Personalized recommendation method for virtual digital humans in enterprise publicity

Abstract

The invention relates to the technical field of enterprise propaganda, and discloses a personalized recommendation method for virtual digital humans in enterprise propaganda. According to the method, multi-modal interaction data such as visual attention data and voice feedback data in the interaction process of a user and a virtual digital human are collected in real time, and an original interaction flow is generated; performing multi-dimensional fusion analysis on the original interaction flow, analyzing a dependency relationship among different dimensions, identifying a hidden association between a user interest mode and a virtual digital human performance feature, labeling an analysis result as an initial interest index, and generating an interest labeling data set with confidence; optimizing model adaptability and dynamically updating a model state by utilizing a fusion analysis result; matching the real-time interaction data with the dynamic interest evolution model, and identifying recommendation candidates and opportunities to obtain recommendation contexts; and mining a resource library based on a matching result, identifying a deep recommendation strategy and an adjustment signal, and tracing the strategy to identify a core factor and an optimization direction, thereby realizing accurate and adaptive personalized recommendation.

Description

Technical Field

[0001] This invention relates to the field of corporate publicity technology, specifically a personalized recommendation method for virtual digital humans in corporate publicity. Background Technology

[0002] In the current field of corporate communication, virtual avatars, with their vivid interactive features, are gradually becoming an important medium connecting businesses and users. As users' demands for personalized experiences continue to rise, how to use virtual avatars to push content that aligns with their interests has become a key focus for businesses in their communication efforts. In existing technologies, the virtual digital human recommendation methods used by some companies in their advertising often rely on preset, fixed recommendation rules. These methods can only make recommendations based on a single dimension of user data, such as determining the interactive content of the virtual digital human to be pushed to the user based solely on the types of advertising content the user has previously clicked. Due to the lack of capturing multi-dimensional data during real-time user interactions, it is often impossible to accurately grasp the user's true interests in different interaction scenarios. Some recommendation methods, while attempting to collect user interaction data, fail to fully integrate and analyze multi-dimensional data by incorporating environmental context during data processing. This makes it difficult to analyze the dependencies between different data dimensions, and consequently, to identify the implicit correlation between user interest patterns and the performance characteristics of the virtual avatar. In such cases, the generated interest metrics lack accuracy and reliability, and recommendation models built based on these metrics cannot dynamically adjust their state according to real-time changes in user interactions, resulting in a significant deviation between the recommendation results and the user's actual needs. Existing recommendation methods, after acquiring recommendation candidates, mostly fail to conduct in-depth mining of the virtual digital human resource library, making it impossible to identify potential deep recommendation strategies and adjustment signals. They also struggle to trace the core factors and optimization directions of recommendation strategies, resulting in a lack of effective basis for iterative optimization of the recommendation system. After long-term use, problems such as homogenization of recommended content and decreased user interaction willingness are likely to occur. This fails to meet the needs of enterprises for personalized recommendation effects in the process of corporate publicity, and is not conducive to improving the influence of corporate publicity and user conversion rate. Summary of the Invention

[0003] The purpose of this invention is to provide a personalized recommendation method for virtual digital humans in corporate promotion, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a personalized recommendation method for virtual digital humans in corporate promotion, the method comprising: Real-time collection of multimodal interaction data during user interaction with virtual digital humans, including visual attention data and voice feedback data, to generate raw interaction streams; The original interaction flow is subjected to multi-dimensional fusion analysis. At the same time, the dependency relationship between different dimensions is analyzed by using environmental context information, the implicit association between user interest patterns and virtual digital human performance characteristics is identified, and the analysis results are labeled as initial interest indicators to generate an interest-labeled dataset with confidence. A dynamic interest evolution model is constructed based on the original interaction flow and interest annotation dataset. The adaptability of the model is optimized by using the fusion analysis results, and the model state is dynamically updated. The system matches real-time user interaction data with a dynamic interest evolution model to identify matching recommendation candidates and marks the timing of recommendations. The matching results are then fed back to the virtual digital human resource library for analysis to obtain the recommendation context. Based on the matching results of the dynamic interest evolution model, the virtual digital human resource database is mined to identify potential deep recommendation strategies and adjustment signals. The recommendation strategies are then traced to identify core factors and optimization directions.

[0005] Preferably, the real-time acquisition of multimodal interaction data during the interaction between the user and the virtual digital human, including visual attention data and voice feedback data, to generate the original interaction stream specifically includes: Monitor and identify target user interaction scenarios, and capture multimodal data of real-time user interactions; The captured multimodal data is classified to distinguish between visual and auditory data, and the data is parsed to extract visual attention data and voice feedback data. The parsed data is merged and standardized to generate a structured original interactive stream.

[0006] Preferably, the step of performing multi-dimensional fusion analysis on the original interaction flow, simultaneously using environmental context information to parse the dependencies between different dimensions, identifying the implicit association between user interest patterns and virtual digital human performance characteristics, and labeling the analysis results as initial interest indicators to generate an interest-labeled dataset with confidence, specifically includes: The original interaction flow is organized according to the interaction dimension, and environmental context information is extracted from the data; Dependency analysis is performed between different interaction dimensions to identify normal and abnormal interest patterns. Abnormal interest patterns are labeled, initial interest indicators are generated, and interest type, intensity, and persistence are recorded. Based on the feature scores of each interest metric, confidence is configured to generate an interest-labeled dataset with confidence scores. All analysis and labeling results are then integrated to output a structured interest-labeled dataset.

[0007] Preferably, the step of constructing a dynamic interest evolution model based on the original interaction flow and interest-labeled dataset, optimizing the model's adaptability using the fusion analysis results, and dynamically updating the model state specifically includes: Feature extraction is performed on the original interaction flow and interest annotation dataset to construct a dynamic interest evolution model, and the dynamic interest evolution model is trained using the extracted features; The parameters and structure of the model are dynamically adjusted based on the real-time fusion analysis results.

