Algorithm information visualization interaction method and system

Abstract

The present invention discloses an algorithm information visualization interaction method and system, which relates to the field of information visualization technology, including: constructing a gradient boosting decision tree model, calculating the feature contribution value of each dimensional feature, and generating a weight reference value; calculating the real-time weight proportion of each dimensional feature; updating the feature weight distribution according to the video content parameters and spatiotemporal feature parameters adjusted by the user, calculating the recommendation weight gain and the comprehensive benefit value in combination with the feature contribution value and the weight reference value, and generating a real-time traffic gain curve; deploying the parameter combination with the highest comprehensive benefit value to the short video platform, monitoring the actual benefit deviation through a real-time feedback mechanism, conducting A / B tests on the experimental group and the control group, and updating the model parameters according to the test results, triggering the parameter rollback mechanism or storing the effective configuration, thereby solving the algorithm black box problem, the low efficiency of manual trial and error, and the parameter feedback delay defect caused by the opacity of feature weights in the existing recommendation system.

Description

Technical Field

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

[0002] With the growing popularity of short video platforms, algorithm recommendation systems have become a key technology for improving user experience and platform activity. Most existing recommendation algorithms rely on historical data and user behavior analysis, building complex models to predict user interests and preferences and recommend relevant video content accordingly. These algorithms typically include content-based recommendations, collaborative filtering recommendations, and hybrid recommendations, aiming to improve the accuracy and personalization of recommendations.

[0003] However, existing recommendation systems are mostly closed black box models that lack a visual presentation and interactive adjustment mechanism for feature weights. Platform operators find it difficult to intuitively understand the contribution of various dimensional features (such as user behavior, video content, and temporal and spatial factors) to recommendation results, and are unable to dynamically adjust parameter configurations according to business needs. As a result, algorithm optimization relies on offline iteration and manual trial and error, which is inefficient and lacks flexibility.

[0004] In addition, the feedback and verification mechanism of the existing system has delay defects. After the parameters are adjusted, large-scale offline testing or long-term data accumulation is required to evaluate the effect. It is difficult to monitor the actual benefit deviation of the recommendation strategy in real time. The lack of an efficient A / B testing framework and automated rollback mechanism makes it difficult to identify and correct invalid configurations 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 solved

[0006] In response to the shortcomings of the existing technology, the present invention provides an algorithm information visualization interaction method and system. By acquiring multi-dimensional feature data of the short video platform in real time and constructing a gradient boosting decision tree model to quantify the feature contribution value, the particle swarm optimization algorithm is combined to dynamically adjust the feature weight ratio, and a three-dimensional visualization map is generated to achieve transparent display of the weight. At the same time, a back propagation neural network and a long short-term memory network are introduced to predict the recommendation weight and distribution, which solves the algorithm black box problem, the low efficiency of manual trial and error, and the parameter feedback delay defect caused by the opaque feature weights in the existing recommendation system.

[0007] (2) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for visualizing and interacting with algorithm information, comprising:

[0009] Acquire video data from short video platforms in real time, build a gradient boosting decision tree model, calculate the feature contribution value of each dimension feature, and generate a weight benchmark value based on the historical weight data of each dimension feature;

[0010] The particle swarm optimization algorithm is used to dynamically calculate the real-time weight ratio of each dimension feature, construct a three-dimensional visualization map, use the back propagation neural network to predict the recommendation weight, and use the long short-term memory network to predict the weight distribution of each dimension feature;

[0011] Update the feature weight distribution based on the video content parameters and spatiotemporal feature parameters adjusted by the user, calculate the recommendation weight gain and comprehensive benefit value by combining the feature contribution value and the weight benchmark value, and generate a real-time traffic gain curve;

[0012] Deploy the parameter combination with the highest comprehensive benefit value to the short video platform, monitor the actual benefit deviation through a real-time feedback mechanism, conduct A / B testing between the experimental group and the control group, and update the model parameters based on the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0013] Furthermore, the real-time acquisition of video data from the short video platform and the construction of a gradient boosting decision tree model specifically include:

[0014] Through the open API interface of the short video platform, the original data stream is captured in real time, including user behavior data, video content data and time-space data; a multi-dimensional feature vector matrix is ​​constructed, including user behavior features, video content features and time-space features; the XGBoost framework is used to build a gradient boosting decision tree model, and the gradient boosting decision tree model is trained based on historical data sets; the TreeSHAP algorithm is used to calculate the feature contribution value of each dimensional feature and generate a feature sensitivity heat map; the weight baseline value of each dimensional feature is generated based on the sliding average method and linear regression model.

