AI dynamic weight analysis-based short video projection flow accurate regulation and control system

Abstract

The invention discloses an AI dynamic weight analysis-based short video projection flow precise regulation and control system, which comprises the following modules: a structured feature acquisition module, which is used for constructing a structured feature input set; the input preprocessing module is used for forming a unified input vector; the traffic target modeling module is used for inputting the unified input vector into the improved PLE network model and outputting a plurality of traffic target prediction results and corresponding dynamic weight vectors; the multi-target fusion scoring module is used for calculating and outputting a comprehensive flow throwing score value; the strategy mapping acquisition module is used for generating a structured flow parameter set; the precise delivery execution module is used for executing precise recommendation delivery operation of the short video content; and the feedback calibration updating module is used for collecting user feedback data and updating parameters of the improved PLE network model. According to the method, improved PLE modeling and a multi-target scoring mechanism are fused, and an intelligent flow casting system suitable for multi-task prediction and dynamic regulation and control is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and advertisement optimization, and particularly relates to a short video streaming precise regulation system based on AI dynamic weight analysis. BACKGROUND

[0002] With the rapid development of short video platform content ecology, personalized recommendation based on user interest and content characteristics has become a key means to improve platform activity and commercial conversion efficiency. Existing short video streaming regulation methods usually rely on a single optimization target (such as click rate or completion rate) to execute content distribution strategies through static rules or simple weighting mechanisms. However, in actual applications, the following problems generally exist:

[0003] Firstly, the diversified propagation effect of short video content cannot be comprehensively evaluated by a single indicator, and there is a lack of collaborative modeling of multiple optimization targets such as like rate, comment rate and conversion rate, resulting in feedback bias and target deviation problems in actual deployment of recommendation results; secondly, traditional models cannot dynamically adjust the importance distribution of each optimization target according to specific content, user behavior and historical feedback data, lacking adaptive modeling capability for multi-task weight, causing insufficient generalization capability of regulation strategies in different situations; thirdly, the closed-loop calibration mechanism between recommendation strategies and user feedback is imperfect, making it difficult to realize continuous optimization of the model and dynamic evolution of the rules, limiting the long-term stability and precision improvement space of the system.

[0004] Therefore, how to provide a short video streaming precise regulation system based on AI dynamic weight analysis is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a short video streaming precise regulation system based on AI dynamic weight analysis. The present application makes full use of the improved PLE network model, multi-objective fusion scoring function and feedback calibration mechanism to realize adaptive modeling of short video content multi-streaming targets and generation of regulation strategies, and has the advantages of high target prediction accuracy, strong strategy matching precision and sustainable optimization of streaming feedback.

[0006] According to the short video streaming precise regulation system based on AI dynamic weight analysis of the embodiment of the present application, the following modules are included:

[0007] The structured feature acquisition module is used to acquire multi-source streaming related data in the short video platform and construct a structured feature input set;

[0008] The input preprocessing module is used to preprocess the structured feature input set to form a unified input vector;

[0009] The flow target modeling module is configured to input the unified input vector into the improved PLE network model, output a plurality of flow target prediction results and corresponding dynamic weight vectors;

[0010] The multi-target fusion scoring module is configured to input the plurality of flow target prediction results and the corresponding dynamic weight vectors into a multi-target fusion scoring function, and calculate a comprehensive flow score value;

[0011] The strategy mapping obtaining module is configured to query a preset score interval mapping rule according to an interval in which the comprehensive flow score value is located, and obtain a structured flow parameter set;

[0012] The precise delivery execution module is configured to perform a precise recommendation delivery operation of the video content based on the structured flow parameter set.

[0013] The feedback calibration updating module is configured to collect user feedback data after the precise recommendation delivery, and update parameters of the improved PLE network model through a feedback calibration mechanism according to a deviation between the user feedback data and the flow target prediction result.

[0014] According to the short video flow precise regulation method based on AI dynamic weight analysis, the following steps are included:

[0015] Collecting multi-source flow related data in a short video platform, and constructing a structured feature input set;

[0016] Pretreating the structured feature input set to form a unified input vector;

[0017] Inputting the unified input vector into the improved PLE network model, and outputting a plurality of flow target prediction results and corresponding dynamic weight vectors;

[0018] Inputting the plurality of flow target prediction results and the corresponding dynamic weight vectors into a multi-target fusion scoring function, and calculating a comprehensive flow score value;

[0019] According to an interval in which the comprehensive flow score value is located, a preset score interval mapping rule is queried to obtain a structured flow parameter set;

[0020] Performing a precise recommendation delivery operation of the video content based on the structured flow parameter set;

[0021] Collecting user feedback data after the precise recommendation delivery, and updating parameters of the improved PLE network model through a feedback calibration mechanism according to a deviation between the user feedback data and the flow target prediction result.

