User operation recommendation method under AI platform

Abstract

The application provides a user operation recommendation method under an AI platform, and belongs to the technical field of large language models.The application acquires multi-dimensional operation data of a user in real time through a distributed data acquisition system, processes page operation records by using a clustering compression algorithm, extracts inquiry content semantic features and historical interaction behavior patterns by using natural language processing and sequence identification algorithms, constructs a user operation intention graph, calculates the correlation strength by using a shortest path algorithm, accurately models the user intention, establishes a multi-modal fusion recommendation model based on a Transformer architecture, constructs an online learning system, calculates the user satisfaction degree by using an intention correlation degree function, dynamically adjusts the model parameters in real time by using a gating weight dynamic adjustment function, and combines offline precalculation and online retrieval optimization to ensure the recommendation accuracy and response speed, thereby solving the technical problem that the user operation recommendation of the AI platform is not accurate enough.

Description

Technical Field

[0001] This invention belongs to the field of large language model technology, and more specifically, relates to a user operation recommendation method under an AI platform. Background Technology

[0002] User action recommendation technology on AI platforms primarily relies on recommendation algorithms based on collaborative filtering and content filtering. These algorithms generate recommendations by analyzing users' historical behavior data and content feature similarity. Traditional recommendation systems typically employ a single-modal feature extraction approach, processing user behavior data, query content, and historical interaction records separately, then using machine learning models to predict the user's likely next action. In AI platform applications, these recommendation systems are widely used in intelligent assistants, knowledge-based question-answering systems, and human-computer interfaces to guide users through complex tasks and improve user experience. However, traditional recommendation methods have significant drawbacks. These include incomplete feature extraction, failing to effectively integrate multi-dimensional information such as user action records, query content, historical interaction data, and user profiles. Existing technologies often use static recommendation models, lacking a deep understanding of user intent and dynamic modeling capabilities, making it difficult to accurately capture the correlation between current actions and historical behaviors. Furthermore, traditional methods lack real-time feedback mechanisms and adaptive optimization capabilities, failing to dynamically adjust model parameters based on user satisfaction and recommendation performance. In existing technologies, due to the lack of multimodal feature fusion mechanisms, user operation intent modeling methods, and real-time feedback optimization mechanisms, recommendation systems cannot accurately understand users' true operation needs and intentions, resulting in a significant deviation between recommendation results and user expectations. In other words, existing technologies suffer from the technical problem of inaccurate user operation recommendations on AI platforms. Summary of the Invention

[0003] In view of this, the present invention provides a user operation recommendation method under an AI platform, which can solve the technical problem that the user operation recommendation of the AI ​​platform is not accurate enough in the prior art.

[0004] This invention is implemented as follows: It provides a user operation recommendation method under an AI platform, comprising: using a distributed data acquisition system to acquire user page operation records, query content, historical interaction data, and user profile information in real time; performing clustering and compression algorithms on the page operation records to remove duplicate and erroneous records; using natural language processing technology to segment and extract semantic features from the query content; simultaneously employing a sequence identification algorithm to perform time-series analysis on historical interaction data to extract user behavior patterns; constructing a user operation intent graph, using the user's current operation, historical behavior, and query content as graph nodes; calculating the association strength between nodes using a shortest path algorithm to find the optimal operation path; and integrating the page operation records, semantic feature vectors, and historical interaction data... A multimodal fusion recommendation model is constructed by inputting mutual data and user profile information. An attention mechanism and a sequence length adaptive optimization function are used for feature fusion and recommendation generation. An online learning system is built to receive user feedback data in real time. User operation satisfaction scores are calculated using an intent relevance function, and gating weights are used to dynamically adjust model parameters when scores fall within different ranges. When a user performs an operation, multi-dimensional information is acquired in real time, and recommendations for the next operation are quickly generated through feature fusion and model prediction. Offline pre-computation and online retrieval optimization functions, along with an approximation algorithm coupling optimization function, are used to improve the speed of recommendation result generation. Recommendations are ranked and filtered based on their relevance and importance, and the most suitable operation recommendation is fed back to the user in real time.

[0005] Specifically, the clustering compression algorithm uses K-means clustering technology to group user page operation records, grouping similar operation records into the same cluster, and extracting cluster centers as representative operation records to achieve data compression. At the same time, it identifies and removes duplicate and erroneous records, reducing storage space occupation and subsequent data processing time. The algorithm calculates similarity distance based on features such as operation type, timestamp, page position, and user identifier.

[0006] Specifically, the sequence identification algorithm marks and encodes historical interaction data in time series. It extracts user behavior patterns by analyzing the time intervals, frequencies, and sequence patterns of user operations, transforming complex time series data into structured feature representations. This provides basic data support for subsequent sequence analysis and feature extraction. The algorithm can identify the periodic and trend characteristics of user behavior patterns.

[0007] Specifically, the shortest path algorithm calculates the minimum weighted path between any two graph nodes in the user's operation intent graph. It then uses Dijkstra's algorithm to traverse the graph structure to find the optimal operation path from the current operation state to the target operation state. The path weight is calculated based on the frequency of the user's historical behavior and the operation success rate. The shortest path sequence output by the algorithm serves as the priority basis for recommended operations.

[0008] Specifically, the multimodal fusion recommendation model is a deep neural network built on the Transformer architecture. The model includes a multi-head attention mechanism to process feature information from different modalities, and achieves feature fusion and recommendation generation through an encoder-decoder structure. The sequence length parameter of the model is dynamically adjusted according to the sequence length adaptive optimization function. The model also integrates a positional encoding mechanism to maintain the temporal relationship of the operation sequence.

[0009] Specifically, the sequence length adaptive optimization function is used to dynamically determine the optimal input sequence length of the multimodal fusion recommendation model. The function intelligently calculates the most suitable sequence length parameter for the current recommendation scenario by comprehensively analyzing multiple dimensions such as user operation complexity score, historical interaction depth index, current task type encoding, system resource utilization rate and user response time preference. The function adopts a weighted fusion strategy to combine and calculate each input parameter according to its importance.

[0010] The intent relevance function is a comprehensive evaluation index calculated based on user operation history, current behavior, and feedback information. The function comprehensively considers data from multiple dimensions such as operation success rate, user dwell time, click conversion rate, and task completion status, and calculates a numerical score reflecting user operation satisfaction through weighted summation. The score range is between 0 and 1.

[0011] Specifically, the dynamic adjustment function for the gating weights is used to adjust the gating mechanism parameters of the neural network. The function calculates an iterative effect index based on multiple data such as user satisfaction feedback data, model prediction accuracy, feature importance score, historical adjustment effect, and current recommendation performance. When the iterative effect index falls within different ranges, different weight adjustment functions are used to adjust the gating parameters. The function achieves adaptive optimization of the gating parameters by dynamically monitoring the changing trend of the iterative effect index.

[0012] Specifically, the offline pre-computation and online retrieval optimization function is used to improve the response speed and computational efficiency of the multimodal fusion recommendation model. The function divides the recommendation calculation process into two stages: offline pre-computation and online fast retrieval. In the offline stage, common user operation patterns, high-frequency query content, and typical interaction sequences are pre-computed in batches and the results are stored in a high-performance caching system. In the online stage, the pre-computation results are quickly retrieved and adjusted in real time in combination with the current user status to generate the final recommendation.

[0013] Specifically, the approximation algorithm coupling optimization function is used to reduce the computational complexity of large-scale user data processing. The function constructs an efficient approximation calculation framework based on locality-sensitive hashing and random sampling techniques. The function includes a data point characteristic coupling term, which comprehensively considers multiple dimensions such as outlier identification, equilibrium point location, representative point selection, and mutation point detection. By weighted combination of data points with different characteristics, a comprehensive index reflecting the data distribution characteristics is obtained.

[0014] Specifically, the distributed data acquisition system uses multiple data acquisition nodes to work in parallel, acquiring and transmitting various operation data of users on the AI ​​platform in real time, ensuring the timeliness and integrity of data acquisition. The system uses a load balancing algorithm to allocate data acquisition tasks to avoid single-node overload, and establishes a data backup mechanism to ensure the reliability and persistent storage of data.

[0015] Specifically, the attention mechanism calculates the relevance score between the query matrix, key matrix, and value matrix, enabling the model to focus on the feature information most relevant to the current operation intent. The calculation of attention weights takes into account the timeliness and importance of user operations, giving higher attention weights to recent operations and allocating additional attention resources to key operation path nodes.

[0016] Furthermore, the structure of the multimodal fusion recommendation model is specifically a deep neural network architecture containing a 12-layer Transformer encoder and a 6-layer decoder. Each encoder layer contains a multi-head self-attention mechanism and a feedforward neural network, with 8 attention heads and a hidden layer dimension of 512. The model input layer receives preprocessed multimodal feature vectors, and the embedding layer maps different types of features to a unified vector space. The encoder is responsible for learning the complex relationships between features, and the decoder generates a recommendation operation sequence based on the encoding results.

