Advertisement intention generation system and method based on user mental driving

Abstract

The invention relates to the technical field of advertisement design and generation, and relates to an advertisement intention generation system and method based on user mental driving, and the system comprises a user mental insight module, an intention strategy mapping module, an intention creativity generation module, a creativity dynamic evolution module, and a cross-domain mental creativity migration function. A mapping relation among a user behavior mode, a psychological state and a potential intention is constructed, and an intention strategy mapping module applies a mental intervention strategy to a user intention node; the intention creativity generation module generates advertisement creativity content capable of touching or reconstructing the intention node of the target user according to a preset mental intervention strategy through a combinatorial optimization algorithm; the creative dynamic evolution module is used for adaptively adjusting the user mental state model, the intention strategy mapping mechanism and model parameters and creative gene segments in the creative generation module, and the advertisement creative which can accurately reach and guide the psychological process of the user is generated through a mental model decoupling and reconstruction mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advertisement design generation, more specifically, it relates to a user mind-driven advertisement intention generation system and method. BACKGROUND

[0002] In recent years, with the rapid development of artificial intelligence (AI), big data analysis and deep learning technology, advertisement creative generation and optimization have entered the era of intelligence. Traditional advertisement creativity relies on the experience, inspiration and judgment of market trends of designers, which is not only inefficient and costly, but also difficult to achieve large-scale personalized creative output, and the insight into the changing market demand and massive user data is also relatively limited.

[0003] In order to solve the above problems, existing technologies continue to explore the use of AI technology to improve the efficiency of advertisement design and generation. For example, in Chinese patent application CN119583878A, an advertisement intelligent design system based on big data is disclosed. The system collects product names, associated materials, uses historical advertisement data and audience ratings to build a scoring model, and adjusts the advertisement appearance time, filters key frames, sets picture and sound parameters based on this, and finally generates an advertisement segment. Through the automatic process, the system aims to reduce manual operation, optimize the rhythm and length of the advertisement, and improve the attractiveness and viewing experience of the advertisement.

[0004] In addition, Chinese patent application CN119006055B also proposes an advertisement content intelligent design method and system. The patent describes an advertisement content generation model based on U-net architecture, which includes N base generators and a global fully connected layer, and can receive advertisement demand vectors (including text, images and style features), generate advertisement content with different advertisement styles through deconvolution and feature fusion operations. The system introduces a noise vector to enhance the diversity of generated content, and obtains an advertisement style feature map through clustering analysis, and then uses a group optimization algorithm to obtain a fusion weight to realize personalized generation of advertisement content.

[0005] The above-mentioned existing technologies have made significant progress in using big data and AI to realize the automatic matching, combination, timing optimization and even preliminary stylization generation of advertisement materials, effectively improving the efficiency of advertisement design, reducing production costs, and to some extent, achieving preliminary personalization. However, these existing technologies still mainly focus on the analysis, matching and generation stage of explicit advertisement elements (such as product features, user tags, popular styles), and the generation logic of their advertisement creativity is more based on data-driven statistical rules and pattern recognition.

[0006] However, the core value of an advertising creative lies in its deep insight into the psychological motivation of consumers, the establishment of emotional connection and the effective guidance of user behavior. Existing methods for intelligent design of advertisements often focus on processing and optimizing explicit and quantifiable data, such as matching more materials, adjusting the length, and applying existing style templates, but they rarely involve the excavation of deep psychological logic and potential motivation behind user behavior, nor do they reconstruct the creative itself based on mental drive. The generation of user behavior is often the result of the interaction of complex psychological activities such as demand, cognition, emotion and decision-making. Simply matching patterns based on user tags or historical behavior sequences cannot touch the user's hidden, nonlinear mental process. Therefore, existing technologies cannot generate high-level creatives that are truly groundbreaking and can touch the user's deep psychology, thus effectively driving user action.

[0007] Although user portraits and historical advertising data are used in the existing technology during the creative generation process, the analysis model is more for predicting user preferences for specific styles or materials rather than deeply understanding the behavior driving mechanism of users in different mental states. There is also a lack of a mechanism to actively and subtly intervene and reconstruct the user's mental model to achieve the advertising goal. In addition, the existing technology has not provided an effective technical solution for how to accurately adapt and guide the advertising creative to the user's specific psychological cognitive nodes, emotional trigger points and final behavior decision chain.