[0015] Furthermore, the step of dynamically calculating the real-time weight proportion of each dimensional feature by the particle swarm optimization algorithm includes:

[0016] Initialize the particle swarm parameters, set the number of particles, number of iterations and 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 speed and position according to the particle swarm optimization algorithm formula, and map the particle swarm optimization results into 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.

[0017] Furthermore, the step of predicting the recommendation weight using the back propagation neural network includes:

[0018] Construct a network structure of input layer, hidden layer and output layer, and use ReLU activation function in the hidden layer; calculate the gradient of input feature to output through back propagation, and the absolute value of normalized gradient is the weight ratio.

[0019] Furthermore, the step of using the long short-term memory network to predict the weight distribution includes:

[0020] Convert the historical spatiotemporal weight sequence into a time window input in a supervised learning format; construct a network structure consisting of an LSTM layer and a fully connected layer to predict the weight prediction value for future time periods.

[0021] Furthermore, the step of calculating the recommendation weight gain and the comprehensive benefit value includes:

[0022] Dynamically update feature weight distribution based on user adjustment parameters, and calculate recommendation weight gain based on feature contribution value:

[0023] ,in, R is the recommendation weight gain, For the d The weight of the feature, t For the moment, For the d The feature contribution value of each feature, D is the total number of features; a real-time traffic gain curve is generated, with the horizontal axis being time and the vertical axis being the recommended weight gain value.

[0024] Furthermore, the comprehensive return value is calculated based on the deviation between the weight base value and the adjusted weight:

[0025] ,in, For the new parameter combination d The weight of the feature, For the d The weighted benchmark value of each feature, Zw is the comprehensive benefit value.

[0026] Furthermore, the steps of deploying parameter combinations and A / B testing include:

[0027] Synchronize the parameter combination with the highest comprehensive benefit value to the short video platform through the API interface, overwriting the default configuration;

[0028] Real-time monitoring of actual profit deviation values. If the deviation threshold is exceeded, the parameter rollback mechanism will be triggered.

[0029] Divide the user ID hash value into experimental and control groups, and calculate the exposure conversion rate, user retention rate, and interactive behavior density;

[0030] A two-sided T test was used to verify the significance of the difference in indicators. If valid, the configuration was saved and the model was retrained.

[0031] Furthermore, the triggering conditions of the parameter rollback mechanism are:

[0032] The actual return deviation value exceeds the preset deviation threshold several times in a row;

[0033] The key indicators of the experimental group failed the significance test or the improvement rate was lower than the baseline;

[0034] Automatically freeze user permissions and restore to historical valid configurations.

[0035] An algorithm information visualization interactive system, comprising:

[0036] The model building module obtains video data from short video platforms in real time, builds a gradient boosting decision tree model, calculates the feature contribution value of each dimension feature, and generates a weight benchmark value based on the historical weight data of each dimension feature;

[0037] The weight distribution module dynamically calculates the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, constructs a three-dimensional visualization map, uses the back propagation neural network to predict the recommendation weight, and uses the long short-term memory network to predict the weight distribution of each dimension feature;

[0038] The revenue evaluation module updates the feature weight distribution based on the video content parameters and spatiotemporal feature parameters adjusted by the user, calculates the recommendation weight gain and comprehensive revenue value by combining the feature contribution value with the weight reference value, and generates a real-time traffic gain curve;

[0039] The parameter deployment module deploys the parameter combination with the highest comprehensive benefit value to the short video platform, monitors the actual benefit deviation through a real-time feedback mechanism, conducts A / B testing between the experimental group and the control group, and updates the model parameters based on the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0040] (3) Beneficial effects

[0041] The present invention provides an algorithm information visualization interaction method and system, which has the following beneficial effects:

[0042] (1) By acquiring short video platform data in real time and building a gradient boosting decision tree model, the contribution value of each dimension feature is effectively calculated, and the benchmark value is generated by combining historical weights, providing data support for algorithm optimization, enhancing the algorithm's understanding of feature importance, and helping to accurately predict recommendation weights, laying the foundation for subsequent dynamic adjustment of feature weights and improving recommendation effects.