[0017] Before establishing the training dataset for the multimodal fusion recommendation model, the process also includes collecting large-scale AI platform user operation log data as basic training samples. The data spans at least 6 months to ensure the diversity and representativeness of the samples. The original log data is cleaned and standardized to remove abnormal operation records and invalid data entries. User operation sequences, query content, page interaction behaviors, and final operation results are extracted as input-output pairs for the training samples.

[0018] The training steps of the multimodal fusion recommendation model specifically involve using the Adam optimizer to optimize model parameters, setting the initial learning rate to 0.0001 and the batch size to 32. The training process is divided into two stages: pre-training and fine-tuning. In the pre-training stage, self-supervised learning is performed using large-scale unlabeled user operation data, learning the inherent rules and patterns of user operations through a masked language model task. In the fine-tuning stage, supervised learning is performed using labeled user operation-recommendation data to optimize the model's recommendation accuracy and relevance.

[0019] Among them, the user operation complexity score is a numerical indicator reflecting the complexity of the operation, which is calculated by analyzing factors such as the number of steps in the user's current operation, the number of functional modules involved, the length of operation time, and the number of decision branches. The value ranges from 0 to 1, and the larger the value, the more complex the operation. The historical interaction depth index is a numerical indicator reflecting the depth of user interaction behavior, which is derived from the statistical analysis of the user's past operation sequences. It is obtained by calculating parameters such as the average length, maximum length, and operation level depth of the user's historical operation sequences.

[0020] This invention addresses the technical problem of inaccurate user operation recommendations on AI platforms by constructing a multimodal fusion recommendation model and employing techniques such as user operation intent graph modeling, attention mechanism feature fusion, and online learning optimization. This invention overcomes the shortcomings of traditional recommendation methods in terms of incomplete feature extraction. Through a multimodal fusion recommendation model, it comprehensively processes user page operation records, semantic features of query content, historical interaction data, and user profile information, achieving deep fusion of multi-dimensional features. The user operation intent graph calculates node association strength using a shortest path algorithm, accurately modeling user operation intent and behavioral paths. The attention mechanism enables the model to focus on features most relevant to the current operation, significantly improving recommendation accuracy. The online learning system optimizes model parameters in real time through an intent association function and a dynamic adjustment mechanism for gating weights, ensuring continuous improvement of recommendation results. This invention solves the technical problem of inaccurate recommendations in existing technologies because it establishes a complete technical framework for multimodal feature fusion, deep user intent modeling, and real-time feedback optimization. It accurately understands user operation needs and intents, generating recommendation results that highly match user expectations, thus significantly improving recommendation accuracy. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0023] like Figure 1 The diagram shown is a flowchart of a user operation recommendation method under an AI platform provided by the present invention. This method includes the following steps:

[0024] S01. A distributed data acquisition system is used to acquire user page operation records, query content, historical interaction data and user profile information in real time, and the page operation records are processed by clustering compression algorithm to remove duplicate and erroneous records.

[0025] S02. Natural language processing technology is used to segment and extract semantic features from the query content. At the same time, a sequence identification algorithm is used to perform time series analysis on historical interaction data to extract user behavior patterns.

[0026] S03. Construct a user operation intent graph, using the user's current operation, historical behavior, and query content as graph nodes. Calculate the association strength between nodes using the shortest path algorithm to find the optimal operation path.

[0027] S04. Input the page operation records, semantic feature vectors, historical interaction data and user profile information into the multimodal fusion recommendation model, and use the attention mechanism and sequence length adaptive optimization function to perform feature fusion and recommendation generation;

[0028] S05. Construct an online learning system to receive user feedback data in real time, use the intent correlation function to calculate the user operation satisfaction score, and use the gating weight dynamic adjustment function to update the model parameters when the score falls into different ranges.

[0029] S06. When the user performs an operation, multi-dimensional information is acquired in real time, and the next operation recommendation is quickly generated through the feature fusion and model prediction process. The offline pre-computation and online retrieval optimization function and the approximation algorithm are coupled optimization functions to improve the speed of recommendation result generation.

[0030] S07. Based on the relevance and importance of the recommendation results, sort and filter the recommended content, and provide the most suitable operation recommendation to the user in real time, forming a complete recommendation closed-loop system.

[0031] Among them, the distributed data acquisition system works in parallel through multiple data acquisition nodes to acquire and transmit various operation data of users on the AI ​​platform in real time, ensuring the timeliness and integrity of data acquisition. The system uses a load balancing algorithm to allocate data acquisition tasks to avoid the occurrence of single node overload, and at the same time establishes a data backup mechanism to ensure the reliability and persistent storage of data.

[0032] Among them, the clustering compression algorithm uses K-means clustering technology to cluster user page operation records, grouping similar operation records into the same cluster, and extracting cluster centers as representative operation records to achieve data compression. At the same time, it identifies and removes duplicate and erroneous records, reducing storage space occupation and subsequent data processing time. The algorithm calculates similarity distance based on features such as operation type, timestamp, page position and user identifier, and dynamically adjusts the number of clusters to achieve the best compression effect.

[0033] Among them, the sequence identification algorithm marks and encodes historical interaction data of time series, extracts user behavior patterns by analyzing the time interval, frequency and sequence pattern of user operations, and transforms complex time series data into structured feature representations, providing basic data support for subsequent sequence analysis and feature extraction. The algorithm can identify the periodic and trend characteristics of user behavior patterns.

[0034] The shortest path algorithm calculates the minimum weight path between any two graph nodes in the user's operation intent graph. It then uses Dijkstra's algorithm to traverse the graph structure to find the optimal operation path from the current operation state to the target operation state. The path weight is calculated based on the frequency of the user's historical behavior and the operation success rate. The shortest path sequence output by the algorithm is used as the priority basis for recommended operations.

[0035] Among them, the multimodal fusion recommendation model is a deep neural network built on the Transformer architecture. The model includes a multi-head attention mechanism to process feature information of different modalities. Feature fusion and recommendation generation are achieved through an encoder-decoder structure. The sequence length parameter of the model is dynamically adjusted according to the sequence length adaptive optimization function. When processing simple operations, the sequence length is set to a smaller value to improve processing speed. When processing complex operations, the sequence length is increased to obtain richer contextual information. The model also integrates a position encoding mechanism to maintain the temporal relationship of the operation sequence.

[0036] The attention mechanism calculates the relevance score between the query matrix, key matrix, and value matrix, enabling the model to focus on the feature information most relevant to the current operation intent. The calculation of attention weights takes into account the timeliness and importance of user operations, giving higher attention weights to recent operations and allocating additional attention resources to key operation path nodes, thereby improving the accuracy of feature fusion and the relevance of recommendation results.

[0037] The sequence length adaptive optimization function is used to dynamically determine the optimal input sequence length for the multimodal fusion recommendation model. This function intelligently calculates the most suitable sequence length parameter for the current recommendation scenario by comprehensively analyzing multiple dimensions such as user operation complexity score, historical interaction depth index, current task type code, system resource utilization, and user response time preference. Inputs include the user operation complexity score derived from the analysis of the user's current operation, the historical interaction depth index obtained from historical interaction data, the current task type code generated by the task classification module, the system resource utilization provided by the system monitoring module, and the user response time preference derived from user behavior analysis. The output is the optimized sequence length value used by the multimodal fusion recommendation model. This value is dynamically adjusted between 64 and 512. The function uses a weighted fusion strategy to combine and calculate each input parameter according to its importance. When the user operation complexity score is high and the system resource utilization is sufficient, a longer sequence length is preferred to obtain richer contextual information. When the user response time preference is high or the system resource utilization is tight, a shorter sequence length is chosen to ensure the real-time performance and smoothness of the recommendation.

[0038] The intent relevance function is a comprehensive evaluation index calculated based on user operation history, current behavior, and feedback information. The function comprehensively considers data from multiple dimensions such as operation success rate, user dwell time, click conversion rate, and task completion status. It calculates a numerical score reflecting user operation satisfaction through weighted summation. The score ranges from 0 to 1. When the score is greater than 0.8, it indicates that the user is highly satisfied, and a reinforcement learning weight adjustment function is used to increase the weight of relevant features. When the score is between 0.5 and 0.8, it indicates that the user is moderately satisfied, and a balanced weight adjustment function is used to maintain the current parameter settings. When the score is less than 0.5, it indicates that the user is dissatisfied, and a penalty weight adjustment function is used to decrease the weight of relevant features and explore new recommendation strategies.

