A breath-type intelligent touch decision method and system based on an intention judgment model

Abstract

A breath-type intelligent touch decision method and system based on an intention judgment model, wherein the method comprises: acquiring a multi-modal behavior signal of a target user in a current session; extracting a semantic feature vector and performing splicing coding to obtain a current behavior feature vector; acquiring a historical intention vector and performing weighted fusion with the current behavior feature vector to obtain a fusion feature vector; inputting the fusion feature vector into a preset intention judgment model to obtain probability values in each preset intention stage; calculating an intention drift value and calculating the similarity between the semantic feature vector and each preset service scene anchor vector to determine a target service scene; calculating an interaction trigger probability; when the interaction trigger probability is greater than a preset trigger threshold, generating an interaction decision scheme of the target user and rendering a corresponding interaction interface on the client according to the interaction decision scheme. The application effectively suppresses invalid rejection operations through dynamic smoothing interface rendering, reduces system redundancy overhead, and improves interaction decision efficiency.

Description

Technical Field

[0001] This application relates to the field of interactive decision-making technology, specifically to a breathing-style intelligent outreach decision-making method and system based on an intent judgment model. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, intelligent human-computer interaction systems (such as intelligent customer service, virtual assistants, and intelligent recommendation and shopping guides) have become widely used in various terminal applications. In order to improve user experience, current intelligent interaction systems are gradually shifting from "passive response" to "proactive service," that is, the system attempts to proactively provide decision-making suggestions or service guidance by understanding user behavior.

[0003] Existing technologies typically collect single user inputs (such as search text or click actions) in real time and feed them into a pre-trained classification model for intent recognition. When the confidence level of a certain intent output by the model reaches a fixed threshold, the system immediately triggers the corresponding interaction logic, such as directly popping up a service prompt on the current page or redirecting to a recommendation page.

[0004] However, in real-world, complex interaction scenarios, user intentions are rarely instantaneous, but rather a continuous evolutionary process accompanied by hesitation, exploration, and reflection. Existing interaction decision-making mechanisms primarily rely on static slice analysis of isolated user behaviors, failing to effectively perceive and measure the dynamic fluctuations and instabilities of a user's psychological state during an interaction. This makes it difficult for the system to accurately grasp the timing of proactive intervention: either triggering interaction prematurely while the user is still in a aimless browsing phase, frequently interrupting the user's normal operations; or failing to respond in a timely manner when the user truly needs guidance, reducing the efficiency of interaction decision-making. Summary of the Invention

[0005] This application provides a breathing-style intelligent outreach decision-making method and system based on an intent judgment model, which solves the technical problem of difficulty in accurately grasping the timing of proactive intervention in interactive decision-making and improves the efficiency of interactive decision-making.

[0006] The first aspect of this application provides a breathing-style intelligent outreach decision-making method based on an intent judgment model. The method includes: periodically acquiring multimodal behavior signals of a target user in the current session according to a preset sliding time window, wherein the multimodal behavior signals include interactive behavior signals and text semantic signals. Semantic feature vectors are extracted from the text semantic signals, and the interaction behavior signals are concatenated and encoded with the semantic feature vectors to obtain the current behavior feature vector of the current sliding time window; Obtain the historical intent vector of the target user within a preset historical time period, and then perform a weighted fusion of the historical intent vector and the current behavior feature vector to obtain a fused feature vector. The fused feature vector is input into a preset intent judgment model to obtain the probability value of the target user being in each preset intent stage within the current sliding time window, and stored in a preset time-series cache queue; Extract the probability values ​​within multiple consecutive sliding time windows in the time-series cache queue, and calculate the intention drift value based on the temporal fluctuation of the probability values; Calculate the similarity between the semantic feature vector and each preset service scenario anchor vector, and determine the service scenario corresponding to the preset service scenario anchor vector whose similarity is greater than a preset similarity threshold as the target service scenario; The probability of interaction triggering is calculated based on the probability value, the intent drift value, and the similarity of each preset intent stage. When the probability of triggering the interaction is greater than the preset trigger threshold, an interaction decision scheme for the target user is generated based on the target service scenario and the probability values ​​of each preset intent stage, and the corresponding interactive interface is rendered on the client according to the interaction decision scheme, wherein the transparency rendering parameter of the interactive interface is dynamically updated gradually according to a preset periodic function.

[0007] Optionally, a semantic feature vector is extracted from the text semantic signal, and the interaction behavior signal is concatenated and encoded with the semantic feature vector to obtain the current behavior feature vector of the current sliding time window, specifically including: Extract the back number, idle time, and tab switching number corresponding to the current swipe time window from the interaction behavior signal, and use a preset activation function to weight and combine the back number, idle time, and tab switching number to obtain the hesitation behavior pattern features; Obtain the page dwell time, scroll depth, and revisit frequency within the current swipe time window from the interaction behavior signal; The page dwell time, scroll depth, revisit frequency, and hesitation behavior pattern features are concatenated with the semantic feature vector, and then linearly transformed using a preset encoding weight matrix and bias term to obtain the current behavior feature vector.

[0008] Optionally, the historical intent vector and the current behavior feature vector are weighted and fused to obtain a fused feature vector, specifically including: Obtain the historical update time corresponding to the historical intent vector, and calculate the time interval between the historical update time and the current time; Based on the time interval, the historical intent weighting coefficient is dynamically calculated using a preset time decay function; Based on the historical intent weight coefficient, the current behavior weight coefficient is calculated, and the sum of the historical intent weight coefficient and the current behavior weight coefficient is one. Multiply the current behavior feature vector by the current behavior weight coefficient to obtain the current behavior weighted vector; Multiply the historical intent vector by the historical intent weight coefficient to obtain the historical intent weighted vector; The current behavior weighted vector and the historical intent weighted vector are summed to obtain the fused feature vector.

[0009] Optionally, based on the time interval, the historical intent weighting coefficient is dynamically calculated using a preset time decay function, specifically including: The time interval is input into the preset time decay function to obtain the initial historical weight; Obtain the historical session interaction depth and historical session duration corresponding to the historical intent vector; If the time interval is less than a preset recent time threshold, and the historical session interaction depth is less than a preset depth threshold, and the historical session duration is less than a preset duration threshold, then the historical intent vector is determined to be an accidental interference intent, and the penalty decay factor is calculated based on the time interval, the historical session interaction depth, and the historical session duration. Multiply the initial historical weight by the penalty decay factor to obtain the historical intention weight coefficient; If the time interval is greater than or equal to the preset recent time threshold, or the historical session interaction depth is greater than or equal to the preset depth threshold, or the historical session duration is greater than or equal to the preset duration threshold, then the initial historical weight is used as the historical intent weight coefficient.

[0010] Optionally, the penalty decay factor is calculated based on the time interval, the historical session interaction depth, and the historical session duration, specifically including: The time interval, the historical session interaction depth, and the historical session duration are normalized and then weighted and summed to obtain a basic penalty factor with a value greater than zero and less than or equal to a preset penalty upper limit, wherein the preset penalty upper limit is less than one. Historical semantic features are extracted from the historical intent vector, and the semantic divergence between the historical semantic features and the semantic feature vector in the current behavior feature vector is calculated. If the semantic divergence is less than or equal to the preset semantic conflict threshold, then the basic penalty factor is used as the penalty decay factor; If the semantic divergence is greater than the preset semantic conflict threshold, then the accidental interference intent and the current intent have heterogeneous semantic noise. The semantic divergence is used as the attenuation index to perform exponential attenuation calculation on the basic penalty factor to obtain the penalty attenuation factor.

[0011] Optionally, the interaction trigger probability is calculated based on the probability value, the intent drift value, and the similarity of each preset intent stage, specifically including: The maximum probability value is extracted from the probability values ​​of each preset intent stage and used as the target intent score, and the similarity is used as the scene matching score; Based on preset intent weights and preset scene weights, the target intent score and the scene matching score are weighted and summed to obtain the initial trigger probability; If the target intent score is greater than a preset high intent threshold and the intent drift value is greater than a preset oscillation drift threshold, then the target user is determined to be in a high-frequency oscillation decision state. When the target user is in the high-frequency oscillation decision state, a preset oscillation suppression floor is obtained, and the intention drift value is used as the exponent of the preset oscillation suppression floor to calculate a suppression coefficient that is dynamically scaled between zero and one. The design value of the preset oscillation suppression floor is greater than zero and less than one. Multiplying the initial trigger probability by the suppression coefficient yields the interaction trigger probability; If the target intent score is less than or equal to the preset high intent threshold, or the intent drift value is less than or equal to the preset oscillation drift threshold, then the intent drift value is multiplied by the preset drift weight to obtain a drift penalty term, and the initial trigger probability is subtracted from the drift penalty term to obtain the interaction trigger probability.

[0012] Optionally, based on the target service scenario and the probability values ​​of each preset intent stage, an interaction decision scheme for the target user is generated, and the corresponding interactive interface is rendered on the client according to the interaction decision scheme, specifically including: Calculate the information entropy of the probability values ​​for each of the preset intention stages; The information entropy is input into a preset periodic mapping function to obtain a dynamic respiratory cycle, wherein the information entropy is positively correlated with the dynamic respiratory cycle; Extract the maximum probability value from the probability values ​​of each preset intention stage, and input the maximum probability value into a preset transparency mapping function to calculate the upper limit of the peak of the gradient transparency; Extract the scene service content corresponding to the target service scene; The dynamic breathing cycle, the peak upper limit, and the scenario service content are combined to generate the interactive decision-making scheme; In response to the interactive decision scheme, the dynamic graphics rendering interface of the client is invoked. By configuring the period parameter of the corresponding preset periodic function in the dynamic rendering interface as the dynamic breathing cycle and configuring the amplitude parameter as the peak value limit, the redrawing frequency of the corresponding rendering thread is controlled by the dynamic breathing cycle. The interactive interface containing the scene service content is rendered, and the transparency rendering parameters of the interactive interface are driven to be dynamically updated in a gradual manner.

[0013] Secondly, embodiments of this application provide a breathing-based intelligent outreach decision system based on an intent judgment model. The breathing-based intelligent outreach decision system based on the intent judgment model includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the breathing-based intelligent outreach decision system based on the intent judgment model to perform the method described in the first aspect and any possible implementation thereof.

[0014] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a breathing-based intelligent outreach decision system based on an intent judgment model, cause the breathing-based intelligent outreach decision system based on the intent judgment model to perform the method described in the first aspect and any possible implementation thereof.

