An artificial intelligence-based virtual interaction method

Abstract

The application relates to the technical field of artificial intelligence interaction and intelligent marketing, and discloses a virtual interaction method based on artificial intelligence, which comprises the following steps: generating a non-explicit demand vector by statistically processing and standardizing the length of pauses, the topic switching frequency and the keyword appearance frequency in user conversation; calculating the user trust degree by combining real-time emotional scoring and information gain weighting; extracting keywords from candidate content and calculating the correlation degree, and generating a quality score by weighting the trust degree; optimizing the strategy execution weight based on the benefit-cost ratio; generating a personalized touch plan by combining channel preference and time activity; and performing normalization and exponential smoothing update on indexes such as click rate and conversion rate, and feeding back the parameters to the demand identification step to realize closed-loop adaptive optimization.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence interaction and intelligent marketing technology, specifically to a virtual interaction method based on artificial intelligence. Background Technology

[0002] With the continuous improvement of artificial intelligence technology and natural language processing capabilities, virtual interaction methods have been widely applied in scenarios such as online customer service, intelligent shopping guides, and marketing outreach. Current mainstream personalized marketing systems typically build user profiles based on users' historical behavior records and explicit feedback (such as clicks, browsing, and purchases), and rely on collaborative filtering, matrix factorization, or deep learning models for recommendations. However, these methods suffer from insufficient capture of users' potential intentions during dialogue and cannot respond promptly to subtle changes in user interest during real-time conversations. Furthermore, most systems use static training weights or preset rules for user trust or interest models, lacking a dynamic adjustment mechanism based on real-time emotional changes, leading to "content cold start" or "content mismatch" phenomena within recommendations. In terms of content generation... In terms of content, traditional approaches often rely on templates or keyword matching, failing to deeply connect content with context, resulting in low relevance between generated information and user concerns and marketing goals. Furthermore, existing strategy execution modules typically use fixed frequency or experience-based coefficient allocation methods, making it impossible to achieve optimal "benefit / cost ratio" allocation in complex environments with multiple strategies, channels, and time periods. Regarding channel reach scheduling, most solutions only push based on historical channel open rates or fixed time windows, ignoring current user activity and trust levels, often leading to inaccurate push timing and channel selection. Finally, the closed-loop evaluation process often only offline summarizes core metrics such as click-through rate and conversion rate, lacking real-time normalization and adaptive parameter updates, preventing the system from automatically adjusting thresholds and weights for each stage based on the latest feedback. To this end, this invention proposes an integrated virtual interaction method that can capture users' non-explicit needs in real-time dialogue, dynamically model user trust, deeply integrate contextual relevance, optimize the execution weight of multiple strategies, and formulate outreach plans in combination with channel preferences and time activity, while also having the ability to normalize effects and update parameters in a closed loop. Summary of the Invention

[0003] This invention provides a virtual interaction method based on artificial intelligence, which helps to solve the problems mentioned in the background art.

[0004] This invention provides the following technical solution: a virtual interaction method based on artificial intelligence, comprising: S1. Statistical analysis and standardization of pause duration, topic switching frequency, and keyword frequency in user dialogues are performed to generate a vector of non-explicit user needs. S2. Based on the aforementioned non-explicit demand vector and real-time sentiment tendency, information gain weighting and processing by a mapping function are used to obtain the user trust score; S3. Extract candidate content and keywords from the dialogue context, calculate their relevance, and combine them with user trust to calculate the content quality score. S4. Evaluate the benefit-cost ratio of each candidate strategy based on content quality scores and trust levels, and allocate strategy execution weights proportionally. S5. Analyze click preferences and user activity levels across different channels and time periods, combine these with user trust levels to generate outreach priorities, and then push notifications according to these priorities. S6. Normalize the click-through rate, conversion rate and other indicators, and use exponential smoothing to update the parameters, and feed the update results back to step S1.

