AI short drama content publishing scheduling method and system based on user behavior prediction

Abstract

The application discloses an AI short drama content release scheduling method and system based on user behavior prediction, relates to the technical field of big data analysis, and comprises the following steps: obtaining user viewing behavior stream data generated by a target short drama in a current episode in an observation window, and inputting the user viewing behavior stream data into a trained behavior-payment opportunity predictor to output estimated churn probability values and estimated payment trigger probability values of users at different progress coordinates of the current episode; generating a dynamic payment unlocking point correction offset amount at a zero-crossing position on a progress axis according to the estimated churn probability values and the estimated payment trigger probability values, in combination with a preset unlocking point offset constraint condition; adjusting a payment unlocking position of the current episode for non-exposed users, and updating a corresponding playing strategy configuration. The technical problem of high user churn rate and payment conversion opportunity loss caused by the static and unified setting of the existing short drama payment unlocking point is solved.

Description

Technical Field

[0001] This application relates to the field of big data analytics technology, specifically to an AI-powered short drama content release and scheduling method and system based on user behavior prediction. Background Technology

[0002] Short dramas, with their short length, fast pace, and tight plots, have become an important part of content consumption. Especially in the mobile internet environment, users' demand for fragmented entertainment content is growing, leading to explosive growth in the short drama market. To achieve commercial monetization, most mainstream short drama platforms currently adopt a "free trial + paid unlock" model. Users can watch the first part of the short drama for free, but need to pay to unlock more exciting parts.

[0003] However, in existing technologies, the pay unlock points for short dramas are usually fixed positions pre-set by platforms or content providers based on experience. This static and uniform approach ignores individual user viewing preferences, behavioral characteristics, and the changing appeal of the content itself. If the pay unlock point is set too early, it may trigger a payment request before users have developed sufficient interest in the plot or invested emotionally, easily causing user resentment and leading to increased user churn. If the pay unlock point is set too late, the platform may lose potential paid conversion opportunities, as some users may leave before the free content ends due to the dragging plot or loss of interest, making it impossible to effectively convert them into paying users.

[0004] Furthermore, the distribution of plot appeal varies among different types of short dramas and different episodes of the same short drama. Fixed payment unlock points are difficult to adapt to dynamic changes and cannot achieve the optimal balance between user retention and paid conversion. Summary of the Invention

[0005] This application provides an AI-powered short drama content release and scheduling method and system based on user behavior prediction, which solves the technical problem that the static and standardized setting of paid unlock points for existing short dramas leads to high user churn rates and lost paid conversion opportunities.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows: Firstly, this application provides an AI-powered short drama content publishing and scheduling method based on user behavior prediction, the method comprising: Acquire user viewing behavior stream data generated within the observation window for the current episode of the target short drama; The viewing behavior stream data is input into a trained behavior-payment timing predictor, which outputs the estimated churn probability and the estimated payment trigger probability of the user at different progress coordinates in the current episode. Based on the estimated churn probability value and the estimated payment trigger probability value, at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint, a dynamic payment unlock point correction offset is generated. The dynamic paid unlock point correction offset is used to adjust the paid unlock position for unexposed users in the current episode, and the corresponding playback strategy configuration is updated.

[0007] Secondly, this application provides an AI-powered short drama content publishing and scheduling system based on user behavior prediction, including: The data acquisition module is used to acquire user viewing behavior stream data generated within the observation window for the current episode of the target short drama. The model training module is used to input the viewing behavior stream data into the trained behavior-payment timing predictor and output the estimated churn probability value and the estimated payment trigger probability value of the user at different progress coordinates in the current episode. The offset calculation module is used to generate a dynamic payment unlock point correction offset based on the estimated churn probability value and the estimated payment trigger probability value at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint conditions. The strategy update module is used to adjust the paid unlock position for unexposed users in the current episode based on the dynamic paid unlock point correction offset, and update the corresponding playback strategy configuration.

[0008] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides an AI-powered short drama content release and scheduling method and system based on user behavior prediction. First, it acquires user viewing behavior stream data generated within the observation window for the current episode of the target short drama, reflecting user interaction and behavioral patterns during viewing. Second, the viewing behavior stream data is input into a trained behavior-payment timing predictor, which outputs the estimated churn probability and estimated payment trigger probability for users at different progress coordinates in the current episode, providing a deeper understanding of user behavior tendencies at different viewing stages from a data perspective. Then, based on the two probability values, a zero-crossing position is found on the progress axis, and combined with preset unlock point offset constraints, a dynamic payment unlock point correction offset is generated, fully considering the dynamic changes in user behavior and differences in content attractiveness. Finally, the dynamic payment unlock point correction offset is used to adjust the payment unlock position for unexposed users in the current episode, and the corresponding playback strategy configuration is updated. This achieves dynamic adjustment of the payment unlock point, enabling flexible optimization of the payment unlock position based on actual user behavior and real-time content feedback, thereby finding the optimal balance between user retention and paid conversion.

[0009] Through the above technical solution, this application achieves the technical effect of dynamically adjusting the location of the paid unlock point based on the user's real-time viewing behavior, thereby achieving the optimal balance between user retention and paid conversion. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating the AI ​​short drama content publishing and scheduling method based on user behavior prediction provided in the embodiments of this application; Figure 2 This is a schematic diagram of the structure of the AI ​​short drama content publishing and scheduling system based on user behavior prediction provided in the embodiments of this application.

[0012] The components represented by each number in the attached diagram are explained below: Data acquisition module 11, model training module 12, offset calculation module 13, policy update module 14. Detailed Implementation

[0013] This application provides an AI-powered short drama content release and scheduling method and system based on user behavior prediction, which addresses the technical problem of high user churn and lost paid conversion opportunities caused by the static and standardized setting of existing short drama paid unlock points.

[0014] Example 1, as Figure 1 As shown in the embodiments of this application, an AI-powered short drama content publishing and scheduling method based on user behavior prediction is provided, including: S10: Obtain user viewing behavior stream data generated within the observation window for the current episode of the target short drama; In this embodiment of the application, by collecting multi-dimensional user viewing behavior stream data, a behavioral profile that reflects the user's real viewing experience and points of interest is constructed. The observation window refers to a specific time period or data volume threshold range used to collect the user's initial viewing behavior data after the current episode of content is published. For example, it is set to be within 2 hours after the content is published, or the observation window will automatically end when 5,000 valid user behavior records are accumulated.

[0015] Specifically, step S10 of the method includes: The user viewing behavior stream data includes the dwell time decay trajectory, the start and end position pairs of speed-up operations, and the frequency of bullet screen emotional polarity transitions. The observation window is divided into multiple consecutive time slices. The ratio of the user's viewing time in each time slice to the length of the corresponding time slice is calculated. The ratios of adjacent time slices are connected sequentially to construct a viewing time decay trajectory. Based on the speed control command triggered by the user on the playback interface, record the first progress coordinate corresponding to the start time of speed increase and the second progress coordinate corresponding to the end time of speed increase; The position binary formed by the first progress coordinate and the second progress coordinate is used as the start and end position pair of the speed-up operation; Collect the bullet screen text stream within the observation window, extract bullet screen text segments at fixed time steps, and input them into the sentiment polarity classification network to obtain the sentiment polarity label for each time step; The frequency of emotional polarity transitions in bullet comments is obtained by counting the cumulative number of times the emotional polarity label changes from negative to positive between adjacent time steps. The user viewing behavior stream data is obtained by summarizing the attenuation trajectory of the dwell time, the start and end positions of the speed-up operation, and the frequency of the emotional polarity transition of the bullet comments.

