Data analysis method and system for smart stage technology performance services

Abstract

The embodiment of the present application provides a data analysis method and system for smart stage technology performing arts services, which obtains audience user feedback graph data related to the target stage performance and performs feature extraction on these data to obtain the feedback knowledge description vector of each performing arts feedback event. Next, frequent pattern item mining is performed on each feedback knowledge description vector to obtain the corresponding frequent pattern item data as audience behavior mining data. Then, the pre-trained stage preference prediction model is used to perform stage preference prediction on the audience behavior mining data to generate corresponding stage preference prediction data. Finally, based on the stage preference prediction data, the corresponding content recommendation information is pushed to the cloud service center of future stage performances associated with the target stage performance. Through the above steps, the audience's stage preferences can be accurately predicted, and personalized content recommendations can be provided accordingly, thereby improving the audience's viewing experience and satisfaction.

Description

Technical Field

[0001] The present application relates to the field of smart stage application technology, and more specifically, to a data analysis method and system applied to smart stage technology performance services. Background Art

[0002] With the development of information technology and network technology, various stage performances such as musicals, dramas, operas, etc. are broadcast to the general audience through the Internet platform, bringing unprecedented convenience. However, in the face of many rich and diverse stage performance resources, how to provide accurate recommendations based on the audience's personal preferences has become a problem that needs to be solved urgently.

[0003] Related technologies mainly rely on information such as users' historical browsing history or purchasing behavior. However, when dealing with complex stage performance recommendation problems, this method often fails to take into account the audience's specific feedback, such as opinions on actor performances, plots, stage design, etc., resulting in recommendation results that cannot accurately meet the audience's needs. Therefore, there is an urgent need for an improved solution that can more deeply understand and analyze audience feedback, thereby generating more accurate stage performance content recommendation results. Summary of the invention

[0004] In view of this, the purpose of this application is to provide a data analysis method and system for smart stage technology performance services.

[0005] According to a first aspect of the present application, a data analysis method for a smart stage technical performance service is provided, which is applied to a smart stage performance service system, and the method includes:

[0006] Obtain x audience user feedback graph data related to the target stage performance, and perform feature extraction on the x audience user feedback graph data to obtain a feedback knowledge description vector of each performance feedback event; each audience user feedback graph data includes y performance feedback events, and the x audience user feedback graph data are respectively associated with different feedback categories;

[0007] Performing frequent pattern item mining on the feedback knowledge description vector of each performance feedback event to obtain corresponding frequent pattern item data as audience behavior mining data;

[0008] Performing stage preference prediction on the audience behavior mining data based on a pre-trained stage preference prediction model to generate corresponding stage preference prediction data;

[0009] Based on the stage preference prediction data, corresponding content recommendation information is pushed to a cloud service center of future stage performances associated with the target stage performance.

[0010] In a possible implementation of the first aspect, the step of extracting features from the x audience user feedback graph data to obtain a feedback knowledge description vector for each performance feedback event includes:

[0011] In the zth feedback category, the first feedback connection vector of the connection feedback event of the qth performance feedback event is analyzed, and the target performance feedback behavior vector of the qth performance feedback event is analyzed; in the zth feedback category, the basic event matching degree of the connection feedback event of the qth performance feedback event with respect to the qth performance feedback event is determined according to the first feedback connection vector and the target performance feedback behavior vector;

[0012] If the basic event matching degree in the zth feedback category is greater than the set matching degree, the first event dependency type is output as the event dependency type of the connection feedback event of the qth performance feedback event for the qth performance feedback event;

[0013] If the basic event matching degree in the z-th feedback category is not greater than the set matching degree, the second event dependency type is output as the event dependency type of the connection feedback event of the q-th performance feedback event for the q-th performance feedback event; z is not greater than x, and q is not greater than y;

[0014] According to the event dependency type, the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event is determined, wherein the first event dependency type indicates that the event dependency vector is determined based on the connection feedback event of the qth performance feedback event, and the first feedback connection vectors are respectively associated in x feedback categories; the second event dependency type indicates that the event dependency vector is determined based on the connection feedback event of the qth performance feedback event, and the first feedback connection vector corresponding to the zth feedback category;

[0015] In the zth feedback category, according to the event dependency vector of the connected feedback event of the qth performance feedback event for the qth performance feedback event, determine the event description vector to be enriched of the qth performance feedback event;

[0016] Based on the event description vector to be enriched, a target event description vector of the qth performance feedback event in the zth feedback category is obtained, and the target event description vector is used for fusion to obtain a feedback knowledge description vector expressing the qth performance feedback event.

[0017] For example, in a possible implementation of the first aspect, the number of connection feedback events of the qth performance feedback event is K;

[0018] The determining, in the zth feedback category, according to the first feedback connection vector and the target performance feedback behavior vector, of the basic event matching degree of the connection feedback event of the qth performance feedback event to the qth performance feedback event comprises:

[0019] In the z-th feedback category, the first feedback connection vectors associated with each of the K connection feedback events are fused with the target performance feedback behavior vector to generate an initial event fusion vector of the q-th performance feedback event;

[0020] In the z-th feedback category, analyzing the initial event fusion sub-vectors between the K first feedback connection vectors and the target performance feedback behavior vector respectively;

[0021] In the zth feedback category, the basic event matching degrees of the K connection feedback events for the qth performance feedback event are determined according to the weights between the K initial event fusion sub-vectors and the initial event fusion vector.

[0022] For example, in a possible implementation of the first aspect, determining, based on the event dependency type, an event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event includes:

[0023] If the event dependency type is the first event dependency type, the first feedback connection vectors associated with the connection feedback event of the qth performance feedback event in the x feedback categories are parsed, the x first feedback connection vectors are fused to generate a feedback connection fusion vector, the feedback connection fusion vector is subjected to feedback category conversion mapping, and an event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event is generated;

[0024] If the event dependency type is the second event dependency type, then the first feedback connection vector corresponding to the connection feedback event of the qth performance feedback event in the zth feedback category is parsed, and the first feedback connection vector in the zth feedback category is output as the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event.

[0025] For example, in a possible implementation of the first aspect, in the zth feedback category, determining the event description vector to be enriched of the qth performance feedback event according to the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event includes:

[0026] In the zth feedback category, according to the event dependency vector of the connection feedback event of the qth performance feedback event with respect to the qth performance feedback event, and the target performance feedback behavior vector of the qth performance feedback event, determine the 0th event matching degree of the connection feedback event of the qth performance feedback event with respect to the qth performance feedback event;

[0027] According to the 0th order event matching degree, the connection feedback event of the qth performance feedback event is weighted with respect to the event dependency vector of the qth performance feedback event to generate an event description vector to be enriched of the qth performance feedback event.

[0028] For example, in a possible implementation of the first aspect, obtaining the target event description vector of the qth performance feedback event in the zth feedback category based on the event description vector to be enriched includes:

[0029] In the zth feedback category, the rth order connection feedback behavior vector corresponding to the connection feedback event of the qth performance feedback event is analyzed, and the rth order event description vector corresponding to the qth performance feedback event is analyzed; the event description vector to be enriched of the qth performance feedback event is used to represent the first order event description vector corresponding to the qth performance feedback event when r is 1;

[0030] In the zth feedback category, the connection feedback event of the qth performance feedback event is parsed for the rth-order event matching degree of the qth performance feedback event, and the rth-order connection feedback behavior vector is weighted according to the rth-order event matching degree to generate an rth-order rich description vector;

[0031] In the zth feedback category, feature enrichment is performed on the rth order event description vector corresponding to the qth performance feedback event according to the rth order enriched description vector to generate the r+1th order event description vector of the qth performance feedback event;

[0032] When r+1 is the threshold feature enrichment round number, the r+1th order event description vector is output as the target event description vector of the qth performance feedback event in the zth feedback category.

[0033] For example, in a possible implementation of the first aspect, the method further includes:

[0034] Analyze the target event description vectors associated with the qth performance feedback event in the x feedback categories, fuse the x target event description vectors corresponding to the qth performance feedback event, and generate a feedback knowledge description vector of the qth performance feedback event.