[0039] The dynamic adjustment function for gating weights is used to adjust the gating mechanism parameters of the neural network. This function calculates an iterative effect index based on multiple data points, including user satisfaction feedback, model prediction accuracy, feature importance scores, historical adjustment effects, and current recommendation performance. Different weight adjustment functions are used to adjust the gating parameters when the iterative effect index falls within different ranges. The iterative effect index uses 0.25, 0.5, and 0.75 as threshold separators to divide the adjustment strategy into four levels. When the index is less than 0.25, it indicates that the current gating parameter effect is poor and requires significant adjustment; an aggressive adjustment function is used to increase the magnitude of the gating weight change and introduce random perturbations to explore new parameter spaces. When the index is between 0.25 and 0.5, it indicates that the effect is average and requires moderate adjustment; a mild adjustment function is used to slightly increase or decrease the gating weight while maintaining consistency in the adjustment direction. When the index is at 0... A value between 5 and 0.75 indicates good performance requiring only minor adjustments. A fine-grained adjustment function is used to optimize the gating weights to further improve performance. When the index is greater than 0.75, it indicates that the current parameters have reached a relatively optimal state. A conservative adjustment function is used to maintain the existing weight settings and make minor adjustments only when necessary. The inputs of the function include user satisfaction feedback data collected by the online learning system, model prediction accuracy calculated by the model evaluation module, feature importance scores generated by the feature analysis module, historical adjustment effects obtained from parameter adjustment history, and current recommendation performance obtained from recommendation effect evaluation. The output is the adjusted gating weight parameters for use by the multimodal fusion recommendation model. The function achieves adaptive optimization of the gating parameters by dynamically monitoring the changing trend of the iterative effect index, ensuring that the neural network can continuously improve the effectiveness of the gating mechanism based on the actual recommendation effect.

[0040] The offline pre-computation and online retrieval optimization functions are used to improve the response speed and computational efficiency of the multimodal fusion recommendation model. The function divides the recommendation calculation process into two stages: offline pre-computation and online fast retrieval. In the offline stage, common user operation patterns, high-frequency query content, and typical interaction sequences are pre-computed in batches and the results are stored in a high-performance caching system. In the online stage, the pre-computation results are retrieved quickly and adjusted in real time in combination with the current user status to generate the final recommendation. The inputs include the user's historical operation feature vector obtained from user behavior analysis, the current operation context information obtained from current operation monitoring, the pre-computation result index generated by offline calculation, the real-time adjustment parameters provided by the real-time adjustment module, and the cache hit rate statistics obtained from cache monitoring. The output is the optimized recommendation results and the corresponding confidence score for the recommendation result ranking. The function adopts a multi-level caching strategy to store pre-computation results of different frequencies in the corresponding cache levels. High-frequency results are stored in memory cache to achieve millisecond-level access, medium-frequency results are stored in solid-state disk cache to ensure second-level response, and low-frequency results are stored in distributed storage system to provide minute-level service. The pre-computation results are refreshed regularly through an intelligent cache update mechanism to ensure the timeliness and accuracy of the recommendation.

[0041] The approximation algorithm coupling optimization function is used to reduce the computational complexity of large-scale user data processing. This function constructs an efficient approximation calculation framework based on locality-sensitive hashing and random sampling techniques. The function includes a data point characteristic coupling term, which comprehensively considers multiple dimensions such as outlier identification, equilibrium point location, representative point selection, and mutation point detection. It calculates a comprehensive index reflecting the data distribution characteristics by weighted combination of data points with different characteristics. Inputs include the original user operation dataset provided by the distributed data acquisition system, the approximation accuracy requirements determined by the system configuration, computational resource constraints derived from resource monitoring, data distribution characteristic parameters obtained from data analysis, and the coupling weight configuration determined by the algorithm configuration. The output is a simplified dataset after approximation and the corresponding error estimate for subsequent recommendation calculations. The function adopts an adaptive sampling strategy to adjust the sampling frequency according to the data density distribution. In dense data areas, the sampling density is reduced to improve efficiency, and in sparse data areas, the sampling density is increased to ensure coverage. The data point characteristic coupling term identifies outliers and assigns them lower weights through Mahalanobis distance and kernel density estimation methods. The equilibrium point is determined by cluster center calculation and assigned a medium weight. Representative points are selected and assigned higher weights through information entropy and variance analysis. Variants are detected by time series change rate and a dynamic weight adjustment strategy is adopted. The entire function significantly reduces computational complexity and resource consumption while ensuring recommendation quality through multi-level approximation calculation and coupling optimization.

[0042] The multimodal fusion recommendation model is structured as a deep neural network architecture comprising a 12-layer Transformer encoder and a 6-layer decoder. Each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network, with 8 attention heads and a hidden layer dimension of 512. The model input layer receives preprocessed multimodal feature vectors, and the embedding layer maps different types of features to a unified vector space. The encoder is responsible for learning the complex relationships between features, and the decoder generates recommendation operation sequences based on the encoding results. The sequence length parameter of the model is dynamically adjusted between 64 and 512 based on the output of the sequence length adaptive optimization function, and is optimized according to the complexity of user operations and computational resources. When the user performs a simple query operation, the sequence length is set to 64 to improve response speed, and when the user performs a complex task operation, the sequence length is expanded to 512 to obtain sufficient contextual information. The model also integrates residual connections and layer normalization mechanisms to improve training stability and convergence speed.

[0043] The steps for establishing the training dataset for the multimodal fusion recommendation model include collecting large-scale AI platform user operation log data as basic training samples, with a data time span covering no less than 6 months to ensure sample diversity and representativeness; cleaning and standardizing the original log data to remove abnormal operation records and invalid data entries; extracting user operation sequences, query content, page interaction behaviors, and final operation results as input-output pairs for training samples; labeling the samples based on the success rate and user operation satisfaction rating, marking high-quality operation sequences as positive samples and low-quality or failed operation sequences as negative samples; performing stratified sampling for different types of operation scenarios to ensure the balance of training data; and constructing validation and test datasets. The validation set accounts for 15% of the total data volume for model parameter tuning and performance evaluation, while the test set accounts for 10% of the total data volume for final model effect verification. All datasets have been anonymized to protect the security of user privacy information.

[0044] The training steps of the multimodal fusion recommendation model include optimizing model parameters using the Adam optimizer, with an initial learning rate of 0.0001 and a batch size of 32. The training process is divided into two stages: pre-training and fine-tuning. In the pre-training stage, self-supervised learning is performed using large-scale unlabeled user operation data. The model learns the inherent rules and patterns of user operations through a masked language model task. The pre-training period is 50 epochs. In the fine-tuning stage, supervised learning is performed using labeled user operation-recommendation data to optimize the model's recommendation accuracy and relevance. The fine-tuning period is 20 epochs. During training, a learning rate decay strategy is adopted, halving the learning rate every 10 epochs to improve the stability of model convergence. Gradient clipping is used to prevent gradient explosion. An early stopping mechanism is set to monitor the validation set performance. Training is terminated early when the validation loss no longer decreases after 5 consecutive epochs to avoid overfitting. The entire training process is performed on a distributed image processing unit cluster, utilizing data parallelism and model parallelism to accelerate training efficiency. After training, the model is quantized and compressed to reduce the computational resource consumption during inference.

[0045] User operation complexity score is a numerical index that reflects the complexity of an operation by comprehensively calculating factors such as the number of steps in the user's current operation, the number of functional modules involved, the length of operation time, and the number of decision branches. The value ranges from 0 to 1, with a larger value indicating a more complex operation.

[0046] The historical interaction depth index is a numerical value that reflects the depth of user interaction behavior, derived from statistical analysis of users' past operation sequences. It is obtained by calculating parameters such as the average length, maximum length, and operation level depth of users' historical operation sequences.

[0047] The current task type encoding refers to the encoding method that maps different types of tasks that a user is performing to numerical identifiers. Through task classification algorithms, different task types such as query, creation, analysis, and interaction are converted into corresponding numerical codes.

[0048] System resource utilization rate refers to the proportion of hardware resources such as image processing unit, memory, central processing unit and storage device used in the current computing system, which is collected and calculated in real time by the system monitoring module;

[0049] User response time preference refers to the user's expected value of system response speed based on the analysis of the user's historical behavior patterns. It is calculated by statistically analyzing the user's waiting time tolerance, operation frequency, and task switching patterns.

[0050] The Iteration Effect Index is a comprehensive evaluation index that reflects the effect of model parameter adjustment. It is calculated by integrating multiple dimensions of data, including user satisfaction feedback data, model prediction accuracy, feature importance score, historical adjustment effect and current recommendation performance, and its value ranges from 0 to 1.

[0051] The data point characteristic coupling term refers to the weight calculation term used in approximation algorithms to comprehensively consider data points with different characteristics. By performing characteristic analysis and weight allocation on different types of data points such as outliers, equilibrium points, representative points, and variant points, a comprehensive index reflecting the data distribution characteristics is formed.

[0052] The specific implementation methods of the above steps are described in detail below.

