Financial marketing large model generation and delivery optimization method based on multi-modal user behavior modeling

Abstract

The application relates to the technical field of artificial intelligence large models, and discloses a financial marketing large model generation and optimization method based on multi-modal user behavior modeling, which comprises the following steps: converting original behavior data of multiple behavior modes into behavior events, obtaining continuously evolved user hidden states and comprehensive excitation intensities through an event coding network; after the comprehensive excitation intensity exceeds the first intensity threshold, taking the peak moment in a preset observation time length as a trigger moment, determining a quantitative depth according to the trigger intensity, performing residual quantization on a query state to obtain a semantic word element sequence; mapping the semantic word element sequence into placeholder word elements and constructing a prompt sequence, inputting the prompt sequence into a financial marketing large model for word element by word element decoding, splicing a risk prompt template to obtain a to-be-launched script; and performing direct launching, frequency reduction launching or manual review according to the uncertainty and two gating thresholds. The application improves the coincidence degree of launching time, the compliance of the script and the controllability of the review.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence large model technology, specifically involving a method for generating and optimizing financial marketing large models based on multimodal user behavior modeling. Background Technology

[0002] Online marketing of financial products relies on timely identification of user interests and personalized outreach. Existing marketing systems typically slice user behavior records within the application at fixed time granularities, use sequence models such as recurrent neural networks or attention networks to learn user representations, and then use click-through rate prediction models to determine whether to target ads. In terms of copywriting, existing practices involve incorporating user tags into prompt words to drive large language models to generate candidate copy in batches, and then relying on sensitive word filtering and manual review to ensure quality.

[0003] The above-mentioned technologies have the following problems in practical applications: First, page browsing, control clicks, transaction logs, and customer service conversations arrive asynchronously and unevenly. Fixed-granularity slicing causes the loss of arrival time sequence information, and the continuous decline of interest cannot be depicted, resulting in a misalignment between the timing of ad placement and the peak of user interest. Second, there is a connection barrier between the continuously valued user representation and the discrete vocabulary of the large language model. Directly concatenating numerical features results in significant semantic loss, and fixed-granularity representation introduces noise when interest evidence is insufficient and loses detail when interest is strong. Third, compliance control is placed after generation and is a probabilistic interception. Violations may still be leaked even after multiple segmentation methods using sub-words, and regenerating after filtering consumes a significant amount of inference computing power. Fourth, the ad placement decision lacks quantification of the reliability of the prediction, and there is no predictable basis for the investment of human review resources. Summary of the Invention

[0004] The main objective of this invention is to provide a method for generating and optimizing large-scale financial marketing models based on multimodal user behavior modeling, in order to solve the problems of asynchronous arrival of multimodal behavioral events, misalignment between the timing of ad placement and peak user interests, difficulty in connecting continuous user representations with the large-scale model vocabulary, lagging copy compliance control, and lack of quantitative basis for manual review workload.

[0005] To solve the above problems, the technical solution of the present invention is implemented as follows: A method for generating and optimizing large-scale financial marketing models based on multimodal user behavior modeling includes the following steps: Step 1: Convert the raw behavioral data generated by the target user under multiple behavioral modalities into behavioral events one by one. Each behavioral event consists of an event type identifier, an event time, and an event attribute vector. Arrange all behavioral events in ascending order of event time to form a behavioral event sequence. Use an event coding network to encode the behavioral event sequence online to obtain the user's latent state that evolves continuously over time between adjacent behavioral events. The intensity readout layer maps the user's latent state into a comprehensive incentive intensity. Step 2: When the comprehensive incentive intensity changes from less than the preset first intensity threshold to greater than the first intensity threshold, the moment when the comprehensive incentive intensity reaches its maximum value within the subsequent preset observation period is determined as the trigger moment. The comprehensive incentive intensity at the trigger moment is recorded as the trigger intensity, and the user's hidden state at the trigger moment is recorded as the query state. The quantization depth is determined based on the trigger intensity. Residual quantization is performed on the query state level by level according to the quantization depth. The index of the codeword selected by the progressive residual quantization is used as the semantic unit of the corresponding level. The semantic units of each level are arranged in sequence to form a semantic unit sequence. Step 3: Map each semantic word in the semantic word sequence to a corresponding placeholder word in the pre-expanded vocabulary of the financial marketing big data model. Concatenate this placeholder word with the product description text of the target financial product and the marketing task instruction text in sequence to form a prompt sequence. Input the prompt sequence into the financial marketing big data model for word-by-word decoding to obtain candidate marketing copy. In each decoding step, set the generation probability of the next candidate word that will cause the compliance state machine to transition from its current state to the hit state to 0, and re-normalize the generation probabilities of the remaining candidate next words. The compliance state machine is pre-constructed using the word sequence corresponding to the violation statement as a pattern string. The hit state is the state in the compliance state machine that completely matches any pattern string. Finally, append a risk warning template corresponding to the target financial product to the end of the candidate marketing copy to obtain the copy to be delivered. Step 4: Encode the text to be delivered into a text vector. Concatenate the query status with the text vector and input it into the matching evidence network to obtain the evidence vector. Calculate the uncertainty based on the evidence vector. When the uncertainty is less than or equal to the predetermined first gate threshold, the text to be delivered is directly pushed to the target user through the delivery channel. When the uncertainty is greater than the first gate threshold but less than or equal to the predetermined second gate threshold, multiply the default exposure frequency limit by a preset reduction factor to obtain the reduced exposure frequency limit, and push the text to be delivered to the target user at the reduced exposure frequency limit. When the uncertainty is greater than the second gate threshold, the text to be delivered is sent to the manual review queue. After manual review and confirmation, it is pushed to the target user. The first gate threshold is less than or equal to the second gate threshold.

[0006] Furthermore, in step 1, the behavioral modalities include page browsing modalities, control clicking modalities, transaction flow modalities, and customer service conversation modalities; the event attribute vectors are obtained by encoding the original behavioral data by the feature encoder corresponding to the behavioral modalities and then mapping it to a unified dimension through a linear layer.

[0007] Further, in step 1, online encoding includes: initializing the user's hidden state before the arrival of the first behavioral event as a zero vector; when each behavioral event arrives, the event encoding network takes the type embedding vector corresponding to the event type identifier of the current behavioral event, the event attribute vector of the current behavioral event, and the user's hidden state at the instant before the arrival of the current behavioral event as input, updates the hidden state after the transition through a gated recurrent unit, obtains a steady-state vector through a linear output head, and obtains a decay rate vector through an output head with softplus activation, where each component of the decay rate vector takes a positive value; in any interval between two adjacent behavioral events... At time t, the difference between the user's hidden state and the steady-state vector corresponding to the previous action event decays exponentially with time, element by element, starting from the event time of the previous action event. The initial value of the difference is the difference between the hidden state after the jump and the steady-state vector corresponding to the previous action event. The decay rate is determined by the corresponding component of the decay rate vector corresponding to the previous action event. The comprehensive stimulus intensity is obtained by performing a linear transformation on the user's hidden state by the intensity readout layer and then activating it with softplus. The event encoding network and the intensity readout layer are pre-trained with the maximum likelihood of the action event sequence under the time-series process framework as the training objective.

[0008] Further, in step 2, determining the quantization depth based on the trigger intensity includes: pre-setting a second intensity threshold and a third intensity threshold, wherein the first intensity threshold is less than the second intensity threshold, and the second intensity threshold is less than the third intensity threshold; when the trigger intensity is less than the second intensity threshold, the quantization depth is 1; when the trigger intensity is greater than or equal to the second intensity threshold and less than the third intensity threshold, the quantization depth is 2; when the trigger intensity is greater than or equal to the third intensity threshold, the quantization depth is 3.