[0008] Preferably, the step of matching real-time collected user interaction data with a dynamic interest evolution model, identifying matching recommendation candidates, marking the recommendation timing, and feeding the matching results back to the virtual digital human resource library for parsing to obtain the recommendation context specifically includes: The real-time user interaction data is input into the dynamic interest evolution model. The dynamic interest evolution model is used to match the real-time input data, analyze the consistency between the real-time data and the expected interest pattern data, and identify potential recommendation candidates. The detected matching recommendation candidates are labeled, and the degree of matching and relevant environmental context information are recorded; Prioritize the recommended candidates and assign a recommendation level to each candidate based on matching strength and frequency; Extract the matching results and feed them back to the virtual digital human resource library. Query the virtual digital human resource library and read the corresponding recommendation context.

[0009] Preferably, the matching results based on the dynamic interest evolution model are used to mine the virtual digital human resource database, identify potential deep recommendation strategies and adjustment signals, and trace the recommendation strategies to identify core factors and optimization directions, specifically including: Collect virtual digital human resource library data related to the matching results and obtain complete environmental context information; Mining data from the virtual digital human resource database to identify potential deep recommendation strategies and adjustment signals, extracting potential strategy clues from the resource database, and identifying strategy types; The identified adjustment signals are traced and analyzed to identify the generation path of the adjustment signals and trace the original source of the adjustment; A comprehensive report is generated based on all the identified information.

[0010] Preferably, the method further includes: Perform time series analysis on the initial interest metrics to capture changes in interest trends and update the interest annotation dataset; The dynamic interest evolution model is further optimized using the updated interest-labeled dataset, forming a closed-loop feedback mechanism.

[0011] Preferably, the step of performing time series analysis on the initial interest indicators, capturing changes in interest trends, and updating the interest annotation dataset specifically includes: Calculate the slope of the initial interest index over time to identify trends of rising or falling interest. Adjust the confidence weights in the interest-labeled dataset based on the slope metric, and regenerate the updated interest-labeled dataset.

[0012] Preferably, the step of further optimizing the dynamic interest evolution model using the updated interest-labeled dataset to form a closed-loop feedback mechanism specifically includes: The updated interest-labeled dataset was used as additional training data to retrain the dynamic interest evolution model; Adjust the recommended candidate recognition threshold during the real-time matching process based on the output of the retrained model.

[0013] Preferably, adjusting the recommended candidate identification threshold during the real-time matching process specifically includes: Based on the performance metrics of the retrained model, the optimal value of the recommendation candidate recognition threshold is dynamically calculated. Applying optimal values ​​to the real-time matching process ensures the accuracy of recommendation timing annotations.

[0014] Compared with the prior art, the beneficial effects of the present invention are: The company's advertised personalized recommendation method for virtual digital humans utilizes real-time collection of multimodal interaction data during user interactions with the virtual digital human, including visual attention data and voice feedback data, to generate raw interaction streams. This comprehensively captures various behavioral information of users during interactions, avoiding the problem of incomplete user interest capture caused by traditional recommendation methods relying on only single-dimensional data. The introduction of multimodal data makes the characterization of user interaction behavior more three-dimensional, more realistically reflecting the user's state in different interaction scenarios, laying the foundation for accurate identification of user interests. When processing the original interaction flow, this method combines environmental context information for multi-dimensional fusion analysis, analyzes the dependencies between different dimensions, and identifies the implicit correlation between user interest patterns and the performance characteristics of virtual digital humans. Simultaneously, the analysis results are labeled as initial interest indicators, generating an interest-labeled dataset with confidence scores. This approach fully considers the influence of environmental factors on user interests, enabling the uncovering of hidden user interest patterns behind surface data. The introduction of confidence scores provides a basis for judging the reliability of interest indicators, facilitating the subsequent construction of recommendation models that better reflect users' actual interests. A dynamic interest evolution model is constructed based on the original interaction flow and interest-labeled dataset. The model's adaptability is optimized and its state is dynamically updated using fusion analysis results, enabling the recommendation model to adjust in real time as the user interaction progresses. Compared to traditional fixed models, the dynamic evolution characteristic allows the model to continuously track changes in user interests. Even if user interests change due to interaction scenarios or the passage of time, the model can promptly perceive and update, ensuring that the recommendation direction remains consistent with the user's current interests and reducing the deviation between recommendation results and user needs. By matching real-time user interaction data with a dynamic interest evolution model, identifying matching recommendation candidates and marking the timing of recommendations, and then feeding the matching results back to the virtual digital human resource library to parse and obtain the recommendation context, this process achieves precise alignment between recommendation candidates and users' real-time needs. Marking the timing of recommendations ensures that content is pushed to users at the most suitable interaction point, increasing user acceptance of recommended content. Obtaining the recommendation context provides richer information support for the virtual digital human to present recommended content, making the display of recommended content more relevant to the interaction scenario and enhancing the user's interactive experience. By mining the virtual human resource database based on the matching results of a dynamic interest evolution model, potential deep recommendation strategies and adjustment signals can be identified. Tracing these strategies helps identify core factors and optimization directions, providing clear guidance for the continuous optimization of the recommendation system. Mining deep strategies within the resource database can further enrich recommendation methods and avoid content homogenization. Tracing adjustment signals and core factors helps identify shortcomings in the recommendation process, clarify optimization directions, and drive continuous improvement of the recommendation system. This, in turn, enhances the personalization of virtual human recommendations in corporate promotion, strengthens user interaction with virtual humans, helps companies better convey promotional information, and improves the effectiveness of corporate advertising. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the personalized recommendation method for virtual digital humans in corporate promotion as described in this invention. Figure 2 A flowchart for generating the raw interaction flow by collecting multimodal interaction data in real time; Figure 3 A flowchart for constructing and optimizing a dynamic interest evolution model. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1 This invention provides a personalized recommendation method for virtual digital humans in corporate promotion. The method includes visual attention data and voice feedback data to generate an original interaction stream. The original interaction stream undergoes multi-dimensional fusion analysis, utilizing environmental context information to parse the dependencies between different dimensions, identifying implicit associations between user interest patterns and virtual digital human performance characteristics. The analysis results are labeled as initial interest indicators, generating a confidence-weighted interest-labeled dataset. A dynamic interest evolution model is constructed based on the original interaction stream and the interest-labeled dataset. The model's adaptability is optimized using the fusion analysis results, and the model state is dynamically updated. Real-time collected user interaction data is matched with the dynamic interest evolution model to identify matching recommendation candidates and label the recommendation timing. The matching results are fed back to the virtual digital human resource library for parsing to obtain the recommendation context. Finally, based on the matching results of the dynamic interest evolution model, the virtual digital human resource library is mined to identify potential deep recommendation strategies and adjustment signals, and the recommendation strategies are traced to identify core factors and optimization directions.