[0053] The specific implementation of step S01 involves constructing a multi-node distributed data acquisition architecture. A load balancing strategy distributes data acquisition tasks to different nodes. Each node monitors user actions such as page clicks, scrolling, input, and dwell time, while simultaneously collecting user query texts, historical dialogue records, and personal preference tags. An event-driven mechanism captures user actions in real time and transmits the data to the central processing unit. Timestamp synchronization technology is used during data transmission to ensure the accuracy of the timing of the action records. The collected page action records are grouped using a K-means clustering algorithm. The algorithm calculates Euclidean distance based on features such as action type, occurrence time, page position, and user identity, grouping actions with a similarity distance less than 0.3 into the same cluster. Data compression is achieved by extracting the cluster centroids as representative action records. Duplicate records and abnormal actions are identified and removed from the dataset. This entire process can compress the original data volume to approximately 60%.

[0054] The specific implementation of step S02 involves using a Chinese word segmentation algorithm to segment the user's query content into words, employing a term frequency-inverse document frequency algorithm to extract keywords and construct semantic feature vectors, with the vector dimension set to 256 dimensions to balance expressive power and computational efficiency. A pre-trained word embedding model maps the segmentation results to a continuous vector space, and cosine similarity is used to calculate the semantic association strength between words. Sequence identification technology is applied to encode historical interaction data. The algorithm analyzes the time interval distribution, frequency change trends, and operation sequence patterns of user operations, extracting operation subsequences of length 10 through a sliding window mechanism to identify periodic patterns and habitual preferences of user behavior, converting the time-series data into a structured feature representation for subsequent analysis.

[0055] The specific implementation of step S03 involves constructing a directed weighted graph structure to represent the user's operational intent relationships. Nodes in the graph include the current operational state, historical behavior records, and query semantic information. Edge weights are calculated based on operational transition probabilities and user preference strength. The Dijkstra shortest path algorithm is used to traverse the graph structure. The algorithm starts from the current operational node and gradually expands to adjacent nodes, calculating cumulative path weights until the target state is reached. Path weights comprehensively consider factors such as operational success rate, user historical selection frequency, and time efficiency. Paths with a success rate greater than 0.8 have a weight of 1.0, paths with a success rate between 0.5 and 0.8 have a weight of 1.5, and paths with a success rate less than 0.5 have a weight of 2.0. The algorithm outputs the operational path sequence with the minimum weight as the recommendation priority, providing users with optimal operational guidance.

[0056] The specific implementation of step S04 involves inputting page operation records, semantic feature vectors, historical interaction data, and user profile information into a multimodal fusion recommendation model based on the Transformer architecture. The model employs a multi-head attention mechanism to process data from different modalities. It generates an attention weight distribution by calculating the correlation between the query matrix, key matrix, and value matrix. The weight coefficient for recent operations is set to 1.2 times, and the weight coefficient for key path nodes is set to 1.5 times, to highlight the impact of important information on the recommendation results. The sequence length adaptive optimization function dynamically determines the optimal sequence length based on parameters such as user operation complexity score, historical interaction depth index, task type encoding, system resource utilization, and response time preference. When the complexity score is higher than 0.7 and system resources are sufficient, the sequence length is set to 512; when the complexity score is lower than 0.3 or system resources are scarce, the sequence length is set to 64; and under medium complexity conditions, the sequence length is set to an adaptive value between 128 and 256.

[0057] The specific implementation of step S05 involves establishing an online learning feedback mechanism to collect real-time user response data to the recommendation results, including click behavior, dwell time, task completion status, and subjective evaluation. The intent correlation function integrates data from dimensions such as operation success rate, user dwell time, click conversion rate, and task completion status, calculating a user operation satisfaction score through a weighted summation method. The weights for operation success rate, dwell time, click conversion rate, and task completion status are set to 0.3, 0.2, and 0.2 respectively. When the satisfaction score is greater than 0.8, a reinforcement learning strategy is used to increase the weights of relevant features by 1.2 times. When the score is between 0.5 and 0.8, the existing parameter settings are maintained. When the score is less than 0.5, the weights of relevant features are reduced to 0.8 times, and a new strategy exploration mechanism is initiated. The gating weight dynamic adjustment function adopts corresponding adjustment strategies according to different ranges of the iterative effect index. When the index is less than 0.25, aggressive adjustment is adopted to increase the weight change range to 20%. When the index is between 0.25 and 0.5, mild adjustment is adopted to control the change range within 10%. When the index is between 0.5 and 0.75, fine adjustment is adopted to control the change range within 5%. When the index is greater than 0.75, conservative adjustment is adopted to maintain the existing settings.

[0058] The specific implementation of step S06 involves a hybrid optimized architecture combining offline pre-computation and online retrieval. In the offline phase, batch pre-computation is performed on high-frequency user operation patterns, common query content, and typical interaction sequences. The calculation results are then stored hierarchically in a multi-level caching system according to access frequency. High-frequency results are stored in a memory cache for millisecond-level access; results accessed more than 100 times per hour are categorized into this category. Medium-frequency results are stored in a solid-state drive cache to ensure second-level response; results accessed between 10 and 100 times per hour are categorized into this category. Low-frequency results are stored in a distributed storage system to provide minute-level service. In the online phase, pre-computation results are quickly retrieved and adjusted in real-time based on the current user status. When the cache hit rate reaches 85% or higher, the pre-computation result is returned directly; when the hit rate is below 85%, the real-time calculation process is initiated. The approximate algorithm coupling optimization function uses locality-sensitive hashing (LSH) to build an efficient computational framework. An adaptive sampling strategy is used to adjust the sampling frequency according to the data density distribution. The sampling rate is set to 10% in dense data regions and 30% in sparse data regions. The data point characteristic coupling term comprehensively considers the weight allocation of outliers, equilibrium points, representative points, and mutant points. The weight of outliers is set to 0.1, the weight of equilibrium points is set to 0.5, the weight of representative points is set to 0.8, and the weight of mutant points adopts a dynamic adjustment strategy.

[0059] The specific implementation of step S07 involves evaluating the relevance and importance of the generated recommendation results, and using a comprehensive ranking algorithm to arrange the recommended content according to priority. The relevance score is calculated based on the degree of matching between the user's current operational intent and the recommended content. Recommended content with a matching degree greater than 0.9 has a relevance score of 1.0, content with a matching degree between 0.7 and 0.9 has a score of 0.8, and content with a matching degree less than 0.7 has a score of 0.5. The importance score considers the criticality of the recommended operation to the completion of the user's task. Key operations have an importance score of 1.0, auxiliary operations have an importance score of 0.6, and optional operations have an importance score of 0.3. A comprehensive ranking score is obtained by weighted fusion of the relevance and importance scores, with a relevance weight of 0.6 and an importance weight of 0.4. The top 5 recommendations with the highest comprehensive scores are selected and fed back to the user in real time, forming a complete recommendation closed-loop system to ensure continuous optimization of the recommendation effect.

[0060] The multimodal fusion recommendation model employs a deep neural network design based on the Transformer architecture. Its overall structure includes an input embedding layer, a positional encoding layer, a 12-layer encoder, a 6-layer decoder, and an output prediction layer. The input embedding layer maps feature data from different modalities to a unified 512-dimensional vector space. It processes page operation records, semantic feature vectors, historical interaction data, and user profile information through linear transformations and activation functions, ensuring compatibility and consistency across different data types. The positional encoding layer uses sine and cosine functions to generate positional encoding vectors, incorporating the temporal relationship information of the operation sequence into the feature representation, thus preserving the temporal order of user operations.

[0061] The encoder consists of 12 stacked encoder layers with identical structures. Each encoder layer contains two main components: a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism uses eight attention heads, each with a dimension of 64, to capture complex relationships between features by computing attention weights in different subspaces in parallel. The feedforward neural network employs a two-layer fully connected structure, with the intermediate layer having a dimension of 2048 and the output layer maintaining a dimension of 512. The ReLU activation function is used to enhance the model's non-linear expressive power. Each encoder layer also integrates residual connections and layer normalization mechanisms. Residual connections alleviate the degradation problem of deep networks, while layer normalization accelerates model convergence and improves training stability.

[0062] The decoder consists of six decoder layers, each containing a masked multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network. The masked multi-head self-attention mechanism prevents information leakage during decoding and ensures the autoregressive properties of the prediction. The encoder-decoder attention mechanism fuses contextual information from the encoder output to guide recommendation generation. The output prediction layer transforms the decoder output into a probability distribution of recommendation operations through linear transformation and a softmax activation function, selecting the operation with the highest probability as the recommendation result.

[0063] The model's sequence length parameter is dynamically adjusted based on an adaptive optimization function, ranging from 64 to 512. For simple query operations, the length is set to 64 to improve response speed; for complex tasks, the length is expanded to 512 to obtain sufficient contextual information; and for medium-complexity operations, the length is set to an intermediate value of 128 or 256. The model has approximately 120 million parameters and achieves efficient training and inference through distributed training and model parallelism.

[0064] The process of building the training dataset includes multiple stages such as data collection, preprocessing, labeling, partitioning, and validation. In the data collection stage, large-scale user behavior data was extracted from the user operation log system of the AI ​​platform. The data time span was set to 8 months to ensure the seasonality and diversity coverage of the samples. The collected raw data includes information such as user click records, page browsing trajectory, query text content, interaction dialogue history, and final operation results, with a total data volume of approximately 5 million operation records.