[0009] Further, in step 2, the progressive residual quantization includes: pre-training a level 1 codebook, a level 2 codebook, and a level 3 codebook, each level containing a preset number of codewords. The codewords have the same dimension as the user's latent state. During the offline training phase, the level 3 codebook updates the codewords using an exponential moving average method with the goal of minimizing the quantization reconstruction error of historical query state samples. In the level 1 codebook, the codeword with the smallest Euclidean distance to the query state is selected as the level 1 codeword, and the index of the level 1 codeword in the level 1 codebook is used as the level 1 semantic term. When the quantization depth is greater than or equal to 2, the query state is subtracted from the level 1 codebook. The Level 1 codeword yields the Level 1 residual. In the Level 2 codebook, the codeword with the smallest Euclidean distance to the Level 1 residual is selected as the Level 2 codeword. The index of the Level 2 codeword in the Level 2 codebook is used as the Level 2 semantic token. When the quantization depth is 3, the Level 1 residual is subtracted from the Level 2 codeword to obtain the Level 2 residual. In the Level 3 codebook, the codeword with the smallest Euclidean distance to the Level 2 residual is selected as the Level 3 codeword. The index of the Level 3 codeword in the Level 3 codebook is used as the Level 3 semantic token. The resulting semantic tokens at each level are arranged in the order of Level 1, Level 2, and Level 3 to form a semantic token sequence.

[0010] Furthermore, in the vocabulary of the financial marketing big data model, each codeword in the 3-level codebook is pre-extended with one corresponding placeholder word. The embedding vector of each placeholder word is initialized with the vector of the corresponding codeword after linear projection. The financial marketing big data model is fine-tuned using a low-rank adaptation method. The parameters updated in the fine-tuning stage consist of the low-rank adaptation parameters and the embedding vectors of all placeholder words, while the base parameters remain frozen. The fine-tuning samples are composed of semantic word sequences obtained by offline replay of historical behavioral event sequences in steps 1 and 2, paired with manually written compliance copy samples. The fine-tuning objective is to minimize the word-level cross-entropy loss of the compliance copy samples. The financial marketing big data model with the input prompt sequence in step 3 is the fine-tuned financial marketing big data model.

[0011] Furthermore, the compliance state machine is pre-constructed in the following way: A set of context prefixes and a set of context suffixes are constructed, consisting of an empty string, high-frequency Chinese characters, and common punctuation marks; each violation expression in the violation expression vocabulary is combined with each context prefix and each context suffix to obtain a concatenated string formed by sequentially connecting context prefixes, violation expressions, and context suffixes; the concatenated string is input into the word segmenter of the financial marketing big data model, and based on the character offset information output by the word segmenter, consecutive words overlapping with the character range of the violation expression are extracted to form the word sequence corresponding to the violation expression; the word sequences obtained from all violations expressions under the combination of all context prefixes and context suffixes are summarized and deduplicated, and the deduplicated word sequences are used as pattern strings to construct an Aho-Corasick automaton as the compliance state machine; the state corresponding to a complete match with any pattern string in the compliance state machine is the hit state, and the other states are safe states.

[0012] Furthermore, in step 3, word-by-word decoding employs a bundle search, where each candidate sequence within the bundle maintains its own compliance state machine state. The compliance state machine state is synchronously copied during candidate sequence expansion. In each decoding step, for each candidate next word in each candidate sequence, the transition state is calculated according to the compliance state machine's transition table. The candidate next word whose transition state is the hit state is the candidate next word whose generation probability is set to 0. The word selection and compliance state machine state update for the current decoding step are completed based on the re-normalized generation probability. The risk warning template is selected from the risk warning template library based on the product risk level of the target financial product.

[0013] Further, in step 4, the copywriting vector is obtained by inputting the copywriting to be delivered into a pre-trained text encoder; the matching evidence network is a multilayer perceptron with softplus activation in the output layer, which outputs a 2-dimensional evidence vector, and the two components of the evidence vector correspond to the amount of evidence for the conversion category and the other categories, respectively; historical delivery records are collected in advance, and each historical delivery record consists of the query status, copywriting vector and real conversion label corresponding to one historical delivery, and all historical delivery records are randomly divided into a training set and a calibration set according to a preset ratio; the matching evidence network is pre-trained using the training set as training samples and the expected mean squared error loss under the evidence deep learning framework as the training objective; the uncertainty is calculated based on the evidence vector, including: adding 1 to each of the two components of the evidence vector and summing them to obtain the total amount of evidence, and using the ratio of the number of categories to the total amount of evidence as the uncertainty, with the number of categories being 2.

[0014] Furthermore, the first and second gate thresholds are predetermined as follows: Uncertainties are calculated for all historical delivery records in the calibration set and arranged in ascending order; the sum of the calibration set sample size plus 1 is multiplied by the difference between 1 and the preset first significance level, and the difference between 1 and the preset second significance level, respectively. The two products are rounded up to obtain the first and second positions; the uncertainty at the first position after ascending order is used as the first gate threshold, and the uncertainty at the second position after ascending order is used as the second gate threshold; if the first significance level is greater than the second significance level, the calibration set sample size is chosen such that the second position is less than or equal to the calibration set sample size. The method also includes: after the campaign is completed, mapping the click events and subscription events in the conversion feedback of the target users to the corresponding event type identifiers, converting them into new behavioral events and writing them into the behavioral event sequence; constructing new historical campaign records by combining the query status, copy vector and real conversion tags corresponding to this campaign, and randomly allocating them to the training set or calibration set according to a preset ratio, with the allocation results remaining fixed; performing incremental training on the matching evidence network with the expanded training set according to a preset period, and then recalculating the uncertainty of all historical campaign records in the expanded calibration set with the incrementally trained matching evidence network, and redetermining the first and second gate thresholds based on the recalculated uncertainty.

[0015] The financial marketing large-scale model generation and delivery optimization method based on multimodal user behavior modeling of the present invention has the following beneficial effects: Compared with existing technologies, this invention models asynchronously arriving multimodal behavioral events in a continuous-time manner, obtains the semantics of event occurrence rate per unit time by comprehensively considering the intensity of the stimulus, and uses a triggering method that combines overshoot detection with peak value selection within a preset observation period to ensure that the delivery action is close to the peak moment of user interest, thus improving the fit between the timing of the delivery and the peak of interest. The quantization depth is adaptively determined according to the trigger intensity. When there is sufficient evidence of interest, deeper residual quantization is used to retain fine-grained preferences, while shallow quantization is used to suppress noise and shorten the prompt sequence when there is limited evidence of interest. Compared with the processing method with fixed quantization depth, this invention achieves a scene-adaptive balance between generating relevance and inference overhead. Low-intensity triggering also eliminates the need for high-order detection. The algorithm incorporates search and residual calculations; compliance constraints are pre-defined in the word-by-word decoding process, enabling deterministic interception of pattern strings already in the database at the word level. Dual enumeration of context prefixes and suffixes expands the coverage of multiple sub-word segmentation paths, ensuring compliance upon generation and reducing retries and rework. The uncertainty gating uses two thresholds determined based on calibration set quantiles, with the traffic proportion of each tier controlled by the significance level. The manual review queue obtains a predictable upper bound on the load, concentrating review resources on deliveries with insufficient evidence. Conversion feedback is fed back as new behavioral events and historical delivery records. The matching evidence network and gating thresholds are updated synchronously according to a preset cycle, forming a closed loop and maintaining the long-term effectiveness of tier division. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the continuous evolution of a user's hidden state between adjacent behavioral events, provided by an embodiment of the present invention. Figure 2 A schematic diagram illustrating the principle of the overhead detection of comprehensive excitation intensity and peak triggering within a preset observation period provided in an embodiment of the present invention; Figure 3 The geometric principle diagram of the query status progressive residual quantization provided in the embodiment of the present invention is shown, wherein (a) is the first level residual quantization and (b) is the second level residual quantization; Figure 4 A schematic diagram of the transition structure of the compliance state machine provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the generation probability of the candidate next word in one decoding step before and after masking, provided in an embodiment of the present invention. (a) is before masking, and (b) is after masking. Figure 6 This is a schematic diagram of the distribution of uncertainty on the calibration set and two gate thresholds provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of quantization depth classification and quantization effect provided in the embodiment of the present invention, wherein (a) is the classification rule of quantization depth with trigger intensity, and (b) is the relative quantization reconstruction error under different quantization depths. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are used to explain the implementation of this invention; processes not detailed in the embodiments can be performed in accordance with conventional methods in the art.