[0016] In this embodiment, the user viewing behavior stream data includes the dwell time decay trajectory, the start and end positions of the speed-up operation, and the frequency of emotional polarity transitions in the bullet comments. First, the construction of the dwell time decay trajectory is explained in detail. For example, when the observation window is divided into 10 time slices, each time slice is 30 seconds long, if the user's actual viewing dwell time in a certain time slice is 25 seconds, then the ratio of that time slice is... The ratios of the 10 time slices are connected in chronological order to form a continuous curve from the start time slice to the end time slice. The change in the slope of the curve can intuitively reflect the decline in user attention over time. The larger the absolute value of the negative slope, the faster the user's dwell time decreases and the more obvious the decline in interest in the current content.

[0017] Secondly, the start and end positions of the speed-up operation are precisely matched to progress coordinates down to the millisecond level. For example, if a user triggers 2x speed playback at 1 minute 20 seconds 300 milliseconds and switches back to normal 1x speed playback at 3 minutes 15 seconds 500 milliseconds, then the first progress coordinate is 00:01:20.300 and the second progress coordinate is 00:03:15.500. This position tuple can be used to analyze the specific content segments that the user believes need to accelerate the pace of the plot.

[0018] Furthermore, in the calculation of the frequency of emotional polarity transitions in bullet comments, a fixed time step is determined based on the statistical distribution of the average sending interval of the bullet comment stream in the short drama and the duration of the emotional turning point in the plot. For example, the fixed time step can be set to 10 seconds. The captured bullet comment text segments are classified into three emotional tags: positive, negative, and neutral. Only the number of transitions from negative to positive is counted. For example, if the bullet comment emotional tag is negative within a certain 10-second time step and becomes positive within the next 10-second time step, it is counted as 1 transition, which is the frequency of emotional polarity transitions in bullet comments.

[0019] Furthermore, by summarizing the decay trajectory of dwell time, the start and end positions of speed-up operations, and the frequency of emotional polarity transitions in bullet comments, the obtained user viewing behavior stream data can be used to quantify the intensity of user group's emotional feedback to plot turning points.

[0020] The construction steps of the emotion polarity classification network include: A collection of bullet screen text samples generated during the playback of historical short dramas was collected, and each bullet screen text sample was labeled with an emotional polarity tag, wherein the emotional polarity tag includes positive emotional markers and negative emotional markers; The set of labeled bullet screen text samples is divided into a training set and a validation set; Construct a sentiment polarity classification network based on a convolutional neural network architecture; Using the barrage text samples in the training set as input and the corresponding sentiment polarity labels as supervision signals, the sentiment polarity classification network is trained in a supervised manner. The cross-entropy loss function is used to calculate the error between the classification output and the supervision signal, and the network weight parameters are updated through backpropagation. When the sentiment polarity classification network achieves a sentiment polarity classification accuracy exceeding a preset accuracy threshold on the validation set and the loss no longer decreases after multiple consecutive training rounds, training is stopped, and the trained sentiment polarity classification network is obtained.

[0021] In this embodiment, firstly, a set of bullet screen text samples is collected from the historical short drama playback process. For example, user bullet screen data of different types of short dramas in the past 6 months are extracted from the platform database to ensure that the samples cover different themes and audience groups. Each bullet screen text sample needs to be labeled with the corresponding emotional polarity tag. Positive emotional tags may include positive emotional expressions such as "happy", "wonderful", and "expectation", while negative emotional tags may include negative emotional expressions such as "bored", "procrastinating", and "disappointed". The labeling process can combine manual labeling with machine pre-labeling. First, the machine performs preliminary labeling based on the emotional dictionary, and then professional labelers review and correct the labeling results to ensure the accuracy of the tags.

[0022] Secondly, the labeled bullet screen text sample set is randomly divided into a training set and a validation set in a 7:3 ratio. The training set is used to learn the model parameters, and the validation set is used to evaluate the model's generalization ability and training effect.

[0023] Then, a sentiment polarity classification network based on a convolutional neural network architecture is constructed. This network may include an embedding layer, a convolutional layer, a pooling layer, and a fully connected layer. The embedding layer converts the input bullet screen text into a low-dimensional dense word vector representation. The convolutional layer extracts local sentiment features in the text through convolutional kernels of different sizes. The pooling layer performs dimensionality reduction on the feature maps output by the convolutional layer. The fully connected layer maps the pooled features to the sentiment polarity classification space.

[0024] Furthermore, during the training phase, the barrage text samples in the training set are used as input, and the corresponding sentiment polarity labels are used as supervision signals. The cross-entropy loss function is used to calculate the error between the network classification output and the true label. The weight parameters of the network are continuously updated through the backpropagation algorithm. During the training process, the sentiment polarity classification accuracy of the model on the validation set is monitored in real time. When the accuracy exceeds the preset accuracy threshold, such as 85%, and the loss value of the validation set no longer decreases in 5 consecutive training rounds, the model is considered to have reached convergence and training is stopped, thus obtaining the trained sentiment polarity classification network, which is used to identify the sentiment polarity of barrage text.

[0025] For example, a sentiment polarity classification network is trained based on a convolutional neural network. The embedding layer receives the input matrix and further maps the word vectors into 256-dimensional feature vectors through a trainable embedding weight matrix, outputting an embedding feature map with a dimension of 50×256. The convolutional layer uses three different sizes of convolutional kernels, such as 3×256, 5×256, and 7×256, with 64 kernels of each size. These kernels perform convolution operations on the embedding feature map to extract local sentiment features from the text. The pooling layer uses max pooling, taking the maximum value along the time dimension for each group of convolutional feature maps to obtain three groups of 64-dimensional pooled feature vectors. These are then concatenated to form a 192-dimensional global sentiment feature vector. The fully connected layer consists of two layers. The first layer maps the 192-dimensional feature vector to 128 dimensions using the ReLU activation function; the second layer maps the 128-dimensional feature vector to 2 dimensions, corresponding to positive and negative sentiment, and outputs the classification probability using the Softmax activation function. During training, the initial learning rate was set to 0.001, and the Adam optimizer was used. After each training round, the classification accuracy was evaluated on the validation set. When the accuracy improvement was less than 0.5% for 5 consecutive rounds, the learning rate was reduced to 0.5 times the original value until the model converged.

[0026] S20: Input the viewing behavior stream data into the trained behavior-payment timing predictor and output the estimated churn probability value and the estimated payment trigger probability value of the user at different progress coordinates in the current episode. In this embodiment of the application, the acquired viewing behavior stream data is input into a trained behavior-payment timing predictor. The behavior-payment timing predictor adopts a deep learning model architecture with dual output branches, which outputs the estimated churn probability value and the estimated payment trigger probability value, respectively.