[0043] (2) The particle swarm optimization algorithm is used to dynamically calculate the feature weight ratio, construct a three-dimensional visualization map, and use the back propagation neural network and long short-term memory network to predict the recommendation weight and weight distribution, thus realizing the real-time dynamic adjustment of the feature weight, improving the flexibility and accuracy of the recommendation algorithm, and at the same time, the visualization map enhances the transparency and interactivity of the algorithm.

[0044] (3) By adjusting the parameters by the user, the feature weight distribution is dynamically updated, and the recommendation weight gain and comprehensive benefit value are calculated by combining the feature contribution value and the weight benchmark value, and a real-time traffic gain curve is generated. This enables the user to actively intervene 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.

[0045] (4) By deploying the parameter combination with the highest comprehensive benefit value to the short video platform, combining the real-time feedback mechanism with A / B testing, monitoring the actual benefit deviation and verifying the effectiveness of the parameters, the continuity and accuracy of the algorithm optimization are ensured, invalid parameters can be rolled back in time, valid configurations can be stored, and the overall recommendation effect and user experience of the platform are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the steps of the algorithm information visualization interaction method of the present invention;

[0047] Figure 2 This is a structural diagram of the algorithm information visualization interaction system of the present invention. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0049] See also Figure 1 The present invention provides an algorithm information visualization interaction method, comprising the following steps:

[0050] Step 1: Obtain video data from the short video platform in real time, build a gradient boosting decision tree model, calculate the feature contribution value of each dimension feature, and generate a weight benchmark value based on the historical weight data of each dimension feature;

[0051] The step 1 includes the following contents:

[0052] Step 101: Use the short video platform's open API to capture the raw data stream of video content in real time, including: user behavior data: viewing time, interactive behavior (likes, comments, shares), user profile (age, region, interest tags); video content data: video content, video length, tag density, keyword coverage, video image (such as color distribution); spatiotemporal data: upload time, geographic location, holiday identifiers (such as weekends, promotional days);

[0053] Step 102: Use the Z-score method to detect and eliminate abnormal data that exceeds the 3σ range (e.g., records where the single viewing duration exceeds 200% of the total video duration). For missing user profile data, use the KNN (K-nearest neighbor) algorithm to interpolate based on the behavioral characteristics of similar users.

[0054] Step 103: Construct a multidimensional feature vector matrix, including user behavior features, video content features, and spatiotemporal features. User behavior features include completion rate, interaction rate (like rate + comment rate + share rate), and user activity (historical behavior frequency); video content features include tag density (frequency of tag appearance per unit time), keyword coverage (ratio of keyword appearances in the video to the total number of words), image complexity (image saturation calculated based on the HSV color space), close-up duration ratio (percentage of close-ups in the first three seconds, such as 80% for agricultural product close-ups), and interactive element insertion frequency (number of interactive elements such as bullet comments and voting inserted per minute, such as once per minute); spatiotemporal features include upload time, geographic location, and holiday flag (Boolean value);

[0055] Min-Max normalization is performed on continuous features (such as viewing time and interaction rate), and the range is scaled to [0,1]. One-Hot encoding is performed on discrete features (such as holiday identifiers and geographic locations) to generate a dummy variable matrix. The constructed multidimensional feature vector matrix is ,in, N is the number of videos, D is the total number of features, R is the set of real numbers;

[0056] Step 104: Use the XGBoost framework to build a gradient boosted 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 square error (MSE), and convert the multidimensional feature vector matrix As input, the actual recommendation weight of the short video platform (such as video exposure and click-through rate) is used as the target variable, and the GBDT model, i.e., the recommendation weight prediction model, is trained based on the historical dataset;

[0057] Step 105: Use the TreeSHAP algorithm to calculate the feature contribution value (SHAP value) of each dimension feature. The feature contribution value reflects the contribution of the feature to the algorithm's recommendation weight. Generate a feature sensitivity heat map based on the feature contribution value. The horizontal axis is the feature name, and the vertical axis is the feature contribution value. The color depth indicates the sensitivity level. The color depth of the heat map represents 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 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).

