An artificial intelligence-based digital marketing global optimization scheduling method and system

Abstract

The application belongs to the technical field of digital marketing, and discloses a digital marketing global optimization scheduling method and system based on artificial intelligence. The method comprises the following steps: a dynamic response window model driven by a push event is established for each marketing channel, a cross-channel shadow effect calculation is triggered in real time after the channel executes the push, a shadow coefficient is calculated based on the response window half-life and the content form similarity, and the effective remaining time of other channels is reduced, and a shadow ability decay propagation mechanism is introduced; the window state after reduction is input into a time sequence scheduling decision model, a target push channel and a push time offset are output synchronously, and the next round of shadow calculation is triggered in a closed loop after delay execution. The application solves the problem of global response rate decline caused by attention competition in multi-channel push, and realizes the joint optimization of cross-channel scheduling and time selection.

Description

Technical Field

[0001] This invention relates to the field of digital marketing technology, and in particular to a digital marketing global optimization scheduling method and system based on artificial intelligence. Background Technology

[0002] In digital marketing practices, businesses typically push marketing content to target users through multiple channels, such as SMS, app push notifications, emails, and social media messages. To improve marketing effectiveness, existing technologies have developed various cross-channel scheduling solutions. These include rule-based prioritization and time scheduling of multi-channel marketing information to allocate limited channel reach resources, and the establishment of cross-channel predictive models to calculate the contribution weight and interaction effect of each channel, thereby optimizing budget allocation. These solutions have achieved cross-channel coordination to a certain extent.

[0003] However, existing technologies generally fail to address a fundamental technical flaw: when multiple channels push marketing content to the same user within a similar timeframe, a momentary attention competition effect arises between channels—once a user's attention is momentarily captured by a push from one channel, the responsiveness of other channels simultaneously declines. Existing rule-based scheduling systems only make instantaneous, single-point decisions, failing to maintain the temporally continuous channel response state and thus unable to predict future state changes. Existing cross-channel effect models focus on long-term macro-level attribution analysis at the budget allocation level, unable to quantify and respond to this momentary attention competition relationship in real time at the event-driven level. This deficiency of "event-level cross-channel attention competition being unable to be perceived and responded to in real time" leads to a continuous decline in the global response rate of existing systems in multi-channel intensive push scenarios, resulting in severe waste of channel resources.

[0004] Therefore, this invention proposes a digital marketing global optimization scheduling method and system based on artificial intelligence. Summary of the Invention

[0005] This invention provides a digital marketing global optimization scheduling method and system based on artificial intelligence. By using event-driven dynamic response windows and real-time calculation of cross-channel masking effects, it solves the problem that the instantaneous competition for attention during multi-channel push cannot be perceived and quantified in real time, realizes joint optimization of cross-channel scheduling and timing, and significantly improves the global cumulative response count.

[0006] This invention provides a digital marketing global optimization scheduling method based on artificial intelligence, comprising: A dynamic response window model is established for each of the multiple marketing channels. The dynamic response window model is updated by push events. The dynamic response window model includes the channel identifier, the response window half-life, and the current effective remaining time. The current effective remaining time decreases over time and with the occurrence of cross-channel blocking events. When any channel executes a push action, the cross-channel masking effect calculation step is triggered in real time: the channel that executes the push action is taken as the source channel, and all other channels except the source channel are taken as target channels. Based on the response window half-life of the source channel, the response window half-life of each target channel, and the content form similarity between the source channel and each target channel, the masking coefficient of the source channel on each target channel is calculated by a preset masking effect function, and the current effective remaining time of the corresponding target channel is reduced according to each masking coefficient obtained. When the current effective remaining time of a target channel is reduced due to a cross-channel blocking event, the blocking capability attenuation factor is calculated based on the ratio of the reduced current effective remaining time of the target channel to the total window length of the target channel. When the target channel triggers the cross-channel blocking effect calculation step for other channels as a source channel, the blocking capability attenuation factor is multiplied by the blocking coefficient of the target channel for other channels to obtain the attenuated blocking coefficient. The system obtains the marketing tasks to be pushed and the dynamic response window status of each channel after reduction at the current moment. The dynamic response window status of each channel after reduction is input into the trained time-series scheduling decision model. The time-series scheduling decision model takes maximizing the cumulative effective response times of all channels within the preset future time window as the optimization objective. It simultaneously outputs the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current moment. Using the push timing offset as a delay parameter, the marketing task to be pushed is sent to the target push channel for execution after the delay push period expires, and the cross-channel masking effect calculation step is triggered after execution.

[0007] Preferably, the initial value of the current effective remaining time is determined jointly by the response window half-life and a preset response probability threshold; the process of the current effective remaining time decreasing over time includes: Based on the response window opening time and response window half-life, the window closing curve is calculated using an exponential decay function. The window closing curve represents the decay process of the response probability over time. The response window opening time is the execution time of the most recent push action of the corresponding channel. The window is closed when the response probability decays to below a preset response probability threshold, and the total window length is the time between the window opening time and the window closing time. The current effective remaining time is the total window length minus the window's open time, and then minus the cumulative time reduced due to cross-channel obscuring events.

[0008] Preferably, the preset occlusion effect function is defined as: The masking coefficient is obtained by dividing the response window half-life of the source channel by the response window half-life of the target channel, and then multiplying it by the inter-channel attention competition factor. The inter-channel attention competition factor is determined based on the similarity of the content format between the source channel and the target channel. The higher the similarity of the content format, the greater the inter-channel attention competition factor. When the similarity of the content format is zero, the inter-channel attention competition factor is zero, so that there is no masking effect between channels with completely different formats.

[0009] Preferably, the current effective remaining time of the corresponding target channel is reduced according to each calculated occlusion coefficient, including: Multiply each masking coefficient by a preset unit masking duration to obtain the effective masking duration of each cross-channel masking event on the corresponding target channel. Subtract the corresponding effective blocking time from the current effective remaining time of each target channel. If the result of the subtraction is negative, set it to zero, so that the corresponding target channel loses its effective response capability at the current moment.

[0010] Preferably, when the current effective remaining time of a target channel is reduced to zero, the masking capability decay factor is zero, so that the channel that has lost its effective response capability no longer has a masking effect on other channels.

[0011] Preferably, after the real-time triggering of the cross-channel masking effect calculation step, a cross-channel synergy effect step is also included: Once any channel executes a push action, the channel that executed the push action will be designated as the source channel, and all other channels except the source channel will be designated as alternative collaboration channels. When the source channel and the alternative collaborative channel meet the preset collaborative conditions, the reduction operation in the cross-channel masking effect calculation step is not executed. Instead, the current effective remaining duration of the alternative collaborative channel is extended according to the preset collaborative effect function. The preset collaboration conditions are: the similarity of the content format between the source channel and the alternative collaboration channel is lower than the preset collaboration judgment threshold, and the correlation between the content of the source channel and the alternative collaboration channel is higher than the preset correlation judgment threshold.

[0012] Preferably, when the push timing offset is positive, during the delayed push period corresponding to the delayed push duration, the time series of the current effective remaining time of the target push channel is continuously monitored, and the decay acceleration of the current effective remaining time is calculated. The decay acceleration is compared with the historical average decay acceleration of the target push channel within a preset historical window to calculate the decay trend deviation. When the decay trend deviation exceeds the preset trend deviation threshold, it is determined that the response window of the target push channel is decaying at an accelerated rate. The current delayed push task is terminated, and the dynamic response window status of each channel is re-input into the timing scheduling decision model to obtain a new target push channel and a new push timing offset.