[0027] The training steps of the behavior-payment timing predictor include: Collect a sample set of historical user viewing behavior streams generated within the observation window for similar short dramas. For each historical user viewing behavior stream sample, label it with historical churn event markers and historical payment event markers aligned with different progress coordinates to form a dual-supervised label set. Construct a time-series prediction network architecture that includes a shared feature extraction layer, a churn probability prediction branch, and a payment trigger prediction branch; The input of the shared feature extraction layer is used to receive user viewing behavior stream data, and the output of the shared feature extraction layer is connected to the input of the churn probability prediction branch and the payment trigger prediction branch, respectively. The output of the churn probability prediction branch is used to output the estimated churn probability value of the user at different progress coordinates in the current episode, and the output of the payment trigger prediction branch is used to output the estimated payment trigger probability value of the user at different progress coordinates in the current episode. Using the historical user viewing behavior stream sample set as input, and the historical churn event tags and historical payment event tags in the dual-supervised tag set as joint supervision signals, the time-series prediction network architecture is subjected to multi-task supervised training until the network converges, thus obtaining a trained behavior-payment timing predictor.

[0028] In this embodiment, firstly, a sample set of historical user viewing behavior streams generated within the observation window for historical episodes of similar short dramas is collected. For example, historical episode data of nearly 1,000 short dramas with the same theme and target audience as the target short drama are selected to ensure that the samples have high similarity and reference value. For each historical user viewing behavior stream sample, historical churn event markers and historical payment event markers are added, aligned with different progress coordinates.

[0029] Specifically, historical churn event markers can be defined as the behavior of a user stopping watching at a certain progress coordinate and not resuming watching within a preset time threshold, such as 5 minutes, which is marked as 1, otherwise as 0; historical payment event markers are defined as the behavior of a user triggering a paid unlock operation at a certain progress coordinate, which is marked as 1, otherwise as 0. A dual-supervised label set containing churn and payment status at different progress coordinates is constructed for each historical sample.

[0030] Secondly, a temporal prediction network architecture is constructed, comprising a shared feature extraction layer, a churn probability prediction branch, and a payment trigger prediction branch. The shared feature extraction layer employs a bidirectional long short-term memory network combined with an attention mechanism to extract deep, dynamic temporal features from user viewing behavior stream data. The bidirectional long short-term memory network can capture the dependencies between behavior sequences, while the attention mechanism can automatically focus on key behavior segments that contribute significantly to churn and payment behavior prediction. The input of the shared feature extraction layer receives preprocessed user viewing behavior stream data, such as concatenating multi-dimensional features like dwell time decay trajectory, speed-up operation start and end positions, and the frequency of bullet screen sentiment polarity transitions.

[0031] Furthermore, the output of the shared feature extraction layer is connected to the inputs of the churn probability prediction branch and the payment trigger prediction branch, respectively. The churn probability prediction branch consists of two fully connected layers. The first layer uses the ReLU activation function, and the second layer uses the Sigmoid activation function, outputting the estimated churn probability value of the user at different progress coordinates in the current set (the value ranges from 0 to 1). The payment trigger prediction branch also uses two fully connected layers, with a similar structure to the churn probability prediction branch, and also outputs the estimated payment trigger probability value using the Sigmoid function.

[0032] During training, a historical user viewing behavior stream sample set is used as input, and historical churn event tags and historical payment event tags from the dual-supervised label set are used as joint supervision signals to perform multi-task supervised training on the time-series prediction network architecture. The loss function is a weighted sum of churn prediction loss and payment prediction loss, where both churn prediction loss and payment prediction loss use the binary cross-entropy loss function. The weight parameters of all layers of the network are updated through the backpropagation algorithm. During training, the overall prediction performance of the model on the validation set, such as the average AUC value, is continuously monitored. When the validation set performance no longer improves and the network loss tends to stabilize, training is stopped, thus obtaining a trained behavior-payment timing predictor. This predictor can output the estimated churn probability value and the estimated payment trigger probability value under different progress coordinates in the current set based on the input user viewing behavior stream data.

[0033] For example, the training steps for the behavior-payment timing predictor are as follows: The samples are divided into a training set and a validation set in an 8:2 ratio. The training set is used for parameter learning, and the validation set is used for model tuning. When constructing the temporal prediction network architecture, the shared feature extraction layer uses a two-layer bidirectional LSTM with 128 hidden units per layer and an input feature dimension of 3. The hidden states output by the LSTM are weighted and summed using an attention mechanism to obtain a shared feature vector of dimension 256. The churn probability prediction branch includes a fully connected layer (e.g., 256→64, ReLU activation) and an output layer (64→1, Sigmoid activation). The payment trigger prediction branch has the same structure but independent parameters. During training, the joint loss function is set with a churn loss weight of 0.6 and a payment loss weight of 0.4. The Adam optimizer is used with an initial learning rate of 0.001. Every 5 rounds, if the validation set AUC does not improve, the learning rate decreases by 50%. Training stops when the AUC improvement is less than 0.001 for 10 consecutive rounds, resulting in the trained behavior-payment timing predictor.

[0034] S30: Based on the estimated churn probability value and the estimated payment trigger probability value, at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint, a dynamic payment unlock point correction offset is generated. In this embodiment, the estimated churn probability curve and the estimated payment trigger probability curve on the progress axis are zero-crossing position detected. Specifically, the estimated churn probability value is compared with the preset churn probability threshold, and then combined with the preset unlock point offset constraint, to generate a dynamic payment unlock point correction offset.

[0035] Specifically, step S30 in the method includes: Calculate the difference between the estimated churn probability value and the estimated payment trigger probability value at different progress coordinates in the current episode, and scan along the progress axis at the positions where the sign of the difference changes from positive to negative or from negative to positive to obtain the zero-crossing positions; If the zero-crossing position is unique, the progress difference between the zero-crossing position and the current fixed-payment unlocking position is used as the initial offset. If the number of zero-crossing positions is greater than one, the zero-crossing position closest to the end time of the observation window is selected, and the progress difference between the selected zero-crossing position and the current fixed payment unlocking position is calculated as the initial offset. If the zero-crossing position does not exist, scan from front to back along the progress axis, and use the progress coordinate of the first time the estimated payment trigger probability value exceeds the estimated churn probability value as the reference, and calculate the progress difference between the reference and the current fixed payment unlock position as the initial offset. The initial offset is limited by a preset unlock point offset constraint to obtain the dynamic paid unlock point correction offset.

[0036] In this embodiment of the application, firstly, the difference between the estimated churn probability value and the estimated payment trigger probability value at different progress coordinates of the current episode is calculated. For example, at progress coordinate t, The sign of the difference D(t) is scanned point by point along the plot progress axis from the starting position to the ending position. When the value of D(t) changes from negative to positive, it means that at that progress coordinate, the user's willingness to pay exceeds the risk of churn for the first time. This is a zero-crossing position. Similarly, if D(t) changes from positive to negative, it is also considered a zero-crossing position. All such progress coordinate points with sign transitions are recorded.

[0037] Secondly, if only one zero-crossing position is obtained after scanning, the progress coordinate of this zero-crossing position is directly compared with the current default fixed paid unlocking position, such as at 60% of the total episode duration. The progress difference between the two is calculated, and this difference is the initial offset. For example, if the fixed paid unlocking position is at 300 seconds and the zero-crossing position is at 280 seconds, the initial offset is -20 seconds, indicating that the paid unlocking point should be moved forward by 20 seconds.