[0065] The data preprocessing stage involves cleaning and standardizing the raw log data, deleting abnormal operation records such as records with excessively long dwell times exceeding 30 minutes, invalid requests submitted repeatedly, and abnormal data generated by system errors, desensitizing user identity information to ensure privacy and security, converting timestamps of different formats to a standard format, and filling missing fields with reasonable default values ​​or interpolation methods.

[0066] During the data annotation phase, samples were graded for quality based on user operation success rate and satisfaction score. Sequences with an operation success rate greater than 0.8 and a user satisfaction score greater than 0.7 were labeled as high-quality positive samples, sequences with operation failures or user satisfaction scores less than 0.3 were labeled as negative samples, and sequences in between were labeled as medium-quality samples. Annotation quality was ensured through a combination of manual annotation and automated evaluation. Manual annotation covered 10% of the samples as a baseline, while automated evaluation processed the remaining 90%.

[0067] During the data partitioning phase, the processed dataset is divided into a 75% training set, a 15% validation set, and a 10% test set. A stratified sampling strategy is employed to ensure a balanced distribution of different operation types and user groups across the datasets. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model selection, and the test set is used for final performance evaluation and effectiveness verification.

[0068] The data validation phase involves quality checks and consistency verification of the constructed dataset, calculating the statistical characteristics and distribution patterns of different modalities, detecting data bias and outliers, and ensuring the representativeness and validity of the training data. Cross-validation is then used to evaluate the stability and reliability of the dataset, ultimately establishing a high-quality dataset that meets the requirements for model training.

[0069] It should be noted that this invention integrates the following three key technical ideas: a sequence length adaptive optimization mechanism, a gating weight dynamic adjustment function, and a two-stage processing architecture of offline pre-computation and online retrieval optimization.

[0070] The effectiveness of the adaptive sequence length optimization mechanism lies in overcoming the limitations of traditional recommendation systems that use fixed sequence lengths. This mechanism intelligently determines the optimal sequence length parameter within the range of 64 to 512 by comprehensively analyzing multiple dimensions such as user operation complexity, historical interaction depth, task type, and system resource status. Compared to the computational resource waste or insufficient contextual information caused by the fixed sequence length in existing technologies, this adaptive mechanism can dynamically adjust according to actual needs. It ensures sufficient contextual information acquisition in complex operation scenarios while improving processing efficiency in simple operations, achieving a balanced optimization between recommendation accuracy and computational efficiency.

[0071] The effectiveness of the dynamic weight adjustment function is reflected in the construction of an adaptive learning mechanism based on multi-dimensional feedback. This function calculates the iterative effect exponent and sets multiple threshold separation points, achieving a gradient parameter update strategy from aggressive adjustment to conservative fine-tuning. Traditional methods often employ fixed parameter update methods, making it difficult to flexibly adjust based on actual results. This mechanism, however, can intelligently select appropriate weight adjustment strategies based on real-time feedback such as user satisfaction and model accuracy, enabling the neural network gating mechanism to continuously optimize and adapt to different recommendation scenario requirements.

[0072] The effectiveness of the two-stage processing architecture lies in resolving the conflict between real-time recommendation and computational complexity. By pre-computing common operation patterns and high-frequency interaction sequences offline, combined with online fast retrieval and real-time adjustment, this architecture avoids the efficiency bottleneck of traditional methods that require complete computation for each recommendation. The introduction of a multi-level caching strategy further optimizes the storage and access efficiency of data at different frequencies, achieving tiered service capabilities from milliseconds to minutes.

[0073] The synergistic effect of these three technical approaches forms an adaptive, efficient, and accurate recommendation system architecture. Adaptive sequence length optimization provides a flexible processing basis for gating adjustment, dynamic gating ensures continuous optimization of model parameters, and the two-stage processing architecture guarantees the real-time responsiveness of the overall system. This synergy enables the system to significantly improve computational efficiency and user experience while maintaining recommendation quality, achieving a technological leap in intelligence, adaptability, and high efficiency compared to traditional fixed-parameter, statically processed recommendation methods.

[0074] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of inaccurate user operation recommendations on AI platforms. Its core principle lies in constructing a multi-layered accuracy assurance mechanism. The logical basis of this mechanism is to systematically improve recommendation accuracy through three key technical links: multimodal feature fusion, accurate modeling of user intent, and real-time optimization feedback.

[0075] Multimodal feature fusion is fundamental to improving recommendation accuracy. Natural language processing (NLP) techniques are used to extract deep semantic features from query content, sequence labeling algorithms analyze temporal patterns in historical interaction data, and clustering compression algorithms process page operation records. These, along with user profile information, constitute a rich, multi-dimensional feature representation. The Transformer-based multimodal fusion recommendation model calculates the correlation between different modal features through a multi-head attention mechanism. An adaptive sequence length optimization function dynamically adjusts model parameters based on operational complexity, ensuring the sufficiency and effectiveness of feature fusion.

[0076] User intent graph modeling is a key technology for accurate recommendation. It constructs a graph structure from the user's current actions, historical behaviors, and queries, and calculates the strength of associations between nodes using a shortest path algorithm to accurately identify the user's intent and target path. An attention mechanism further enhances the ability to identify key action paths, enabling the model to precisely pinpoint the user's true needs.

[0077] A real-time optimization feedback mechanism ensures continuous improvement in recommendation accuracy. The online learning system calculates user operation satisfaction scores through an intent relevance function, and the gating weight dynamic adjustment function updates model parameters in real time based on user feedback and model performance indicators, forming a closed-loop optimization mechanism. The combination of offline pre-computation and online retrieval optimization, by pre-calculating common operation patterns and combining them with real-time adjustments, improves response speed while maintaining recommendation accuracy. This technical solution is logical because it forms a complete accuracy assurance chain from feature extraction and intent modeling to feedback optimization, with each step specifically designed to improve recommendation accuracy.

[0078] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0079] The specific implementation of step S01 involves using a distributed data acquisition system to acquire user page operation records, query content, historical interaction data, and user profile information in real time, and then processing the page operation records using a clustering compression algorithm. The similarity distance calculation of the K-means clustering compression algorithm is specifically represented as follows:

[0080]

[0081] In the formula, d ij x is the Euclidean distance between the i-th operation record and the j-th operation record; ik x is the k-th feature value of the i-th operation record; jk is the k-th feature value of the j-th operation record; n is the number of feature dimensions, with a value of 4, including operation type, timestamp, page position, and user identifier.

[0082] The specific calculation for cluster center updates is as follows:

[0083]

[0084] In the formula, c m S is the center point of the m-th cluster; m Let S be the set of operation records in the m-th cluster; m | represents the number of operation records in the cluster; x i Let be the feature vector of the i-th operation record in the cluster.

[0085] The parameter acquisition method is as follows: x ik The data is acquired in real-time, including the following steps: Step 1: Deploy a page listener to capture user actions such as clicking, scrolling, inputting, and hovering; Step 2: Record the timestamp and page coordinates of the actions; Step 3: Extract the user's identity and action type code. The value of n ranges from 3 to 6, determined based on the actual number of features.

[0086] The specific implementation of step S02 involves using natural language processing technology to segment and extract semantic features from the query content, while simultaneously employing a sequence labeling algorithm to perform time-series analysis on historical interaction data. The specific representation of the term frequency-inverse document frequency feature weight calculation is as follows:

[0087]

[0088] In the formula, w t，d The feature weights of word t in document d; tf t，dLet be the frequency of word t in document d; N be the total number of documents; df t d represents the number of documents containing the word t; t represents a specific word unit; d represents the document the user is querying.

[0089] The semantic similarity calculation is specifically represented as follows:

[0090]

[0091] In the formula, sim(v1, v2) is the cosine similarity between two word vectors; v1 and v2 are the corresponding word embedding vectors; ∥v1∥ and ∥v2∥ are the Euclidean norms of the vectors, respectively.

[0092] The time-series analysis calculation of user behavior pattern extraction is specifically represented as follows:

[0093]

[0094] In the formula, P pattern Δt represents the user behavior pattern feature value; m is the length of the operation sequence, with a value of 10; Δt i T represents the time interval for the i-th operation. total f represents the total time span. i s represents the frequency of the i-th operation; i Let α be the order weight of the i-th operation; i ,β i γ i The weight coefficient for the i-th position is 0.4, 0.3, and 0.3 respectively; i represents the position index in the operation sequence.

[0095] The parameter acquisition method is: tf t，d The data is obtained through statistical calculation, which involves traversing the document to count the frequency of words; N and df t v1 and v2 were obtained through document set statistics; v1 and v2 were obtained through a pre-trained word embedding model with a vector dimension of 256; Δt i Calculated by the timestamp difference in the user's operation log; i Calculated by counting the number of times each operation appears in historical data; s i Determined based on the position of the operation in the sequence.

