Product optimization method and device based on user feedback, equipment, medium and product

Abstract

The present application relates to the technical field of product optimization, and discloses a product optimization method and device based on user feedback, equipment, medium and product.The present application performs correlation analysis on the multi-modal evaluation data of users, realizes semantic correlation of text, voice and pictures, and obtains user feedback data with high reliability.A plurality of feedback opinions to be processed are extracted from the user feedback data of a plurality of users, multi-dimensional classification of the feedback opinions to be processed is realized by using a knowledge graph architecture, and classification refinement of the feedback opinions to be processed is realized.The scores of the feedback opinions to be processed in a plurality of basic dimensions are evaluated by using the multi-dimensional classification results, and the quantum state priority of the feedback opinions to be processed is obtained according to the quantum amplitude of each basic dimension, so that the priority judgment is more flexible.Finally, each feedback opinion to be processed is sorted based on the quantum state priority, a target product optimization scheme is generated according to the sorting result, and product optimization is performed, thereby improving the user experience of the product.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of product optimization, and in particular to a product optimization method and device based on user feedback, equipment, medium and product. BACKGROUND

[0002] As the core basis for product iteration, the analysis quality of user feedback directly determines the accuracy of user experience optimization of product functions. However, with the increasing complexity of product functions, the traditional user feedback processing mode of manual sampling and keyword retrieval has been difficult to meet the needs of technological development.

[0003] The prior art analyzes user opinions through clustering processing and other methods by capturing a large amount of user text comment data, but has the defect of only analyzing text content without checking screenshots related to text comments, which easily misses key information and cannot accurately restore the fault scenario. Moreover, when processing user feedback, the prior art uses the same set of priority rules for all functional problems and scores all user feedback, which cannot dynamically determine the priority in combination with the actual scene and cannot accurately identify the fault problems that users urgently need to solve. SUMMARY

[0004] The present application provides a product optimization method and device based on user feedback, equipment, medium and product to solve the problem of ignoring user picture comments and being unable to dynamically determine the priority of user feedback in the prior art.

[0005] In a first aspect, the present application provides a product optimization method based on user feedback, which comprises:

[0006] Correlating picture data, voice data or first text data of user evaluation of a product to obtain user feedback data, and obtaining a plurality of feedback opinions to be processed according to user feedback data of a plurality of users;

[0007] Based on a pre-trained knowledge graph architecture, a multi-dimensional classification result of each feedback opinion to be processed is obtained;

[0008] Based on the multi-dimensional classification result, the score of the feedback opinion to be processed in a plurality of basic dimensions is determined;

[0009] The quantum amplitude of each basic dimension is determined, and the quantum state priority of the feedback opinion to be processed is obtained according to the score and quantum amplitude of the feedback opinion to be processed in each basic dimension; wherein the quantum amplitude is used to represent the weight of the corresponding basic dimension;

[0010] Each feedback opinion to be processed is sorted based on the quantum state priority, a target product optimization scheme is generated according to the sorting result, and the product is optimized according to the target product optimization scheme.

[0011] This invention performs correlation analysis on multimodal user evaluation data, achieving semantic association between text, voice, and images to obtain highly reliable user feedback data. Multiple feedback opinions to be processed are extracted from user feedback data from multiple users. A knowledge graph architecture is used to achieve multidimensional classification of these feedback opinions, refining the classification and making it more aligned with real-world scenarios. The multidimensional classification results are used to evaluate the scores of the feedback opinions across multiple basic dimensions. Based on the quantum amplitude of each basic dimension, the quantum state priority of each feedback opinion is obtained, thus integrating the multidimensional classification results into the quantum state priority calculation process, making priority determination more flexible. Finally, each feedback opinion is ranked based on its quantum state priority, and a target product optimization plan is generated based on the ranking results to optimize the product and improve the user experience.

[0012] In one optional implementation, the multidimensional classification results include the product functions, scenario tags, and multiple user needs associated with each piece of feedback to be processed, including preset core needs; based on the multidimensional classification results, the scores of the feedback to be processed in multiple basic dimensions are determined, including:

[0013] Identify the product functions associated with the feedback to be processed, and obtain the basic value score of the feedback based on the business data of the product functions.

[0014] Based on the multiple user needs associated with the feedback to be processed, at least one knowledge graph path is determined from the feedback to be processed to the preset core needs. Based on the at least one knowledge graph path, the causal strength score of the feedback to be processed is obtained.

[0015] Determine the scene tags associated with the feedback to be processed, and obtain the scene urgency score of the feedback to be processed based on the scene usage time data and scene accident data corresponding to the scene tags.

[0016] Based on the feedback to be processed and the preset core requirements, the requirement matching score of the feedback to be processed is obtained.

[0017] This invention quantifies and evaluates feedback to be processed based on mutually orthogonal fundamental values, causal strength, scenario urgency, and demand matching, so as to avoid the limitations of single-dimensional evaluation when calculating the quantum state priority of feedback to be processed and ensure the credibility of priority evaluation.

[0018] In one optional implementation, a causal strength score for the feedback to be processed is obtained based on at least one knowledge graph path, including:

[0019] Based on at least one knowledge graph path, obtain the target knowledge graph path with the shortest path length;

[0020] a first basic score is obtained according to a path length of the target knowledge graph path;

[0021] a second basic score is obtained according to an associated edge occurrence frequency of the target knowledge graph path;

[0022] The first basic score and the second basic score are linearly weighted to obtain a causal strength score of the feedback opinion to be processed.

[0023] Based on the knowledge graph architecture, the shortest path length of the feedback opinion to be processed reaching the preset core demand and the associated edge occurrence frequency are determined, the correlation strength between the feedback opinion to be processed and the preset core demand is quantitatively analyzed according to the shortest path length, and the shorter the path is, the more directly the feedback opinion to be processed can affect the preset core demand. The mentioned associated edge occurrence frequency is used to quantitatively analyze the mentioned mentioned frequency of a user demand. Then, the shortest path length and the associated edge occurrence frequency are evaluated by scores to obtain the causal strength score of the feedback opinion to be processed, and the causal strength between the feedback opinion to be processed and the user demand is quantified.

[0024] In an optional implementation, a demand matching degree score of the feedback opinion to be processed is obtained according to the feedback opinion to be processed and the preset core demand, including:

[0025] A third basic score corresponding to the preset core demand is determined, and a semantic similarity between the feedback opinion to be processed and the preset core demand is calculated;

[0026] The third basic score is modified according to the semantic similarity to obtain the demand matching degree score of the feedback opinion to be processed.

[0027] Based on the demand type of the preset core demand, the third basic score is determined, and the third basic score is modified according to the semantic similarity between the feedback opinion to be processed and the preset core demand to obtain the demand matching degree score, so that the semantic correlation and matching degree between the feedback opinion and the core demand are deeply understood, and the expected product improvement direction of the user is highly matched with the defined core demand.

[0028] In an optional implementation, a basic value score of the feedback opinion to be processed is obtained according to business data of a product function, including:

[0029] According to the business data of the product function, a function type, a user renewal correlation degree and a feedback user proportion of the product function are obtained, wherein the function type includes a paid function and a non-paid function;

[0030] Based on the function type and the user renewal correlation degree, a basic value weight is determined;

[0031] According to the basic value weight and the feedback user proportion, the basic value score of the feedback opinion to be processed is calculated.

[0032] The application identifies a product function associated with the feedback to be processed, calculates a basic value score of the feedback to be processed according to a function type of the product function, a user renewal correlation degree and a feedback user proportion, thereby quantitatively evaluating the basic value of the feedback to be processed, and optimizing the product by considering the basic value of the product function, and ensuring that resources can be preferentially invested in high-value product functions.

[0033] In an optional implementation, a quantum amplitude of each basic dimension is determined, and a quantum state priority of the feedback to be processed is obtained according to the score and the quantum amplitude of the feedback to be processed in each basic dimension, comprising:

[0034] A quantum amplitude constraint function is constructed, and the quantum amplitude of each basic dimension is obtained according to the quantum amplitude constraint function;

[0035] The score and the quantum amplitude of the feedback to be processed in each basic dimension are weighted and summed to obtain the quantum state priority of the feedback to be processed.

[0036] The application determines the quantum amplitude through the quantum amplitude constraint function, balances and adjusts each basic dimension, and can capture the interaction between the basic dimensions compared with the traditional linear weighting, thereby realizing dynamic adjustment of the weight in different scenarios. Furthermore, the score and the quantum amplitude of each basic dimension are weighted and summed to calculate the quantum state priority, thereby providing an improvement direction for product optimization.

[0037] In an optional implementation, each feedback to be processed is sorted based on the quantum state priority, comprising:

[0038] A constraint condition for the feedback to be processed is constructed, and an initial sorting sequence is determined based on the size order of the quantum state priority of each feedback to be processed;

[0039] The initial sorting sequence is adjusted based on the quantum state priority and the constraint condition to obtain a sorting result meeting the constraint condition.

[0040] The application firstly performs preliminary sorting on the feedback to be processed by using the size order of the quantum state priority to obtain an initial sorting sequence, and then adjusts the initial sorting sequence to obtain a sorting result meeting the constraint condition, thereby avoiding falling into local optimization by only sorting according to the quantum state priority.

[0041] In an optional implementation, a target product optimization scheme is generated according to the sorting result, comprising:

[0042] screen the plurality of to-be-processed feedback opinions according to the sorting result, obtain a target to-be-processed feedback opinion, and determine a cause-and-effect label of the target to-be-processed feedback opinion; wherein the cause-and-effect label comprises an influence result of the target to-be-processed feedback opinion on user experience;

[0043] input the target to-be-processed feedback opinion and the cause-and-effect label and the scene label of the target to-be-processed feedback opinion into a pre-trained product optimization model, and obtain a plurality of product optimization schemes;

[0044] screen the plurality of product optimization schemes, and obtain a target product optimization scheme.