[0015] Fourthly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a breathing-based intelligent outreach decision system based on an intent judgment model, causes the breathing-based intelligent outreach decision system based on the intent judgment model to execute the method described in the first aspect and any possible implementation thereof.

[0016] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: 1. By acquiring multimodal behavioral signals according to a preset sliding time window and weightedly fusing the current behavioral feature vector with the historical intent vector, a coherent perception of the user's continuous interaction state is achieved. Based on this, the fused feature vector is input into a preset intent judgment model to obtain probability values ​​for each preset intent stage, which are then stored in a preset temporal cache queue. Furthermore, intent drift values ​​are calculated based on the temporal fluctuations of probability values ​​within consecutive adjacent windows, quantifying the dynamic fluctuations and instability of the user's psychological state during interaction and truly reflecting the continuous evolution of intent. Further, the target service scenario is determined by combining the similarity between the semantic feature vector and the preset service scenario anchor vector. The interaction trigger probability is calculated by integrating the probability values ​​of each preset intent stage, the intent drift value, and the aforementioned similarity. This allows for precise timing of proactive intervention based on the user's dynamic intent evolution, avoiding premature interaction that interrupts normal user operation or untimely response, thus improving interaction decision-making efficiency. Finally, when triggering an interaction, the transparency rendering parameters of the interaction interface are dynamically updated according to a preset periodic function. This dynamic and smooth interface rendering effectively suppresses invalid rejection operations, reducing system redundancy overhead while improving interaction decision-making efficiency.

[0017] 2. Further, by extracting the number of back navigations, idle time, and tab switching times corresponding to the current swipe window from the interaction behavior signals, and weighting these indicators using a preset activation function, the hesitation behavior pattern features are obtained. This precisely quantifies the user's scattered surface interaction actions into high-order features reflecting the user's conflicted and uncertain psychology. Simultaneously, by combining the obtained page dwell time, scroll depth, and revisit frequency within the current swipe window, the explicit attention depth of the user to the current content is captured. Furthermore, the page dwell time, scroll depth, revisit frequency, and hesitation behavior pattern features are concatenated with the semantic feature vector, and a linear transformation is performed using a preset encoding weight matrix and bias term to obtain the current behavior feature vector. This process achieves deep fusion and feature mapping of explicit attention indicators, implicit psychological state features, and specific text semantics, enhancing the representational ability of the current behavior feature vector and providing multi-dimensional data support for more accurate calculation of intent drift values ​​and grasping the timing of interaction triggers.

[0018] 3. Further, by obtaining the historical update time corresponding to the historical intent vector and calculating the time interval with the current time, this is input into a preset time decay function to obtain the initial historical weight, reflecting the objective law that the reference value of historical information decreases over time. Based on this, this solution obtains the historical session interaction depth and historical session duration in one step. When the time interval is less than a preset recent time threshold, and both the historical session interaction depth and historical session duration are less than the corresponding thresholds, the historical intent vector is determined to be an accidental interference intent. A penalty decay factor is calculated based on the above dimensions, and multiplied by the initial historical weight to obtain the final historical intent weight coefficient; otherwise, the initial historical weight is directly used as the historical intent weight coefficient. Subsequently, the current behavior weight coefficient is calculated based on this historical intent weight coefficient, and the current behavior feature vector and the historical intent vector are weighted respectively to obtain the current behavior weighted vector and the historical intent weighted vector, and finally summed to obtain the fused feature vector. It achieves a dynamic balance between historical experience and current behavior in terms of timeliness. Through multi-dimensional threshold judgment, it cleverly identifies and suppresses noise interference caused by short-term and shallow interactions such as accidental touches or rapid exits by users, improves the purity and representation accuracy of the fused feature vector, and thus provides a more reliable and realistic input basis for subsequent intent judgment models. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a breathing-based intelligent outreach decision-making method based on an intent judgment model, as described in an embodiment of this application. Figure 2 This is a schematic diagram of the process for determining the fused feature vector in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a breathing-type intelligent outreach decision system based on an intent judgment model provided in an embodiment of this application.

[0020] Explanation of reference numerals in the attached drawings: 301, Central Processing Unit; 302, Read-Only Memory; 303, Random Access Memory; 304, Bus; 305, Input / Output Interface; 306, Input Section; 307, Output Section; 308, Storage Section; 309, Communication Section; 310, Driver; 311, Removable Media. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0022] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0023] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0024] Figure 1 This is a flowchart illustrating a breathing-style intelligent outreach decision-making method based on an intent judgment model, as described in an embodiment of this application.

[0025] Please see Figure 1 This application provides a breathing-style intelligent outreach decision-making method based on an intent judgment model, the method comprising: S101. Periodically acquire the multimodal behavior signals of the target user in the current session according to a preset sliding time window. The multimodal behavior signals include interactive behavior signals and text semantic signals. In existing intelligent interaction systems, if single input data is collected only when the user makes a clear submission action, the system can only obtain discrete static slices, completely losing the subtle state changes of the user during continuous operation. In order to achieve dynamic perception of the user's continuous interaction state, step S101 periodically acquires the target user's multimodal behavior signals in the current session according to a preset sliding time window.

[0026] To capture data continuously and smoothly along the timeline, the system introduces a preset sliding time window, which is a data capture interval that moves forward along the timeline by a fixed length and a fixed step size. The reason for using a sliding window instead of a fixed, non-overlapping window is that the sliding mechanism preserves overlapping data between adjacent time periods, thus avoiding abruptly truncating continuous user interactions at time boundaries. The preset sliding time window includes a window length parameter and a sliding step size parameter. The window length parameter is set to a range of 3 to 5 seconds, based on the average reaction time of a human completing a basic cognitive action and generating a corresponding interactive action. The sliding step size parameter is set to a range of 1 to 2 seconds, based on ensuring a 50% to 60% data overlap between adjacent windows to maintain a smooth transition of temporal characteristics. The current session refers to the continuous interaction lifecycle of the target user from opening the terminal application interface until actively closing the interface or exceeding a preset silence period of 30 minutes, based on the session timeout standard for conventional web applications.

[0027] During system operation, the front-end acquisition module periodically advances the preset sliding time window according to the set sliding step size, and records in real time all multimodal behavioral signals generated by the target user within the current preset sliding time window. The multimodal behavioral signals are composed of interactive behavioral signals and text semantic signals.

[0028] For interactive behavior signals, the front-end acquisition module listens to Document Object Model (DOM) events and records the action data generated during physical interactions between the target user and the terminal interface in real time. To comprehensively depict the user's browsing state and indecisiveness, the interactive behavior signals specifically include the target user's back navigation count, idle time, tab switching count, page dwell time, scroll depth, and revisit frequency within the current scrolling time window. Back navigation count records the back actions in the browser history; idle time measures the number of milliseconds without any mouse or keyboard operation; tab switching count counts the frequency of changes in the focused window; page dwell time records the active browsing time of the current page; scroll depth extracts the maximum pixel ratio of vertical page scrolling; and revisit frequency counts the number of times the same Uniform Resource Locator (URL) is repeatedly loaded.

[0029] For text semantic signals, the system reads the text value of the input box or extracts the text label of the clicked control to obtain the linguistically meaningful character data entered or triggered by the target user during the interaction. Text semantic signals specifically cover search keywords, speech-to-text transcription, and navigation bar click text, reflecting the user's explicit cognitive needs.

[0030] The advancement logic of the preset sliding time window is expressed by the following formula: W(n)=[T_start+(n-1)*T_step,T_start+(n-1)*T_step+T_len]; In the formula, W(n) represents the nth preset sliding time window, T_start represents the start timestamp of the current session, T_step represents the sliding step size parameter, T_len represents the window length parameter, and n is a positive integer greater than zero.

[0031] By periodically acquiring multimodal behavioral signals as described above, the system not only ensures the temporal continuity of the underlying data by utilizing the overlapping characteristics of the sliding time window, but also constructs a multi-dimensional data foundation by integrating physical interaction signals and cognitive text semantic signals, providing high-density, seamless raw data support for the subsequent accurate quantification of intent drift values.

[0032] Taking an intelligent e-commerce shopping guide scenario as an example, assuming the window length parameter is set to 4 seconds and the sliding step size parameter is set to 2 seconds, within the preset sliding time window from second 0 to second 4, the system acquires interactive behavior signals including the target user scrolling the page down by 80% (scroll depth), the mouse stopping for 1500 milliseconds (idle time), and switching between two different product tabs twice (tab switching count). Simultaneously, the system acquires textual semantic signals including the target user entering "sunscreen" in the search box and clicking the filter button with the text label "suitable for sensitive skin." When the time progresses to second 2, the system begins collecting data within the preset sliding time window from second 2 to second 6. At this point, the data from second 2 to second 4 is shared between the two adjacent windows, ensuring the continuous recording of the user's product comparison actions.

[0033] S102. Extract semantic feature vectors from the text semantic signal, and concatenate and encode the interaction behavior signal with the semantic feature vectors to obtain the current behavior feature vector of the current sliding time window; After acquiring multimodal behavioral signals containing interactive behavior signals and textual semantic signals, feature extraction and fusion encoding are required to transform them into a mathematical expression that the model can process. However, the original interactive behavior signals are complex in dimensions and have diverse physical meanings. Simply concatenating them directly with high-dimensional semantic feature vectors can easily create a semantic gap in the feature space, failing to accurately depict the user's true psychological state within the current sliding time window. Therefore, a refined hierarchical processing of the interactive behavior signals was performed, aiming to extract scattered actions reflecting the user's hesitation and exploration psychology into high-order pattern features, and combine them with behavioral indicators reflecting the user's explicit attention, performing deep linear mapping and fusion encoding with textual semantics. Specifically, the following steps are included: Extract the back number, idle time, and tab switching number corresponding to the current swipe time window from the interaction behavior signal, and use a preset activation function to weight and combine the back number, idle time, and tab switching number to obtain the hesitation behavior pattern features; Obtain the page dwell time, scroll depth, and revisit frequency within the current swipe time window from the interaction behavior signal; The page dwell time, scroll depth, revisit frequency, and hesitation behavior pattern features are concatenated with the semantic feature vector, and then linearly transformed using a preset encoding weight matrix and bias term to obtain the current behavior feature vector.