[0018] The financial marketing large-scale model generation and delivery optimization method based on multimodal user behavior modeling provided in this invention is deployed and executed on the marketing service platform of a financial institution. The marketing service platform includes behavior access nodes, model inference nodes, and delivery channel services. The behavior access nodes handle online computation of event encoding networks and residual quantization. The model inference nodes are configured with two graphics processors with 80GB of dedicated memory each, handling the inference service of the financial marketing large-scale model. During the offline training phase, a training server equipped with eight identical graphics processors is used. The following example, a marketing push targeting fund products in a mobile banking application, illustrates the implementation process with reference to the accompanying drawings.

[0019] Raw behavioral data is collected with user authorization, enters a message queue, and is consumed by the behavior access node. User identity information is anonymized before entering the behavior access node. The behavior access node writes the latest evolution parameters of each target user into an in-memory database for trigger judgment and query status reading. The model inference node and the behavior access node are connected through an internal service interface. The query status generated by the trigger, the quantified semantic word sequence, and the target financial product information are sent to the model inference node through the internal service interface.

[0020] The online links of the entire method include, in sequence: continuous encoding of behavioral events, upward detection and peak triggering of comprehensive incentive intensity, residual quantification of query state, construction of prompt sequence and word-by-word meta-decoding constrained by compliance state machine, concatenation of risk prompt templates, and three-tier delivery based on uncertainty; the offline links are responsible for training the event encoding network and intensity readout layer, the 3-level codebook, the financial marketing big model and matching evidence network, as well as the construction of the compliance state machine and two gating thresholds. The following sections describe each link in sequence.

[0021] Target users continuously generate raw behavioral data across four behavioral modalities: page browsing, control clicking, transaction history, and customer service conversation. Raw behavioral data for the page browsing modality includes the page category number and dwell time; for the control clicking modality, it includes the control number and click time; for the transaction history modality, it includes the transaction amount and product type number; and for the customer service conversation modality, it consists of the text content of one round of conversation. Each piece of raw behavioral data is converted into a behavioral event, which consists of an event type identifier, an event time, and an event attribute vector. The page browsing modality, control click modality, transaction flow modality, and customer service conversation modality correspond to event type identifiers one, two, three, and four, respectively. The event time is the timestamp of the raw behavioral data when it was stored in the database, accurate to the second. The event attribute vector is output by the feature encoder corresponding to the behavioral modality. The feature encoder of the page browsing modality consists of a page category embedding layer and a dwell time binning embedding layer concatenated with two layers of feedforward network. The feature encoder of the control click modality is a control number embedding layer. The feature encoder of the transaction flow modality consists of a transaction amount binning embedding layer and a product type embedding layer concatenated with the transaction amount binned logarithmically. The feature encoder of the customer service conversation modality is a 12-layer pre-trained text encoder, and its output sentence vector is taken. Dwell time is divided into four bins: 0-5 seconds, 5-30 seconds, 30-120 seconds, and over 120 seconds. Transaction amount is divided into eight bins based on a base-10 logarithmic value. This binning embedding transforms continuous quantities into discrete representations robust to scale changes, preventing extreme values ​​from dominating the event attribute vector. The outputs of the four feature encoders are each mapped to a unified 64-dimensional dimension via a linear layer, enabling the subsequent event encoding network to process behavioral events from different behavioral modalities with the same set of parameters. Heterogeneous information undergoes spatial alignment before entering sequence modeling.

[0022] All behavioral events are arranged in ascending order of event time to form a behavioral event sequence. Two behavioral events with the same event time are arranged in the order they were added to the database. If the order of addition to the database is still indistinguishable, they are arranged in ascending order of event type identifier. The arrangement rules are consistent on both the offline training and online service sides to avoid obtaining sequences with different discrepancies on both sides for the same batch of data.

[0023] The core of the event encoding network is the gated recurrent unit (GRU). Upon arrival of each behavioral event, the 32-dimensional type embedding vector corresponding to the event type identifier is concatenated with the 64-dimensional event attribute vector to form a 96-dimensional input. This input, along with the 128-dimensional user latent state instant before arrival, is fed into the GRU to update the post-transition latent state. The post-transition latent state is then processed by a linear output head to obtain a steady-state vector, and by an output head with softplus activation to obtain a decay rate vector. Both the steady-state vector and the decay rate vector are 128-dimensional. The reason for using softplus activation for the decay rate vector is that the softplus output is always positive and the gradient is smooth. If any component of the decay rate vector takes a negative value, the exponential term in subsequent evolution will diverge over time, causing the user latent state to lose stability and the overall stimulus intensity to become meaningless. The user latent state before the arrival of the first behavioral event is initialized as a zero vector. The type embedding vector comes from a 4x32 learnable embedding table. The behavioral events generated by the transformation feedback mapping reuse the row and feature encoders of the corresponding behavioral modality without further expanding the embedding table. Between two adjacent behavioral events, the user's hidden state evolves continuously according to formula (1): (1); In formula (1), h(t) represents the user's hidden state at time t; t i h represents the event time of the i-th behavioral event, where i is the sequence number of the behavioral event in the sequence of behavioral events; i The hidden state after the transition corresponding to the i-th action event; g i δ is the steady-state vector corresponding to the i-th behavioral event; i Let be the decay rate vector corresponding to the i-th behavioral event; ⊙ represents element-wise multiplication, exp represents the element-wise exponential operation, and formula (1) is in t i The event time is consistent with the (i+1)th behavioral event. The latent state after the transition describes the immediate impact of the behavioral event, the steady-state vector describes the level after the impact fades, and the decay rate vector allows different dimensions of user interests to fade at different rates. For example, dimensions related to time-limited activities fade quickly, while dimensions related to long-term financial preferences fade slowly. Figure 1The evolution of two components in the user's latent state under the influence of four behavioral events is presented. Each behavioral event causes a jump in the component's value, which then converges towards the index of the corresponding component in the steady-state vector. The dashed line indicates the position of the component corresponding to the steady-state vector within one interval. Taking component 1 as an example, the behavioral event at 0.6 hours causes component 1 to jump from 0 to 1.6, and then decays towards the steady-state value of 0.9. The behavioral events at 1.8 hours, 3.2 hours, and 4.1 hours respectively raise component 1 to 2.3, 3.0, and 3.4, with corresponding steady-state values ​​of 1.2, 1.8, and 2.0. Overall, the pattern shows that interest is gradually increased by continuous behaviors and slowly decreases during intervals. The intervals between adjacent behavioral events vary, and the evolution process is handled by the same set of formulas, thus replacing fixed-granularity slices with continuous-time modeling.

[0024] In engineering implementation, the behavior access node caches only the hidden state after the transition, the steady-state vector, the decay rate vector, and the event time corresponding to the most recent behavior event for each target user. The user's hidden state at any time is calculated instantly according to formula (1). The storage overhead is independent of the historical number of behavior events. The continuous evolution form has engineering advantages over time-step recursion. The comprehensive incentive intensity is read from the user's hidden state by the intensity readout layer according to formula (2): (2); In formula (2), λ(t) is the comprehensive excitation intensity at time t; w is the 128-dimensional weight vector of the intensity readout layer; T represents the transpose; and ln is the natural logarithm. Formula (2) is equivalent to applying softplus activation to the linear transformation result, and the output is always positive, which meets the requirements of the time-series process framework for the value of the intensity function. The comprehensive excitation intensity thus obtains the dimension of the event occurrence rate per unit time, which is expressed as times / hour in the example.