[0045] The application combines the specific content of the target to-be-processed feedback opinion, the cause-and-effect label and the scene label, uses the pre-trained product optimization model to mine the cause-and-effect correlation and the scene correlation between user feedback, outputs a plurality of product optimization schemes, and selects a suitable target product optimization scheme, thereby improving the accuracy and scene adaptability of the product optimization scheme.

[0046] In an optional implementation, screening the plurality of product optimization schemes to obtain the target product optimization scheme comprises:

[0047] inputting each product optimization scheme into a pre-trained optimization prediction model to obtain an optimization effect of each product optimization scheme;

[0048] screening the plurality of product optimization schemes based on the optimization effect to obtain the target product optimization scheme.

[0049] The application predicts the optimization effect of each product optimization scheme through the pre-trained optimization prediction model, selects a target product optimization scheme with the optimal comprehensive effect to optimize the product, so as to quickly make the corresponding product function achieve the user expected effect and reduce the user feedback rate.

[0050] In an optional implementation, the user feedback data is obtained by associating picture data, voice data or first text data of the user evaluating the product, and the user feedback data comprises:

[0051] performing feature extraction on the picture data to obtain image feature labels;

[0052] convert the voice data into second text data, and associate the image feature labels, the second text data or the first text data to obtain the user feedback data.

[0053] The application realizes unified processing and multi-modal semantic association of text, voice and picture data by extracting image feature tags of picture data and converting voice data of user evaluation into second text data, obtains user feedback data, and significantly improves the integrity, accuracy and effectiveness of user evaluation data processing, thereby providing high-quality data support for subsequent user feedback opinion analysis and solution.

[0054] In a second aspect, the application provides a product optimization device based on user feedback, which comprises:

[0055] A first processing module is configured to associate picture data, voice data or first text data of user evaluation of a product to obtain user feedback data, and obtain a plurality of feedback opinions to be processed based on user feedback data of a plurality of users.

[0056] A second processing module is configured to obtain a multi-dimensional classification result of each feedback opinion to be processed based on a pre-trained knowledge graph architecture.

[0057] A third processing module is configured to determine a score of the feedback opinion to be processed in a plurality of basic dimensions based on the multi-dimensional classification result.

[0058] A fourth processing module is configured to determine a quantum amplitude of each basic dimension, and obtain a quantum state priority of the feedback opinion to be processed based on the score and the quantum amplitude of the feedback opinion to be processed in each basic dimension; wherein the quantum amplitude is used to represent the weight of the corresponding basic dimension.

[0059] A fifth processing module is configured to sort each feedback opinion to be processed based on the quantum state priority, generate a target product optimization scheme based on the sorting result, and optimize the product based on the target product optimization scheme.

[0060] In a third aspect, the application provides an electronic device, which comprises a memory and a processor, the memory and the processor are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the product optimization method based on user feedback of the first aspect or any of the corresponding embodiments thereof.

[0061] In a fourth aspect, the application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are used to make a computer execute the product optimization method based on user feedback of the first aspect or any of the corresponding embodiments thereof.

[0062] In a fifth aspect, the application provides a computer program product, which comprises computer instructions, and the computer instructions are used to make a computer execute the product optimization method based on user feedback of the first aspect or any of the corresponding embodiments thereof.

[0063] The beneficial effects of the present application are:

[0064] The present application correlates and analyzes multi-modal evaluation data of users, realizes semantic correlation of text, voice and pictures, and obtains user feedback data with high reliability. A plurality of feedback opinions to be processed are extracted from user feedback data of a plurality of users, multi-dimensional classification of the feedback opinions to be processed is realized by using a knowledge graph architecture, classification refinement of the feedback opinions to be processed is realized, and the classification is more consistent with the actual scene. The multi-dimensional classification result is used to evaluate the scores of the feedback opinions to be processed in a plurality of basic dimensions, and a quantum state priority of the feedback opinions to be processed is obtained according to the quantum amplitude of each basic dimension, so that the multi-dimensional classification result is integrated into the calculation process of the quantum state priority, and the priority judgment is more flexible. Finally, each feedback opinion to be processed is sorted based on the quantum state priority, a target product optimization scheme is generated according to the sorting result, and the product is optimized to improve the user experience of the product. BRIEF DESCRIPTION OF DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings needed in the specific embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0066] Figure 1 is a schematic diagram of an application scenario according to an embodiment of the present application;

[0067] Figure 2 is a first flowchart of a product optimization method according to an embodiment of the present application;

[0068] Figure 3 is a second flowchart of a product optimization method according to an embodiment of the present application;

[0069] Figure 4 is a third flowchart of a product optimization method according to an embodiment of the present application;

[0070] Figure 5 is a structural block diagram of a product optimization device according to an embodiment of the present application;

[0071] Figure 6 is a hardware structure schematic diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0072] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0073] It can be understood that, before using the technical solutions disclosed in the embodiments of the present application, the type of personal information involved in the present application, the use range, the use scenario and the like should be informed to the user and the authorization of the user should be obtained through appropriate means according to relevant laws and regulations.

[0074] The terms "first", "second" are only for descriptive purpose, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0075] As an optional application scenario of the embodiments of the present application, as shown in Figure 1 The product optimization system based on user feedback can include at least one terminal device and at least one server, Figure 1 The system is exemplarily shown in the figure, which includes a computer 101, a mobile terminal 102 and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110. The terminal device can be specifically a smart phone, a tablet computer, a notebook computer, a palm computer, and can also be a desktop computer, a game console, a smart television, a smart wearable device, a vehicle-mounted terminal, a VR (Virtual Reality) device, an AR (Augmented Reality) device, etc. The server 103 can be an independent physical server, or a server cluster or a distributed system, or a cloud server providing cloud services. The network 110 can be a wired network or a wireless network, and examples thereof include but are not limited to the Internet, an enterprise intranet, a local area network, a wide area network, a mobile communication network and a combination thereof.

[0076] With the deep transformation of the automotive industry towards intelligence and networking, the function boundary of the intelligent cockpit gradually expands from the traditional basic services such as central control entertainment and navigation to a complex system covering multi-modal interaction (voice, touch, gesture, etc.), scenario-based services (active navigation, health monitoring), personalized cockpit settings (seats, atmosphere lights, sound effect linkage), etc. The intelligent cockpit integrates cockpit domain controller, multi-screen interaction, voice assistant, vehicle-mounted sensor, etc. The feedback data generated by users in actual use presents the characteristics of large volume, multiple forms (text, voice, pictures, operation logs, etc.) and close scene association. As the core basis for product iteration, the analysis quality of user feedback directly determines the accuracy of cockpit experience optimization.

[0077] However, as the functional complexity of the intelligent cockpit increases, the traditional user feedback processing mode mainly based on manual sampling and keyword retrieval has been difficult to meet the needs of technological development. In the face of millions of user feedback data, it is difficult to capture the relevance of cross-scenario problems and quickly respond to feedback changes brought by new functions, which causes problems such as mismatch and timeliness between user demand and product improvement. However, when the prior art tries to use a large model to solve the problem of the traditional user feedback processing mode, there are still problems of ignoring non-text information and priority judgment being not flexible enough.

[0078] The embodiment of the present application provides a product optimization method based on user feedback, which is used for associatively analyzing multi-modal evaluation data of users to obtain user feedback data with high reliability. A plurality of feedback opinions to be processed are extracted from user feedback data of a plurality of users, and a knowledge graph architecture is used to realize multi-dimensional classification of the feedback opinions to be processed. The multi-dimensional classification results are used to evaluate scores of the feedback opinions to be processed in a plurality of basic dimensions, and a quantum state priority of the feedback opinions to be processed is obtained according to quantum amplitudes of each basic dimension. Finally, each feedback opinion to be processed is sorted based on the quantum state priority, and a target product optimization scheme is generated according to the sorting result to optimize the product. The present application can realize semantic association analysis of text, voice and pictures, and through classification and refinement of the feedback opinions to be processed, the actual scene is more fitted, and the multi-dimensional classification results are integrated into the calculation process of the quantum state priority, so that the priority judgment is more flexible.

[0079] According to the embodiment of the present application, a product optimization method based on user feedback is provided. It should be noted that the steps shown in the flowchart can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0080] In the present embodiment, a product optimization method based on user feedback is provided, which can be used for the terminal device described above, such as a computer, a tablet computer, etc.Figure 2 is a flowchart of a product optimization method based on user feedback according to an embodiment of the present application, as shown in the figure, the flow includes the following steps: Figure 2

[0081] Step S201, the picture data, voice data or first text data of the user evaluating the product are associated to obtain user feedback data, and a plurality of feedback opinions to be processed are obtained according to the user feedback data of a plurality of users.

[0082] Specifically, taking the intelligent cockpit as an example, the user can evaluate the function design, use experience, fault problem, etc. of the intelligent cockpit through the feedback function built in the car machine system, the mobile APP matched with the intelligent cockpit, the social media and network platform, the application store, the customer service system, etc. Through the API data interface, the evaluation data of the feedback function of the car machine system, the mobile APP and other applications are collected, and various evaluation data of the social media and network platform are crawled, including but not limited to customer complaint data, application store comments, social media feedback, user research text, customer service dialogue record, etc. Text, picture, voice and other evaluation data.

[0083] The related technology uses a large model to analyze the text content, although it can process a large amount of text content in batches, but the utilization of user evaluation data is not sufficient, and key information is easily missed. For example, when the user evaluates "car machine black screen screenshot + text description", analyzing the text alone or analyzing the picture alone will miss the key information, and only looking at the text description of the functional abnormality without checking the screenshot of the black screen moment, it is also impossible to restore the interface state and operation scene when the fault is triggered, and the data reliability is low.

[0084] The embodiment establishes the association of evaluation data of different data types, such as picture data, voice data or first text data, to obtain user feedback data, and improves the reliability of user feedback data. And collect the user feedback data of a plurality of users, and extract a plurality of feedback opinions to be processed from them.

[0085] Step S202, based on the pre-trained knowledge graph architecture, a multi-dimensional classification result of each feedback opinion to be processed is obtained.