[0013] Preferably, the time-series scheduling decision model is built based on a constrained optimization network, and the training process of the constrained optimization network includes: Retrieve historical push records. Each historical push record contains the dynamic response window status of each channel at the time of the push, the actual push channel selected, the actual push timing, and the cumulative number of valid responses from all channels within the preset time window as a global response tag. The constrained optimization network is trained with the goal of maximizing the global response label, the dynamic response window state of each channel at the push time as input, and the push channel selection and push timing offset as joint output. The loss function of the constrained optimization network includes a structural penalty term, which is defined as follows: when the current effective remaining time of the target push channel selected by the constrained optimization network is lower than the preset minimum time threshold, a penalty value proportional to the difference between the current effective remaining time and the preset minimum time threshold is applied to the loss function, so as to force the constrained optimization network to learn a strategy to avoid pushing when the channel response capability is insufficient.

[0014] Preferably, the structure of the constrained optimization network includes: A shared feature extraction layer is used to encode the dynamic response window status of each channel into a channel-level feature vector. The channel-level feature vector contains information on the current effective remaining duration, response window half-life, and duration that has been opened for the corresponding channel. The channel selection branch takes the concatenated vector as input and outputs the probability distribution of each channel being selected as the target push channel. The concatenated vector is formed by concatenating the channel-level feature vectors of all channels. The timing prediction branch takes the concatenated vector and the intermediate layer features of the channel selection branch as joint inputs and outputs continuous values ​​as the push timing offset. The timing prediction branch's input depends on the intermediate layer features of the channel selection branch, making the prediction of the push timing offset constrained by the response window state of the selected target push channel, thus forming a coupled decision between channel selection and push timing.

[0015] This invention provides an artificial intelligence-based digital marketing global optimization scheduling system, comprising: The window maintenance module is used to establish dynamic response window models for multiple marketing channels. The dynamic response window models are updated by push events. The dynamic response window model includes the channel identifier, the response window half-life, and the current valid remaining time. The current valid remaining time decreases over time and with the occurrence of cross-channel blocking events. The masking calculation module is used to trigger the cross-channel masking effect calculation step in real time after any channel executes a push action: the channel that executes the push action is taken as the source channel, and all other channels except the source channel are taken as target channels. Based on the response window half-life of the source channel, the response window half-life of each target channel, and the content form similarity between the source channel and each target channel, the masking coefficient of the source channel on each target channel is calculated by a preset masking effect function, and the current effective remaining time of the corresponding target channel is reduced according to each calculated masking coefficient. The attenuation propagation module is used to calculate the attenuation factor of the occlusion capability when the current effective remaining time of a target channel is reduced due to a cross-channel occlusion event. The attenuation factor is calculated based on the ratio of the reduced current effective remaining time of the target channel to the total window length of the target channel. When the target channel triggers the cross-channel occlusion effect calculation step for other channels as a source channel, the attenuation factor of the occlusion capability is multiplied by the occlusion coefficient of the target channel for other channels to obtain the attenuated occlusion coefficient. The scheduling decision module is used to obtain the marketing tasks to be pushed and the dynamic response window status of each channel after reduction at the current time. The dynamic response window status of each channel after reduction is input into the trained time-series scheduling decision model. The time-series scheduling decision model takes maximizing the cumulative effective response times of all channels within a preset future time window as the optimization objective. It simultaneously outputs the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current time. The delayed execution module is used to send the marketing task to be pushed to the target push channel for execution after the delayed push period expires, using the push timing offset as the delay parameter, and triggering the execution of the masking calculation module.

[0016] The beneficial effects of this invention compared to existing technologies are as follows: By establishing a dynamic response window model driven by push events, this invention, for the first time, models cross-channel attention competition as a quantifiable "masking effect." After push notifications are sent to any channel, the masking coefficient is calculated in real time, and the current effective remaining time of other channels is reduced. At the same time, a masking capability decay propagation mechanism is introduced to synchronously reduce the masking impact of the weakened channel. Finally, the optimal push channel and delayed push duration are synchronously output based on the real-time remaining response capabilities of each channel through a time-series scheduling decision model. This fundamentally solves the problem of the continuous decline in global response rate caused by the inability to perceive and respond to event-level cross-channel attention competition in existing multi-channel push systems, and achieves joint global optimization of cross-channel scheduling and push timing selection.

[0017] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a digital marketing global optimization scheduling method based on artificial intelligence, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the interaction between the lifecycle and occlusion effect of the dynamic response window model in this embodiment of the invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] like Figure 1 and Figure 2 As shown, this invention provides an embodiment of an artificial intelligence-based digital marketing global optimization scheduling method, comprising: A dynamic response window model is established for each of the multiple marketing channels. The dynamic response window model is updated by push events. The dynamic response window model includes the channel identifier, the response window half-life, and the current effective remaining time. The current effective remaining time decreases over time and with the occurrence of cross-channel blocking events. When any channel executes a push action, the cross-channel masking effect calculation step is triggered in real time: the channel that executes the push action is taken as the source channel, and all other channels except the source channel are taken as target channels. Based on the response window half-life of the source channel, the response window half-life of each target channel, and the content form similarity between the source channel and each target channel, the masking coefficient of the source channel on each target channel is calculated by a preset masking effect function, and the current effective remaining time of the corresponding target channel is reduced according to each masking coefficient obtained. When the current effective remaining time of a target channel is reduced due to a cross-channel blocking event, the blocking capability attenuation factor is calculated based on the ratio of the reduced current effective remaining time of the target channel to the total window length of the target channel. When the target channel triggers the cross-channel blocking effect calculation step for other channels as a source channel, the blocking capability attenuation factor is multiplied by the blocking coefficient of the target channel for other channels to obtain the attenuated blocking coefficient. The system obtains the marketing tasks to be pushed and the dynamic response window status of each channel after reduction at the current moment. The dynamic response window status of each channel after reduction is input into the trained time-series scheduling decision model. The time-series scheduling decision model takes maximizing the cumulative effective response times of all channels within the preset future time window as the optimization objective. It simultaneously outputs the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current moment. Using the push timing offset as a delay parameter, the marketing task to be pushed is sent to the target push channel for execution after the delay push period expires, and the cross-channel masking effect calculation step is triggered after execution.

[0022] In this embodiment, marketing channels refer to the digital reach channels that enterprises use to push marketing content to target customers, such as SMS channels, app push channels, email channels, and social media messaging channels. Each marketing channel has independent communication protocols, push frequency limits, and user reach characteristics, which together constitute a multi-channel marketing system.

[0023] In this embodiment, establishing dynamic response window models for multiple marketing channels means constructing an independent data structure for each marketing channel to describe its remaining effective response capability at the current moment. The dynamic response window model serves as the unified foundation for subsequent cross-channel masking effect calculations and time-series scheduling decisions, ensuring that the state of each channel can be independently tracked and updated.

[0024] In this embodiment, the dynamic response window model is updated by push events. This means that the state changes of the dynamic response window model are not triggered by periodic polling, but are directly triggered by the execution event of the push action. When a channel executes a push action, the push event immediately triggers the state update of the dynamic response window model for that channel and all other channels.

