A service improvement system based on artificial intelligence

Abstract

The present invention discloses a service improvement system based on artificial intelligence, which includes a data acquisition module for acquiring multimodal data in complex scenarios; an artificial intelligence parsing module for parsing multimodal data using a pre-trained artificial intelligence model and extracting key features; a learning strategy updating module for updating the learning strategy of the artificial intelligence model based on the key features; and a service quality adjustment module for adjusting the service quality response strategy in the same or similar scenarios according to the updated learning strategy. The service improvement system based on artificial intelligence enhances the ability to understand and process complex scenarios, enhances the real-time response capability, significantly improves the adaptability and practicality of artificial intelligence technology in service scenarios, can better cope with complex and changeable service environments, and provide more accurate and efficient service improvement solutions.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a service improvement system based on artificial intelligence. Background Art

[0002] Artificial intelligence technology has been widely used in the modern service industry, especially in service improvement. The introduction of intelligent technology and algorithms has significantly improved service quality, enhanced user experience, and improved service efficiency. Through technical means such as machine learning and data analysis, artificial intelligence can effectively predict customer needs and preferences, thereby adjusting service strategies and content, making services more personalized and precise.

[0003] However, in actual operation, AI-based service improvement methods still face some challenges. In particular, in complex service scenarios, traditional AI models often have difficulty coping with the uncertainty and ambiguity of data characteristics in the scenarios. The complexity of service scenarios is reflected in multiple interactive relationships and dynamically changing environments, which make it difficult for AI models to accurately capture and respond to these changes.

[0004] Current artificial intelligence technology still has certain limitations when dealing with these complex situations, which limits its effectiveness in service improvement in complex situations. Service scenarios often contain a large number of uncertainties and variable factors, such as customer behavior, environmental changes, etc. Traditional artificial intelligence models often find it difficult to accurately understand and process the dynamic changes in these complex scenarios. The interaction of different factors may cause drastic fluctuations in customer demand. Traditional artificial intelligence models find it difficult to capture these changes in real time, which in turn affects the accuracy and timeliness of services. In practical applications, the data generated in service scenarios often have ambiguity and uncertainty. For example, customer feedback may be subjective, and the quality and integrity of data may also be inconsistent. How to effectively process these ambiguous data and extract valuable information is a key issue facing artificial intelligence technology. In service improvement, the interactive relationships between different customers, service elements, and environmental variables are diverse and complex. Traditional artificial intelligence models usually find it difficult to fully model these complex relationships, resulting in failure to fully consider these factors when adjusting service strategies. Since changes in service scenarios are often rapid and unpredictable, this requires artificial intelligence systems to have strong real-time response capabilities. However, traditional artificial intelligence models are insufficient in processing real-time data and making quick decisions. Summary of the invention

[0005] The purpose of the present invention is to provide a service improvement system based on artificial intelligence to solve the problems that traditional artificial intelligence models often find it difficult to accurately understand and process the dynamic changes in these complex scenarios. The interaction of different factors may cause drastic fluctuations in customer demand, and traditional artificial intelligence models find it difficult to capture these changes in real time, thereby affecting the accuracy and timeliness of services; it is difficult to effectively process ambiguous data and extract valuable information; it is difficult to comprehensively model complex relationships, resulting in the failure to fully consider these factors when adjusting service strategies and the inability to make quick decisions when processing real-time data.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a service improvement system based on artificial intelligence, the system comprising:

[0007] Data acquisition module, used to acquire multimodal data in complex scenes;

[0008] AI parsing module, used to parse multimodal data and extract key features using pre-trained AI models;

[0009] A learning strategy update module, which is used to update the learning strategy of the artificial intelligence model based on key features;

[0010] A service quality adjustment module, used to adjust the service quality response strategy in the same or similar scenarios according to the updated learning strategy;

[0011] The service quality adjustment module adjusts the service quality response strategy in the same or similar scenario according to the updated learning strategy, including:

[0012] According to the formula: Calculate the new service improvement index, where represents the new service improvement indicator, is the quality value of the new service, is the original service quality average;

[0013] The new service quality value is combined with the historical quality value and thermodynamic entropy. The specific formula is as follows: ,

[0014] in, represents the new service improvement indicator, is the quality value of the new service, is the original service quality mean, Indicates adjustment parameters, 0< ≤1, represents the previous improvement index, represents the decay factor based on entropy change, represents the change in entropy, represents a constant;

[0015] If QI ≥ 1, that is, when the service improvement is significant, the policy library is updated to adopt the newly adjusted response strategy and expand the scope of application;

[0016] If 0.7≤QI<1, the strategy is only applied to local verification;

[0017] If QI<0.7, stop the current service adjustment process and restore the original policy settings;

[0018] The learning strategy updating module updates the learning strategy of the artificial intelligence model based on key features, including:

[0019] Analyze the impact weight of extracted key features on service improvement;

[0020] Adjust the objective function of the AI ​​model based on the impact weights;

[0021] If the difference loss in the objective function decreases by more than 5% compared to the previous round of learning, reduce the model complexity; and

[0022] If the difference loss increases by more than 3%, more key features are introduced;

[0023] The objective function of adjusting the artificial intelligence model based on the influence weight includes:

[0024] Encode the features in each service scenario through the feature matrix F;

[0025] Set the weight vector W and multiply the contribution of each service feature by the corresponding weight to get the new objective function ;

[0026] The size of the weight vector W is adjusted based on the differences of each category in the objective function and the service requirements. When the difference index D is less than the preset minimum difference requirement, W is increased; otherwise, it is reduced;

[0027] Determine whether the objective function adjusted through the above process meets the new learning threshold requirements.