[0086] Specifically, the pre-trained knowledge graph architecture and the semantic analysis capability of the large language model are used to classify each feedback opinion to be processed to obtain a multi-dimensional classification result, which includes the product function, scene label and multiple user demands associated with each feedback opinion to be processed.

[0087] ​In some embodiments, the knowledge graph architecture includes an entity layer and a relationship layer, wherein the entity layer contains feedback entities, function entities, scenario entities and demand entities, and the relationship layer reflects the association relationship between the feedback entities and other entities. Based on the pre-trained knowledge graph architecture, a large language model is used to identify the feedback entity corresponding to the feedback opinion to be processed; according to the function entity associated with the feedback entity, the product function associated with the feedback opinion to be processed is determined, and then the problem module is automatically located; according to the scenario entity associated with the feedback entity, the scenario label of the feedback opinion to be processed is obtained, the scenario is refined, and general classification is avoided; according to the demand entity associated with the feedback entity, a plurality of user demands associated with the feedback opinion to be processed are determined, and the user experience short board is helped to be located by the R&D personnel.

[0088] It should be noted that the user demand includes a preset core demand and a preset intermediate demand, and whether the user demand is an intermediate demand or a core demand can be determined by a pre-stored product strategy document library, and the product strategy document library includes pre-defined core demands and intermediate demands. In the knowledge graph architecture, each user demand can have an association relationship with other user demands, for example, a feedback opinion to be processed can be associated with a preset intermediate demand, and the preset intermediate demand is associated with a preset core demand, that is, each feedback opinion to be processed can be associated with a plurality of user demands.

[0089] In this embodiment, by using the knowledge graph architecture and the semantic analysis capability of the large language model, accurate classification and dynamic adaptation of the feedback opinion to be processed are realized from multiple dimensions (function, scenario and demand), and according to the identified feedback entity type of the feedback opinion to be processed, the product function positioning, scenario label identification and user demand identification of the feedback opinion to be processed are realized in combination with the association relationship of the relationship layer.

[0090] In some embodiments, the meta-learning technology (Model-Agnostic Meta-Learning, MAML) can be used to realize the rapid self-adaptation of the knowledge graph, realize the effect of rapidly adding feedback entities and automatically establishing the association relationship between the feedback entities and other entities, and adapt to new functions and new scenarios without retraining.

[0091] Exemplarily, a plurality of tasks are sampled from the historical knowledge graph, each task containing a target feedback entity (simulating a new feedback entity) and a plurality of related triples, each triple containing the association relationship between the target feedback entity and a related function entity or a scene entity or a demand entity, and the triples of each task are divided into a support set and a query set. The support set and the query set of each task are used to train the initial knowledge graph architecture to obtain a pre-trained knowledge graph architecture, so that the pre-trained knowledge graph architecture can autonomously build and update the classification system, and realize automatic addition of the new feedback entity and its associated entities. The specific principle of the meta-learning technology can be referred to the description of the related technology, and will not be repeated here.

[0092] The related technology realizes fault problem classification by using a fixed problem classification directory, and cannot automatically adapt to new product functions. When a new function is added, the structure of the problem classification directory needs to be manually adjusted. For example, after adding the "vehicle AR navigation" function, when a new problem such as "vehicle AR navigation conversation confusion" occurs, a new sub-item needs to be manually added to the original problem classification directory, which is cumbersome to operate.

[0093] The pre-trained knowledge graph architecture is used to perform multi-dimensional classification on the feedback opinions to be processed, so as to realize product function positioning, scene label identification and user demand identification of the feedback opinions to be processed. Compared with the traditional classification rule, the pre-trained knowledge graph architecture can realize automatic addition of new feedback entities and their associated entities without retraining the entire knowledge graph architecture.

[0094] In step S203, scores of the feedback opinions to be processed in a plurality of basic dimensions are determined based on the multi-dimensional classification results.

[0095] Specifically, the feedback opinions to be processed are scored in a plurality of basic dimensions according to the product functions, scene labels and a plurality of user demands associated with the feedback opinions to be processed, wherein the plurality of basic dimensions include but are not limited to basic value, causal strength, scene urgency and demand matching degree, and each basic dimension is independent of each other.

[0096] In step S204, the quantum amplitude of each basic dimension is determined, and the quantum state priority of the feedback opinions to be processed is obtained according to the score and the quantum amplitude of the feedback opinions to be processed in each basic dimension.

[0097] Specifically, the quantum state priority of the feedback opinions to be processed is calculated by using a quantum state system, and the quantum state system contains four independent basic dimensions. Each basic dimension is scored by an interpretable quantification method. The calculation function of the quantum state priority is:

[0098]

[0099] wherein, represents the quantum state priority, representing a base value score, representing a quantum amplitude of the base value, representing a causal strength score, representing a quantum amplitude of the causal strength, representing a scenario urgency score, representing a quantum amplitude of the scenario urgency, representing a demand matching degree score, representing a quantum amplitude of the demand matching degree. It should be noted that, are four orthogonal (independent of each other) base dimension quantum state scores, and quantum amplitudes are used to represent the weights of the corresponding base dimensions.

[0100] The related technology uses the same set of priority rules to score all user feedback, calculates the priority in a fixed weight manner, and does not distinguish between user scenarios. For example, northern users need accurate seat heating in winter, while southern users may pay more attention to real-time road conditions of the vehicle navigation. The related technology will evaluate the priority according to the same weight, and will ignore the difference that the northern user needs to solve the seat heating first and the southern user needs to focus on the vehicle navigation problem, resulting in that the priority evaluation result does not match the user use scenario.

[0101] The embodiment breaks through the limitation of the traditional linear weighting model through quantum state priority, and solves the scenario heterogeneity problem in user feedback priority evaluation. Compared with the disadvantage of the traditional model using fixed weight for all scenarios, the scheme deeply integrates the classification characteristics such as base value, causal strength, scenario urgency and demand matching degree of the feedback opinions to be processed into the calculation process of quantum state priority, and realizes dynamic evaluation of quantum state priority based on real user feedback data.

[0102] In step S205, each feedback opinion to be processed is sorted based on the quantum state priority, and a target product optimization scheme is generated according to the sorting result, and the product is optimized according to the target product optimization scheme.

[0103] Specifically, the feedback opinions to be processed are screened based on the quantum state priority, and a target product optimization scheme is generated for the feedback opinions to be processed with high quantum state priority, and the product is optimized by using the target product optimization scheme, thereby improving the user experience.

[0104] ​​​This embodiment provides a product optimization method based on user feedback. It performs correlation analysis on multimodal user evaluation data, achieving semantic association between text, voice, and images to obtain highly reliable user feedback data. Multiple feedback opinions to be processed are extracted from the user feedback data of multiple users. A knowledge graph architecture is used to achieve multidimensional classification of these feedback opinions, refining the classification and making it more aligned with real-world scenarios. The multidimensional classification results are used to evaluate the scores of the feedback opinions on multiple basic dimensions. Based on the quantum amplitude of each basic dimension, the quantum state priority of the feedback opinions is obtained, thus integrating the multidimensional classification results into the quantum state priority calculation process, making priority determination more flexible. Finally, each feedback opinion to be processed is ranked based on its quantum state priority, and a target product optimization plan is generated based on the ranking results, thereby optimizing the product and improving the user experience.

[0105] This embodiment provides a product optimization method based on user feedback, which can be used in the aforementioned terminal devices, such as computers and tablets. Figure 3 This is a flowchart of a product optimization method based on user feedback according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps:

[0106] Step S301: Associate the image data, voice data, or first text data of user reviews of the product to obtain user feedback data. Based on the user feedback data of multiple users, obtain multiple feedback opinions to be processed.

[0107] Specifically, step S301 includes:

[0108] Step S3011: Extract features from the image data to obtain image feature labels.

[0109] Specifically, for each user, a trained image feature extraction model is used to identify the content of the image data evaluated by that user, and image feature labels are obtained.

[0110] In some embodiments, an original image feature extraction model and an image dataset are obtained. The image dataset includes various image data of the product and image feature labels annotated for each image data. The original image feature extraction model is trained based on the image dataset to obtain a trained image feature extraction model.

[0111] It should be noted that image feature labels include, but are not limited to, “screen distortion”, “black screen”, “reflection”, “indicator light on”, and “indicator light off”. The specific labels can be set according to the various states that occur during product use. The image feature extraction model can be a ResNet model, and its training process can be referred to the description of relevant technologies, which will not be elaborated here.

[0112] Step S3012, convert the voice data into second text data, associate the image feature label, the second text data or the first text data, and obtain user feedback data.

[0113] Specifically, the voice data of the user evaluation is converted into second text data using a voice-to-text model, such as the Whisper model. Then, the image feature label corresponding to the corresponding user evaluation data, the first text data, and the second text data converted by the voice are associated, and output in the structure of "text description + image feature label", to obtain the user feedback data of the user. Among them, the text description in the user feedback data includes the first text data and the second text data.

[0114] In some embodiments, a contrastive language-image pre-training (CLIP) model can be used to calculate the similarity between the image feature label of a certain user evaluation data and the text description (second text data and first text data). When the similarity reaches a set similarity threshold (for example, 0.8, which can be adjusted according to actual needs), it is confirmed that the text, picture and voice data of the user evaluation describe the same scene, and then the image feature label, the second text data and the first text data are associated, thereby establishing a connection between the evaluation data of different data types, realizing cross-modal semantic association, and improving the credibility of the user feedback data.

[0115] The present embodiment realizes unified processing and multi-modal semantic association of text, voice and picture data by extracting the image feature label of the picture data and converting the voice data of the user evaluation into second text data, obtains user feedback data, and significantly improves the completeness, accuracy and effectiveness of the user evaluation data processing, providing high-quality data support for subsequent user feedback analysis and solution.

[0116] Step S3013, obtaining a plurality of feedback opinions to be processed according to a plurality of user feedback data.