[0018] Example 1: See Figure 2 Monitoring and identifying the target user interaction scenario is the starting point of this process. The system automatically identifies the nature of the current interaction through a scenario perception module, such as whether it is a product demonstration, Q&A consultation, or entertainment interaction. This module makes judgments based on predefined scenario rules and real-time environmental sensor input. Capturing multimodal data of real-time user interaction requires the coordinated work of multiple hardware devices. High-resolution cameras continuously track user facial expressions and body movements, infrared eye trackers accurately record the trajectory of visual focus movement, and high-fidelity microphone arrays capture user speech content and tone changes. All devices ensure data consistency through a precise time synchronization controller. Classifying the captured multimodal data is a crucial step in organizing the mixed raw signals. Visual layer data is fed into a dedicated image processing pipeline, which includes facial feature point detection algorithms and gaze direction estimation models to analyze the user's attention focus area and emotional state changes. Auditory layer data, after noise reduction and enhancement, is converted into text information by a speech recognition engine. Simultaneously, an acoustic feature analysis module extracts the emotional color and excitement index of the speech. This layered processing method ensures professional analysis quality for different types of data. The integration and standardization of the parsed data is a systematic integration process. The designed data fusion engine receives structured outputs from visual and auditory channels, aligns and stitches them according to a unified time frame, and standardization operations include data format conversion, unit normalization and outlier filtering. Finally, it generates a raw interactive stream with a rigorous structure and accurate timestamps. This data stream provides a clean and well-organized input source for subsequent in-depth analysis.

[0019] Organizing the original interaction flow according to interaction dimensions is a prerequisite for analysis. The system decomposes the data flow into independent but related data channels such as visual attention, voice feedback, and behavioral response dimensions, with each dimension containing complete time-series information. Extracting environmental context information from the data requires examining the overall background of the interaction, including the content theme currently displayed by the virtual human, the user's historical interaction records, the duration of the session, and interfering factors in the environment. This contextual information is encoded as structured metadata and stored in association with the original interaction flow. Dependency analysis between different interaction dimensions employs multivariate time-series correlation calculation. The system detects the temporal correlation patterns between changes in visual attention and voice feedback content, identifying normal interest patterns that conform to expectations and abnormal patterns that deviate from expectations. Abnormal patterns may manifest as user avoidance reactions to specific topics or sudden shifts in interest. After identifying normal and abnormal interest patterns, abnormal patterns need to be highlighted. During the highlighting process, the specific type of interest, such as content preference type or interaction style tendency, is recorded, along with quantified interest intensity indicators and persistence over different time periods. Generating initial interest metrics is a comprehensive judgment process. The system generates quantitative metrics representing user interests based on convergent evidence from multi-dimensional data. Each metric is accompanied by a type label, intensity value, and duration stamp, forming a preliminary interest profile. Assigning confidence scores based on the characteristics of each interest metric is a calculation process based on data quality and consistency. Interest signals from multiple dimensions and with high temporal consistency are assigned higher confidence scores, while signals with a single dimension or contradictory signals are assigned lower confidence scores. The confidence score calculation considers the completeness of data collection and the signal-to-noise ratio of the signal.

[0020] Generating a dataset of interest-annotated data with confidence scores requires integrating all analysis results. The system binds initial interest metrics to corresponding confidence scores and retains the original terminal data and environmental context used to generate these metrics, forming a complete chain of evidence. The final output of this embodiment is a structured interest-annotated dataset. The dataset uses a hierarchical storage structure: the top layer is a summary of interest metrics, and the lower layer contains the original terminal data supporting these metrics and analysis process logs. This structure ensures both data utilization efficiency and traceability. The entire implementation process forms a complete closed loop from multimodal data acquisition to deep interest analysis, while a rigorous data processing workflow ensures the reliability and practicality of the analysis results. In the data acquisition stage, the acquisition of visual attention data relies on advanced computer vision algorithms. These algorithms can identify the screen coordinates and dwell time of the user's gaze points in real time, while combining the positional information of interface elements to determine the actual content area the user is focusing on. The processing of voice feedback data is more complex, requiring the differentiation between meaningful user voice input and background noise, and the identification of keywords and emotional tendencies in the voice. These processes need to balance real-time performance and accuracy to avoid impacting the interactive experience due to processing delays. The challenge of multimodal data fusion lies in the temporal alignment and semantic integration of different modal data. The system adopts a dynamic time warping algorithm to solve the time asynchrony problem caused by the acquisition delay of different sensors, and establishes a cross-modal semantic mapping table to realize the correlation analysis between visual attention and voice feedback.