[0025] In this embodiment, the channel identifier is a unique identification code for each marketing channel, used to distinguish different push channels among multiple channels. The response window half-life is a parameter characterizing the time required for the user response probability to decay to half its peak value after a channel push; different channels can be configured with different response window half-life values. The current effective remaining time is the remaining effective response capability of the channel from the current moment, in minutes. The initial value of the current effective remaining time is jointly determined by the response window half-life and a preset response probability threshold. When the response probability decays exponentially to below the preset response probability threshold, the window closes; the time between the window opening time and the window closing time is the initial value of the current effective remaining time.

[0026] In this embodiment, a cross-channel masking event refers to the instantaneous attention competition effect on other channels caused by a push action executed by any channel. This event is triggered by the push action of the source channel, resulting in a reduction of the target channel's current effective remaining time. Cross-channel masking events are one of the driving factors of the decreasing current effective remaining time in the dynamic response window model.

[0027] In this embodiment, the current effective remaining time decreases over time and with the occurrence of cross-channel blocking events, comprising two layers of reduction mechanisms. The first layer is natural decay: as time goes on, the current effective remaining time decreases continuously according to an exponential decay function, simulating the natural fading of user attention over time. The second layer is event-driven reduction: when push actions from other channels trigger cross-channel blocking events, the current effective remaining time is further reduced. The two reduction mechanisms work together to ensure that the current effective remaining time always reflects the latest remaining response capability of the channel.

[0028] In this embodiment, the execution of a push action by any channel refers to the completion of the operation of pushing marketing content to the target user by one of the channels in the multi-channel system. The execution time of the push action is the opening time of the response window of the channel's dynamic response window model, and it is also the starting event that triggers the calculation of the global cross-channel masking effect.

[0029] In this embodiment, the higher the similarity of content formats, the stronger the competition between the two channels for user attention resources.

[0030] In this embodiment, the larger the masking coefficient, the stronger the masking effect of the source channel on the target channel.

[0031] In this embodiment, the masking capability attenuation factor is calculated based on the ratio of the current effective remaining time of the target channel after reduction to the total window length of the target channel. The masking capability attenuation factor ranges from 0 to 1. When the current effective remaining time of the target channel equals the total window length, the masking capability attenuation factor is 1; when the current effective remaining time of the target channel is reduced to zero, the masking capability attenuation factor is 0. The masking capability attenuation factor reflects the proportion of the target channel's current remaining response capability to its maximum response capability.

[0032] In this embodiment, when the target channel, as a source channel, triggers the cross-channel masking effect calculation step for other channels, the masking capability attenuation factor is multiplied by the masking coefficient of the target channel for other channels to obtain the attenuated masking coefficient. This mechanism ensures that the masking effect of a target channel whose response capability has been severely weakened as a source channel on other channels is correspondingly reduced, avoiding the unrealistic outcome that a weakened channel would still exert the same intensity of masking effect on other channels.

[0033] In this embodiment, the marketing task to be pushed refers to a specific marketing content that the enterprise needs to push to target customers, including the marketing content itself, the target customer identifier, and the marketing content type identifier. The marketing task to be pushed is the input condition that triggers the timing scheduling decision.

[0034] In this embodiment, the current dynamic response window state of each channel after reduction refers to the latest set of values ​​for parameters such as the current effective remaining duration, response window half-life, and response window opening time of each channel after natural decay and the most recent round of cross-channel masking events, based on the dynamic response window models of all marketing channels at the current moment. This state is the direct input to the time-series scheduling decision model.

[0035] In this embodiment, the preset future time window refers to the time interval from the current moment to a specified future duration, such as the next 24 hours. This time window is the evaluation range for the time-series scheduling decision model to perform global optimization. When making a decision, the model comprehensively considers the possible response effects of all channels within this window.

[0036] In this embodiment, the cumulative number of valid responses across all channels refers to the total number of valid user responses obtained from push notifications across all marketing channels within a preset future time window. Valid response behaviors include preset target behaviors such as clicking on push content, completing a purchase, and filling out a form. The cumulative number of valid responses is the global optimization objective function value of the time-series scheduling decision model.

[0037] In this embodiment, the reduced dynamic response window state of each channel is input into a trained time-series scheduling decision model. The model aims to maximize the cumulative effective response count of all channels within a preset future time window, simultaneously outputting the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current moment, in minutes. Synchronous output means the model generates two decision variables simultaneously during a single inference process, rather than performing channel selection and timing decisions separately. A positive push timing offset indicates delayed push, while zero indicates immediate push. During inference, the model performs forward prediction based on the current remaining response capacity of each channel, selecting the channel and timing combination that maximizes the global cumulative effective response count when pushing at the current moment or at some future moment.

[0038] In this embodiment, the push timing offset is used as a delay parameter. After the delayed push duration expires, the marketing task to be pushed is sent to the target push channel for execution, and a cross-channel masking effect calculation step is triggered after execution. When the push timing offset is zero, the push is executed immediately and the masking calculation is triggered; when the push timing offset is positive, a delay timer is started, and the push is executed and the masking calculation is triggered when the delayed push duration expires. The cross-channel masking effect calculation step triggered after execution forms a closed loop, so that the impact of this push on other channels is incorporated into the input state of the next round of scheduling decisions in real time.

[0039] Furthermore, the initial value of the current effective remaining time is determined jointly by the response window half-life and the preset response probability threshold; the process by which the current effective remaining time decreases over time includes: Based on the response window opening time and response window half-life, the window closing curve is calculated using an exponential decay function. The window closing curve represents the decay process of the response probability over time. The response window opening time is the execution time of the most recent push action of the corresponding channel. The window is closed when the response probability decays to below a preset response probability threshold, and the total window length is the time between the window opening time and the window closing time. The current effective remaining time is the total window length minus the window's open time, and then minus the cumulative time reduced due to cross-channel obscuring events.

[0040] In this embodiment, the initial value of the current effective remaining duration refers to the starting value of the current effective remaining duration at the moment the response window just opens. This initial value is equal to the total window length, representing the maximum effective response capability duration of the channel when it is not affected by any natural attenuation or cross-channel blocking. After initialization, the current effective remaining duration decreases from the initial value over time and with the occurrence of cross-channel blocking events.

[0041] In this embodiment, the preset response probability threshold is a pre-defined probability threshold value, ranging from 0 to 1, for example, 0.1. The preset response probability threshold is used to determine whether the dynamic response window is closed. When the response probability decays to below this threshold, it is determined that the effective response capability of the channel has been exhausted, and the window is subsequently closed.

[0042] In this embodiment, the initial value of the current effective remaining time is determined jointly by the response window half-life and a preset response probability threshold. Specifically, an exponential decay function is constructed based on the response window half-life, and the time required for the response probability to decay from its peak value of 1 to the preset response probability threshold is calculated. This time length is the total window length, which is also the initial value of the current effective remaining time. The response window half-life determines the decay rate; a shorter half-life results in faster decay and a smaller initial value. The preset response probability threshold determines the decay termination condition; a lower threshold results in a larger total window length and a larger initial value.