[0025] The event encoding network and the intensity readout layer are jointly trained before deployment. The training samples are the de-identified historical user behavior event sequences, for example, 1.2 million sequences, with the 512 most recent behavior events extracted from each sequence. The training objective is to maximize the likelihood of the behavior event sequence within the time-series process framework, and the log-likelihood is calculated according to formula (3): (3); In formula (3), L is the log-likelihood of a sequence of behavioral events; N is the total number of behavioral events in this sequence; t NLet N be the event time of the Nth behavioral event. The summation term encourages the overall incentive intensity to be high at the time the actual behavioral event occurs, while the integral term inhibits the overall incentive intensity from being too high during periods without events. Under the combined effect of these two terms, the overall incentive intensity obtains the semantics of the event occurrence rate. Subsequent measurements of the sufficiency of user behavior evidence by trigger intensity are based on this semantics. The integral term has no analytical solution and is approximated using a Monte Carlo method: 20 time points are uniformly sampled between each pair of adjacent behavioral events, and the average of the overall incentive intensity at the 20 time points is multiplied by the interval length as an approximation of the integral for the corresponding interval. The integral term is obtained by summing all intervals. The optimizer used is Adam, with a learning rate of 0.001, a batch size of 128, and 10 training rounds, taking approximately 3 hours on 8 GPUs. After training, it was observed that the components of the decay rate vector are concentrated in the range of 0.2 to 4 per hour, corresponding to half-lives ranging from tens of minutes to tens of hours. This is consistent with the experience of short-term impulses and long-term preferences coexisting in marketing scenarios, and also confirms the necessity of setting the decay rate on an element-by-element basis.

[0026] During the online service phase, the event coding network updates instantly upon the arrival of behavioral events. The marketing service platform calculates the current value of the comprehensive incentive intensity at a fixed interval of once every 10 seconds, based on the arrival time of each behavioral event. The first intensity threshold is set to 2.0 times / hour, which can be determined based on approximately the 70th percentile of the historical intensity value distribution. When the comprehensive incentive intensity changes from less than the first intensity threshold to greater than the first intensity threshold, an upward crossing is determined, and a preset observation period of 30 minutes is initiated. The upward crossing can occur at the instant the behavioral event causes a jump, or it can be detected by comparing the values ​​of two adjacent calculation times. The moment when the comprehensive incentive intensity reaches its maximum value within the preset observation period is determined as the trigger moment. The comprehensive incentive intensity at the trigger moment is recorded as the trigger intensity, and the user's hidden state at the trigger moment is recorded as the query state. According to formula (1), the evolution between adjacent behavioral events monotonically tends towards a steady-state vector. The maximum value within the preset observation period will only appear at the moment when the behavioral event causes a jump or at the beginning of the preset observation period. In implementation, it is only necessary to compare and record each time a jump occurs, without the need for intensive scanning of continuous time. When two maximum values ​​of the same size appear within the preset observation period, the moment with the earlier time is taken as the trigger moment. The trigger moment is based on two points: the comprehensive incentive intensity at the moment of upward movement must be near the first intensity threshold, and it is impossible to distinguish the magnitude of the current interest. The peak moment provides both a comparable intensity magnitude and the time position of the interest peak, so that the subsequent delivery action is close to the moment when the user's attention is most concentrated. When the comprehensive incentive intensity falls back within the preset observation period and then moves upward again, it is included in the current observation and the preset observation period is not restarted. After the preset observation period ends and one trigger is completed, a 2-hour cooldown period is set. During the cooldown period, the upward movement judgment is suspended to avoid repeated triggering of the same wave of interest. During the cooldown period, the behavioral event is still written into the behavioral event sequence and the user's hidden state is updated normally. Only the upward movement judgment is suspended. Figure 2 A complete triggering process is given: the behavioral event at 1.30 hours changes the overall incentive intensity from about 1.4 times / hour to 2.9 times / hour, completing the crossing of the first intensity threshold. The gray area represents the preset observation period of 0.5 hours thereafter. During the observation period, the behavioral events at 1.55 hours and 1.72 hours raise the overall incentive intensity to 4.6 times / hour and 5.9 times / hour respectively. 5.9 times / hour is the peak value within the preset observation period. 1.72 hours is determined as the trigger time, and 5.9 times / hour is recorded as the trigger intensity.

[0027] The second intensity threshold is set to 3.5 triggers / hour, and the third intensity threshold is set to 5.0 triggers / hour. These thresholds can be set based on approximately the 90th and 97th percentiles of historical intensity value distributions, respectively. All three intensity thresholds maintain a strictly increasing trend. The quantization depth is 1 when the trigger intensity is less than the second intensity threshold, 2 when the trigger intensity is greater than or equal to the second intensity threshold but less than the third intensity threshold, and 3 when the trigger intensity is greater than or equal to the third intensity threshold. See the grading rules below. Figure 7 (a) Figure 2 In the example, the trigger intensity of 5.9 times / hour exceeds the third intensity threshold, so the quantization depth is set to 3. The quantization depth is determined by trigger intensity because: trigger intensity reflects the sufficiency of evidence regarding current user behavior; when evidence is insufficient, the components attributable to stable interests in the query state are limited, and deep residuals are mainly noise. Quantizing the semantic units obtained from deep residuals can inject noise into the large-scale financial marketing model, and deep semantic units corresponding to low trigger intensity are sparse and insufficiently trained in fine-tuning samples; when evidence is sufficient, the fine-grained preferences carried by deep residuals are reliable, and longer prompt sequences can be used to obtain generated relevance. The quantization depth is determined first, and then residual quantization is performed level by level according to the quantization depth. When the quantization depth is 1, the calculation of the first-level residual and the nearest neighbor search on the second and third-level codebooks are directly omitted. The quantization depth also determines the number of placeholder units in the prompt sequence. When the quantization depth is 1, the prompt sequence is 2 units shorter than when it is 3, resulting in shorter decoding prefixes and faster first-unit generation in weak interest scenarios.

[0028] The Level 1, Level 2, and Level 3 codebooks each contain 256 codewords, with each codeword and the user's latent state having 128 dimensions. The training samples for the codebooks are historical query state samples, obtained by first training the event encoding network and the strength readout layer, then performing offline replay of historical user behavior event sequences, using the same uplink judgment, preset observation duration, and peak rules as online services to obtain historical trigger times, and taking the user's latent state at each historical trigger time. For example, approximately 3 million historical query state samples are accumulated. The Level 3 codebook aims to minimize the quantization reconstruction error of historical query state samples. Each level of the codebook is initialized using k-means++. During training, a complete Level 3 residual quantization is performed on each historical query state sample. The Level 3 codebook is updated synchronously using an exponential moving average: maintaining two statistics for each codeword—a cumulative count and a cumulative sum. After each training batch, these two statistics are multiplied by a decay coefficient of 0.99, and then added to the number of samples allocated to the corresponding codeword in the current batch and the sample vector sum. The codeword is then calculated as the ratio of the cumulative sum to the cumulative count. Specifically, the Level 1 codebook is initialized using k-means increment with 100,000 randomly selected historical query state samples. The Level 2 and Level 3 codebooks are initialized in the same way with the Level 1 and Level 2 residuals generated in the first round of traversal, respectively. The training batch size is 1024. After traversing the full sample for 5 rounds, the codeword drift of each level of the codebook is negligible. Exponential moving average updates allow the codewords to slowly shift with the data distribution, ensuring a stable training process and decoupling from gradient optimization. Offline training is fixed at 3 levels, while online truncation is based on quantization depth. The two methods are compatible: the complete reconstruction goal enables the 3-level codebook to form a coarse-to-fine division of labor. The Level 1 codewords carry the main position of the query state, while the Level 2 and Level 3 codewords successively carry progressively smaller residuals. Online truncation of the first 1, the first 2, or all 3 levels yields a nested coarse-to-fine representation. The truncation action does not violate the training premise of any level of codebook.