[0038] If multiple zero-crossing positions are detected, such as when there are multiple intersections between willingness to pay and risk of churn during the plot development, the zero-crossing position that is closest to the end of the observation window, i.e. the end of the current episode, is selected. Since the closer to the end of the episode, the more complete the user's overall plot experience of the current episode is, and the higher the reference value of their behavioral feedback, the progress difference between the most recent zero-crossing position and the fixed payment unlock position is calculated as the initial offset.

[0039] If no zero-crossing position is found during the entire progress axis scan, i.e., D(t) is always negative or always positive, meaning the probability of triggering payment is consistently lower than the probability of churn or consistently higher than the probability of churn, then the progress axis is scanned from front to back to find the progress coordinate point where the estimated probability of triggering payment P(t) first exceeds the estimated probability of churn L(t) as the reference position. If P(t) is always higher than L(t), then the starting position of the set is used as the reference; if P(t) is always lower than L(t), then the progress coordinate corresponding to the relative maximum value of P(t) within the observation window is used as the reference, and the progress difference between this reference position and the fixed payment unlock position is calculated as the initial offset.

[0040] Furthermore, after obtaining the initial offset, a limiting process is performed using preset unlock point offset constraints. These constraints typically include upper and lower limits on the offset, such as stipulating that the dynamically corrected offset must not exceed ±60 seconds, and the corrected paid unlock point must not be earlier than 20% of the total episode duration or later than 80% of the total episode duration. If the initial offset exceeds these limits, it is truncated to the boundary value, ultimately yielding a dynamically corrected paid unlock point offset that meets the constraints.

[0041] Furthermore, the preset unlocking point offset constraint includes: The absolute value of the dynamic paid unlock point correction offset shall not exceed the first proportional threshold of the total progress of the current episode. The paid unlock position after offset adjustment does not exceed the second proportional threshold range intercepted forward from the end point of the current episode's content progress; When the offset of the dynamic paid unlock point is corrected, if the directions of two adjacent adjustments are opposite and the magnitudes both exceed the third proportional threshold, then the current offset update is paused and the previous paid unlock position is maintained unchanged.

[0042] In this embodiment, firstly, the absolute value of the dynamic payment unlock point correction offset does not exceed a first percentage threshold of the total progress of the current episode. For example, if the first percentage threshold is set to 15%, and the total duration of the current episode is 500 seconds, then the absolute value of the correction offset will not exceed a certain limit. This is to prevent drastic fluctuations in the location of the paid unlock point due to excessive single offset, which could affect the stability of the user's viewing experience.

[0043] Secondly, the paid unlock position after offset adjustment should not exceed the second percentage threshold range intercepted from the end of the current episode's content progress. For example, the second percentage threshold range is set to 20% to 80% of the total duration of the current episode, meaning the earliest adjusted paid unlock point cannot be earlier than [the specified threshold]. No later than To ensure that paid unlock points are within a reasonable range of plot development, avoid them appearing too early and affecting the user's understanding of the plot, or appearing too late and causing the user to churn and lose the opportunity to trigger payment.

[0044] Finally, when generating the dynamic paid unlock point correction offset, if two adjacent adjustments are in opposite directions and both exceed the third proportional threshold (for example, if the third proportional threshold is set to 5%, or 25 seconds, and the previous adjustment was to shift backward by 30 seconds, and this adjustment is to shift forward by 35 seconds, with both directions being opposite and both exceeding 25 seconds), it is determined that the current user behavior prediction may have abnormal fluctuations or noise interference. In this case, the current offset update is paused, and the previous paid unlock position is kept unchanged to ensure the smoothness and reliability of the paid unlock point adjustment and avoid frequent large-scale reverse adjustments that may confuse users.

[0045] Specifically, the calculation steps for the first proportional threshold, the second proportional threshold range, and the third proportional threshold include: Calculate the average percentage of the distance between the paid unlock position and the end point of the content progress for similar short dramas in the past, and use it as the benchmark offset percentage. The average decay slope of the decay trajectory of the duration of the broadcast history episodes of the target short drama within the observation window is obtained, and the first proportion threshold is calculated in combination with the benchmark offset ratio. The frequency of barrage emotional polarity transitions in the current episode of the target short drama within the observation window is collected. Combined with the baseline offset ratio, the upper and lower limits of the second proportional threshold interval are determined. For each unit increase in the frequency of barrage emotional polarity transitions, the upper and lower limits of the second proportional threshold interval decrease synchronously according to a fixed shrinkage step. The cumulative value of the coverage width between the first and second progress coordinates of all speed-up operations within the observation window of the current set is calculated, and the ratio of the cumulative value to the total progress of the current set is used as the third ratio threshold.

[0046] In this embodiment, firstly, the average percentage of the distance between the paid unlock position and the end point of the content progress in historical similar short dramas is calculated as the baseline offset percentage. Specifically, the coordinates of the paid unlock positions for all episodes are extracted from the historical similar short drama dataset. The time distance from the paid unlock position to the end point of the content progress for each episode is calculated. Then, this distance is divided by the total duration of the corresponding episode to obtain the offset percentage for each episode. Finally, the arithmetic mean of the offset percentages for all episodes is taken to obtain the baseline offset percentage. For example, if a similar short drama contains 100 episodes, and the offset percentages for each episode are 10%, 12%, 15%, etc., and the average value is 13%, then the baseline offset percentage is 13%.

[0047] Secondly, the average decay slope of the dwell time decay trajectory of the historical episodes of the target short drama within the observation window is obtained. Combined with the benchmark offset ratio, the first proportional threshold is calculated. The dwell time decay trajectory reflects the changing trend of the dwell time per unit time during the viewing process. The larger the decay slope, the faster the user's attention declines, and the stricter the restriction on the offset should be.

[0048] Specifically, for each historical episode of the target short drama that has been broadcast, within its observation window, a dwell time sequence is constructed with time as the horizontal axis and average user dwell time as the vertical axis. The decay slope of this sequence is fitted by linear regression, and then the arithmetic mean of the decay slopes of all historical episodes is calculated to obtain the mean decay slope. The mean decay slope is then calculated with the baseline offset ratio, for example, the first ratio threshold = mean decay slope × baseline offset ratio.

[0049] Secondly, the frequency of emotional polarity transitions in bullet comments refers to the number of times the emotional polarity of the bullet comment text changes significantly within the observation window, such as from positive to negative, from negative to positive, or from neutral to positive / negative. A higher frequency of transitions indicates more intense plot conflict or greater user emotional fluctuations. Therefore, the reasonable range for paid unlock points should be appropriately narrowed to focus on key plot points. For example... ; .

[0050] Furthermore, the fixed shrinkage step size is determined by the statistical distribution of the relationship between the adjustment range of the paid unlock position and the change in the user completion rate in similar short dramas in the past. This makes the second proportion threshold range narrow smoothly as the frequency of the jump in the emotional polarity of the bullet screen increases. While constraining the offset range of the paid point, it avoids the range shrinkage from being too abrupt and causing the adjustment to fail. For example, it is taken as 1%.