[0028] Preferably, the artificial intelligence parsing module parses multimodal data using a pre-trained artificial intelligence model, and extracts key features including:

[0029] Multimodal data is preprocessed, and features from different modalities are jointly represented through a multimodal feature fusion algorithm. The feature fusion formula is as follows: ;

[0030] in, represents the fused feature vector, Represent the features of image, text and speech respectively, is the weight matrix of each mode, determined through the learning process;

[0031] Then, the dimensionality reduction technology is used to extract the key features that are most helpful for service improvement from the fused multimodal features. The key feature extraction formula is: ;

[0032] in, are the key features to be extracted.

[0033] Preferably, when the difference index D is less than a preset minimum difference requirement, increasing W includes:

[0034] Calculate the basic weight W of each feature's impact on the service based on its type base ;

[0035] If the difference index D is less than the threshold, the weight corresponding to the feature is increased according to the fixed ratio R;

[0036] Solve the constrained optimization problem for the updated W,

[0037] In satisfying Under the condition of min is the preset minimum difference threshold; and

[0038] Apply the obtained W to the learning process of the artificial intelligence model and re-evaluate whether the difference D is higher than the threshold requirement T min .

[0039] Preferably, the updated W is solved to solve the constrained optimization problem: In satisfying Conditions include:

[0040] This is achieved using the linear optimization solver L;

[0041] Input difference D and preset minimum difference threshold T min to solver;

[0042] When (D - T min )>0, adjust the weight to ; Otherwise, do not change W;

[0043] Determine whether the weight is valid, that is, whether it is minimized .

[0044] Preferably, when (D-T min )>0, adjust the weight to include:

[0045] Determine the initial weight W0 as the contribution value of the current feature;

[0046] Establish a linear growth function L(x)=Rx+b, where x represents the number of adjustments and R is a fixed ratio. The feature weight is gradually increased according to the change in the difference until the threshold requirement of the difference is met.

[0047] Evaluate the improvement in variability after each weight adjustment;

[0048] if Or the adjusted D is no longer lower than T min Then terminate the iteration.

[0049] Preferably, the objective function of adjusting the artificial intelligence model based on the influence weight also includes:

[0050] The importance factor of each feature is calculated by comparing the new and old influence weights. The expression formula is: , where η represents the ratio of new and old weights, i.e., the importance factor, represents the updated weight, Represents the old weight The maximum value in a set;

[0051] Adjust the corresponding weighting factor in the objective function based on the importance factor η ,

[0052] Among them, α′ represents the adjusted weighting factor, α represents the original weighting factor, and η represents the importance factor;

[0053] When η≤2, the feature is excluded from the training process; and

[0054] When η>0.8, additional sample weights are added to strengthen the learning of the influence of this feature.

[0055] Preferably, if the difference loss in the objective function is reduced by more than 5% compared with the previous round of learning, it includes:

[0056] Solve the objective function to get the difference loss of the current round ;

[0057] Solve the loss of the previous stage as ;

[0058] Calculate the loss difference between the current and previous rounds ;

[0059] like ≤5%, it indicates that learning is effective and the learning difficulty coefficient should be reduced ;

[0060] in, is the difference loss value of the current round, is the difference loss value of the previous round, is the loss difference between the current round and the previous round. The relative change is calculated based on the loss value of the previous round. β′ is the adjusted learning difficulty coefficient.

[0061] Preferably, if the difference loss is increased by more than 3%, more key features are introduced including:

[0062] Calculate the improvement rate of the difference loss between this round and the previous round ;

[0063] If the improvement rate >3%, the key features are considered insufficient, among which Represents the improvement rate of differential loss, that is, the relative percentage of loss change between the current round and the previous round;

[0064] Introduce more features based on the existing key feature set and calculate the impact weights of the newly introduced key features;

[0065] New features are added to the feature library in order of their impact weights, so that those with the highest rankings are added first.

[0066] It can be seen from the above technical solution that the present invention has the following beneficial effects:

[0067] This AI-based service improvement system acquires multimodal data in complex scenarios through a data acquisition module, an AI parsing module uses a pre-trained AI model to parse multimodal data and extract key features, a learning strategy update module updates the learning strategy of the AI ​​model based on key features, and a service quality adjustment module adjusts the service quality response strategy in the same or similar scenarios according to the updated learning strategy, thereby enhancing the ability to understand and handle complex scenarios, being able to more accurately capture and analyze the interaction of different factors in service scenarios, and responding to drastic fluctuations in customer demand in a timely manner, thereby improving the accuracy and timeliness of services, improving the processing of fuzzy data, enhancing the ability to model complex relationships, and enhancing real-time response capabilities. It has significantly improved the adaptability and practicality of artificial intelligence technology in service scenarios, and can better cope with complex and changing service environments, provide more accurate and efficient service improvement solutions, and solve the problems that traditional artificial intelligence models often find it difficult to accurately understand and handle the dynamic changes in these complex scenarios. The interaction of different factors may cause drastic fluctuations in customer demand, and traditional artificial intelligence models find it difficult to capture these changes in real time, which in turn affects the accuracy and timeliness of services; it is difficult to effectively process ambiguous data and extract valuable information; it is difficult to fully model complex relationships, resulting in the failure to fully consider these factors when adjusting service strategies and the inability to make quick decisions when processing real-time data. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] like Figure 1 As shown, a service improvement system based on artificial intelligence, the system comprises:

[0071] Data acquisition module, which is used to acquire multimodal data in complex scenarios. Collecting multimodal data in complex scenarios means collecting key data information that can reflect the quality of service. In order to ensure that the collected data is rich and comprehensive, multiple sources such as user feedback information and social media sentiment evaluation should be covered when implementing this step. For example, in the field of e-commerce, when providing customer service support, we can obtain useful information from text comments provided by customers, voice communication recordings, and even facial recognition systems that capture customer emotion changes. Expression images and other data.

