Algorithm information visualization interaction method and system

Abstract

The invention discloses an algorithm information visualization interaction method and system, and relates to the technical field of information visualization, and the method comprises the steps: constructing a gradient lifting decision tree model, calculating the feature contribution value of each dimension feature, and generating a weight reference value; calculating a real-time weight ratio of each dimension feature; updating feature weight distribution according to video content parameters and spatio-temporal feature parameters adjusted by the user, calculating a recommended weight gain and a comprehensive profit value in combination with a feature contribution value and a weight reference value, and generating a real-time flow gain curve; and deploying the parameter combination with the highest comprehensive income value to a short video platform, monitoring the actual income deviation through a real-time feedback mechanism, carrying out A / B test on an experimental group and a control group, updating model parameters according to a test result, triggering a parameter rollback mechanism or storing effective configuration, and carrying out real-time analysis. The problem of algorithm black box, the problem of low manual trial and error efficiency and the defect of parameter feedback delay caused by non-transparent feature weight of an existing recommendation system are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of information visualization, and particularly to an algorithm information visualization interaction method and system. Background Art

[0002] Today, with the increasing prosperity of short video platforms, algorithm recommendation systems have become a key technology to improve user experience and platform activity. Most existing recommendation algorithms rely on historical data and user behavior analysis, and build complex models to predict users' interest preferences, and recommend relevant video content accordingly. These algorithms usually include various methods such as content-based recommendation, collaborative filtering recommendation, and hybrid recommendation, aiming to improve the accuracy and personalization of recommendations.

[0003] However, most existing recommendation systems are closed black-box models, lacking a visualization presentation and interactive adjustment mechanism for feature weights. It is difficult for platform operators to intuitively understand the contribution of each dimension feature (such as user behavior, video content, spatio-temporal factors) to the recommendation result, and they are even less able to dynamically adjust parameter configurations according to business requirements, resulting in algorithm optimization relying on offline iteration and manual experience trial and error, with low efficiency and insufficient flexibility.

[0004] In addition, the feedback and verification mechanism of existing systems has a delay defect. After parameter adjustment, it is necessary to evaluate the effect through large-scale offline testing or long-term data accumulation, and it is difficult to monitor the actual revenue deviation of the recommendation strategy in real time. There is a lack of an efficient A / B testing framework and an automated rollback mechanism, resulting in ineffective configurations being difficult to identify and correct in a timely manner, which may cause fluctuations in recommendation effects and even a decline in user experience. Summary of the Invention

[0005] (1) Technical Problems to be Solved Aiming at the deficiencies of the existing technology, the present invention provides an algorithm information visualization interaction method and system. By real-time obtaining multi-dimensional feature data of a short video platform and constructing a gradient boosting decision tree model to quantify the feature contribution value, combining the particle swarm optimization algorithm to dynamically adjust the feature weight ratio, generating a three-dimensional visualization map to realize the transparent display of weights, and at the same time introducing a backpropagation neural network and a long short-term memory network to predict the recommendation weights and distributions, the problems of algorithm black-box caused by opaque feature weights, low efficiency of manual trial and error, and parameter feedback delay defects in existing recommendation systems are solved.

[0006] (2) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An algorithm information visualization interaction method, including: Real-time obtaining video data of a short video platform, constructing a gradient boosting decision tree model, calculating the feature contribution value of each dimension feature, and generating a weight reference value in combination with the historical weight data of each dimension feature; Dynamically calculate the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, construct a three-dimensional visualization map, use the backpropagation neural network to predict the recommendation weight, and use the long short-term memory network to predict the weight distribution of each dimension feature; According to the video content parameters and spatio-temporal feature parameters adjusted by the user, update the feature weight distribution, calculate the recommendation weight gain and the comprehensive benefit value by combining the feature contribution value and the weight benchmark value, and generate a real-time traffic gain curve; Deploy the parameter combination with the highest comprehensive benefit value to the short video platform, monitor the actual benefit deviation through the real-time feedback mechanism, conduct A / B testing on the experimental group and the control group, and update the model parameters according to the test results, trigger the parameter rollback mechanism or store the effective configuration.