[0008] Therefore, it is necessary to develop an innovative advertising creative generation technology that can break through the limitations of existing technologies relying only on explicit data and pattern matching, deeply excavate the mental structure and subconscious motivation behind user behavior, and map this to the promotional action goal of the advertisement. Through the mechanism of mental model decoupling and reconstruction, an advertising creative that can accurately reach and guide the user's psychological process is generated. SUMMARY

[0009] To this end, the purpose of the present application is to provide a user mind-driven advertising intention generation system and method, which generates an advertising creative that can accurately reach and guide the user's psychological process through the mechanism of mental model decoupling and reconstruction.

[0010] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0011] A user mind-driven advertising intention generation system, the system comprising:

[0012] A user mind insight module for analyzing user information and constructing the mapping relationship between user behavior patterns, psychological states and potential intentions through multi-modal data; the user mind insight module is specially provided with:

[0013] A mental state encoder is configured to receive explicit behavior sequence data, physiological feature signals and / or speech tone data of a user, and output a high-dimensional user mental state vector through deep neural network processing The vector encodes the user's immediate psychological state

[0014] An intention node decoupler is configured to analyze and decouple an intention node set representing the user's core psychological driving force , which respectively correspond to the user's motivation, cognition and emotion three dimensions of latent intention

[0015] An intention strategy mapping module is configured to receive an advertising target set by an advertiser, and accurately associate the advertising target with the intention node set of the target user group, to determine the mental intervention strategy applied to the user intention node to achieve the advertising target

[0016] An intention creative generation module is configured to receive the user's intention node set and target intervention strategy, decouple the preset key advertising creative elements into creative gene fragments, and then optimize the combination mode and action energy of the creative gene fragments according to the predetermined mental intervention strategy through a combination optimization algorithm, to generate advertising creative content that can touch or reconstruct the target user intention node

[0017] A creative dynamic evolution module is configured to receive effect feedback data after the actual placement of the advertisement, and based on the feedback data, implement closed-loop iterative optimization to adaptively adjust the model parameters and creative gene fragments in the user mental state model, intention strategy mapping mechanism and creative generation module

[0018] The application further provides that the mental state encoder of the user mental insight module captures the context association and potential semantic information in the user behavior sequence

[0019] The intention node decoupler uses a variational autoencoder to map the high-dimensional mental state vector

[0020] to a low-dimensional latent space, from which the interpretable core intention nodes are extracted .

[0021] The application further provides that the mental state encoder uses a self-attention mechanism to assign dynamic attention weights to the input physiological feature signals and / or speech tone information when processing the user behavior sequence data, to highlight the psychological suggestion information highly related to the current user behavior pattern and to weight and integrate it into the user mental state vector .

[0022] The application is further configured that the intention strategy mapping module is provided with a mapping rule of target-intention driven intervention; the rule decomposes the overall target of the advertising task into the mental influence strength requirement on the specific intention node ,

[0023] That is, , is the intention node subset most related to the advertising target.

[0024] The application is further configured that the intention creative generation module comprises a creative gene library, which stores the minimum creative feature units with independent psychological influence, each segment is expressed in the form of semantic attribute vector and psychological drive vector,

[0025] The combination optimization algorithm selects the optimal combination of creative gene segments from the creative gene library through a generative adversarial network to maximize the mental touch effect of the generated advertising content on the target intention node , ,

[0026] That is .

[0027] The application is further configured that the combination optimization algorithm comprises a subconscious anchoring optimization mechanism, which identifies the key intention node in the user's mental state vector that is most sensitive to the target intervention strategy ;

[0028] And on this sensitive node, the combination of creative gene segments that can produce the maximum cognitive activation or emotional resonance is preferentially generated.

[0029] The application is further configured that the creative dynamic evolution module realizes a user mental feedback closed loop; the closed loop comprises:

[0030] The mental state prediction unit receives the current intention node state of the user and the generated advertising creative content, and predicts the mental state evolution trajectory that the user may generate thereafter ;

[0031] The effect evaluation and calibration unit compares the actual advertising delivery effect data, including user conversion rate, stay time, emotional feedback, etc. as real evolution signals with the predicted trajectory; the system adopts a gradient descent model to calibrate and update the core parameters of the user mental state model, the intention strategy mapping model and the creative generation model,

[0032] If the advertising content fails to effectively activate the target intention node, the system will adjust the combination weight of the creative gene segments used in the content generation process. ​​

[0033] The application is further provided that: the creative gene in the creative dynamic evolution module self-adapts learning mechanism, quantifies the user mind feedback signal as a reward signal R of reinforcement learning;

[0034] Using the reward signal, the generation model is optimized by SARS algorithm to learn the optimal creative gene fragment selection and combination strategy.