[0005] Optionally, the statistical and standardized processing of pause duration, topic switching frequency, and keyword frequency in user dialogues to generate a user's non-explicit demand vector includes: Three types of behavioral indicators are collected in real time during the dialogue between the user and the virtual character: including pause duration. Unit: seconds; Topic switching frequency Unit: times / minute; Statistical by category Number of times similar keywords appear ; Obtain historical dialogue logs for three types of behavioral indicators, and calculate the mean for each behavioral indicator. and standard deviation ,in ; The index value collected at the current moment Mapped to standard scores: ; in: : No. The original behavioral indicator values ​​include pause duration. Frequency of topic switching Number of times keywords appear ; , :index Historical sample mean and standard deviation; Standardized demand components; Obtain the dimensionless demand intensity vector ; The confidence level coefficient is determined based on the 95th percentile of the normal distribution. ; Set the first The demand trigger threshold for each original behavioral indicator value : ; For each raw behavioral indicator value Its corresponding demand trigger threshold Compare: like If the value is significant, then the behavioral indicator is considered a significant demand; otherwise, it is considered to have no significant demand. The original behavioral indicator values ​​corresponding to all identified significant needs Keep the rest and discard the rest, outputting a non-explicit demand vector consisting of significant demand components. .

[0006] Optionally, the step of obtaining a user trust score based on the non-explicit demand vector and real-time sentiment tendency, using information gain weighting and processing through a mapping function, includes: The emotional lexicon scoring method is used to score the emotional content of the dialogue between users and virtual characters, and to obtain the user's emotional score corresponding to the real-time emotional tendency. ; Based on the non-explicit demand vector Each demand vector component With Emotional Score As a feature set ; Calculate the user feedback of each feature in the feature set. The contribution was determined and normalized to obtain the weights. : ; in: , ;No. or the Each input feature includes And emotional score ; User positive feedback tags; :feature right The information gain is calculated by discretizing the historical dialogue. Calculate user trust level : ; in: : Sigmoid function.

[0007] Optionally, the step of extracting candidate content and dialogue context keywords, calculating their relevance, and weighting the content quality score by combining user trust levels includes: Extracting the keyword set of candidate content segments With dialogue context keyword set ; Calculate the overlap rate : ; Calculate content quality score : ; in: User trust level; Trust weight, with a value of [0,1], is obtained through a grid search of ROI (Return on Investment) based on historical A / B testing.

[0008] Optionally, the step of comprehensively evaluating the benefit-cost ratio of each candidate strategy based on content quality scores and trust levels, and allocating strategy execution weights proportionally, includes: For each candidate strategy Calculate expected returns : ; in: Candidate Strategy Corresponding content quality score; : The obtained candidate strategies Execution costs; Normalize all expected returns: ; in: Candidate Strategy Execution weight; according to The strategies are executed sequentially from largest to smallest.

[0009] Optionally, the press The strategies are executed sequentially from largest to smallest, including: Let the scheduling cycle length be... A time slice; For each candidate strategy Calculate the number of pieces occupied: ; Will repeat Second-rate, repeat Second-rate, repeat Second-rate,……, repeat Next, the length is arranged as follows The execution list, for insufficient For the portion where the candidate strategy with the lowest weight is selected, fill it; if the length of the execution list is greater than... If the number of fragments exceeds the limit, then duplicate entries of the corresponding candidate strategies are removed sequentially in ascending order of execution weight, until the execution list length equals the limit. ; Based on the execution list, at each time slice, the corresponding candidate strategy is selected from the execution list and executed.

[0010] Optionally, the step of statistically analyzing click preferences across different channels and user activity levels over various time periods, combining this with user trust levels to generate outreach priorities, and then executing push notifications according to these priorities includes: Obtain user trust rating ; Obtain user click logs from different channels ; Get user's historical conversation volume time series ,in Numbered by hour. This refers to the number of conversations during that time period. Acquiring channels to reach candidates ; Total clicks: ; in: :channel Historical click count; Calculate the user's preference score for the channel: ; Set the time observation window to Select the most recent Hourly conversation data; Calculate the number of sessions in each time period: ; Normalized activity: ; For each channel With time period Calculate the overall score: ; All channels With time period according to Sort the candidates in descending order to generate a candidate list; Execute push tasks in the order of the candidate list.