[0096] The specific implementation of step S03 involves constructing a user action intent graph, using the user's current action, historical behavior, and query content as graph nodes, and calculating the association strength between nodes using a shortest path algorithm. The weight calculation of the Dijkstra shortest path algorithm is specifically represented as follows:

[0097] w edge =α×(1-p success )+β×(1-ffreq )+γ×t time ;

[0098] In the formula, w edge p represents the weight of the edge in the graph. success f represents the success rate of the operation, ranging from 0 to 1. freq Select the frequency for the user's history, with a value range of 0 to 1; t time is the operation time efficiency coefficient, with a value ranging from 0 to 1; β, β, and γ are weighting coefficients, with values ​​of 0.4, 0.3, and 0.3, respectively.

[0099] The specific representation of path distance accumulation is as follows:

[0100] dist[v]=min(dist[v],dist[u]+w uv );

[0101] In the formula, dist[v] is the shortest distance from the starting node to node v; dist[u] is the shortest distance from the starting node to node u; w uv Let be the edge weight from node u to node v; u and v represent the source node and the target node in the graph, respectively.

[0102] The parameter acquisition method is as follows: p success Historical statistics were used to calculate the success rate of each operation path; freq Obtained through user behavior log statistics; t time It is calculated based on the ratio of the operation completion time to the standard time.

[0103] The specific implementation of step S04 involves inputting the page operation records, semantic feature vectors, historical interaction data, and user profile information into the multimodal fusion recommendation model. The attention weight calculation is specifically represented as follows:

[0104]

[0105] In the formula, Q is the query matrix; K is the key matrix; V is the value matrix; d k is the dimension of the key vector, with a value of 64; softmax is the normalization exponential function; T represents the matrix transpose operation.

[0106] The specific calculation for adjusting the timeliness weight is as follows:

[0107]

[0108] In the formula, w time For timeliness weighting; λ decay t is the time decay coefficient, with a value of 0.1; elapsedThe elapsed time is in hours; e is the base of the natural logarithm.

[0109] The specific calculation of the sequence length adaptive optimization function is as follows:

[0110] L seq =λ1C complex +λ2D depth +λ3T type +λ4(1-R resource )+λ5P response +∈ adapt ;

[0111] In the formula, L seq C is the optimal sequence length. complex Score the complexity of user operations, with a value ranging from 0 to 1; D depth This is a historical interaction depth indicator, with a value range of 0 to 1; T type Encodes the current task type, with a value range of 0 to 1; R resource P represents the system resource utilization rate, with a value ranging from 0 to 1. response λ1 represents user response time preference, with a value ranging from 0 to 1; λ1, λ2, λ3, λ4, and λ5 are weight parameters, with values ​​of 0.25, 0.2, 0.2, 0.2, and 0.15 respectively; ∈ adapt This is an adaptive adjustment term, with a value range of -10 to 10.

[0112] The parameter acquisition method is as follows: C complex The data is obtained through calculation, which involves analyzing the number of operation steps, the number of functional modules, the operation time, and the number of decision branches; D depth Calculated by statistically analyzing the length characteristics of users' historical operation sequences; R resource The system monitoring module obtains real-time usage data for CPU, memory, and storage devices; elapsed Calculated by the difference between the current time and the time the operation occurred; ∈ adapt The range is -10 to 10.

[0113] The specific implementation of step S05 involves constructing an online learning system to receive user feedback data in real time. The intent relevance function is specifically represented as follows:

[0114]

[0115] In the formula, S satisfaction User satisfaction rating is assigned, with a value ranging from 0 to 1; p success For the success rate of the operation; t stay The actual time the user stays; t standard Standard stay time; r click Click-through rate; ccomplete The task completion score is determined by a value of 0 or 1; w1, w2, w3, and w4 are weighting coefficients with values ​​of 0.3, 0.2, 0.3, and 0.2 respectively; δ feedback This is a feedback adjustment term, with a value range of -0.1 to 0.1.

[0116] The iterative effect exponent calculation of the gated weight dynamic adjustment function is specifically expressed as follows:

[0117] I iteration =ξ1S satisfaction +ξ2A accuracy +ξ3F importance +ξ4H history +ξ5P performance +μ effect ;

[0118] In the formula, I iteration The iterative effect index, with a value range of 0 to 1; A accuracy To improve the model's prediction accuracy;

[0119] F importance For feature importance scoring; H history Score the historical adjustment effect; P performance The current recommended performance score is given; ξ1, ξ2, ξ3, ξ4, and ξ5 are weight parameters, with values ​​of 0.25, 0.2, 0.2, 0.2, and 0.15 respectively; μ effect This is the effect moderating term, with a value range of -0.05 to 0.05.

[0120] The parameter acquisition method is as follows: t stay and t standard Obtained through user behavior monitoring; click Calculated by analyzing user click behavior; A accuracy δ is calculated by comparing the model's predicted results with the actual results. feedback and μ effect Adjustments were made based on user feedback.

[0121] The specific implementation of step S06 involves acquiring multi-dimensional information in real time when the user performs an operation, and quickly generating a recommendation for the next operation through the feature fusion and model prediction process. The cache hit rate calculation of the offline pre-computation and online retrieval optimization functions is specifically represented as follows:

[0122]

[0123] In the formula, H cache N represents the cache hit rate. hit N represents the number of cache hits. total This represents the total number of queries.

[0124] The specific representation of the data point characteristic coupling term calculation for the approximate algorithm's coupled optimization function is as follows:

[0125] C coupling =v1W outlier +φ2W balance +φ3W representative +φ4W variant +η coupling ;

[0126] In the formula, C coupling For data point characteristic coupling terms; W outlier W represents the weight of outliers, with a value of 0.1. balance The balance point weight is set to 0.5; W representative The weight of the representative point is 0.8; W variant The weights for the mutation points are dynamically adjusted; φ1, φ2, φ3, and φ4 are coupling coefficients, with values ​​of 0.15, 0.25, 0.35, and 0.25 respectively; η coupling This is the coupling error term, with a value range of -0.02 to 0.02.

[0127] The parameter acquisition method is as follows: N hit and N total Obtained through system log statistics; W outlier Identification is performed using Mahalanobis distance; W representative Determined through information entropy and variance analysis; η coupling Adjust dynamically based on data distribution characteristics.

[0128] The specific implementation of step S07 involves ranking and filtering the recommended content based on the relevance and importance of the recommendation results. The comprehensive ranking score is calculated as follows:

[0129] Score final =ω1Score relevance +ω2Score importance +ζ ranking ;

[0130] In the formula, Score final Score is used for comprehensive ranking and scoring. relevance This is a relevance score, ranging from 0 to 1; importance The importance score ranges from 0 to 1; ω1 and ω2 are weighting coefficients, with values ​​of 0.6 and 0.4 respectively; ζ ranking This is a sorting adjustment item, with a value range of -0.05 to 0.05.

[0131] The specific calculation of the correlation score is as follows:

[0132]

[0133] In the formula, M match The degree of matching between the user's current action intent and the recommended content is obtained through semantic similarity calculation.

[0134] The importance score is calculated as follows:

[0135]

[0136] The parameter acquisition method is as follows: M match The similarity is obtained by calculating the cosine similarity between the user action vector and the recommended content vector; the action type classification is determined through task dependency analysis; ζ ranking Adjusted based on user's historical preferences.

[0137] To better understand and implement this invention, the following is a specific application scenario of the invention, embodiment 2: Researchers use the method of this invention to process vouchers on an intelligent consumer voucher recommendation platform. This platform mainly provides consumers with intelligent recommendation services for various types of consumer vouchers such as restaurant vouchers, shopping vouchers, entertainment vouchers, and transportation vouchers. The user group is mainly urban residents and office workers.

[0138] Researchers first deployed a distributed data collection system comprising 12 data collection nodes, each responsible for monitoring user activity on the consumer voucher platform. During the initial data collection phase, the system recorded real-time activity data from 2,500 users over four months, collecting approximately 5.2 million page operation records, 2.86 million voucher query records, and 12.5 million historical redemption and usage records. After processing using a K-means clustering compression algorithm, operations with a similarity distance less than 0.3 were grouped into the same cluster, ultimately compressing the original data to approximately 3.12 million valid records, a compression ratio of 60%.

[0139] In analyzing user activity logs, researchers found that the main user activity types included browsing coupons, filtering and searching, claiming coupons, usage records, and feedback. (See Table 1 for details.)