[0035] For example, in a possible implementation of the first aspect, the method includes:

[0036] Parsing x training user feedback graph data; each training user feedback graph data includes y template performance feedback events; the x training user feedback graph data are respectively associated with different feedback categories;

[0037] Input the x training user feedback graph data into the initial graph neural network, and in the zth feedback category of the initial graph neural network, parse the template connection feedback event of the qth template performance feedback event for the template event dependency type of the qth template performance feedback event; z is not greater than x; q is not greater than y;

[0038] According to the template event dependency type, determine the template event dependency vector of the template connection feedback event of the qth template performance feedback event for the qth template performance feedback event; the template event dependency type includes a first template event dependency type and a second template event dependency type; the first template event dependency type represents that the template event dependency vector is determined based on the template connection feedback event of the qth template performance feedback event, and the first template feedback connection vectors respectively associated in x feedback categories; the second template event dependency type represents that the template event dependency vector is determined based on the template connection feedback event of the qth template performance feedback event, and the first template feedback connection vector corresponding to the qth feedback category; in the zth feedback category, determine the template event description vector to be enriched of the qth template performance feedback event based on the template event dependency vector of the qth template performance feedback event of the qth template performance feedback event;

[0039] Based on the template event description vector to be enriched, a predicted target event description vector of the qth template performance feedback event in the zth feedback category is obtained; the predicted target event description vector is used to fuse to obtain a template feedback knowledge description vector expressing the qth template performance feedback event;

[0040] When generating the template feedback knowledge description vectors associated with each of the y template performance feedback events, determining the matching degree of a template event combination in the y template performance feedback events according to the template feedback knowledge description vectors associated with each of the y template performance feedback events; the template event combination includes two different template performance feedback events in the y template performance feedback events;

[0041] According to the matching degree of the template event combination, the initial graph neural network is updated to generate a graph neural network;

[0042] Wherein, the y template performance feedback events correspond to e template event combinations; e is a positive integer;

[0043] The updating of the initial graph neural network according to the matching degree of the template event combination to generate a graph neural network includes:

[0044] Among the e template event combinations, the template event combinations that have event connection relationships in the x training user feedback graph data are output as positive template event combinations, and the template event combinations that do not have event connection relationships in the x training user feedback graph data are output as negative template event combinations;

[0045] Obtaining a training cost based on a matching degree of the positive template event combination and a matching degree of the negative template event combination;

[0046] The initial graph neural network is updated according to the training cost to generate a graph neural network.

[0047] According to the second aspect of the present application, a smart stage performance service system is provided, which includes a processor and a readable storage medium, wherein the readable storage medium stores a program, which, when executed by the processor, implements the aforementioned data analysis method applied to smart stage technology performance services.

[0048] According to the third aspect of the present application, a computer-readable storage medium is provided, in which computer-executable instructions are stored. When the computer-executable instructions are monitored to be executed, the aforementioned data analysis method applied to smart stage technology performing services is implemented.

[0049] According to any of the above aspects, in the present application, the audience user feedback graph data related to the target stage performance is obtained, and the features of these data are extracted to obtain the feedback knowledge description vector of each performance feedback event. Then, the frequent pattern item mining is performed on each feedback knowledge description vector to obtain the corresponding frequent pattern item data as the audience behavior mining data. Then, the stage preference prediction is performed on the audience behavior mining data using the pre-trained stage preference prediction model to generate the corresponding stage preference prediction data. Finally, based on the stage preference prediction data, the corresponding content recommendation information is pushed to the cloud service center of the future stage performances associated with the target stage performance. Through the above steps, the audience's stage preferences can be accurately predicted, and personalized content recommendations can be provided accordingly, thereby improving the audience's viewing experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 A flow chart of a data analysis method for smart stage technology performance services provided in an embodiment of the present application;

[0052] Figure 2 A schematic diagram of the component structure of a smart stage performing arts service system provided in an embodiment of the present application for implementing the above-mentioned data analysis method applied to smart stage technical performing arts services is shown. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present application.

[0054] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0055] Figure 1 The flowchart of the data analysis method applied to the smart stage technology performance service provided by the embodiment of the present application is shown. It should be understood that in other embodiments, the order of some steps in the data analysis method applied to the smart stage technology performance service of this embodiment can be interchanged according to actual needs, or some steps can be omitted or deleted. The detailed steps of the data analysis method applied to the smart stage technology performance service are introduced as follows.

[0056] Step S110, obtaining x audience user feedback graph data related to the target stage performance, and performing feature extraction on the x audience user feedback graph data to obtain a feedback knowledge description vector of each performance feedback event.

[0057] In this embodiment, each audience user feedback graph data includes y performance feedback events, and the x audience user feedback graph data are respectively associated with different feedback categories.

[0058] Each performance feedback event may include comments on songs, dances, costumes, props, lighting, etc. Among these x performance feedback events, each performance feedback event has one or more mainly associated feedback categories. For example, one performance feedback event may focus mainly on music, and another performance feedback event may focus mainly on dance.

[0059] For example, suppose 300 pieces of audience feedback on a musical "Cats" are collected, that is, x=300. Each performance feedback event contains an evaluation of five aspects of the performance, including songs, dances, costumes, props, and lighting, that is, y=5. So, although each performance feedback event contains evaluations of these five aspects, each performance feedback event may focus on different aspects. For example, the first performance feedback event may focus more on songs, so its main associated category is "songs." The second performance feedback event may focus more on dances, so its main associated category is "dance." In this way, the audience's feedback can be more accurately understood and analyzed based on the main associated categories of each performance feedback event.

[0060] Then, a machine learning model can be used to extract various event description vectors, such as the sentiment polarity of the evaluation, the strength of the evaluation, the topics mentioned, etc., which can be combined into a vector as the feedback knowledge description vector of each performing arts feedback event.

[0061] Step S120, performing frequent pattern item mining on the feedback knowledge description vector of each performance feedback event to obtain corresponding frequent pattern item data as audience behavior mining data.

[0062] For example, data mining algorithms such as Apriori or FP-Growth can be used to mine frequent pattern items in these feedback knowledge description vectors. For example, it may be found that among the audiences who give positive feedback, they may often first praise the melody of the song, then the innovation of the lyrics, and finally the performance skills of the performer. Such a sequence constitutes a frequent pattern item.

[0063] Step S130 , performing stage preference prediction on the audience behavior mining data based on a pre-trained stage preference prediction model to generate corresponding stage preference prediction data.

[0064] Frequent pattern mining is a technique in data mining, which is mainly used to discover patterns or items that appear frequently in a large number of data sets. Frequent pattern mining of the feedback knowledge description vector of each performance feedback event can be understood as: looking for evaluation patterns or characteristics that frequently appear in the audience's feedback.

[0065] For example, these frequent pattern items can be used as inputs to a pre-trained stage preference prediction model, which may be a deep learning model, such as a recurrent neural network (RNN) or a long short-term memory network (LSTM), which can process this type of sequential input data and predict future stage preferences based on historical behavior patterns. For example, it may be predicted that audiences who first focus on the melody, then the innovation of the lyrics and the skills of the performers when evaluating may be more inclined to appreciate musicals with innovative elements and high-quality performances. In a further example, suppose there is a feedback knowledge description vector containing audience evaluations of various aspects of a musical performance (such as songs, dances, costumes, etc.). In this feedback knowledge description vector, there may be many elements, such as "beautiful melody", "deeply touching lyrics", "wonderful dance", etc. Then use a frequent pattern item mining algorithm to analyze these feedback knowledge description vectors, trying to find out which evaluations or features often appear together.

[0066] For example, it may be found that among the audiences who gave positive comments, they often mentioned "beautiful melody" and "the lyrics are deeply touching" at the same time. This forms a frequent pattern item, that is, the common behavior pattern of the audience when making comments. In this way, this frequent pattern item can be used as audience behavior mining data for further analysis and prediction.

[0067] In general, through frequent pattern mining, we can discover the common patterns and trends of audiences when evaluating performances, which is very helpful for understanding audience preferences, improving performance effects, and even predicting audience feedback on future performances.

[0068] Step S140: Push corresponding content recommendation information to a cloud service center of future stage performances associated with the target stage performance based on the stage preference prediction data.

[0069] For example, these stage preference prediction data can be sent to a cloud service center associated with the target stage performance. The cloud service center manages various future stage performance information associated with the "Cats" musical. Based on the stage preference prediction data, the cloud service center can generate personalized content recommendation information and recommend future performances that may be of interest to the audience through emails, APP push, etc., such as recommending a new musical that excels in innovative melody and lyrics and has outstanding performances.