[0021] The quality of environmental context information parsing directly impacts the accuracy of interest recognition. The system not only considers the current interaction context but also establishes a personalized baseline by referencing the user's historical behavior patterns, thereby more accurately determining the meaning of the current interest signal. During dependency analysis, the system pays particular attention to causal relationships between different dimensions of data rather than simple correlations. For example, whether a user's positive verbal feedback on a certain content occurs immediately after their visual attention is focused is crucial; this temporal causal relationship reflects the true interest association better than simple simultaneous occurrence. The generation of interest indicators is not a simple threshold judgment but a classification process based on machine learning. The system uses a pre-trained interest recognition model to comprehensively judge multi-dimensional features, and the model is fine-tuned according to different application scenarios to adapt to the needs of enterprise promotion. The confidence allocation mechanism is the essence of this embodiment. It allows the system to transparently express the uncertainty of each interest judgment. High-confidence interest indicators can be directly used for recommendation decisions, while low-confidence indicators can trigger deeper data collection or interactive confirmation.

[0022] Example 2: See Figure 3The feature extraction process is conducted on the original interaction flow and interest-annotated dataset, extracting representative pattern features from the time series. These features include, but are not limited to, the variation patterns of user gaze duration, the fluctuation cycle of emotional intensity in voice feedback, and behavioral sequence features at different interaction stages. The architecture of the dynamic interest evolution model adopts a neural network structure with memory units, which can capture the temporal dependence of user interests. The model training process uses interest datasets with temporal annotations and minimizes the difference between predicted interest patterns and actual observations through backpropagation. Dynamically adjusting model parameters and structure based on real-time fusion analysis results is a continuous optimization process. When multi-dimensional fusion analysis identifies new interest patterns or significant changes in existing patterns, the model triggers a parameter update mechanism. This update includes both subtle adjustments to connection weights and adaptive reconstruction of the network structure. The model's adaptive mechanism can automatically adjust its internal representation according to changes in the distribution of input data. When a long-term shift in user interest patterns is detected, the model initiates a structural optimization process, increasing or decreasing the number of hidden units or adjusting the network depth to better capture new interest features.

[0023] Inputting real-time user interaction data into the dynamic interest evolution model requires rigorous data preprocessing. Newly input multimodal data first undergoes the same standardization process as the training data, then is converted into a tensor format acceptable to the model, ensuring complete consistency between the input data and the model's expected data specifications. When using the dynamic interest evolution model to match and analyze real-time input data, the model calculates the similarity between current interaction features and learned interest patterns. This calculation considers not only static feature matching but also assesses the consistency of dynamic trends, thereby identifying potential recommendation candidates. Analyzing the consistency between real-time data and expected interest patterns employs a multi-scale matching strategy, focusing on both local matching of short-term interaction fragments and typical interest patterns, as well as examining global consistency exhibited during long-term interactions. This multi-granularity analysis improves the accuracy of recommendation candidate identification. Labeling detected matching recommendation candidates requires establishing a comprehensive metadata recording system. Each labeled candidate is accompanied by detailed information such as a matching score, matching time point, duration, and matching feature vector. When recording the degree of matching and related environmental context information, the system captures environmental variables such as the virtual digital human's state, user sentiment indicators, and interaction scenario type at the time of matching. This contextual information helps to understand the actual meaning of the matching results. The priority of recommended candidates is evaluated using a composite scoring mechanism based on matching strength and frequency. Matching strength is determined by the distance metric and pattern overlap ratio in the feature space, while frequency is calculated by counting the number of times the same or similar candidates appear within a specific time window. The weighted combination of these two factors forms the final priority score.

[0024] Assigning recommendation levels to each candidate requires establishing a tiered standard system. The system sets multiple recommendation level thresholds based on application needs, mapping composite scores to different recommendation levels. Higher-level candidates receive priority recommendation, while lower-level candidates may enter a wait-and-see state. Extracting matching results and feeding them back to the virtual human resource library requires designing an efficient data interface. Matching results include candidate identifiers, recommendation levels, matching details, and timestamps. This information is transmitted to the resource library query interface in the form of standardized data packets. When querying the virtual human resource library, the system generates database search conditions based on the candidate's feature vectors. The search process utilizes inverted index technology to accelerate query speed, ensuring rapid result retrieval in real-time interactive scenarios. Reading the corresponding recommendation context requires the resource library to provide rich content association information. This context includes specific materials for the recommended content, suggested presentation methods, expected interaction effect data, and historical recommendation effect records. Obtaining complete context provides sufficient basis for subsequent recommendation strategy formulation.

[0025] During the feature extraction phase, the system employs a multi-level feature extraction strategy. Low-level features capture specific details of interactive behavior, mid-level features abstract behavioral sequence patterns, and high-level features integrate long-term interest trends. This hierarchical feature representation enables the model to comprehensively understand user interests from micro to macro levels. During model training, a rolling time window training data selection strategy is used, consistently employing the most recent interaction data as the primary training sample while retaining representative historical data to prevent the model from forgetting important patterns. This data strategy balances the model's adaptability to new patterns with its memory capacity for historical patterns. Computational efficiency during real-time matching is ensured through various optimization techniques. Pruning techniques are used during model inference to reduce computational load, and the matching algorithm employs approximate nearest neighbor search to improve retrieval speed. These optimizations ensure smooth operation even in resource-constrained real-time environments. The reliability assessment mechanism for matching results generates a credibility index for each matching result. This index comprehensively considers model confidence, data quality factors, and matching consistency. Matching results with low credibility are marked as requiring manual review or further verification. The dynamic adjustment mechanism for recommendation levels continuously optimizes the level classification criteria based on actual recommendation feedback. When the actual interaction effect of recommended content at a certain level is found to be consistently better or worse than expected, the system automatically adjusts the threshold range corresponding to that level, forming a feedback-based optimization loop. The query optimization of the virtual digital human resource library employs caching strategies and preloading mechanisms. The context of frequently queried recommended content is cached in memory, while context predicted to be needed based on the current interaction scenario is preloaded. These strategies significantly reduce query latency.