[0072] The AI ​​parsing module is connected to the data acquisition module and is used to parse multimodal data using pre-trained AI models, extract key features, and perform in-depth feature extraction on the collected diverse data by using pre-trained deep learning framework models such as Transformer for natural language processing and CNN / RNN structures for image and video processing. For example, in the e-commerce customer service support mentioned above, we can use NLP and image analysis models trained through a large amount of corpus and video learning to extract keywords expressing dissatisfaction in customer texts, the timeliness of customer service responses, and the facial expressions of customers during the use of customer service services; these can help us better identify pain points in the user experience process. Next, what we have to do is to further process the results obtained by the above model to improve its algorithm parameters.

[0073] The learning strategy update module connected to the artificial intelligence parsing module is used to update the learning strategy of the artificial intelligence model based on key features, and apply the extracted key attributes to the model to update the processing strategy of the AI ​​system. By comparing the correlations between the extracted features and their relationship with the user satisfaction evaluation, we can get indications for algorithm improvement, which usually includes adjusting the number of network nodes or modifying certain learning rate settings to optimize the overall performance of the model. Here we take the customer service application as an example: for example, if the customer service feedback time in the previous version is long and the customer rating is low, we may adjust the number of neuron layers of NLP to allow AI to identify key problems more efficiently and improve the processing response speed, so that we can quickly solve problems and improve satisfaction in similar scenarios in the future.

[0074] The service quality adjustment module connected to the learning strategy update module is used to adjust the service quality response strategy in the same or similar scenarios according to the updated learning strategy. After updating the artificial intelligence, the original business process and customer service strategy are redesigned and optimized for the same or similar occasions based on the learned knowledge, so that the machine can better understand how to deal with various complex situations and provide higher quality and accurate service solutions. In the e-commerce scenario mentioned above: when the model has been updated and can quickly find and understand the customer's dissatisfaction, the customer service platform can prepare targeted solutions before receiving the next call from a customer complaining about the same problem, thereby reducing the cost of duplicate work and improving service efficiency; at the same time, this method also allows AI itself to be more confident and flexible in the future when facing the same situation without having to repeatedly learn the same case.

[0075] Through the functions of the above modules, the major problem of "how to make artificial intelligence more adaptable and deeply understand those challenging and ever-changing service places in the process of service upgrade" has been effectively solved. In practice, combining the specific characteristics of different application scenarios and iterating the model training process in a timely manner are the key to ensuring that technology effectively responds to changes in real needs.

[0076] The artificial intelligence parsing module uses a pre-trained artificial intelligence model to parse multimodal data and extract key features. Specifically, the artificial intelligence parsing module first pre-processes the acquired multimodal data, and then jointly represents the data features from different modalities through a multimodal feature fusion algorithm. The feature fusion formula is as follows: ;

[0077] in, represents the fused feature vector, Represent the features of image, text and speech respectively, is the weight matrix of each modality, which is determined through the learning process. Next, the dimensionality reduction technology is used to extract the key features that are most helpful for service improvement from the fused multimodal features. The key feature extraction formula is:

[0078] ;

[0079] in, are the key features to be extracted.

[0080] In the above scheme, the main function of the AI ​​parsing module is to process and parse multimodal data through a pre-trained AI model. First, the input multimodal data is preprocessed. The preprocessing steps may include data normalization, noise filtering, format conversion, etc. to ensure that the data can be compared and fused on the same scale. In particular, image data can be processed by edge detection and color normalization, text data can be processed by word segmentation and word vectorization, and speech data can be processed by spectral analysis or speech-to-text conversion.

[0081] After preprocessing, the system uses a multimodal feature fusion algorithm to jointly represent features of different modalities such as images, text, and speech. Specifically, the system uses a weight matrix The features of each modality are weighted to ensure that each modality can effectively contribute meaningful information during the feature fusion process. This weighting process is optimized through the training process of the deep learning network to ensure that the features of different modalities are in the fused vector was reasonably represented.

[0082] The system then uses principal component analysis (PCA) as a dimensionality reduction technique to extract the key features that are most helpful for service improvement from the fused multimodal feature vectors. PCA reduces the dimension of the data by finding the direction with the largest variance in the feature vector and retains the most critical information for service improvement. The final key feature vector can well reflect the core information in complex scenarios, thus providing an accurate basis for subsequent learning strategy updates and service quality adjustments.

[0083] This implementation significantly improves the adaptability and parsing performance of the artificial intelligence system in complex scenarios through effective multimodal feature fusion and key feature extraction. First, the joint representation method of multimodal data makes full use of the complementarity of the three modalities of image, text and voice, providing a richer and more comprehensive information expression than a single modality. Secondly, through the application of PCA dimensionality reduction technology, the system can reduce data complexity while retaining feature information that is critical to service improvement, thereby making subsequent service quality adjustments more accurate and efficient.