[0043] In this embodiment, the exponential decay function is a mathematical function in which the value decreases exponentially over time, and its decay rate is controlled by the response window half-life. After each half-life, the response probability value decays to half of its original value.

[0044] In this embodiment, a window closing curve is calculated using an exponential decay function based on the response window opening time and the response window half-life. The window closing curve characterizes the decay process of the response probability over time. The response window opening time is the execution time of the most recent push action on the corresponding channel, at which time the response probability is at its peak of 1. Starting from the response window opening time, the response probability continuously decreases over time according to the exponential decay function, forming a curve that decreases from 1 to 0; this curve is the window closing curve. The response window half-life determines the steepness of the window closing curve; the shorter the half-life, the steeper the curve.

[0045] In this embodiment, the response probability refers to the estimated probability that a user will make an effective response to the push content from a certain channel, and the value ranges from 0 to 1. The response probability peaks at 1 when the response window is open, indicating that the user's willingness to respond is at its highest when the push is first executed; the response probability continuously decreases along the window closing curve over time, indicating that the user's attention naturally fades over time; when other channels trigger cross-channel blocking events, the response probability of the target channel decreases accordingly, which is manifested as an additional reduction in the current effective remaining time.

[0046] In this embodiment, the window is defined as closing when the response probability decays to below a preset response probability threshold. The duration from the opening of the response window to the closing of the window is defined as the total window length. The total window length is the total time that the channel takes from window opening to window closing under natural decay conditions, representing the maximum effective response time of the channel when it is not affected by any cross-channel occlusion events. The total window length is determined by two parameters: the response window half-life and the preset response probability threshold.

[0047] In this embodiment, the window open duration refers to the length of time that has elapsed from the moment the response window was opened to the current moment. The window open duration continues to increase over time, and when the window open duration equals the total window length, the current effective remaining duration decreases to zero simply due to natural decay.

[0048] In this embodiment, the cumulative duration reduced due to cross-channel blocking events refers to the sum of the effective blocking durations corresponding to all cross-channel blocking events experienced by the target channel from the moment the response window opens to the current moment. The effective blocking duration generated by each cross-channel blocking event is equal to the blocking coefficient multiplied by a preset unit blocking duration. This cumulative value continues to accumulate as cross-channel blocking events continue to occur.

[0049] In this embodiment, the current effective remaining time is the total window length minus the window's open time, and then minus the cumulative time reduced due to cross-channel blocking events. The portion of the total window length minus the window's open time reflects the remaining effective time of the channel under natural decay conditions; the further subtraction of the cumulative time reduced due to cross-channel blocking events reflects the final effective response time of the channel under the combined effects of natural decay and cross-channel blocking. When the calculation result is negative, the current effective remaining time is zero.

[0050] Furthermore, the preset occlusion effect function is defined as: The masking coefficient is obtained by dividing the response window half-life of the source channel by the response window half-life of the target channel, and then multiplying it by the inter-channel attention competition factor. The inter-channel attention competition factor is determined based on the similarity of the content format between the source channel and the target channel. The higher the similarity of the content format, the greater the inter-channel attention competition factor. When the similarity of the content format is zero, the inter-channel attention competition factor is zero, so that there is no masking effect between channels with completely different formats.

[0051] In this embodiment, the inter-channel attention competition factor is determined based on the content format similarity between the source channel and the target channel. The higher the content format similarity, the larger the inter-channel attention competition factor. The inter-channel attention competition factor is a coefficient ranging from 0 to 1, used to quantify the intensity of user attention competition between two channels due to similar content formats. When the content formats of the two channels are completely identical, the inter-channel attention competition factor takes the maximum value of 1, indicating that there is a perfect competition for user attention resources between the two channels; when the content formats of the two channels are completely different, the inter-channel attention competition factor takes the value of 0, indicating that there is no attention competition between the two channels and no masking effect occurs.

[0052] In this embodiment, the content format similarity between the source channel and the target channel refers to the degree of similarity between the content format of the marketing content pushed by the source channel and the content format of the marketing content typically carried by the target channel. Content formats include four presentation methods: text, images, videos, and audio. The content format similarity is calculated as follows: First, the content format of each channel is encoded into a 4-dimensional content format feature vector. The four dimensions correspond to the proportion of text, images, videos, and audio, respectively. The value of each dimension is the frequency of occurrence of that content format in the channel's historical content pushes, and the sum of the four dimensions is 1. For channels carrying multiple content formats, each dimension is assigned a value according to its usage frequency within that channel. Then, the cosine similarity between the content format feature vector of the source channel and the content format feature vector of the target channel is calculated. This cosine similarity is used as the content format similarity value, ranging from 0 to 1. When the content format feature vectors of the source channel and the target channel are completely identical, the content format similarity is 1; when their content format feature vectors are orthogonal, the content format similarity is 0.

[0053] Furthermore, based on each calculated masking coefficient, the current effective remaining duration of the corresponding target channel is reduced, including: Multiply each masking coefficient by a preset unit masking duration to obtain the effective masking duration of each cross-channel masking event on the corresponding target channel. Subtract the corresponding effective blocking time from the current effective remaining time of each target channel. If the result of the subtraction is negative, set it to zero, so that the corresponding target channel loses its effective response capability at the current moment.

[0054] In this embodiment, the preset unit shielding duration is a fixed time length set in advance, in minutes, such as 5 minutes. The preset unit shielding duration is used to convert the shielding coefficient into effective shielding duration, and is a conversion factor between the shielding coefficient and the actual reduction amount. When the shielding coefficient is 1, the effective shielding duration generated by a cross-channel shielding event is equal to the preset unit shielding duration; when the shielding coefficient is 0, the effective shielding duration is 0.

[0055] In this embodiment, the current effective remaining time of each target channel is subtracted from the corresponding effective blocking time. If the subtraction result is negative, it is set to zero, causing the corresponding target channel to lose its effective response capability at the current moment. This operation means that when the impact intensity of a cross-channel blocking event is large enough to exhaust the current effective remaining time of a target channel in one go, the target channel immediately enters a closed state, and its current effective remaining time is reduced to zero. A closed channel no longer has effective response capability and will not participate in scheduling decisions as a candidate push channel until it performs a push action again and reopens its response window. Simultaneously, when the current effective remaining time of a target channel is reduced to zero, its blocking capability decay factor is also set to zero, and the channel no longer has a blocking effect on other channels.

[0056] Furthermore, when the current effective remaining time of a target channel is reduced to zero, the masking capability decay factor is zero, so that the channel that has lost its effective response capability no longer has a masking effect on other channels.

[0057] In this embodiment, when the current effective remaining time of a target channel is reduced to zero, the obscuring ability decay factor is zero, ensuring that the channel, having lost its effective responsiveness, no longer obscures other channels. This boundary condition is designed based on the following logic: reducing the current effective remaining time to zero means that the channel has completely lost its effective responsiveness at the current moment, and the user's attention resources for that channel have been exhausted. In this state, even if the channel initiates a push activity to other channels as a source channel, it cannot have a substantial impact on the attention competition of other channels because it no longer possesses the ability to capture user attention. Setting the obscuring ability decay factor to zero blocks the propagation path of obscuring effects from closed channels, making the calculation of cross-channel obscuring effects more consistent with the true laws of user attention allocation.