[0034] In real-world interaction scenarios, users often exhibit unconscious, fragmented movements when facing decision-making difficulties or in the exploratory phase. Directly inputting raw, low-level action data into the model makes it difficult for the model to effectively understand the user's deeper psychological state. To accurately quantify users' hesitation and uncertainty, the system extracts the number of back actions, idle time, and tab switching frequency corresponding to the current swipe window from the interaction behavior signals. The number of back actions records the back actions in the browser history, the idle time measures the number of milliseconds of silence without any mouse or keyboard operation, and the tab switching frequency counts the frequency of changes in the focus window. The system assigns corresponding weight coefficients to the number of back actions, idle time, and tab switching frequency, with the weight coefficients ranging from 0.1 to 0.5, based on statistical analysis of the correlation between various actions and the final unconverted results in a large number of historical interaction logs. The system multiplies the number of back actions, idle time, and tab switching frequency by their corresponding weight coefficients and sums them. The sum is then input into a preset activation function for non-linear mapping to obtain the characteristics of the hesitation behavior pattern. The preset activation function uses the Sigmoid function, which can compress the results of infinitely wide linear combinations into the range of 0 to 1, avoiding abnormal extreme values ​​from interfering with subsequent calculations. Hesitation behavior pattern features, as a high-level indicator for quantifying user indecisiveness, effectively extract fragmented surface interactions into deep features reflecting psychological states. The calculation logic is expressed by the following formula: H = Sigmoid(w1*C_back + w2*T_idle + w3*C_tab); where H represents the hesitation behavior pattern feature, C_back represents the number of back navigations, T_idle represents the idle time, C_tab represents the number of tab switching attempts, and w1, w2, and w3 represent the corresponding weight coefficients.

[0035] Beyond implicit hesitation, the system also needs to measure the user's explicit attention to the current content. Therefore, the system simultaneously acquires page dwell time, scroll depth, and revisit frequency within the current scrolling time window from interaction behavior signals. Page dwell time records the active browsing time of the current page, scroll depth extracts the maximum pixel ratio of vertical page scrolling, and revisit frequency counts the number of times the same Uniform Resource Locator (URL) is repeatedly loaded. Page dwell time, scroll depth, and revisit frequency comprehensively capture the user's explicit attention depth to the current content from three dimensions: time, space, and frequency.

[0036] Because explicit attention metrics, implicit psychological state features, and textual semantic signals reside in completely different feature spaces, direct combination can create a severe semantic gap, hindering model convergence. To achieve deep fusion and spatial alignment of multidimensional features, the system first extracts fixed-dimensional semantic feature vectors from the textual semantic signals using a pre-trained language model. Subsequently, the system concatenates page dwell time, scroll depth, revisit frequency, and hesitation behavior patterns with the semantic feature vectors along the feature dimension, forming a high-dimensional joint feature vector. To eliminate differences in scales and extract deep-level correlation features, the system multiplies the joint feature vector with a pre-defined encoding weight matrix and adds a bias term, performing a linear transformation to obtain the current behavior feature vector for the current scrolling time window. The pre-defined encoding weight matrix and bias term are network parameters of the fully connected layer in a deep neural network, obtained through backpropagation training on massive historical multimodal interaction samples, primarily serving to reduce feature dimensionality and align cross-modal semantics. Through concatenation and linear transformation operations, the system achieves deep fusion of low-level behavioral metrics, high-order psychological features, and specific textual semantics, significantly enhancing the representational power of the current behavior feature vector. The linear transformation logic is expressed by the following formula: V_current = W_enc * Concat(T_stay, D_scroll, F_revisit, H, V_semantic) + B_enc; where V_current represents the current behavior feature vector, W_enc represents the preset encoding weight matrix, Concat represents the concatenation operation, T_stay represents the page stay duration, D_scroll represents the scroll depth, F_revisit represents the revisit frequency, H represents the hesitation behavior pattern feature, V_semantic represents the semantic feature vector, and B_enc represents the bias term.

[0037] Taking the smartphone purchase scenario as an example, within the current scrolling time window, the system extracted the following data: the user retook the page 2 times, had an idle time of 2000 milliseconds, and switched tabs 3 times. The system multiplied these values ​​by their corresponding weights and input them into the Sigmoid function, calculating a hesitation behavior pattern feature of 0.75, indicating that the user is in a highly indecisive state. Simultaneously, the system obtained the page dwell time of 3.5 seconds, the scroll depth of 60%, and the revisit frequency of 2 times. The system concatenates the values ​​of 0.75, 3.5, 0.6, and 2, along with the semantic feature vector extracted from the search term "camera phone recommendation," and finally multiplies it by a preset encoding weight matrix and adds a bias term, outputting a 256-dimensional current behavior feature vector. This provides rich and detailed multi-dimensional data support for the subsequent accurate calculation of intent drift values.

[0038] S103. Obtain the historical intent vector of the target user within a preset historical time period, and perform weighted fusion of the historical intent vector and the current behavior feature vector to obtain a fused feature vector. After obtaining the current behavior feature vector, relying solely on the data within the current sliding time window for decision-making can easily lead to a limited local perspective, neglecting the continuity of user intent and the ongoing influence of historical context. To comprehensively depict the evolution of user intent, the system needs to incorporate historical intent vectors within a preset historical time period for comprehensive consideration. However, the reference value of historical information is not constant but follows the objective law of gradually diminishing over time. Using a fixed fusion ratio can easily lead to distant historical noise interfering with current decisions, or excessive dilution of recent important historical information. To achieve a dynamic balance between historical experience and current behavior in terms of timeliness, a dynamic weight allocation mechanism based on time intervals is introduced to control the degree of interference of historical information on current decisions. Specifically, this includes steps S201-S206. Figure 2 This is a flowchart illustrating the process of determining the fused feature vector in an embodiment of this application. The following is a summary of the process. Figure 2 A detailed explanation of step S103 is provided below: S201. Obtain the historical update time corresponding to the historical intent vector, and calculate the time interval between the historical update time and the current time; In continuous intelligent interaction, users' psychological states and potential needs are constantly evolving. To accurately assess the effective reference value of previously generated intent features for current decision-making, a time dimension must be introduced as a core metric. Without precise time span measurement, the system will be unable to distinguish between recently generated short-term intents and long-term intents that have already expired, leading to a loss of timeliness in subsequent feature fusion. Therefore, the system first obtains the historical update time corresponding to the historical intent vector. The historical update time refers to the absolute timestamp recorded when the system last completed intent recognition calculation and vectorized the result for writing to the storage medium, primarily used as the baseline starting point for subsequent timeliness decay calculations.

[0039] In practice, when the system reads the target user's historical intent vector within a preset historical time period from a preset time-series cache queue, it simultaneously extracts the timestamp field bound to the historical intent vector as the historical update time. Subsequently, the system calls the underlying operating system's clock interface to obtain the absolute timestamp of the current moment as the current time. Finally, the system performs a subtraction operation, subtracting the historical update time from the current time to calculate the time difference, i.e., the time interval. The time interval is typically quantified in seconds or milliseconds.

[0040] By precisely calculating time intervals, the system assigns specific time scales to abstract historical intentions, providing a basic numerical input for quantifying the timeliness of historical information through a decay function, and ensuring that the multimodal feature fusion process has strict time sensitivity.

[0041] Taking e-commerce shopping guide scenario as an example, the system reads the historical update time corresponding to the historical intent vector as 10:15:30 am, and then calls the system clock to get the current time as 10:15:45 am. By subtracting the two, the time interval is calculated to be 15 seconds, indicating that the historical intent occurred within a very short time window from the present, providing a time basis for assigning higher fusion weights in the future.

[0042] S202. Based on the time interval, the historical intent weight coefficient is dynamically calculated using a preset time decay function; After calculating the time interval, relying solely on a single time dimension for exponential decay calculation can lead to an extreme and common interference scenario: when a target user makes an accidental touch or quickly exits the program, the extremely short time interval means that a simple time decay function will assign a very high weight to the noise generated by this accidental interference, severely polluting the fused feature vector. To eliminate short-term and shallow invalid historical information, this solution introduces a multi-dimensional intent validity verification logic on top of the basic time decay mechanism. Specifically, the system not only calculates the initial historical weight based on the time interval but also combines the depth of historical session interactions and the duration of historical sessions to accurately identify accidental interference intents. Furthermore, it dynamically applies a penalty decay factor to correct the initial historical weights, ensuring that the final output historical intent weight coefficients truly reflect the effective reference value of historical experience. This includes: The time interval is input into the preset time decay function to obtain the initial historical weight; Obtain the historical session interaction depth and historical session duration corresponding to the historical intent vector; If the time interval is less than a preset recent time threshold, and the historical session interaction depth is less than a preset depth threshold, and the historical session duration is less than a preset duration threshold, then the historical intent vector is determined to be an accidental interference intent, and the penalty decay factor is calculated based on the time interval, the historical session interaction depth, and the historical session duration. Multiply the initial historical weight by the penalty decay factor to obtain the historical intention weight coefficient; If the time interval is greater than or equal to the preset recent time threshold, or the historical session interaction depth is greater than or equal to the preset depth threshold, or the historical session duration is greater than or equal to the preset duration threshold, then the initial historical weight is used as the historical intent weight coefficient.

[0043] To establish a basic fusion ratio based purely on the time dimension, the system first inputs the time intervals calculated in previous steps into a preset time decay function to obtain initial historical weights. The preset time decay function employs a negative exponential function with a base of the natural constant. Utilizing the initially steep and then gradually decreasing decay characteristics of the negative exponential curve, it objectively simulates the natural forgetting pattern of information over time. The specific calculation logic is expressed by the following formula: W_init = exp(-lambda * T_interval), where W_init represents the initial historical weights, exp represents the exponential function with a base of the natural constant, lambda represents the decay constant, and T_interval represents the time interval. By introducing the preset time decay function, the system can initially quantify the time-sensitive value of historical information, providing a benchmark value for subsequent multi-level corrections.

[0044] However, relying solely on the time decay mechanism is prone to misjudgment. When a target user accidentally clicks a link and quickly exits, the preset time decay function assigns a very high initial historical weight due to the extremely short time proximity of the event, resulting in significant contamination of the current features by invalid noise. To measure the true interaction quality of historical intents, the system further obtains the historical session interaction depth and historical session duration corresponding to the historical intent vector. Historical session interaction depth refers to the weighted statistical value of the percentage of page scroll pixels triggered and the total number of effective clicked controls within the session period in which the target user generated the historical intent vector, reflecting the user's exploration degree of historical content from both spatial and actional dimensions. Historical session duration refers to the absolute number of milliseconds during which the target user kept the page active in the foreground within the historical session period, reflecting the user's attention investment cost from a temporal dimension. By simultaneously extracting the above two dimensions of indicators, the system constructs a multi-dimensional data barrier for evaluating the effectiveness of historical intents.