[0058] Step 106: Based on the historical weight data of each dimension feature (e.g., the past 30 days), use the sliding average method (e.g., window size = 7) to generate a historical weight change curve. Align the feature contribution value with the historical weight curve by timestamp, and use a linear regression model to model the relationship between the feature contribution value and the historical weight as follows: ,in, Indicates the weight reference value, Shap represents the feature contribution value, is the intercept term, is the regression coefficient, , , and Generated using least squares fitting;

[0059] When using, combine the contents of steps 101 to 106:

[0060] By acquiring short video platform data in real time and building a gradient boosting decision tree model, we can effectively calculate the contribution value of features in each dimension, and generate benchmark values ​​based on historical weights. This provides data support for algorithm optimization, enhances the algorithm's understanding of feature importance, and helps to accurately predict recommendation weights, laying the foundation for subsequent dynamic adjustment of feature weights and improving recommendation effects.

[0061] Step 2: 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 back propagation neural network to predict the recommendation weight, and use the long short-term memory network to predict the weight distribution of each dimension feature;

[0062] The second step includes the following contents:

[0063] Step 201: Based on the multi-dimensional feature vector matrix (including user behavior features, video content features, and spatiotemporal features) and feature sensitivity heat map generated in Step 1, extract the real-time weight benchmark value of each dimension, dynamically calculate the real-time weight ratio of each dimension feature 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 a 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);

[0064] The real-time weight ratio of each dimension feature is dynamically calculated through the particle swarm optimization algorithm, including:

[0065] Initialize the particle swarm parameters and set the number of particles (e.g., 50). Each particle is mapped to a feature dimension (e.g., "label density," "completion rate," or "interaction rate"). The particle position represents the real-time weight of the feature, and the speed represents the weight change trend. Set the number of iterations (e.g., 100) and the inertia weight (e.g., 0.8). The inertia weight adopts a linear attenuation strategy (e.g., decreasing from 0.8 to 0.4 with 100 iterations). Limit the particle speed range to [-0.1, 0.1]. The objective function is to maximize the recommendation weight: ,in, For the n The predicted recommendation weight value of samples, For the n The historical recommendation weight value of samples, n Indicates the n A sample (such as a short video), N represents the number of samples;

[0066] According to the particle swarm optimization algorithm formula, the particle speed and position are updated, and the weight distribution of each dimension feature is calculated. For example, the weight of "label 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 is: , ,in, Indicates the i The particle in k The first iteration d Dimensional speed, Indicates the i The particle in k The first iteration d Dimensional position, Indicates the i The particle in k The optimal position of the individual at the iteration, Indicates the k The global optimal position of group g at the iteration, w represents the inertia weight, represents the cognitive learning factor, represents the social learning factor, , , 、 is a random number between [0,1], used to introduce randomness to prevent particles from falling into a fixed pattern;

[0067] Step 202: Construct a back propagation neural network (BPNN) model, including: Input layer: multi-dimensional feature vector matrix ( nodes); Hidden layer: adopts a two-layer fully connected structure, with the number of neurons in each layer being , the activation function is ReLU; output layer: predict recommendation weight (single node), use historical data for training, the loss function is mean square error, and calculate the gradient of input features to output through back propagation. The larger the absolute value of the gradient, the more significant the influence of the feature on the recommendation weight. The absolute value of the gradient is normalized to the weight ratio: ,in, Indicates the d The weight ratio of the features, D represents the total number of features, Indicates the d The gradient value of the feature;

[0068] Step 203: Obtain the historical spatiotemporal weight sequence (e.g., the weight change of "completion rate" in the past 7 days) and convert it into supervised learning format: , , is the input feature vector, which represents the historical weight value within a time window. Represents the time step T The weight value of l is the length of the time window; using a long short-term memory network, including 2 LSTM layers and 1 fully connected layer, the number of neurons in each layer is , the activation function is tanh, the weight prediction value for a period of time in the future (such as 24 hours) is predicted, and the loss function is the root mean square error;