[0117] Specifically, the user feedback opinions of a plurality of users are split and clustered to obtain a plurality of feedback clusters, and the center feedback opinion of each feedback cluster is identified to obtain a plurality of feedback opinions to be processed.

[0118] In the embodiment, for each user, the text description in the user feedback data of the user contains at least one feedback opinion, and the image feature label contains a sub-feature label corresponding to each feedback opinion. For example, the text description of a piece of user feedback data is "seat heating failure, navigation slow to wake up", and the image feature label includes "seat heating button light off" and "interface unresponsive", wherein "seat heating button light off" is a sub-feature label corresponding to "seat heating failure", and "interface unresponsive" is a sub-feature label corresponding to "navigation slow to wake up".

[0119] In some embodiments, for each user feedback data, the text description in the user feedback data is semantically split into multiple feedback opinions by using punctuation segmentation and a large language model, and a CLIP model is used to identify a sub-feature label corresponding to each feedback opinion, so as to ensure the matching degree of the feedback opinions and the sub-feature labels. Further, multiple feedback opinions with consistent sub-feature labels and a semantic similarity exceeding a semantic similarity threshold are clustered into the same feedback cluster. Through the above steps of splitting and clustering the user feedback opinions, multiple feedback clusters are obtained.

[0120] In some embodiments, for each feedback cluster, a sentence vector of each feedback opinion in the feedback cluster is determined, an arithmetic mean of all sentence vectors in the cluster is calculated to obtain a geometric center vector. The feedback opinion with the highest cosine similarity to the geometric center vector is determined as the center feedback opinion of the feedback cluster. By identifying the center feedback opinion of each feedback cluster, multiple to-be-processed feedback opinions are obtained.

[0121] In some embodiments, a large language model is used to predict the causal label and potential value of each center feedback opinion, and the center feedback opinions are screened according to the causal label and the potential value to obtain multiple to-be-processed feedback opinions. The causal label is used to represent the impact of the feedback opinion on the user experience, and the potential value is used to represent the benefits of solving the feedback opinion.

[0122] Exemplarily, each central feedback opinion is input into the large language model, and a thinking chain prompt word is constructed to guide the large language model to analyze the influence result of the central feedback opinion on the user experience and the benefits brought by solving the central feedback opinion. For example, the large language model identifies that the "navigation wake-up is slow", and deduces that the influence result is "navigation response delay", and then the causal label of "navigation wake-up is slow" is "navigation wake-up is slow navigation response delay". Meanwhile, combined with historical data, similar cases and domain knowledge, the potential value after solving the problem is quantified as "navigation delay rate decreases by 60%". Finally, the large language model filters high-value central feedback opinions based on the causal label and potential value of each central feedback opinion, and obtains a plurality of feedback opinions to be processed. Thus, by using a two-dimensional evaluation system of causal label and potential value, the specific, operability, causal correlation and demand urgency of the central feedback opinion are comprehensively considered for semantic filtering, and the high-value feedback opinions to be processed are deeply analyzed and filtered, thereby improving the processing efficiency and accuracy of user feedback.

[0123] In step S302, based on the pre-trained knowledge graph architecture, a multi-dimensional classification result of each feedback opinion to be processed is obtained. For details, please refer to Figure 2 The step S202 of the embodiment shown is not repeated here.

[0124] In step S303, based on the multi-dimensional classification result, the score of the feedback opinion to be processed in the plurality of basic dimensions is determined.

[0125] Specifically, the above step S303 includes:

[0126] In step S3031, the product function associated with the feedback opinion to be processed is determined, and the basic value score of the feedback opinion to be processed is obtained according to the business data of the product function.

[0127] In some optional embodiments, the above step S3031 includes:

[0128] In step a1, the function type, user renewal correlation and feedback user proportion of the product function are obtained according to the business data of the product function; wherein the function type includes a paid function and a non-paid function.

[0129] Specifically, the business data of the product function is obtained, including but not limited to the business data center (including paid function list, user renewal data) and user behavior log (total number of users, number of feedback users).

[0130] In some embodiments, according to the paid function list, it is judged whether the function type of the product function associated with the feedback to be processed is a paid function or a non-paid function. According to the ratio between the number of feedback users and the total number of users, the proportion of feedback users is obtained. According to the user renewal data, the proportion of renewal users of the product function associated with the feedback to be processed is obtained, and if the proportion of renewal users is greater than the threshold value (for example, 0.2) of the proportion of renewal users, the user renewal association degree is set to 1; otherwise, the user renewal association degree is set to 0.

[0131] Step a2, based on the function type and the user renewal association degree, the basic value weight is determined.

[0132] Specifically, a multi-level weighting algorithm is used to superimpose the weight corresponding to the function type and the weight corresponding to the user renewal association degree to obtain the basic value weight. For example, the weight of the paid function can be 1.2, the weight of the non-paid function can be 0.8, the weight corresponding to the user renewal association degree when it is set to 1 is 0.3, and the weight corresponding to the user renewal association degree when it is set to 0 is 0.

[0133] Step a3, according to the basic value weight and the proportion of feedback users, the basic value score of the feedback to be processed is calculated.

[0134] Specifically, according to the product of the basic value weight and the proportion of feedback users, the original score is calculated, and through the minimum-maximum normalization, the original score is mapped to 0-1 to obtain the basic value score of the feedback to be processed. It should be noted that the maximum value of the original score can be obtained by multiplying the maximum basic value weight and the maximum proportion of feedback users, and the minimum value of the original score can be obtained by multiplying the minimum basic value weight and the minimum proportion of feedback users.

[0135] In some embodiments, the product function associated with the feedback to be processed is automatically identified, the daily business data of the product function is synchronized through the ETL tool, the original score is calculated according to the business data, and the normalization interface is called to automatically calculate the basic value score of the feedback to be processed and output.

[0136] The embodiment identifies the product function associated with the feedback to be processed, calculates the basic value score of the feedback to be processed according to the function type, the user renewal association degree and the proportion of feedback users of the product function, and quantitatively evaluates the basic value of the feedback to be processed, so as to consider the basic value of the product function to optimize the product and ensure that resources can be preferentially invested in high-value product functions.

[0137] Step S3032, based on the plurality of user demands associated with the feedback to be processed, at least one knowledge graph path from the feedback to be processed to the preset core demand is determined, and according to the at least one knowledge graph path, the causal strength score of the feedback to be processed is obtained.

[0138] Specifically, since each user demand can have a correlation with other user demands, i.e., there is a path connection. For each feedback opinion to be processed, at least one knowledge graph path with the feedback opinion as the starting point and the preset core demand as the ending point is determined by traversing the knowledge graph architecture.

[0139] In some optional embodiments, the step S3032 comprises:

[0140] Step b1, obtaining a target knowledge graph path with the shortest path length from the at least one knowledge graph path.

[0141] Specifically, the knowledge graph path comprises at least one edge, and the length of each edge can be set to 1, so as to obtain the path length of the knowledge graph path, and the target knowledge graph path with the shortest path length is selected from the at least one knowledge graph path.

[0142] Step b2, obtaining a first basic score according to the path length of the target knowledge graph path.

[0143] Specifically, the target knowledge graph path is scored based on the path length, and the first basic score is obtained. For example, the first basic score is 0.9 when the path length is 1, the first basic score is 0.7 when the path length is 2, and the first basic score is 0.4 when the path length is greater than or equal to 3.

[0144] Step b3, calculating the correlation edge occurrence frequency of the target knowledge graph path, and obtaining a second basic score according to the correlation edge occurrence frequency.

[0145] In some embodiments, the occurrence frequency of each edge of the at least one knowledge graph path is counted, and the occurrence frequency of each edge is summed to obtain a total occurrence frequency (i.e., the total path length of all knowledge graph paths). The occurrence frequency of each edge of the target knowledge graph path is summed to obtain a target path edge occurrence frequency. The ratio between the target path edge occurrence frequency and the total occurrence frequency is calculated to obtain the correlation edge occurrence frequency of the target knowledge graph path.

[0146] Specifically, the second basic score is scored according to the size of the correlation edge occurrence frequency. For example, the second basic score increases by 0.05 for every 5% increase in the correlation edge occurrence frequency.

[0147] Step b4, linearly weighting the first basic score and the second basic score to calculate a causal strength score of the feedback opinion to be processed.

[0148] Specifically, the first and second basic scores are weighted and summed, for example, the first basic score weight can be 0.4 and the second basic score weight can be 0.6, and the result is mapped to 0-1 through min-max normalization to obtain the causal strength score of the feedback opinion to be processed.

[0149] In some embodiments, the nodes and edges in the knowledge graph are updated periodically (such as weekly) through a crawler and an NLP analysis tool, a graph computing algorithm is called to automatically count the path length and associated edge occurrence frequency of the target knowledge graph path, and the weighted formula is substituted to complete the automatic calculation and update of the causal strength.

[0150] The present embodiment is based on a knowledge graph architecture to determine the shortest path length and associated edge occurrence frequency of the feedback processing opinion to the preset core demand, and to quantitatively analyze the correlation strength between the feedback processing opinion and the preset core demand according to the shortest path length. The shorter the path, the more directly the feedback processing opinion can affect the preset core demand. The occurrence frequency of the associated edge is used to quantitatively analyze the mentioned frequency of a user's demand. Further, by scoring the shortest path length and the occurrence frequency of the associated edge, the causal strength score of the feedback processing opinion is obtained, and the causal strength between the feedback processing opinion and the user demand is quantified.

[0151] Step S3033, determine the scene label associated with the feedback opinion to be processed, and obtain the scene emergency degree score of the feedback opinion to be processed according to the scene usage time length data and the scene accident data corresponding to the scene label.

[0152] Specifically, the scene usage time length proportion corresponding to the scene label is obtained through the scene usage time length data of the user behavior analysis platform; and the scene accident proportion corresponding to the scene label is obtained through the scene accident data recorded in the historical complaint database, such as the failure complaint and safety accident association record.