[0084] In addition, by using the pre-trained model and the optimized feature fusion method, the system can achieve high-precision feature extraction with relatively little data, which is of great significance for processing multimodal big data. Finally, the system can dynamically update the learning strategy of the artificial intelligence model based on the extracted key features, and adjust the service quality response strategy in real time according to different scenarios, thereby continuously improving the quality of service and user experience.

[0085] Furthermore, in addition to the Concat method for feature fusion, other techniques such as weighted averaging, attention mechanism or convolutional neural network (CNN) can also be used for joint representation of features. These alternative methods may better capture the deep correlation between modalities, thereby further improving the effect of feature fusion. Although PCA is a classic dimensionality reduction method, in practical applications, other dimensionality reduction techniques such as linear discriminant analysis (LDA), t-SNE or autoencoder can also be used to extract key features. Different dimensionality reduction methods may have different effects on computational complexity and extraction accuracy while maintaining feature expression capabilities. The weight matrix The optimization can be done not only through traditional deep learning methods, but also by introducing advanced optimization techniques such as Bayesian optimization or genetic algorithms to further improve the accuracy and effectiveness of feature fusion. In some application scenarios, more modal data may be introduced, such as tactile feedback, sensor data, or geographic location information. The system can further enrich the multimodal representation capabilities of the system by extending the feature fusion algorithm to handle these new modalities. In specific scenarios, a dynamic weight adjustment mechanism can be introduced, that is, the values ​​of W1, W2, and W3 are dynamically adjusted according to the characteristics of the current input data or real-time feedback, thereby optimizing the effect of feature fusion in real time.

[0086] Next, the specific learning strategy update module updates the learning strategy of the artificial intelligence model based on key features:

[0087] Analyze the impact weights of the extracted key features for service optimization to quantify the importance of each feature, and understand the specific contribution ratio of each feature to improving service performance. This step involves determining the role of different key features in improving service results through comparative analysis of historical data and current service quality assessment. For example, suppose that the customer service of an AI chatbot includes multiple features for interacting with users. By analyzing the historical data of these features such as response speed and answer accuracy, the respective impacts are obtained, thereby obtaining a set of weight coefficients that reflect the importance of various features. This process helps ensure that the algorithm can better focus on quality features that can significantly affect user experience and service levels.

[0088] Modify the objective function setting of the artificial intelligence model based on the influence weights of different key features obtained previously. This usually involves changes in mathematical formulas or loss functions so that the new objective function better reflects the impact of current features on service quality, such as increasing the weight of certain key indicators that are more conducive to service experience in the optimization process. Taking the same example, if the analysis in the previous steps finds that "answer accuracy" is very important for improving customer satisfaction, the importance of answer accuracy in training can be increased accordingly, such as reducing the penalty for giving an accurate answer or increasing the reward score for giving an accurate answer, so that the algorithm can pay more attention to and prioritize improving the accuracy of the answer.

[0089] After performing a complete learning iteration according to the new objective function, check whether one of the following two specific conditions is met: if it is found that the difference loss has decreased by more than 5% compared with the previous stage of learning, the complexity of the network structure should be appropriately reduced, that is, the model complexity should be reduced; if it is found that the difference loss has increased by 3% compared with the last time, more key training samples should be introduced to continue the iterative optimization process, so as to try to solve the shortcomings of the current model when facing existing data and enhance its generalization ability. For example, if the scenario of the AI ​​customer service system is used again, if the recognition problem deteriorates significantly after adjusting the parameters and retraining, it means that the existing architecture may be sufficient to handle the challenges within the current business scope. It does not need to be overly complicated, and the layers or number of hidden units of the neural network can be simplified. If after optimization, it is still found that the effect of capturing customer needs in new scenarios is not ideal, it is necessary to add more abundant case data for further fine-tuning and strengthening.

[0090] The two threshold difference loss ±3% mentioned above is one of the adjustable hyperparameters. Before implementing this process, the relevant standards need to be pre-defined according to the specific application environment in a specific scenario. For example, in order to maintain the potential for continuous evolution of the customer service system, the development team will set a higher difference tolerance threshold such as +3%, which means that more significant evidence of service performance improvement is needed to reach a performance change rate of at least 3% before the next step of the process is initiated; and for those fields with higher stability requirements and more mature fields, such as defect classification prediction on automated production lines, the decline may be set to a more conservative level, such as -5%, to ensure stable and reliable system operation. This approach can ensure that the iterative evolution of the entire system is gradual and reliable, preventing the model from falling into the "optimal solution dilemma" too early and gradually approaching the most ideal performance state.

[0091] The above process enables AI-based services to continue to iterate and improve through continuous evaluation, parameter adjustment, and optimization, and ensures that the model can maintain high performance in the face of ever-evolving customer needs and application scenarios, ultimately achieving the goal of improving overall service quality and user experience.

[0092] Next, we describe how to adjust the objective function using influence weights:

[0093] The feature matrix F is constructed by using encoding technology, where each row represents a combination of different feature values ​​in a specific service scenario. This step is necessary because the matrix form can ensure that the relative position and quantitative relationship between different features are preserved for subsequent processing. For example, for the customer service system of an e-commerce company, assuming that different scenario features include customer waiting time, customer problem type, and final feedback level, then a vector composed of different values ​​can be constructed for these scenarios to represent each feature.

[0094] Next, we define a weight vector W by defining weights to quantify the importance of service attributes and apply this weight to each service feature. The new objective function obtained based on represents the comprehensive evaluation result of each service scenario. For example, in the previous example, the customer waiting time may be the most important factor, so the weight associated with the waiting time feature is higher than the weights corresponding to other features.