[0026] Example 3: Collecting virtual human resource library data related to matching results requires establishing an efficient data association mechanism. The system retrieves all relevant data items in the resource library based on the recommendation candidate identifiers in the matching results, including structured information such as content metadata, user interaction history, and effect evaluation records. Obtaining complete environmental context information requires integrating multiple data sources. The system extracts basic environmental variables such as interaction time, duration, and user device type from session logs, obtains system load and response time data from the performance monitoring system, and obtains attribute information of the currently displayed content of the virtual human from the content management platform. These multi-dimensional environmental variables together constitute the complete context required for analysis. Multi-modal analysis methods are used to mine the virtual human resource library data. Association rule mining algorithms are used to discover symbiotic relationships between different recommended content, sequence pattern analysis technology is used to identify effective recommended content ordering, and cluster analysis is used to discover feature combinations with similar effects. Identifying potential deep recommendation strategies requires going beyond surface associations to explore causal relationships. The system compares and analyzes the feature differences between successful and unsuccessful recommendation cases to find combinations of factors that statistically significantly affect recommendation performance. These combinations of factors may involve the synergistic effects of multiple dimensions such as content features, presentation timing, and interaction methods. Potential strategy cues are extracted from the resource library using a graph-based representation method. Entities such as recommended content, user responses, and environmental factors are represented as nodes, and the relationships between them are represented as edges. The graph traversal algorithm is used to discover path patterns that connect to successful recommendation results.

[0027] Identifying strategy types requires establishing a comprehensive classification system. The system categorizes discovered strategies into main types such as content-driven, timing-sensitive, and personalized adaptation, with each category further subdivided into specific strategy subcategories. This classification system facilitates subsequent strategy management and application. Tracing the identified adjustment signals employs a reverse reasoning mechanism, starting from observed changes in effects and tracing back along data dependencies to gradually pinpoint the initial decision point or data change point leading to the adjustment. The tracing process records the complete causal chain and supporting evidence for each step. Identifying the generation path of the adjustment signal requires reconstructing the decision history. By replaying historical decision logs and data state change sequences, the system accurately reconstructs the complete temporal process from the initial interaction event to the final adjustment signal. This reconstruction helps in understanding the internal logic of the system's behavior. Tracing the original source of the adjustment requires distinguishing between direct triggering factors and root causes. The system not only records the most recent event directly leading to the adjustment but also deeply analyzes the background conditions and historical factors that led to the event. This deep attribution analysis can uncover systemic optimization opportunities. A comprehensive report, generated based on all identified information, is presented in a structured document format. The report includes standard sections such as an executive summary, a description of the analytical methods, key findings, and strategic recommendations, along with detailed data appendices to support the verification of the report's conclusions. The report generation process combines automated templates with manual review to ensure that the report maintains a standardized information structure while containing in-depth professional insights.

[0028] In evaluating the effectiveness of a strategy, the system uses the following quantitative indicators to measure its impact:

[0029] Where: symbol The strategy influence index measures the standardized impact of a specific recommendation strategy on user experience. (Symbol) This indicates the total number of scenarios in which the strategy was applied, with the symbol [symbol missing]. It is the first The weighting coefficients for each application scenario are determined based on the scenario's importance and data quality. (Symbol) Indicates the first The changes in key performance indicators before and after strategy implementation in each scenario. (Symbol) It represents the number of time points for effect observation, with the symbol... It is the first Effect observations at each time point, symbol It is the average of the observed effects. This calculation formula considers both the magnitude and stability of the policy effect, avoiding misjudgments due to individual outliers.

[0030] During the data collection phase, the system establishes a distributed data acquisition pipeline capable of processing multiple resource library query requests corresponding to matching results in parallel. The query process employs an incremental loading strategy to avoid system pressure caused by loading large amounts of data at once. Integrating environmental context information faces challenges related to data heterogeneity and temporal alignment. The system uses a unified timeline to synchronize all context data and maps data from different sources to a unified semantic model through a data transformation layer, ensuring the consistency and interpretability of context information. The identification process for deep recommendation strategies introduces a strategy utility evaluation metric, which comprehensively considers the application frequency, success rate, and stability of the strategy. Only strategy patterns that demonstrate good performance in multiple scenarios are labeled as deep strategies. The extraction of strategy cues not only focuses on the effectiveness of individual strategies but also analyzes the synergistic or antagonistic relationships between strategies. This relationship analysis helps to construct more systematic strategy combination schemes.

[0031] When generating comprehensive reports, the system places particular emphasis on the actionability and prioritization of findings. The reports clearly indicate which findings can be immediately applied to optimize the recommendation system, which require further validation, and which require infrastructure support. This tiered approach helps decision-makers allocate optimization resources rationally. The entire implementation process embodies a complete transformation path from data to insights. Through systematic data mining and deep analysis, raw matching results are transformed into valuable strategic knowledge, providing a scientific basis for the continuous optimization of the recommendation system. The design of the analytical methods fully considers the specificities of corporate promotional scenarios, focusing not only on improving short-term interaction effects but also on establishing long-term user relationships and conveying brand value. Domain knowledge constraints are introduced during strategy identification. The system's built-in corporate promotional goal model performs compliance checks and value alignment assessments on the identified strategies, ensuring that the recommended strategies align with the company's promotional objectives and brand image requirements.