[0007] Further, the specific steps of obtaining the video data of the short video platform in real time and constructing the gradient boosting decision tree model include: Real-time capture the original data stream through the open API interface of the short video platform, including user behavior data, video content data and spatio-temporal data; construct a multi-dimensional feature vector matrix, including user behavior features, video content features and spatio-temporal features; use the XGBoost framework to construct a gradient boosting decision tree model, and train the gradient boosting decision tree model based on the historical data set; use the TreeSHAP algorithm to calculate the feature contribution value of each dimension feature, and generate a feature sensitivity heat map; generate the weight benchmark value of each dimension feature based on the moving average method and the linear regression model.

[0008] Further, the steps of dynamically calculating the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm include: Initialize the particle swarm parameters, set the number of particles, the number of iterations and the inertia weight, and map the particle position to the real-time weight ratio of the feature; define the objective function as maximizing the recommendation weight; update the particle velocity and position according to the particle swarm optimization algorithm formula, and map the particle swarm optimization result to a three-dimensional visualization map, with the X-axis as the feature dimension, the Y-axis as the weight value, and the Z-axis as the time series.

[0009] Further, the steps of predicting the recommendation weight by the backpropagation neural network include: Construct the network structure of the input layer, hidden layer and output layer, and use the ReLU activation function for the hidden layer; calculate the gradient of the input feature to the output through backpropagation, and normalize the absolute value of the gradient to the weight ratio.

[0010] Further, the steps of predicting the weight distribution by using the long short-term memory network include: Convert the historical spatio-temporal weight sequence into a time window input in the supervised learning format; construct a network structure including an LSTM layer and a fully connected layer to predict the weight prediction value in the future time period.

[0011] Further, the steps of calculating the recommended weight gain and the comprehensive income value include: Dynamically update the feature weight distribution according to the user-adjusted parameters, and calculate the recommended weight gain in combination with the feature contribution value: , where Rw is the recommended weight gain, is the weight of the d th feature, t is the time, is the d th feature contribution value, D is the total number of features; generate a real-time traffic gain curve, with the horizontal axis being time and the vertical axis being the recommended weight gain value.

[0012] Further, calculate the comprehensive income value based on the deviation degree between the weight benchmark value and the adjusted weight: , where is the weight of the d th feature under the new parameter combination, is the weight benchmark value of the d th feature, and Zw is the comprehensive income value.

[0013] Further, the steps of deploying the parameter combination and A / B testing include: Synchronize the parameter combination with the highest comprehensive income value to the short video platform through the API interface to overwrite the default configuration; Real-time monitor the actual income deviation value, and trigger the parameter rollback mechanism if it exceeds the deviation threshold; Divide the experimental group and the control group based on the user ID hash value, and count the exposure conversion rate, user retention rate, and interaction behavior density; Use the two-sided T-test to verify the significance of the index difference. If it is effective, store the configuration and retrain the model.

[0014] Further, the triggering conditions of the parameter rollback mechanism are: The actual income deviation value exceeds the preset deviation threshold continuously for several times; The key indicators of the experimental group fail to pass the significance test or the improvement amplitude is lower than the baseline; Automatically freeze the user permissions and restore to the historical effective configuration.

[0015] An algorithm information visualization interaction system includes: A model construction module that obtains the video data of the short video platform in real time, constructs a gradient boosting decision tree model, calculates the feature contribution values of each dimension feature, and generates a weight benchmark value in combination with the historical weight data of each dimension feature; The weight allocation module dynamically calculates the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, constructs a three-dimensional visualization map, predicts the recommended weight using a backpropagation neural network, and predicts the weight distribution of each dimension feature using a long short-term memory network; The revenue evaluation module updates the feature weight distribution according to the video content parameters and spatio-temporal feature parameters adjusted by the user, calculates the recommended weight gain and the comprehensive revenue value by combining the feature contribution value and the weight benchmark value, and generates a real-time traffic gain curve; The parameter deployment module deploys the parameter combination with the highest comprehensive revenue value to the short video platform, monitors the actual revenue deviation through a real-time feedback mechanism, conducts A / B tests on the experimental group and the control group, and updates the model parameters according to the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0016] (3) Beneficial effects The present invention provides an algorithm information visualization interaction method and system, which has the following beneficial effects: (1) By obtaining short video platform data in real time and constructing a gradient boosting decision tree model, effectively calculating the feature contribution value of each dimension, and generating a benchmark value in combination with historical weights, it provides data support for algorithm optimization, enhances the algorithm's understanding of feature importance, helps to accurately predict the recommended weight, and lays a foundation for subsequent dynamic adjustment of feature weights and improvement of recommendation effects.