[0011] Optionally, the step of normalizing metrics such as click-through rate and conversion rate, and using exponential smoothing to update parameters, and feeding the update results back to step S1, includes: Set an evaluation cycle, and within each evaluation cycle: Get the click-through rate of push notifications. And obtain the conversion rate of effective conversions after user clicks. ; Normalization is performed: ; in: : Core indicators after normalization; The minimum and maximum values ​​of this indicator observed within the same evaluation period; Set minimum acceptable level ; Get the generated values ​​of each candidate strategy in the most recent period. and and with If the comparison is lower than If so, the execution rights of this strategy will be reset to zero in the next cycle; For threshold Perform exponential smoothing updates: ; in: : No. The threshold at the next iteration; Behavioral indicators in the most recent assessment period The average normalized strength; The learning rate controls the speed at which the historical threshold is fused with the new average standardized strength. : The updated threshold; Each Send it back to S1 as the updated trigger threshold.

[0012] The present invention has the following beneficial effects: 1. By analyzing the pause duration during dialogues between users and virtual characters. Frequency of topic switching and frequency of keyword occurrence This step involves real-time data collection and standardized processing of three behavioral indicators, enabling the identification of potential interests even when users do not explicitly state their needs; based on historical averages... and standard deviation Calculate the demand trigger threshold and the current indicator value and Direct comparison is only possible when... Demand flags are only triggered when demand exceeds historical normal ranges, thus avoiding misjudgments caused by oversensitivity or improper setting of empirical coefficients; the generated dimensionless demand intensity vector Not only does it eliminate the difference in units of measurement, but it also takes into account the comparability between various indicators, enabling downstream modules to evaluate multi-dimensional demand signals on the same scale. In addition, in high-concurrency real-time dialogue scenarios, this method only requires a small amount of computing resources to complete the high-precision capture of users' potential needs, providing a solid input foundation for subsequent trust assessment and content recommendation. 2. By combining the non-explicit demand vector output in step S1 with the real-time sentiment score Together as features, information gain (IG) is used to analyze the positive user feedback for each feature. The predicted contribution is measured and normalized to generate weights. Then, the weighted sum is mapped to the (0,1) interval using the Sigmoid mapping function, and the user trust score is output. This method is more adaptable and data-driven than traditional static models based on preset rules or experience weights. It can dynamically reflect the user's acceptance of the virtual character's recommended content during the dialogue process, preventing static models from failing in real-time scenarios; sentiment scoring. It is implemented using a lightweight sentiment dictionary scoring and normalization algorithm, which can run online without the need for a deep learning model, reducing system dependence; the information gain weights are calculated once from offline historical data and can be reused efficiently to ensure response speed in the online phase; The value combines smoothness and interpretability, and can be directly interfaced with downstream content generation and strategy optimization modules; therefore, this step, while taking into account both real-time performance and accuracy, provides a reliable measurement basis for subsequent content selection and strategy decisions. 3. Calculate the overlap rate between candidate content fragments and the set of keywords in the dialogue context. Then, combined with user trust levels The optimal result obtained by grid search Weighted content quality score ; here It can measure the relevance of candidate content to the current conversation topic, ensuring content relevance; This indicates the user's willingness to accept the terms; a balance should be struck between the two to avoid being dominated by a single dimension. By using a grid search to determine ROI (Return on Investment) through historical A / B testing, we ensure that the weighting is based on actual results. Compared with traditional template-based or manually set ratios, this step combines effect verification with parameter selection, so that each content generation has a data-driven optimization mechanism. This provides a quantitative standard for subsequent strategy optimization, allowing the system to prioritize content most likely to generate conversions; this process only requires offline A / B test results statistics and can be periodically recalibrated. This step keeps the model in sync with business metrics; therefore, it ensures that the quality of generated content is continuously upgraded while generating relevant content efficiently. 4. By scoring based on content quality and trust Candidate strategies Calculate expected returns Then Normalized policy execution weights This allows for the proportional allocation of strategy call frequencies; furthermore, by... Mapped to the actual number of executions per cycle This approach maintains the fairness of proportional allocation while also ensuring predictability; the dynamic optimization process balances benefits and costs. ,by By measuring cost-effectiveness, this step automatically highlights high-efficiency and low-cost strategies. This step is data-driven throughout, avoiding subjective biases caused by human settings. All algorithms used are calculated online, and resource consumption is controllable. Therefore, this step effectively improves the efficiency and flexibility of strategy application, enabling the system to always make decisions around the optimal ROI in complex scenarios. 5. By adopting channel click preferences Activity level during time period Combined, and then multiplied by user trust level Generate reach priority For all channels With time period The combination is prioritized; this method ensures that the most preferred channels and most active time periods are selected for push notifications, while also considering the user's current level of trust; compared to relying solely on channel open rates or fixed time periods, this step provides a multi-dimensional comprehensive evaluation, resulting in more accurate push decisions; this process does not require large-scale model support and can be implemented based on simple statistics; it significantly improves the click-through rate and conversion rate of marketing outreach, while reducing the risk of user churn; 6. Normalize core metrics such as CTR and CVR. And update the trigger threshold using the exponential smoothing formula. and policy lower bound The updated results are promptly fed back to the preceding steps; normalized metrics can be used to dynamically eliminate inefficient strategies and support budget and channel allocation decisions; exponential smoothing updates combined with learning rates It strikes a balance between historical stability and responsiveness to new trends; the online adaptive closed-loop mechanism enables the system to iterate continuously at the entire process level, avoiding the lag of offline batch updates; the algorithm is lightweight and has no deep dependencies, and can be frequently executed in actual services to achieve real-time optimization. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the basic process of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example, refer to Figure 1 A virtual interaction method based on artificial intelligence, comprising: S1. Statistical analysis and standardization of pause duration, topic switching frequency, and keyword frequency in user dialogues are performed to generate a vector of non-explicit user needs. S2. Based on the aforementioned non-explicit demand vector and real-time sentiment tendency, information gain weighting and processing by a mapping function are used to obtain the user trust score; S3. Extract candidate content and keywords from the dialogue context, calculate their relevance, and combine them with user trust to calculate the content quality score. S4. Evaluate the benefit-cost ratio of each candidate strategy based on content quality scores and trust levels, and allocate strategy execution weights proportionally. S5. Analyze click preferences and user activity levels across different channels and time periods, combine these with user trust levels to generate outreach priorities, and then push notifications according to these priorities. S6. Normalize the click-through rate, conversion rate and other indicators, and use exponential smoothing to update the parameters, and feed the update results back to step S1.