[0070] Based on the above steps, by obtaining the audience user feedback graph data related to the target stage performance and performing feature extraction on these data, the feedback knowledge description vector of each performance feedback event is obtained. Next, frequent pattern item mining is performed on each feedback knowledge description vector to obtain the corresponding frequent pattern item data as audience behavior mining data. Then, the pre-trained stage preference prediction model is used to predict the stage preference of the audience behavior mining data to generate the corresponding stage preference prediction data. Finally, based on the stage preference prediction data, the corresponding content recommendation information is pushed to the cloud service center of future stage performances associated with the target stage performance. Through the above steps, the audience's stage preferences can be accurately predicted, and personalized content recommendations can be provided accordingly, thereby improving the audience's viewing experience and satisfaction.

[0071] In a possible implementation, step S110 may include:

[0072] Step S111, in the zth feedback category, parse the first feedback connection vector of the connection feedback event of the qth performance feedback event, and parse the target performance feedback behavior vector of the qth performance feedback event.

[0073] Step S112, in the zth feedback category, determining the basic event matching degree of the connection feedback event of the qth performance feedback event to the qth performance feedback event according to the first feedback connection vector and the target performance feedback behavior vector.

[0074] For example, suppose that feedback on songs is being analyzed (i.e., the first feedback category, so z=1), and the "theme song performance" (i.e., the first performance feedback event, so q=1) is selected for in-depth analysis. From all performance feedback events that mention "theme song performance", the connection feedback events are obtained, such as the protagonist's singing skills, sound effects, etc., and the first feedback connection vectors of these connection feedback events are parsed, which may include characteristics such as the emotional polarity and intensity of the feedback. At the same time, the target performance feedback behavior vector of the "theme song performance" itself is also parsed, which may include the audience's overall evaluation and reaction to the theme song performance.

[0075] For example, the target performance feedback behavior vector of "theme song performance" and the first feedback connection vector of its linked feedback event (such as the main character's singing skills, sound effects, etc.) can be compared to determine the basic event matching degree between them. For example, if the audience often mentions "the main character's singing skills" when highly praising "theme song performance", then the matching degree between the two will be very high.

[0076] Step S113: if the basic event matching degree in the zth feedback category is greater than the set matching degree, the first event dependency type is output as the event dependency type of the connection feedback event of the qth performance feedback event for the qth performance feedback event.

[0077] For example, assuming that the set matching threshold is 0.7, if the matching degree between "theme song performance" and "protagonist's singing skills" reaches 0.8, then it is considered that there is a strong dependency relationship between the two (i.e., the first event dependency type). On the contrary, if the matching degree between "theme song performance" and "sound effects" is only 0.6, then it is considered that there is a weak dependency relationship between the two (i.e., the second event dependency type).

[0078] Step S114: if the basic event matching degree in the zth feedback category is not greater than the set matching degree, the second event dependency type is output as the event dependency type of the connection feedback event of the qth performance feedback event for the qth performance feedback event. z is not greater than x, and q is not greater than y.

[0079] Step S115, according to the event dependency type, determine the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event, the first event dependency type represents that the event dependency vector is determined based on the connection feedback event of the qth performance feedback event, and the first feedback connection vectors associated with each other in the x feedback categories. The second event dependency type represents that the event dependency vector is determined based on the connection feedback event of the qth performance feedback event, and the first feedback connection vector corresponding to the zth feedback category.

[0080] For example, based on the event dependency type obtained in the previous step, we can determine the degree of influence of each joint feedback event (such as "the main character's singing skills" or "sound effects") on "theme song performance", thereby generating an event dependency vector. This event dependency vector can help understand the importance of each joint feedback event to the target performance feedback event.

[0081] Step S116, in the zth feedback category, according to the event dependency vector of the connected feedback event of the qth performance feedback event with respect to the qth performance feedback event, determine the event description vector to be enriched of the qth performance feedback event.

[0082] The event dependency vector is used to enhance the original feedback knowledge description vector and generate a new event description vector to be enriched. For example, if the "protagonist's singing skills" have a very high impact on the "theme song performance", then in the event description vector to be enriched, the elements related to the "protagonist's singing skills" will be weighted and enhanced.

[0083] Step S117, based on the event description vector to be enriched, obtain the target event description vector of the qth performance feedback event in the zth feedback category, and the target event description vector is used for fusion to obtain the feedback knowledge description vector expressing the qth performance feedback event.

[0084] For example, the event description vector to be enriched is further processed to obtain the final target event description vector. This target event description vector fully reflects the characteristics and influencing factors of the performance feedback event "theme song performance" in the song feedback category, and can be used as the feedback knowledge description vector of "theme song performance" for subsequent steps (such as frequent pattern item mining, stage preference prediction, etc.).

[0085] Here is an overall example:

[0086] 1. In the zth feedback category, parse the first feedback connection vector of the connection feedback event of the qth performance feedback event, and parse the target performance feedback behavior vector of the qth performance feedback event.

[0087] Suppose we are analyzing audience feedback on songs from the musical Cats (i.e., the first feedback category, so z=1), especially the theme song "Memory" (i.e., the first performance feedback event, so q=1). We collect all feedback related to "Memory", such as the main character's singing skills, whether the melody is moving, sound effects, etc., and parse out various features to form the first feedback connection vector, such as emotional polarity, evaluation intensity, etc. At the same time, we also parse out the target performance feedback behavior vector of "Memory" itself, such as overall evaluation, reaction, etc.

[0088] 2. In the zth feedback category, determine the basic event matching degree of the connection feedback event of the qth performance feedback event to the qth performance feedback event based on the first feedback connection vector and the target performance feedback behavior vector.

[0089] In this step, the correlation between "Memory" and other connection feedback events (such as the protagonist's singing skills) is compared. Assuming that the matching threshold is set to 0.7, if the correlation between "Memory" and "the protagonist's singing skills" reaches 0.8, then they are considered to have a high degree of matching. This means that when the audience evaluates "Memory", they usually consider the protagonist's singing skills.

[0090] 3. If the basic event matching degree in the zth feedback category is greater than the set matching degree, the first event dependency type is output as the event dependency type of the connection feedback event of the qth performance feedback event for the qth performance feedback event.

[0091] Since the matching degree between “Memory” and “the singing skills of the protagonist” exceeds the threshold, this relationship is defined as the first event dependency type, that is, the audience’s evaluation of “Memory” is highly dependent on their evaluation of the protagonist’s singing skills.

[0092] 4. According to the event dependency type, determine the event dependency vector of the connection feedback event of the qth performance feedback event with respect to the qth performance feedback event.

[0093] In this step, an event dependency vector is generated to represent the impact of each link feedback event (such as the protagonist's singing skills) on "Memory". If the protagonist's singing skills are highly correlated with "Memory", then in this vector, the protagonist's singing skills will have a larger weight.

[0094] 5. In the zth feedback category, determine the event description vector to be enriched of the qth performance feedback event based on the event dependency vector of the connected feedback event of the qth performance feedback event for the qth performance feedback event.

[0095] Using the above event dependency vector, a to-be-enriched event description vector can be generated, which can more comprehensively reflect the audience's evaluation of "Memory". For example, if the protagonist's singing skills have a greater weight in the event dependency vector, then in the to-be-enriched event description vector, "the protagonist's singing skills" will be given greater importance.

[0096] 6. Based on the event description vector to be enriched, obtain the target event description vector of the qth performance feedback event in the zth feedback category, and the target event description vector is used for fusion to obtain the feedback knowledge description vector expressing the qth performance feedback event.

[0097] Finally, the target event description vector is further optimized based on the event description vector to be enriched to generate the final target event description vector. This target event description vector can fully and accurately reflect the audience's evaluation of "Memory". For example, it may include the audience's evaluation of the song melody, lyrics, the main character's singing skills and other aspects as well as the degree of correlation between them.

[0098] On this basis, the target event description vectors associated with the qth performance feedback event in the x feedback categories can be parsed, and the x target event description vectors corresponding to the qth performance feedback event can be fused to generate a feedback knowledge description vector of the qth performance feedback event. This feedback knowledge description vector can be used as the feedback knowledge description vector of "Memory" for subsequent steps (such as frequent pattern item mining, stage preference prediction, etc.).