[0058] Furthermore, after the real-time triggering of the cross-channel masking effect calculation step, a cross-channel synergy effect step is also included: Once any channel executes a push action, the channel that executed the push action will be designated as the source channel, and all other channels except the source channel will be designated as alternative collaboration channels. When the source channel and the alternative collaborative channel meet the preset collaborative conditions, the reduction operation in the cross-channel masking effect calculation step is not executed. Instead, the current effective remaining duration of the alternative collaborative channel is extended according to the preset collaborative effect function. The preset collaboration conditions are: the similarity of the content format between the source channel and the alternative collaboration channel is lower than the preset collaboration judgment threshold, and the correlation between the content of the source channel and the alternative collaboration channel is higher than the preset correlation judgment threshold.

[0059] In this embodiment, the preset synergy effect function is used to calculate the strength of the positive synergy effect generated by the source channel on the candidate synergy channel. The inputs to the preset synergy effect function are the response window half-life of the source channel, the response window half-life of the candidate synergy channel, and the content correlation degree between the source channel and the candidate synergy channel; the output is a synergy coefficient. The synergy coefficient is calculated by dividing the response window half-life of the source channel by the sum of the response window half-lives of the source channel and the candidate synergy channel, and then multiplying by the content correlation degree between the source channel and the candidate synergy channel. The larger the synergy coefficient, the stronger the positive synergy effect of the source channel on the candidate synergy channel.

[0060] In this embodiment, extending the current effective remaining time of the candidate collaboration channel according to a preset synergy effect function specifically includes: multiplying the synergy coefficient by a preset unit synergy duration to obtain the effective synergy duration; and adding the current effective remaining time of the candidate collaboration channel to the effective synergy duration. The preset unit synergy duration is a pre-set fixed time length in minutes, such as 5 minutes. When the synergy coefficient is 1, the effective synergy duration generated by a single cross-channel synergy effect is equal to the preset unit synergy duration; when the synergy coefficient is 0, the effective synergy duration is 0.

[0061] In this embodiment, the content format similarity between the source channel and the alternative collaborative channels refers to the degree of similarity between the content format of the marketing content pushed by the source channel and the content format of the marketing content typically carried by the alternative collaborative channels. Content formats include text, images, videos, audio, and other presentation methods. The content format similarity is calculated as follows: the content format types of the source channel and the alternative collaborative channels are respectively encoded into content format feature vectors; the cosine similarity between the two content format feature vectors is calculated; and this cosine similarity is used as the value of the content format similarity, ranging from 0 to 1.

[0062] In this embodiment, the preset collaboration judgment threshold is a pre-defined content format similarity threshold, ranging from 0 to 1, for example, 0.3. The preset collaboration judgment threshold is used to determine whether the source channel and the candidate collaboration channel meet the first condition for triggering a collaboration effect. When the content format similarity between the source channel and the candidate collaboration channel is lower than the preset collaboration judgment threshold, it indicates that the content format differences between the two channels are sufficiently large, there is no direct attention competition, and the prerequisite for generating a collaboration effect is met. When the content format similarity is higher than or equal to the preset collaboration judgment threshold, the two channels are still primarily in a competitive relationship, and no collaboration effect is triggered.

[0063] In this embodiment, the content relevance between the source channel and the alternative collaborative channels refers to the degree of complementarity between the marketing content pushed by the source channel and the marketing content typically carried by the alternative collaborative channels in terms of theme, scenario, or function. The content relevance is calculated as follows: First, a pre-defined three-level content theme tag classification system is constructed. The first-level classification includes four main categories: product promotion, brand promotion, user care, and event notifications. The second-level classification is further subdivided under the first-level classification, and the third-level classification is further subdivided into specific product categories or service types. Each channel is assigned one or more third-level content theme tags based on the main categories of its historically pushed content. Then, the content theme tags of each channel are encoded into content theme feature vectors. The vector dimension is equal to the sum of the number of all third-level categories in the three-level classification system, and the value of each dimension is the frequency of content pushed by the channel under that third-level category. Finally, the cosine similarity between the content theme feature vectors of the source channel and the content theme feature vectors of the alternative collaborative channels is calculated. This cosine similarity is used as the content relevance value, ranging from 0 to 1. The higher the content relevance, the easier it is for the content of the two channels to generate positive associations and complementary effects in the user's perception.

[0064] In this embodiment, the preset association determination threshold is a pre-set content association degree threshold, ranging from 0 to 1, for example, 0.6. The preset association determination threshold is used to determine whether the source channel and the alternative collaborative channel meet the second condition for triggering a synergistic effect. When the content association degree between the source channel and the alternative collaborative channel is higher than the preset association determination threshold, it indicates that the content of the two channels has a sufficiently strong complementary relationship, and the user's willingness to respond to the alternative collaborative channel after being exposed to the content of the source channel will significantly increase; when the content association degree is lower than or equal to the preset association determination threshold, the complementary relationship between the two channels is insufficient to trigger a synergistic effect.

[0065] Furthermore, when the push timing offset is positive, during the delayed push period corresponding to the delayed push duration, the time series of the current effective remaining time of the target push channel is continuously monitored, and the decay acceleration of the current effective remaining time is calculated. The decay acceleration is compared with the historical average decay acceleration of the target push channel within a preset historical window to calculate the decay trend deviation. When the decay trend deviation exceeds the preset trend deviation threshold, it is determined that the response window of the target push channel is decaying at an accelerated rate. The current delayed push task is terminated, and the dynamic response window status of each channel is re-input into the timing scheduling decision model to obtain a new target push channel and a new push timing offset.

[0066] In this embodiment, the time series of the current effective remaining time of the target push channel is continuously monitored, and the decay acceleration of the current effective remaining time is calculated. The decay acceleration is calculated as follows: during the delayed push period, the current effective remaining time value of the target push channel is collected at fixed time intervals to form a time series; the decay rate of the current effective remaining time between two adjacent collection times is calculated; then the difference between two adjacent decay rates is calculated, and this difference divided by the collection time interval is the decay acceleration. The decay acceleration reflects the changing trend of the decay rate of the current effective remaining time. When the decay acceleration is negative, it indicates that the decay rate is accelerating, and the response window of the target push channel is deteriorating rapidly.

[0067] In this embodiment, the preset historical window is a pre-defined backtracking time interval, such as the past 30 days. The preset historical window is used to statistically analyze the average decay acceleration of the target push channel under normal historical conditions, serving as a benchmark for judging whether the current decay trend is abnormal.

[0068] In this embodiment, the historical average decay acceleration of the target push channel within a preset historical window refers to the average decay acceleration of the target push channel in each response window when it was in a similar remaining time interval to the current time, within the time range of the preset historical window tracing back from the current moment. The similar remaining time interval refers to a time range within which the current effective remaining time fluctuates by a preset percentage, such as 10%. The historical average decay acceleration represents the baseline level of the target push channel's acceleration under normal decay mode.

[0069] In this embodiment, the attenuation acceleration is compared with the historical average attenuation acceleration of the target push channel within a preset historical window to calculate the attenuation trend deviation. The attenuation trend deviation is calculated as follows: the difference between the current attenuation acceleration and the historical average attenuation acceleration is calculated, the absolute value of this difference is taken, and this absolute value is divided by the absolute value of the historical average attenuation acceleration to obtain the attenuation trend deviation. A larger attenuation trend deviation indicates a more severe deviation between the current attenuation mode and the historical normal attenuation mode.