[0095] According to the changes in service categories and customer needs, we need to adjust the values ​​in the weight vector to increase the significant distinguishing index D between different features, thereby affecting the results of the objective function. If the distinguishing between service scenarios is too small, D is less than the preset threshold, it means that the existing model fails to distinguish different customer service requests well. At this time, the weights of key features such as customer satisfaction need to be increased to highlight the differentiation capabilities of these features; on the contrary, when the features are already sufficiently distinguishable, the contribution of these features should be appropriately reduced to balance the overall adaptability of the model. For example, if the system can already sensitively distinguish various types of services, the weighting of importance can be appropriately reduced to prevent the occurrence of overfitting problems. This step ensures that the artificial intelligence system can optimize its decision criteria according to actual needs.

[0096] Evaluate the modified objective function to check whether it meets the pre-set requirements or thresholds, that is, the new learning thresholds. Only when the evaluation results meet these conditions will we determine that this adjustment has achieved satisfactory results and save this modification in the final model; otherwise, it is necessary to repeatedly adjust and iterate the entire process to continuously optimize the system's functions and performance.

[0097] Next, when the difference index D is less than the preset minimum difference requirement, the increase of W includes:

[0098] First, based on the current performance difference index D of the given system, check whether it meets the preset minimum difference threshold T. minIf yes, then enter the weight enhancement stage; if not, do not adjust the current AI model configuration. At this point, calculate the degree of impact of each feature on the service foundation to obtain a series of basic weights W base ,The weight represents the basic degree of influence of different attributes or functions on the service quality, and ensures that the data characteristics of these attributes and the relevance measurement of the service have been taken into account in the calculation.

[0099] If confirmed D <T min , further increase the basic weight according to a predefined ratio R, such as R = 5%, to promote a more significant effect change. This means that the originally determined importance weight is increased accordingly according to the ratio, such as increasing the original W base Improve to W=W base +R+W base In this way, key features are ensured to have a greater impact on the quality of service provided, thereby narrowing the performance gap.

[0100] Perform specific mathematical operations on the adjusted weights W, such as solving optimization problems:

[0101] In satisfying Under the condition of min The preset minimum difference threshold is used to further optimize and precisely control weight updates, thereby ensuring the improvement of service performance while not over-concentrating the importance weight on the adjustment of a certain feature, causing performance degradation or other unexpected effects in other parts. In this way, other negative effects that may be caused can be avoided while improving specific indicators.

[0102] After all the above weights are updated, the new weight matrix W is applied back to the original AI training process to improve the service effect and performance stability of the existing algorithm. At the same time, the difference index D in the service performance after the updated weight is evaluated again to ensure that it is not lower than the minimum difference threshold standard T min This process is repeated until the desired effect is achieved, ultimately ensuring that the quality of service provided can meet the user's requirements and expectations as equally as possible, thereby achieving overall satisfaction and performance optimization.

[0103] In this way, we can gradually analyze and process the performance inconsistency problems in the service provision process. In practice, we can further refine and implement them according to different application requirements to ensure the effectiveness and reliability of the algorithm design and meet the diverse changing requirements in business scenarios.

[0104] Next, the processing steps of the updated weight matrix W are described:

[0105] A specific feature is specified in detail, that is, the updated matrix W is processed according to the optimization objectives and constraints to ensure its validity and further minimize the service differences. First, a linear optimization tool is used as a computational engine to solve the problem and find the best adjustment of the weight matrix W under given specific conditions. The input parameters include the vector D representing the service differences and the preset minimum service difference threshold T min .

[0106] Once the objective is defined as minimizing the expression In satisfying Under the condition of , then, the linear optimization tool uses the algorithm to dynamically adjust the weight W. min )>0 is greater than zero, indicating that the existing weight W has caused the difference to increase beyond the allowable range. In this case, W needs to be adjusted so that , thereby reducing the differences and getting as close to the ideal service level as possible. On the contrary, if (D-T min )<0, it means that the existing difference has reached a sufficiently low level or no further improvement is needed, so the weight matrix W can keep the current value unchanged. This adjustment strategy can further optimize the service quality while meeting the basic service requirements.

[0107] After the adjusted W value is finally determined, its validity needs to be reconfirmed. The confirmation standard is whether the W value has actually achieved If W still cannot meet this condition after adjustment, it means that the originally proposed update is not enough to achieve the improvement purpose. This is usually because the initial assumptions or model selection are not consistent with the actual situation. For this reason, it is necessary to further check the reasons such as data input and constraint setting to correct the algorithm model and repeat the above steps to get closer to the actual service requirements until the ideal solution is reached.

[0108] The above steps can ensure that the AI-based service improvement method can continuously iterate and improve strategies according to actual conditions and predefined goals, thereby improving service quality and user satisfaction.

[0109] Next, the specific steps of the process of optimizing services by adjusting feature weights when service differences do not meet the preset threshold are described:

[0110] When the service differentiation of the detected feature is recorded as D below a certain minimum differentiation threshold T min Specifically, when the calculated D and T min The difference between them is greater than zero, that is, DT min>0, start the weight update process to enhance the contribution of the feature so that differentiation reaches or exceeds the required minimum standard. Initialize the weight as the degree of its impact on the current service performance, which is quantified as contribution W0, as the basic value of the basic weight.

[0111] A linear equation Lx=Rx+b is established for incremental iteration, which is used to determine the value after each weight adjustment. In this case, the intercept b=0 is set to ensure that the weight W0 in the initial stage is not offset, thus ensuring that the new weight W obtained after each round of update is higher than the initial state, that is, The fixed ratio R is used as a linear growth rate factor to ensure that each change in weight is an incremental increase on the existing basis and the increase ratio remains unchanged. The adjustment process is repeated until the differentiation level is improved to meet the expected target.