[0017] (2) By dynamically calculating the feature weight ratio through the particle swarm optimization algorithm, constructing a three-dimensional visualization map, and using a backpropagation neural network and a long short-term memory network to predict the recommended weight and weight distribution, it realizes the real-time dynamic adjustment of feature weights, improves the flexibility and accuracy of the recommendation algorithm, and at the same time the visualization map enhances the algorithm transparency and interactivity.

[0018] (3) By dynamically updating the feature weight distribution through user-adjusted parameters, calculating the recommended weight gain and the comprehensive revenue value by combining the feature contribution value and the weight benchmark value, and generating a real-time traffic gain curve, it realizes the user's active intervention in the algorithm, can quickly respond to market changes, improve the recommendation effect, and intuitively display the adjustment effect through the traffic gain curve to assist decision-making.

[0019] (4) By deploying the parameter combination with the highest comprehensive revenue value to the short video platform, combining the real-time feedback mechanism and A / B tests, monitoring the actual revenue deviation and verifying the effectiveness of the parameters, it ensures the sustainability and accuracy of algorithm optimization, can roll back invalid parameters in time, store the effective configuration, and improve the overall recommendation effect and user experience of the platform. Description of the drawings

[0020] Figure 1 It is a schematic diagram of the steps of the algorithm information visualization interaction method of the present invention; Figure 2 This is a schematic structural diagram of the algorithm information visualization interaction system of the present invention. Specific embodiments

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to Figure 1 , the present invention provides an algorithm information visualization interaction method, including the following steps: Step 1: Real-time obtain video data of a short video platform, construct a gradient boosting decision tree model, calculate the feature contribution values of each dimension feature, and generate a weight reference value in combination with the historical weight data of each dimension feature; The said Step 1 includes the following contents: Step 101: Through the open API interface of the short video platform, real-time capture the original data stream of the video content, including: user behavior data: viewing duration, interaction behaviors (liking, commenting, sharing), user portraits (age, region, interest tags); video content data: video content, video duration, tag density, keyword coverage rate, video picture (such as color distribution); spatio-temporal data: upload time, geographical location, holiday identifier (such as weekend, promotion day); Step 102: Use the Z-score method to detect and remove abnormal data beyond the 3σ range (such as records with a single viewing duration exceeding 200% of the total video duration), and for the missing user portrait data, use the KNN (K-Nearest Neighbor) algorithm to interpolate based on the behavior characteristics of similar users; Step 103: Construct a multi-dimensional feature vector matrix, including user behavior features, video content features, and spatio-temporal features. User behavior features include completion rate, interaction rate (liking rate + commenting rate + sharing rate), user activity (historical behavior frequency); video content features include tag density (the frequency of tag appearance per unit time), keyword coverage rate (the proportion of the number of keyword appearances in the video to the total number of words), picture complexity (calculating the picture saturation based on the HSV color space), the proportion of the duration of close-up shots (the proportion of the first three seconds of close-up shots, such as the proportion of agricultural product close-ups being 80%), the insertion frequency of interaction elements (the number of interaction elements such as bullet screens and votes inserted per minute, such as inserting 1 time per minute); spatio-temporal features include upload period, geographical location, holiday identifier (Boolean value); Min-Max normalization is performed on continuous features (such as viewing duration and interaction rate), and the range is scaled to [0, 1]. One-Hot encoding is performed on discrete features (such as holiday identification and geographical location) to generate a dummy variable matrix. The constructed multi-dimensional feature vector matrix is , where N is the number of videos, D is the total number of features, R is the set of real numbers; Step 104: Use the XGBoost framework to construct a Gradient Boosting Decision Tree (GBDT) model. Set the model parameters, including the learning rate (such as 0.1), the maximum depth (such as 5), and the number of trees (such as 100). The objective function is the Mean Squared Error (MSE). Take the multi-dimensional feature vector matrix as the input, and use the actual recommendation weights of the short video platform (such as video exposure and click-through rate) as the target variable. Train the GBDT model based on the historical dataset, that is, the recommendation weight prediction model; Step 105: Use the TreeSHAP algorithm to calculate the feature contribution values (SHAP values) of each dimension feature. The feature contribution value reflects the contribution degree of the feature to the algorithm recommendation weight. Generate a feature sensitivity heatmap based on the feature contribution values. The horizontal axis is the feature name, the vertical axis is the size of the feature contribution value, and the color depth represents the sensitivity level. The color depth of the heatmap represents the size of the influence of different features on the recommendation weight. The darker the color, the greater the influence of the feature. For example, the feature contribution value of the label density is 0.42, indicating that its influence on the recommendation weight is significantly higher than that of the completion rate (feature contribution value = 0.18); Step 106: Based on the historical weight data of each dimension feature (such as the past 30 days), use the moving average method (such as window size = 7) to generate a historical weight change curve. Align the feature contribution value with the historical weight curve according to the timestamp. Model the relationship between the feature contribution value and the historical weight through a linear regression model as: , where represents the weight benchmark value, Shap represents the feature contribution value, is the intercept term, is the regression coefficient, , , and are generated by fitting using the least squares method; When using, combine the content of steps 101 to 106: By obtaining short video platform data in real time and constructing a gradient boosting decision tree model, the contribution values of features in each dimension are effectively calculated, and a benchmark value is generated in combination with historical weights, providing data support for algorithm optimization, enhancing the algorithm's understanding of feature importance, helping to accurately predict recommendation weights, and laying a foundation for subsequent dynamic adjustment of feature weights and improvement of recommendation effects.