[0029] The execution process of one online quantization is as follows: In the first-level codebook, the codeword with the smallest Euclidean distance to the query state is selected as the first-level codeword, and the index of the first-level codeword in the first-level codebook is the first-level semantic term; when the quantization depth is greater than or equal to 2, the query state is subtracted from the first-level codeword to obtain the first-level residual, and the codeword with the smallest Euclidean distance to the first-level residual in the second-level codebook is selected as the second-level codeword, and the index of the second-level codeword in the second-level codebook is the second-level semantic term; when the quantization depth is 3, the first-level residual is subtracted from the second-level codeword to obtain the second-level residual, and the codeword with the smallest Euclidean distance to the second-level residual in the third-level codebook is selected as the third-level codeword, and the index of the third-level codeword in the third-level codebook is the third-level semantic term. Figure 3 In (a), the geometric relationship of the first-level residual quantization is given by two-dimensional projection. The solid arrow points to the selected first-level codeword, and the dashed arrow is the first-level residual. Figure 3Figure (b) shows the correspondence between the second-level residual quantization and the residual space. The second-level codeword approximates the first-level residual, and the remaining second-level residual has been significantly reduced. Figure 3 (a) also shows the remaining Level 1 codewords that were not selected. The selected Level 1 codewords are marked with circles. The Level 1 residual is the difference vector between the query state and the Level 1 codeword. Figure 3 In triangle (b), the triangle represents the codeword in the second-level codebook. Quantization repeats the same nearest neighbor selection in the residual space. On a verification set consisting of approximately 3 million historical query state samples, the relative quantization reconstruction errors at quantization depths of 1, 2, and 3 are approximately 0.34, 0.15, and 0.07, respectively. The relative quantization reconstruction errors are normalized based on the norm of the query state and averaged over the samples, showing a decreasing trend at each level. Figure 7 (b) The obtained semantic lexical units at each level are arranged in the order of level 1, level 2, and level 3 to form a semantic lexical sequence.

[0030] The foundation of the financial marketing model is a large language model with a decoder structure containing approximately 7 billion parameters and a vocabulary of approximately 120,000 words. 768 placeholder words are pre-added at the end of the vocabulary, corresponding one-to-one with the 768 codewords in the three-level codebook. In the level l codebook, the codeword with index m corresponds to the (l-1)×256+m-th placeholder word in the extended sequence, where l ranges from 1 to 3 and m ranges from 1 to 256. For example, in the level l codebook, the codeword with index 37 corresponds to the 293rd placeholder word. The embedding vector of each placeholder word is initialized with the vector of the corresponding codeword after linear projection. The linear projection matrix maps the 128-dimensional codeword to a 4096-dimensional word embedding space, obtained through random initialization followed by a fixed method. Initialization with codeword projection ensures that placeholder words corresponding to codewords with close Euclidean distances are close to each other in the word embedding space, resulting in faster and smoother convergence during the fine-tuning stage. Placeholder words only enter the prompt sequence through semantic word mapping and do not participate in the segmentation of ordinary text by the word segmenter; the two do not interfere with each other.

[0031] The financial marketing model employs a low-rank adaptation method for fine-tuning. Low-rank adaptation is applied to the query projection matrix and value projection matrix of each attention layer, with a rank of 16 and a scaling factor of 32. The parameters updated during the fine-tuning phase consist of low-rank adaptation parameters and embedding vectors of 768 placeholder terms, while the base parameters remain frozen. This significantly reduces training and memory overhead compared to full-parameter fine-tuning. Fine-tuning samples are composed of semantic terminology sequences obtained through offline replay of historical behavioral event sequences, paired with manually written compliant copy samples. The compliant copy samples come from existing copy approved in the institution's historical campaigns and copy supplemented by the marketing team. Offline replay also determines the quantization depth based on trigger strength, ensuring that the semantic terminology sequence length distribution seen during the fine-tuning phase is consistent with the online service. For example, 120,000 fine-tuning samples are constructed. The fine-tuning objective is to minimize the word-level cross-entropy loss of compliant copywriting samples. The learning rate is set to 0.0001, and the training is conducted for 3 epochs. The word-level cross-entropy loss is calculated only for the words corresponding to compliant copywriting samples, and the hint sequence part is excluded by loss masking. Thus, the model's fitting ability is focused on the copywriting generation itself. After fine-tuning, the low-rank adaptation parameters are merged back into the base weights, and no additional matrix multiplication overhead is introduced during online inference.

[0032] During online service, each semantic lexical in the semantic lexical sequence is mapped to a corresponding placeholder lexical, which is sequentially concatenated with the product description text and marketing task instruction text of the target financial product to form a prompt sequence. An example prompt sequence is as follows: the 87th, 293rd, and 713th placeholder lexicals, the product description text "XX mixed fund, risk level R3, maximum drawdown in the past year 6.2%, maximum stock position 40%", and the marketing task instruction text "Generate a push notification with no more than 60 characters, in a stable tone". The product description text and marketing task instruction text are derived from product operation configurations, with a combined length controlled within 120 lexicals, and the total length of the prompt sequence does not exceed 123 lexicals. The reason for using placeholder lexicals instead of directly concatenating numerical vectors is that semantic lexicals are isomorphic to text lexicals after entering the vocabulary. The attention mechanism can directly establish a connection between the placeholder lexicals and the lexicals of the product description text and marketing task instruction text. The variable-length semantic lexical sequence also naturally adapts to the profile granularity that changes with the trigger intensity.

[0033] The compliance state machine is pre-built offline. The list of prohibited expressions compiles expressions forbidden by regulatory requirements and the institution's internal control checklist, exemplarily containing approximately 200 entries, including terms such as "guaranteed principal," "guaranteed profit," "risk-free," "promised returns," and "guaranteed principal and interest." The context prefix set consists of one empty string, 300 high-frequency Chinese characters, and 14 commonly used punctuation marks, totaling 315 elements. The high-frequency Chinese characters are the 300 most frequently occurring characters from the institution's historical marketing copy corpus; the context suffix set is constructed in the same way. Each prohibited expression is combined with each context prefix and each context suffix to obtain a concatenated string of approximately 200 × 315 × 315 characters, totaling approximately 19.84 million strings, all of which are fed into the word segmenter of the financial marketing big data model. The word segmenter used to construct the pattern string is the same and the same version as the word segmenter used for online decoding. The lexical numbers of the pattern string and the lexical numbers produced in the decoding step are in the same lexical space; only word-by-word matching is meaningful. The word segmenter outputs character offset information along with the word sequence, i.e., the start and end positions of each word in the concatenated string. Based on the character offset information, it extracts consecutive words that overlap with the character range of the violation, forming the word sequence corresponding to the violation in the current context. The necessity of double-sided enumeration stems from the contextual relevance of sub-word segmentation: the first character of the violation may merge with the preceding character into one word, and the last character may merge with the following character into one word. The word sequence obtained by segmenting the violation separately cannot cover these two types of cross-boundary segmentation. However, the truncation criterion for overlapping character ranges is symmetrical to both left and right boundaries, and words merged across either boundary will be included. After summarizing and deduplicating all the combined word sequences, approximately 26,000 pattern strings are obtained for example.