[0016] The process of statistically analyzing and standardizing the pause duration, topic switching frequency, and keyword frequency in user conversations to generate a vector of non-explicit user needs includes: Three types of behavioral indicators are collected in real time during the dialogue between the user and the virtual character: including pause duration. Unit: seconds; Topic switching frequency Unit: times / minute; Statistical by category Number of times similar keywords appear ; Obtain historical dialogue logs for three types of behavioral indicators, and calculate the mean for each behavioral indicator. and standard deviation ,in ; The index value collected at the current moment Mapped to standard scores: ; in: : No. The original behavioral indicator values ​​include pause duration. Frequency of topic switching Number of times keywords appear ; , :index Historical sample mean and standard deviation; Standardized demand components; Obtain the dimensionless demand intensity vector ; The confidence level coefficient is determined based on the 95th percentile of the normal distribution. ; Decision made The degree of strictness, The larger the threshold, the higher the threshold, and the stricter the conditions for determining significant demand; Set the first The demand trigger threshold for each original behavioral indicator value : ; For each raw behavioral indicator value Its corresponding demand trigger threshold Compare: like If the value is significant, then the behavioral indicator is considered a significant demand; otherwise, it is considered to have no significant demand. The original behavioral indicator values ​​corresponding to all identified significant needs Keep the rest and discard the rest, outputting a non-explicit demand vector consisting of significant demand components. By analyzing the pause duration during user-virtual character dialogues... Frequency of topic switching and frequency of keyword occurrence This step involves real-time data collection and standardized processing of three behavioral indicators, enabling the identification of potential interests even when users do not explicitly state their needs; based on historical averages... and standard deviation Calculate the demand trigger threshold and the current indicator value and Direct comparison is only possible when... Demand flags are only triggered when demand exceeds historical normal ranges, thus avoiding misjudgments caused by oversensitivity or improper setting of empirical coefficients; the generated dimensionless demand intensity vector Not only does it eliminate the difference in units of measurement, but it also takes into account the comparability between various indicators, enabling downstream modules to evaluate multi-dimensional demand signals on the same scale. In addition, in high-concurrency real-time dialogue scenarios, this method only requires a small amount of computing resources to complete the high-precision capture of users' potential needs, providing a solid input foundation for subsequent trust assessment and content recommendation. The process of obtaining a user trust score based on the non-explicit demand vector and real-time sentiment tendency, using information gain weighting and processing through a mapping function, includes: The emotional lexicon scoring method is used to score the emotional content of the dialogue between users and virtual characters, and to obtain the user's emotional score corresponding to the real-time emotional tendency. The aforementioned emotion dictionary scoring method is a commonly used existing technology. Based on the non-explicit demand vector Each demand vector component With Emotional Score As a feature set ; Calculate the user feedback of each feature in the feature set. The contribution was determined and normalized to obtain the weights. : ; in: ;No. or the Each input feature includes And emotional score ; Positive user feedback tags are obtained by capturing user feedback behaviors in a preset feedback window that interact with the virtual character's output. These behaviors include clicking, purchasing, adding to favorites, adding to cart, continuing inquiry, requesting details, closing, exiting, and ignoring. When a user performs any of these positive actions within the preset feedback window, such as clicking, purchasing, adding to favorites, adding to cart, continuing inquiry, or requesting details, a positive user feedback tag is assigned to the corresponding interaction sample. A positive feedback label is assigned when the user does not perform a positive action within the preset feedback window, or performs any of the non-positive actions such as closing, exiting, or ignoring. Marked as a non-positive tag; :feature right The information gain is calculated by discretizing the historical dialogue, specifically as follows: A large number of interaction samples were extracted from historical dialogue logs, and each sample contained feature values. and corresponding feedback ; For continuous features Perform binning, dividing into equal-width or equal-frequency groups. each interval ; Among them, the first Number of bins for each continuous feature for: ; in, Indicates the first The total number of valid samples for each continuous feature in historical dialogue samples; Indicates the first The number of different values ​​of a continuous