[0022] Optionally, the collecting multi-source flow related data in the short video platform and constructing the structured feature input set specifically includes:

[0023] Collect multi-source data related to the flow in the short video platform, the multi-source data including user behavior data, video content attribute data and historical flow feedback data, and construct an original data set;

[0024] Perform field screening and format standardization processing on the original data set, retain flow target associated fields, and unify field naming specifications;

[0025] Perform structured mapping operation according to the type of the field, including setting the arrangement order of the field in the input vector, labeling the source category and time index information of the field, and obtaining a structured feature input set.

[0026] Optionally, the structured feature input set is preprocessed to form a unified input vector, specifically including:

[0027] According to the field source category in the structured feature input set, the field is divided into user feature field, video content feature field and behavior interaction feature field;

[0028] Perform embedding encoding processing on the user feature field, wherein an index mapping relationship table is constructed for the category type field and an embedding index table is established to generate an embedding vector, and the numerical type field is normalized and directly used as an input dimension to generate a user embedding vector;

[0029] Perform embedding encoding processing on the video content feature field, wherein an embedding index table is constructed for the label type field and an embedding vector is generated, and the text type field is segmented and a semantic vector is generated using a text encoder to constitute a video content embedding vector;

[0030] Perform normalization processing on the behavior interaction feature field and form a behavior interaction feature vector according to the field order;

[0031] Concatenate the user embedding vector, the video content embedding vector and the behavior interaction feature vector, and label the source category of each vector to form a unified input vector.

[0032] Optionally, the unified input vector is input into the improved PLE network model to output a plurality of flow target prediction results and corresponding dynamic weight vectors, specifically including:

[0033] Construct an improved PLE network model, the improved PLE network model including a multi-layer shared expert tower, a multi-task specialized expert tower, an expert cross information interaction structure, a gate structure based on attention and a multi-task output fusion layer;

[0034] The unified input vector is input into a multi-layer shared expert tower composed of a plurality of parallel general expert sub-networks, each general expert sub-network including two layers of full connection structure and residual connection mechanism, and a ReLU activation function is used to extract general semantic features of the input vector, and a plurality of global shared expert feature vectors are output to form a global shared feature representation;

[0035] For a plurality of optimization target tasks in short video streaming, including click rate prediction, complete play rate prediction, like rate prediction, comment rate prediction and conversion rate prediction, five types of tasks, corresponding task identification codes are set respectively, the task identification codes are represented by one-hot encoding structure, and are converted into task identification embedding vectors through a trainable embedding layer;

[0036] The task identification embedding vector and the unified input vector are spliced and fused to form a task input representation, and the task input representation is input into a corresponding multi-task specialized expert tower, the specialized expert tower including two layers of full connection structure and ReLU activation function, to generate an initial task specialized representation.

[0037] Optionally, the expert cross information interaction structure, the attention-based gating structure and the multi-task output fusion layer specifically include:

[0038] The initial task specialized representation is input into the expert cross information interaction structure, which is based on the initial task specialized representation, calculates the feature similarity value between any two optimization target tasks using cosine similarity, constructs a similarity matrix, selects other optimization target tasks with similarity higher than a preset threshold from the similarity matrix as adjacent tasks, and extracts their initial optimization target task specialized representations to form a task cross information set;

[0039] The task cross information set is input into the cross information fusion structure, and the similarity weight between the adjacent task and the current optimization target task is calculated based on the attention mechanism, the specialized representations in the task cross information set are weighted and fused, and the enhanced specialized representation of the current optimization target task is obtained;

[0040] The enhanced specialized representation of the current optimization target task and its corresponding global shared feature representation are spliced and input into a dynamic fusion structure with Sigmoid gating weight, the fusion proportion of different information sources is adaptively adjusted according to the feature importance, and a fusion representation vector is output;

[0041] The fusion representation vector is input into a multi-task output fusion layer, the multi-task output fusion layer including a prediction branch and a weight branch;

[0042] The prediction branch includes two layers of full connection layer and Sigmoid activation function, and outputs a plurality of streaming target prediction results, the prediction value representing the streaming response probability of the corresponding target.

[0043] The weight branch includes a full connection layer and a Softmax normalization structure, and outputs a dynamic weight vector corresponding to multiple flow target pairs. The weight value represents the importance of each target in the final regulation strategy.

[0044] Optionally, the multiple flow target prediction results and the corresponding dynamic weight vectors are input into a multi-target fusion scoring function to calculate a comprehensive flow score value, specifically including:

[0045] The multiple optimization target task prediction results and the corresponding dynamic weight vectors are normalized. The dynamic weight vector is standardized by using a Softmax function, so that the weight sum of the five types of optimization target tasks is 1, forming a dynamic weight distribution vector.