[0034] An Aho-Corasick automaton is constructed as the compliance state machine using all deduplicated pattern strings: First, a lexical-level prefix tree is built from the set of pattern strings. Then, a failure pointer is calculated for each state in breadth-first order, pointing to the state that matches the longest proper suffix of the lexical sequence corresponding to the current state. During the construction process, the output set propagates along the failure pointers. If the state pointed to by the failure pointer of any state corresponds to a complete match of a pattern string, the matching information is incorporated into the current state. States with non-empty output sets are all recorded as hit states, and the remaining states are safe states. Matches completed through failure transitions are thus guaranteed to be detected. The compliance state machine in the example has approximately 48,000 states. Figure 4Taking 4 pattern strings generated from the non-compliant expressions "capital preservation" and "capital preservation and interest guarantee" as an example, a transition structure is provided, where solid lines are transition edges, the corresponding tokens are marked on the edges, double circles represent hit states, the dashed line is the failure pointer pointing from state three to state one, the token sequence corresponding to state three ends with the token "guarantee", which is exactly the first token of the pattern string, and the failure pointers of the remaining states all point to the initial state, which is omitted in the figure. The function of the failure pointer can be illustrated by the scenario where the non-compliant expression "capital preservation" appears twice in a row: the corresponding 4 tokens are "guarantee", "capital", "guarantee" and "capital" in sequence. After decoding the first 3 tokens, the compliance state machine reaches state three. The 4th token "capital" has no corresponding transition edge in state three, the matching rolls back to state one along the failure pointer, and then reaches hit state two via the transition edge, so that the interrupted spanning matching will not be missed.

[0035] To decouple the overhead of online decoding from the vocabulary size, a masked token set is precomputed offline for each state of the compliance state machine: traverse all tokens in the vocabulary, calculate the post-transition state after reading the corresponding token by first checking the transition edge, and rolling back along the failure pointer if there is no transition edge, and add the token whose post-transition state is a hit state to the masked token set of the current state. Statistics show that the masked token set of each state contains less than 3 tokens on average, the storage overhead of all masked token sets is on the order of tens of megabytes, which is stored in the shared memory indexed by state number, and is shared by multiple concurrent decoding paths without being copied as the beam width increases. The online token-by-token decoding adopts beam search with a beam width of 4. Each candidate sequence in the beam maintains one compliance state machine state, and the compliance state machine state is copied synchronously when the candidate sequence is extended. At each decoding step, the masked token set corresponding to the state of the current candidate sequence is queried, the generation probability of candidate next tokens within the set is set to 0, the generation probabilities of the remaining candidate next tokens are renormalized, and then the token selection and compliance state machine state update are completed according to the renormalized generation probabilities. The extra latency caused by masking in a single decoding step is less than 0.2 milliseconds. The compliance state machine state is only an integer number, and the overhead of copying it along with the candidate sequence is negligible. Decoding terminates when an end-of-generation token is generated or the length limit of 64 tokens is reached, and the candidate sequence with the highest cumulative log-probability score in the beam is output as the candidate marketing copy.

[0036] Figure 5 An example of one decoding step is given: the generated text ends with the token "guarantee", and the compliance state machine is in Figure 4 state one in, Figure 5 where (a) shows the generation probabilities of some candidate next tokens before masking, and the candidate next token "capital" will cause state one to transition to a hit state, Figure 5In (b), the result is obtained by setting the generation probability of the character "ben" (本) to 0 and performing renormalization, where the model naturally continues writing among the remaining candidate next word elements, for example, selecting the character "chi" (持) to generate the term "baochi" (保持). Compared with post-generation filtering, zero-setting at the decoding end realizes deterministic interception of the stored pattern strings at the word level, and produces usable copy in a single decoding pass, eliminating the computational overhead of regenerating the entire content after a filtering hit; renormalization ensures that the probability distribution over the remaining candidate next word elements remains valid, keeping the continuation smooth.

[0037] After the decoding is completed and the candidate marketing copy is obtained, a risk prompt template is selected from the risk prompt template library according to the product risk level of the target financial product, and is spliced to the end of the candidate marketing copy to obtain the copy to be delivered. The risk prompt template library pre-stores 5 templates corresponding to product risk levels R1 to R5, and an example of the template corresponding to the R3 level is "The market carries risks, and investment requires caution. The past performance of funds does not indicate their future performance". Figure 2 The candidate marketing copy generated by the example trigger is: "The steady allocation direction you follow has new updates recently: XX hybrid fund focuses on interest rate bonds and high-grade credit bonds, with a purchase discount as low as 10% of the original rate, click to view details".

[0038] The ad copy is fed into a 12-layer pre-trained Chinese text encoder. The output 768-dimensional sentence vector is used as the ad copy vector. The 128-dimensional query state vector and the 768-dimensional ad copy vector are concatenated to form an 896-dimensional vector, which is then input into the matching evidence network. The matching evidence network is a 3-layer multilayer perceptron with hidden layer dimensions of 256 and 64 respectively. The output layer outputs a 2-dimensional vector, which is then activated by softplus to obtain the evidence vector. The two components correspond to the amount of evidence for the conversion category and other categories, respectively. The larger the component value, the more sufficient the evidence supporting the corresponding category. The concatenation of the query state and the ad copy vector allows the matching evidence network to see information from both the user side and the ad copy side simultaneously. It learns the conversion evidence in the combined sense of both, rather than scoring one side in isolation. The training of the matching evidence network relies on historical campaign records. Each historical campaign record consists of the query status, copy vector, and real conversion tag corresponding to one historical campaign. The real conversion tag is determined based on whether the target user made a purchase within 7 days after the campaign. The 7-day attribution window takes into account both the natural latency of conversion behavior and the timeliness of feedback. For example, 500,000 records are collected and randomly divided into a training set of 400,000 records and a calibration set of 100,000 records at a preset ratio of 8:2. The conversion sample accounts for about 6%. During training, the original ratio is maintained and no resampling is performed, thus preserving the absolute level of evidence quantity and the true sparsity of conversion behavior. The division is performed before training. The calibration set does not participate in any training stage of the matching evidence network. This is the premise for the subsequent quantile threshold to control the proportion of online traffic. The training objective is the expected mean squared error loss under the evidence deep learning framework: the evidence vector is determined to be a Dirichlet distribution. The squared error between the one-hot vector of the real conversion tag and the class probability is expected under the Dirichlet distribution, and then summed over the two classes as the loss of a single sample. The optimizer used is Adam, with a learning rate of 0.001, a batch size of 512, and 20 training rounds. During inference, the uncertainty is calculated based on the evidence vector using formula (4): (4); In formula (4), u is the uncertainty; K is the number of categories, which is taken as 2; e k Let S be the k-th component of the evidence vector; S is the total amount of evidence. The larger the total amount of evidence, the more supporting evidence the matching evidence network accumulates for the combination of the current query state and the text vector, and the smaller the uncertainty. For combinations that deviate from the training distribution, both components approach 0 simultaneously, and the uncertainty approaches 1, naturally giving a signal of insufficient evidence. This is also why deep learning evidence is chosen instead of softmax confidence: softmax output may still give a confidence level close to 1 on out-of-distribution samples, while explicit modeling of the amount of evidence expresses "don't know" as a measurable quantity.

[0039] The first and second gate thresholds are predetermined based on the calibration set. Uncertainty is calculated for all 100,000 historical delivery records in the calibration set and sorted in ascending order. The first and second positions are calculated using formula (5): (5); In formula (5), r1 is the first ordinal position, r2 is the second ordinal position; n is the number of samples in the calibration set; α1 is the first significance level, which is 0.2; α2 is the second significance level, which is 0.05; This indicates rounding up. When n is 100000, r1 is 80001 and r2 is 95001. After ascending order, the uncertainty at the 80001st position (0.45) is used as the first gate threshold, and the uncertainty at the 95001st position (0.60) is used as the second gate threshold. The distribution of uncertainty on the calibration set and the positions of the two gate thresholds are shown in [the diagram]. Figure 6 The right tail of the distribution corresponds to sparse evidence combinations, and the sample size decreases rapidly as uncertainty increases. Two gating thresholds divide the right tail into two levels: reduced-frequency delivery and manual review. The second order does not exceed the requirement that the number of samples in the calibration set (n+1)×(1-α2) does not exceed n, i.e., n is greater than or equal to 19. The calibration set of 100,000 samples is far above the lower limit. The calibration set and the online delivery samples come from the same distribution, and the calibration set did not participate in the training. The proportion of online sample uncertainty exceeding the first and second gating thresholds is thus controlled near the first and second significance levels, respectively. The traffic proportions of the three levels—direct delivery, reduced-frequency delivery, and manual review—are approximately 80%, 15%, and 5%, respectively, and the manual review queue obtains a predictable upper bound on the load.