[0032] Example 4: Time series analysis is performed on initial interest indicators to capture changes in interest trends, and the interest-labeled dataset is updated. The updated dataset is then used to optimize the dynamic interest evolution model, forming a closed-loop feedback mechanism. In a specific example of an automotive company's promotional scenario, a virtual digital human introduces a new electric vehicle model to users. The system continuously collects user interaction data from the past two weeks, including changes in attention to three main topics: battery technology, intelligent driving, and interior design. Calculating the slope of the initial interest indicators over time requires constructing a time series model. The system performs linear fitting on the intensity value of each interest indicator on a daily basis, and the resulting slope value reflects the rate of increase or decrease in user interest. A threshold mechanism is set to identify the rising or falling trends of interest; when the absolute value of the slope continuously exceeds a predetermined threshold, the system determines that a significant trend exists.

[0033] The confidence weights in the interest-annotated dataset are dynamically adjusted based on the slope index. For interest indicators showing a clear upward trend, the system will appropriately increase their confidence weights, as continuously strengthening interest patterns usually have higher predictive value. Regenerating the updated interest-annotated dataset requires reconstructing the data storage structure. The system adds a trend dimension, including metadata such as trend direction, trend strength, and trend persistence, while retaining the original interest indicators, forming an enhanced dataset. See Table 1, which shows the changes in a user's interest indicators for three promotional points over time and the corresponding confidence adjustments.

[0034] Table 1: Time Series Analysis and Confidence Adjustment of Interest Indicators

[0035] Using the updated interest-labeled dataset as additional training data requires data balancing. The system appropriately weights samples with obvious trends to ensure the model can better learn the evolutionary patterns of interests. Retraining the dynamic interest evolution model employs incremental learning, fine-tuning the parameters based on the original model to avoid the computational overhead and historical knowledge loss associated with complete retraining. A performance evaluation mechanism is established by adjusting the recommendation candidate recognition threshold in the real-time matching process based on the retrained model output. The system tests the recommendation accuracy and coverage under different threshold settings on the validation set to select the optimal balance point.

[0036] The slope metric is calculated using a rolling time window strategy. The system calculates not only the global slope but also the local slope within a recent specific time period. The global slope reflects long-term trends, while the local slope captures short-term changes; combining the two improves the accuracy of trend judgment. A volatility filter is introduced into the trend identification process. For interest metrics with large slopes but drastic fluctuations, the system lowers their trend reliability score to avoid misjudgments due to data noise. The confidence weight adjustment considers trend consistency; trends that maintain the same direction over multiple consecutive time periods receive larger weight adjustments because consistent trends have higher predictive stability. The updated interest labeling dataset is version-managed, with each version recording a complete data change history and the reasons for adjustments, facilitating subsequent auditing and model behavior analysis. During retraining, the model structure is adaptively adjusted. For interest patterns with obvious trend characteristics, the model increases the weight of the temporal attention mechanism to improve its ability to capture sequence patterns. The adjustment of the recommendation candidate identification threshold adopts a hierarchical strategy. Different interest types and trend intensities correspond to different threshold standards. Interest types with upward trends are assigned lower thresholds to capture recommendation opportunities, while interest types with downward trends are assigned higher thresholds to avoid over-recommendation. In the automotive advertising example, the continuously rising interest in battery technology led to a gradual increase in its confidence weight. After retraining, the model became more sensitive to this type of interest, enabling it to identify recommendation opportunities earlier when similar patterns reappeared. The declining interest in intelligent driving triggered a system adjustment strategy, lowering the recommendation priority of related content and redirecting resources to topics of greater user interest. The moderately rising interest in interior design received a moderate increase in confidence, and the system adopted a balanced approach in its recommendation strategy, neither overemphasizing nor completely ignoring it.

[0037] In time series analysis, the outlier detection mechanism avoids interference from short-term fluctuations in trend judgment. The system uses robust statistical methods to identify and appropriately handle outliers, ensuring that the slope calculation reflects the true trend rather than noise. A decay factor is introduced in the confidence weight adjustment, giving more recent data higher weight than historical data, ensuring the system's timely response to changes in user interests. Model retraining employs elastic weight consolidation technology, retaining the memory of important historical patterns while incorporating new trend features, preventing the model from forgetting important historical patterns as it adapts to new trends. The threshold adjustment process considers the diversity and balance of recommended content, avoiding excessive concentration on a few content areas with obvious upward trends and maintaining reasonable coverage of recommended content. The closed-loop feedback mechanism includes manual supervision; automatic adjustments of important parameters require verification through business logic to ensure that the system optimization direction aligns with promotional goals. In trend analysis, the system not only focuses on changes in interest intensity but also analyzes changes in interest persistence and attention patterns, providing a more comprehensive interest evolution map through multi-dimensional trend analysis. Confidence adjustment considers the phased characteristics of trends, employing different adjustment strategies for early and mature trends. Early trends are given more observation space, while mature trends are directly converted into weight adjustments. Knowledge distillation is applied during model updates to extract key knowledge from the old model and transfer it to the new model, maintaining the stability of the system's recommendations. Threshold optimization employs a multi-objective balancing approach, considering multiple metrics such as click-through rate, interaction depth, and user satisfaction to find the optimal threshold setting. During implementation, an A / B testing framework is used to verify the adjustments. New parameter settings are tested on a small scale with limited traffic before being rolled out nationwide, ensuring the robustness of the system optimization.