feature in historical dialogue samples; Indicates the first The number of samples with positive user feedback labels in each continuous feature-corresponding sample; Indicates the first The number of samples in a continuous feature corresponding to a sample where the user's positive feedback label is a non-positive label; This indicates rounding down to the square root of the total number of samples. This is used to ensure that at least two intervals are formed in order to distinguish between lower and higher feature levels; This is used to prevent the number of bins from exceeding the sample size, the number of different values, or the statistical range that the feedback labels can support; Calculate the overall entropy : ; in, ; Calculate conditional entropy : Calculate the sample proportion within this interval: ; Calculate the sub-entropy of this interval: ; The conditional entropy is obtained by accumulating the values: ; Calculate information gain ; Calculate user trust level : ; in: The Sigmoid function. This is achieved by combining the non-explicit demand vector output from step S1 with the real-time sentiment score. Together as features, information gain (IG) is used to analyze the positive user feedback for each feature. The predicted contribution is measured and normalized to generate weights. Then, the weighted sum is mapped to the (0,1) interval using the Sigmoid mapping function, and the user trust score is output. This method is more adaptable and data-driven than traditional static models based on preset rules or experience weights. It can dynamically reflect the user's acceptance of the virtual character's recommended content during the dialogue process, preventing static models from failing in real-time scenarios; sentiment scoring. It is implemented using a lightweight sentiment dictionary scoring and normalization algorithm, which can run online without the need for a deep learning model, reducing system dependence; the information gain weights are calculated once from offline historical data and can be reused efficiently to ensure response speed in the online phase; The value combines smoothness and interpretability, and can be directly interfaced with downstream content generation and strategy optimization modules; therefore, this step, while taking into account both real-time performance and accuracy, provides a reliable measurement basis for subsequent content selection and strategy decisions. The process of extracting candidate content and keywords from the dialogue context, calculating their relevance, and weighting the content quality score by combining user trust levels includes: Extracting the keyword set of candidate content segments With dialogue context keyword set ; Calculate the overlap rate : ; Calculate content quality score : ; in: User trust level; Trust weight, with a value of [0,1], is obtained through a grid search of ROI (Return on Investment) based on historical A / B testing; its value determination process is as follows: The ROI (Return on Investment) from actual marketing metrics is used as a measure of effectiveness. Set the grid range: ; Multiple sets of context and candidate content were extracted from historical dialogues to form a test content library, and the number of push content items was set to [number missing]. ; For each ,according to Before scoring and selecting a score Push content to the device; Calculate the ROI of this experiment on the same user group; Select the one that maximizes ROI The value serves as the trust weight. The overlap rate is calculated between candidate content fragments and the set of keywords in the dialogue context. Then, combined with user trust levels The optimal result obtained by grid search Weighted content quality score ; here It can measure the relevance of candidate content to the current conversation topic, ensuring content relevance; This indicates the user's willingness to accept the terms; a balance should be struck between the two to avoid being dominated by a single dimension. By using a grid search to determine ROI (Return on Investment) through historical A / B testing, we ensure that the weighting is based on actual results. Compared with traditional template-based or manually set ratios, this step combines effect verification with parameter selection, so that each content generation has a data-driven optimization mechanism. This provides a quantitative standard for subsequent strategy optimization, allowing the system to prioritize content most likely to generate conversions; this process only requires offline A / B test results statistics and can be periodically recalibrated. This step keeps the model in sync with business metrics; therefore, it ensures that the quality of generated content is continuously upgraded while generating relevant content efficiently. The process of comprehensively evaluating the benefit-cost ratio of each candidate strategy based on content quality scores and trust levels, and allocating strategy execution weights proportionally, includes: For each candidate strategy Calculate expected returns : ; in: Candidate Strategy Corresponding content quality score; : The obtained candidate strategies Execution costs; Normalize all expected returns: ; in: Candidate Strategy Execution weight; according to Each strategy is invoked and executed sequentially from largest to smallest.