[0070] In this embodiment, the preset trend deviation threshold is a pre-set attenuation trend deviation threshold, such as 50%. When the attenuation trend deviation exceeds the preset trend deviation threshold, it is determined that the attenuation pattern of the current effective remaining time of the target push channel has significantly deviated from the historical normal pattern, and the response window is experiencing abnormal accelerated attenuation. Continuing to wait for the delayed push time to expire before executing the push will face the risk of the response window closing prematurely.

[0071] In this embodiment, the current delayed push task is terminated, and the dynamic response window status of each channel is re-input into the timing scheduling decision model to obtain a new target push channel and a new push timing offset. When it is determined that the response window of the target push channel is decaying rapidly, the system immediately cancels the current delayed push task without waiting for the original delayed push duration to expire. Then, the latest dynamic response window status of all channels at the current moment is collected, including the status after natural decay during the delayed push period and the impact of possible new cross-channel blocking events, and these statuses are re-input into the timing scheduling decision model. The timing scheduling decision model re-infers based on the latest global window status and outputs a new target push channel and a new push timing offset. The new push timing offset may be zero, indicating immediate push; or it may be a positive value, indicating entering a new round of delayed push. This mechanism ensures that when the channel status deteriorates abnormally, the system can adjust the scheduling scheme in a timely manner to avoid the target push channel losing its effective response capability when the original delayed push duration expires.

[0072] Furthermore, the time-series scheduling decision model is built based on a constrained optimization network. The training process of the constrained optimization network includes: Retrieve historical push records. Each historical push record contains the dynamic response window status of each channel at the time of the push, the actual push channel selected, the actual push timing, and the cumulative number of valid responses from all channels within the preset time window as a global response tag. The constrained optimization network is trained with the goal of maximizing the global response label, the dynamic response window state of each channel at the push time as input, and the push channel selection and push timing offset as joint output. The loss function of the constrained optimization network includes a structural penalty term, which is defined as follows: when the current effective remaining time of the target push channel selected by the constrained optimization network is lower than the preset minimum time threshold, a penalty value proportional to the difference between the current effective remaining time and the preset minimum time threshold is applied to the loss function, so as to force the constrained optimization network to learn a strategy to avoid pushing when the channel response capability is insufficient.

[0073] In this embodiment, the timing scheduling decision model is built based on a constrained optimization network. A constrained optimization network is a deep neural network that introduces constraints during training. Its network structure includes a shared feature extraction layer, a channel selection branch, and a timing prediction branch. The shared feature extraction layer uses a fully connected neural network to encode the dynamic response window state of each channel into a fixed-dimensional channel-level feature vector. The channel selection branch uses a Softmax classifier, taking the concatenated vector (composed of all channel-level feature vectors) as input and outputting the probability distribution of each channel being selected as the target push channel. The timing prediction branch uses a fully connected regression network, taking the concatenated vector and the intermediate layer features of the channel selection branch as joint input and outputting a continuous value as the push timing offset. The input of the timing prediction branch depends on the intermediate layer features of the channel selection branch, forming a coupled decision between channel selection and push timing.

[0074] In this embodiment, historical push records refer to complete records of push tasks actually executed by the system over a past period. Each historical push record corresponds to a real push scheduling execution process, recording all key information from scheduling decisions to execution completion, and serves as the source of sample data required for training the constraint optimization network. The collection time range of historical push records typically covers at least 30 days of system operation to ensure the diversity and representativeness of the samples.

[0075] In this embodiment, each historical push record includes the dynamic response window status of each channel at the time of push, the actually selected push channel, the actual push timing, and the cumulative number of valid responses from all channels within a preset time window as a global response tag. The dynamic response window status of each channel at the time of push refers to the values ​​of parameters such as the current remaining effective duration, response window half-life, and active duration for all channels when the scheduling decision corresponding to that record is made. The actually selected push channel and the actual push timing are the execution results of the historical scheduling decision. The global response tag refers to the cumulative number of valid responses obtained by all channels from the actual push timing until the end of the preset time window due to this and subsequent push actions. The global response tag reflects the true global effect of this scheduling decision within the entire time window.

[0076] In this embodiment, the optimization objective is to maximize the global response label. The dynamic response window state of each channel at the push time is used as input, and the push channel selection and push timing offset are used as joint outputs to train the constrained optimization network. The training process employs a batch gradient descent optimization algorithm, with each batch of training samples containing multiple historical push records. During training, the dynamic response window state of each channel at the push time in each historical push record is input into the constrained optimization network. The network's forward propagation outputs the predicted push channel probability distribution and push timing offset. The prediction results are compared with the actual push channels and actual push times in the historical records to calculate the basic loss. The basic loss includes the cross-entropy loss of the channel selection branch and the mean squared error loss of the timing prediction branch. The sum of the basic loss and the structural penalty term constitutes the total loss, which is used to update the weight parameters of the constrained optimization network through backpropagation. Training is iterated multiple times until the total loss converges.

[0077] In this embodiment, the preset minimum duration threshold is a pre-defined critical value for the current effective remaining duration, in minutes, such as 10 minutes. The preset minimum duration threshold is used to define a state where the channel's response capability is insufficient. When the current effective remaining duration of the target push channel selected by the constraint optimization network is lower than this threshold, it indicates that the channel's remaining response capability is already at a dangerous level. Forcing a push at this point may result in an actual response rate far lower than expected.

[0078] In this embodiment, the loss function of the constrained optimization network includes a structural penalty term, defined as follows: when the current effective remaining time of the target push channel selected by the constrained optimization network is lower than a preset minimum time threshold, a penalty value proportional to the difference between the current effective remaining time and the preset minimum time threshold is applied to the loss function. This forces the constrained optimization network to learn a strategy to avoid forcibly pushing when the channel's response capacity is insufficient. The penalty value is calculated by subtracting the current effective remaining time from the preset minimum time threshold, and multiplying the difference by a preset penalty coefficient. The lower the current effective remaining time, the larger the difference, and the larger the penalty value. During training, when the constrained optimization network attempts to select a channel whose response capacity is nearly exhausted as the target push channel, the structural penalty term significantly increases the total loss, forcing the network parameters to update in the direction of avoiding such selections through gradient backpropagation. After sufficient training, the constrained optimization network can automatically learn a strategy of prioritizing push when the channel's response capacity is sufficient and selecting other channels or delaying push when the channel's response capacity is insufficient.

[0079] Furthermore, the structure of the constrained optimization network includes: A shared feature extraction layer is used to encode the dynamic response window status of each channel into a channel-level feature vector. The channel-level feature vector contains information on the current effective remaining duration, response window half-life, and duration that has been opened for the corresponding channel. The channel selection branch takes the concatenated vector as input and outputs the probability distribution of each channel being selected as the target push channel. The concatenated vector is formed by concatenating the channel-level feature vectors of all channels. The timing prediction branch takes the concatenated vector and the intermediate layer features of the channel selection branch as joint inputs and outputs continuous values ​​as the push timing offset. The timing prediction branch's input depends on the intermediate layer features of the channel selection branch, making the prediction of the push timing offset constrained by the response window state of the selected target push channel, thus forming a coupled decision between channel selection and push timing.