[0069] When using, combine the contents of step 201 to step 203:

[0070] The particle swarm optimization algorithm is used to dynamically calculate the feature weight ratio, construct a three-dimensional visualization map, and use the back propagation neural network and long short-term memory network to predict the recommendation weight and weight distribution, realizing real-time dynamic adjustment of feature weights, improving the flexibility and accuracy of the recommendation algorithm, and at the same time, the visualization map enhances the transparency and interactivity of the algorithm.

[0071] Step 3: Update the feature weight distribution based on the video content parameters and spatiotemporal feature parameters adjusted by the user. Calculate the recommendation weight gain and comprehensive benefit value by combining the feature contribution value with the weight benchmark value to generate a real-time traffic gain curve.

[0072] The step three includes the following contents:

[0073] Step 301: Adjust parameters based on the 3D visualization graph. Adjustable parameters include video content parameters and spatiotemporal feature parameters. Video content parameters include tag density, keyword coverage, image complexity, close-up duration ratio, and interactive element insertion frequency. Spatiotemporal feature parameters include upload time and geographic location.

[0074] Step 302: When the user adjusts the parameters (e.g., increasing the duration of "agricultural product close-up" from 50% to 80%), the particle swarm optimization algorithm formula in step 2 is used to update the particle speed and position, and the weight distribution of each dimension feature under the new parameter combination is calculated (e.g., the weight of "close-up duration ratio" increases from 15% to 19.7%).

[0075] Step 303: Based on the feature contribution values ​​calculated by the TreeSHAP algorithm in step 1 (e.g., the contribution value of "label 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 (e.g., the weight of "label density" increases from 0.25 to 0.38), the recommendation weight gain is calculated using the formula: ,in, R represents the recommendation weight gain, Indicates the d The weight of the feature, t For the moment, Indicates the d The feature contribution value of each feature, D Indicates the total number of features. When the user adjusts the parameters, the feature weight value is updated in real time, and the recommended weight gain is recalculated to generate a traffic gain curve. The horizontal axis is time and the vertical axis is the recommended weight gain value (interval [0,1]).

[0076] Step 304: The user can adjust multiple parameters simultaneously (e.g., reduce "label density" to 50%, increase "close-up duration ratio" to 80%), obtain the weight baseline value of each dimension feature in step 1, and calculate the comprehensive benefit value based on the weight distribution of each dimension feature under the new parameter combination: ,in, 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 benefit value;

[0077] When using, combine the contents of steps 301 to 304:

[0078] By dynamically updating feature weight distribution through user adjustment of parameters, the recommendation weight gain and comprehensive benefit value are calculated by combining the feature contribution value and the weight benchmark value, and a real-time traffic gain curve is generated. This enables users to actively intervene in the algorithm, quickly respond to market changes, improve recommendation effects, and intuitively display the adjustment effects through the traffic gain curve to assist decision-making.

[0079] Step 4: 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 between the experimental group and the control group, and update the model parameters based on the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0080] The fourth step includes the following contents:

[0081] Step 401: Based on the comprehensive benefit value generated in step 3, the parameter combination with the highest comprehensive benefit value after multiple adjustments is deployed to the short video platform through a real-time feedback mechanism. Specifically, the following steps are included:

[0082] Synchronize user-adjusted parameters (such as "label density" and "close-up duration ratio") to the platform through the API interface, overwriting the original default parameter configuration;

[0083] After the recommendation weight is updated, every time window Δ t (e.g. 30 minutes) Extract video exposure, click-through rate, and user stay time to calculate actual revenue deviation value : ,in, Pw is the predicted recommendation weight, Aw The actual recommended weight. If the actual return deviation exceeds the deviation threshold, the parameter rollback mechanism is triggered and the parameter configuration is restored to the previous state.