[0153] Further, the fourth basic score is obtained according to the scene usage time length proportion, and the fourth basic score is corrected based on the scene accident proportion to obtain the fifth basic score, for example, the fifth basic score = scene usage time length proportion x 1.0 + scene accident proportion x 0.6. The fifth basic score is mapped to the interval of 0-1 to obtain the scene emergency degree score of the feedback opinion to be processed.

[0154] In some embodiments, the user scene usage data is collected in real time, the daily scene usage time length proportion is counted, the scene accident proportion corresponding to the scene label is automatically identified, and the scene usage time length proportion and the scene accident proportion are substituted into the formula to realize the daily automatic update and early warning of the scene emergency degree.

[0155] The embodiment comprehensively considers the scene use duration data and the scene accident data, quantitatively evaluates the scene emergency degree of the feedback opinion to be processed, obtains a scene emergency degree score, ensures that the problem with high user frequency and high safety risk can be processed preferentially to avoid single index deviation, and can early identify a potential high-risk scene through scene accident correlation analysis.

[0156] In step S3034, a demand matching degree score of the feedback opinion to be processed is obtained according to the feedback opinion to be processed and the preset core demand.

[0157] In some optional embodiments, a third basic score corresponding to the preset core demand is determined, and a semantic similarity between the feedback opinion to be processed and the preset core demand is calculated. The third basic score is corrected according to the semantic similarity, and a demand matching degree score of the feedback opinion to be processed is obtained.

[0158] Specifically, the demand type of the user demand can include a safety type, an experience type and the like. The third basic score corresponding to the preset core demand is obtained based on the demand type corresponding to the preset core demand. For example, the third basic score of the safety type is 0.9, and the third basic score of the experience type is 0.6. The third basic score can be set by a product committee according to actual demand, and the present application is not limited thereto.

[0159] Further, the third basic score is corrected according to the semantic similarity between the feedback opinion to be processed and the preset core demand. For example, the third basic score is increased by 0.05 when the semantic similarity is increased by 10%, and the result is mapped to the interval of 0-1 to obtain the demand matching degree score. The semantic similarity between the feedback opinion to be processed and the preset core demand can be calculated by using a BERT model, and details are described in the related technical description.

[0160] In some embodiments, the third basic score of each demand type is updated according to product strategy adjustment every quarter. After the third basic score is updated, the third basic score is automatically synchronized to a scoring system. The feedback opinion to be processed is connected to a semantic analysis interface of the BERT model in real time to obtain the semantic similarity between the feedback opinion to be processed and the preset core demand. The demand matching degree score is automatically calculated and output according to the rules.

[0161] The embodiment determines the third basic score based on the demand type of the preset core demand, and corrects the third basic score by using the semantic similarity between the feedback opinion to be processed and the preset core demand to obtain the demand matching degree score, so as to deeply understand the semantic association and matching degree between the feedback opinion and the core demand, and ensure that the product improvement direction expected by the user is highly consistent with the defined core demand.

[0162] In the above embodiment, the to-be-processed feedback opinions are quantitatively evaluated from the mutually orthogonal basic values, the causal strength, the scene urgency and the demand matching degree, so as to avoid the limitation of single-dimensional evaluation and guarantee the credibility of the priority evaluation when calculating the quantum state priority of the to-be-processed feedback opinions.

[0163] In step S304, the quantum amplitude of each basic dimension is determined, and the quantum state priority of the to-be-processed feedback opinions is obtained according to the score of the to-be-processed feedback opinions in each basic dimension and the quantum amplitude.

[0164] Specifically, the step S304 includes:

[0165] In step S3041, a quantum amplitude constraint function is constructed, and the quantum amplitude of each basic dimension is obtained according to the quantum amplitude constraint function.

[0166] Specifically, the quantum amplitude of each basic dimension is 、 、 and satisfies the normalized quantum amplitude constraint function, that is, The optimal solution is found on the boundary of the quantum amplitude constraint function by the Lagrange multiplier method, and the quantum amplitude of each basic dimension is adjusted to adapt to different scene demands. For details, refer to the detailed description of the related art, which will not be repeated here.

[0167] In some embodiments, the quantum amplitude is automatically adjusted based on the real-time feedback trend and the classification features of the feedback opinions. For example, the real-time feedback trend is monitored by a sliding window algorithm (such as the feedback amount growth rate and the complaint severity mean in a 7-day sliding window), and when the feedback opinions in a feedback cluster satisfy “growth rate ≥ 150% and severity mean ≥ 0.7”, the quantum amplitude adjustment is triggered.

[0168] In step S3042, the score and the quantum amplitude of the to-be-processed feedback opinions in each basic dimension are weighted and summed to obtain the quantum state priority of the to-be-processed feedback opinions.

[0169] Specifically, according to , the score and the quantum amplitude of each basic dimension are weighted and summed to obtain the quantum state priority of the to-be-processed feedback opinions.

[0170] In this embodiment, the quantum amplitude is determined by the quantum amplitude constraint function, and the basic dimensions are balanced and adjusted. Compared with the traditional linear weighting, the interaction between the basic dimensions can be better captured, and the weight dynamic adjustment in different scenes can be realized. Then, the quantum state priority is calculated by weighting and summing the score and the quantum amplitude of each basic dimension, which provides an improvement direction for product optimization.

[0171] Step S305, sort each feedback opinion based on the quantum state priority, generate a target product optimization scheme according to the sorting result, and optimize the product according to the target product optimization scheme. For details, please refer to Figure 2 Step S205 of the embodiment shown in FIG. 2 will not be described here.

[0172] The product optimization method based on user feedback provided in this embodiment performs correlation analysis on multi-modal evaluation data of users, realizes semantic correlation of text, voice and pictures, and obtains user feedback data with high reliability. A plurality of feedback opinions to be processed are extracted from user feedback data of a plurality of users, multi-dimensional classification of the feedback opinions to be processed is realized by using a knowledge graph architecture, and classification refinement of the feedback opinions to be processed is realized, which is more consistent with the actual scene. The feedback opinions to be processed are quantitatively evaluated from basic value, causal strength, scene urgency and demand matching degree, and quantum amplitudes of each basic dimension are determined according to a quantum amplitude constraint function to obtain quantum state priorities of the feedback opinions to be processed, which are more consistent with the actual scene. Finally, each feedback opinion to be processed is sorted based on the quantum state priority, a target product optimization scheme is generated according to the sorting result, and the product is optimized, thereby improving the user experience of the product.

[0173] In this embodiment, a product optimization method based on user feedback is provided, which can be used for the terminal device described above, such as a computer, a tablet computer, etc. Figure 4 The flowchart of the product optimization method based on user feedback according to the embodiment of the present application is shown in FIG. 1, which includes the following steps: Figure 4

[0174] Step S401, correlate picture data, voice data or first text data evaluated by users to obtain user feedback data, and obtain a plurality of feedback opinions to be processed according to user feedback data of a plurality of users. For details, please refer to Figure 3 Step S301 of the embodiment shown in FIG. 2 will not be described here.

[0175] Step S402, based on the pre-trained knowledge graph architecture, obtain multi-dimensional classification results of each feedback opinion to be processed. For details, please refer to Figure 3 Step S302 of the embodiment shown in FIG. 2 will not be described here.

[0176] Step S403, based on the multi-dimensional classification results, determine scores of the feedback opinions to be processed in a plurality of basic dimensions. For details, please refer to Figure 3 Step S303 of the embodiment shown in FIG. 2 will not be described here.

[0177] ​In step S404, quantum amplitudes of each basic dimension are determined, and a quantum state priority of the feedback to be processed is obtained according to the scores of each basic dimension of the feedback to be processed and the quantum amplitudes. The quantum amplitude is used to represent the weight of the corresponding basic dimension. For details, please refer to Figure 3 In step S304 of the embodiment shown, no further description is given here.

[0178] In step S405, each feedback to be processed is sorted based on the quantum state priority, and a target product optimization scheme is generated according to the sorting result, and product optimization is performed according to the target product optimization scheme.

[0179] Specifically, step S405 includes the following steps.

[0180] In step S4051, a constraint condition for the feedback to be processed is constructed, and an initial sorting sequence is determined based on the size order of the quantum state priority of each feedback to be processed.

[0181] Specifically, first, the feedback to be processed is preliminarily sorted according to the size order of the quantum state priority, and an initial sorting sequence is obtained.

[0182] In this embodiment, the constraint condition is constructed, which is used to adjust the initial sorting sequence. For example, the constraint condition can be that the quantum state priority of the safety class feedback opinion is greater than or equal to 0.8, and the processing time limit of the feedback to be processed is less than or equal to 72 hours.

[0183] In step S4052, the initial sorting sequence is adjusted based on the quantum state priority and the constraint condition, and a sorting result satisfying the constraint condition is obtained.

[0184] Specifically, a quantum annealing algorithm is used to solve the global optimal priority sequence under multiple constraint conditions, the initial sorting sequence is adjusted, and a sorting result satisfying the constraint condition is obtained.

[0185] In some embodiments, the quantum annealing algorithm maps the sorting problem of the feedback to be processed into an energy state optimization problem of a quantum bit, first encodes the sorting position of each feedback as a quantum bit configuration, calculates a priority objective function using the quantum bit configuration, and constructs an energy function including the priority objective function and the constraint condition. Then, in the quantum annealing process, the quantum tunneling effect is used to break out of the local optimal trap and explore all possible sorting combinations in parallel to obtain a sorting result satisfying multiple constraint conditions. For details of the quantum annealing algorithm, please refer to the description of related technologies, which will not be described here.

[0186] The traditional greedy algorithm is easy to fall into local optimization by sorting in a single dimension. The quantum annealing algorithm in this embodiment simulates the quantum tunneling effect and can cross the local extreme point in the energy valley. Compared with the traditional single numerical score, it can more accurately guide the demand iteration priority in different scenarios. For example, although the absolute number of "seat heating" feedback in winter in the north is less than that of "navigation interface optimization", the greedy algorithm may underestimate its priority due to the small amount of feedback, while the quantum annealing algorithm can adjust the results combined with the constraint conditions to find the optimal solution in the global range and avoid the loss of user experience caused by local optimization.