[0045] After acquiring multi-dimensional evaluation metrics, the system executes strict conditional judgment logic. The system determines whether the time interval is less than a preset recent time threshold, whether the depth of historical session interaction is less than a preset depth threshold, and whether the duration of the historical session is less than a preset duration threshold. The preset recent time threshold is set between 30 and 60 seconds, based on defining a short time window that is likely to directly interfere with the current operation; the preset depth threshold is set at 20% scrolling or less than 2 valid clicks, based on the minimum action limit for a typical user to complete a preliminary content overview; and the preset duration threshold is set between 3 and 5 seconds, based on industry-recognized bounce rate standards. When all three conditions are met simultaneously, it indicates that although the historical intent occurred within a very recent timeframe, the user neither browsed deeply nor invested sufficient time. Based on this, the system determines the historical intent vector as an accidental interference intent. Accidental interference intent represents invalid characteristics resulting from unconscious accidental touches or rapid bounce behavior by the user. Through the aforementioned stringent joint threshold determination, the system accurately identifies and intercepts underlying noisy data disguised as highly timely data, preventing invalid features from overshadowing the main data in subsequent fusion stages.

[0046] Taking a news and information application browsing scenario as an example. Assume the system calculates a time interval of 10 seconds and inputs a preset time decay function to obtain an initial historical weight of 0.85. Subsequently, the system obtains a historical session interaction depth of 10% and a historical session duration of 2 seconds. The system compares 10 seconds with a preset recent time threshold (e.g., 30 seconds), 10% with a preset depth threshold (e.g., 20%), and 2 seconds with a preset duration threshold (e.g., 5 seconds). Since all three indicators are less than their corresponding thresholds, the system determines that the historical intent vector generated 10 seconds ago belongs to an accidental interference intent where the user accidentally clicked on a news headline and immediately exited. This provides a basis for triggering the subsequent penalty mechanism, preventing the initial historical weight of 0.85 from overshadowing the current true reading intent.

[0047] After determining that the historical intent vector belongs to accidental interfering intent, directly forcing the historical weights to zero can easily disrupt the temporal smoothness of the underlying feature sequence, leading to abrupt numerical changes in the model input. To achieve flexible suppression and precise removal of interfering noise, the system constructs a multi-dimensional dynamic penalty mechanism. Specifically, the system first comprehensively considers the time interval, historical session interaction depth, and historical session duration to calculate the penalty base. Then, it further introduces semantic conflict verification across time windows. By quantifying the divergence between historical and current semantics, it applies exponential decay to interfering terms with heterogeneous semantic noise, thereby calculating the final penalty decay factor. Specifically, this includes: The time interval, the historical session interaction depth, and the historical session duration are normalized and then weighted and summed to obtain a basic penalty factor with a value greater than zero and less than or equal to a preset penalty upper limit, wherein the preset penalty upper limit is less than one. Historical semantic features are extracted from the historical intent vector, and the semantic divergence between the historical semantic features and the semantic feature vector in the current behavior feature vector is calculated. If the semantic divergence is less than or equal to the preset semantic conflict threshold, then the basic penalty factor is used as the penalty decay factor; If the semantic divergence is greater than the preset semantic conflict threshold, then the accidental interference intent and the current intent have heterogeneous semantic noise. The semantic divergence is used as the attenuation index to perform exponential attenuation calculation on the basic penalty factor to obtain the penalty attenuation factor.

[0048] To unify the diverse underlying metrics into a common numerical space for comprehensive evaluation, the system first performs maximum-minimum normalization on the time interval, historical session interaction depth, and historical session duration, linearly mapping each metric to the range of zero to one. Subsequently, the system assigns corresponding weight coefficients to the three normalized metrics and performs a weighted summation. To prevent insufficient penalty from effectively suppressing noise, or excessive penalty from causing weights to drop to zero and disrupting the temporal smoothness of the underlying feature sequence, the system truncates the weighted summation result, obtaining a basic penalty factor that is greater than zero and less than or equal to a preset penalty upper limit. The preset penalty upper limit is set between 0.3 and 0.5, based on the principle that historical weights identified as interference items are reduced by at least 50%, while retaining a slight numerical gradient to maintain the stability of the model's backpropagation.

[0049] Relying solely on behavioral metrics makes it difficult to distinguish whether a target user accidentally clicked on completely unrelated content or was making a quick comparison among similar content. Therefore, the system needs to introduce content-level conflict verification. The system extracts historical semantic features representing historical context from the historical intent vector and uses the cosine distance algorithm to calculate the cosine difference between the historical semantic features and the semantic feature vector in the current behavioral feature vector in a multi-dimensional vector space. The result is defined as semantic divergence. Semantic divergence, as a quantitative indicator measuring the directional difference between two text features in vector space, indicates that the semantic content of the two interactions is less related.

[0050] After obtaining the semantic divergence, the system compares it with a preset semantic conflict threshold. The preset semantic conflict threshold is set between 0.4 and 0.6, based on the empirical cosine distance boundary used in natural language processing to determine whether two short texts are not strongly correlated. If the semantic divergence is less than or equal to the preset semantic conflict threshold, it indicates that although the historical operation exhibited short-term behavioral characteristics, it still has a certain semantic connection with the current operation. The system adopts a conservative penalty strategy, directly using the basic penalty factor as the penalty decay factor. If the semantic divergence is greater than the preset semantic conflict threshold, the system determines that there is heterogeneous semantic noise between the accidental interference intent and the current intent. Heterogeneous semantic noise refers to invalid feature data that completely deviates from the current user's true intent in the semantic space, usually generated by the target user accidentally clicking on completely unrelated page links. For interference items containing heterogeneous semantic noise, the system uses the semantic divergence as the decay exponent of the negative exponent term, performs exponential decay calculation on the basic penalty factor, thereby further significantly compressing the value of the penalty decay factor to obtain the final penalty decay factor.

[0051] The calculation logic of exponential decay is expressed by the following formula: F_penalty=F_base*exp(-D_semantic); in the formula, F_penalty represents the penalty decay factor, F_base represents the base penalty factor, exp represents the exponential function with the natural constant as the base, and D_semantic represents the semantic divergence.

[0052] Taking an e-commerce platform browsing scenario as an example, the system normalizes and weights the extremely short time intervals, shallow historical session interaction depths, and extremely short historical session durations, calculating a basic penalty factor of 0.4. The system extracts the text corresponding to historical semantic features as "smartphone" and the text corresponding to the current semantic feature vector as "phone case." The calculated semantic divergence is 0.2, less than the preset semantic conflict threshold of 0.5. The system determines that the preceding and following operations are related and directly uses 0.4 as the penalty decay factor. In another scenario, if the text corresponding to the historical semantic feature is "fresh fruit" (accidentally touched by the target user while swiping the screen), and the text corresponding to the current semantic feature vector is "phone case," the calculated semantic divergence is as high as 1.2, greater than the preset semantic conflict threshold of 0.5. The system determines the presence of heterogeneous semantic noise and substitutes 1.2 as the decay exponent into the exponential decay formula, calculating a penalty decay factor of approximately 0.12. Through this mechanism, the system applies a more severe weight reduction to completely irrelevant accidental touch noise, ensuring the purity of the current true intent feature.

[0053] S203. Based on the historical intent weight coefficient, calculate the current behavior weight coefficient, wherein the sum of the historical intent weight coefficient and the current behavior weight coefficient is one; After obtaining the historical intent weight coefficients, which have undergone multi-dimensional verification and correction, the system needs to determine the proportion of the current behavior feature vector in the final fused features. Without normalizing the weights, directly adding feature vectors of different magnitudes can easily lead to a drastic expansion of the fused feature value range, causing scale shift in the deep learning model during backpropagation. Scale shift indicates that the numerical distribution of the input data deviates from the sensitive range of the neural network activation function, easily causing gradient vanishing or exploding, leading to the model's inability to converge properly. To maintain the consistency of the numerical distribution before and after feature fusion, the system calculates the current behavior weight coefficient based on the historical intent weight coefficients and strictly constrains the sum of the historical intent weight coefficients and the current behavior weight coefficients to be one. Specifically, the calculation logic is to directly subtract the historical intent weight coefficients from the constant one. Through this rigorous complementary calculation, the system ensures that historical experience and current behavior form a dynamic zero-sum game in the feature space, avoiding uncontrolled feature scale.

[0054] S204. Multiply the current behavior feature vector by the current behavior weight coefficient to obtain the current behavior weighted vector; After establishing a complementary weight allocation scheme, the system performs numerical scaling on the current and historical feature vectors respectively. The system performs a scalar multiplication operation on the current behavior feature vector and the current behavior weight coefficient, that is, multiplies the value of each dimension in the current behavior feature vector by the current behavior weight coefficient to obtain the current behavior weighted vector.

[0055] S205. Multiply the historical intent vector by the historical intent weight coefficient to obtain the historical intent weighted vector; Simultaneously, the system performs the same scalar multiplication operation on the historical intent vector and the historical intent weight coefficient to obtain a historical intent weighted vector. The scalar multiplication operation adjusts the magnitude of the feature vectors according to their timeliness and importance without changing their original direction, thus amplifying the current true intent and effectively suppressing historical interference noise.

[0056] S206. Summing the current behavior weighted vector with the historical intent weighted vector yields the fused feature vector.

[0057] After numerical scaling, the system performs element-wise summation on the current behavior weighted vector and the historical intent weighted vector to obtain a fused feature vector. The element-wise summation linearly superimposes the two feature vectors, which reside in the same dimensional space, along their corresponding dimensions. Through this summation, the system perfectly integrates the continuity features representing historical context with the immediacy features representing the current operation into a unified vector representation. The final output fused feature vector smoothly inherits effective historical semantics while keenly capturing the latest developments in the current behavior, providing high-quality input data with both temporal coherence and a high signal-to-noise ratio for subsequent pre-defined intent judgment models.

[0058] The complete logic of the above weight calculation and feature fusion is represented by the following formulas: W_current = 1 - W_history; V_current_weighted = W_current * V_current; V_history_weighted = W_history * V_history; V_fusion = V_current_weighted + V_history_weighted. In these formulas, W_current represents the current behavior weight coefficient, W_history represents the historical intent weight coefficient, V_current_weighted represents the current behavior weighted vector, V_current represents the current behavior feature vector, V_history_weighted represents the historical intent weighted vector, V_history represents the historical intent vector, and V_fusion represents the fused feature vector.