[0099] For example, suppose we are processing feedback on the song "Memory" from the musical "Cats", and there are two feedback categories: 1) singing skills 2) sound effects. Now, we have obtained their respective associated target event description vectors, for example:

[0100] Target event description vector of singing skills: [0.8, 0.7]

[0101] Target event description vector of sound effect: [0.6, 0.5]

[0102] Next, the two target event description vectors are fused to generate the feedback knowledge description vector. A simple fusion method is to directly concatenate:

[0103] Feedback knowledge description vector: [0.8, 0.7, 0.6, 0.5]

[0104] In this way, the feedback knowledge description vector of Memory more comprehensively includes the audience's feedback on the song, while also taking into account the information of different feedback categories (singing skills and sound effects).

[0105] In actual operation, fusion may involve more complex calculations, such as using deep learning models for feature extraction and fusion, but the basic idea is the same: to generate a comprehensive feedback knowledge description vector based on relevant information.

[0106] Based on the above steps S111 to S117, first, in a given feedback category, the feedback connection vector of the linked feedback event of the performance feedback event and the target performance feedback behavior vector are parsed. Then, the system determines the matching degree of the linked feedback event of the performance feedback event for the basic event based on these vectors. If the matching degree is greater than the set threshold, the system uses the first event dependency type as the dependency type of the event; otherwise, the second event dependency type is selected. Next, the system determines the event dependency vector according to the event dependency type, and uses this vector to determine the description vector of the event to be enriched. Finally, based on the description vector of the event to be enriched, the system obtains the target event description vector, which can be used to fuse and obtain the feedback knowledge description vector that expresses the feedback event. Through the above steps, valuable information can be effectively extracted from a large amount of feedback data, thereby improving the understanding of user feedback and optimizing the decision-making process, that is, it can more accurately understand and represent the audience's stage feedback, thereby providing strong support for the improvement of stage performances and personalized recommendations.

[0107] For example, in a possible implementation, the number of linked feedback events of the qth performance feedback event is K. For example, in the song feedback category of the musical Cats, "Memory" is selected as the target performance feedback event. Assume that three closely related linked feedback events are found: the main character's singing skills, sound effects, and song melody. Therefore, K=3.

[0108] Step S117 may include:

[0109] Step S1171, in the zth feedback category, the first feedback connection vectors associated with each of the K connection feedback events are fused with the target performance feedback behavior vector to generate an initial event fusion vector of the qth performance feedback event.

[0110] For example, the first feedback connection vectors related to the three linked feedback events (the protagonist's singing skills, sound effects, and song melody) are taken out, which may include information such as the evaluation intensity and emotional polarity of each event. Then, these first feedback connection vectors are fused with the target performance feedback behavior vector of "Memory" mentioned above to generate an initial event fusion vector.

[0111] For example, the following is a specific scenario example:

[0112] Assume that we are processing the feedback of the song "Memory" in the musical "Cats", and there are three connected feedback events: 1) the singing skills of the protagonist 2) the sound effects 3) the melody of the song. Now, their respective first feedback connection vectors have been obtained, for example:

[0113] The first feedback contact vector of the protagonist's singing skills: [0.8, 0.7]

[0114] First feedback contact vector of sound effect: [0.6, 0.5]

[0115] First feedback contact vector of song melody: [0.9, 0.8]

[0116] At the same time, there is also a target performance feedback behavior vector, for example it is [0.7, 0.6].

[0117] A simple way to fuse is to concatenate all the vectors, for example:

[0118] Initial event fusion vector: [0.8, 0.7]+[0.6, 0.5]+[0.9, 0.8]+[0.7, 0.6].

[0119] In this way, the initial event fusion vector of “Memory” contains the audience’s feedback on the song more comprehensively, while also taking into account the information of related linked feedback events (the protagonist’s singing skills, sound effects, and song melody).

[0120] Of course, in practice, fusion may involve more complex calculations, such as using deep learning models for feature extraction and fusion, but the basic idea is the same: generating a comprehensive vector containing multiple feedbacks based on relevant information.

[0121] Step S1172, in the z-th feedback category, parse the initial event fusion sub-vectors between the K first feedback connection vectors and the target performance feedback behavior vector respectively.

[0122] For example, the relationship between the first feedback connection vector of each linked feedback event (the protagonist's singing skills, sound effects, and song melody) and the target performance feedback behavior vector of "Memory" is further analyzed to form an initial event fusion sub-vector. For example, an initial event fusion sub-vector may be obtained to represent the correlation between the protagonist's singing skills and "Memory", another initial event fusion sub-vector may represent the correlation between the sound effects and "Memory", and another sub-vector may represent the correlation between the song melody and "Memory".

[0123] Step S1173, in the zth feedback category, according to the weights between the K initial event fusion sub-vectors and the initial event fusion vector, determine the basic event matching degrees of the K connection feedback events for the qth performance feedback event respectively.

[0124] For example, the matching degree of each connection feedback event to the basic event "Memory" is determined according to the weight between the initial event fusion sub-vector and the initial event fusion vector. For example, if the protagonist's singing skills have a larger weight in the initial event fusion vector, then its matching degree to the basic event "Memory" will be correspondingly higher.

[0125] For example, the following is a specific scenario example:

[0126] Assume that we are processing the feedback of the song "Memory" in the musical "Cats", and there are three linked feedback events: 1) the singing skills of the protagonist 2) the sound effects 3) the melody of the song. In the previous steps, their respective initial event fusion sub-vectors (assuming they are [0.8, 0.7], [0.6, 0.5] and [0.9, 0.8] respectively) and the overall initial event fusion vector (assuming it is [3.0, 2.6]) have been obtained.

[0127] Next, we need to determine the weights between each initial event fusion sub-vector and the initial event fusion vector. This weight can be calculated in a variety of ways, for example, the cosine similarity or Euclidean distance between them can be calculated. Assuming that cosine similarity is used, the calculation results are as follows:

[0128] The weight of the protagonist’s singing skills: cos([0.8, 0.7], [3.0, 2.6]) = 0.95

[0129] Weight of sound effect: cos([0.6, 0.5], [3.0, 2.6]) = 0.88

[0130] The weight of the song melody: cos([0.9, 0.8], [3.0, 2.6]) = 0.96

[0131] Finally, these weights are used as the basic event matching:

[0132] Basic event matching degree of the protagonist's singing skills: 0.95

[0133] Basic event matching of sound effects: 0.88

[0134] Basic event matching of song melody: 0.96

[0135] In this way, the basic event matching degree of each connection feedback event to "Memory" is obtained, and this matching degree represents the correlation between each connection feedback event and "Memory".

[0136] For example, in a possible implementation, step S115 may include:

[0137] Step S1151, if the event dependency type is the first event dependency type, then parse the first feedback connection vectors associated with the connection feedback event of the qth performance feedback event in the x feedback categories, fuse the x first feedback connection vectors to generate a feedback connection fusion vector, perform feedback category conversion mapping on the feedback connection fusion vector, and generate an event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event.

[0138] In the previous step, it has been determined that "Memory" and "The main character's singing skills" have a high degree of match (i.e., the first event dependency type). Now, the first feedback connection vector related to "The main character's singing skills" will be parsed (for example, it may contain features such as emotional polarity and evaluation intensity), and then fused with the corresponding vectors in other feedback categories to generate a feedback connection fusion vector. Finally, this fusion vector is converted and mapped to the feedback category to generate an event dependency vector to represent the degree of influence of "The main character's singing skills" on "Memory".

[0139] The following is a specific scenario example:

[0140] Assume that we are processing feedback for the song "Memory" in the musical "Cats", and there are two linked feedback events: 1) the singing skills of the protagonist 2) the sound effects. Now, we have obtained the first feedback connection vectors that are associated with each other in the two feedback categories (e.g., singing skills and sound effects), for example:

[0141] The first feedback contact vector of the protagonist's singing skills: [0.8, 0.7]

[0142] First feedback contact vector of sound effect: [0.6, 0.5]

[0143] Next, these two vectors are fused to generate a feedback connection fusion vector. A simple fusion method is to add them directly:

[0144] Feedback connection fusion vector: [0.8, 0.7] + [0.6, 0.5] = [1.4, 1.2]

[0145] Then, feedback category conversion mapping is required. This mapping process may involve some transformation or projection operations, such as through neural networks or other machine learning models. The specific operation depends on the actual application scenario and requirements. For simplicity, it can be assumed that the feedback category conversion mapping is to multiply each element of the vector by 2:

[0146] Event dependency vector: 2*[1.4, 1.2] = [2.8, 2.4]

[0147] In this way, the event dependency vector of "Memory" is obtained. This vector contains the information of all the linked feedback events related to "Memory" (the protagonist's singing skills and sound effects), and is uniformly processed through feedback category conversion mapping, thus reflecting the audience's evaluation of "Memory" more comprehensively.