[0046] Based on the multiple optimization target task prediction results, independent evaluation indexes of short video content in click rate prediction, complete playback rate prediction, like rate prediction, comment rate prediction, and conversion rate prediction are calculated respectively, forming a five-task evaluation result set.

[0047] According to the dynamic weight distribution vector and the five-task evaluation result set, a weighted summary method is used to calculate the comprehensive flow score value of each short video content.

[0048] Optionally, according to the interval in which the comprehensive flow score value is located, a preset score interval mapping rule is queried to obtain a structured flow parameter set, specifically including:

[0049] The comprehensive flow score value of each short video content is compared with a score interval set to match its belonging score interval identifier. The score interval set is a plurality of non-overlapping interval segments divided according to the size of the score value, and the interval boundary value is preset by the system.

[0050] According to the score interval identifier, the corresponding structured flow parameter set is retrieved in the score interval mapping rule table. The score interval mapping rule table is constructed by historical flow performance data. Each record in the rule table includes a one-to-one correspondence relationship between the score interval identifier and the structured flow parameter set.

[0051] Optionally, the user feedback data after the accurate recommendation is collected, and the parameters of the improved PLE network model are updated and improved through a feedback calibration mechanism according to the deviation between the user feedback data and the flow target prediction result, specifically including:

[0052] Based on the collected strategy execution feedback information and the corresponding structured flow parameter set, an optimization sample of the score interval mapping rule is constructed. The optimization sample takes the comprehensive flow score value as an index to establish an associated mapping relationship between the flow parameter configuration and the strategy execution result.

[0053] performing a statistical analysis operation on the historical optimization samples under each scoring interval, and calculating the mapping performance score of each field parameter combination and core index performance in the interval;

[0054] According to the mapping performance score, a sliding window strategy and an effect dominant weight mechanism are used to dynamically update the parameter configuration in the current scoring interval mapping rule table, and the update result covers the original parameter set;

[0055] When the number of optimization samples is lower than a preset threshold, the rule update of the scoring interval is suspended and marked as a to-be-verified state.

[0056] The beneficial effects of the present application are:

[0057] The present application introduces an improved PLE network model, constructs a multi-layer shared expert tower and a multi-task specialized expert tower structure, effectively improves the information interaction efficiency between various optimization objectives (including click rate, completion rate, like rate, comment rate and conversion rate) in the short video streaming task. Through the joint modeling of task identification embedding vector and task specialized input representation, the expression ability of each task to input features is enhanced, and further through the expert cross information interaction structure and the attention-based gating mechanism, the effective fusion of specialized representation between adjacent tasks is realized, which improves the expression accuracy and generalization ability of the model in the multi-objective collaborative modeling scene.

[0058] In addition, the multi-objective fusion scoring function proposed by the present application can dynamically adjust the weight distribution according to the importance of each optimization objective, and generate more reasonable comprehensive streaming score value. The matching streaming parameter set and the structured streaming parameter set realize the fine mapping from the scoring result to the streaming strategy, improve the matching degree of the streaming parameters and the landing execution effect. The feedback calibration mechanism is based on the deviation between the user's real response and the predicted result, and dynamically optimizes the scoring interval mapping rule combined with the sliding window and the effect dominant mechanism, which has the ability of continuous self iteration and regulation.

[0059] In summary, the present application not only improves the modeling accuracy of different streaming targets of the short video platform, but also significantly enhances the response ability of the system to actual feedback and the adaptability of the streaming strategy, and has good practical value and promotion potential in the large-scale short video streaming scene. BRIEF DESCRIPTION OF DRAWINGS

[0060] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. In the drawings:

[0061] Figure 1 a structural schematic diagram of a short video streaming accurate regulation system based on AI dynamic weight analysis proposed by the present application;

[0062] Figure 2 A flowchart of a short video streaming precision regulation method based on AI dynamic weight analysis is proposed for the present application.

[0063] Figure 3 A schematic diagram of an improved PLE network model in a short video streaming precision regulation method based on AI dynamic weight analysis is proposed for the present application. DETAILED DESCRIPTION

[0064] The present application will now be further described in detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0065] REFERENCE Figure 1 A short video streaming precision regulation system based on AI dynamic weight analysis includes the following modules:

[0066] The structured feature acquisition module is used to acquire multi-source streaming related data in the short video platform and construct a structured feature input set.

[0067] The input preprocessing module is used to preprocess the structured feature input set to form a unified input vector.

[0068] The streaming target modeling module is used to input the unified input vector into the improved PLE network model and output multiple streaming target prediction results and corresponding dynamic weight vectors.