[0017] The method The strategies are executed sequentially from largest to smallest, including: Let the scheduling cycle length be... A time slice; For each candidate strategy Calculate the number of pieces occupied: ; Will repeat Second-rate, repeat Second-rate, repeat Second-rate,……, repeat Next, the length is arranged as follows The execution list, for insufficient For the portion where the candidate strategy with the lowest weight is selected, fill it; if the length of the execution list is greater than... If the number of fragments exceeds the limit, then duplicate entries of the corresponding candidate strategies are removed sequentially in ascending order of execution weight, until the execution list length equals the limit. ; Based on the execution list, at each time slice, the corresponding candidate strategy is selected from the execution list and executed. This is done by scoring based on content quality. and trust Candidate strategies Calculate expected returns Then Normalized policy execution weights This allows for the proportional allocation of strategy call frequencies; furthermore, by... Mapped to the actual number of executions per cycle This approach maintains the fairness of proportional allocation while also ensuring predictability; the dynamic optimization process balances benefits and costs. ,by By measuring cost-effectiveness, this step automatically highlights high-efficiency and low-cost strategies. This step is data-driven throughout, avoiding subjective biases caused by human settings. All algorithms used are calculated online, and resource consumption is controllable. Therefore, this step effectively improves the efficiency and flexibility of strategy application, enabling the system to always make decisions around the optimal ROI in complex scenarios. The process of statistically analyzing click preferences across different channels and user activity levels over various time periods, combined with user trust levels, generates a reach priority, and push notifications are executed according to this priority, including: Obtain user trust rating ; Obtain user click logs from different channels ; Get user's historical conversation volume time series ,in Numbered by hour. This refers to the number of conversations during that time period. Acquiring channels to reach candidates ; Total clicks: ; in: :channel Historical click count; Calculate the user's preference score for the channel: ; Set the time observation window to Select the most recent Hourly conversation data; Calculate the number of sessions in each time period: ; Normalized activity: ; For each channel With time period Calculate the overall score: ; All channels With time period according to Sort the candidate list in descending order; Execute push notifications according to the candidate list order. This is achieved by employing channel click preferences. Activity level during time period Combined, and then multiplied by user trust level Generate reach priority For all channels With time period The combination is prioritized; this method ensures that the most preferred channels and most active time periods are selected for push notifications, while also considering the user's current level of trust; compared to relying solely on channel open rates or fixed time periods, this step provides a multi-dimensional comprehensive evaluation, resulting in more accurate push decisions; this process does not require large-scale model support and can be implemented based on simple statistics; it significantly improves the click-through rate and conversion rate of marketing outreach, while reducing the risk of user churn; The step of normalizing metrics such as click-through rate and conversion rate, and updating parameters using exponential smoothing, and then feeding the updated results back to step S1, includes: Set an evaluation cycle, and within each evaluation cycle: Get the click-through rate of push notifications. And obtain the conversion rate of effective conversions after user clicks. Effective conversions include users making purchases after clicking; Normalization is performed: ; in: : The core indicator after normalization, used for cross-indicator comparison; The minimum and maximum values ​​of this indicator observed within the same evaluation period; Set minimum acceptable level ; Get the generated strategies for each candidate strategy in the most recent period. and and with Comparison, if lower If so, the execution rights of this strategy will be reset to zero in the next cycle; For threshold Perform exponential smoothing updates: ; in: : No. The threshold at the next iteration; Behavioral indicators in the most recent assessment period The average normalized strength; The learning rate controls the fusion speed between historical thresholds and the new average standardized strength. Larger values ​​result in higher weights for new data and faster system response, but are also prone to excessive fluctuations. Smaller values ​​result in conservative updates, smooth and stable performance, but a lag in adapting to new trends. : The updated threshold; Each The data is sent back to S1 as the updated trigger threshold to adjust the sensitivity of the "significant demand" determination. This is achieved by normalizing core metrics such as CTR and CVR. And update the trigger threshold using the exponential smoothing formula. and policy lower bound The updated results are promptly fed back to the preceding steps; normalized metrics can be used to dynamically eliminate inefficient strategies and support budget and channel allocation decisions; exponential smoothing updates combined with learning rates It strikes a balance between historical stability and responsiveness to new trends; the online adaptive closed-loop mechanism enables the system to iterate continuously at the entire process level, avoiding the lag of offline batch updates; the algorithm is lightweight and has no deep dependencies, and can be frequently executed in actual services to achieve real-time optimization.