[0148] Note that in actual operation, the fusion and mapping process may involve more complex calculations and deep learning models, but the basic idea is the same: generating a comprehensive event dependency vector based on relevant information.

[0149] Step S1152, if the event dependency type is the second event dependency type, then parse the first feedback connection vector corresponding to the connection feedback event of the qth performance feedback event in the zth feedback category, and output the first feedback connection vector in the zth feedback category as the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event.

[0150] For example, if the matching degree between "Memory" and "Sound Effects" does not reach the set threshold (i.e., it belongs to the second event dependency type), the first feedback connection vector related to "Sound Effects" will be analyzed separately and directly used as its event dependency vector for "Memory", indicating the degree of influence of "Sound Effects" on "Memory".

[0151] For example, in a possible implementation, step S116 may include:

[0152] Step S1161, in the zth feedback category, based on the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event, and the target performance feedback behavior vector of the qth performance feedback event, determine the 0th event matching degree of the connection feedback event of the qth performance feedback event for the qth performance feedback event.

[0153] Still taking the above example, this step is to calculate the 0th order event matching between "Memory" and each linked feedback event (such as the protagonist's singing skills, sound effects, and song melody). This metric is based on the previously obtained event dependency vector and the target performance feedback behavior vector of "Memory". For example, if the protagonist's singing skills have a large weight in the event dependency vector, and the audience's feedback also strongly indicates that they highly appreciate the protagonist's singing, then the protagonist's singing skills will have a very high matching degree with the 0th order event of "Memory".

[0154] The 0-order event matching here may refer to the similarity or correlation between each connection feedback event and the target performance feedback event in the initial state without any processing or weighting. The following is an example of a specific scenario:

[0155] Suppose we are processing feedback for the song "Memory" from the musical "Cats", and there are two connected feedback events: 1) the singing skills of the protagonist 2) the sound effects. Now, we have obtained the event dependency vector of "Memory" (e.g. [2.8, 2.4]) and the target performance feedback behavior vector (e.g. [0.7, 0.6]).

[0156] Next, we need to determine the matching degree of each connection feedback event for the 0th-order event of "Memory". This matching degree can be calculated in many ways, such as calculating their cosine similarity or Euclidean distance. Assuming that cosine similarity is used, and the initial feedback connection vectors of each connection feedback event are [0.8, 0.7] (the protagonist's singing skills) and [0.6, 0.5] (sound effects), the calculation results are as follows:

[0157] The 0th-order event matching degree of the protagonist’s singing skills: cos([0.8, 0.7], [2.8, 2.4]) = 0.98

[0158] The 0th-order event matching degree of the sound effect: cos([0.6, 0.5], [2.8, 2.4]) = 0.95

[0159] In this way, the 0th-order event matching degree of each connection feedback event to "Memory" is obtained. This matching degree represents the correlation between each connection feedback event and "Memory" in the initial state.

[0160] Step S1162: weighting the connection feedback event of the qth performance feedback event with respect to the event dependency vector of the qth performance feedback event according to the 0th order event matching degree, and generating the event description vector to be enriched of the qth performance feedback event.

[0161] For example, based on the 0th-order event matching degree obtained in the previous step, each event dependency vector is weighted to generate an event description vector to be enriched. For example, if the protagonist's singing skills have a very high matching degree with the 0th-order event of "Memory", then in the event description vector to be enriched, the relevant elements of "the protagonist's singing skills" will be weighted and enhanced. This event description vector to be enriched more comprehensively reflects the audience's evaluation of "Memory" and the various factors that affect their evaluation.

[0162] The following is a specific scenario example:

[0163] Assume that we are processing feedback for the song "Memory" from the musical "Cats", and there are two linked feedback events: 1) the main character's singing skills 2) the sound effects. In the previous steps, we have obtained their respective 0-order event matching degrees (e.g., the main character's singing skills are 0.98, and the sound effects are 0.95), as well as the event dependency vector of "Memory" (e.g., [2.8, 2.4]).

[0164] Next, the event dependency vector needs to be weighted. There are many ways to weight the event. A simple way is to use the event matching degree as the weight and directly multiply it by the event dependency vector:

[0165] Weighted event dependency vector: 0.98*[2.8, 2.4]+0.95*[2.8, 2.4]=[5.446, 4.572]

[0166] In this way, the event description vector to be enriched for "Memory" is obtained. This vector is a weighted event dependency vector, which more comprehensively reflects the audience's evaluation of "Memory".

[0167] In practice, the weighting process may involve more complex calculations, such as using deep learning models for feature extraction and fusion, but the basic idea is the same: generating a comprehensive event description vector based on the weights of relevant information.

[0168] For example, in a possible implementation, step S117 may include:

[0169] Step S1171, in the zth feedback category, parse the rth order connection feedback behavior vector corresponding to the connection feedback event of the qth performance feedback event, and parse the rth order event description vector corresponding to the qth performance feedback event. The event description vector to be enriched of the qth performance feedback event is used to represent the first order event description vector corresponding to the qth performance feedback event when r is 1.

[0170] For example, suppose we are processing the feedback of "Memory", which is the qth performance feedback event. When r=1, the event description vector to be enriched of "Memory" is obtained from the previous step. Now, we will parse the first-order connection feedback behavior vectors of each connection feedback event related to "Memory" (such as the protagonist's singing skills, sound effects, and song melody) and associate them with the event description vector to be enriched.

[0171] Step S1172, in the zth feedback category, parse the connection feedback event of the qth performance feedback event for the rth-order event matching degree of the qth performance feedback event, weight the rth-order connection feedback behavior vector according to the rth-order event matching degree, and generate an rth-order rich description vector.

[0172] For example, the first-order event matching of each connection feedback event to "Memory" is calculated, and then the connection feedback behavior vector is weighted based on these event matching degrees to generate a first-order rich description vector. Exemplarily, in this step, the event matching degree is used as a weight, multiplied with the connection feedback behavior vector or other forms of weighted combination to generate a new, more expressive description vector. The following is a simplified example:

[0173] Assume that we are processing the feedback of the song "Memory" in the musical "Cats" (i.e., the qth performance feedback event), and there are two connected feedback events: 1) the singing skills of the protagonist 2) the sound effects. Now, their event matching degrees and connected feedback behavior vectors have been obtained at the rth order.

[0174] For example:

[0175] -The r-th order event matching degree of the protagonist’s singing skill is 0.8, and its connection feedback behavior vector is [0.6, 0.7] (assuming that this vector is used to represent evaluation intensity and emotional polarity).

[0176] -The r-th order event matching degree of the sound effect is 0.5, and its connection feedback behavior vector is [0.4, 0.9].

[0177] Based on this information, the r-th order connection feedback behavior vector can be weighted according to the r-th order event matching degree, as shown below:

[0178] The weighted vector of the protagonist’s singing skills: 0.8*[0.6, 0.7] = [0.48, 0.56]

[0179] Weight vector of sound effects: 0.5*[0.4, 0.9] = [0.2, 0.45]

[0180] These weighted vectors can then be combined to generate the r-th order rich description vector. A common approach is to simply add them together:

[0181] - r-th order rich description vector: [0.48, 0.56] + [0.2, 0.45] = [0.68, 1.01]

[0182] In this way, the r-th order rich description vector of “Memory” is obtained, which integrates the information of all related link feedback events and is weighted according to their event matching degree, thus reflecting the audience’s evaluation of “Memory” more comprehensively.

[0183] Step S1173, in the zth feedback category, feature enrichment is performed on the rth order event description vector corresponding to the qth performance feedback event based on the rth order enriched description vector to generate the r+1th order event description vector of the qth performance feedback event.

[0184] For example, the first-order enriched description vector is used to enrich the first-order event description vector of "Memory" to generate a more descriptive second-order event description vector. Exemplarily, this feature enrichment process can usually be completed through mathematical operations such as weighted sum, convolution, etc.