[0023] Step 2: Dynamically calculate the real-time weight ratios of features in each dimension through the particle swarm optimization algorithm, construct a three-dimensional visualization map, predict recommendation weights using a backpropagation neural network, and predict the weight distribution of features in each dimension using a long short-term memory network; The above Step 2 includes the following content: Step 201: Based on the multi-dimensional feature vector matrix (including user behavior features, video content features, and spatio-temporal features) and the feature sensitivity heat map generated in Step 1, extract the real-time weight benchmark values of each dimension, dynamically calculate the real-time weight ratios of features in each dimension through the particle swarm optimization algorithm (PSO), and construct a three-dimensional visualization map, where the weight value of each particle is mapped to a point in the three-dimensional coordinate system (the X-axis is the feature dimension, the Y-axis is the weight value, and the Z-axis is the time series); Dynamically calculating the real-time weight ratios of features in each dimension through the particle swarm optimization algorithm specifically includes: Initialize the particle swarm parameters, set the number of particles (such as 50), map each particle to a feature dimension (such as "tag density", "completion rate", "interaction rate"), the particle position represents the real-time weight ratio of the feature, the velocity represents the weight change trend, set the number of iterations (such as 100) and the inertia weight (such as 0.8), the inertia weight adopts a linear decay strategy (such as decreasing from 0.8 to 0.4, with 100 iterations), limit the particle velocity range to [-0.1, 0.1], and the objective function is to maximize the recommendation weight: , where, is the predicted recommendation weight value of the n th sample, is the historical recommendation weight value of the n th sample, n represents the n th sample (such as a short video), N represents the number of samples; Update the particle velocity and position according to the particle swarm optimization algorithm formula, and calculate the weight distribution of features in each dimension. For example, the weight of "tag density" is adjusted from the initial value of 0.25 to 0.38, and the weight of "completion rate" is adjusted from 0.18 to 0.12. The formula: , , where, represents the i th particle at the k th iteration and the dThe velocity in dimension represents the i -th particle's k -th dimensional position at the d -th iteration, represents the individual optimal position of the i -th particle at the k -th iteration, represents the global optimal position of swarm g at the k -th iteration, w represents the inertia weight, represents the cognitive learning factor, represents the social learning factor, , , and are random numbers between [0, 1], used to introduce randomness and prevent particles from falling into a fixed pattern; Step 202: Construct a backpropagation neural network (BPNN) model, including: Input layer: a multi-dimensional feature vector matrix ( nodes); Hidden layer: adopting a two-layer fully connected structure, with the number of neurons in each layer being , and the activation function being ReLU; Output layer: predicting the recommended weight (single node), training with historical data, the loss function being the mean squared error, calculating the gradient of the input features with respect to the output through backpropagation, the greater the absolute value of the gradient, the more significant the influence of the feature on the recommended weight, and normalizing the absolute value of the gradient to the weight ratio: , where represents the weight ratio of the d -th feature, D represents the total number of features, represents the gradient value of the d -th feature; Step 203: Obtain the historical spatio-temporal weight sequence (such as the weight change of "completion rate" in the past 7 days) and convert it into a supervised learning format: , , is the input feature vector, representing the historical weight values within a time window, represents the time step T 's weight value, l is the length of the time window; Using a long short-term memory network, including 2 LSTM layers and 1 fully connected layer, with the number of neurons in each layer being , and the activation function being tanh, predicting the weight prediction value for a future period (such as 24 hours), and the loss function being the root mean squared error; When in use, combine the content of Steps 201 to 203: The proportion of feature weights is dynamically calculated through the particle swarm optimization algorithm, a three-dimensional visualization map is constructed, and the backpropagation neural network and long short-term memory network are used to predict the recommendation weights and weight distributions, realizing the real-time dynamic adjustment of feature weights, improving the flexibility and accuracy of the recommendation algorithm. At the same time, the visualization map enhances the transparency and interactivity of the algorithm.