[0187] The embodiment first uses the size order of the quantum state priority to preliminarily sort the feedback opinions to be processed to obtain an initial sorting sequence, and then adjusts the initial sorting sequence to obtain a sorting result that meets the constraint condition, thereby avoiding falling into local optimization by sorting only according to the quantum state priority.

[0188] In step S4053, the target feedback opinions are obtained by screening the multiple feedback opinions to be processed according to the sorting result, and the causal labels of the target feedback opinions are determined; wherein the causal labels include the influence of the target feedback opinions on the user experience.

[0189] Specifically, the feedback opinion at the top of the sorting result is taken as the target feedback opinion, and the target feedback opinion is processed preferentially, and the causal label of the target feedback opinion derived by the large language model is obtained.

[0190] In some embodiments, the number of target feedback opinions can be multiple, that is, in the case of sufficient resources and manpower, product optimization is carried out based on multiple target feedback opinions at the same time.

[0191] In step S4054, the target feedback opinions and the causal labels and scene labels of the target feedback opinions are input into the pre-trained product optimization model to obtain multiple product optimization schemes.

[0192] Specifically, the target feedback opinions, the causal labels and scene labels of the target feedback opinions are taken as inputs, and the pre-trained product optimization model (such as the Stable Diffusion model) is used to analyze the function improvement direction, experience optimization details and associated impact, and multiple product optimization schemes are output, including hardware optimization schemes, software iteration schemes and service optimization schemes.

[0193] It should be noted that the product optimization model can receive a text description and generate a structured improvement scheme text. The product optimization model is pre-trained through intelligent cockpit field professional knowledge and vehicle function module knowledge, so that the pre-trained product optimization model can generate product optimization schemes in different directions (hardware optimization, software iteration, and service optimization) through prompt words. The specific training process can refer to the detailed description of related technologies, and will not be described here.

[0194] The related technology only analyzes and processes user feedback through keywords, cannot find the internal relationship between feedback, and is prone to misjudge the fault source. The embodiment combines the specific content of the target feedback to be processed, the cause and effect label and the scene label, uses the pre-trained product optimization model to mine the cause and effect association and scene association between user feedback, outputs multiple product optimization schemes, and selects a suitable target product optimization scheme, thereby improving the accuracy and scene adaptability of the product optimization scheme.

[0195] In step S4055, multiple product optimization schemes are screened to obtain a target product optimization scheme.

[0196] In some optional embodiments, each product optimization scheme is input into a pre-trained optimization prediction model to obtain the optimization effect of each product optimization scheme. Based on the optimization effect, the multiple product optimization schemes are screened to obtain the target product optimization scheme.

[0197] In some embodiments, the optimization prediction model (for example, an XGBoost prediction model) is trained using historical feedback data to simulate the feedback changes after improvement using the corresponding product optimization scheme, and a pre-trained optimization prediction model is obtained. The product optimization scheme is input into the pre-trained optimization prediction model to obtain the predicted optimization effect, which includes but is not limited to feedback reduction rate, R&D / execution man-hours, and efficiency ratio (the ratio between feedback reduction rate and R&D / execution man-hours). The specific adjustment can be combined with the actual scene.

[0198] In some embodiments, by comparing the optimization effect of each product optimization scheme, the target product optimization scheme with the optimal comprehensive effect is obtained. For example, the product optimization scheme with the largest efficiency ratio is taken as the target product optimization scheme, and in the case of the same efficiency ratio, the product optimization scheme with higher feedback reduction rate or smaller R&D / execution man-hours is taken as the target product optimization scheme.

[0199] In step S4056, the product is optimized according to the target product optimization scheme.

[0200] For example, if the target product optimization scheme is a hardware optimization scheme, the related product function can be optimized from the aspect of hardware improvement to achieve the user's expected effect and reduce the user feedback rate.

[0201] The related technology only feeds back the score to the user and does not give a specific improvement scheme, and a special person needs to redesign and improve the verification process, so that the problem cycle from discovery to closure is lengthened.

[0202] The embodiment predicts the optimization effect of each product optimization scheme by the pre-trained optimization prediction model, selects a target product optimization scheme with the optimal comprehensive effect from among them for product optimization, so as to quickly make the corresponding product function reach the user's expected effect, and reduce the user feedback rate.

[0203] The product optimization scheme of the present application will be described in detail below in conjunction with a specific application embodiment.

[0204] Step 1, obtain the user's original evaluation data: the first text data "the seat heating does not react in winter, and the navigation shouts for half a day and does not come out", the voice data "tried three times to heat, and the navigation wake-up delay is 5 seconds", and the picture data (the seat heating button light is off, and the navigation interface has no wake-up response).

[0205] The voice data is converted into the second text data "the seat heating does not work three times, and the navigation wake-up delay is 5 seconds" by the Whisper model. The key information of the picture data is identified by using the ResNet model, and the "seat heating button failure" and "navigation does not wake up" image feature labels are labeled, and finally the "text description: seat heating does not react, navigation wake-up delay + image feature label: seat heating button failure, navigation does not wake up" is generated. At the same time, the CLIP model is used to establish the semantic mapping of "seat heating does not react-seat heating button failure" and "navigation wake-up delay-navigation does not wake up", to ensure the consistency of multi-modal information.

[0206] Step 2, judge that the above user feedback opinion is a composite feedback, and split it into "seat heating does not react" and "navigation wake-up delay" two feedback opinions. Combine the historical data, and classify "seat heating does not react" and similar "seat heating button light is off" and "heating lamp failure" and the like into the same feedback cluster.

[0207] The cause and effect label of "seat heating does not react" output by the large model thinking chain is "seat heating does not react → winter vehicle warm-keeping is insufficient", and the potential value is "seat heating complaint rate decreases by 50%"; the cause and effect label of slow navigation wake-up is "navigation wake-up delay → navigation response delay", and the potential value is "navigation delay rate decreases by 60%". The large language model determines that the above two feedback opinions are both to-be-processed feedback opinions based on the operability (both have a clear fault point), the urgency (the heating demand in winter is urgent) and the like.

[0208] Step 3, taking the "seat heating non-response" classification process as an example, the entity layer sets up four types of entities such as feedback entities (such as seat heating failure), functional entities (seat heating module), scenario entities (winter driving in the north), and demand entities (warmth and safety), the relationship layer establishes the product function to which "seat heating non-response" belongs, the scenario tag that occurs, and the associated user demand, forming a structured link. Through knowledge graph and large language model classification, such as "seat heating non-response" being classified as hardware failure, vehicle body comfort system, winter northern static scenario, and member warmth and safety demand, providing data support for subsequent processes.

[0209] By adopting MAML technology, when a new "steering wheel heating failure" is added, the knowledge graph can automatically add feedback entities, functional entities, and associated winter scenarios, without the need for retraining.

[0210] Step 4, select "seat heating non-response (suggestion 1)" and "navigation wake-up delay (suggestion 2)" as two feedback opinions to be processed, calculate the scores of the four-dimensional basic dimensions of the two suggestions according to the specified algorithm, and calculate the priority step by step to solve the priority misjudgment problem under scenario heterogeneity.

[0211] Step 4.1, basic value score Calculation.

[0212] Taking suggestion 1 as an example, the product function of seat heating is a paid function (weight 1.2), and the feedback affects user renewal, with a user renewal correlation degree of 1 (additional 0.3 weight), and the feedback user accounts for 10% of the winter users in the north. After calculation, the original score is (1.2+0.3)*10%=0.15, and the value range of the original score is , the normalized final basic value score of suggestion 1 is 0.1. Suggestion 2 is a non-paid function (weight 0.8), and the feedback user accounts for 20%. After 0.8*20% calculation and normalization, the basic value score of suggestion 2 is 0.1.

[0213] Step 4.2, causal intensity score Calculation.

[0214] The knowledge graph path of suggestion 1 is "heating failure → warmth demand (core demand)", the path length is 1 (corresponding to the first basic score 0.9), the associated edge appears frequency is 80% (every 5% increase in score increases 0.05, and the final second basic score is 0.8), and the calculation is 0.9*0.4+0.8*0.6=0.84. The maximum and minimum value range is , and the normalized causal intensity score of suggestion 1 is 0.9; the path length of suggestion 2 (corresponding to the first basic score of 0.7), the occurrence frequency of the associated edge is 60% (each increase of 5% of the score increases 0.05, and the final second basic score is 0.6), and the causal strength score is calculated as 0.7*0.4+0.6*0.6=0.64 after normalization. The score is 0.6.

[0215] Step 4.3, scene emergency score The score is calculated.

[0216] Suggestion 1 corresponds to the winter vehicle scene, the scene duration accounts for 40%, and the scene accident accounts for 10% with low-temperature frostbite. The calculation is 0.4*1.0+0.1*0.6=0.46, and the corresponding maximum and minimum value range is , and the normalized scene emergency score is 0.3; the scene duration of suggestion 2 accounts for 30% and has no accident association, and the calculation is 0.3*1.0+0*0.6=0.3, and the normalized scene emergency score of suggestion 2 is 0.2.

[0217] Step 4.4, demand matching score The score is calculated.

[0218] The current product strategy is "winter safety breakthrough (core demand)", and the third basic score of the safety class user demand is 0.9. The semantic similarity of suggestion 1 to the core demand is 90% (the semantic similarity increases by 10% and the additional score is 0.05, and the final additional score is 0.45), and the calculation is 0.9+0.45=1.35. The corresponding maximum and minimum value range is , and the normalized demand matching score of suggestion 1 is 0.9; suggestion 2 corresponds to the experience class demand (the third basic score is 0.6), and the semantic similarity to the corresponding core demand is 80% (the semantic similarity increases by 10% and the additional score is 0.05, and the final additional score is 0.4), and the calculation is 0.6+0.4=1.0. The normalized demand matching score of suggestion 2 is 0.5.