[0035] The application is further provided that: it contains cross-domain mind creative transfer function; the function refers to that the system first learns creative elements driving user psychological reaction and their performance rules in different mind states from non-advertising fields;

[0036] Then, these learned psychological activation patterns and creative combination paradigms are applied to advertising generation to create more powerful and memorable advertising content, thereby improving the breaking through propagation ability of advertising.

[0037] A user mind driven advertising intention generation method, comprising the following steps:

[0038] S1, mind decoupling and intention recognition: analyzing user multi-modal data, accurately identifying the core intention node of the user;

[0039] S2, target-intervention strategy derivation: deriving specific mind intervention strategies required for the intention nodes according to the advertising target;

[0040] S3, intention creative generation: combining and optimizing creative gene fragments according to the identified intention nodes and derived mind intervention strategies to generate advertising creative content capable of producing specific mind influence;

[0041] S4, effect feedback and model self-optimization: real-time tracking of advertising delivery effect, and using feedback data for dynamic calibration and optimization of user mind model, intention strategy and creative generation model, forming a continuously evolving intelligent generation capability.

[0042] Compared with the prior art, the application has the following advantages:

[0043] Through in-depth user mind insight, the fuzzy user psychological state is converted into operable intention nodes, and accurate intervention strategies are formulated accordingly. On this basis, the system uses innovative creative gene concept and powerful combination optimization algorithm (especially subconscious anchoring mechanism) to generate highly personalized and psychologically touching advertising content. Most importantly, the system can continuously learn and evolve through dynamic evolution module and cross-domain creative transfer, realizing intelligent, automated and maximized advertising effect. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1System framework diagram of the present application. DETAILED DESCRIPTION

[0045] Reference Figure 1 Further description is made to an embodiment of the present application, a user mind-driven advertisement intention generation system and method, which specifically comprises: a user mind insight module, an intention strategy mapping module, an intention creative generation module, a creative dynamic evolution module, and a cross-field mind creative migration function.

[0046] The user mind insight module is to understand the deep psychological state and potential intention of the user from multi-dimensional and multi-modal user data.

[0047] The mind state encoder is responsible for receiving the original multi-modal data input by the user and converting it into a high-dimensional vector that can quantify and summarize the user's immediate psychological state. The input data includes user behavior, physiological signals, and speech tone, which are external manifestations of the user's internal state.

[0048] The user explicit behavior sequence data includes the user's click stream, page browsing history, search query, interaction (comments, likes, sharing), purchase record, APP usage path, etc. These data are integrated through time series to form an ordered behavior sequence.

[0049] The physiological feature signals include heart rate (HR), heart rate variability (HRV), skin conductivity (GSR / EDA), brain waves (EEG, if available), eye tracking data (gaze duration, eye movement pattern), etc. These signals directly reflect the user's physiological arousal level, emotional tension, attention focus, etc.

[0050] The speech tone data includes speaking speed, pitch, intensity, intonation, pause frequency, etc. These are important clues for assessing the user's emotional state (such as excitement, frustration, anger, calm) and cognitive load.

[0051] In order to process these heterogeneous and time-series data, a multi-modal fusion deep learning architecture is adopted, based on Transformer or a special multi-modal fusion network.

[0052] Behavior sequence processing: Consider using long short-term memory network (LSTM), gated recurrent unit (GRU), or Transformer Encoder. Transformer can better capture the contextual relevance and potential semantic information in the behavior sequence (such as the intention conversion of the user in a specific scenario) due to its self-attention mechanism, without relying on the sequence order like RNN (although the order is important).

[0053] For example, a user's continuous clicks on product A compared to a search for 'A brand review' can reveal more about their underlying cognition (validation) or motivation (purchase decision).