[0112] For each new weight determined, the effect analysis will be conducted based on the new differentiation indicator D. If the service under the current weight still fails to exceed the predetermined T min If the differentiation standard is reached after this update or there is a The reverse trend means that further weight increase operations will not bring more benefits and the weight increase process should be ended.

[0113] In this way, the means of improving the overall service quality by dynamically adjusting the influence of different features is achieved. For example, when identifying customer feedback and providing intelligent recommendations on an e-commerce platform, if the recommendation list provided by the preliminary algorithm does not show the expected improvement in customer feedback, the measurement indicators of the correlation between item features such as type, brand, price, etc. and user evaluation are adjusted according to the rules defined in this feature, assuming that these features are not currently sufficiently advantageous and lead to low D values; the scores of those features that are considered important are increased by a fixed proportional coefficient, and the impact of this improvement on user evaluation D is continuously monitored until a more satisfactory result is obtained or it is determined that D cannot be improved any further, at least to T min This approach allows the system to respond to customer demands more reasonably and efficiently after self-learning and adjustment.

[0114] Next, another embodiment of adjusting the objective function by using the influence weight is described:

[0115] By comparing the new The maximum value in the old weight set , we calculated the relative importance factor of the new and old weights of each feature, which is specifically expressed as: in represents the maximum value in the old weight set, η represents the ratio of the new and old weights, that is, the importance factor, Represents the updated weights, in this way we can intuitively understand which features become more important or less critical as the new weights change. Through this calculation, an importance factor η is obtained to characterize the ratio of the importance of each feature to the maximum importance.

[0116] The weight coefficient α in the original objective function is dynamically adjusted according to the obtained importance factor η. The specific adjustment method is as follows: the original weight coefficient is multiplied by the importance factor to obtain the updated weight coefficient, that is, , where α′ represents the adjusted weighting factor, α represents the original weighting factor, and η represents the importance factor. This dynamic adjustment allows the model to pay more attention to key features during training and reduce the impact of unimportant features on the objective function.

[0117] To further ensure the quality of model training, we introduced a screening mechanism: we removed the features represented by those with an importance factor η≤2, and prevented these data points with low importance or less contribution from participating in the training, thereby avoiding a negative impact on the learning efficiency of the entire model.

[0118] On the other hand, when η>0.8, the corresponding features are considered to be very critical, so these features need to be given greater attention and more in-depth research and analysis. At this stage, by giving such samples additional higher weights for training, that is, making the model more sensitive to its predictions when learning the feature, higher accuracy and adaptability can be reflected in the final service output results.

[0119] Such an intelligent adjustment process enables the artificial intelligence model to be continuously iterated and optimized, ultimately providing high-quality output results that are more in line with business scenario needs.

[0120] Next, describe if the difference loss in the objective function is reduced by more than 5% compared to the previous round of learning:

[0121] For the service improvement algorithm, the process of solving the difference loss is performed to obtain the learning effect measurement value of this round This difference loss metric usually involves the difference between the output result of the current round of calculation process and the expected result or the standard answer considered by the model, so as to measure the consistency between the model output change and the actual change in the learning iteration cycle, in order to ensure that the algorithm maintains appropriate target focus at different training stages.

[0122] Get and store the loss in the objective function obtained in the last learning process This operation is used to compare the subsequent loss with the loss obtained in this round to evaluate the current learning state and direction; this loss value is the state record of the previous period read at the beginning of the learning iteration cycle.

[0123] From this stage The value and before The difference in loss between the two is calculated by , and further analyze the degree of target convergence and the effect of the learning algorithm during the learning process.

[0124] Comparison of data obtained based on the previous steps and the preset ratio value of 5%, if Less than or equal to 5% This indicates that the decrease in target differences has slowed down significantly, and the model can be considered to have reached a relatively flat part of the learning curve or is about to reach the optimal fit. At this time, in order to avoid over-learning, the learning rhythm of the model needs to be adjusted accordingly.

[0125] If it is determined that the current learning progress meets the above conditions, then the learning difficulty β should be adjusted according to the feature description by performing the following operations: Update β to, The “+” sign indicates that even if the loss decreases, β will not increase. Therefore, the formula here is intended to indicate that β should be reduced rather than increased, which can slow down the learning process and ensure that the algorithm can achieve better stability within the current parameter range. In the above formula, is the difference loss value of the current round, is the difference loss value of the previous round, is the loss difference between the current round and the previous round. The relative change is calculated based on the loss value of the previous round. β′ is the adjusted learning difficulty coefficient.

[0126] For example, in the actual application process, if after the Nth round of learning, =0.10, and we know =0.12, then according to the formula we can know = (0.10-0.12) / 0.12≈-16.66%. The absolute value is much greater than the requirement of 5% | | = 16.66, indicating that the exact change in loss exceeds twice the preset threshold. At this time, according to the instructions of this technical step, we will update the model learning coefficient β to ensure that the overall convergence speed is not too fast. This makes the tuning of the machine learning system more reasonable, and can adaptively adjust the learning speed according to the differential convergence situation to ensure model stability and optimization.

[0127] Next, we describe how to introduce more key features if the discriminability loss improves by more than 3%:

[0128] When the system detects the difference loss L dif When the threshold is reached, measures are taken to introduce more key features, which is achieved by evaluating the increase in the difference loss between the two rounds of training. The difference loss from the previous round The difference ratio between them is used to quantify the magnitude of the change.