[0140] Table 1: Distribution Statistics of User Voucher Transaction Types

[0141] Operation type operation frequency percentage Average execution time (seconds) Coupon browsing 672400 21.5％ 28.6 Filter query 561300 18.0％ 45.3 Claiming Operation 452800 14.5％ 12.8 Merchant Search 593600 19.0％ 36.7 Usage Records 418200 13.4％ 158.4 Evaluation and feedback 421700 13.6％ 72.9

[0142] By analyzing user queries using natural language processing (NLP) techniques, researchers extracted 4267 keywords using a term frequency-inverse document frequency (IF-FCN) algorithm. High-frequency keywords included food deals, shopping discounts, movie tickets, and coffee delivery. The semantic feature vectors were set to 256 dimensions, and a pre-trained word embedding model mapped the query content to a continuous vector space. When analyzing historical interaction data using a sequence identification algorithm with an operation sequence length of 10, the extracted user behavior patterns showed that 73% of users exhibited clear consumption preferences and usage habits.

[0143] Researchers constructed a user action intent graph containing 2148 nodes, with node types including current coupon type preference, historical usage records, and query semantic information. When calculating the association strength between nodes using the Dijkstra shortest path algorithm, the weight coefficients α, β, and γ were set to 0.4, 0.3, and 0.3, respectively. In path weight calculation, paths with a coupon type usage success rate greater than 0.8 were assigned a weight of 1.0, paths with a success rate between 0.5 and 0.8 were assigned a weight of 1.5, and paths with a success rate less than 0.5 were assigned a weight of 2.0. After algorithm optimization, the average operation path length from query to redemption was shortened from the initial 5.4 steps to 3.7 steps, demonstrating significant path optimization performance, as shown in Table 2.

[0144] Table 2 Comparison of the Optimization Effect of Recommendation Path for Consumption Vouchers

[0145]

[0146]

[0147] The researchers constructed a multimodal fusion recommendation model using the Transformer architecture, comprising a 12-layer encoder and a 6-layer decoder, with 8 attention heads and a hidden layer dimension of 512. The model's training dataset contains 6.24 million training samples, with the training set comprising 75%, the validation set 15%, and the test set 10%. In the adaptive sequence length optimization, the sequence length is set to 512 when the user query complexity score is higher than 0.7 and system resources are sufficient, and to 64 when the complexity score is lower than 0.3 or system resources are limited.

[0148] The model was trained using the Adam optimizer with an initial learning rate of 0.0001 and a batch size of 32. The pre-training phase lasted 50 epochs, and the fine-tuning phase lasted 20 epochs. In the attention mechanism calculation, the weight coefficient for recently used records was set to 1.2 times, and the weight coefficient for frequently used coupons was set to 1.5 times. In the timeliness weight adjustment, the time decay coefficient λ... decay When set to 0.1, the weight of usage records exceeding 72 hours is reduced to 50% of the original value.

[0149] The online learning system developed by the researchers collects user feedback data in real time. The weight coefficients w1, w2, w3, and w4 of the intent-relatedness function are set to 0.3, 0.2, 0.3, and 0.2, respectively. When the user satisfaction score is greater than 0.8, the system uses a reinforcement learning strategy to increase the weight of relevant features by 1.2 times. When the score is between 0.5 and 0.8, the existing parameter settings are maintained. When the score is less than 0.5, the weight of relevant features is decreased to 0.8 times. The gating weight dynamic adjustment function adopts corresponding adjustment strategies according to different ranges of the iterative effect index. When the index is less than 0.25, the weight change range is set to 20%, and when the index is between 0.25 and 0.5, the change range is controlled within 10%, as shown in Table 3.

[0150] Table 3: Distribution of User Satisfaction Ratings for Voucher Use

[0151] Rating range Number of users percentage Regulation strategy 0.8～1.0 975 39.0％ Strengthen learning and empowerment 0.5～0.8 1075 43.0％ Keep existing parameters 0.0～0.5 450 18.0％ Penalty weight reduction

[0152] In the recommendation result generation phase, researchers employed a hybrid optimization architecture combining offline pre-computation and online retrieval. High-frequency coupons, with a query frequency exceeding 150 times per hour, were stored in a memory cache for millisecond-level access. Mid-frequency coupons, with a query frequency between 15 and 150 times per hour, were stored in a solid-state drive cache to ensure second-level response. The cache hit rate reached 88.6%, exceeding the set threshold of 85%, and the system directly returned a high proportion of pre-computed results.

[0153] In the approximate algorithm coupling optimization function, the sampling rate is set to 10% for dense data regions and 30% for sparse data regions. In the weight allocation of the data point characteristic coupling term, outlier points are weighted at 0.1, equilibrium points at 0.5, representative points at 0.8, and the weight of mutation points is dynamically adjusted. Abnormal usage behavior is identified through Mahalanobis distance calculation, and typical user groups are determined through information entropy and variance analysis. The overall computational complexity is reduced from O(n log n). 2 The value is reduced to O(nlogn).

[0154] During the content recommendation ranking phase, the weighting coefficients ω1 and ω2 for relevance score and importance score were set to 0.6 and 0.4, respectively. When the user's query intent matched the recommended coupon category with a score greater than 0.9, the relevance score was set to 1.0; when the score was between 0.7 and 0.9, it was set to 0.8; and when the score was less than 0.7, it was set to 0.5. In the importance score, coupons for essential needs were set to 1.0, coupons for interests were set to 0.6, and coupons for optional uses were set to 0.3. The system selected the top 5 recommendations with the highest overall scores and presented them to the user, as shown in Table 4.

[0155] Table 4. Verification of the Accuracy of Consumption Voucher Recommendation Results

[0156]

[0157]

[0158] After four months of implementation, researchers conducted a comprehensive evaluation of the system's performance. Recommendation response time was reduced from an average of 3.2 seconds using traditional methods to 2.1 seconds, an improvement of approximately 34%. User coupon usage efficiency improved by approximately 16%, and user satisfaction increased from 71.8% in traditional recommendation systems to 84.2%, an improvement of 17.3%. System resource utilization remained at around 78%, and the computational load was evenly distributed.

[0159] In the validation of the adaptive sequence length optimization, the average sequence length for complex queries was 342, and the average sequence length for simple queries was 76. The system can dynamically adjust parameters according to query complexity. The weight distribution of the attention mechanism shows that the average attention weight for recently used records is 0.71, and the average attention weight for historical records is 0.29, which meets the expected effect of adjusting the timeliness weight; as shown in Table 5.

[0160] Table 5 Performance Comparison and Evaluation of Consumption Voucher Recommendation System

[0161] Performance indicators Traditional recommendation methods Method of the present invention Increase Recommended response time (seconds) 3.2 2.1 34.4％ Recommendation accuracy 71.8％ 84.2％ 17.3％ User satisfaction 69.2％ 81.5％ 17.8％ Coupon usage efficiency 74.6％ 86.5％ 16.0％ Cache hit rate 73.4％ 88.6％ 20.7％

[0162] During the implementation of the online learning mechanism, the system processed approximately 1.85 million user feedback data points. The average user operation satisfaction score calculated by the intent-relatedness function was 0.763, with 39% of users scoring above 0.8, triggering reinforcement learning strategies. The average iterative effect exponent of the gating weight dynamic adjustment function was 0.598, with most adjustment strategies employing a fine-tuning mode, controlling weight changes within 5%.

[0163] The model parameter update frequency was set to trigger an update process once every 1500 new feedback data points collected. During the four-month operation, the system performed 1233 model parameter updates, with 91.2% of the updates improving model performance. Dynamic adjustment of feature importance scores showed that the weight of features related to restaurant vouchers increased from an initial 0.62 to 0.76, and the weight of features related to shopping vouchers increased from 0.58 to 0.73, demonstrating the adaptability of online learning to changes in user preferences.

[0164] Through time-series analysis extracted from user behavior patterns, researchers found significant differences in user preferences for coupon types across different time periods. Users were more inclined to search for coffee and breakfast coupons in the morning, with searches for restaurant coupons peaking during lunchtime. Demand for milk tea and dessert coupons increased during afternoon tea, while restaurant coupons saw the highest recommendation success rate during dinner. Entertainment and shopping coupons were used approximately 42% more frequently on weekends compared to weekdays.

[0165] User profile clustering analysis divided 2,500 users into six main groups: food lovers, avid shoppers, entertainment consumers, commuters, home-based consumers, and general consumers. Significant differences in coupon preferences were observed among these user groups. Food lovers showed an 89.3% usage rate for restaurant coupons, avid shoppers for shopping mall coupons for 85.7%, and entertainment consumers for cinema coupons for 82.4% and KTV coupons for 79.6%, as shown in Table 6.

[0166] Table 6. Statistics on Coupon Usage Preferences of Different User Groups

[0167] User Group Number of people percentage Main coupon preferences Average usage Food lovers 425 17.0％ Restaurant vouchers, food delivery vouchers 89.3％ Shopping expert 375 15.0％ Shopping mall coupons, brand coupons 85.7％ Entertainment Consumers 350 14.0％ Cinema vouchers, KTV vouchers 81.0％ Commuters 450 18.0％ Transportation vouchers, breakfast vouchers 87.2％ Home-based consumers 400 16.0％ Life service vouchers, supermarket vouchers 83.5％ Comprehensive consumer 500 20.0％ Use all types of coupons in a balanced manner 78.9％

[0168] In terms of computational efficiency optimization, the approximate algorithm coupling optimization function reduces the computational complexity of large-scale user data processing from O(n^2) to O(n^2). 2 The computation time for processing 2.5 million user data entries was reduced to O(nlogn), decreasing from 68 minutes to 26 minutes. The application of locality-sensitive hashing improved the similarity calculation efficiency by approximately 38%, while random sampling reduced the data processing volume to 23% of the original, maintaining an accuracy of over 96%.