[0054] The feedforward network performs further nonlinear transformation on the self-attention output. Layer normalization and residual connection stabilize training and accelerate convergence. After multiple layers of Transformer Encoder, the output is denoted as. For continuous physiological signals (e.g., HRV, GSR), a convolutional neural network (CNN) can be used to extract local time-frequency features, or a one-dimensional CNN / RNN can be used for physiological signal input.

[0055] Voice prosody data processing: Acoustic features are often extracted first and then input into RNN / LSTM / GRU or Transformer.

[0056] Fusion strategies:

[0057] Early fusion: Concatenate feature vectors of different modalities at the input layer or the first layer (if they are at the same time step).

[0058] Mid-level fusion: After independent encoding of each modality, fusion is performed at the intermediate layer, for example, through an attention mechanism such as the attention described above to highlight the psychological cue information.

[0059] Late fusion (LateFusion): Each modality is independently encoded and classified / regressed by its own classifier / regressor, and the final output is fused.

[0060] When considering the dynamic weights of physiological and voice data on the behavior sequence, an attention mechanism can be used. After processing by multiple layers of deep network (e.g., Transformer layer stacking or fused RNN / Transformer), the feature sequences of all modalities are aggregated into a fixed-length high-dimensional user mental state vector This vector encodes the user's comprehensive psychological state at the current time, including their possible emotions, attention levels, cognitive load, interest levels, etc.

[0061] Intention node decoupler: This module receives the high-dimensional vector output by the mental state encoder and uses a latent variable model (LVM) to decouple it into a set of intention nodes representing the user's underlying core psychological driving forces, which correspond to the three dimensions of user motivation, cognition, and emotion.

[0062] Variational autoencoder (VAE) is an ideal choice because it can map high-dimensional data to a low-dimensional, statistically meaningful latent space and allow us to extract interpretable parts from it.

[0063] An encoder maps high-dimensional mental state vectors to a distribution over a latent space,

[0064] The encoder learns parameters to predict the mean and variance (or log variance ) of the latent variable :

[0065] Here, the latent variable z is a low-dimensional vector with dimensionality less than This low-dimensional latent space is designed to carry the expression of the user's core intent.

[0066] The latent space z of the VAE is designed to capture a low-dimensional, meaningful representation of the data. In this system, this latent space is further structured into three independent (or weakly correlated) intent dimensions, each corresponding to a user intent node:

[0067] Motivation node (M): represents the user's intrinsic drive, goal orientation, need satisfaction (e.g. want to buy, want to learn, want to entertain). Cognition node (C): represents the user's thinking pattern, information processing style, decision logic (e.g. focus on price, emphasize quality, need detailed information, easily persuaded).

[0068] Emotion node (E): represents the user's emotional state, affective preference, psychological pleasure (e.g. like surprises, hate interruptions, pursue novelty, seek security).

[0069] This decoupling can be achieved by introducing a structured latent space in the design of the VAE: design a VAE whose latent space z is explicitly divided into three subspaces:

[0070] By adding a regularization term to the loss function, encourage the three subspaces to independently capture information in their respective dimensions.

[0071] The loss function:

[0072]

[0073] where is the divergence weight, to achieve decoupling, modify the KL divergence term: Here

[0074] , ,​​​​ It can be understood as Mapping to three independent or partially independent low-dimensional spaces.

[0075] Or, more directly, if z is a vector, we can split it into And impose KL divergence constraints on each part separately, or let them be as independent as possible in the encoding and decoding process.

[0076] From a specific dimension or subspace in the low-dimensional latent space z, extract nodes that represent the user's core intentions:

[0077]

[0078]

[0079] Here Can be a simple linear mapping, MLP, or a specific sub-vector extracted from z. These nodes can be continuous vector representations, or they can be discretized into intention categories (e.g. The value of

[0080] Mathematical representation: (d is much smaller than D).

[0081] The latent space of VAE has certain interpretability, meaning that changes in latent dimensions (here, intention nodes) correspond to understandable changes in user state. If The value of a certain dimension increases, it may indicate that the user's purchase intention is increasing; if The value of a certain dimension changes, it may indicate that the user's preference for information depth has changed.

[0082] Intention strategy mapping module: This module bridges the marketing goals of advertisers and the intrinsic psychological needs of users. It receives the advertising goals set by advertisers and accurately associates them with the intention node set of the target user group (obtained from the user mind insight module), thereby deriving specific mind intervention strategies.

