digital marketing user profile generation method proposed in the above embodiment.

[0096] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0097] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the deep learning-based digital marketing user profile generation method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0098] In summary, this invention achieves precise quantitative perception of users' instantaneous psychological states by standardizing the user's fine-grained interaction event stream and simultaneously extracting the user's real-time cognitive load quantification score and visual attention focus distribution vector using a deep temporal neural network. The quantified cognitive state is then weighted and fused with the user's static long-term interest profile to generate a dynamic profile reflecting the user's real-time context and communicability. By inputting this dynamic profile into a counterfactual interpretation engine, the information entropy density and time phase of preset marketing actions are deduced in the information entropy optimization space. The optimal parameter combination that resonates with the user's current cognitive rhythm is identified, generating directly executable personalized communication strategies. By executing these strategies and continuously updating the model using feedback data, a complete perception-decision-learning process is formed. This invention introduces deep modeling of real-time cognitive load and attention state into the user profile, upgrading it from describing what users are interested in to a dynamic strategy generation tool that guides when and how to communicate optimally. This solves the problem of insufficient adaptability in marketing communication timing and form caused by the lack of assessment of users' real-time psychological states in existing technologies, improving the accuracy of digital marketing, user experience, and communication efficiency.

[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for generating digital marketing user profiles based on deep learning, characterized in that: include, The original fine-grained interaction event stream is standardized to output a standardized time series of user micro-behaviors; The standardized time series of user micro-behaviors is input into a deep temporal neural network, which outputs a quantitative score of user cognitive load and a distribution vector of visual attention focus. Based on the user's cognitive load quantification score and visual attention focus distribution vector, the static long-term interest profile is weighted and focused to generate a contextualized dynamic user profile vector. By inputting contextualized dynamic user profile vectors into the counterfactual interpretation engine, the information entropy density and time phase of preset marketing actions are extrapolated to identify the optimal parameter combination that resonates between the information delivery path and the user's cognitive rhythm. Based on the optimal parameter combination, a personalized communication strategy package is output. Marketing actions are executed based on the resonance parameters in the personalized communication strategy package, and the deep temporal neural network and counterfactual explanation engine are updated based on feedback data.

2. The method for generating digital marketing user profiles based on deep learning as described in claim 1, characterized in that: The original fine-grained interaction event stream is standardized to output a standardized time series of user micro-behaviors, including the following steps: Data cleaning is performed on the original fine-grained interaction event stream to obtain the cleaned original fine-grained interaction event stream, and session segmentation is performed to obtain independent session fragments; Perform time alignment on each individual session segment to obtain time-aligned session segments, and perform feature vectorization to transform the event at each time point into a feature vector; Perform sequence standardization on the feature vectors to output a standardized time series of user micro-behaviors.

3. The method for generating digital marketing user profiles based on deep learning as described in claim 2, characterized in that: The standardized time series of user micro-behaviors is input into a deep temporal neural network, which outputs a quantitative score of user cognitive load and a vector of visual attention focus. The steps include: Standardized user micro-behavior time series are input into the one-dimensional convolutional layer of a one-dimensional convolutional gated recurrent hybrid neural network to extract local temporal features, resulting in a local temporal feature vector; The local temporal feature vector is input into the gated recurrent unit layer of a one-dimensional convolutional gated recurrent hybrid neural network to capture long-term dependencies and obtain long-term dependency feature vectors. The long-time dependent feature vector is input into the multi-head self-attention mechanism layer of a one-dimensional convolutional gated recurrent hybrid neural network and then subjected to attention weighting across time steps to obtain attention-weighted feature vectors. The attention-weighted feature vector is fed into the cognitive load score output head of a one-dimensional convolutional gated recurrent hybrid neural network, and mental bandwidth assessment and neural information flow compression perception are performed on the attention-weighted feature vector to generate a user cognitive load quantification score. The attention-weighted feature vector is fed into the attention distribution output head of a one-dimensional convolutional gated recurrent hybrid neural network, and visual field reconstruction and spatial attention weight diffusion are performed on the attention-weighted feature vector to generate a visual attention focus distribution vector.