[0038] Example 5: In a virtual digital human tutoring scenario at an online education institution, the system continuously optimizes its recommendation strategy to improve learning engagement by analyzing students' evolving interest patterns in different knowledge points. Using the updated interest-labeled dataset as additional training data requires a data filtering mechanism. The system selects sample points with significant trend changes from the time-series analysis results. These sample points contain key signals of user interest shifts and are highly valuable for model retraining. The retraining of the dynamic interest evolution model employs a combination of incremental learning and elastic weight consolidation. While absorbing new trend patterns, the model retains its memory of historically important patterns through the elastic weight consolidation algorithm. In a linear algebra tutoring scenario, the system found that students' interest in vector application problems showed a continuous upward trend, while their interest in abstract proof problems gradually declined. These trend changes were encoded into the updated dataset. The model retraining process uses a phased strategy. First, the output layer of the model is fine-tuned using the new dataset to quickly adapt to trend changes. Then, the hidden layer parameters are gradually adjusted to achieve deeper model optimization. Adjusting the recommendation candidate identification threshold in the real-time matching process based on the retrained model output requires establishing a performance monitoring system. The system simulates recommendation effects under different threshold settings in a test environment, monitoring metrics including recommendation hit rate, interaction depth, and user dwell time. A multi-objective optimization method is used to dynamically calculate the optimal value of the recommendation candidate identification threshold based on the retrained model's performance metrics. The system needs to balance recommendation accuracy and coverage, avoiding missed recommendation opportunities due to excessively high thresholds or excessive invalid recommendations due to excessively low thresholds. A gradual update strategy is adopted when applying the optimal value to the real-time matching process. The new threshold is first tested with a small user base, and its application scope is gradually expanded after verifying its effectiveness with actual interaction data.

[0039] In the online education example, the rising interest in vector application questions leads to an appropriate lowering of their recommendation threshold. When the model detects that a student's interest in related content has reached the new threshold, a recommendation is triggered even if the absolute value of the interest intensity is not high, thus seizing teaching opportunities earlier. Conversely, a declining interest in abstract proof questions corresponds to an upward adjustment of the threshold; recommendations are only made when the interest intensity reaches a higher standard to avoid student boredom. The threshold adjustment process considers content relevance, setting association threshold rules for related content within the same knowledge system. When interest in the main knowledge point increases, the recommendation threshold for its related content is adjusted accordingly. The data weighting strategy during model retraining reflects the differences in trend importance, assigning higher weights to trend samples with large slopes and strong persistence, enabling the model to respond more quickly to significant changes. A context-adaptive mechanism is introduced into threshold calculation, using differentiated threshold calculation parameters for different learning stages and knowledge modules. For example, the threshold is appropriately relaxed during the introduction of new knowledge points to explore interest, while the threshold is increased during the review stage for precise recommendations. The threshold application in the real-time matching process adopts a dynamic floating method. The system fine-tunes the threshold in real time based on the recent recommendation effect. When the effect is good, the threshold is maintained or relaxed, and when the effect declines, the threshold standard is tightened.

[0040] Personalized needs in educational scenarios are met through a threshold grouping strategy. The system establishes different threshold profiles for different types of learners. For example, a lower threshold is set for active, exploratory learners to encourage discovery, while a higher threshold is set for passive learners to ensure recommendation quality. The synergy between model optimization and threshold adjustment is enhanced through a feedback loop. Changes in recommendation behavior resulting from threshold adjustments generate new interaction data, which is fed back into model training to initiate a new optimization cycle. During retraining, the model structure is flexibly adjusted. When a structural change in interest patterns is detected, the system automatically increases network capacity or adjusts the attention mechanism to improve the ability to capture complex trends. Threshold optimization considers long-term effects. The system not only evaluates immediate interaction metrics but also tracks the impact of recommended content on long-term learning outcomes, avoiding the pursuit of short-term interactions at the expense of learning effectiveness. When applying optimal values ​​to the real-time environment, safety boundaries are set, and the threshold change range is controlled within a reasonable range to prevent system fluctuations caused by excessive single adjustments. In specific implementation, the system uses a shadow mode to verify the adjustment effect. Before formal application, the new threshold is run in parallel to compare the differences in recommendation decisions under the new and old thresholds, and the adjustment effect is predicted through comparative analysis. Model version management ensures system rollback capability, saving complete historical versions after each major adjustment, allowing for rapid restoration to a stable state when unexpected problems arise in a new version. Through continuous monitoring, analysis, and adjustments, the virtual digital human tutoring system can adapt to changes in student interests.

[0041] The different subject characteristics of educational institutions are reflected through subject adapters. Logically demanding subjects like mathematics, physics, and chemistry are handled with different trend analysis parameters and threshold calculation rules compared to humanities subjects. Learning progress is incorporated into threshold adjustment considerations; the system sets different threshold adjustment strategies for prior knowledge points, current key knowledge points, and extended knowledge points. A hybrid model is used to balance group interest trends with individual differences; the system considers the overall interest changes of the student population while retaining the ability to respond to individual specific interest patterns. Real-time matching efficiency is ensured through threshold indexing technology; the system establishes fast retrieval indexes for different threshold ranges, avoiding the performance pressure of full-scale matching calculations. Transparent recording of threshold adjustments aids in teaching analysis; the system fully records the reasons, magnitude, and effects of each threshold adjustment, providing data support for optimizing teaching strategies. A dynamic optimization mechanism enables the virtual digital human to continuously improve tutoring strategies as its understanding of students deepens, much like a human teacher. Seasonal changes are reflected in threshold adjustments; the system identifies changes in student interest patterns during special periods such as winter and summer vacations and exam weeks, adjusting threshold parameters in advance to adapt to specific needs. A multi-dimensional threshold coordination mechanism avoids recommendation conflicts. When multiple pieces of content meet the recommendation criteria at the same time, the system intelligently selects the best recommendation based on the content priority and the student's current status.