[0083] The overall goal of the advertiser (such as: improving brand awareness, promoting product sales, increasing user activity) is first analyzed, and a product sales promotion goal can be further divided into: improving user awareness of product value, stimulating user purchase desire, reducing user purchase concerns, etc.

[0084] The system establishes a target-driven intervention mapping rule, which divides the overall goal of the advertising task Into the mind influence intensity of specific intention nodes ​ ) requires, wherein, is the subset of intent nodes that are most relevant (i.e., critical to influence) to the ad goal.

[0085] is a quantitative indicator of how much psychological influence needs to be exerted on the intent nodes to drive user behavior towards the ad goal. This strength indicator can be a continuous value from 0 to 1, or a more granular five-level / ten-level rating.

[0086] Based on the above mapping relationship, the system formulates a series of targeted mental intervention strategies for the intent node set of the user group, which define how to manipulate or reconstruct the user's intent nodes through ad creatives and presentation methods to make them develop towards the direction of the ad goal.

[0087] If the goal is to increase sales, and the user shows weak purchase desire, the intervention strategy strengthens the perception of scarcity or highlights the use value .

[0088] If the goal is to enhance brand trust, the user shows information uncertainty, and the intervention strategy provides authoritative endorsement or displays detailed user reviews .

[0089] If the goal is to enhance brand favorability, the user shows indifference or negative emotions, and the intervention strategy arouses positive emotional associations or creates surprise experiences .

[0090] This module usually contains a knowledge graph or rule engine that stores the relationships between different ad goals and user intent nodes, as well as the corresponding intervention strategy library for each relationship.

[0091] Intentional creative generation module: This is the core link of converting insights and strategies into specific ad content. This module receives the user's intent node set and the determined mental intervention strategy, decouples the preset key ad creative elements (such as visual elements, copy, music, interaction forms, etc.) into minimal creative gene fragments, and through combination optimization algorithms, optimizes the combination way of creative gene fragments and their role energy in the creative according to the target strategy, to generate ad creative content that can significantly touch or reconstruct the target user's intent nodes.

[0092] Creative gene library: The creative gene library stores the smallest creative feature units (creative gene fragments) that have independent psychological influence and can be combined. Each fragment is represented in structured data, including:

[0093] Semantic attribute vector: describes the surface meaning, content attribute of the creative snippet. A visual snippet can be a warm-toned family scene, a copy snippet can be a time-limited offer, not to be missed.

[0094] Psychological drive vector: [highlight] describes the potential psychological impact direction and strength of this snippet on the user's specific intent node (motivation, cognition, emotion). For example:

[0095] Snippet A: [family warm picture + family reunion moment copy] -> psychological drive vector:

[0096] Snippet B: [countdown to the last 24 hours + immediate purchase CTA] -> psychological drive vector:

[0097] Selection criteria for creative gene snippets: as fine-grained as possible, such as a word, a sentence, a picture element, a transition effect.

[0098] Combination optimization algorithm: find the optimal combination of creative gene snippets , so that the generated advertising content has the greatest impact on the target intent node specified by the intent strategy mapping module . Generative adversarial network (GAN); the generator receives a noise vector (or the representation of the combined creative gene snippets here) and conditional information (target intent node, intervention strategy), and generates advertising creative content (such as images, videos, text); the discriminator (Discriminator, D) tries to distinguish whether the generated content is real (from the real advertising creative library) or generated.

[0099] Optimization goal: G aims to generate content that can deceive D, while maximizing .

[0100] Optimization objective function: , is the set of gene snippets selected and combined from the creative gene library.

[0101] The measure is the reach effect of the combined creative on the target intent node , and this performance function can be output by a pre-trained psychological impact evaluation model.

[0102] The model can be a separate prediction model, input the sum of the psychological drive vectors (or after more complex interaction models), and output the impact strength on .

[0103] Assume where is a creative gene segment. Each has a psychological drive vector (M is the intention node dimension, 3 nodes each have multiple sub-dimensions).

[0104]

[0105] where Combine can be a simple vector addition, weighted average, or a more complex learning function that considers the interaction between gene segments:

[0106]

[0107] Based on the target intention node and intervention strategy, preliminarily filter the gene segments that match the target, and use genetic algorithms, simulated annealing, reinforcement learning, or GAN-based generation to search for the optimal .