[0059] Taking an intelligent customer service dialogue scenario as an example. Assume the system, after initial attenuation and penalty calculations, obtains a historical intent weight coefficient of 0.3. The system subtracts 0.3 from 1 to calculate the current behavior weight coefficient as 0.7. Subsequently, the system multiplies all elements of a 256-dimensional current behavior feature vector by 0.7 to obtain a current behavior weighted vector; simultaneously, it multiplies all elements of a 256-dimensional historical intent vector by 0.3 to obtain a historical intent weighted vector. Finally, the system adds the two weighted vectors one by one across all 256 dimensions, outputting a final 256-dimensional fused feature vector. Because the current behavior weight is dominant, the direction of the fused feature vector in multidimensional space will be more biased towards the user's latest question intent, while retaining 30% of the historical dialogue context information. This allows the customer service system to accurately answer the current question without forgetting the key background information mentioned by the user in the previous round.

[0060] S104. Input the fused feature vector into the preset intent judgment model to obtain the probability value of the target user being in each preset intent stage within the current sliding time window, and store it in the preset time-series cache queue. After obtaining the fused feature vector, the system faces the challenge that the fused feature vector is essentially a high-dimensional implicit mathematical expression, lacking intuitive business interpretability. The downstream interactive rendering engine cannot directly execute decisions based on the high-dimensional vector. In order to transform the abstract mathematical features into specific business instructions, the system inputs the fused feature vector into a preset intent judgment model. Through the non-linear mapping capability of deep neural networks, it calculates the probability value of the target user being in each preset intent stage within the current sliding time window.

[0061] The preset intent judgment model employs a multilayer perceptron network architecture combined with a normalized exponential function. The model input receives a fixed-dimensional fused feature vector and contains three fully connected hidden layers. ReLU activation functions are used between these hidden layers to introduce non-linear expressiveness, and a Dropout layer is configured to randomly discard some neuron connections to prevent overfitting. The model output is connected to a Softmax classification layer, and the output dimension strictly corresponds to the number of preset intent stages. The preset intent stages are divided into four discrete business states based on the user's psychological evolution path: no explicit intent stage, information exploration stage, deep comparison stage, and decision conversion stage.

[0062] Before deployment, the pre-defined intent judgment model undergoes rigorous offline training. The system extracts massive amounts of multimodal user behavior data from historical interaction logs and calculates corresponding fusion feature vectors as training sample features according to the aforementioned logic. Simultaneously, based on the user's final conversion result (e.g., whether an order is placed, whether a form is submitted) and page dwell time, the system uses heuristic rules to reverse-engineer and label the true intent stage corresponding to each time window. For example, if a user adds an item to their cart or initiates a payment control within the next 5 seconds, the current time window is reverse-labeled as the "decision conversion stage"; if a user repeatedly scrolls up and down within the same long list page without clicking for more than 10 seconds, it is reverse-labeled as the "information exploration stage." The true intent stage is converted into a one-hot encoding format as a supervision label. During the training phase, the model outputs a predicted probability distribution via forward propagation. The system uses the cross-entropy loss function to calculate the error between the predicted probability distribution and the supervision label, and executes the backpropagation algorithm through the Adam optimizer, iteratively updating the weight matrix of the fully connected layer until the cross-entropy loss value converges to a stable range.

[0063] In the online inference phase, the system inputs the fused feature vectors calculated in real time into the trained preset intent judgment model. The fused feature vectors pass through the hidden layers of the multilayer perceptron for feature dimensionality reduction and abstraction, finally reaching the Softmax classification layer. The Softmax classification layer maps the multidimensional real-valued outputs to a probability distribution with a sum of one, obtaining four probability values ​​for the target user within the current sliding time window: no clear intent stage, information exploration stage, deep comparison stage, and decision conversion stage. The calculation logic of the Softmax classification layer is expressed by the following formula: P_i=exp(Z_i) / Sum(exp(Z_j)); where P_i represents the probability value of being in the i-th preset intent stage, Z_i represents the original output value of the i-th neuron in the model output layer, exp represents the exponential function with the natural constant as the base, and Sum represents the summation of the exponential results over all preset intent stages.

[0064] After obtaining the probability values, relying solely on the probability distribution at a single point in time makes it difficult for the system to capture the dynamic evolution of user intent. Therefore, the system stores the calculated probability values ​​in a preset temporal cache queue. This preset temporal cache queue employs a first-in, first-out (FIFO) circular memory data structure, specifically responsible for temporarily storing the inference results of multiple consecutive sliding time windows in chronological order. The maximum length of the preset temporal cache queue is set to 5 to 10 sliding time windows, a setting designed to cover the typical time span for a regular user to complete short-term decision-making within a single page, ensuring sufficient temporal context while avoiding excessive memory consumption.

[0065] Taking the hotel booking scenario on an online travel platform as an example, the system inputs a 256-dimensional fused feature vector containing user browsing actions and search term semantics into a preset intent judgment model. After forward computation, the model outputs probability values ​​for four preset intent stages: 10% for no clear intent, 20% for information exploration, 60% for in-depth comparison, and 10% for decision conversion. The system packages these four probability values ​​into a data frame and pushes it to the tail of a preset temporal cache queue with a maximum length of 8. If the queue is full, the oldest data frame is automatically pushed out. By continuously storing probability values, the system successfully transforms the instantaneous intent state into a continuous intent evolution trajectory, providing a complete data slice for subsequent calculation of intent drift values.

[0066] S105. Extract the probability values ​​within multiple consecutive adjacent sliding time windows in the time-series cache queue, and calculate the intention drift value based on the numerical fluctuation of the probability values ​​in the time series. Probability values ​​at a single point in time only reflect a static psychological slice and cannot reveal the psychological wavering and hesitation of the target user during the interaction process. If the target user frequently switches between different intent stages, it indicates that they are facing decision-making difficulties and urgently need the system to actively intervene and provide assistance. In order to quantify the instability of the target user's potential demands within a continuous interaction cycle, the system introduces an intent drift value for measurement. The underlying principle of the intent drift value is based on time series analysis. By calculating the geometric distance of the probability distribution vectors at adjacent time points, the abstract psychological wavering is transformed into a specific numerical volatility rate, which is mainly used as a dynamic basis for subsequent judgments on whether to actively trigger interaction decisions.

[0067] In practice, the system extracts probability values ​​from multiple consecutive sliding time windows in chronological order from a preset time-series cache queue. The number of time windows is set to 3 to 5, based on the principle that it can form an effective time series comparison without smoothing out short-term drastic fluctuations due to excessively long periods. The system combines the probability values ​​of each preset intent stage within each sliding time window into a probability distribution vector. Subsequently, the system calculates the Euclidean distance between the probability distribution vectors corresponding to two adjacent sliding time windows to obtain the single-step fluctuation difference. Finally, the system calculates the arithmetic mean of all single-step fluctuation differences and outputs the intent drift value. The calculation logic is expressed by the following formulas: V_diff(t)=sqrt(Sum((P(t,i)-P(t-1,i))^2)); V_drift=Sum(V_diff(t)) / (N-1). In the formula, V_diff(t) represents the single-step fluctuation difference between the t-th sliding time window and the previous sliding time window, P(t,i) represents the probability value of the i-th preset intention stage within the t-th sliding time window, Sum represents the summation operation, sqrt represents the square root operation, V_drift represents the intention drift value, and N represents the total number of extracted sliding time windows.

[0068] Through the above calculations, the system successfully upgraded static intent classification into dynamic psychological trajectory tracking, providing dynamic quantitative indicators for accurately grasping the timing of interactions.

[0069] Taking an e-commerce shopping scenario as an example, the system extracts probability values ​​from three consecutive sliding time windows from a pre-set time-series cache queue. In the first time window, the probability of the information exploration stage is 80%; in the second time window, the probability of the deep comparison stage suddenly increases to 70%; in the third time window, the probability of the information exploration stage drops back to 60%. The system combines these probability values ​​into a vector and calculates the Euclidean distance between adjacent vectors, finally averaging them to obtain an intent drift value of 0.65. A value of 0.65 indicates that the target user is repeatedly jumping between exploration and comparison, in a highly conflicted state, providing strong data support for the system to proactively pop up a customer service window for shopping guides.

[0070] In other alternative embodiments, the intention drift value can also be obtained by calculating the relative entropy (i.e., KL divergence) between probability distribution vectors within adjacent sliding time windows. Using relative entropy to measure the asymmetry of the difference between two probability distributions can also achieve accurate quantification of the degree of intention fluctuation.

[0071] S106. Calculate the similarity between the semantic feature vector and each preset service scenario anchor vector, and determine the service scenario corresponding to the preset service scenario anchor vector with a similarity greater than a preset similarity threshold as the target service scenario. In the preceding steps, the system has already grasped the target user's intent stage and level of psychological conflict. However, the intent stage alone is insufficient to determine the specific business content the system should provide. For example, if the target user is in the "deep comparison stage," the system must clarify whether the user is comparing product parameters or after-sales policies in order to make an effective interaction decision. To accurately converge the target user's freely expressed natural language into a standard business track that the system can process, the system needs to perform semantic-to-scenario mapping and matching.

[0072] Before performing the matching calculation, the system backend pre-encodes the features of historical high-frequency standard business documents and customer service Q&A corpora using a pre-trained language model, extracting high-dimensional vectors representing the semantics of each specific business category center, which are defined as preset service scenario anchor vectors. The principle of introducing preset service scenario anchor vectors is to transform discrete text classification labels into continuous vector space coordinates, mainly used as benchmark targets for subsequent semantic matching.

[0073] In practice, the system calls the cosine similarity algorithm to calculate the cosine of the angle between the semantic feature vector extracted in the current sliding time window and the anchor vectors of each preset service scenario stored in the database in the multidimensional vector space. The cosine similarity algorithm evaluates the semantic closeness by measuring the difference in the directions of two vectors. The calculation logic is expressed by the following formula: Sim(V_sem,V_anchor_k)=(V_sem·V_anchor_k) / (||V_sem||*||V_anchor_k||); In the formula, Sim represents the similarity, V_sem represents the semantic feature vector, V_anchor_k represents the k-th preset service scenario anchor vector, the numerator represents the dot product of the two vectors, and the denominator represents the product of the magnitudes of the two vectors.