[0051] Finally, the ratio of the cumulative value to the total progress of the current session is calculated and used as the third proportional threshold. The start and end positions of the speed-up operation reflect the time periods during which users actively adjust their viewing speed, and the coverage width is the duration of each speed-up operation period. The larger the cumulative coverage width, the stronger the user's personalized demand for the pacing of the plot, and the allowable fluctuation range for adjusting paid unlock points should be appropriately relaxed.

[0052] For example, if the total progress of the current set is 500 seconds, and there are 3 pairs of start and end positions for the speed-up operation within the observation window, namely [100 seconds, 150 seconds], [200 seconds, 280 seconds], and [350 seconds, 400 seconds], with coverage widths of 50 seconds, 80 seconds, and 50 seconds respectively, and a cumulative value of 50 + 80 + 50 = 180 seconds, then... This means that the magnitude of two consecutive adjustments is allowed to fluctuate within a range not exceeding 36% of the total progress of the current set. If it exceeds this proportion and in the opposite direction, the update will be paused.

[0053] S40: Adjust the paid unlock position for unexposed users in the current episode using the dynamic paid unlock point correction offset, and update the corresponding playback strategy configuration.

[0054] In this embodiment, the default paid unlock position for unexposed users in the current episode is adjusted in real time by dynamically adjusting the paid unlock point correction offset. Specifically, the dynamic paid unlock point correction offset is added to the current fixed paid unlock position to obtain the adjusted actual paid unlock position coordinates, which are then associated and bound with the playback identifier of the current episode to generate an updated playback strategy.

[0055] Specifically, step S40 in the method includes: The current fixed payment unlock position is algebraically superimposed with the dynamic payment unlock point correction offset to obtain the candidate payment unlock position. Determine whether the candidate paid unlock position falls within the second proportional threshold range intercepted forward from the end point of the current episode's content progress; If the candidate payment unlock position falls within the second ratio threshold range, then the candidate payment unlock position is used as the updated payment unlock position. If the candidate paid unlock position exceeds the upper limit of the second proportional threshold range, then the upper limit of the second proportional threshold range will be used as the updated paid unlock position. If the candidate paid unlock position is lower than the lower limit of the second proportional threshold range, then the lower limit of the second proportional threshold range will be used as the updated paid unlock position. The updated paid unlock location is associated with the playback identifier of the current episode to generate an updated playback strategy configuration item; After the observation window ends, receive playback requests for the current episode from newly arrived unexposed users, and read the updated paid unlock position from the updated playback strategy configuration. The updated paid unlock position is sent to the player of unexposed users, and the player is controlled to trigger the paid interception interface when the playback progress reaches the updated paid unlock position.

[0056] In this embodiment, firstly, the current fixed-payment unlock position and the dynamic-payment unlock point correction offset are algebraically superimposed. For example, if the current fixed-payment unlock position is 300 seconds and the dynamic correction offset is -20 seconds, then the candidate unlock position is... .

[0057] Secondly, it is determined whether the candidate position falls within the preset second proportion threshold range. Assuming the total duration of the current episode is 500 seconds, the second proportion threshold range is 20% to 80%, i.e., 100 seconds to 400 seconds. 280 seconds falls within this range, so 280 seconds is directly used as the updated paid unlock position. If the calculated result of the candidate position is 80 seconds, which is lower than the lower limit of the range by 100 seconds, then 100 seconds is used as the updated position; if the result is 420 seconds, which is higher than the upper limit of the range by 400 seconds, then 400 seconds is used as the updated position.

[0058] After determining the updated paid unlock location, it is associated with the unique playback identifier of the current episode and stored in the playback strategy configuration database to form a new configuration item. When the observation window ends and a new unexposed user requests to play the episode, the updated paid unlock location bound to the episode ID is read from the database and sent to the user's playback client via the interface. After receiving the location information, the client player automatically pops up the paid unlock interface when the playback progress bar reaches 280 seconds, prompting the user to make a payment to continue watching the subsequent content.

[0059] Furthermore, after updating the corresponding playback strategy configuration, it also includes: Obtain actual viewing behavior feedback data of unexposed users at the new paid unlock location, wherein the actual viewing behavior feedback data includes the actual churn event occurrence progress coordinates and the actual payment event trigger progress coordinates within the new observation window; Based on the estimated churn probability and estimated payment trigger probability of users at different progress coordinates in the current set output by the behavior-payment timing predictor, obtain the peak progress coordinate of the estimated churn probability and the peak progress coordinate of the estimated payment trigger probability. Calculate the first deviation between the actual churn event occurrence progress coordinate and the peak progress coordinate of the estimated churn probability value, and the second deviation between the actual payment event trigger progress coordinate and the peak progress coordinate of the estimated payment trigger probability value; Based on the weighted result of the first deviation and the second deviation, adjust the output response offset of the churn probability prediction branch and the payment trigger prediction branch in the behavior-payment timing predictor. The adjusted output response offset is added to the calculation process of the estimated churn probability value and the estimated payment trigger probability value in subsequent episodes.

[0060] In this embodiment, firstly, after the new paid unlock location is deployed, the actual viewing behavior data of unexposed users is continuously collected and transmitted back to the server in real time through the tracking and reporting mechanism of the user's playback client. Specifically, this includes the actual churn time point within the new observation window, which is usually within a 30% episode duration interval before and after the adjustment of the paid unlock location. For example, if the user stops playing and exits at 250 seconds, this 250 seconds is the progress coordinate of the actual churn event. It also includes the time point when the user actually triggers the payment operation, such as clicking the payment button at 285 seconds, which is the progress coordinate of the actual payment event trigger.

[0061] Secondly, from the output of the behavior-payment timing predictor, extract the progress coordinate corresponding to the maximum value of the estimated churn probability value L(t) on the entire progress axis for the current episode, i.e., the peak progress coordinate of the estimated churn probability value, assumed to be 260 seconds, and the progress coordinate corresponding to the maximum value of the estimated payment trigger probability value P(t), i.e., the peak progress coordinate of the estimated payment trigger probability value, assumed to be 290 seconds.

[0062] Next, calculate the first deviation and the second deviation. The first deviation is the difference between the actual churn event occurrence time coordinate and the estimated churn probability value peak time coordinate, i.e. The second deviation is the difference between the actual payment event triggering progress coordinate and the peak progress coordinate of the estimated payment triggering probability value, i.e. A negative deviation indicates that the actual event occurred earlier than the predicted peak, while a positive deviation indicates that it occurred later than the predicted peak.

[0063] Then, the output response offset of the predictor is adjusted based on the weighted result of the first and second biases. The weighting coefficients are determined based on the importance analysis of the impact of churn events and paid events on the final revenue in historical data. For example, the weight of paid events is usually higher than that of churn events. Assuming the weight of churn is 0.3 and the weight of paid events is 0.7, then the weighted bias = (-10 seconds) × 0.3 + (-5 seconds) × 0.7 = -6.5 seconds. If the weighted bias is negative, it means that the actual event occurred earlier than the prediction, and the peak values ​​of the churn probability and the peak values ​​of the paid event trigger probability in the subsequent output of the predictor need to be shifted forward by the corresponding value; if it is positive, it is shifted backward.