[0080] In this embodiment, the dynamic response window state of each channel is encoded into a channel-level feature vector. This vector contains information about the current remaining effective duration, response window half-life, and active duration of the corresponding channel. The encoding process is performed by a shared feature extraction layer. This shared feature extraction layer is a fully connected neural network. The input layer receives the three values ​​for each channel: the current remaining effective duration, response window half-life, and active duration. The input passes through two fully connected hidden layers, each containing 128 neurons, using the ReLU activation function. The output layer outputs a fixed-dimensional channel-level feature vector with 64 dimensions. All channels share the parameters of the same shared feature extraction layer, ensuring that the states of different channels are mapped to the same feature space.

[0081] In this embodiment, the "activated duration" information refers to the length of time elapsed from the moment the response window opens to the current moment for the corresponding channel, measured in minutes. The activated duration information reflects the time interval since the channel's most recent push notification. A shorter activated duration indicates that the channel has just performed a push notification, and user attention is still high; a longer activated duration indicates that the channel has not pushed notifications for a long time, and user attention has significantly decreased. The activated duration information, used in conjunction with the current effective remaining duration, enables the model to distinguish between two different states: "having just pushed notifications and having ample remaining duration" and "having not pushed notifications for a long time and having naturally decaying remaining duration."

[0082] In this embodiment, the splicing vector refers to a one-dimensional vector formed by concatenating the channel-level feature vectors of all channels end-to-end according to a preset channel order. Assuming there are 5 channels, and each channel's channel-level feature vector has 64 dimensions, the total dimension of the splicing vector is 320 dimensions. The preset channel order is fixed during system initialization to ensure that the feature vectors of each channel are in the same position within the splicing vector during each splicing process.

[0083] In this embodiment, the channel selection branch takes the concatenated vector as input and outputs the probability distribution of each channel being selected as the target push channel. The channel selection branch employs a Softmax classifier structure, containing one fully connected hidden layer and one Softmax output layer. The fully connected hidden layer contains 64 neurons, using the ReLU activation function to map the concatenated vector to intermediate layer features. The Softmax output layer contains the same number of neurons as the channels, converting the intermediate layer features into a probability value for each channel being selected, with the sum of all probabilities being 1. The channel with the highest probability value in the probability distribution output by the channel selection branch is the target push channel predicted by the model.

[0084] In this embodiment, the timing prediction branch takes the concatenated vector and the intermediate layer features of the channel selection branch as joint inputs, and outputs a continuous value as the push timing offset. The timing prediction branch adopts a fully connected regression network structure, containing two fully connected hidden layers. The first hidden layer receives the concatenation result of the concatenated vector and the intermediate layer features of the channel selection branch, and contains 128 neurons, with the ReLU activation function. The second hidden layer contains 64 neurons, with the ReLU activation function. The output layer contains one neuron, has no activation function, and directly outputs a continuous value as the push timing offset, in minutes. The push timing offset ranges from non-negative real numbers, with 0 indicating immediate push and positive values ​​indicating delayed push duration.

[0085] In this embodiment, the input of the timing prediction branch depends on the intermediate layer features of the channel selection branch, making the prediction of the push timing offset constrained by the response window state of the selected target push channel, thus forming a coupled decision between channel selection and push timing. The intermediate layer features of the channel selection branch already contain information highly relevant to the channel selection decision, especially the feature expression of the channel that the model tends to select in the current window state. By accessing these intermediate layer features, the timing prediction branch can perceive which target push channel is most likely to be selected and what the response window state of that channel is, thereby fully considering whether the current effective remaining time of the channel is sufficient, whether a delay in push is needed, and how long the delay is most appropriate when predicting the push timing offset. If the two branches are independent of each other, the timing prediction branch cannot know which channel will be selected for push, which may result in a large delay push duration for channels whose response capacity is almost exhausted; the coupled decision structure ensures that the push timing offset always adapts to the response window state of the target push channel.

[0086] This invention provides an embodiment of an artificial intelligence-based digital marketing global optimization scheduling system, comprising: The window maintenance module is used to establish dynamic response window models for multiple marketing channels. The dynamic response window models are updated by push events. The dynamic response window model includes the channel identifier, the response window half-life, and the current valid remaining time. The current valid remaining time decreases over time and with the occurrence of cross-channel blocking events. The masking calculation module is used to trigger the cross-channel masking effect calculation step in real time after any channel executes a push action: the channel that executes the push action is taken as the source channel, and all other channels except the source channel are taken as target channels. Based on the response window half-life of the source channel, the response window half-life of each target channel, and the content form similarity between the source channel and each target channel, the masking coefficient of the source channel on each target channel is calculated by a preset masking effect function, and the current effective remaining time of the corresponding target channel is reduced according to each calculated masking coefficient. The attenuation propagation module is used to calculate the attenuation factor of the occlusion capability when the current effective remaining time of a target channel is reduced due to a cross-channel occlusion event. The attenuation factor is calculated based on the ratio of the reduced current effective remaining time of the target channel to the total window length of the target channel. When the target channel triggers the cross-channel occlusion effect calculation step for other channels as a source channel, the attenuation factor of the occlusion capability is multiplied by the occlusion coefficient of the target channel for other channels to obtain the attenuated occlusion coefficient. The scheduling decision module is used to obtain the marketing tasks to be pushed and the dynamic response window status of each channel after reduction at the current time. The dynamic response window status of each channel after reduction is input into the trained time-series scheduling decision model. The time-series scheduling decision model takes maximizing the cumulative effective response times of all channels within a preset future time window as the optimization objective. It simultaneously outputs the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current time. The delayed execution module is used to send the marketing task to be pushed to the target push channel for execution after the delayed push period expires, using the push timing offset as the delay parameter, and triggering the execution of the masking calculation module.

[0087] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A digital marketing global optimization scheduling method based on artificial intelligence, characterized in that, include: Establish dynamic response window models for multiple marketing channels, and update the dynamic response window models driven by push events; The dynamic response window model includes a channel identifier, a response window half-life, and the current effective remaining time; the current effective remaining time decreases over time and with the occurrence of cross-channel blocking events. When any channel executes a push action, the cross-channel masking effect calculation step is triggered in real time: the channel that executes the push action is taken as the source channel, and all other channels except the source channel are taken as target channels. Based on the response window half-life of the source channel, the response window half-life of each target channel, and the content form similarity between the source channel and each target channel, the masking coefficient of the source channel on each target channel is calculated by a preset masking effect function, and the current effective remaining time of the corresponding target channel is reduced according to each masking coefficient obtained. When the current effective remaining time of a target channel is reduced due to a cross-channel blocking event, the blocking capability attenuation factor is calculated based on the ratio of the reduced current effective remaining time of the target channel to the total window length of the target channel. When the target channel triggers the cross-channel blocking effect calculation step for other channels as a source channel, the blocking capability attenuation factor is multiplied by the blocking coefficient of the target channel for other channels to obtain the attenuated blocking coefficient. The system obtains the marketing tasks to be pushed and the dynamic response window status of each channel after reduction at the current moment. The dynamic response window status of each channel after reduction is input into the trained time-series scheduling decision model. The time-series scheduling decision model takes maximizing the cumulative effective response times of all channels within the preset future time window as the optimization objective. It simultaneously outputs the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current moment. Using the push timing offset as a delay parameter, the marketing task to be pushed is sent to the target push channel for execution after the delay push period expires, and the cross-channel masking effect calculation step is triggered after execution.