[0108] In the process of combination optimization, this mechanism aims to identify the user's mental state vector , which is most sensitive to the target intervention strategy (i.e. the key intention node ). Once this trigger point is identified, the algorithm will prioritize generating creative gene segment combinations that can produce the maximum cognitive activation or emotional resonance at this node.

[0109] Identify Analyze Relevance to intervention strategy: check the relevance or sensitivity of each dimension of (particularly the decoupled intention nodes ) to the current intervention strategy (the goal is to affect a certain node ).

[0110] This can be done by training a small regression model to predict the loudness of a certain dimension to the target intervention strategy, or directly analyzing which intention node in has the current value that best or least meets the requirements of the intervention strategy.

[0111] Once is determined, the combination optimization algorithm (whether it is a GAN generator or other search algorithm) will adjust its search direction and prioritize exploring creative gene segment combinations that produce the maximum cognitive activation (such as triggering users to think deeply, solving certain cognitive conflicts) or emotional resonance (such as touching users' specific emotional memories, creating strong emotional connections).

[0112] This mechanism ensures that the ad creative can hit the user's current psychological pain points or itch points, achieving the best dissemination effect.

[0113] Creative dynamic evolution module: This module is the learning brain of the system, responsible for closed-loop iterative optimization of the entire system based on user feedback data after the actual placement of the ad. Through continuous adaptive adjustment, it continuously improves the accuracy of the user's mental state model, the effectiveness of the intent strategy mapping, and the generation quality of the creative generation model.

[0114] User mental feedback loop contains two core units:

[0115] Mental state prediction unit: receives the user's current intent node state (output from the intent node decoupler) and the generated (or placed) ad creative content to predict the user's mental state evolution trajectory after that , usually a sequence model (Transformer) or dynamic Bayesian network that learns the user's psychological state change pattern after receiving a specific type of creative stimulus.

[0116] Input:

[0117] Output:

[0118] Effect evaluation and calibration unit: compares the actual effect feedback data after ad placement with the predicted trajectory, calculates the error, and updates the model parameters accordingly.

[0119] Actual effect data includes user conversion rate (click, purchase), dwell time, interaction rate, viewing completion, emotional feedback (subjective rating, comment sentiment analysis), and repeat purchase rate.

[0120] Model calibration uses optimization algorithms such as gradient descent (or its variants such as Adam, RMSprop) to update the core parameters of the user's mental state model (such as the encoder, decoupler), the intent strategy mapping model (such as the mapping rule library, association weight), and the creative generation model (such as the generator / discriminator of GAN, gene selection algorithm).

[0121] Prediction error: the core is to calculate the difference between the actual observed user mental state and use it as a loss term,

[0122] Loss function:

[0123]

[0124] Where, measures the effectiveness of the creative For target intent node The gap between the actual impact and the expected impact.

[0125] If the advertising content is The combined generation failed to effectively activate the target intent node. This means that the combination or gene fragment has low efficiency.

[0126] The system will adjust the combined weights of the creative gene fragments used in the content generation process. This may mean reducing the weight of invalid fragments or adding others that are more likely to trigger [the content]. The weight of fragments is determined, and even in the next generation, fragments that are opposite to or complementary to the current inefficient fragments are given priority.

[0127] Creative Gene Adaptive Learning Mechanism: This mechanism quantifies user mental feedback signals (i.e., advertising effectiveness) into a reward signal R for reinforcement learning (RL), and uses the RL algorithm to optimize the generative model, enabling it to learn better creative gene fragment selection and combination strategies.

[0128] SARSA is an online, on-policy reinforcement learning algorithm that considers the current state, current action, reward, next state, and next action simultaneously when updating its policy. This allows it to learn smoother, more conservative policies, avoiding extreme choices in uncertain situations.

[0129] State: The current state of the user's intent node. In addition to contextual information such as the current iteration step and advertising stage, specific creative gene fragment combinations are selected. (It could be a strategy, i.e., a probability distribution, to select which segments).

[0130] SARSA Algorithm Flow:

[0131] Initialization: Strategy or neural network parameters ).

[0132] Starting from the initial state S; according to the current strategy Choose an action A (i.e., choose a creative gene fragment combination); execute action A, observe the reward R and the next state S'.

[0133] SARS-A core strategy Select the next action A' in state S'.