[0040] When the uncertainty is less than or equal to the first threshold of 0.45, the ad copy is directly pushed to the target user through the ad delivery channel. When the uncertainty is greater than 0.45 but less than or equal to the second threshold of 0.60, the default exposure frequency limit is multiplied by a preset reduction factor to obtain the reduced exposure frequency limit. In the example, the default exposure frequency limit is 6 times per week, the preset reduction factor is 0.5, and the reduced exposure frequency limit is 3 times per week. The ad copy is pushed to the target user according to the reduced exposure frequency limit, balancing conversion opportunities and user experience with a lower disturbance limit. When the uncertainty is greater than 0.60, the ad copy is sent to the manual review queue. After manual review and confirmation, it is pushed to the target user. The manual review queue is arranged in descending order of uncertainty. Reviewers prioritize ad copy with the weakest evidence. Ad copy that passes review is pushed according to the default exposure frequency limit and is returned to the historical ad delivery record as usual. Ad copy rejected by review is not delivered and no historical ad delivery record is generated. Figure 2For example, the copy to be delivered and the query status are matched by the evidence network to obtain the evidence vector (6.2, 1.8). The total amount of evidence is 10. According to formula (4), the uncertainty is 0.2, which is less than the first gate threshold. It is directly pushed to the target user.

[0041] The end-to-end latency from triggering to completing the delivery decision is mainly composed of word-by-word decoding. Generating a bundle of about 4 to 60 words takes approximately 1.2 seconds on two GPUs. The total time for residual quantization's three nearest neighbor searches, masked word set queries, and forward pass of the matching evidence network is less than 10 milliseconds. The time consumed in the decision-making process is negligible compared to the preset observation period of 30 minutes, ensuring that the interest vertex closest to the trigger is not missed due to the decision itself.

[0042] After the deployment is completed, the conversion feedback of the target users flows back to the two stages of the method. On the one hand, the click event in the conversion feedback is mapped to the event type identifier two corresponding to the control click modality, and the subscription event is mapped to the event type identifier three corresponding to the transaction flow modality. It is converted into a new behavioral event and written into the behavioral event sequence to participate in the calculation of the subsequent comprehensive incentive intensity. On the other hand, the query status, copy vector and real conversion tag corresponding to this deployment constitute a new historical deployment record. It is randomly allocated to the training set or calibration set according to the preset ratio of 8 to 2. The allocation result remains fixed and no historical deployment record migrates between the two sets. According to the preset cycle of once a week, the expanded training set is first used to perform incremental training on the matching evidence network. The learning rate is 0.0001. The training is carried out for 2 rounds. The training is completed in about 20 minutes on a graphics processor and does not affect the online service. Then, the uncertainty of all historical deployment records in the expanded calibration set is recalculated with the incrementally trained matching evidence network, and the first gate threshold and the second gate threshold are re-determined according to the order of formula (5). The order of updating the network first and then resetting the threshold is necessary: ​​the value of uncertainty changes with the parameters of the matching evidence network, and the threshold must correspond to the same version of the network that generates online uncertainty so that the traffic proportion of each level will not deviate from the set significance level due to the network update.

[0043] Based on the above implementation methods, several alternative implementation methods exist. The preset observation duration can be set between 10 and 60 minutes; a shorter value is suitable for user groups with active transaction flow modalities. The number of intensity threshold tiers can be 2 or 4, with the codebook level and placeholder word number adjusted accordingly. A hysteresis mechanism can supplement the cooling-off period, requiring the overall incentive intensity to fall below 0.8 times the first intensity threshold before resuming the upward triggering judgment, further suppressing repeated triggering near the threshold. The gated recurrent unit in the event encoding network can be replaced with a long short-term memory unit. The intensity readout layer can be expanded to 4 layers, outputting the submodal incentive intensities of the 4 behavioral modalities respectively. The overall incentive intensity is obtained by weighted summation using learnable weights, and the weights and the event encoding network are jointly trained under the same maximum likelihood objective. The linear projection matrix can be jointly trained with a 3-level codebook, replacing the fixed linear projection matrix after random initialization. The method is fixed; the context prefix set and context suffix set can be expanded quarterly based on the segmentation statistics in the online generated logs; after the illegal expression vocabulary is updated, the concatenated string segmentation and pattern string construction can be re-executed, and the construction process is completed within a few minutes, and it takes effect immediately after replacing the compliance state machine; the number of sampling moments for Monte Carlo approximation can be adjusted between 10 and 50, the denser the sampling, the more accurate the integral approximation, and the training time will increase accordingly; the risk warning template library can be replaced in real time as regulatory requirements are updated, and the replacement will not affect the decoding process; the number of parameters of the financial marketing big model can be selected between 1.3 billion and 10 billion; the pre-trained text encoder can be replaced with other encoders that output fixed-length sentence vectors; the preset calibration cycle can be shortened to once a day, further tightening the fit between the traffic proportion and significance level of each level.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating and optimizing a large-scale financial marketing model based on multimodal user behavior modeling, characterized in that, Includes the following steps: Step 1: Convert the raw behavioral data generated by the target user under multiple behavioral modalities into behavioral events one by one. Each behavioral event consists of an event type identifier, an event time, and an event attribute vector. Arrange all behavioral events in ascending order of event time to form a behavioral event sequence. Use an event coding network to encode the behavioral event sequence online to obtain the user's latent state that evolves continuously over time between adjacent behavioral events. The comprehensive incentive intensity is obtained by performing a linear transformation on the user's latent state by the intensity readout layer and then activating it with softplus. The comprehensive incentive intensity represents the event occurrence rate per unit time. The event encoding network and the intensity readout layer are pre-trained with the maximum likelihood of the behavioral event sequence in the temporal point process framework as the training objective; Step 2: When the comprehensive incentive intensity changes from less than the preset first intensity threshold to greater than the first intensity threshold, the moment when the comprehensive incentive intensity reaches its maximum value within the subsequent preset observation period is determined as the trigger moment, the comprehensive incentive intensity at the trigger moment is recorded as the trigger intensity, and the user's hidden state at the trigger moment is recorded as the query state. The quantization depth is determined based on the trigger strength. Residual quantization is performed on the query status level by level according to the quantization depth. The index of the codeword selected by the progressive residual quantization is used as the semantic unit of the corresponding level. The semantic units of each level are arranged in order to form a semantic unit sequence. Step 3: Map each semantic word in the semantic word sequence to a corresponding placeholder word in the pre-expanded vocabulary of the financial marketing big data model, and concatenate it sequentially with the product description text of the target financial product and the marketing task instruction text to form a prompt sequence; input the prompt sequence into the financial marketing big data model for word-by-word decoding to obtain candidate marketing copy; in each decoding step of word-by-word decoding, set the generation probability of the next candidate word that will cause the compliance state machine to transition from the current state to the hit state to 0, and renormalize the generation probability of the remaining next candidate words. The compliance state machine is pre-constructed using the sequence of words corresponding to the violation as pattern strings. The hit state is the state in the compliance state machine that completely matches any pattern string. A risk warning template corresponding to the target financial product is appended to the end of the candidate marketing copy to obtain the copy to be delivered. Step 4: Encode the text to be delivered into a text vector. Concatenate the query status with the text vector and input it into the matching evidence network to obtain the evidence vector. Calculate the uncertainty based on the evidence vector. When the uncertainty is less than or equal to the predetermined first gate threshold, the text to be delivered is directly pushed to the target user through the delivery channel. When the uncertainty is greater than the first gate threshold but less than or equal to the predetermined second gate threshold, multiply the default exposure frequency limit by a preset reduction factor to obtain the reduced exposure frequency limit, and push the text to be delivered to the target user at the reduced exposure frequency limit. When the uncertainty is greater than the second gate threshold, the text to be delivered is sent to the manual review queue. After manual review and confirmation, it is pushed to the target user. The first gate threshold is less than or equal to the second gate threshold.