[0018] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0019] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A virtual interaction method based on artificial intelligence, characterized in that, include: S1. Statistical analysis and standardization of pause duration, topic switching frequency, and keyword frequency in user dialogues are performed to generate a vector of non-explicit user needs. S2. Based on the aforementioned non-explicit demand vector and real-time sentiment tendency, information gain weighting and processing by a mapping function are used to obtain the user trust score; S3. Extract candidate content and keywords from the dialogue context, calculate their relevance, and combine them with user trust to calculate the content quality score. S4. Evaluate the benefit-cost ratio of each candidate strategy based on content quality scores and trust levels, and allocate strategy execution weights proportionally. S5. Analyze click preferences and user activity levels across different channels and time periods, generate outreach priorities based on user trust levels, and execute push notifications according to these priorities; S6. Normalize the click-through rate, conversion rate and other indicators, and use exponential smoothing to update the parameters, and feed the update results back to step S1.

2. The virtual interaction method based on artificial intelligence according to claim 1, characterized in that, The process of statistically analyzing and standardizing the pause duration, topic switching frequency, and keyword frequency in user conversations to generate a vector of non-explicit user needs includes: Three types of behavioral indicators are collected in real time during the dialogue between the user and the virtual character: including pause duration. Unit: seconds; Topic switching frequency Unit: times / minute; Statistical by category Number of times similar keywords appear ; Obtain historical dialogue logs for three types of behavioral indicators, and calculate the mean for each behavioral indicator. and standard deviation ,in ; The index value collected at the current moment Mapped to standard scores: ； in: : No. The original behavioral indicator values ​​include pause duration. Frequency of topic switching Number of times keywords appear ; :index Historical sample mean and standard deviation; Standardized demand components; Obtain the dimensionless demand intensity vector ; The confidence level coefficient is determined based on the 95th quantile of the normal distribution. ; Set the first The demand trigger threshold for each original behavioral indicator value : ； For each raw behavioral indicator value Its corresponding demand trigger threshold Compare: like If the value is significant, then the behavioral indicator is considered a significant demand; otherwise, it is considered to have no significant demand. The original behavioral indicator values ​​corresponding to all identified significant needs Keep the rest and discard the rest, outputting a non-explicit demand vector consisting of significant demand components. .