[0064] Finally, the calculated adjusted output response offset is used as a correction parameter and added to the calculation of the estimated churn probability value L(t) and the estimated payment trigger probability value P(t) for subsequent episodes. For example, when predicting the peak churn probability of the next episode, the original prediction coordinate was 300 seconds. After adding the offset, it is adjusted to 300 seconds + (-6.5 seconds) = 293.5 seconds, enabling the prediction model to dynamically adapt to the changing trends of actual user behavior, gradually improving prediction accuracy, and thus optimizing the dynamic adjustment effect of the payment unlock point in subsequent episodes.

[0065] In summary, compared to existing technologies, this application introduces a user behavior prediction mechanism to construct a dynamic offset for paid unlocking points. Combined with multi-dimensional proportional thresholds, it achieves fine-grained adjustment of the paid unlocking position, effectively solving the problem that traditional fixed pay points are difficult to adapt to users' personalized viewing habits. By establishing an adjustment foundation through a baseline offset ratio, and dynamically calculating threshold ranges using user behavior characteristics such as the average slope of dwell time decay, the frequency of bullet screen emotional polarity transitions, and the coverage width of speed-up operations, it ensures both the flexibility of pay point adjustments and avoids frequent reverse adjustments caused by abnormal fluctuations through a third proportional threshold constraint, thus improving the stability of the user experience.

[0066] Meanwhile, the prediction model is continuously optimized through feedback data from actual behavior, enabling the setting of paid unlock points to dynamically evolve with changes in user behavior, thus realizing the transformation of AI short drama content release scheduling from static configuration to dynamic intelligence.

[0067] In summary, the embodiments of this application have at least the following technical effects: This application provides an AI-powered short drama content release and scheduling method based on user behavior prediction. First, it acquires user viewing behavior stream data generated within the observation window for the current episode of the target short drama, reflecting user interaction and behavioral patterns during viewing. Second, the viewing behavior stream data is input into a trained behavior-payment timing predictor, which outputs the estimated churn probability and estimated payment trigger probability for users at different progress coordinates in the current episode, providing a deeper understanding of user behavior tendencies at different viewing stages from a data perspective. Then, based on the two probability values, a zero-crossing position is found on the progress axis, and combined with preset unlock point offset constraints, a dynamic payment unlock point correction offset is generated, fully considering the dynamic changes in user behavior and differences in content attractiveness. Finally, the dynamic payment unlock point correction offset is used to adjust the payment unlock position for unexposed users in the current episode, and the corresponding playback strategy configuration is updated. This achieves dynamic adjustment of the payment unlock point, enabling flexible optimization of the payment unlock position based on actual user behavior and real-time content feedback, thereby finding the optimal balance between user retention and paid conversion.

[0068] Through the above technical solution, this application achieves the technical effect of dynamically adjusting the location of the paid unlock point based on the user's real-time viewing behavior, thereby achieving the optimal balance between user retention and paid conversion.

[0069] Example 2, as Figure 2 As shown, based on the same inventive concept as the AI ​​short drama content publishing and scheduling method based on user behavior prediction provided in Embodiment 1, this application also provides an AI short drama content publishing and scheduling system based on user behavior prediction, including: Data acquisition module 11 is used to acquire user viewing behavior stream data generated within the observation window for the current episode of the target short drama; Model training module 12 is used to input the viewing behavior stream data into the trained behavior-payment timing predictor and output the estimated churn probability value and the estimated payment trigger probability value of the user at different progress coordinates in the current episode. The offset calculation module 13 is used to generate a dynamic payment unlock point correction offset based on the estimated loss probability value and the estimated payment trigger probability value, at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint conditions. The strategy update module 14 is used to adjust the paid unlock position for unexposed users in the current episode based on the dynamic paid unlock point correction offset, and update the corresponding playback strategy configuration.

[0070] In one embodiment, the data acquisition module 11 is specifically used for: The user viewing behavior stream data includes the dwell time decay trajectory, the start and end position pairs of speed-up operations, and the frequency of bullet screen emotional polarity transitions. The observation window is divided into multiple consecutive time slices. The ratio of the user's viewing time in each time slice to the length of the corresponding time slice is calculated. The ratios of adjacent time slices are connected sequentially to construct a viewing time decay trajectory. Based on the speed control command triggered by the user on the playback interface, record the first progress coordinate corresponding to the start time of speed increase and the second progress coordinate corresponding to the end time of speed increase; The position binary formed by the first progress coordinate and the second progress coordinate is used as the start and end position pair of the speed-up operation; Collect the bullet screen text stream within the observation window, extract bullet screen text segments at fixed time steps, and input them into the sentiment polarity classification network to obtain the sentiment polarity label for each time step; The frequency of emotional polarity transitions in bullet comments is obtained by counting the cumulative number of times the emotional polarity label changes from negative to positive between adjacent time steps. The user viewing behavior stream data is obtained by summarizing the attenuation trajectory of the dwell time, the start and end positions of the speed-up operation, and the frequency of the emotional polarity transition of the bullet comments.

[0071] Furthermore, in one embodiment of the application, the construction steps of the emotion polarity classification network include: A collection of bullet screen text samples generated during the playback of historical short dramas was collected, and each bullet screen text sample was labeled with an emotional polarity tag, wherein the emotional polarity tag includes positive emotional markers and negative emotional markers; The set of labeled bullet screen text samples is divided into a training set and a validation set; Construct a sentiment polarity classification network based on a convolutional neural network architecture; Using the barrage text samples in the training set as input and the corresponding sentiment polarity labels as supervision signals, the sentiment polarity classification network is trained in a supervised manner. The cross-entropy loss function is used to calculate the error between the classification output and the supervision signal, and the network weight parameters are updated through backpropagation. When the sentiment polarity classification network achieves a sentiment polarity classification accuracy exceeding a preset accuracy threshold on the validation set and the loss no longer decreases after multiple consecutive training rounds, training is stopped, and the trained sentiment polarity classification network is obtained.

[0072] Furthermore, in one embodiment of the application, the training steps of the behavior-payment timing predictor include: Collect a sample set of historical user viewing behavior streams generated within the observation window for similar short dramas. For each historical user viewing behavior stream sample, label it with historical churn event markers and historical payment event markers aligned with different progress coordinates to form a dual-supervised label set. Construct a time-series prediction network architecture that includes a shared feature extraction layer, a churn probability prediction branch, and a payment trigger prediction branch; The input of the shared feature extraction layer is used to receive user viewing behavior stream data, and the output of the shared feature extraction layer is connected to the input of the churn probability prediction branch and the payment trigger prediction branch, respectively. The output of the churn probability prediction branch is used to output the estimated churn probability value of the user at different progress coordinates in the current episode, and the output of the payment trigger prediction branch is used to output the estimated payment trigger probability value of the user at different progress coordinates in the current episode. Using the historical user viewing behavior stream sample set as input, and the historical churn event tags and historical payment event tags in the dual-supervised tag set as joint supervision signals, the time-series prediction network architecture is subjected to multi-task supervised training until the network converges, thus obtaining a trained behavior-payment timing predictor.