2. The AI-based digital marketing global optimization scheduling method according to claim 1, characterized in that, The initial value of the current effective remaining time is determined by the response window half-life and the preset response probability threshold; the process of the current effective remaining time decreasing over time includes: Based on the response window opening time and response window half-life, the window closing curve is calculated using an exponential decay function. The window closing curve represents the decay process of the response probability over time. The response window opening time is the execution time of the most recent push action of the corresponding channel. The window is closed when the response probability decays to below a preset response probability threshold, and the total window length is the time between the window opening time and the window closing time. The current effective remaining time is the total window length minus the window's open time, and then minus the cumulative time reduced due to cross-channel obscuring events.

3. The AI-based digital marketing global optimization scheduling method according to claim 1, characterized in that, The preset occlusion effect function is defined as follows: The masking coefficient is obtained by dividing the response window half-life of the source channel by the response window half-life of the target channel, and then multiplying it by the inter-channel attention competition factor. The inter-channel attention competition factor is determined based on the similarity of the content format between the source channel and the target channel. The higher the similarity of the content format, the greater the inter-channel attention competition factor. When the similarity of the content format is zero, the inter-channel attention competition factor is zero, so that there is no masking effect between channels with completely different formats.

4. The AI-based digital marketing global optimization scheduling method according to claim 3, characterized in that, Based on each calculated masking coefficient, the current effective remaining time of the corresponding target channel is reduced, including: Multiply each masking coefficient by a preset unit masking duration to obtain the effective masking duration of each cross-channel masking event on the corresponding target channel. Subtract the corresponding effective blocking time from the current effective remaining time of each target channel. If the result of the subtraction is negative, set it to zero, so that the corresponding target channel loses its effective response capability at the current moment.

5. The AI-based digital marketing global optimization scheduling method according to claim 1, characterized in that, When the current effective remaining time of a target channel is reduced to zero, the masking capability decay factor is zero, so that the channel that has lost its effective response capability no longer has a masking effect on other channels.

6. The AI-based digital marketing global optimization scheduling method according to claim 1, characterized in that, Following the real-time triggering of the cross-channel masking effect calculation step, a cross-channel synergy effect step is also included: Once any channel executes a push action, the channel that executed the push action will be designated as the source channel, and all other channels except the source channel will be designated as alternative collaboration channels. When the source channel and the alternative collaborative channel meet the preset collaborative conditions, the reduction operation in the cross-channel masking effect calculation step is not executed. Instead, the current effective remaining duration of the alternative collaborative channel is extended according to the preset collaborative effect function. The preset collaboration conditions are: the similarity of the content format between the source channel and the alternative collaboration channel is lower than the preset collaboration judgment threshold, and the correlation between the content of the source channel and the alternative collaboration channel is higher than the preset correlation judgment threshold.

7. The AI-based digital marketing global optimization scheduling method according to claim 1, characterized in that, When the push timing offset is positive, during the delayed push period corresponding to the delayed push duration, the time series of the current effective remaining time of the target push channel is continuously monitored, and the decay acceleration of the current effective remaining time is calculated. The decay acceleration is compared with the historical average decay acceleration of the target push channel within a preset historical window to calculate the decay trend deviation. When the decay trend deviation exceeds the preset trend deviation threshold, it is determined that the response window of the target push channel is decaying at an accelerated rate. The current delayed push task is terminated, and the dynamic response window status of each channel is re-input into the timing scheduling decision model to obtain a new target push channel and a new push timing offset.

8. The AI-based digital marketing global optimization scheduling method according to claim 1, characterized in that, The time-series scheduling decision model is built upon a constrained optimization network. The training process of the constrained optimization network includes: Retrieve historical push records. Each historical push record contains the dynamic response window status of each channel at the time of the push, the actual push channel selected, the actual push timing, and the cumulative number of valid responses from all channels within the preset time window as a global response tag. The constrained optimization network is trained with the goal of maximizing the global response label, the dynamic response window state of each channel at the push time as input, and the push channel selection and push timing offset as joint output. The loss function of the constrained optimization network includes a structural penalty term, which is defined as follows: when the current effective remaining time of the target push channel selected by the constrained optimization network is lower than the preset minimum time threshold, a penalty value proportional to the difference between the current effective remaining time and the preset minimum time threshold is applied to the loss function, so as to force the constrained optimization network to learn a strategy to avoid pushing when the channel response capability is insufficient.

9. The AI-based digital marketing global optimization scheduling method according to claim 8, characterized in that, The structure of a constrained optimization network includes: A shared feature extraction layer is used to encode the dynamic response window status of each channel into a channel-level feature vector. The channel-level feature vector contains information on the current effective remaining duration, response window half-life, and duration that has been opened for the corresponding channel. The channel selection branch takes the concatenated vector as input and outputs the probability distribution of each channel being selected as the target push channel. The concatenated vector is formed by concatenating the channel-level feature vectors of all channels. The timing prediction branch takes the concatenated vector and the intermediate layer features of the channel selection branch as joint inputs and outputs continuous values ​​as the push timing offset. The timing prediction branch's input depends on the intermediate layer features of the channel selection branch, making the prediction of the push timing offset constrained by the response window state of the selected target push channel, thus forming a coupled decision between channel selection and push timing.

10. A digital marketing global optimization and scheduling system based on artificial intelligence, characterized in that, include: The window maintenance module is used to create dynamic response window models for multiple marketing channels. The dynamic response window models are updated by push events. The dynamic response window model includes a channel identifier, a response window half-life, and the current effective remaining time; the current effective remaining time decreases over time and with the occurrence of cross-channel blocking events. The masking calculation module is used to trigger the cross-channel masking effect calculation step in real time after any channel executes a push action: the channel that executes the push action is taken as the source channel, and all other channels except the source channel are taken as target channels. Based on the response window half-life of the source channel, the response window half-life of each target channel, and the content form similarity between the source channel and each target channel, the masking coefficient of the source channel on each target channel is calculated by a preset masking effect function, and the current effective remaining time of the corresponding target channel is reduced according to each calculated masking coefficient. The attenuation propagation module is used to calculate the attenuation factor of the occlusion capability when the current effective remaining time of a target channel is reduced due to a cross-channel occlusion event. The attenuation factor is calculated based on the ratio of the reduced current effective remaining time of the target channel to the total window length of the target channel. When the target channel triggers the cross-channel occlusion effect calculation step for other channels as a source channel, the attenuation factor of the occlusion capability is multiplied by the occlusion coefficient of the target channel for other channels to obtain the attenuated occlusion coefficient. The scheduling decision module is used to obtain the marketing tasks to be pushed and the dynamic response window status of each channel after reduction at the current time. The dynamic response window status of each channel after reduction is input into the trained time-series scheduling decision model. The time-series scheduling decision model takes maximizing the cumulative effective response times of all channels within a preset future time window as the optimization objective. It simultaneously outputs the target push channel and push timing offset. The push timing offset represents the delayed push duration relative to the current time. The delayed execution module is used to send the marketing task to be pushed to the target push channel for execution after the delayed push period expires, using the push timing offset as the delay parameter, and triggering the execution of the masking calculation module.

Images