[0024] Step 3: According to the video content parameters and spatio-temporal feature parameters adjusted by the user, update the feature weight distribution, calculate the recommendation weight gain and the comprehensive benefit value by combining the feature contribution value and the weight benchmark value, and generate a real-time traffic gain curve. The above Step 3 includes the following contents: Step 301: Perform parameter adjustment based on the three-dimensional visualization map. The adjustable parameters include video content parameters and spatio-temporal feature parameters. Among them, the video content parameters include tag density, keyword coverage rate, picture complexity, the proportion of close-up shot duration, and the insertion frequency of interactive elements. The spatio-temporal feature parameters include the upload time period and geographical location. Step 302: When the user adjusts the parameters (such as increasing the proportion of the "agricultural product close-up" duration from 50% to 80%), use the particle swarm optimization algorithm formula in Step 2 to update the particle velocity and position, and calculate the weight distribution of each dimension feature under the new parameter combination (such as the weight of the "proportion of close-up shot duration" increases from 15% to 19.7%). Step 303: Based on the feature contribution values calculated by the TreeSHAP algorithm in Step 1 (such as the contribution value of "tag density" is 0.42 and the contribution value of "completion rate" is 0.18), combined with the weight distribution dynamically optimized by the PSO algorithm in Step 2 (such as the weight of "tag density" increases from 0.25 to 0.38), calculate the recommendation weight gain. The formula is: , where Rw represents the recommendation weight gain, represents the weight of the d th feature, t is the time, represents the d th feature contribution value, D represents the total number of features; when the user adjusts the parameters, the feature weight values are updated in real time, and the recommendation weight gain is recalculated to generate a traffic gain curve, with the horizontal axis being time and the vertical axis being the recommendation weight gain value (the interval is [0,1]); Step 304: The user can adjust multiple parameters simultaneously (such as reducing the "tag density" to 50% and increasing the "proportion of close-up shot duration" to 80%), obtain the weight benchmark values of each dimension feature in Step 1, and calculate the comprehensive benefit value by combining the weight distribution of each dimension feature under the new parameter combination: , where represents the weight of the dth feature under the new parameter combination, represents the weight benchmark value of the d-th feature, and Zw represents the comprehensive income value; When in use, combine the content of steps 301 to 304: By dynamically updating the feature weight distribution through user parameter adjustment, calculating the recommended weight gain and comprehensive income value by combining the feature contribution value and the weight benchmark value, generating a real-time traffic gain curve, it realizes the active intervention of users in the algorithm, can quickly respond to market changes, improve the recommendation effect, and visually display the adjustment effect through the traffic gain curve to assist decision-making.