[0073] In one embodiment, the offset calculation module 13 is specifically used for: Calculate the difference between the estimated churn probability value and the estimated payment trigger probability value at different progress coordinates in the current episode, and scan along the progress axis at the positions where the sign of the difference changes from positive to negative or from negative to positive to obtain the zero-crossing positions; If the zero-crossing position is unique, the progress difference between the zero-crossing position and the current fixed-payment unlocking position is used as the initial offset. If the number of zero-crossing positions is greater than one, the zero-crossing position closest to the end time of the observation window is selected, and the progress difference between the selected zero-crossing position and the current fixed payment unlocking position is calculated as the initial offset. If the zero-crossing position does not exist, scan from front to back along the progress axis, and use the progress coordinate of the first time the estimated payment trigger probability value exceeds the estimated churn probability value as the reference, and calculate the progress difference between the reference and the current fixed payment unlock position as the initial offset. The initial offset is limited by a preset unlock point offset constraint to obtain the dynamic paid unlock point correction offset.

[0074] Furthermore, the preset unlocking point offset constraint includes: The absolute value of the dynamic paid unlock point correction offset shall not exceed the first proportional threshold of the total progress of the current episode. The paid unlock position after offset adjustment does not exceed the second proportional threshold range intercepted forward from the end point of the current episode's content progress; When the offset of the dynamic paid unlock point is corrected, if the directions of two adjacent adjustments are opposite and the magnitudes both exceed the third proportional threshold, then the current offset update is paused and the previous paid unlock position is maintained unchanged.

[0075] Furthermore, the calculation steps for the first proportional threshold, the second proportional threshold range, and the third proportional threshold include: Calculate the average percentage of the distance between the paid unlock position and the end point of the content progress for similar short dramas in the past, and use it as the benchmark offset percentage. The average decay slope of the decay trajectory of the duration of the broadcast history episodes of the target short drama within the observation window is obtained, and the first proportion threshold is calculated in combination with the benchmark offset ratio. The frequency of barrage emotional polarity transitions in the current episode of the target short drama within the observation window is collected. Combined with the baseline offset ratio, the upper and lower limits of the second proportional threshold interval are determined. For each unit increase in the frequency of barrage emotional polarity transitions, the upper and lower limits of the second proportional threshold interval decrease synchronously according to a fixed shrinkage step. The cumulative value of the coverage width between the first and second progress coordinates of all speed-up operations within the observation window of the current set is calculated, and the ratio of the cumulative value to the total progress of the current set is used as the third ratio threshold.

[0076] In one embodiment, the policy update module 14 is specifically used for: The current fixed payment unlock position is algebraically superimposed with the dynamic payment unlock point correction offset to obtain the candidate payment unlock position. Determine whether the candidate paid unlock position falls within the second proportional threshold range intercepted forward from the end point of the current episode's content progress; If the candidate payment unlock position falls within the second ratio threshold range, then the candidate payment unlock position is used as the updated payment unlock position. If the candidate paid unlock position exceeds the upper limit of the second proportional threshold range, then the upper limit of the second proportional threshold range will be used as the updated paid unlock position. If the candidate paid unlock position is lower than the lower limit of the second proportional threshold range, then the lower limit of the second proportional threshold range will be used as the updated paid unlock position. The updated paid unlock location is associated with the playback identifier of the current episode to generate an updated playback strategy configuration item; After the observation window ends, receive playback requests for the current episode from newly arrived unexposed users, and read the updated paid unlock position from the updated playback strategy configuration. The updated paid unlock position is sent to the player of unexposed users, and the player is controlled to trigger the paid interception interface when the playback progress reaches the updated paid unlock position.

[0077] Furthermore, after updating the corresponding playback strategy configuration, it also includes: Obtain actual viewing behavior feedback data of unexposed users at the new paid unlock location, wherein the actual viewing behavior feedback data includes the actual churn event occurrence progress coordinates and the actual payment event trigger progress coordinates within the new observation window; Based on the estimated churn probability and estimated payment trigger probability of users at different progress coordinates in the current set output by the behavior-payment timing predictor, obtain the peak progress coordinate of the estimated churn probability and the peak progress coordinate of the estimated payment trigger probability. Calculate the first deviation between the actual churn event occurrence progress coordinate and the peak progress coordinate of the estimated churn probability value, and the second deviation between the actual payment event trigger progress coordinate and the peak progress coordinate of the estimated payment trigger probability value; Based on the weighted result of the first deviation and the second deviation, adjust the output response offset of the churn probability prediction branch and the payment trigger prediction branch in the behavior-payment timing predictor. The adjusted output response offset is added to the calculation process of the estimated churn probability value and the estimated payment trigger probability value in subsequent episodes.

Claims

1. An AI-powered short drama content release and scheduling method based on user behavior prediction, characterized in that, The method includes: Acquire user viewing behavior stream data generated within the observation window for the current episode of the target short drama; The viewing behavior stream data is input into a trained behavior-payment timing predictor, which outputs the estimated churn probability and the estimated payment trigger probability of the user at different progress coordinates in the current episode. Based on the estimated churn probability value and the estimated payment trigger probability value, at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint, a dynamic payment unlock point correction offset is generated. The dynamic paid unlock point correction offset is used to adjust the paid unlock position for unexposed users in the current episode, and the corresponding playback strategy configuration is updated.

2. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 1, characterized in that, Obtain user viewing behavior stream data generated within the observation window for the current episode of the target short drama, including: The user viewing behavior stream data includes the dwell time decay trajectory, the start and end position pairs of speed-up operations, and the frequency of bullet screen emotional polarity transitions. The observation window is divided into multiple consecutive time slices. The ratio of the user's viewing time in each time slice to the length of the corresponding time slice is calculated. The ratios of adjacent time slices are connected sequentially to construct a viewing time decay trajectory. Based on the speed control command triggered by the user on the playback interface, record the first progress coordinate corresponding to the start time of speed increase and the second progress coordinate corresponding to the end time of speed increase; The position binary formed by the first progress coordinate and the second progress coordinate is used as the start and end position pair of the speed-up operation; Collect the bullet screen text stream within the observation window, extract bullet screen text segments at fixed time steps, and input them into the sentiment polarity classification network to obtain the sentiment polarity label for each time step; The frequency of emotional polarity transitions in bullet comments is obtained by counting the cumulative number of times the emotional polarity label changes from negative to positive between adjacent time steps. The user viewing behavior stream data is obtained by summarizing the attenuation trajectory of the dwell time, the start and end positions of the speed-up operation, and the frequency of the emotional polarity transition of the bullet comments.

3. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 2, characterized in that, The steps for constructing the sentiment polarity classification network include: A collection of bullet screen text samples generated during the playback of historical short dramas was collected, and each bullet screen text sample was labeled with an emotional polarity tag, wherein the emotional polarity tag includes positive emotional markers and negative emotional markers; The set of labeled bullet screen text samples is divided into a training set and a validation set; Construct a sentiment polarity classification network based on a convolutional neural network architecture; Using the barrage text samples in the training set as input and the corresponding sentiment polarity labels as supervision signals, the sentiment polarity classification network is trained in a supervised manner. The cross-entropy loss function is used to calculate the error between the classification output and the supervision signal, and the network weight parameters are updated through backpropagation. When the sentiment polarity classification network achieves a sentiment polarity classification accuracy exceeding a preset accuracy threshold on the validation set and the loss no longer decreases after multiple consecutive training rounds, training is stopped, and the trained sentiment polarity classification network is obtained.

4. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 1, characterized in that, The training steps for the behavior-payment timing predictor include: Collect a sample set of historical user viewing behavior streams generated within the observation window for similar short dramas. For each historical user viewing behavior stream sample, label it with historical churn event markers and historical payment event markers aligned with different progress coordinates to form a dual-supervised label set. Construct a time-series prediction network architecture that includes a shared feature extraction layer, a churn probability prediction branch, and a payment trigger prediction branch; The input of the shared feature extraction layer is used to receive user viewing behavior stream data, and the output of the shared feature extraction layer is connected to the input of the churn probability prediction branch and the payment trigger prediction branch, respectively. The output of the churn probability prediction branch is used to output the estimated churn probability value of the user at different progress coordinates in the current episode, and the output of the payment trigger prediction branch is used to output the estimated payment trigger probability value of the user at different progress coordinates in the current episode. Using the historical user viewing behavior stream sample set as input, and the historical churn event tags and historical payment event tags in the dual-supervised tag set as joint supervision signals, the time-series prediction network architecture is subjected to multi-task supervised training until the network converges, thus obtaining a trained behavior-payment timing predictor.

5. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 1, characterized in that, Based on the estimated churn probability value and the estimated payment trigger probability value, at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint, a dynamic payment unlock point correction offset is generated, including: Calculate the difference between the estimated churn probability value and the estimated payment trigger probability value at different progress coordinates in the current episode, and scan along the progress axis at the positions where the sign of the difference changes from positive to negative or from negative to positive to obtain the zero-crossing positions; If the zero-crossing position is unique, the progress difference between the zero-crossing position and the current fixed-payment unlocking position is used as the initial offset. If the number of zero-crossing positions is greater than one, the zero-crossing position closest to the end time of the observation window is selected, and the progress difference between the selected zero-crossing position and the current fixed payment unlocking position is calculated as the initial offset. If the zero-crossing position does not exist, scan from front to back along the progress axis, and use the progress coordinate of the first time the estimated payment trigger probability value exceeds the estimated churn probability value as the reference, and calculate the progress difference between the reference and the current fixed payment unlock position as the initial offset. The initial offset is limited by a preset unlock point offset constraint to obtain the dynamic paid unlock point correction offset.

6. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 1, characterized in that, The preset unlocking point offset constraint conditions include: The absolute value of the dynamic paid unlock point correction offset shall not exceed the first proportional threshold of the total progress of the current episode. The paid unlock position after offset adjustment does not exceed the second proportional threshold range intercepted forward from the end point of the current episode's content progress; When the offset of the dynamic paid unlock point is corrected, if the directions of two adjacent adjustments are opposite and the magnitudes both exceed the third proportional threshold, then the current offset update is paused and the previous paid unlock position is maintained unchanged.

7. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 6, characterized in that, The calculation steps for the first proportional threshold, the second proportional threshold range, and the third proportional threshold include: Calculate the average percentage of the distance between the paid unlock position and the end point of the content progress for similar short dramas in the past, and use it as the benchmark offset percentage. The average decay slope of the decay trajectory of the duration of the broadcast history episodes of the target short drama within the observation window is obtained, and the first proportion threshold is calculated in combination with the benchmark offset ratio. The frequency of barrage emotional polarity transitions in the current episode of the target short drama within the observation window is collected. Combined with the baseline offset ratio, the upper and lower limits of the second proportional threshold interval are determined. For each unit increase in the frequency of barrage emotional polarity transitions, the upper and lower limits of the second proportional threshold interval decrease synchronously according to a fixed shrinkage step. The cumulative value of the coverage width between the first and second progress coordinates of all speed-up operations within the observation window of the current set is calculated, and the ratio of the cumulative value to the total progress of the current set is used as the third ratio threshold.

8. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 1, characterized in that, Adjust the paid unlock position for unexposed users in the current episode using the dynamic paid unlock point correction offset, and update the corresponding playback strategy configuration, including: The current fixed payment unlock position is algebraically superimposed with the dynamic payment unlock point correction offset to obtain the candidate payment unlock position. Determine whether the candidate paid unlock position falls within the second proportional threshold range intercepted forward from the end point of the current episode's content progress; If the candidate payment unlock position falls within the second ratio threshold range, then the candidate payment unlock position is used as the updated payment unlock position. If the candidate paid unlock position exceeds the upper limit of the second proportional threshold range, then the upper limit of the second proportional threshold range will be used as the updated paid unlock position. If the candidate paid unlock position is lower than the lower limit of the second proportional threshold range, then the lower limit of the second proportional threshold range will be used as the updated paid unlock position. The updated paid unlock location is associated with the playback identifier of the current episode to generate an updated playback strategy configuration item; After the observation window ends, receive playback requests for the current episode from newly arrived unexposed users, and read the updated paid unlock position from the updated playback strategy configuration. The updated paid unlock position is sent to the player of unexposed users, and the player is controlled to trigger the paid interception interface when the playback progress reaches the updated paid unlock position.

9. The AI ​​short drama content release and scheduling method based on user behavior prediction according to claim 1, characterized in that, After updating the corresponding playback strategy configuration, it also includes: Obtain actual viewing behavior feedback data of unexposed users at the new paid unlock location, wherein the actual viewing behavior feedback data includes the actual churn event occurrence progress coordinates and the actual payment event trigger progress coordinates within the new observation window; Based on the estimated churn probability and estimated payment trigger probability of users at different progress coordinates in the current set output by the behavior-payment timing predictor, obtain the peak progress coordinate of the estimated churn probability and the peak progress coordinate of the estimated payment trigger probability. Calculate the first deviation between the actual churn event occurrence progress coordinate and the peak progress coordinate of the estimated churn probability value, and the second deviation between the actual payment event trigger progress coordinate and the peak progress coordinate of the estimated payment trigger probability value; Based on the weighted result of the first deviation and the second deviation, adjust the output response offset of the churn probability prediction branch and the payment trigger prediction branch in the behavior-payment timing predictor. The adjusted output response offset is added to the calculation process of the estimated churn probability value and the estimated payment trigger probability value in subsequent episodes.

10. An AI-powered short drama content publishing and scheduling system based on user behavior prediction, characterized in that: The method for scheduling the release of AI-powered short drama content based on user behavior prediction as described in any one of claims 1-9 includes: The data acquisition module is used to acquire user viewing behavior stream data generated within the observation window for the current episode of the target short drama. The model training module is used to input the viewing behavior stream data into the trained behavior-payment timing predictor and output the estimated churn probability value and the estimated payment trigger probability value of the user at different progress coordinates in the current episode. The offset calculation module is used to generate a dynamic payment unlock point correction offset based on the estimated churn probability value and the estimated payment trigger probability value at the zero-crossing position on the progress axis, combined with the preset unlock point offset constraint conditions. The strategy update module is used to adjust the paid unlock position for unexposed users in the current episode based on the dynamic paid unlock point correction offset, and update the corresponding playback strategy configuration.

Images