[0042] 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.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A personalized recommendation method for virtual digital humans in corporate promotion, characterized in that, The method includes: Real-time collection of multimodal interaction data during user interaction with virtual digital humans, including visual attention data and voice feedback data, to generate raw interaction streams; The original interaction flow is subjected to multi-dimensional fusion analysis. At the same time, the dependency relationship between different dimensions is analyzed by using environmental context information, the implicit association between user interest patterns and virtual digital human performance characteristics is identified, and the analysis results are labeled as initial interest indicators to generate an interest-labeled dataset with confidence. A dynamic interest evolution model is constructed based on the original interaction flow and interest annotation dataset. The adaptability of the model is optimized by using the fusion analysis results, and the model state is dynamically updated. The system matches real-time user interaction data with a dynamic interest evolution model to identify matching recommendation candidates and marks the timing of recommendations. The matching results are then fed back to the virtual digital human resource library for analysis to obtain the recommendation context. Based on the matching results of the dynamic interest evolution model, the virtual digital human resource database is mined to identify potential deep recommendation strategies and adjustment signals. The recommendation strategies are then traced to identify core factors and optimization directions.

2. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 1, characterized in that, The real-time acquisition of multimodal interaction data during the interaction between the user and the virtual digital human, including visual attention data and voice feedback data, to generate the original interaction stream, specifically includes: Monitor and identify target user interaction scenarios, and capture multimodal data of real-time user interactions; The captured multimodal data is classified to distinguish between visual and auditory data, and the data is parsed to extract visual attention data and voice feedback data. The parsed data is merged and standardized to generate a structured original interactive stream.

3. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 2, characterized in that, The process involves multi-dimensional fusion analysis of the original interaction flow, utilizing environmental context information to parse dependencies between different dimensions, identifying implicit associations between user interest patterns and virtual digital human performance characteristics, and labeling the analysis results as initial interest indicators to generate a confidence-labeled interest-labeled dataset. Specifically, this includes: The original interaction flow is organized according to the interaction dimension, and environmental context information is extracted from the data; Dependency analysis is performed between different interaction dimensions to identify normal and abnormal interest patterns. Abnormal interest patterns are labeled, initial interest indicators are generated, and interest type, intensity, and persistence are recorded. Based on the feature scores of each interest metric, confidence is configured to generate an interest-labeled dataset with confidence scores. All analysis and labeling results are then integrated to output a structured interest-labeled dataset.

4. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 3, characterized in that, The construction of a dynamic interest evolution model based on the original interaction flow and interest annotation dataset, the optimization of the model's adaptability using fusion analysis results, and the dynamic updating of the model state specifically include: Feature extraction is performed on the original interaction flow and interest annotation dataset to construct a dynamic interest evolution model, and the dynamic interest evolution model is trained using the extracted features; The parameters and structure of the model are dynamically adjusted based on the real-time fusion analysis results.

5. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 4, characterized in that, The process involves matching real-time collected user interaction data with a dynamic interest evolution model to identify matching recommendation candidates and mark the recommendation timing. The matching results are then fed back to the virtual digital human resource library for parsing to obtain the recommendation context. Specifically, this includes: The real-time user interaction data is input into the dynamic interest evolution model. The dynamic interest evolution model is used to match the real-time input data, analyze the consistency between the real-time data and the expected interest pattern data, and identify potential recommendation candidates. The detected matching recommendation candidates are labeled, and the degree of matching and relevant environmental context information are recorded; Prioritize the recommended candidates and assign a recommendation level to each candidate based on matching strength and frequency; Extract the matching results and feed them back to the virtual digital human resource library. Query the virtual digital human resource library and read the corresponding recommendation context.

6. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 5, characterized in that, The matching results based on the dynamic interest evolution model are used to mine the virtual digital human resource database, identify potential deep recommendation strategies and adjustment signals, and trace the recommendation strategies to identify core factors and optimization directions, specifically including: Collect virtual digital human resource library data related to the matching results and obtain complete environmental context information; Mining data from the virtual digital human resource database to identify potential deep recommendation strategies and adjustment signals, extracting potential strategy clues from the resource database, and identifying strategy types; The identified adjustment signals are traced and analyzed to identify the generation path of the adjustment signals and trace the original source of the adjustment; A comprehensive report is generated based on all the identified information.

7. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 1, characterized in that, The method further includes: Perform time series analysis on the initial interest metrics to capture changes in interest trends and update the interest annotation dataset; The dynamic interest evolution model is further optimized using the updated interest-labeled dataset, forming a closed-loop feedback mechanism.

8. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 7, characterized in that, The step of performing time-series analysis on the initial interest metrics, capturing changes in interest trends, and updating the interest annotation dataset specifically includes: Calculate the slope of the initial interest index over time to identify trends of rising or falling interest. Adjust the confidence weights in the interest-labeled dataset based on the slope metric, and regenerate the updated interest-labeled dataset.

9. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 8, characterized in that, The step of further optimizing the dynamic interest evolution model using the updated interest-labeled dataset to form a closed-loop feedback mechanism specifically includes: The updated interest-labeled dataset was used as additional training data to retrain the dynamic interest evolution model; Adjust the recommended candidate recognition threshold during the real-time matching process based on the output of the retrained model.

10. The personalized recommendation method for virtual digital humans in corporate promotion according to claim 9, characterized in that, The adjustment of the recommended candidate identification threshold in the real-time matching process specifically includes: Based on the performance metrics of the retrained model, the optimal value of the recommendation candidate recognition threshold is dynamically calculated. Applying optimal values ​​to the real-time matching process ensures the accuracy of recommendation timing annotations.

Images