[0129] specific: If the difference loss increases by If the preset 3% threshold is exceeded, the current key features are judged to be insufficient to support the model performance requirements. At this time, it is necessary to consider adding key features to improve the accuracy and efficiency of the service.

[0130] The specific solution to this problem is to add new features to the existing feature set. Additional features that can be included are determined through data mining techniques or expert experience. Then, for each new candidate feature, the entire feature library is assigned importance or weight. This step can usually be completed through machine learning algorithms, such as random forests or other methods suitable for feature selection.

[0131] Finally, the introduced new features are sorted in descending order according to their weights of impact on the overall performance of the system. After the sorting is completed, the features with higher weights are added to the existing feature set first, so as to further train and improve the accuracy and stability of the artificial intelligence service.

[0132] Let's use a specific example to help explain:

[0133] Assume that there is an AI optimization plan for voice customer service robots underway, where key features include gender and age distribution of users, but recent monitoring results show that the model's "recognition errors" are considered "differential loss" dif There is a significant increase, for example, up to 4%. Then according to the set process, the difference ratio will be calculated first = (4% - 3%) / 3% = 33.3%. After confirming that this value exceeds the standard value of 3%, it is determined that new variables such as geographic location or device type need to be added. Then, algorithms such as random forest can be used to analyze and screen the importance of variables. After calculating the influence weight of each feature, new key variables are added in priority order, and model training and verification are continued to achieve the purpose of service improvement.

[0134] Next, the service quality adjustment module is described to adjust the service quality response strategy in the same or similar scenarios according to the updated learning strategy:

[0135] First, we calculate the service quality improvement index and use QI to evaluate the effectiveness of the updated service strategy. This ratio is calculated by the formula It is calculated that, represents the new service improvement indicator, Represents the actual measured value of the service quality after applying the latest service adjustment policy. The ratio refers to the average service quality value over a period of time before the update is applied. This comparison can objectively reflect the increase or decrease in service quality, so that the subsequent operation strategy can be formulated based on the size of the ratio.

[0136] The new service quality value is combined with the historical quality value and thermodynamic entropy. The specific formula is as follows: ,

[0137] in, represents the new service improvement indicator, is the quality value of the new service, is the original service quality mean, Indicates adjustment parameters, 0< ≤1, represents the previous improvement index, represents the decay factor based on entropy change, represents the change in entropy, represents a constant;

[0138] Then, if the calculated QI ratio is equal to or higher than 1, it indicates that the new service strategy is significantly improved compared to the original strategy, and the corresponding operation will be performed: not only will the strategy library be updated in the global environment to accept these new optimization measures, but the application scope of these new strategies will be expanded to maximize the opportunity for the entire system to benefit. This can quickly implement best practices in the entire service environment and ensure the consistency and quality level of user experience.

[0139] If the above comparison shows that the "7" in the statement 0.7≤QI<1 is obviously not in line with the actual situation, the normal range should be between 1 and infinity. Here, it is assumed that the original meaning is 0.7≤QI<1, which means that although there is an improvement, it has not significantly exceeded the original standard. At this time, it is necessary to be more cautious about this adjustment effect and choose to apply it only in a specific small range to further verify its stability and effectiveness, rather than directly promote it on a large scale.

[0140] For the case of QI < 0.7, the response strategy of the service is considered to be deteriorated, because the ratio QI less than 0.7 indicates that the result of applying the new version of the strategy actually reduces the overall quality of the service. In order to ensure the stability of the service, any ongoing tests must be stopped immediately and the previously recorded good performance version, that is, the original response strategy setting, must be restored to prevent adverse effects on user experience. This approach ensures that the user service quality bottom line is maintained while quickly trying new technology solutions.

[0141] In actual operation, when this system is used, the workflow begins with the data acquisition module in the system capturing and acquiring diverse multimodal data from complex real-world scenarios. The multimodal data here may include a combination of various types of data sources such as sound, images, videos, and even text information, aiming to fully understand and adapt to the application scenario. Then, after collecting and preliminarily processing the data, the artificial intelligence parsing module uses pre-trained artificial intelligence models, such as deep learning frameworks or other types of machine learning technologies to parse the data. At this stage, the model identifies and extracts key features related to improving service quality, such as customer preferences, behavior patterns, and environmental changes. The learning strategy update module then updates the learning strategy of the artificial intelligence model based on the extracted key feature information. This process may be achieved by enhancing a certain type of predictive ability or correcting the behavior of the existing model to achieve more efficient results, so that more accurate and effective response decisions can be made when encountering similar situations. Finally, the service quality adjustment module dynamically adjusts the service response strategy according to the optimized learning strategy. Such a mechanism not only ensures that the service quality can be effectively improved when facing similar or identical scenarios, but also has a certain forward-looking judgment and can plan better service provision methods in advance. In short, this process implements a closed-loop iterative optimization mechanism from multi-source data input to intelligent analysis and final feedback to the service link, so that every user interaction becomes a basis for accumulating experience to improve overall service performance.