3. The virtual interaction method based on artificial intelligence according to claim 1, characterized in that, The process of obtaining a user trust score based on the non-explicit demand vector and real-time sentiment tendency, using information gain weighting and processing through a mapping function, includes: The emotional lexicon scoring method is used to score the emotional content of the dialogue between users and virtual characters, and to obtain the user's emotional score corresponding to the real-time emotional tendency. ; Based on the non-explicit demand vector Each demand vector component With Emotional Score As a feature set ; Calculate the user feedback of each feature in the feature set. The contribution was determined and normalized to obtain the weights. : ； in: , ;No. or the Each input feature includes And emotional score ; : Positive user feedback tags are obtained by capturing user feedback behavior within a preset feedback window. When a user performs a positive action, the corresponding user's interaction sample is displayed. The Y value of the user in the corresponding interaction sample is marked as a positive label when the user does not perform a positive action or performs a negative action. :feature right The information gain is calculated by discretizing the historical dialogue. Calculate user trust level : ； in: : Sigmoid function.

4. The virtual interaction method based on artificial intelligence according to claim 1, characterized in that, The process of extracting candidate content and keywords from the dialogue context, calculating their relevance, and weighting the content quality score by combining user trust levels includes: Extracting the keyword set of candidate content fragments With dialogue context keyword set ; Calculate the overlap rate : ； Calculate content quality score : ； in: User trust level; Trust weight, with a value of [0,1], is obtained through a grid search of ROI (Return on Investment) based on historical A / B testing.

5. The virtual interaction method based on artificial intelligence according to claim 1, characterized in that, The process of comprehensively evaluating the benefit-cost ratio of each candidate strategy based on content quality scores and trust levels, and allocating strategy execution weights proportionally, includes: For each candidate strategy Calculate expected returns : ； in: Candidate Strategy Corresponding content quality score; : The obtained candidate strategies Execution costs; Normalize all expected returns: ； in: Candidate Strategy Execution weight; according to The strategies are executed sequentially from largest to smallest.

6. The virtual interaction method based on artificial intelligence according to claim 5, characterized in that, The method The strategies are executed sequentially from largest to smallest, including: Let the scheduling cycle length be... A time slice; For each candidate strategy Calculate the number of pieces occupied: ； Will repeat Second-rate, repeat Second-rate, repeat Second-rate,……, repeat Next, the length is arranged as follows The execution list, for insufficient For the portion where the candidate strategy with the lowest weight is selected, fill it; if the length of the execution list is greater than... If the number of fragments exceeds the limit, then duplicate entries of the corresponding candidate strategies are removed sequentially in ascending order of execution weight, until the execution list length equals the limit. ; Based on the execution list, at each time slice, the corresponding candidate strategy is selected from the execution list and executed.

7. The virtual interaction method based on artificial intelligence according to claim 1, characterized in that, The process involves statistically analyzing click preferences across different channels and user activity levels over various time periods, combining this with user trust levels to generate outreach priorities, and then executing push notifications according to these priorities. This includes: Obtain user trust rating ; Obtain user click logs from different channels ; Get user's historical conversation volume time series ,in Numbered by hour. This refers to the number of conversations during that time period. Acquiring channels to reach candidates ; Total clicks: ； in: :channel Historical click count; Calculate the user's preference score for the channel: ； Set the time observation window to Select the most recent Hourly conversation data; Calculate the number of sessions in each time period: ; Normalized activity: ； For each channel With time period Calculate the overall score: ； All channels With time period according to Sort the candidate list in descending order; Execute push tasks in the order of the candidate list.

8. The virtual interaction method based on artificial intelligence according to claim 1, characterized in that, The step of normalizing metrics such as click-through rate and conversion rate, and updating parameters using exponential smoothing, and then feeding the updated results back to step S1, includes: Set an evaluation cycle, and within each evaluation cycle: Get the click-through rate of push notifications. And obtain the conversion rate of effective conversions after user clicks. ; Normalization is performed: ； in: : Core indicators after normalization; The minimum and maximum values ​​of this indicator observed within the same evaluation period; Set minimum acceptable level ; Get the generated values ​​of each candidate strategy in the most recent period. and and with If the comparison is lower than If so, the execution rights of this strategy will be reset to zero in the next cycle; For threshold Perform exponential smoothing updates: ； in: : No. The threshold at the next iteration; Behavioral indicators in the most recent assessment period The average normalized strength; The learning rate controls the speed at which the historical threshold is fused with the new average standardized strength. : The updated threshold; Each Send it back to S1 as the updated trigger threshold.

Images