[0134] Update strategy:

[0135]

[0136] Training a neural network to estimate Q values,

[0137]

[0138] where are parameters of the target network, let repeat until termination.

[0139] Through SARSA, the system can learn a policy that maximizes cumulative rewards over its lifetime, i.e., generates the most effective creative combinations for the user intent nodes.

[0140] Cross-domain mental creative transfer function: By learning the creative elements in non-advertising domains (such as art, film, literature, games, social media content, etc.) that can effectively drive user psychological responses and their performance patterns in different mental states, and then transferring these mature psychological activation patterns and creative combination paradigms to advertising generation. The ultimate goal is to create more powerful, novel, memorable, and viral advertising content, breaking through the limitations of traditional advertising and improving the breaking-through communication ability of advertising.

[0141] Cross-domain data collection and analysis: Collect data from art reviews, popular film plot analysis, popular song lyrics, best-selling book content, viral social media content (short videos, memes), and excellent game design.

[0142] Feature extraction: Apply similar techniques to the user mental insight module to analyze elements in these contents (such as narrative structure, visual style, emotional expression, humor, conflict setting, suspense creation, color use, music rhythm, etc.).

[0143] Psychological response model construction: Train the model to explain how these non-advertising creative elements interact with human psychology (such as curiosity, empathy, surprise, fear, joy, etc.) and user mental states (such as specific emotions, cognitive processing patterns), and quantify their psychological activation effectiveness.

[0144] Psychological activation patterns and creative combination paradigm learning: Identify from cross-domain data those creative patterns or psychological anchors that repeatedly appear and produce strong user psychological responses. The narrative structure of sudden plot twists often enhances user cognitive engagement and expectations; warm color tones and soft music often evoke positive emotional resonance.

[0145] Mental state association: Learn the activation effect of specific creative patterns on different mental states (corresponding Some humor styles work best when users are in a relaxed state, but some humor styles work best when users are in a highly alert The state may be counterproductive.

[0146] Creative gene migration and reconfiguration:

[0147] Gene library expansion: extract and transform effective creative elements learned from cross-domain into new creative gene fragments, and give them corresponding semantic attributes and psychological drive vectors, and store or update the creative gene library. Transform the learned psychological activation patterns and combination paradigms into new rules or optimization targets for guiding the generation of advertising creativity. If it is found that creating a slight cognitive dissonance can effectively improve the user's memory of a certain type of information, then when generating an advertisement, prefer to combine gene fragments that can create this effect.

[0148] Application in advertising generation: when executing the intentional creative generation module, the system will preferentially consider gene fragments and knowledge migrated from cross-domain.

[0149] The gain term from cross-domain knowledge can be added to the function, or the migrated gene is given a higher weight in the gene fragment library.

[0150] If a cross-domain gene fragment has been proven to effectively improve the effectiveness of a target intent node , then in optimization: ;

[0151] Through such cross-domain migration, advertising content can break out of the traditional advertising routine and present in a way that is more in line with human nature, more narrative, emotional or intellectual, thereby standing out in the ocean of information and more easily triggering the user's active dissemination and deep memory.

[0152] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make usual changes and replacements within the scope of the technical solutions of the present application, which should be included in the protection scope of the present application.​

Claims

1. A user mind-driven based advertisement intention generation system, characterized in that, The system comprises: A user mental insight module for analyzing user information and constructing a mapping relationship between user behavior patterns, psychological states and potential intentions through multi-modal data; the user mental insight module is particularly provided with: A mental state encoder is used to receive explicit behavior sequence data, physiological feature signals and / or voice tone data of a user, and output a high-dimensional user mental state vector through deep neural network processing The vector encodes the user's instant psychological state An intent node decoupler, based on latent variable model, analyzes and decouples out a set of intent nodes representing the core psychological drivers of the user , respectively corresponding to the latent intents of the three dimensions of motivation, cognition and emotion of the user. An intention strategy mapping module for receiving an advertising target set by an advertiser and accurately associating the advertising target with an intention node set of a target user group to determine a mental intervention strategy for the user intention node to achieve the advertising target; An intention creative generation module for receiving the intention node set of the user and the target intervention strategy, decoupling preset key advertising creative elements into creative gene segments, and then optimizing the combination mode and action energy of the creative gene segments according to a predetermined mental intervention strategy by a combination optimization algorithm to generate advertising creative content capable of reaching or reconstructing the target user intention node; A creative dynamic evolution module for receiving effect feedback data after actual advertising is launched and implementing closed-loop iterative optimization based on the feedback data to adaptively adjust model parameters and creative gene segments in the user mental state model, the intention strategy mapping mechanism and the creative generation module.