[0025] Step Four: Deploy the parameter combination with the highest comprehensive income value to the short video platform, monitor the actual income deviation through a real-time feedback mechanism, conduct A / B testing on the experimental group and the control group, and update the model parameters according to the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0026] The said Step Four includes the following content: Step 401: Generate the comprehensive income value based on Step Three, and deploy the parameter combination with the highest comprehensive income value after multiple adjustments to the short video platform through the real-time feedback mechanism, specifically including: Synchronize the parameters adjusted by the user (such as "tag density", "proportion of close-up shot duration") to the platform through the API interface to overwrite the original default parameter configuration; After the recommended weight is updated, every time window Δ t (such as 30 minutes) extract the video exposure volume, click-through rate, and user stay duration, and calculate the actual income deviation value : where, Pw is the predicted recommended weight, Aw is the actual recommended weight. If the actual income deviation value exceeds the deviation threshold, trigger the parameter rollback mechanism to restore to the previous parameter configuration state; Step 402: Divide the traffic into an experimental group (applying new parameters) and a control group (retaining the original parameters) according to the user ID hash value, and count the key indicator differences during the experimental period (such as 24 hours), including: exposure conversion rate, user retention rate, interaction behavior density. Conduct a two-sided t-test (significance level α = 0.05) on the data of the experimental group and the control group. If the p-value < α and the indicators of the experimental group increase by more than the baseline (such as click-through rate + 5%), it is determined that the parameter adjustment is effective; Step 403: Store the effective parameter configuration and results of the experimental group in the MongoDB database, mark the spatio-temporal context (such as "promotion day configuration", "regional differentiation strategy"), and based on the newly generated behavioral data, retrain the GBDT model in Step 1 and the LSTM network in Step 2, update the feature contribution value and weight prediction value. If the actual revenue deviation value exceeds the deviation threshold for multiple consecutive times (such as 3 times), automatically freeze the user permissions and trigger the parameter rollback mechanism; When in use, combine the content of Steps 401 to 403: By deploying the parameter combination with the highest comprehensive revenue value to the short video platform, combining the real-time feedback mechanism and A / B testing, monitoring the actual revenue deviation and verifying the parameter effectiveness, the continuity and accuracy of the algorithm optimization are ensured, invalid parameters can be rolled back in a timely manner, effective configurations can be stored, and the overall recommendation effect and user experience of the platform can be improved.

[0027] Please refer to Figure 2 , the present invention also provides an algorithm information visualization interaction system, including: a model construction module, a weight allocation module, a revenue evaluation module, and a parameter deployment module; Among them, the model construction module obtains the video data of the short video platform in real time, constructs a gradient boosting decision tree model, calculates the feature contribution value of each dimension feature, and generates a weight benchmark value in combination with the historical weight data of each dimension feature; The weight allocation module dynamically calculates the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, constructs a three-dimensional visualization map, predicts the recommendation weight using the backpropagation neural network, and predicts the weight distribution of each dimension feature using the long short-term memory network; The revenue evaluation module updates the feature weight distribution according to the video content parameters and spatio-temporal feature parameters adjusted by the user, calculates the recommendation weight gain and comprehensive revenue value in combination with the feature contribution value and the weight benchmark value, and generates a real-time traffic gain curve; The parameter deployment module deploys the parameter combination with the highest comprehensive revenue value to the short video platform, monitors the actual revenue deviation through the real-time feedback mechanism, conducts A / B testing on the experimental group and the control group, and updates the model parameters according to the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0028] In the application, several formulas involved are calculated by taking their numerical values after dimensionless, and the formula is a formula obtained by collecting a large amount of data for software simulation to approximate the real situation as much as possible. The coefficients in the formula are set by those skilled in the art according to the actual situation.

[0029] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those of ordinary skill in the art will appreciate that the units and algorithm steps of the examples described in connection with the embodiments disclosed herein can be implemented in a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution.

[0030] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units. They may be located in one place or distributed over multiple network units. Some or all of the units can be selected according to actual needs to achieve the objectives of the solution of this embodiment.

[0031] As described above, the specific implementation manners of the present application are only described, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. An algorithm information visualization interaction method, characterized in that: Including: Real-time obtain video data of the short video platform, construct a gradient boosting decision tree model, calculate the feature contribution values of each dimension feature, and generate a weight benchmark value in combination with the historical weight data of each dimension feature; Dynamically calculate the real-time weight proportion of each dimension feature through the particle swarm optimization algorithm, construct a three-dimensional visualization map, use the backpropagation neural network to predict the recommended weight, and use the long short-term memory network to predict the weight distribution of each dimension feature; According to the video content parameters and spatio-temporal feature parameters adjusted by the user, update the feature weight distribution, calculate the recommended weight gain and comprehensive income value in combination with the feature contribution value and the weight benchmark value, and generate a real-time traffic gain curve; Deploy the parameter combination with the highest comprehensive income value to the short video platform, monitor the actual income deviation through the real-time feedback mechanism, conduct A / B tests on the experimental group and the control group, and update the model parameters according to the test results, trigger the parameter rollback mechanism or store the effective configuration.