[0185] Assume that we still take the feedback of the song "Memory" in the musical "Cats" as an example. In the previous step, the r-th order rich description vector has been obtained, for example, it is [0.68, 1.01].

[0186] Now, there is the r-th order event description vector of "Memory", for example it is [0.5, 0.6].

[0187] A simple method for feature enrichment is to add the r-th order enriched description vector directly to the r-th order event description vector to generate the r+1-th order event description vector:

[0188] r+1th order event description vector: [0.5, 0.6] + [0.68, 1.01] = [1.18, 1.61]

[0189] In this way, the r+1th order event description vector of “Memory” contains the audience’s feedback on the song more comprehensively, while also taking into account the information of related feedback events (such as the protagonist’s singing skills and sound effects).

[0190] In practice, feature enrichment may involve more complex calculations, such as using deep learning models for feature extraction and fusion, but the basic idea is the same: enriching and enhancing the original event description vector based on relevant information.

[0191] Step S1174, when r+1 is the threshold feature enrichment round number, the r+1th order event description vector is output as the target event description vector of the qth performance feedback event in the zth feedback category.

[0192] For example, assuming that the threshold feature enrichment round number is set to 2, then after completing the above steps, the target event description vector of "Memory" will be obtained. This vector is a comprehensive description of the audience's evaluation of "Memory", including their feedback on the protagonist's singing skills, sound effects, and song melody.

[0193] It is worth noting that the expressions such as r-order and 0-order in the above description are used to describe the iterative process or a certain hierarchical relationship in the algorithm. Generally, order can be understood as a step, level or number of iterations.

[0194] In the above context: the 0th order event matching may refer to the event matching degree in the initial state without any processing or weighting. The rth order usually refers to the result obtained after r rounds of processing or iteration. For example, the rth order connection feedback behavior vector is the connection feedback behavior vector obtained after r rounds of processing.

[0195] For example, in a possible implementation, the following further introduces the training steps of the graph neural network for executing the aforementioned steps S111 to S117, which may include:

[0196] Step S101, parsing x training user feedback graph data. Each training user feedback graph data includes y template performance feedback events. The x training user feedback graph data are respectively associated with different feedback categories.

[0197] For example, suppose there are 3 training user feedback graph data, each graph contains 5 template performance feedback events, and these template performance feedback events may be associated with different feedback categories, such as singing skills, sound effects, stage design, etc.

[0198] Step S102, input the x training user feedback graph data into the initial graph neural network, and parse the template connection feedback event of the qth template performance feedback event for the template event dependency type of the qth template performance feedback event in the zth feedback category of the initial graph neural network. z is not greater than x. q is not greater than y.

[0199] For example, this initial graph neural network can be trained by deep learning techniques, such as being composed of multiple layers of neurons, to extract effective features and patterns from the data. For example, suppose that the template performance feedback event "the protagonist's singing skills" is being processed. It is necessary to determine other connected feedback events related to this event, such as "sound effects", and parse out their event dependency types. This type may be based on a variety of factors, such as the specific content of the feedback, emotional tendencies, etc.

[0200] Step S103: determining the template event dependency vector of the template connection feedback event of the qth template performance feedback event with respect to the qth template performance feedback event according to the template event dependency type.

[0201] The template event dependency type includes a first template event dependency type and a second template event dependency type. The first template event dependency type represents that the template event dependency vector is determined based on the template connection feedback event of the qth template performance feedback event, and the first template feedback connection vectors respectively associated in the x feedback categories. The second template event dependency type represents that the template event dependency vector is determined based on the template connection feedback event of the qth template performance feedback event, and the first template feedback connection vector corresponding to the zth feedback category.

[0202] This step generates an event dependency vector, which represents the relationship between each link feedback event and the target performance feedback event. This event dependency vector may be determined by the template feedback connection vector, which may be resolved in all feedback categories or in a specific feedback category.

[0203] Step S104, in the zth feedback category, according to the template connection feedback event of the qth template performance feedback event with respect to the template event dependency vector of the qth template performance feedback event, determine the template event description vector to be enriched of the qth template performance feedback event.

[0204] In this step, an event description vector to be enriched is generated based on the event dependency vector generated in the previous step. This vector more comprehensively represents the target performance feedback event and also contains information about related connection feedback events.

[0205] Step S105 , based on the to-be-enriched template event description vector, obtain a predicted target event description vector of the qth template performance feedback event in the zth feedback category.

[0206] The predicted target event description vector is used to fuse to obtain a template feedback knowledge description vector that expresses the qth template performance feedback event.

[0207] For example, using a graph neural network to process the template event description vector to be enriched to obtain the predicted target event description vector can be regarded as the network's understanding or interpretation of the event "the protagonist's singing skills". For example, the predicted target event description vector is fused to generate a template feedback knowledge description vector, which contains more information, such as multiple aspects of the audience's evaluation of the "protagonist's singing skills".

[0208] Step S106, when generating the template feedback knowledge description vectors associated with each of the y template performance feedback events, determine the matching degree of the template event combination in the y template performance feedback events according to the template feedback knowledge description vectors associated with each of the y template performance feedback events. The template event combination includes two different template performance feedback events in the y template performance feedback events.

[0209] For example, the similarity or correlation between the two feedback events "the singing skills of the protagonist" and "the sound effects" to a certain extent can be calculated.

[0210] Step S107: updating the initial graph neural network according to the matching degree of the template event combination to generate a graph neural network.

[0211] For example, the matching degree obtained in the previous step is used to update the initial graph neural network to generate a trained graph neural network. This graph neural network can process new performance feedback events more accurately and generate more precise event description vectors and matching degrees.

[0212] The y template performance feedback events correspond to e template event combinations, where e is a positive integer.

[0213] Step S107 may specifically be: among the e template event combinations, the template event combinations that have event connection relationships in the x training user feedback graph data are output as positive template event combinations, and the template event combinations that do not have event connection relationships in the x training user feedback graph data are output as negative template event combinations. Then, based on the matching degree of the positive template event combination and the matching degree of the negative template event combination, a training generation is obtained, and the initial graph neural network is updated according to the training cost to generate a graph neural network.

[0214] For example, this step is to better understand the relationship between template performance feedback events through graph neural network training. The following is a specific scenario example:

[0215] Suppose we are processing feedback for the song "Memory" in the musical "Cats". There are five template performance feedback events: 1) the main character's singing skills 2) sound effects 3) stage design 4) lighting effects 5) audience interaction. Based on these five events, we can generate 10 template event combinations (for example, [main character's singing skills, sound effects], [main character's singing skills, stage design], etc.).

[0216] Next, these template event combinations are divided into two categories:

[0217] Template event combinations that have event connection relationships in the training user feedback graph data are output as positive template event combinations.

[0218] Template event combinations for which there is no event connection relationship in the training user feedback graph data are output as negative template event combinations.

[0219] For example, if the audience's feedback data often mentions "the main character's singing skills" and "sound effects", then this combination may be regarded as a positive template event combination. If "lighting effects" and "audience interaction" rarely appear in the same feedback, then this combination may be regarded as a negative template event combination.

[0220] Then, the training cost is calculated based on the matching degree of the positive template event combination and the negative template event combination. The training cost can be regarded as the gap between the network prediction result and the actual result, which is usually used to guide the update and optimization of the network.

[0221] Finally, the initial graph neural network is updated according to the training cost to generate a trained graph neural network. This trained graph neural network can more accurately understand and process new performance feedback events.

[0222] In a possible implementation, the step of training the stage preference prediction model includes:

[0223] Step A110, obtaining a template audience behavior mining data sequence, wherein the template audience behavior mining data sequence includes a template audience behavior mining data subsequence for each audience user, and annotated stage preference data respectively associated with each audience user.

[0224] For example, we collected a large amount of audience behavior mining data after stage performances, including their comments and evaluations on various aspects such as singing, sound, and stage design. At the same time, we also asked these audiences to fill out questionnaires about their preferences for stage performances as annotated stage preference data. For example, some audiences prefer gorgeous stage designs, while others pay more attention to the singing quality of songs.

[0225] Step A120, based on the template audience behavior mining data sequence, the first deep learning network model is subjected to traversal loop knowledge learning to generate a second deep learning network model after traversal loop knowledge learning, and the target typical preference features established by the second deep learning network model for each audience user are obtained.