4. The method for generating digital marketing user profiles based on deep learning as described in claim 3, characterized in that: Based on the user's cognitive load quantification score and visual attention focus distribution vector, a weighted and focused fusion of static long-term interest profiles is performed to generate contextualized dynamic user profile vectors, including the following steps: Calculate the dynamic fusion weight coefficient based on the user's cognitive load quantification score; Based on the visual attention focus distribution vector, the content region is filtered from the real-time detected user short-term interest vector to obtain the filtered user short-term interest vector. The static long-term interest profile vector is weighted based on the dynamic fusion weight coefficient to obtain the adjusted static long-term interest profile vector. The adjusted static long-term interest profile vector is merged with the filtered short-term user interest vector to generate a contextualized dynamic user profile vector.

5. The method for generating digital marketing user profiles based on deep learning as described in claim 4, characterized in that: The contextualized dynamic user profile vector is input into the counterfactual interpretation engine to extrapolate the information entropy density and time phase of preset marketing actions, and to identify the optimal parameter combination that resonates between the information delivery path and the user's cognitive rhythm. This includes the following steps: Input the contextualized dynamic user profile vector into the counterfactual interpretation engine, and construct the information entropy optimization space based on the real-time communication appropriateness label and interest focus in the contextualized dynamic user profile vector; Within the information entropy optimization space, the counterfactual explanation engine takes minimizing cognitive friction and maximizing intention convergence as its dual objectives, and performs collaborative perturbation and virtual simulation on the information entropy density parameter and time phase parameter of each preset marketing action in the preset marketing action template library. The virtual simulation process simulates the expected impact of different combinations of information entropy density parameters and time phase parameters on the cognitive state represented by contextualized dynamic user profile vectors, and evaluates the matching degree between information delivery path and user cognitive rhythm. By comparing the matching degree evaluation results under different parameter combinations, the counterfactual explanation engine identifies the combination of information entropy density parameter and time phase parameter that maximizes the matching degree, and obtains the optimal parameter combination that resonates between the information transmission path and the user's cognitive rhythm.

6. The method for generating digital marketing user profiles based on deep learning as described in claim 5, characterized in that: Based on the optimal parameter combination, a personalized communication strategy package is output, including the following steps: Based on the information entropy density parameter in the optimal parameter combination, the content presentation complexity of the preset marketing action template is adjusted to generate the tuned marketing content. Based on the time phase parameter in the optimal parameter combination, the planned push time of the tuned marketing content is set, and the tuned marketing content and the planned push time are encapsulated to output a personalized communication strategy package.

7. The method for generating digital marketing user profiles based on deep learning as described in claim 6, characterized in that: Marketing actions are executed based on the resonance parameters in the personalized communication strategy package, and the deep temporal neural network and counterfactual interpretation engine are updated based on feedback data, including the following steps: Based on the planned push times in the personalized communication strategy package, execute marketing actions for the adjusted marketing content at the corresponding times; Collect user feedback data on the adjusted marketing content, including clicks, browsing time, and conversion behavior; The feedback data is combined with the standardized user micro-behavior time series, contextualized dynamic user profile vectors and optimal parameters used in the process of generating personalized communication strategy packages to form a feedback sample. The deep temporal neural network is incrementally trained using feedback samples to update its parameters. The weights of the dual objective function in the counterfactual explanation engine are adaptively adjusted using feedback samples to update the inference evaluation criteria of the counterfactual explanation engine.

8. A deep learning-based digital marketing user profile generation system, based on the deep learning-based digital marketing user profile generation method according to any one of claims 1 to 7, characterized in that: This includes a standardization module, which standardizes the raw fine-grained interaction event stream and outputs a standardized time series of user micro-behaviors. The evaluation module inputs standardized time series of user micro-behaviors into a deep temporal neural network and outputs a quantitative score of user cognitive load and a vector of visual attention focus. The fusion module, based on the user's cognitive load quantification score and visual attention focus distribution vector, performs weighted and focused fusion of static long-term interest profiles to generate contextualized dynamic user profile vectors. The deduction module inputs contextualized dynamic user profile vectors into the counterfactual interpretation engine to deduce the information entropy density and time phase of preset marketing actions, and identify the optimal parameter combination that resonates with the information transmission path and the user's cognitive rhythm. Assemble the module and output a personalized communication strategy package based on the optimal parameter combination; The learning module executes marketing actions based on the resonance parameters in the personalized communication strategy package and updates the deep temporal neural network and counterfactual explanation engine based on feedback data.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based digital marketing user profile generation method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based digital marketing user profile generation method according to any one of claims 1 to 7.

Images