[0074] After calculating all similarity scores, the system compares each score with a preset similarity threshold. The preset similarity threshold is set between 0.75 and 0.85, based on the empirical lower bound for determining high semantic consistency between two short texts in natural language processing. This effectively filters out low-level interference from fuzzy matching while tolerating the diversity and non-standardization of natural language expression by target users. The system then selects preset service scenario anchor vectors with similarity scores greater than the preset similarity threshold and maps these anchor vectors to the service scenarios in the database, defining them as the target service scenarios. Through this vector space similarity matching mechanism, the system successfully identifies the business domain that the target user is truly interested in, providing clear business context support for generating specific interactive interface content.

[0075] Take an e-commerce platform's intelligent shopping guide system as an example. A target user enters "How to send back clothes that are too big?" into the search box. The system extracts the corresponding semantic feature vector and calculates the cosine similarity with anchor vectors of preset service scenarios stored in the backend, such as "logistics tracking," "after-sales returns and exchanges," and "invoice issuance application." The calculation results show a similarity of 0.21 with "logistics tracking," 0.05 with "invoice issuance application," and a high similarity of 0.88 with "after-sales returns and exchanges." The system compares 0.88 with the preset similarity threshold of 0.80, determines a successful match, and thus identifies "after-sales returns and exchanges" as the target service scenario. This ensures that the subsequent pop-up interactive interface includes business components for filling in the return address or courier pickup, rather than irrelevant product recommendations.

[0076] S107. Calculate the interaction trigger probability based on the probability value, the intent drift value, and the similarity of each preset intent stage; After determining the target user's intent state, psychological fluctuation level, and target service scenario, the core decision the system faces is whether to proactively initiate interaction with the target user at the current moment. Relying solely on a single-dimensional indicator can easily lead to interaction being too early, interrupting the user's normal browsing, or too late, missing conversion opportunities. To achieve precise control over interaction timing, the system needs to comprehensively quantify and evaluate the probability values ​​of each preset intent stage (representing the clarity of intent), the intent drift value (representing psychological fluctuation), and the similarity (representing the accuracy of business matching), constructing a unified interaction trigger probability as the final decision-making criterion. By introducing a multi-dimensional joint calculation mechanism, the system can dynamically apply corresponding exponential suppression or linear penalty logic for different situations, such as when the user is in a high-frequency oscillating decision-making state or a normal browsing state, thereby calculating the interaction trigger probability that matches the current actual interaction needs. The specific calculation process includes: The maximum probability value is extracted from the probability values ​​of each preset intent stage and used as the target intent score, and the similarity is used as the scene matching score; Based on preset intent weights and preset scene weights, the target intent score and the scene matching score are weighted and summed to obtain the initial trigger probability; If the target intent score is greater than a preset high intent threshold and the intent drift value is greater than a preset oscillation drift threshold, then the target user is determined to be in a high-frequency oscillation decision state. When the target user is in the high-frequency oscillation decision state, a preset oscillation suppression floor is obtained, and the intention drift value is used as the exponent of the preset oscillation suppression floor to calculate a suppression coefficient that is dynamically scaled between zero and one. The design value of the preset oscillation suppression floor is greater than zero and less than one. Multiplying the initial trigger probability by the suppression coefficient yields the interaction trigger probability; If the target intent score is less than or equal to the preset high intent threshold, or the intent drift value is less than or equal to the preset oscillation drift threshold, then the intent drift value is multiplied by the preset drift weight to obtain a drift penalty term, and the initial trigger probability is subtracted from the drift penalty term to obtain the interaction trigger probability.

[0077] To establish a basic trigger evaluation benchmark, the system first needs to extract dominant factors from the multi-dimensional intent distribution. The system extracts the highest probability value from the probability values ​​of each preset intent stage calculated in the preceding steps, defining this highest probability value as the target intent score. Simultaneously, the system directly uses the similarity between the semantic feature vector calculated in the preceding steps and the preset service scenario anchor vector corresponding to the target service scenario as the scenario matching score. Subsequently, the system introduces preset intent weights and preset scenario weights, performing a weighted summation of the target intent score and the scenario matching score to obtain the initial trigger probability. The sum of the preset intent weights and preset scenario weights is strictly limited to one, typically set to 0.6 and 0.4 respectively. This is based on the fact that the clarity of intent dominates interaction decisions, while scenario matching accuracy serves as a secondary verification condition. Through weighted summation, the system integrates the intensity of intent and the matching confidence of the business scenario into a single scalar value, providing a basic probability level for subsequent dynamic adjustments.

[0078] After obtaining the initial trigger probability, the system needs to classify the context by combining the fluctuation characteristics of the time series to cope with the complex and ever-changing psychological states of users. The system determines whether the target intent score is greater than a preset high intent threshold and simultaneously determines whether the intent drift value is greater than a preset oscillation drift threshold. The preset high intent threshold is set to a range of 0.75 to 0.85, based on distinguishing whether the target user has a clear and strong potential demand; the preset oscillation drift threshold is set to a range of 0.4 to 0.6, based on defining the numerical boundary of a sharp reversal in the target user's psychological state within adjacent sliding time windows. When both conditions are met simultaneously, the system determines that the target user is in a high-frequency oscillating decision-making state. The high-frequency oscillating decision-making state represents the target user's rapid and repeated comparison and switching behavior between multiple similar options or decision branches while having a strong conversion intention. The underlying logic stems from the cognitive overload of the target user when facing information overload, and is mainly used to trigger the subsequent nonlinear inhibition mechanism to prevent the system from abruptly disturbing the target user when the target user is highly focused on comparison.

[0079] When the target user is in a state of high-frequency oscillation decision-making, the system adopts an exponentially flexible withdrawal strategy. The system obtains a pre-configured preset oscillation suppression floor, the design value of which is strictly limited to a range greater than zero and less than one, typically between 0.5 and 0.7. This setting is based on ensuring that the floor produces a convergent decay effect during exponential calculation. The system uses the intent drift value as the exponent of the preset oscillation suppression floor and performs a power operation to calculate a suppression coefficient that dynamically scales between zero and one. The larger the intent drift value, the closer the calculated suppression coefficient is to zero. Subsequently, the system multiplies the initial trigger probability with the suppression coefficient to obtain the final interaction trigger probability. By introducing an exponential suppression mechanism, the system can non-linearly and significantly reduce the trigger probability according to the deepening of the target user's entanglement, thereby delaying the interaction timing until the target user has completed the comparison and the fluctuations have subsided.

[0080] In another typical scenario, if the target intent score is less than or equal to a preset high intent threshold, or the intent drift value is less than or equal to a preset oscillation drift threshold, it indicates that the target user is in a normal, stable browsing state or a low-intent exploration state. For this typical scenario, the system adopts a linear, mild penalty strategy. The system multiplies the intent drift value by a preset drift weight to obtain a drift penalty item. The preset drift weight is set to a range of 0.1 to 0.2, designed to ensure the penalty magnitude remains within a controllable fine-tuning range, avoiding excessive changes to the initial trigger probability. Finally, the system performs a subtraction operation, subtracting the drift penalty item from the initial trigger probability to obtain the final interaction trigger probability. Through this linear penalty logic, the system transforms minor psychological fluctuations into a moderate reduction in the trigger probability, ensuring that the system maintains appropriate restraint when the target user's intent is not yet fully stable.

[0081] The logic for calculating the interaction trigger probability in the above two scenarios is expressed by the following formula: P_initial=W_intent*S_intent+W_scene*S_scene; #Scenario 1: High-frequency oscillation decision-making state C_suppress=Base_suppress^V_drift; P_trigger_1=P_initial*C_suppress.

[0082] #Scenario 2: Normal Stable State Penalty_drift=V_drift*W_drift; P_trigger_2=P_initial-Penalty_drift.

[0083] In the formula, P_initial represents the initial trigger probability, W_intent represents the preset intent weight, S_intent represents the target intent score, W_scene represents the preset scene weight, S_scene represents the scene matching score, C_suppress represents the suppression coefficient, Base_suppress represents the preset oscillation suppression base, V_drift represents the intent drift value, P_trigger_1 and P_trigger_2 represent the interaction trigger probabilities in the two scenarios, Penalty_drift represents the drift penalty term, and W_drift represents the preset drift weight.

[0084] Taking online ticketing as an example, the system extracts the highest probability that the target user is in the "decision conversion stage," with a value of 0.85, as the target intent score; the scenario matching score is 0.90. The system performs a weighted summation with weights of 0.6 and 0.4, resulting in an initial trigger probability of 0.87. Subsequently, the system detects an intent drift value of 0.55. The system compares 0.85 with a preset high intent threshold of 0.80 and 0.55 with a preset oscillation drift threshold of 0.50, determining that the target user is in a high-frequency oscillating decision-making state (e.g., the target user repeatedly switches between two flights at similar times to check refund and change rules). The system obtains a preset oscillation suppression base of 0.6, and calculates 0.6 raised to the power of 0.55 to obtain a suppression coefficient of approximately 0.75. The system multiplies 0.87 by 0.75, ultimately outputting an interaction trigger probability of 0.6525. Because the probability is dynamically suppressed, the system does not pop up a pop-up urging the user to place an order, avoiding interrupting the target user's comparison and thinking process.

[0085] Optionally, when the interaction trigger probability is greater than a preset trigger threshold, before generating the interaction decision scheme for the target user based on the target service scenario and the probability values ​​of each preset intent stage, the method further includes: obtaining the historical touch timestamp of the last triggered interaction of the target user, and calculating the actual interval between the historical touch timestamp and the current time; extracting the target intent stage corresponding to the maximum probability value from the probability values ​​of each preset intent stage; obtaining a preset basic cooling time, and matching a corresponding dynamic cooling coefficient according to the target intent stage, wherein the closer the target intent stage is to the decision conversion stage, the smaller the dynamic cooling coefficient; multiplying the preset basic cooling time by the dynamic cooling coefficient to obtain an adaptive cooling threshold; if the actual interval time is greater than the adaptive cooling threshold, then allowing the generation instruction of the interaction decision scheme; if the actual interval time is less than or equal to the adaptive cooling threshold, then intercepting the generation instruction of the interaction decision scheme.

[0086] After calculating the probability of interaction triggering and determining that it exceeds a preset triggering threshold, if the system immediately and unconditionally executes interaction rendering, it is highly likely to bombard the same user with frequent visuals in a short period of time, causing user resentment and churn. In order to strike a balance between providing intelligent assistance and protecting user experience, the system introduces an adaptive cooling mechanism based on the intent stage before generating interaction decision schemes.