[0069] The multi-target fusion scoring module is used to input the multiple streaming target prediction results and corresponding dynamic weight vectors into a multi-target fusion scoring function to calculate a comprehensive streaming score value.

[0070] The strategy mapping acquisition module is used to query the preset scoring interval mapping rule according to the interval in which the comprehensive streaming score value is located to obtain a structured streaming parameter set.

[0071] The precise placement execution module is used to perform precise recommendation and placement operations of video content based on the structured streaming parameter set.

[0072] The feedback calibration update module is used to collect user feedback data after precise recommendation and placement, and update the parameters of the improved PLE network model through a feedback calibration mechanism according to the deviation between the user feedback data and the streaming target prediction results.

[0073] REFERENCE Figures 2-3 A short video streaming precision regulation method based on AI dynamic weight analysis includes the following steps:

[0074] Step 1: Acquire multi-source streaming related data in the short video platform and construct a structured feature input set.

[0075] Step two: preprocessing the structured feature input set to form a unified input vector;

[0076] Step three: inputting the unified input vector into the improved PLE network model to output multiple flow target prediction results and corresponding dynamic weight vectors;

[0077] Step four: inputting the multiple flow target prediction results and corresponding dynamic weight vectors into the multi-target fusion scoring function to calculate the comprehensive flow score value;

[0078] Step five: according to the interval where the comprehensive flow score value is located, querying the preset score interval mapping rule to obtain the structured flow parameter set;

[0079] Step six: performing accurate recommendation and placement operation of video content based on the structured flow parameter set;

[0080] Step seven: collecting user feedback data after accurate recommendation and placement, and updating the parameters of the improved PLE network model through feedback calibration mechanism according to the deviation between the user feedback data and the flow target prediction result.

[0081] In the embodiment, the multi-source flow-related data in the short video platform is collected to construct a structured feature input set, which specifically includes:

[0082] Collecting multi-source data related to flow in the short video platform, the multi-source data including user behavior data, video content attribute data and historical flow feedback data, to construct an original data set;

[0083] Performing field filtering and format standardization processing on the original data set, retaining flow target associated fields, and unifying field naming specifications;

[0084] Performing structured mapping operation according to the type of field, including setting the arrangement order of field in input vector, labeling the source category and time index information of field, to obtain the structured feature input set.

[0085] The embodiment collects multi-source data related to flow in the short video platform, and performs field filtering, format standardization and structured mapping operation to construct a structured feature input set. This method can fully integrate heterogeneous data such as user behavior, video content attribute and historical flow feedback, provide a unified and standardized input basis for subsequent modeling, effectively improve the data quality and flow prediction accuracy of the model, and enhance the integration and generalization ability of the system to multi-source data.

[0086] In the embodiment, the structured feature input set is preprocessed to form a unified input vector, which specifically includes:

[0087] According to the field source category in the structured feature input set, the fields are divided into user feature fields, video content feature fields and behavior interaction feature fields;

[0088] An embedding encoding process is performed on the user feature fields, wherein an index mapping relationship table is constructed for category type fields and an embedding index table is established to generate embedding vectors, a normalization process is performed on numerical value type fields and the fields are directly used as input dimensions to generate user embedding vectors;

[0089] An embedding encoding process is performed on the video content feature fields, wherein an embedding index table is constructed for label type fields and embedding vectors are generated, a word segmentation process is performed on text type fields and a text encoder is used to generate semantic vectors to constitute video content embedding vectors;

[0090] A normalization process is performed on the behavior interaction feature fields and behavior interaction feature vectors are formed in field order;

[0091] The user embedding vectors, video content embedding vectors and behavior interaction feature vectors are spliced, and each segment of the vectors is labeled with its source category to form a unified input vector.

[0092] The embodiment classifies the fields in the structured feature input set, respectively performs embedding encoding, normalization and splicing fusion operations, and constructs a unified input vector. This method can retain the semantic features and structural information of fields from different sources, improve the expression ability and interpretability of the input representation, and facilitate the model to fully utilize multi-dimensional information when facing complex flow scenarios, thereby improving the overall prediction performance and control accuracy.