[0219] Step 4.5, adjust 、 、 and four quantum amplitudes by the Lagrange multiplier method to ensure , and the final adjustment result is as follows:

[0220] Suggestion 1: =0.32 (corresponding to |A =0.1), =0.31 (corresponding to |B = 0.9), = 0.21 (corresponding to |C = 0.3), = 0.84 (corresponding to |D = 0.9), 0.32 2 + 0.31 2 + 0.21 2 + 0.84 2 ≈ 1, satisfying the normalization constraint.

[0221] Proposal 2: = 0.45 (corresponding to |A = 0.1), = 0.43 (corresponding to |B = 0.6), = 0.22 (corresponding to |C = 0.2), = 0.75 (corresponding to |D = 0.5), 0.45 2 + 0.43 2 + 0.22 2 + 0.75 2 ≈ 1, satisfying the normalization constraint.

[0222] Step 4.6, set constraints: the lower limit of the priority of the quantum state of the explicit safety class feedback opinion ≥ 0.8, the processing time limit of the pending feedback opinion ≤ 72 hours, etc.

[0223] The traditional greedy algorithm only sorts by "feedback user proportion", and the feedback proportion of proposal 2 (20%) is higher than that of proposal 1 (10%), which will put proposal 2 first, but it ignores the scene demand of "heating failure affecting user safety" in winter in the north, which is easy to lead to priority misjudgment.

[0224] The quantum annealing algorithm simulates the quantum tunneling effect, crossing the local optimum of the single dimension of "feedback quantity", combining the scene characteristics of "winter safety demand", and calculating the priority comprehensively in four-dimensional dimensions and quantum amplitude. Taking and the above constraints to construct an energy function, and solving the optimal priority processing target pending feedback opinion through algorithm. Finally, "seat heating non-response" is ranked as the highest priority target pending feedback opinion, avoiding misjudgment caused by scene heterogeneity.

[0225] Step 5, taking the causal label "seat heating non-response → winter driving insufficient warmth" and the scene label "winter driving scene" of "seat heating non-response" as input parameters, output 3 types of product optimization schemes through StableDiffusion model:

[0226] Solution 1 (hardware optimization): upgrade the seat heating module temperature control chip, add preheating compensation logic below -10 DEG C environment; Solution 2 (software iteration): develop a fault self-test function, when the seat heating button light is off, a pop-up window prompts "heating module failure, please restart the vehicle"; Solution 3 (service optimization): extend the member user seat heating warranty period to 3 years, covering the winter high-frequency use cycle.

[0227] Further, based on the historical feedback data of "seat heating failure" in the past 2 years (including 10 similar improvement cases), an XGBoost prediction model is trained, and the predicted optimization effect of the three product optimization solutions is obtained:

[0228] Solution 1 (hardware optimization): the predicted feedback reduction rate is 80%, the R&D / execution man-hours required are 50 hours, and the efficiency ratio is 1.6; Solution 2 (software iteration): the predicted feedback reduction rate is 60%, the R&D / execution man-hours required are 30 hours, and the efficiency ratio is 2.0; Solution 3 (service optimization): the predicted feedback reduction rate is 40%, the R&D / execution man-hours required are 20 hours, and the efficiency ratio is 2.0.

[0229] Since the efficiency ratios of solution 2 and solution 3 are the same (both 2.0), but solution 3 needs to link the after-sales department to adjust the warranty policy, and the implementation cycle needs 15 days; Solution 2 only needs software iteration, and the implementation cycle is only 7 days, which can better meet the requirement of "≤48 hours processing time", so finally solution 2, i.e. developing a fault self-test function, is selected as the target product optimization solution, and the seat heating function is optimized.

[0230] The application integrates various feedback forms such as text, sound and picture through multi-modal feedback fusion, for example, when processing user feedback of text evaluation and fuzzy positioning screenshots, the system will standardize different types of feedback information and perform semantic association, and filter the feedback opinions to be processed from the standardized user feedback data.

[0231] The application automatically adjusts feedback classification rules through dynamic knowledge graph classification, and when user feedback is not covered, the system uses a self-evolution mechanism to automatically generate new classification nodes, reduces the update frequency of classification rules, and significantly reduces the cost of manual maintenance.

[0232] The application quantifies the importance of feedback by using quantum priority modeling, calculates the quantum state priority of the current feedback problem according to the existing user feedback problem, solves the scene difference problem, and improves the accuracy of priority evaluation.

[0233] The application establishes the causal relationship between user feedback and product problems through causal cognitive mining, so that the improvement direction is more accurate, automatically outputs a feasible product optimization scheme, and simulates the optimization effect, so that the whole process does not need manual secondary transformation, and the efficiency is improved obviously.

[0234] The entire processing flow of the application is completely based on user feedback data, without calling hardware data such as vehicle-mounted CAN bus, to realize efficient coverage of new feedback classification, reduce priority judgment error, greatly shorten analysis period, and significantly improve product upgrade conversion rate.

[0235] In the embodiment, a product optimization device based on user feedback is also provided, which is used to realize the above-mentioned embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that realizes a predetermined function. Although the device described in the following embodiments is preferably realized in software, hardware, or a combination of software and hardware is also possible and is conceived.

[0236] The embodiment provides a product optimization device based on user feedback, as shown in Figure 5 The device includes:

[0237] The first processing module 501 is configured to associate picture data, voice data or first text data of a user evaluating a product to obtain user feedback data, and obtain a plurality of feedback opinions to be processed according to user feedback data of a plurality of users.

[0238] The second processing module 502 is configured to obtain a multi-dimensional classification result of each feedback opinion to be processed based on a pre-trained knowledge graph architecture.

[0239] The third processing module 503 is configured to determine a score of the feedback opinion to be processed in a plurality of basic dimensions based on the multi-dimensional classification result.

[0240] The fourth processing module 504 is configured to determine a quantum amplitude of each basic dimension, and obtain a quantum state priority of the feedback opinion to be processed according to the score and the quantum amplitude of the feedback opinion to be processed in each basic dimension. The quantum amplitude is used to represent the weight of the corresponding basic dimension.

[0241] The fifth processing module 505 is configured to sort each feedback opinion to be processed based on the quantum state priority, generate a target product optimization scheme according to the sorting result, and optimize the product according to the target product optimization scheme.

[0242] In some optional embodiments, the first processing module 501 is further configured to:

[0243] perform feature extraction on the picture data to obtain image feature labels;

[0244] convert the voice data into second text data, and associate the image feature labels, the second text data or the first text data to obtain the user feedback data.

[0245] In some optional embodiments, the multi-dimensional classification result includes a product function, a scene label and a plurality of user demands associated with each feedback opinion to be processed, and the user demands include preset core demands; the third processing module 503 is further configured to:

[0246] determine a product function associated with the feedback opinion to be processed, and obtain a basic value score of the feedback opinion to be processed according to business data of the product function;

[0247] determine at least one knowledge graph path from the feedback opinion to be processed to a preset core demand based on a plurality of user demands associated with the feedback opinion to be processed, and obtain a causality strength score of the feedback opinion to be processed according to the at least one knowledge graph path;

[0248] determine a scene label associated with the feedback opinion to be processed, and obtain a scene emergency degree score of the feedback opinion to be processed according to scene use time length data and scene accident data corresponding to the scene label;

[0249] obtain a demand matching degree score of the feedback opinion to be processed according to the feedback opinion to be processed and the preset core demand.

[0250] In some optional embodiments, the third processing module 503 is further configured to:

[0251] obtain a target knowledge graph path with the shortest path length according to the at least one knowledge graph path;

[0252] obtain a first basic score according to the path length of the target knowledge graph path;

[0253] calculate an association edge occurrence frequency of the target knowledge graph path, and obtain a second basic score according to the association edge occurrence frequency;

[0254] linearly weight the first basic score and the second basic score to obtain the causality strength score of the feedback opinion to be processed.

[0255] In some optional embodiments, the third processing module 503 is further configured to:

[0256] determine a third basic score corresponding to the preset core demand, and calculate a semantic similarity between the feedback opinion to be processed and the preset core demand;

[0257] correct the third basic score according to the semantic similarity to obtain the demand matching degree score of the feedback opinion to be processed.

[0258] In some optional embodiments, the third processing module 503 is further configured to:

[0259] According to the product function service data, a function type, a user renewal correlation degree, and a feedback user proportion of the product function are obtained; the function type includes a paid function and a non-paid function;

[0260] Based on the function type and the user renewal correlation degree, a basic value weight is determined.

[0261] According to the basic value weight and the feedback user proportion, a basic value score of the feedback opinion to be processed is calculated.

[0262] In some optional embodiments, the fourth processing module 504 is further configured to:

[0263] A quantum amplitude constraint function is constructed, and quantum amplitudes of each basic dimension are obtained according to the quantum amplitude constraint function;

[0264] The scores and the quantum amplitudes of the feedback opinion to be processed in each basic dimension are weighted and summed to obtain a quantum state priority of the feedback opinion to be processed.

[0265] In some optional embodiments, the fifth processing module 505 is further configured to:

[0266] A constraint condition for the feedback opinion to be processed is constructed, and an initial sorting sequence is determined based on a size order of the quantum state priority of each feedback opinion to be processed;

[0267] Based on the quantum state priority and the constraint condition, the initial sorting sequence is adjusted to obtain a sorting result that meets the constraint condition.

[0268] In some optional embodiments, the fifth processing module 505 is further configured to:

[0269] According to the sorting result, the multiple feedback opinions to be processed are screened to obtain a target feedback opinion to be processed, and a causal label of the target feedback opinion to be processed is determined; the causal label includes an influence result of the target feedback opinion to be processed on user experience;

[0270] The target feedback opinion to be processed and the causal label and a scenario label of the target feedback opinion to be processed are input into a pre-trained product optimization model to obtain multiple product optimization schemes;

[0271] The multiple product optimization schemes are screened to obtain a target product optimization scheme.