2. The algorithm information visualization and interaction method according to claim 1, characterized in that: The specific steps of the real-time obtaining video data of the short video platform and constructing a gradient boosting decision tree model include: Real-time capture the original data stream through the open API interface of the short video platform, including user behavior data, video content data and spatio-temporal data; construct a multi-dimensional feature vector matrix, including user behavior features, video content features and spatio-temporal features; use the XGBoost framework to construct a gradient boosting decision tree model, and train the gradient boosting decision tree model based on the historical data set; use the TreeSHAP algorithm to calculate the feature contribution values of each dimension feature, and generate a feature sensitivity heat map; generate the weight benchmark value of each dimension feature based on the moving average method and the linear regression model.

3. The algorithm information visualization interaction method according to claim 1, wherein: The steps of dynamically calculating the real-time weight proportion of each dimension feature through the particle swarm optimization algorithm include: Initialize the particle swarm parameters, set the number of particles, the number of iterations and the inertia weight, and map the particle position to the real-time weight proportion of the feature; define the objective function as maximizing the recommended weight; update the particle velocity and position according to the particle swarm optimization algorithm formula, and map the particle swarm optimization result to a three-dimensional visualization map, with the X-axis as the feature dimension, the Y-axis as the weight value, and the Z-axis as the time series.

4. An algorithm information visualization interaction method according to claim 3, characterized in that: The steps of the backpropagation neural network predicting the recommended weight include: Construct a network structure of the input layer, hidden layer and output layer, and use the ReLU activation function for the hidden layer; calculate the gradient of the input feature to the output through backpropagation, and normalize the absolute value of the gradient to the weight proportion.

5. The algorithm information visualization interaction method according to claim 4, wherein: The steps of using the long short-term memory network to predict the weight distribution include: Convert the historical spatio-temporal weight sequence into a time window input in the supervised learning format; construct a network structure including an LSTM layer and a fully connected layer, and predict the weight prediction value in the future time period.

6. The algorithm information visualization interaction method according to claim 1, wherein: The steps of calculating the recommended weight gain and comprehensive income value include: Dynamically update the feature weight distribution according to the user-adjusted parameters, and calculate the recommended weight gain in combination with the feature contribution value: , where Rw is the recommended weight gain, is the weight of the d th feature, t is the time, is the d th feature contribution value, D is the total number of features; Generate a real-time traffic gain curve, with the horizontal axis being time and the vertical axis being the recommended weight gain value.

7. An algorithm information visualization interaction method according to claim 6, characterized in that: Calculate the comprehensive income value based on the deviation degree between the weight benchmark value and the adjusted weight: , where is the weight of the d th feature under the new parameter combination, is the weight benchmark value of the d th feature, and Zw is the comprehensive income value.

8. An algorithm information visualization interaction method according to claim 1, characterized in that: The steps of deploying the parameter combination and A / B testing include: Synchronize the parameter combination with the highest comprehensive income value to the short video platform through the API interface to overwrite the default configuration; Monitor the actual income deviation value in real time. If it exceeds the deviation threshold, trigger the parameter rollback mechanism; Divide the experimental group and the control group based on the user ID hash value, and count the exposure conversion rate, user retention rate, and interaction behavior density; Use the two-sided T-test to verify the significance of the index difference. If it is effective, store the configuration and retrain the model.

9. The algorithm information visualization interaction method according to claim 8, characterized in that: The trigger condition of the parameter rollback mechanism is: The actual income deviation value exceeds the preset deviation threshold continuously for several times; The key indicators of the experimental group fail the significance test or the improvement amplitude is lower than the baseline; Automatically freeze the user permissions and restore them to the historical valid configuration.

10. An algorithm information visualization and interaction system for implementing the method according to any one of claims 1 to 9, characterized in that: Include: The model construction module obtains the video data of the short video platform in real time, constructs a gradient boosting decision tree model, calculates the feature contribution value of each dimension feature, and generates a weight reference value in combination with the historical weight data of each dimension feature; The weight allocation module dynamically calculates the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, constructs a three-dimensional visualization map, predicts the recommended weight using the backpropagation neural network, and predicts the weight distribution of each dimension feature using the long short-term memory network; The income evaluation module updates the feature weight distribution according to the video content parameters and spatio-temporal feature parameters adjusted by the user, calculates the recommended weight gain and comprehensive income value in combination with the feature contribution value and the weight reference value, and generates a real-time traffic gain curve; The parameter deployment module deploys the parameter combination with the highest comprehensive income value to the short video platform, monitors the actual income deviation through the real-time feedback mechanism, conducts the A / B test of the experimental group and the control group, and updates the model parameters according to the test results, triggers the parameter rollback mechanism or stores the valid configuration.

Images