[0226] For example, the template audience behavior mining data sequence collected above is input into a first deep learning network model in an initial state for training. This first deep learning network model can process sequence data and extract useful features and patterns from it. In multiple iterations, the first deep learning network model will continuously optimize its parameters so that the gap between the predicted results and the actual results becomes smaller and smaller. After a series of knowledge learning steps of traversal cycles, a second deep learning network model after traversal cycle knowledge learning is generated. At the same time, the target typical preference features established by this second deep learning network model for each audience user are obtained. These target typical preference features represent the unique preferences of each audience for stage performances. For example, it may be found that "the protagonist's singing skills", "sound effects" and "stage design" are the most important features because they have the greatest impact on predicting the audience's stage preferences. Then, these features are set as fixed parameters in the second deep learning network model.

[0227] Step A130, setting each target typical preference feature as each fixed typical preference feature in the first deep learning network model, and performing traversal cycle transfer knowledge learning on the first deep learning network model based on the template audience behavior mining data sequence and the second deep learning network model, to generate a target deep learning network model after traversal cycle knowledge learning.

[0228] Next, these target typical preference features are set as fixed parameters in the first deep learning network model, and the first deep learning network model is again traversed and cycled through for transfer knowledge learning based on the template audience behavior mining data sequence and the second deep learning network model. This means that the knowledge learned by the second deep learning network model when learning audience behavior mining data is transferred to the scenario of processing other audience behavior mining data. This process enables the first deep learning network model to better understand and process new audience behavior mining data.

[0229] Step A140, generating a stage preference prediction model according to the target preference description unit in the target deep learning network model.

[0230] In a possible implementation manner, the first deep learning network model includes a first preference description unit and a first output unit, and when performing a traversal knowledge learning on the first deep learning network model, the following operations are performed:

[0231] 1. Load the extracted first template audience behavior mining data into the first preference description unit to generate a first preference description feature corresponding to the first template audience behavior mining data.

[0232] For example, the first template audience behavior mining data is selected, which may include audience users' behavior mining data on "the main character's singing skills", "sound effects", "stage design" and other aspects. These first template audience behavior mining data are loaded into the first preference description unit of the first deep learning network model. The first preference description unit may be an embedding layer or a convolutional layer, which is used to extract useful features from the first template audience behavior mining data. After processing, the first preference description feature is obtained.

[0233] 2. Based on the first output unit, the feature deviations between the first preference description feature and each currently established target typical preference feature are calculated respectively, and based on each feature deviation, a first output result of the audience behavior mining data for the first template is obtained.

[0234] For example, the first output unit may be a fully connected layer or a softmax layer, which is used to convert the first preference description feature into a prediction result. The feature deviation between the first preference description feature and the currently established typical preference features of each target (such as "like gorgeous stage design", "pay attention to the quality of song performance", etc.) is calculated, that is, the gap between them.

[0235] 3. Update the network layer weight information of the first deep learning network model and the typical preference features of each target based on the prediction error between the first output result and the labeled stage preference data corresponding to the first template audience behavior mining data.

[0236] For example, the difference between the first output result and the first audience's labeled stage preference data is compared, that is, the prediction error. Then, based on this prediction error, the network weight information is updated through the back propagation algorithm, and the typical preference features of each target are also updated.

[0237] In a possible implementation, step A130 may include:

[0238] Step A131, setting each target typical preference feature as each fixed typical preference feature in the first deep learning network model, and performing traversal loop knowledge learning on the first deep learning network model based on the template audience behavior mining data sequence, to generate a third deep learning network model after traversal loop knowledge learning.

[0239] For example, in the traversal cycle knowledge learning process of the previous step, it may be found that "the main character's singing skills", "sound effects" and "stage design" are the most important features because they have the greatest impact on predicting the audience's stage preferences. Then, these features are set as fixed parameters in the first deep learning network model. Next, the above-mentioned template audience behavior mining data sequence and stage preference annotation data are used again to perform another round of traversal cycle knowledge learning on the first deep learning network model with fixed feature parameters in the previous step, generating a third deep learning network model.

[0240] Step A132, based on the template audience behavior mining data sequence and the second deep learning network model, the third deep learning network model is subjected to traversal cycle transfer knowledge learning to generate a target deep learning network model after traversal cycle knowledge learning.

[0241] For example, based on the second deep learning network model, the third deep learning network model is subjected to traversal cycle transfer knowledge learning. This process will enable the network model to better understand and process new audience behavior mining data, and ultimately generate a target deep learning network model.

[0242] In a possible implementation manner, the first deep learning network model includes a second preference description unit and a second output unit, and when performing a traversal knowledge learning on the first deep learning network model, the following operations are performed:

[0243] 1. Load the extracted second template audience behavior mining data into the second preference description unit to generate second preference description features corresponding to the second template audience behavior mining data.

[0244] 2. Load the second preference description feature into the second output unit, and calculate the first feature deviation between the second preference description feature and each of the target typical preference features.

[0245] 3. Obtain a second output result of the training and learning of the audience behavior mining data for the second template based on the deviations of each first feature.

[0246] 4. Update the network layer weight information of the first deep learning network model based on the prediction error between the second output result and the labeled stage preference data corresponding to the second template audience behavior mining data.

[0247] In a possible implementation, the third deep learning network model includes a third preference description unit and a third output unit, and when the third deep learning network model is subjected to a traversal transfer knowledge learning, the following operations are performed:

[0248] 1. Load the extracted third template audience behavior mining data into the second deep learning network model to generate the third preference description features extracted by the second deep learning network model.

[0249] 2. Input the third template audience behavior mining data into the third preference description unit after traversal loop knowledge learning to generate a corresponding fourth preference description feature, and based on the third output unit, based on the fourth preference description feature and the typical preference features of each target, generate a third output result of training learning.

[0250] 3. Calculate the first transfer learning Loss function value based on the second feature deviation between the third preference description feature and the fourth preference description feature, and calculate the first Loss function value based on the prediction error between the third output result and the labeled stage preference data corresponding to the third template audience behavior mining data.

[0251] 4. Update the network layer weight information of the third deep learning network model based on the first transfer learning Loss function value and the first Loss function value.

[0252] In a possible implementation, when performing a traversal transfer knowledge learning on the first deep learning network model, the following operations are performed:

[0253] 1. Load the extracted fourth template audience behavior mining data into the second deep learning network model to generate the fifth preference description features extracted by the second deep learning network model, and load the fourth template audience behavior mining data into the first deep learning network model to generate the extracted fourth preference description features and the third output result of training learning.

[0254] 2. Calculate the second transfer learning Loss function value based on the feature deviation between the fifth preference description feature and the fourth preference description feature, and calculate the second Loss function value based on the prediction error between the third output result and the labeled stage preference data corresponding to the fourth template audience behavior mining data.

[0255] 3. Update the network layer weight information of the first deep learning network model based on the second transfer learning Loss function value and the second Loss function value.

[0256] In a possible implementation, the first deep learning network model includes a second preference description unit and a second output unit, and the fourth template audience behavior mining data is loaded into the first deep learning network model to generate the extracted fourth preference description feature and the third output result of the training learning, including:

[0257] 1. Load the fourth template audience behavior mining data into the second preference description unit to generate extracted fourth preference description features.

[0258] 2. Based on the second output unit, based on the fourth preference description feature and the typical preference features of each target, generate a third output result of training learning.

[0259] Further, Figure 2 FIG. 1 is a schematic diagram showing the hardware structure of a smart stage performance service system 100 for implementing the method provided in an embodiment of the present application. Figure 2 As shown, the smart stage performance service system 100 may include at least one processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 and a controller 108 for communication functions. It can be understood by those skilled in the art that Figure 2 The structure shown is only for illustration and does not limit the structure of the smart stage performance service system 100. For example, the smart stage performance service system 100 may also include Figure 2 More or fewer components as shown, or with Figure 2 Different configurations are shown.

[0260] The memory 104 can be used to store software programs and modules of application software, such as program instructions corresponding to the above-mentioned method embodiments in the embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, realizing the above-mentioned data analysis method applied to smart stage technology performance services. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the smart stage performance service system 100 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0261] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the smart stage performance service system 100. In one example, the transmission device 106 includes a network adapter, which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency module, which is used to communicate with the Internet wirelessly.