2. The method according to claim 1, characterized in that, In step 1, the behavioral modalities include page browsing modalities, control clicking modalities, transaction flow modalities, and customer service conversation modalities; the event attribute vectors are obtained by encoding the original behavioral data by the feature encoder corresponding to the behavioral modalities and then mapping it to a unified dimension through a linear layer.

3. The method according to claim 1, characterized in that, In step 1, online encoding includes: initializing the user's hidden state before the arrival of the first behavioral event as a zero vector; when each behavioral event arrives, the event encoding network takes the type embedding vector corresponding to the event type identifier of the current behavioral event, the event attribute vector of the current behavioral event, and the user's hidden state at the instant before the arrival of the current behavioral event as input, updates the hidden state after the transition through a gated recurrent unit, obtains the steady-state vector through a linear output head, and obtains the decay rate vector through an output head with softplus activation, where each component of the decay rate vector takes a positive value; at any time between two adjacent behavioral events, the difference between the user's hidden state and the steady-state vector corresponding to the previous behavioral event decays exponentially with time from the event time of the previous behavioral event, the initial value of the difference is the difference between the hidden state after the transition and the steady-state vector corresponding to the previous behavioral event, and the element-wise decay rate is determined by the corresponding component of the decay rate vector corresponding to the previous behavioral event.

4. The method according to claim 1, characterized in that, In step 2, determining the quantization depth based on the trigger strength includes: pre-setting a second strength threshold and a third strength threshold, where the first strength threshold is less than the second strength threshold and the second strength threshold is less than the third strength threshold; when the trigger strength is less than the second strength threshold, the quantization depth is 1; when the trigger strength is greater than or equal to the second strength threshold and less than the third strength threshold, the quantization depth is 2; and when the trigger strength is greater than or equal to the third strength threshold, the quantization depth is 3.

5. The method according to claim 4, characterized in that, In step 2, the progressive residual quantization includes: pre-training a Level 1 codebook, a Level 2 codebook, and a Level 3 codebook. Each level codebook contains a preset number of codewords, and the codewords have the same dimension as the user's latent state. During the offline training phase, the Level 3 codebook updates the codewords using an exponential moving average method with the goal of minimizing the quantization reconstruction error of historical query state samples. In the Level 1 codebook, the codeword with the smallest Euclidean distance to the query state is selected as the Level 1 codeword, and the index of the Level 1 codeword in the Level 1 codebook is used as the Level 1 semantic term. When the quantization depth is greater than or equal to 2, the query state is subtracted from the Level 1 codeword. The codewords yield the first-level residual. In the second-level codebook, the codeword with the smallest Euclidean distance to the first-level residual is selected as the second-level codeword. The index of the second-level codeword in the second-level codebook is used as the second-level semantic token. When the quantization depth is 3, the second-level codeword is subtracted from the first-level residual to obtain the second-level residual. In the third-level codebook, the codeword with the smallest Euclidean distance to the second-level residual is selected as the third-level codeword. The index of the third-level codeword in the third-level codebook is used as the third-level semantic token. The resulting semantic tokens at each level are arranged in the order of first-level, second-level, and third-level to form a semantic token sequence.

6. The method according to claim 5, characterized in that, In the vocabulary of the financial marketing big model, each codeword in the 3-level codebook is pre-extended with one corresponding placeholder word. The embedding vector of each placeholder word is initialized with the vector of the corresponding codeword after linear projection. The financial marketing big model is fine-tuned using a low-rank adaptation method. The parameters updated in the fine-tuning stage consist of the low-rank adaptation parameters and the embedding vectors of all placeholder words, while the base parameters remain frozen. The fine-tuning sample consists of a semantic lexical sequence obtained by offline replay of historical behavioral event sequences in steps 1 and 2, paired with manually written compliance copy samples. The fine-tuning objective is to minimize the lexical-level cross-entropy loss of the compliance copy samples. The financial marketing big model with input prompt sequence in step 3 is the fine-tuned financial marketing big model.

7. The method according to claim 1, characterized in that, The compliance state machine is pre-built in the following way: a set of context prefixes and a set of context suffixes consisting of an empty string, high-frequency Chinese characters and common punctuation marks are constructed; each violation expression in the violation expression vocabulary is combined with each context prefix and each context suffix to obtain a concatenated string formed by sequentially connecting the context prefix, the violation expression and the context suffix. The concatenated string is input into the word segmenter of the financial marketing big data model. Based on the character offset information output by the word segmenter, continuous word units that overlap with the character range of the violation are extracted to form the word unit sequence corresponding to the violation. The word sequences obtained from all combinations of all context prefixes and context suffixes of all illegal expressions are summarized and deduplicated. The Aho-Corasick automaton is constructed using the deduplicated word sequences as the pattern string, which serves as the compliance state machine. In the compliance state machine, the state that completely matches any pattern string is the hit state, and the rest are safe states.

8. The method according to claim 1, characterized in that, In step 3, word-by-word decoding uses a bundle search. Each candidate sequence in the bundle maintains its own compliance state machine state. When expanding the candidate sequence, the compliance state machine state is copied synchronously. In each decoding step, for each candidate next word of each candidate sequence, the transition state is calculated according to the transition table of the compliance state machine. The candidate next word whose transition state is the hit state is the candidate next word whose generation probability is set to 0. The word selection and compliance state machine state update of the current decoding step are completed according to the renormalized generation probability. The risk warning template is selected from the risk warning template library based on the product risk level of the target financial product.

9. The method according to claim 1, characterized in that, In step 4, the copywriting vector is obtained by inputting the copywriting to be delivered into a pre-trained text encoder; the matching evidence network is a multilayer perceptron with softplus activation in the output layer, which outputs a 2-dimensional evidence vector. The two components of the evidence vector correspond to the evidence amount of the conversion category and other categories, respectively; historical delivery records are collected in advance. Each historical delivery record consists of the query status, copywriting vector, and real conversion label corresponding to one historical delivery. All historical delivery records are randomly divided into a training set and a calibration set according to a preset ratio; the matching evidence network is pre-trained using the training set as training samples and the expected mean squared error loss under the evidence deep learning framework as the training objective. The uncertainty calculation based on the evidence vector includes: adding 1 to each of the two components of the evidence vector and summing them to obtain the total amount of evidence. The ratio of the number of categories to the total amount of evidence is taken as the uncertainty, and the number of categories is taken as 2.

10. The method according to claim 9, characterized in that, The first and second gate thresholds are predetermined as follows: Uncertainties are calculated for all historical delivery records in the calibration set and sorted in ascending order; the sum of the calibration set sample size plus 1 is multiplied by the difference between 1 and the preset first significance level, and the difference between 1 and the preset second significance level, respectively. The two products are rounded up to obtain the first and second rank; the uncertainty at the first rank after ascending order is used as the first gate threshold, and the uncertainty at the second rank after ascending order is used as the second gate threshold; if the first significance level is greater than the second significance level, the calibration set sample size is chosen such that the second rank is less than or equal to the calibration set sample size. The method also includes: after the campaign is completed, mapping the click events and subscription events in the conversion feedback of the target users to the corresponding event type identifiers, converting them into new behavioral events and writing them into the behavioral event sequence; constructing new historical campaign records by combining the query status, copy vector and real conversion tags corresponding to this campaign, and randomly allocating them to the training set or calibration set according to a preset ratio, with the allocation results remaining fixed; performing incremental training on the matching evidence network with the expanded training set according to a preset period, and then recalculating the uncertainty of all historical campaign records in the expanded calibration set with the incrementally trained matching evidence network, and redetermining the first and second gate thresholds based on the recalculated uncertainty.

Images