[0093] In the embodiment, the improved PLE network model is input with the unified input vector, and outputs a plurality of flow target prediction results and corresponding dynamic weight vectors, specifically including:

[0094] An improved PLE network model is constructed, and the improved PLE network model includes a multi-layer shared expert tower, a multi-task specialized expert tower, an expert cross information interaction structure, an attention-based gating structure and a multi-task output fusion layer;

[0095] The unified input vector is input into the multi-layer shared expert tower, and the multi-layer shared expert tower is composed of a plurality of parallel general expert sub-networks, each general expert sub-network includes two layers of full connection structure and residual connection mechanism, and adopts a ReLU activation function to extract general semantic features of the input vector, output a plurality of global shared expert feature vectors, and form a global shared feature representation;

[0096] For multiple optimization target tasks in short video streaming, including click rate prediction, complete play rate prediction, like rate prediction, comment rate prediction and conversion rate prediction, five types of tasks are set with corresponding task identification codes, which are represented by one-hot encoding structure and converted into task identification embedding vectors through a trainable embedding layer;

[0097] The task identification embedding vector and the unified input vector are spliced and fused to form a task input representation, which is input into the corresponding multi-task specialized expert tower, which includes two fully connected structures and ReLU activation functions, to generate an initial task specialized representation.

[0098] The improved PLE network model is introduced in this embodiment, combined with the multi-layer shared expert tower and multi-task specialized expert tower structure, to complete the mapping of the unified input vector to the multiple streaming target prediction results and dynamic weight vectors. This method significantly enhances the model's ability to recognize differences and commonalities between streaming tasks, and improves prediction accuracy through multi-task learning mechanism, effectively supporting joint modeling and weight dynamic allocation of multiple streaming targets such as click rate, complete play rate and conversion rate, providing a more reasonable input basis for subsequent control strategies.

[0099] In this embodiment, the expert cross information interaction structure, attention-based gating structure and multi-task output fusion layer specifically include:

[0100] The initial task specialized representation is input into the expert cross information interaction structure, which calculates the feature similarity value between any two optimization target tasks based on the initial task specialized representation using cosine similarity, constructs a similarity matrix, selects other optimization target tasks with similarity higher than a preset threshold as adjacent tasks from the similarity matrix, extracts their initial optimization target task specialized representations, and forms a task cross information set;

[0101] The task cross information set is input into the cross information fusion structure, which calculates the similarity weight between the adjacent tasks and the current optimization target task based on the attention mechanism, and weights and fuses the specialized representations in the task cross information set to obtain an enhanced specialized representation of the current optimization target task;

[0102] The enhanced specialized representation of the current optimization target task and its corresponding global shared feature representation are spliced and input into the dynamic fusion structure with Sigmoid gating weight, which adaptively adjusts the fusion proportion of different information sources according to feature importance, and outputs a fusion representation vector;

[0103] The fusion representation vector is input into the multi-task output fusion layer, which includes a prediction branch and a weight branch;

[0104] The prediction branch includes two fully connected layers and a Sigmoid activation function, and outputs multiple prediction results of the flow target. The prediction value represents the probability of the flow response of the corresponding target.

[0105] The weight branch includes one fully connected layer and a Softmax normalization structure, and outputs a dynamic weight vector corresponding to multiple flow targets. The weight value represents the importance of each target in the final regulation strategy.

[0106] In this embodiment, by constructing an expert cross-information interaction structure, an attention-based gating structure, and a multi-task output fusion layer, the specialized representations of different flow targets can be efficiently shared and enhanced based on feature similarity. In the fusion stage, a Sigmoid gating mechanism is introduced to dynamically adjust the importance ratio of global and local information. This method effectively improves the modeling correlation and fusion flexibility between tasks, enhances the processing capability of the prediction model for the multi-task coupling relationship in the complex short video flow environment, and thus improves the accuracy of the overall prediction and the stability of the regulation strategy.

[0107] In this embodiment, the multiple flow target prediction results and the corresponding dynamic weight vectors are input into a multi-target fusion scoring function to calculate a comprehensive flow score value, which specifically includes:

[0108] The multiple optimization target task prediction results and the corresponding dynamic weight vectors are normalized. The Softmax function is used to standardize the dynamic weight vector, so that the weight sum of the five types of optimization target tasks is 1, forming a dynamic weight distribution vector.

[0109] Based on the multiple optimization target task prediction results, independent evaluation indicators of the short video content in the click rate prediction, the complete playback rate prediction, the like rate prediction, the comment rate prediction, and the conversion rate prediction are calculated respectively, forming a five-task evaluation result set.

[0110] According to the dynamic weight distribution vector and the five-task evaluation result set, a weighted summary method is used to calculate the comprehensive flow score value Score of each short video content:

[0111]

[0112] wherein w i represents the dynamic weight value of the i-th optimization target task, R i represents the independent evaluation indicator value of the i-th optimization target task.

[0113] The embodiment introduces a multi-objective fusion scoring function, normalizes and weights the multiple optimization target prediction results and the corresponding dynamic weight vector, and forms a distinguishable video comprehensive streaming score value. This method makes full use of the dynamic weight distribution information output by the improved PLE network model, makes the scoring result more personalized and accurate, effectively reflects the actual contribution of each target task to the strategy, and thus provides a more reliable scoring basis for subsequent strategy mapping and accurate delivery.