[0262] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of the embodiments of the present application are described. Other embodiments are within the scope of the attached claims. In some cases, the actions or steps recorded in the claims can be performed according to an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0263] Each embodiment in the embodiments of the present application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the above different embodiments, since they are basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0264] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program. The above program may be stored in a computer-readable storage medium, and the above storage medium may be a read-only memory, a disk or an optical disk, etc.

Claims

1. A data analysis method applied to smart stage technology performance services, characterized in that: The method comprises: Obtain x audience user feedback graph data related to the target stage performance, and perform feature extraction on the x audience user feedback graph data to obtain a feedback knowledge description vector of each performance feedback event; each audience user feedback graph data includes y performance feedback events, and the x audience user feedback graph data are respectively associated with different feedback categories; Performing frequent pattern item mining on the feedback knowledge description vector of each performance feedback event to obtain corresponding frequent pattern item data as audience behavior mining data; Performing stage preference prediction on the audience behavior mining data based on a pre-trained stage preference prediction model to generate corresponding stage preference prediction data; Pushing corresponding content recommendation information to a cloud service center of future stage performances associated with the target stage performance based on the stage preference prediction data; The step of extracting features from the x audience user feedback graph data to obtain a feedback knowledge description vector for each performance feedback event includes: In the zth feedback category, the first feedback connection vector of the connection feedback event of the qth performance feedback event is analyzed, and the target performance feedback behavior vector of the qth performance feedback event is analyzed; In the zth feedback category, the basic event matching degree of the connection feedback event of the qth performance feedback event with respect to the qth performance feedback event is determined based on the first feedback connection vector and the target performance feedback behavior vector; If the basic event matching degree in the zth feedback category is greater than the set matching degree, the first event dependency type is output as the event dependency type of the connection feedback event of the qth performance feedback event for the qth performance feedback event; If the basic event matching degree in the z-th feedback category is not greater than the set matching degree, the second event dependency type is output as the event dependency type of the connection feedback event of the q-th performance feedback event for the q-th performance feedback event; z is not greater than x, and q is not greater than y; According to the event dependency type, the event dependency vector of the connection feedback event of the qth performance feedback event for the qth performance feedback event is determined, wherein the first event dependency type indicates that the event dependency vector is determined based on the connection feedback event of the qth performance feedback event, and the first feedback connection vectors are respectively associated in x feedback categories; the second event dependency type indicates that the event dependency vector is determined based on the connection feedback event of the qth performance feedback event, and the first feedback connection vector corresponding to the zth feedback category; In the zth feedback category, according to the event dependency vector of the connected feedback event of the qth performance feedback event for the qth performance feedback event, determine the event description vector to be enriched of the qth performance feedback event; Based on the event description vector to be enriched, a target event description vector of the qth performance feedback event in the zth feedback category is obtained, and the target event description vector is used for fusion to obtain a feedback knowledge description vector expressing the qth performance feedback event.

2. The data analysis method applied to smart stage technology performance services according to claim 1 is characterized in that: The training steps of the stage preference prediction model include: Acquire a template audience behavior mining data sequence, wherein the template audience behavior mining data sequence includes template audience behavior mining data subsequences for each audience user, and annotated stage preference data associated with each audience user; According to the template audience behavior mining data sequence, the first deep learning network model is subjected to traversal cycle knowledge learning to generate a second deep learning network model after traversal cycle knowledge learning, and the target typical preference features established by the second deep learning network model for each audience user are obtained; Setting each target typical preference feature as each fixed typical preference feature in the first deep learning network model, and performing traversal cycle transfer knowledge learning on the first deep learning network model based on the template audience behavior mining data sequence and the second deep learning network model to generate a target deep learning network model after traversal cycle knowledge learning; A stage preference prediction model is generated according to the target preference description unit in the target deep learning network model.

3. The data analysis method applied to smart stage technology performance services according to claim 2 is characterized in that: The first deep learning network model includes a first preference description unit and a first output unit. When the first deep learning network model is subjected to a traversal knowledge learning, the following operations are performed: Loading the extracted first template audience behavior mining data into the first preference description unit to generate a first preference description feature corresponding to the first template audience behavior mining data; According to the first output unit, the feature deviations between the first preference description feature and each currently established target typical preference feature are respectively calculated, and according to each feature deviation, a first output result of the audience behavior mining data for the first template is obtained; Based on the prediction error between the first output result and the labeled stage preference data corresponding to the first template audience behavior mining data, the network layer weight information of the first deep learning network model and the typical preference features of each target are updated.

4. The data analysis method applied to smart stage technology performance services according to claim 2 or 3 is characterized in that: The method of setting each target typical preference feature as each fixed typical preference feature in the first deep learning network model, and performing traversal cycle transfer knowledge learning on the first deep learning network model according to the template audience behavior mining data sequence and the second deep learning network model, and generating a target deep learning network model after traversal cycle knowledge learning, includes: Setting each target typical preference feature as each fixed typical preference feature in the first deep learning network model, and performing traversal cycle knowledge learning on the first deep learning network model according to the template audience behavior mining data sequence, to generate a third deep learning network model after traversal cycle knowledge learning; According to the template audience behavior mining data sequence and the second deep learning network model, the third deep learning network model is subjected to traversal cycle transfer knowledge learning to generate a target deep learning network model after traversal cycle knowledge learning.

5. The data analysis method applied to smart stage technology performance services according to claim 4 is characterized in that: The first deep learning network model includes a second preference description unit and a second output unit. When the first deep learning network model is subjected to a traversal knowledge learning, the following operations are performed: Loading the extracted second template audience behavior mining data into the second preference description unit to generate a second preference description feature corresponding to the second template audience behavior mining data; Loading the second preference description feature into the second output unit, and respectively calculating the first feature deviation between the second preference description feature and each of the target typical preference features; Obtaining a second output result of the training and learning of the audience behavior mining data for the second template according to each first feature deviation degree; The network layer weight information of the first deep learning network model is updated according to the prediction error between the second output result and the labeled stage preference data corresponding to the second template audience behavior mining data.

6. The data analysis method for smart stage technology performance service according to claim 4 is characterized in that: The third deep learning network model includes a third preference description unit and a third output unit. When the third deep learning network model is subjected to a traversal transfer knowledge learning, the following operations are performed: Loading the extracted third template audience behavior mining data into the second deep learning network model to generate a third preference description feature extracted by the second deep learning network model; Input the third template audience behavior mining data into the third preference description unit after traversal loop knowledge learning to generate a corresponding fourth preference description feature, and generate a third output result of training learning according to the third output unit, the fourth preference description feature and the typical preference features of each target; Calculating a first transfer learning Loss function value according to a second feature deviation between the third preference description feature and the fourth preference description feature, and calculating a first Loss function value according to a prediction error between the third output result and the labeled stage preference data corresponding to the third template audience behavior mining data; Update the network layer weight information of the third deep learning network model according to the first transfer learning Loss function value and the first Loss function value.

7. The data analysis method applied to smart stage technology performance services according to claim 2 or 3 is characterized in that: When performing a traversal transfer knowledge learning on the first deep learning network model, the following operations are performed: Loading the extracted fourth template audience behavior mining data into the second deep learning network model to generate the fifth preference description feature extracted by the second deep learning network model, and loading the fourth template audience behavior mining data into the first deep learning network model to generate the extracted fourth preference description feature and the third output result of training and learning; Calculating a second transfer learning Loss function value according to the feature deviation between the fifth preference description feature and the fourth preference description feature, and calculating a second Loss function value according to the prediction error between the third output result and the labeled stage preference data corresponding to the fourth template audience behavior mining data; Update the network layer weight information of the first deep learning network model according to the second transfer learning Loss function value and the second Loss function value.

8. The data analysis method applied to smart stage technology performance services according to claim 7 is characterized in that: The first deep learning network model includes a second preference description unit and a second output unit, and the fourth template audience behavior mining data is loaded into the first deep learning network model to generate the extracted fourth preference description feature and the third output result of the training learning, including: loading the fourth template audience behavior mining data into the second preference description unit to generate an extracted fourth preference description feature; According to the second output unit, a third output result of training learning is generated according to the fourth preference description feature and the typical preference features of each target.

9. A smart stage performance service system, characterized in that: The smart stage performance service system includes a processor and a readable storage medium, wherein the readable storage medium stores a program, which, when executed by the processor, implements the data analysis method applied to smart stage technical performance services as described in any one of claims 1-8.

Images