[0272] In some optional embodiments, the fifth processing module 505 is further configured to:

[0273] Each product optimization scheme is input into a pre-trained optimization prediction model to obtain an optimization effect of each product optimization scheme;

[0274] Based on the optimization effect, a plurality of product optimization schemes are screened to obtain a target product optimization scheme.

[0275] The product optimization device based on user feedback provided by the embodiments of the present application can execute the product optimization method based on user feedback provided by any of the embodiments of the present application, and has the corresponding function modules and beneficial effects of the execution method. The further function description of the above-mentioned various modules and units is the same as that of the corresponding embodiments, and will not be repeated here.

[0276] Figure 6 A structural schematic diagram of an electronic device provided by the embodiments of the present application is provided.

[0277] The following will be specifically referred to Figure 6 which shows a structural schematic diagram of an electronic device suitable for being used to implement the embodiments of the present application. The electronic device can include a processor (such as a central processor, a graphics processor, etc.) 601, which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 602 or programs loaded from a storage 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for operation of the electronic device are also stored. The processor 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0278] Generally, the following devices can be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 can allow the electronic device to communicate with other devices wirelessly or by wire to exchange data. Although Figure 6 The electronic device with various devices is shown, but it should be understood that it is not required to implement or have all the shown devices, and more or less devices can be alternatively implemented or possessed.

[0279] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program in accordance with embodiments of the present application. For example, embodiments of the present application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for carrying out the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via the communication device 609, or installed from the memory 608, or installed from the ROM 602. When the computer program is executed by the processor 601, the above-mentioned functions defined in the user feedback based product optimization method of embodiments of the present application are performed.

[0280] Figure 6 The electronic device shown is merely an example and should not impose any limitation on the functions and use range of embodiments of the present application.

[0281] Embodiments of the present application also provide a computer-readable storage medium, the above-mentioned method according to embodiments of the present application can be implemented in hardware, firmware, or as computer code recordable on a storage medium, or as computer code originally stored in a remote storage medium or non-transitory machine-readable storage medium and to be stored in a local storage medium downloaded through a network, so that the method described herein can be processed by such software stored on a storage medium using a general-purpose computer, a special-purpose processor, or programmable or special-purpose hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk or a solid state disk, etc.; further, the storage medium can also include a combination of the above-mentioned types of memories. It can be understood that the computer, processor, microprocessor controller or programmable hardware includes a storage component that can store or receive software or computer code, which, when accessed and executed by the computer, processor or hardware, implements the user feedback based product optimization method shown in the above-mentioned embodiments.

[0282] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, the method and / or technical solutions according to the present application can be invoked or provided. Those skilled in the art should understand that the form of computer program instructions in computer-readable medium includes but is not limited to source file, executable file, installation package file, etc., and accordingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer executes the corresponding compiled program after compiling the instructions, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to the computer.

[0283] While embodiments of the application have been described in connection with the preferred embodiments of the various figures, those of ordinary skill in the art will appreciate that various modifications and changes can be made without departing from the spirit and scope of the application, and that such modifications and changes fall within the scope of the appended claims.

Claims

1. A product optimization method based on user feedback, characterized in that, The method includes: By correlating the image data, voice data, or primary text data of user reviews of products, user feedback data is obtained, and multiple feedback opinions to be processed are obtained based on the user feedback data from multiple users. Based on a pre-trained knowledge graph architecture, multi-dimensional classification results are obtained for each feedback opinion to be processed. Based on the multidimensional classification results, the scores of the feedback to be processed in multiple basic dimensions are determined; The quantum amplitude of each fundamental dimension is determined, and the quantum state priority of the feedback to be processed is obtained based on the score and quantum amplitude of each fundamental dimension; wherein, the quantum amplitude is used to characterize the weight of the corresponding fundamental dimension. Each feedback opinion to be processed is sorted based on the quantum state priority, a target product optimization plan is generated based on the sorting result, and the product is optimized based on the target product optimization plan. The multidimensional classification results include the product functions, scenario tags, and multiple user needs associated with each piece of feedback to be processed, including preset core needs; based on the multidimensional classification results, the scores of the feedback to be processed in multiple basic dimensions are determined, including: Identify the product function associated with the feedback to be processed, and obtain the basic value score of the feedback to be processed based on the business data of the product function. Based on multiple user needs associated with the feedback to be processed, at least one knowledge graph path is determined from the feedback to be processed to the preset core needs. Based on the at least one knowledge graph path, the causal strength score of the feedback to be processed is obtained. Determine the scene tags associated with the feedback to be processed, and obtain the scene urgency score of the feedback to be processed based on the scene usage time data and scene accident data corresponding to the scene tags. Based on the feedback to be processed and the preset core requirements, a requirement matching score is obtained for the feedback to be processed.

2. The product optimization method based on user feedback according to claim 1, characterized in that, The step of obtaining the causal strength score of the feedback opinion to be processed based on at least one knowledge graph path includes: Based on the at least one knowledge graph path, obtain the target knowledge graph path with the shortest path length; The first basic score is obtained based on the path length of the target knowledge graph path; Calculate the frequency of occurrence of associated edges in the target knowledge graph path, and obtain the second basic score based on the frequency of occurrence of associated edges; The causal strength score of the feedback opinion to be processed is calculated by linearly weighting the first base score and the second base score.

3. The product optimization method based on user feedback according to claim 1, characterized in that, The step of obtaining a demand matching score for the feedback to be processed based on the feedback to be processed and the preset core requirements includes: Determine the third basic score corresponding to the preset core requirement, and calculate the semantic similarity between the feedback opinion to be processed and the preset core requirement; The third basic score is corrected based on the semantic similarity to obtain the demand matching score of the feedback opinion to be processed.

4. The product optimization method based on user feedback according to claim 1, characterized in that, The process of obtaining the basic value score of the feedback to be processed based on the business data of the product functions includes: Based on the business data of the product functions, the function type, user renewal correlation, and feedback user ratio of the product functions are obtained; wherein, the function type includes paid functions and non-paid functions; Based on the function type and the user renewal correlation, determine the basic value weight; The basic value score of the feedback opinion to be processed is calculated based on the basic value weight and the proportion of feedback users.

5. The product optimization method based on user feedback according to any one of claims 1-4, characterized in that, The process of determining the quantum amplitude of each fundamental dimension, and obtaining the quantum state priority of the feedback to be processed based on the score and quantum amplitude of each fundamental dimension, includes: Construct a quantum amplitude constraint function, and obtain the quantum amplitude of each fundamental dimension based on the quantum amplitude constraint function; The quantum state priority of the feedback to be processed is obtained by weighted summing of the scores and quantum amplitudes of the feedback in each basic dimension.

6. The product optimization method based on user feedback according to any one of claims 1-4, characterized in that, The process of sorting each feedback opinion to be processed based on the quantum state priority includes: Construct constraints for the feedback to be processed, and determine the initial sorting sequence based on the order of the quantum state priorities of each feedback to be processed; Based on the quantum state priority and the constraints, the initial sorting sequence is adjusted to obtain a sorting result that satisfies the constraints.

7. The product optimization method based on user feedback according to claim 6, characterized in that, The step of generating an optimization plan for the target product based on the ranking results includes: Based on the sorting results, multiple pending feedback opinions are filtered to obtain target pending feedback opinions, and causal labels for target pending feedback opinions are determined; wherein, the causal labels include the impact of the target pending feedback opinions on user experience; The target feedback to be processed, along with its causal and scenario labels, are input into a pre-trained product optimization model to obtain various product optimization solutions. Multiple product optimization schemes were screened to obtain the target product optimization scheme.

8. The product optimization method based on user feedback according to claim 7, characterized in that, The process of screening multiple product optimization schemes to obtain the target product optimization scheme includes: Each product optimization scheme is input into a pre-trained optimization prediction model to obtain the optimization effect of each product optimization scheme; Based on the optimization results, various product optimization schemes are screened to obtain the target product optimization scheme.

9. The product optimization method based on user feedback according to any one of claims 1-4, characterized in that, The process of associating image data, voice data, or first text data of user reviews of products to obtain user feedback data includes: Feature extraction is performed on the image data to obtain image feature labels; The voice data is converted into second text data, and then associated with the image feature tags, the second text data, or the first text data to obtain user feedback data.

10. A product optimization device based on user feedback, characterized in that, The device includes: The first processing module is used to associate the image data, voice data or first text data of user evaluation products to obtain user feedback data, and to obtain multiple feedback opinions to be processed based on the user feedback data of multiple users. The second processing module is used to obtain multi-dimensional classification results for each feedback opinion to be processed based on the pre-trained knowledge graph architecture. The third processing module is used to determine the scores of the feedback to be processed in multiple basic dimensions based on the multidimensional classification results. The fourth processing module is used to determine the quantum amplitude of each basic dimension, and to obtain the quantum state priority of the feedback to be processed based on the score and quantum amplitude of each basic dimension; wherein, the quantum amplitude is used to characterize the weight of the corresponding basic dimension. The fifth processing module is used to sort each feedback opinion to be processed based on the quantum state priority, generate a target product optimization plan based on the sorting result, and optimize the product based on the target product optimization plan. The multidimensional classification results include the product functions, scenario tags, and multiple user needs associated with each piece of feedback to be processed, including preset core needs; the third processing module is also used for: Identify the product function associated with the feedback to be processed, and obtain the basic value score of the feedback to be processed based on the business data of the product function. Based on multiple user needs associated with the feedback to be processed, at least one knowledge graph path is determined from the feedback to be processed to the preset core needs. Based on the at least one knowledge graph path, the causal strength score of the feedback to be processed is obtained. Determine the scene tags associated with the feedback to be processed, and obtain the scene urgency score of the feedback to be processed based on the scene usage time data and scene accident data corresponding to the scene tags. Based on the feedback to be processed and the preset core requirements, a requirement matching score is obtained for the feedback to be processed.

11. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the product optimization method based on user feedback as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the product optimization method based on user feedback as described in any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes computer instructions for causing a computer to perform the product optimization method based on user feedback as described in any one of claims 1 to 9.

Images