[0114] In the embodiment, the structured streaming parameter set is obtained by querying the preset scoring interval mapping rule according to the interval in which the comprehensive streaming score value is located, and specifically includes:

[0115] The comprehensive streaming score value of each short video content is compared with a scoring interval set, and the scoring interval identifier to which it belongs is matched, the scoring interval set being a plurality of non-overlapping interval segments divided according to the size of the score value, and the interval boundary value being a system preset;

[0116] According to the scoring interval identifier, the corresponding structured streaming parameter set is searched in the scoring interval mapping rule table, and the scoring interval mapping rule table is constructed by historical streaming performance data, and each record in the rule table includes a one-to-one correspondence between the scoring interval identifier and the structured streaming parameter set.

[0117] The embodiment realizes a scoring-driven parameter strategy selection mechanism by constructing a scoring interval mapping rule table and matching the corresponding structured streaming parameter set according to the comprehensive streaming score value to which it belongs. This method effectively maps the historical streaming performance and the current scoring result, makes the control strategy traceable and adjustable, and improves the rationality of parameter configuration and the flexibility of strategy execution.

[0118] In the embodiment, the user feedback data after accurate recommendation and delivery is collected, and the parameters of the improved PLE network model are updated according to the deviation between the user feedback data and the streaming target prediction result through a feedback calibration mechanism, and specifically includes:

[0119] Based on the collected strategy execution feedback information and the corresponding structured streaming parameter set, an optimization sample of the scoring interval mapping rule is constructed, the optimization sample taking the comprehensive streaming score value as an index, and an associated mapping relationship between the streaming parameter configuration and the strategy execution result is established;

[0120] The historical optimization sample under each scoring interval is subjected to statistical analysis operation, and the mapping efficiency score S of each field parameter combination and core index performance in the interval is calculated i :

[0121] S i =α·Norm(E i)+ β · Norm(C i )+ γ · Norm(T i );

[0122] wherein E i represents the corresponding exposure efficiency, C i represents the click rate improvement rate, T i represents the content conversion score, which is weighted and aggregated by the like rate, the comment rate and the conversion rate, and a, b and g are index control weights, and Norm(·) represents a normalization operation;

[0123] According to the mapping performance score, a sliding window strategy and an effect dominant weight mechanism are used to dynamically update the parameter configuration in the current scoring interval mapping rule table, and the update result covers the original parameter set.

[0124] When the number of optimization samples is lower than a preset threshold, the rule update of the scoring interval is suspended and a to-be-verified state is marked.

[0125] The embodiment constructs an optimization sample set of the scoring interval mapping rule, evaluates the mapping performance score of each parameter combination, and dynamically adjusts the interval parameter configuration in combination with the sliding window and the effect dominant mechanism, thereby realizing continuous optimization of the structured parameter rule. This method improves the response ability of the system to strategy feedback, enables the model to continuously adapt and evolve while maintaining stability, and effectively enhances the intelligent control ability and user feedback adaptation ability of the short video streaming system.

[0126] Example 1

[0127] In order to verify the feasibility of the present application in implementation, the present application is applied to an intelligent streaming system of a certain short video platform, 12000 popular video contents of the platform in the past three months are selected as test samples, covering multiple content fields (makeup, education, game, life, food, etc.), and modeling is performed in combination with multi-source data such as user historical behavior, content label and streaming feedback information.

[0128] In actual deployment, first, the feature extraction operation is performed on the above-mentioned video data and user behavior data to generate a unified input vector, and the improved PLE network model is input. The model structure includes a multi-layer shared expert tower, a multi-task specialized expert tower, an expert cross information interaction structure, an attention-based gate structure and a multi-task output fusion layer, which supports joint modeling of five target tasks of click rate (CTR), complete play rate (WPR), like rate (LPR), comment rate (CPR) and conversion rate (CVR). The model output includes the prediction values of the five streaming targets and their corresponding dynamic weights, which are further fused into a comprehensive streaming score value.

[0129] The system maps the score value to a preset score interval rule, automatically acquires a corresponding structured push flow parameter set, and executes a push strategy. On the basis of user real interaction feedback, the system synchronously collects response data to construct a sample pool, and periodically optimizes the mapping rule and task prediction weight based on a strategy update module.

[0130] The following are the index comparison results of the method of the application and the traditional static rule method under the same platform and the same video pool:

[0131] Table 1 Short video intelligent push flow effect comparison data table

[0132]

[0133] As can be seen from Table 1, the method of the application shows significant advantages in five core task indicators, with an average CTR increase of 6.8 percentage points, close to an 80% increase, representing a significant increase in user interest in recommended content. WPR and LPR increased by 5.24% and 2.42%, respectively, indicating that users have a stronger perception of viewing depth and content quality. CPR and CVR also increased by more than 1%, showing good interactivity and commercial conversion effect. The recommendation hit rate increased from 61.2% to 78.4%, enhancing the platform recommendation accuracy. The strategy feedback update frequency was 156 times, reflecting the dynamic adjustment capability and intelligent response mechanism of the method of the application.