[0087] Specifically, the system first obtains the historical timestamp of the last interaction with the target user and calculates the actual interval between that time and the current time. Then, the system extracts the target intent stage corresponding to the highest probability value from the probability values ​​of each preset intent stage. The system backend pre-configures a preset base cooldown time (e.g., 60 seconds) and assigns a dynamically decreasing cooldown coefficient to different preset intent stages. The rationale is as follows: when the user is in the "no clear intent stage" or "information exploration stage," their tolerance for disturbance is extremely low, requiring a longer period of non-disturbance; therefore, the dynamic cooldown coefficient is set to a value greater than one (e.g., 2.0 or 1.5). Conversely, when the user is in the "deep comparison stage" or "decision conversion stage," their thirst for effective information increases dramatically; at this time, the cooldown period should be shortened to prevent missing conversion opportunities; therefore, the dynamic cooldown coefficient is set to a value less than one (e.g., 0.7 or 0.5).

[0088] The system multiplies the preset base cooling duration by the matched dynamic cooling coefficient to calculate the adaptive cooling threshold. Finally, the system compares the actual interval duration with the adaptive cooling threshold. Only when the actual interval duration exceeds the adaptive cooling threshold will the system allow the generation command of the interactive decision scheme; otherwise, the system will silently intercept the trigger at the underlying level, thereby achieving precise avoidance of high-frequency invalid disturbances.

[0089] S108. When the interaction trigger probability is greater than the preset trigger threshold, an interaction decision scheme for the target user is generated based on the target service scenario and the probability values ​​of each preset intent stage, and the corresponding interaction interface is rendered on the client according to the interaction decision scheme, wherein the transparency rendering parameter of the interaction interface is dynamically updated gradually according to a preset periodic function.

[0090] After calculating the final interaction trigger probability and determining that it exceeds the preset trigger threshold, the system confirms that the current moment is the optimal time for intervention and assistance. However, traditional static pop-ups or abrupt interface transitions can easily disrupt the user's immersive browsing experience and trigger resistance. To achieve seamless and restrained visual guidance, the system abandons the fixed-parameter interface rendering mode and instead constructs a dynamic visual feedback mechanism that adapts to the user's current cognitive load and the clarity of their intent. Specifically, the system not only extracts the specific business content corresponding to the target service scenario but also deeply mines the distribution characteristics contained in the probability values ​​of each preset intent stage. It transforms the distribution index representing the degree of intent confusion into a temporal parameter controlling the interface's gradual change cycle and transforms the extreme values ​​representing intent intensity into visual parameters controlling the interface's visibility. This generates a highly personalized interaction decision scheme and drives the client's underlying rendering engine to present the interactive interface in a gradual, dynamic update manner similar to "breathing." Specifically, this includes: Calculate the information entropy of the probability values ​​for each of the preset intention stages; The information entropy is input into a preset periodic mapping function to obtain a dynamic respiratory cycle, wherein the information entropy is positively correlated with the dynamic respiratory cycle; Extract the maximum probability value from the probability values ​​of each preset intention stage, and input the maximum probability value into a preset transparency mapping function to calculate the upper limit of the peak of the gradient transparency; Extract the scene service content corresponding to the target service scene; The dynamic breathing cycle, the peak upper limit, and the scenario service content are combined to generate the interactive decision-making scheme; In response to the interactive decision scheme, the dynamic graphics rendering interface of the client is invoked. By configuring the period parameter of the corresponding preset periodic function in the dynamic rendering interface as the dynamic breathing cycle and configuring the amplitude parameter as the peak value limit, the redrawing frequency of the corresponding rendering thread is controlled by the dynamic breathing cycle. The interactive interface containing the scene service content is rendered, and the transparency rendering parameters of the interactive interface are driven to be dynamically updated in a gradual manner.

[0091] To accurately measure the degree of confusion in the target user's current psychological state, the system first calculates the information entropy of the probability values ​​for each preset intention stage. Information entropy, a classic mathematical tool in information theory for measuring system uncertainty, objectively reflects the discrete state of a multidimensional probability distribution. The specific calculation logic is expressed by the following formula: H = -Sum(P_i * log2(P_i)); where H represents the information entropy, P_i represents the probability value of the i-th preset intention stage, Sum represents the summation of the calculation results for all preset intention stages, and log2 represents the logarithmic function to the base 2. The more uniform the probability distribution of each intention stage, the larger the calculated information entropy value, indicating that the target user is in a state of high confusion and extremely unclear potential needs.

[0092] After obtaining the information entropy, the system inputs it into a preset periodic mapping function to calculate the dynamic breathing cycle. The dynamic breathing cycle is defined as the time span required for the interactive interface's transparency to complete a full gradient from dark to light and then back to dark. The preset periodic mapping function uses a direct proportional linear mapping logic to ensure a strict positive correlation between information entropy and the dynamic breathing cycle. Specifically, the higher the information entropy, the longer the mapped dynamic breathing cycle. From a cognitive psychology perspective of human-computer interaction, when the target user's intentions are wavering and the information entropy is high, the brain is under a high cognitive load. At this time, using short-cycle, high-frequency flashing visual cues would severely interfere with the user's visual focus, exacerbating visual fatigue and psychological anxiety. By extending the dynamic breathing cycle, the system presents a slow, soothing gradient guidance effect at the visual level, providing users with ample time to process information. Simultaneously, at the underlying physical performance level, a longer dynamic breathing cycle directly corresponds to a lower rendering thread redraw frequency, effectively reducing the power consumption and memory usage of the mobile device's graphics processor and avoiding interface lag caused by high-frequency redrawing.

[0093] After determining the gradation rhythm in the time dimension, the system further determines the visual salience boundary in the spatial dimension. The system extracts the highest probability value from the probability values ​​of each preset intent stage and inputs this maximum probability value into a preset transparency mapping function to calculate the upper limit of the gradation transparency peak. The preset transparency mapping function also employs positive correlation mapping logic: the higher the maximum probability value, the stronger the dominant intent, and the higher the calculated upper limit of the peak, allowing the interactive interface to more clearly capture the user's attention when gradation reaches its brightest point. Conversely, if the dominant intent is weak, the upper limit of the peak is lowered, keeping the interactive interface in a semi-transparent, low-key state to avoid overshadowing the main visual element.

[0094] Subsequently, based on the target service scenario determined in the preceding steps, the system extracts the corresponding scenario service content from the business database. This scenario service content includes specific business text, operation buttons, and jump links. The system packages the calculated dynamic breathing cycle, peak upper limit, and extracted scenario service content into a standard data structure to generate a complete interaction decision scheme. In addition to the default breathing-style overlay, the interface presentation in the interaction decision scheme can dynamically switch to an intelligent dialogue invitation window, a domain expert business card push card, or a highlighted guidance overlay of the current page's core content, depending on the target intent score range, to match different levels of user demand.

[0095] Finally, the system sends the interactive decision-making scheme to the client. In response to the received interactive decision-making scheme, the system calls the client's underlying dynamic graphics rendering interface. The system configures the period parameter of the preset periodic function (usually a sine function to ensure the continuity of the numerical derivative) built into the dynamic graphics rendering interface as a dynamic breathing period, and then... The amplitude parameter is configured to the peak limit. After configuration, the system uses the dynamic breathing cycle as a benchmark to control the redraw frequency of the corresponding rendering thread, rendering the interactive interface containing scene service content in the specified area of ​​the screen. During rendering, the system drives the transparency rendering parameter of the interactive interface to gradually and dynamically update according to the continuous values ​​output by the preset periodic function, thus presenting a visual interactive effect on the screen like a steady breathing motion.

[0096] Taking the intelligent customer service scenario of a financial management application as an example, the system calculates that the probability distribution of the four intent stages is extremely even, resulting in an information entropy as high as 1.8. The system inputs 1.8 into a preset periodic mapping function, calculating a dynamic breathing cycle of 4 seconds. At the same time, the extracted maximum probability value is only 0.3, and inputting it into a preset transparency mapping function yields a peak upper limit of 40%. The system extracts "Financial Product Comparison Assistant" as the scenario service content, generates an interactive decision-making scheme, and sends it out. After receiving the scheme, the client calls the rendering interface to slowly and faintly display the floating window of "Financial Product Comparison Assistant" in the lower right corner of the screen with a period of 4 seconds and a maximum transparency of 40%. Due to the extremely slow fading and low maximum transparency, the floating window will not glaringly interrupt the user's current in-depth reading of the complex yield table, while providing a gentle auxiliary entry point at the edge of the user's line of sight, achieving uninterrupted intelligent guidance.

[0097] Optionally, after rendering the corresponding interactive interface on the client according to the interactive decision scheme, the method further includes: listening to the feedback behavior signals of the target user on the interactive interface, the feedback behavior signals including ignore signals, click signals, and conversion signals; configuring corresponding preset reward weights for the ignore signals, click signals, and conversion signals respectively, and calculating the interaction benefit value based on the feedback behavior signals and the preset reward weights; updating the network parameters of the preset intent judgment model through backpropagation using a policy gradient algorithm based on the interaction benefit value; calculating the actual interaction conversion rate within a preset evaluation period, calculating the difference between the actual interaction conversion rate and the preset target conversion rate, and dynamically adjusting the preset trigger threshold based on the difference using a preset step size.

[0098] Furthermore, to enable the breathing-style intelligent outreach decision-making system to continuously learn and adaptively evolve, and to avoid strategy aging due to long-term operation, the system continuously monitors the target user's feedback behavior signals on the interactive interface after rendering. These feedback behavior signals are specifically categorized into ignore signals (such as closing a pop-up or timeout without action), click signals (such as clicking to view details), and conversion signals (such as submitting a form or completing payment).

[0099] The system assigns corresponding preset reward weights to the three types of signals. For example, the ignore signal corresponds to a negative weight (e.g., -1.0), the click signal corresponds to a medium positive weight (e.g., +1.0), and the conversion signal corresponds to a high positive weight (e.g., +5.0). The weight settings strictly correspond to the actual contribution value of different feedback behaviors to the business transaction. Based on the actual captured feedback behavior signals and their corresponding preset reward weights, the system calculates the interaction reward value for a single interaction. This interaction reward value, as a scalar value quantifying the effect of a single interaction, is directly used as the reward feedback in reinforcement learning. Subsequently, the system calls the policy gradient algorithm, using the interaction reward value as input to the reward function, to backpropagate and update the parameters of the fully connected layer network at the bottom of the preset intent judgment model. This makes the model more inclined to output the probability distribution of intents that yield high rewards in future predictions.