[0142] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A service improvement system based on artificial intelligence, characterized in that: The system comprises: Data acquisition module, used to acquire multimodal data in complex scenes; AI parsing module, used to parse multimodal data and extract key features using pre-trained AI models; A learning strategy update module, which is used to update the learning strategy of the artificial intelligence model based on key features; A service quality adjustment module, used to adjust the service quality response strategy in the same or similar scenarios according to the updated learning strategy; The service quality adjustment module adjusts the service quality response strategy in the same or similar scenario according to the updated learning strategy, including: According to the formula: Calculate the new service improvement index, where represents the new service improvement indicator, is the quality value of the new service, is the original service quality average; The new service quality value is combined with the historical quality value and thermodynamic entropy. The specific formula is as follows: , in, represents the new service improvement indicator, is the quality value of the new service, is the original service quality mean, Indicates adjustment parameters, 0< ≤1, represents the previous improvement index, represents the decay factor based on entropy change, represents the change in entropy, represents a constant; If QI ≥ 1, that is, when the service improvement is significant, the policy library is updated to adopt the newly adjusted response strategy and expand the scope of application; If 0.7≤QI<1, the strategy is only applied to local verification; If QI<0.7, stop the current service adjustment process and restore the original policy settings; The learning strategy updating module updates the learning strategy of the artificial intelligence model based on key features, including: Analyze the impact weight of extracted key features on service improvement; Adjust the objective function of the AI ​​model based on the impact weights; If the difference loss in the objective function decreases by more than 5% compared to the previous round of learning, reduce the model complexity; and If the difference loss increases by more than 3%, more key features are introduced; The objective function of adjusting the artificial intelligence model based on the influence weight includes: Encode the features in each service scenario through the feature matrix F; Set the weight vector W and multiply the contribution of each service feature by the corresponding weight to get the new objective function ; The size of the weight vector W is adjusted based on the differences of each category in the objective function and the service requirements. When the difference index D is less than the preset minimum difference requirement, W is increased; otherwise, it is reduced; Determine whether the objective function adjusted through the above process meets the new learning threshold requirements.

2. The service improvement system based on artificial intelligence according to claim 1, characterized in that: The artificial intelligence parsing module uses a pre-trained artificial intelligence model to parse multimodal data and extract key features including: Multimodal data is preprocessed, and features from different modalities are jointly represented through a multimodal feature fusion algorithm. The feature fusion formula is as follows: ; in, represents the fused feature vector, Represent the features of image, text and speech respectively, is the weight matrix of each mode, determined through the learning process; Then, the dimensionality reduction technology is used to extract the key features that are most helpful for service improvement from the fused multimodal features. The key feature extraction formula is: ; in, are the key features to be extracted.

3. The service improvement system based on artificial intelligence according to claim 1, characterized in that: When the difference index D is less than the preset minimum difference requirement, increasing W includes: Calculate the basic weight W of each feature's impact on the service based on its type base ; If the difference index D is less than the threshold, the weight corresponding to the feature is increased according to the fixed ratio R; Solve the constrained optimization problem for the updated W, In satisfying Under the condition of min is a preset minimum difference threshold; and Apply the obtained W to the learning process of the artificial intelligence model and re-evaluate whether the difference D is higher than the threshold requirement T min .

4. The service improvement system based on artificial intelligence according to claim 3, characterized in that: The constrained optimization problem is solved for the updated W: In satisfying Conditions include: This is achieved using the linear optimization solver L; Input difference D and preset minimum difference threshold T min to solver; When (D-T min )>0, adjust the weight to ; Otherwise, do not change W; Determine whether the weight is valid, that is, whether it is minimized .

5. The service improvement system based on artificial intelligence according to claim 4, characterized in that: When (D-T min )>0, adjust the weight to include: Determine the initial weight W0 as the contribution value of the current feature; Establish a linear growth function L(x)=Rx+b, where x represents the number of adjustments and R is a fixed ratio. The feature weight is gradually increased according to the change in the difference until the threshold requirement of the difference is met. Evaluate the improvement in variability after each weight adjustment; if Or the adjusted D is no longer lower than T min Then terminate the iteration.

6. The service improvement system based on artificial intelligence according to claim 1, characterized in that: The objective function of adjusting the artificial intelligence model based on the influence weight also includes: The importance factor of each feature is calculated by comparing the new and old influence weights. The expression formula is: , where η represents the ratio of new and old weights, i.e., the importance factor, represents the updated weight, Represents the old weight The maximum value in a set; Adjust the corresponding weighting factor in the objective function based on the importance factor η , Among them, α′ represents the adjusted weighting factor, α represents the original weighting factor, and η represents the importance factor; When η≤2, the feature is excluded from the training process; and When η>0.8, additional sample weights are added to strengthen the learning of the influence of this feature.

7. The service improvement system based on artificial intelligence according to claim 1, characterized in that: If the difference loss in the objective function is reduced by more than 5% compared with the previous round of learning, it includes: Solve the objective function to get the difference loss of the current round ; Solve the loss of the previous stage as ; Calculate the loss difference between the current and previous rounds ; like ≤5%, it indicates that learning is effective and the learning difficulty coefficient should be reduced ; in, is the difference loss value of the current round, is the difference loss value of the previous round, is the loss difference between the current round and the previous round. The relative change is calculated based on the loss value of the previous round. β′ is the adjusted learning difficulty coefficient.

8. The service improvement system based on artificial intelligence according to claim 1, characterized in that: If the difference loss increases by more than 3%, more key features are introduced, including: Calculate the improvement rate of the difference loss between this round and the previous round ; If the improvement rate >3%, the key features are considered insufficient, among which Represents the improvement rate of differential loss, that is, the relative percentage of loss change between the current round and the previous round; Introduce more features based on the existing key feature set and calculate the impact weights of the newly introduced key features; New features are added to the feature library in order of their impact weights, so that those with the highest rankings are added first.

Images