[0134] Overall, this embodiment verifies the advantages of the application in multi-target collaborative optimization, multi-task modeling, weight dynamic allocation, and structured strategy generation. It not only improves the push flow accuracy and user experience, but also enhances the system's strategy adaptive ability, providing strong technical support for short video platform intelligent recommendation and advertisement push.

[0135] The above describes only the preferred specific embodiments of the application, but the protection scope of the application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A short video streaming precision regulation system based on AI dynamic weight analysis, characterized in that, The method comprises the following modules: A structured feature collection module is configured to collect multi-source streaming related data in a short video platform and construct a structured feature input set; An input preprocessing module is configured to preprocess the structured feature input set and form a unified input vector; A streaming target modeling module is configured to input the unified input vector into an improved PLE network model and output a plurality of streaming target prediction results and corresponding dynamic weight vectors; A multi-target fusion scoring module is configured to input the plurality of streaming target prediction results and corresponding dynamic weight vectors into a multi-target fusion scoring function and calculate a comprehensive streaming score value; A strategy mapping acquisition module is configured to query a preset scoring interval mapping rule according to an interval in which the comprehensive streaming score value is located and obtain a structured streaming parameter set; A precise delivery execution module is configured to perform a precise recommendation and delivery operation of video content based on the structured streaming parameter set; A feedback calibration update module is configured to collect user feedback data after the precise recommendation and delivery and update parameters of the improved PLE network model through a feedback calibration mechanism according to a deviation between the user feedback data and the streaming target prediction result.

2. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 1, wherein The modules are connected through the following methods: Collect multi-source streaming related data in a short video platform and construct a structured feature input set; Preprocess the structured feature input set and form a unified input vector; Input the unified input vector into an improved PLE network model and output a plurality of streaming target prediction results and corresponding dynamic weight vectors; Input the plurality of streaming target prediction results and corresponding dynamic weight vectors into a multi-target fusion scoring function and calculate a comprehensive streaming score value; Query a preset scoring interval mapping rule according to an interval in which the comprehensive streaming score value is located and obtain a structured streaming parameter set; Perform a precise recommendation and delivery operation of video content based on the structured streaming parameter set; Collect user feedback data after the precise recommendation and delivery and update parameters of the improved PLE network model through a feedback calibration mechanism according to a deviation between the user feedback data and the streaming target prediction result.

3. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 2, wherein The method of collecting multi-source streaming related data in a short video platform and constructing a structured feature input set specifically comprises: Collect multi-source data related to streaming in a short video platform, wherein the multi-source data comprises user behavior data, video content attribute data and historical streaming feedback data, and an original data set is constructed; Perform field screening and format standardization processing on the original data set, retain streaming target associated fields, and unify field naming specifications; Perform a structured mapping operation according to the type of the field, including setting the arrangement order of the field in the input vector, labeling the source category and time index information of the field, and obtaining a structured feature input set.

4. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 2, wherein The method of preprocessing the structured feature input set and forming a unified input vector specifically comprises: According to the field source category in the structured feature input set, the field is divided into a user feature field, a video content feature field and a behavior interaction feature field; performing embedding coding processing on the user feature field, wherein an index mapping table is constructed for a category type field and an embedding index table is established to generate an embedding vector, a numerical value type field is normalized and directly used as an input dimension to generate a user embedding vector; performing embedding coding processing on the video content feature field, wherein an embedding index table is constructed for a label type field and an embedding vector is generated, a text type field is segmented and a semantic vector is generated using a text encoder to form a video content embedding vector; performing normalization processing on the behavior interaction feature field and grouping the behavior interaction feature vectors in sequence to form a behavior interaction feature vector; concatenating the user embedding vector, the video content embedding vector and the behavior interaction feature vector, and labeling the source category of each vector to form a unified input vector.

5. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 2, wherein The improved PLE network model is inputted with the unified input vector, and a plurality of prediction results of the streaming target and corresponding dynamic weight vectors are outputted, and the specific process includes the following steps: An improved PLE network model is constructed, and the improved PLE network model includes a multi-layer shared expert tower, a multi-task specialized expert tower, an expert cross information interaction structure, an attention-based gating structure and a multi-task output fusion layer. The unified input vector is inputted into the multi-layer shared expert tower, and the multi-layer shared expert tower is composed of a plurality of parallel general expert sub-networks, each general expert sub-network includes two layers of full connection structure and residual connection mechanism, adopts ReLU activation function to extract general semantic features of the input vector, outputs a plurality of global shared expert feature vectors to form a global shared feature representation; For a plurality of optimization target tasks in the short video streaming, including click rate prediction, complete playback rate prediction, like rate prediction, comment rate prediction and conversion rate prediction, five types of tasks, corresponding task identification codes are set respectively, the task identification code adopts one-hot encoding structure, and is converted into a task identification embedding vector through a trainable embedding layer; The task identification embedding vector and the unified input vector are spliced and fused to form a task input representation, and the task input representation is inputted into the corresponding multi-task specialized expert tower, and the specialized expert tower includes two layers of full connection structure and ReLU activation function to generate an initial task specialized representation.

6. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 5, wherein The expert cross information interaction structure, the attention-based gating structure and the multi-task output fusion layer specifically include: The initial task specialized representation is inputted into the expert cross information interaction structure, the expert cross information interaction structure is based on the initial task specialized representation, calculates the feature similarity value between any two optimization target tasks using cosine similarity, constructs a similarity matrix, selects other optimization target tasks with a similarity higher than a preset threshold as adjacent tasks from the similarity matrix, extracts their initial optimization target task specialized representations to form a task cross information set; The task cross information set is inputted into the cross information fusion structure, the similarity weight between the adjacent tasks and the current optimization target task is calculated based on the attention mechanism, the specialized representations in the task cross information set are weighted and fused to obtain an enhanced specialized representation of the current optimization target task; The enhanced specialized representation of the current optimization target task is spliced with the corresponding global shared feature representation, and input into a dynamic fusion structure with a Sigmoid gating weight, and a fusion proportion of different information sources is adaptively adjusted according to feature importance, and a fusion representation vector is output; The fusion representation vector is input into a multi-task output fusion layer, and the multi-task output fusion layer includes a prediction branch and a weight branch; The prediction branch includes two fully connected layers and a Sigmoid activation function, and outputs a plurality of prediction results of the flow target, and the prediction value represents the probability of the flow response of the corresponding target; The weight branch includes a fully connected layer and a Softmax normalization structure, and outputs a dynamic weight vector corresponding to a plurality of flow targets, and the weight value represents the importance of each target in the final regulation strategy.

7. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 2, wherein The plurality of flow target prediction results and the corresponding dynamic weight vector are input into a multi-target fusion scoring function, and a comprehensive flow score value is calculated, specifically including: The plurality of optimization target task prediction results and the corresponding dynamic weight vector are normalized, the dynamic weight vector is standardized by using a Softmax function, so that the weight sum of the five types of optimization target tasks is 1, and a dynamic weight distribution vector is formed; Based on the plurality of optimization target task prediction results, independent evaluation indexes of short video content in click rate prediction, complete playback rate prediction, like rate prediction, comment rate prediction and conversion rate prediction are calculated respectively, and a five-task evaluation result set is constructed; According to the dynamic weight distribution vector and the five-task evaluation result set, a weighted summary method is used to calculate the comprehensive flow score value of each short video content.

8. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 2, wherein According to the interval in which the comprehensive flow score value is located, a preset score interval mapping rule is queried to obtain a structured flow parameter set, specifically including: The comprehensive flow score value of each short video content is compared with a score interval set, and the score interval identifier to which it belongs is matched, the score interval set is a plurality of non-overlapping interval segments divided according to the score value, and the interval boundary value is preset by the system; According to the score interval identifier, the corresponding structured flow parameter set is retrieved in the score interval mapping rule table, and the score interval mapping rule table is constructed by historical flow performance data, and each record in the rule table includes a one-to-one correspondence between the score interval identifier and the structured flow parameter set.

9. The short video streaming precision regulation system based on AI dynamic weight analysis of claim 2, wherein The user feedback data after the accurate recommendation is collected, and the parameters of the improved PLE network model are updated and improved through a feedback calibration mechanism according to the deviation between the user feedback data and the flow target prediction result, specifically including: Based on the collected strategy execution feedback information and the corresponding structured flow parameter set, an optimization sample of the score interval mapping rule is constructed, the optimization sample is indexed by the comprehensive flow score value, and an associated mapping relationship between the flow parameter configuration and the strategy execution result is established; Statistical analysis is performed on the historical optimization sample in each score interval, and the mapping efficiency score of each field parameter combination and the core index performance in the interval is calculated. According to the mapping performance score, a sliding window strategy and an effect dominant weight mechanism are used to dynamically update the parameter configuration in the current score interval mapping rule table, and the update result covers the original parameter set; When the optimization sample quantity is lower than the preset threshold, the rule update of the score interval is suspended and a to-be-verified state is marked.

Images