[0100] Simultaneously, the system performs adaptive threshold adjustments at the macro-strategy level. The system periodically calculates the actual interaction conversion rate within a preset evaluation period (e.g., 24 hours) and the difference between it and the preset target conversion rate. If the actual interaction conversion rate is lower than the preset target conversion rate, it indicates that the current system triggers too frequently and is not accurate enough. The system will then increase the preset trigger threshold by a preset step size (e.g., 0.05). Conversely, if the actual interaction conversion rate is much higher than the preset target conversion rate, it indicates that the system is too conservative and may be missing potential customers. The system will then decrease the preset trigger threshold by the same preset step size. The preset step size of 0.05 is set to ensure the smoothness of threshold adjustments and avoid drastic policy fluctuations. Through this dual-track self-evolutionary mechanism of micro-parameter updates and macro-threshold adjustments, the system achieves closed-loop optimization of the interaction decision-making strategy.

[0101] Please see Figure 3 This is a schematic diagram of the structure of a breathing-type intelligent outreach decision system based on an intent judgment model in an embodiment of this application.

[0102] It should be noted that, Figure 3 The structure of the breathing-style intelligent outreach decision system based on an intent judgment model shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0103] like Figure 3 As shown, a breathing-style intelligent outreach decision-making system based on an intent judgment model includes a central processing unit 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory 302 or a program loaded from a storage section 308 into a random access memory 303, such as executing the methods described in the above embodiments. The random access memory 303 also stores various programs and data required for system operation. The central processing unit 301, the read-only memory 302, and the random access memory 303 are interconnected via a bus 304. An input / output interface 305 is also connected to the bus 304.

[0104] The following components are connected to the input / output interface 305: an input section 306 including audio input devices, push-button switches, etc.; an output section 307 including an LCD display, audio output devices, indicator lights, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the input / output interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 310 as needed so that computer programs read from it can be installed into the storage section 308 as needed.

[0105] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit 301, it performs the various functions defined in the present invention.

[0106] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, flash memory, optical fiber, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0107] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0108] Specifically, the breathing-based intelligent outreach decision system based on an intent judgment model in this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the breathing-based intelligent outreach decision method based on an intent judgment model provided in the above embodiment.

[0109] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the breathing-based intelligent outreach decision system based on the intent judgment model described in the above embodiments; or it may exist independently and not assembled into the breathing-based intelligent outreach decision system based on the intent judgment model. The storage medium carries one or more computer programs, which, when executed by a processor of the breathing-based intelligent outreach decision system based on the intent judgment model, cause the breathing-based intelligent outreach decision system based on the intent judgment model to implement the breathing-based intelligent outreach decision method based on the intent judgment model provided in the above embodiments.

Claims

1. A breathing-style intelligent outreach decision-making method based on an intent judgment model, characterized in that, The method includes: The target user's multimodal behavior signals in the current session are periodically acquired according to a preset sliding time window. The multimodal behavior signals include interactive behavior signals and text semantic signals. Semantic feature vectors are extracted from the text semantic signals, and the interaction behavior signals are concatenated and encoded with the semantic feature vectors to obtain the current behavior feature vector of the current sliding time window; Obtain the historical intent vector of the target user within a preset historical time period, and then perform a weighted fusion of the historical intent vector and the current behavior feature vector to obtain a fused feature vector. The fused feature vector is input into a preset intent judgment model to obtain the probability value of the target user being in each preset intent stage within the current sliding time window, and stored in a preset time-series cache queue; Extract the probability values ​​within multiple consecutive sliding time windows in the time-series cache queue, and calculate the intention drift value based on the temporal fluctuation of the probability values; Calculate the similarity between the semantic feature vector and each preset service scenario anchor vector, and determine the service scenario corresponding to the preset service scenario anchor vector whose similarity is greater than a preset similarity threshold as the target service scenario; The probability of interaction triggering is calculated based on the probability value, the intent drift value, and the similarity of each preset intent stage. When the probability of triggering the interaction is greater than the preset trigger threshold, an interaction decision scheme for the target user is generated based on the target service scenario and the probability values ​​of each preset intent stage, and the corresponding interactive interface is rendered on the client according to the interaction decision scheme, wherein the transparency rendering parameter of the interactive interface is dynamically updated gradually according to a preset periodic function.

2. The method according to claim 1, characterized in that, The step of extracting semantic feature vectors from the text semantic signal and concatenating and encoding the interaction behavior signal with the semantic feature vectors to obtain the current behavior feature vector of the current sliding time window specifically includes: Extract the back number, idle time, and tab switching number corresponding to the current swipe time window from the interaction behavior signal, and use a preset activation function to weight and combine the back number, idle time, and tab switching number to obtain the hesitation behavior pattern features; Obtain the page dwell time, scroll depth, and revisit frequency within the current swipe time window from the interaction behavior signal; The page dwell time, scroll depth, revisit frequency, and hesitation behavior pattern features are concatenated with the semantic feature vector, and then linearly transformed using a preset encoding weight matrix and bias term to obtain the current behavior feature vector.

3. The method according to claim 1, characterized in that, The step of weightedly fusing the historical intent vector with the current behavior feature vector to obtain a fused feature vector specifically includes: Obtain the historical update time corresponding to the historical intent vector, and calculate the time interval between the historical update time and the current time; Based on the time interval, the historical intent weight coefficient is dynamically calculated using a preset time decay function; Based on the historical intent weight coefficient, the current behavior weight coefficient is calculated, and the sum of the historical intent weight coefficient and the current behavior weight coefficient is one. Multiply the current behavior feature vector by the current behavior weight coefficient to obtain the current behavior weighted vector; Multiply the historical intent vector by the historical intent weight coefficient to obtain the historical intent weighted vector; The current behavior weighted vector and the historical intent weighted vector are summed to obtain the fused feature vector.

4. The method according to claim 3, characterized in that, The step of dynamically calculating the historical intent weight coefficient based on the time interval using a preset time decay function specifically includes: The time interval is input into the preset time decay function to obtain the initial historical weight; Obtain the historical session interaction depth and historical session duration corresponding to the historical intent vector; If the time interval is less than a preset recent time threshold, and the historical session interaction depth is less than a preset depth threshold, and the historical session duration is less than a preset duration threshold, then the historical intent vector is determined to be an accidental interference intent, and the penalty decay factor is calculated based on the time interval, the historical session interaction depth, and the historical session duration. Multiply the initial historical weight by the penalty decay factor to obtain the historical intention weight coefficient; If the time interval is greater than or equal to the preset recent time threshold, or the historical session interaction depth is greater than or equal to the preset depth threshold, or the historical session duration is greater than or equal to the preset duration threshold, then the initial historical weight is used as the historical intent weight coefficient.

5. The method according to claim 4, characterized in that, The calculation of the penalty decay factor based on the time interval, the historical session interaction depth, and the historical session duration specifically includes: The time interval, the historical session interaction depth, and the historical session duration are normalized and then weighted and summed to obtain a basic penalty factor with a value greater than zero and less than or equal to a preset penalty upper limit, wherein the preset penalty upper limit is less than one. Historical semantic features are extracted from the historical intent vector, and the semantic divergence between the historical semantic features and the semantic feature vector in the current behavior feature vector is calculated. If the semantic divergence is less than or equal to the preset semantic conflict threshold, then the basic penalty factor is used as the penalty decay factor; If the semantic divergence is greater than the preset semantic conflict threshold, then the accidental interference intent and the current intent have heterogeneous semantic noise. The semantic divergence is used as the attenuation index to perform exponential attenuation calculation on the basic penalty factor to obtain the penalty attenuation factor.

6. The method according to claim 1, characterized in that, The calculation of the interaction trigger probability based on the probability value, intent drift value, and similarity of each preset intent stage specifically includes: The maximum probability value is extracted from the probability values ​​of each preset intent stage and used as the target intent score, and the similarity is used as the scene matching score; Based on preset intent weights and preset scene weights, the target intent score and the scene matching score are weighted and summed to obtain the initial trigger probability; If the target intent score is greater than a preset high intent threshold and the intent drift value is greater than a preset oscillation drift threshold, then the target user is determined to be in a high-frequency oscillation decision state. When the target user is in the high-frequency oscillation decision state, a preset oscillation suppression floor is obtained, and the intention drift value is used as the exponent of the preset oscillation suppression floor to calculate a suppression coefficient that is dynamically scaled between zero and one. The design value of the preset oscillation suppression floor is greater than zero and less than one. Multiplying the initial trigger probability by the suppression coefficient yields the interaction trigger probability; If the target intent score is less than or equal to the preset high intent threshold, or the intent drift value is less than or equal to the preset oscillation drift threshold, then the intent drift value is multiplied by the preset drift weight to obtain a drift penalty term, and the initial trigger probability is subtracted from the drift penalty term to obtain the interaction trigger probability.

7. The method according to claim 1, characterized in that, The step of generating an interaction decision scheme for the target user based on the target service scenario and the probability values ​​of each preset intent stage, and rendering the corresponding interactive interface on the client according to the interaction decision scheme, specifically includes: Calculate the information entropy of the probability values ​​for each of the preset intention stages; The information entropy is input into a preset periodic mapping function to obtain a dynamic respiratory cycle, wherein the information entropy is positively correlated with the dynamic respiratory cycle; Extract the maximum probability value from the probability values ​​of each preset intention stage, and input the maximum probability value into a preset transparency mapping function to calculate the upper limit of the peak of the gradient transparency; Extract the scene service content corresponding to the target service scene; The dynamic breathing cycle, the peak upper limit, and the scenario service content are combined to generate the interactive decision-making scheme; In response to the interactive decision scheme, the dynamic graphics rendering interface of the client is invoked. By configuring the period parameter of the corresponding preset periodic function in the dynamic rendering interface as the dynamic breathing cycle and configuring the amplitude parameter as the peak value limit, the redrawing frequency of the corresponding rendering thread is controlled by the dynamic breathing cycle. The interactive interface containing the scene service content is rendered, and the transparency rendering parameters of the interactive interface are driven to be dynamically updated in a gradual manner.

8. A breathing-style intelligent outreach decision-making system based on an intent judgment model, characterized in that, The breathing-based intelligent outreach decision system based on the intent judgment model includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the breathing-based intelligent outreach decision system based on the intent judgment model to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the breathing-based intelligent outreach decision system based on the intent judgment model, the breathing-based intelligent outreach decision system based on the intent judgment model performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the breathing-based intelligent outreach decision system based on the intent judgment model, the breathing-based intelligent outreach decision system based on the intent judgment model performs the method as described in any one of claims 1-7.

Images