[0169] The offline pre-computation module processed approximately 36,000 common coupon recommendation patterns, accounting for 87% of the total recommendation patterns. In the three-level caching architecture, the memory cache stores the 14% of most frequently accessed results, the solid-state drive cache stores 44% of the mid-frequency results, and the distributed storage system handles 42% of the low-frequency results. This overall caching strategy keeps the system's average query response time below 280 milliseconds.

[0170] The multimodal fusion recommendation model demonstrates good fusion performance when processing page operation records, query content, historical usage data, and user profile information. In the weight distribution of page operation features, browsing behavior accounts for 32%, search behavior for 28%, filtering behavior for 25%, and claiming behavior for 15%. After representing query semantic features using word embedding vectors, the average semantic similarity with historical usage preferences reaches 0.734, showing a strong correlation.

[0171] The weighting of the attention mechanism among different modal features shows that the current operational intent has a weight of 0.35, historical usage records have a weight of 0.28, user profile information has a weight of 0.22, and query semantic features have a weight of 0.15. This weighting allows the recommendation results to reflect both the user's immediate needs and long-term preferences, significantly improving the personalization of the recommendations.

[0172] To address the core technical challenge of intelligent recommendation of consumer vouchers, traditional methods primarily employ collaborative filtering algorithms based on users' historical behavior and content recommendation methods based on voucher category attributes. Collaborative filtering methods make recommendations by analyzing the usage behavior of similar users, but suffer from recommendation lag in consumption scenarios where user preferences change rapidly. Content-based recommendation methods match vouchers based on attributes such as merchant, type, and discount amount, but struggle to capture users' immediate needs and changes in usage scenarios.

[0173] The main technological advancements of this invention compared to traditional methods are reflected in several aspects. First, by deeply integrating user page operation behavior, query semantics, historical usage records, and personal profile information through multimodal fusion technology, it overcomes the limitations of traditional methods with their single information dimension. Second, the attention mechanism based on the Transformer architecture can dynamically identify users' immediate needs and long-term preferences, improving recommendation accuracy by 17.3% compared to traditional static recommendation methods. Third, the sequence length adaptive optimization function dynamically adjusts model parameters according to query complexity, improving computational efficiency by approximately 16% compared to traditional methods with fixed parameters. Finally, the online learning mechanism optimizes the recommendation strategy through real-time feedback, solving the problem that traditional methods cannot quickly adapt to changes in user preferences, resulting in a 17.8% increase in user satisfaction. Through approximate algorithm coupling optimization, the system's computational complexity is reduced from O(n^2) to O(n^2). 2 The efficiency is reduced to O(nlogn), which is a 20.7% improvement compared to the traditional method's 73.4%.

[0174] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A user operation recommendation method under an AI platform, characterized in that, include: A distributed data acquisition system is used to acquire user page operation records, query content, historical interaction data, and user profile information in real time. The page operation records are then processed using a clustering compression algorithm to remove duplicate or erroneous records. This clustering compression algorithm employs K-means clustering to group user page operation records, grouping similar operation records into the same cluster. Data compression is achieved by extracting cluster centers as representative operation records, while simultaneously identifying and removing duplicate or erroneous records, reducing storage space usage and subsequent data processing time. The algorithm calculates similarity distance based on operation type, timestamp, page location, and user identifier features. ， in, For the first The operation record and the first Euclidean distance between operation records For the first The first operation record 1 eigenvalue, For the first The first operation record 1 eigenvalue, With a feature dimension of 4, operations with a similarity distance of less than 0.3 are grouped into the same cluster, and data compression is achieved by extracting the center point of each cluster as a representative operation record. Natural language processing technology is used to segment and extract semantic features from the query content, while sequence identification algorithm is used to perform time series analysis on historical interaction data to extract user behavior patterns. Construct a user action intent graph, using the user's current action, historical behavior, and query content as graph nodes, and calculate the association strength between nodes using the shortest path algorithm to find the optimal operation path; Page operation records, semantic feature vectors, historical interaction data, and user profile information are input into a multimodal fusion recommendation model. An attention mechanism and a sequence length adaptive optimization function are used for feature fusion and recommendation generation. An online learning system is constructed to receive user feedback data in real time. An intent-related function is used to calculate a user operation satisfaction score. This intent-related function comprehensively considers data such as operation success rate, user dwell time, click-through rate, and task completion status, and calculates a numerical score reflecting user operation satisfaction through a weighted summation. ， in, User satisfaction ratings range from 0 to 1. To improve the success rate of the operation, The actual time the user stays. Standard stay time, For click-through rate, The score for task completion is either 0 or 1. The weighting coefficients are set to values ​​of 0.3, 0.2, 0.3, and 0.2 respectively. The feedback adjustment term has a value range of -0.1 to 0.

1. When the score falls into different ranges, the gated weight dynamic adjustment function is used to update the model parameters. When a user performs an action, multi-dimensional information is acquired in real time. A recommendation for the next action is generated through feature fusion and model prediction. The approximate algorithm coupling optimization function is represented as follows: ， in, For data point characteristic coupling terms, The weight of outliers is set to 0.

1. The weight for the equilibrium point is set to 0.

5. The weight of the representative point is set to 0.

8. The weights of mutation points are dynamically adjusted. The coupling coefficients were set to values ​​of 0.15, 0.25, 0.35, and 0.25, respectively. The coupling error term ranges from -0.02 to 0.

02. The data point characteristic coupling term comprehensively considers outlier identification, equilibrium point location, representative point selection, and variant point detection. It calculates a comprehensive index reflecting the data distribution characteristics by weighting and combining data points with different characteristics. The function adopts an adaptive sampling strategy to adjust the sampling frequency according to the data density distribution. The sampling rate is set to 10% in dense data areas and 30% in sparse data areas. The recommended content is sorted and filtered based on its relevance and importance, and the most suitable action recommendation is fed back to the user in real time.

2. The user operation recommendation method under the AI ​​platform according to claim 1, characterized in that, The sequence identification algorithm specifically marks and encodes historical interaction data in time series, extracts user behavior patterns by analyzing the time intervals, frequencies, and sequence patterns of user operations, and converts time series data into structured feature representations.

3. The user operation recommendation method under the AI ​​platform according to claim 2, characterized in that, The shortest path algorithm specifically calculates the minimum weighted path between any two graph nodes in the user's operation intent graph. It then uses Dijkstra's algorithm to traverse the graph structure to find the optimal operation path from the current operation state to the target operation state. The path weight is calculated based on the frequency of the user's historical behavior and the operation success rate. The shortest path sequence output by the algorithm serves as the priority basis for recommended operations.

4. The user operation recommendation method under the AI ​​platform according to claim 3, characterized in that, The multimodal fusion recommendation model is specifically a deep neural network built on the Transformer architecture. The model includes a multi-head attention mechanism to process feature information from different modalities, and achieves feature fusion and recommendation generation through an encoder-decoder structure. The sequence length parameter of the model is dynamically adjusted according to the sequence length adaptive optimization function.

5. The user operation recommendation method under the AI ​​platform according to claim 4, characterized in that, The sequence length adaptive optimization function is specifically used to dynamically determine the optimal input sequence length of the multimodal fusion recommendation model. The function calculates the most suitable sequence length parameter for the current recommendation scenario by comprehensively analyzing user operation complexity scores, historical interaction depth indicators, current task type encoding, system resource utilization, and user response time preferences. The function uses a weighted fusion strategy to combine and calculate each input parameter according to its importance.

6. The user operation recommendation method under the AI ​​platform according to claim 5, characterized in that, The aforementioned dynamic adjustment function for gating weights is specifically used to adjust the gating mechanism parameters of the neural network. The function calculates an iterative effect index based on user satisfaction feedback data, model prediction accuracy, feature importance scores, historical adjustment effects, and current recommendation performance data. When the iterative effect index falls within different ranges, different weight adjustment functions are used to adjust the gating parameters. The function achieves adaptive optimization of the gating parameters by dynamically monitoring the changing trend of the iterative effect index.

7. The user operation recommendation method under the AI ​​platform according to claim 6, characterized in that, The offline pre-computation and online retrieval optimization function divides the recommendation calculation process into two stages: offline pre-computation and online fast retrieval. In the offline stage, common user operation patterns, frequently asked questions, and typical interaction sequences are pre-computed in batches, and the results are stored in a high-performance caching system. In the online stage, the pre-computation results are retrieved and adjusted in real time based on the current user status to generate the final recommendation.

Images