2. The user mind-driven advertisement intention generation system according to claim 1, wherein, The mental state encoder of the user mental insight module captures the context correlation and potential semantic information in the user behavior sequence; The intent node decoupler utilizes a variational autoencoder to decouple high-dimensional mental state vectors , mapping to a low-dimensional latent space from which interpretable core intent nodes are extracted . 3.The user mind-driven advertisement intention generation system of claim 2, wherein, The mental state encoder, when processing the user behavior sequence data, uses a self-attention mechanism to give dynamic attention weights to the input physiological feature signals and / or voice tone information, so as to highlight the psychological hint information highly related to the current user behavior mode and weight and fuse the psychological hint information into the user mental state vector .

4. The user mind-driven advertisement intention generation system of claim 1, wherein, The intention strategy mapping module is provided with a mapping rule of target-intention driven intervention; the rule decomposes the overall target of the advertising task into the mental influence strength requirement on the specific intention node . ​​ That is, wherein is the subset of intent nodes most relevant to the advertising target.

5. The user mind-driven advertisement intention generation system of claim 1, wherein, The intention creative generation module includes a creative gene library storing minimized creative feature units with independent psychological influence, each segment being represented in the form of a semantic attribute vector and a psychological driving vector, A combination optimization algorithm selects the optimal combination of creative gene segments from a creative gene library by generating a generative adversarial network to maximize the mental touch effectiveness of the generated advertisement content on the target intent node to maximize the mental touch effectiveness of the generated advertisement content on the target intent node , That is .

6. The user mind-driven advertisement intention generation system according to claim 5, wherein, The combination optimization algorithm includes a subconscious anchoring optimization mechanism that identifies key intention nodes in a user mental state vector that are most sensitive to a target intervention strategy ; and at the sensitive node, a creative gene segment combination capable of producing the maximum cognitive activation or emotional resonance is preferentially generated.

7. The user mind-driven advertisement intention generation system of claim 1, wherein, The creative dynamic evolution module realizes a user mental feedback closed loop; the closed loop comprises: mental state prediction unit: receives the current intention node state of the user and the generated advertisement creative content, and predicts the mental state evolution trajectory that the user is likely to generate thereafter ; An effect evaluation and calibration unit: actual advertising effect data, including user conversion rate, stay time, emotional feedback, etc., are used as real evolution signals and compared with the predicted trajectory; the system calibrates and updates the core parameters of the user mental state model, the intention strategy mapping model and the creative generation model using a gradient descent model, if the advertising content fails to effectively activate the target intention node, the system will adjust the combination weight of the creative gene segments used in the content generation process.

8. The user mind-driven advertisement intention generation system according to claim 7, wherein, The creative gene adaptive learning mechanism in the creative dynamic evolution module quantizes the user mental feedback signal into a reinforcement learning reward signal R; Using the reward signal, the generation model is optimized through the SARS A algorithm to learn a better creative gene segment selection and combination strategy.

9. The user mind-driven advertisement intention generation system of claim 1, wherein, It also includes a cross-domain mental creative transfer function; the function refers to that the system first learns creative elements driving user psychological responses and their performance rules in different mental states from non-advertising fields; then, the learned psychological activation patterns and creative combination paradigms are applied to advertising generation to create more impactful and memorable advertising content, thereby improving the breakthrough communication ability of the advertising.

10. A user mental driving-based advertising intention generation method, comprising the following steps: S1, mental decoupling and intention recognition: analyzing multi-modal data of users and accurately identifying core intention nodes of users; S2, target-intervention strategy derivation: according to the advertisement target, derive the specific mental intervention strategy required for the intention nodes; S3, intention creative generation: according to the identified intention nodes and the derived mental intervention strategy, combine and optimize the creative gene fragments to generate advertisement creative content capable of producing specific mental influence; S4, effect feedback and model self-optimization: real-time tracking of advertisement delivery effect, and using feedback data for dynamic calibration and optimization of user mental model, intention strategy and creative generation model, forming a continuously evolving intelligent generation capability.

Images