[0084] Step 402: Divide the traffic into an experimental group (applying the new parameters) and a control group (retaining the original parameters) based on the user ID hash value. Compare the key metrics within the experimental period (e.g., 24 hours), including impression conversion rate, user retention rate, and interactive behavior density. A two-sided T-test (significance level α = 0.05) is performed on the experimental and control group data. If the p-value is less than α and the improvement in the experimental group metric exceeds the baseline (e.g., click-through rate +5%), the parameter adjustment is considered effective.

[0085] Step 403: Store the effective parameter configurations and results of the experimental group in the MongoDB database, marking the spatiotemporal context (e.g., "promotional day configuration," "regional differentiation strategy"). Based on the newly generated behavioral data, retrain the GBDT model from step 1 and the LSTM network from step 2, updating the feature contribution values ​​and weight prediction values. If multiple consecutive parameter adjustments (e.g., three times) result in actual revenue deviations exceeding the deviation threshold, the user's permissions are automatically frozen and the parameter rollback mechanism is triggered.

[0086] When using, combine the contents of step 401 to step 403:

[0087] By deploying the parameter combination with the highest comprehensive benefit value to the short video platform, combining real-time feedback mechanism and A / B testing, monitoring actual benefit deviations and verifying parameter validity, we ensure the continuity and accuracy of algorithm optimization, and can promptly roll back invalid parameters, store valid configurations, and improve the overall recommendation effect and user experience of the platform.

[0088] See also Figure 2 ,The present invention also provides an algorithm information visualization interactive system, including: a model construction module, a weight distribution module, a benefit evaluation module and a parameter deployment module;

[0089] The model building module acquires video data from short video platforms in real time, builds a gradient boosting decision tree model, calculates the feature contribution value of each dimensional feature, and generates a weight benchmark value based on the historical weight data of each dimensional feature.

[0090] The weight distribution module dynamically calculates the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, constructs a three-dimensional visualization map, uses the back propagation neural network to predict the recommendation weight, and uses the long short-term memory network to predict the weight distribution of each dimension feature;

[0091] The revenue evaluation module updates the feature weight distribution based on the video content parameters and spatiotemporal feature parameters adjusted by the user, calculates the recommendation weight gain and comprehensive revenue value by combining the feature contribution value with the weight reference value, and generates a real-time traffic gain curve;

[0092] The parameter deployment module deploys the parameter combination with the highest comprehensive benefit value to the short video platform, monitors the actual benefit deviation through a real-time feedback mechanism, conducts A / B testing between the experimental group and the control group, and updates the model parameters based on the test results, triggering the parameter rollback mechanism or storing the effective configuration.

[0093] In the application, the several formulas involved are all calculated by taking their numerical values ​​after removing the dimensions, and the formula is a formula obtained by collecting a large amount of data and performing software simulation to obtain the latest real situation. The coefficients in the formula are set by technical personnel in this field according to actual conditions.

[0094] The above embodiments can be implemented in whole or in part through software, hardware, firmware, or any other 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 skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0095] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0096] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A method for interactive visualization of algorithm information, characterized by: include: Acquire video data from short video platforms in real time, build a gradient boosting decision tree model, calculate the feature contribution value of each dimension feature, and generate a weight benchmark value based on the historical weight data of each dimension feature; The particle swarm optimization algorithm is used to dynamically calculate the real-time weight ratio of each dimension feature, construct a three-dimensional visualization map, use the back propagation neural network to predict the recommendation weight, and use the long short-term memory network to predict the weight distribution of each dimension feature; Update the feature weight distribution based on the video content parameters and spatiotemporal feature parameters adjusted by the user, calculate the recommendation weight gain and 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 a real-time feedback mechanism, conduct A / B testing between the experimental group and the control group, and update the model parameters based on the test results, triggering the parameter rollback mechanism or storing the effective configuration.

2. The algorithm information visualization interactive method according to claim 1, characterized in that: The specific steps of obtaining video data from a short video platform in real time and building a gradient boosting decision tree model include: Through the open API interface of the short video platform, the original data stream is captured in real time, including user behavior data, video content data and time-space data; a multi-dimensional feature vector matrix is ​​constructed, including user behavior features, video content features and time-space features; the XGBoost framework is used to build a gradient boosting decision tree model, and the gradient boosting decision tree model is trained based on historical data sets; the TreeSHAP algorithm is used to calculate the feature contribution value of each dimensional feature and generate a feature sensitivity heat map; the weight baseline value of each dimensional feature is generated based on the sliding average method and linear regression model.

3. The algorithm information visualization interactive method according to claim 1, characterized in that: The step of dynamically calculating the real-time weight proportion of each dimension feature by the particle swarm optimization algorithm includes: Initialize the particle swarm parameters, set the number of particles, number of iterations and 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 speed and position according to the particle swarm optimization algorithm formula, and map the particle swarm optimization results into 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. The algorithm information visualization interactive method according to claim 3, characterized in that: The step of predicting the recommendation weight by the back propagation neural network includes: Construct a network structure of input layer, hidden layer and output layer, and use ReLU activation function in the hidden layer; calculate the gradient of input feature to output through back propagation, and the absolute value of normalized gradient is the weight ratio.

5. The algorithm information visualization interactive method according to claim 4, characterized in that: The step of using the long short-term memory network to predict the weight distribution includes: Convert the historical spatiotemporal weight sequence into a time window input in a supervised learning format; construct a network structure consisting of an LSTM layer and a fully connected layer to predict the weight prediction value for future time periods.

6. The algorithm information visualization interactive method according to claim 1, characterized in that: The step of calculating the recommendation weight gain and the comprehensive benefit value includes: Dynamically update feature weight distribution based on user adjustment parameters, and calculate recommendation weight gain based on feature contribution value: ,in, R is the recommendation weight gain, For the d The weight of the feature, t For the moment, For the d The feature contribution value of each feature, D is the total number of features; a real-time traffic gain curve is generated, with the horizontal axis being time and the vertical axis being the recommended weight gain value.

7. The algorithm information visualization interactive method according to claim 6, characterized in that: The comprehensive return value is calculated based on the deviation between the weight base value and the adjusted weight: ,in, For the new parameter combination d The weight of the feature, For the d The weighted benchmark value of each feature, Zw is the comprehensive benefit value.

8. The algorithm information visualization interactive method according to claim 1, characterized in that: The steps of deploying parameter combinations and A / B testing include: Synchronize the parameter combination with the highest comprehensive benefit value to the short video platform through the API interface, overwriting the default configuration; Real-time monitoring of actual profit deviation values. If the deviation threshold is exceeded, the parameter rollback mechanism will be triggered. Divide the user ID hash value into experimental and control groups, and calculate the exposure conversion rate, user retention rate, and interactive behavior density; A two-sided T test was used to verify the significance of the difference in indicators. If valid, the configuration was saved and the model was retrained.

9. The algorithm information visualization interactive method according to claim 8, characterized in that: The triggering conditions of the parameter rollback mechanism are: The actual return deviation value exceeds the preset deviation threshold several times in a row; The key indicators of the experimental group failed the significance test or the improvement rate was lower than the baseline; Automatically freeze user permissions and restore to historical valid configurations.

10. An algorithm information visualization interactive system, used to implement the method according to any one of claims 1 to 9, characterized in that: include: The model building module obtains video data from short video platforms in real time, builds a gradient boosting decision tree model, calculates the feature contribution value of each dimension feature, and generates a weight benchmark value based on the historical weight data of each dimension feature; The weight distribution module dynamically calculates the real-time weight ratio of each dimension feature through the particle swarm optimization algorithm, constructs a three-dimensional visualization map, uses the back propagation neural network to predict the recommendation weight, and uses the long short-term memory network to predict the weight distribution of each dimension feature; The revenue evaluation module updates the feature weight distribution based on the video content parameters and spatiotemporal feature parameters adjusted by the user, calculates the recommendation weight gain and comprehensive revenue value by combining the feature contribution value with the weight reference value, and generates a real-time traffic gain curve; The parameter deployment module deploys the parameter combination with the highest comprehensive benefit value to the short video platform, monitors the actual benefit deviation through a real-time feedback mechanism, conducts A / B testing between the experimental group and the control group, and updates the model parameters based on the test results, triggering the parameter rollback mechanism or storing the effective configuration.

Images