Advertisement putting optimization method and system based on AI prediction of media data trend

Abstract

The application discloses an advertisement putting optimization method and system based on AI prediction of media data trends, relates to the technical field of advertisement putting optimization, and aims at the problem of media trend misjudgment caused by historical best-selling residual context in a live broadcast scene.The method collects multi-modal data to generate a sold-out best-selling semantic vestige chain after identifying the exit of a sold-out best-selling product, executes a migration stripping optimization algorithm, cooperates with multi-dimensional parameters to construct a fitness function, divides live broadcast segments into a best-selling residual set and a current commodity bearing set, constructs an original media trend vector and executes a mask operation, strips a shell trend migration part to generate a shell cut-off type net media trend, and extracts a scalar parameter representing net intensity.Finally, a blocking instruction is issued to release a budget for the residual set, the budget is injected into the bearing set, and the scalar parameter is combined to generate a real-time bidding bid.The application realizes accurate decoupling of real transaction intention and residual semantics, and avoids invalid consumption of an advertisement budget.

Description

Technical Field

[0001] This invention relates to the field of advertising optimization technology, specifically to an advertising optimization method and system based on AI-predicted media data trends. Background Technology

[0002] With the rapid popularization of live-stream e-commerce, online advertising optimization has fully shifted to high-concurrency real-time live-stream segment sequences. In live-stream bidding networks, media data exhibits highly dynamic attributes and is deeply intertwined with multimodal contexts. Within the same live-stream room, the host will continuously switch between multiple recommended products in a short period of time, while viewers simultaneously interact with the live-stream content through various means such as sending bullet comments, commenting, clicking on product cards, and adding items to their shopping carts. Public research and industry practice show that user conversion and purchasing behavior in live-stream scenarios are driven by multiple dimensions of factors, including basic exposure, host's delivery style, real-time interactive atmosphere, and scarce expressions. Current media popularity trends fluctuate dramatically with the switching of featured products and the evolution of the interactive context. To improve the return on investment of advertising resources, mainstream intelligent advertising systems widely integrate multimodal contextual information such as video footage, audio streams, and dynamic subtitles, attempting to accurately anchor the timing of ad placement by understanding the semantic context of the current live-stream segment in real time, thereby executing real-time ad bidding and resource retargeting.

[0003] Existing advertising optimization technologies commonly suffer from misjudgments of trend attribution when dealing with the high-frequency switching scenarios of live-stream e-commerce due to a lack of multi-dimensional semantic decoupling capabilities. Current technologies typically employ static temporal attribution logic, assuming that all media buzz and interactive behavior occurring on the timeline automatically belong to the currently featured product. In a real live-streaming environment, when a highly popular product sells out, is removed from the main promotion sequence, due to the inertia of the host's verbal promotion, residual user interactions, and delayed page prompts, this sold-out product often continues to dominate the context in subsequent live-stream segments, forming a semantic shell of a sold-out product. At this point, the transaction object in the live-stream has changed, but the semantic center of the media has not shifted synchronously. Because existing advertising systems rely only on surface-level timeline alignment and cannot separate the false buzz caused by the residual popularity of past best-selling products, they are highly susceptible to misjudging subsequent media data trends that have been continuously hijacked by sold-out best-sellers as the targetable trends for currently featured products. The aforementioned flaws in the attribution logic directly lead to the bidding engine generating blind positive feedback, mistakenly allocating a large amount of advertising budget and high bids to live segments that have been occupied by residual semantics and have lost their ability to actually generate transactions, ultimately resulting in serious misallocation of advertising resources and waste of budget. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an advertising optimization method and system based on AI-predicted media data trends. This solves the problem of misjudging media trends in live streaming scenarios due to the lack of multi-dimensional semantic decoupling capabilities, which leads to misallocation of advertising resources and wasted budgets.

[0005] To achieve the above objectives, the present invention provides the following technical solution: After identifying that a sold-out best-selling product has been removed from the main promotion sequence, the live broadcast session is segmented to obtain the subsequent live broadcast segment sequence. Multimodal data is collected to calculate the comprehensive heritage carrying capacity, and the sold-out best-selling semantic heritage chain is generated by concatenating them in chronological order. Based on the semantic legacy chain of sold-out best-selling products, a migration stripping optimization algorithm is performed on the subsequent live broadcast segment sequence. The target fitness function is constructed by integrating the amount of interpretable content and the amount of legacy absorption to obtain the optimal segment assignment scheme vector containing the assignment flag. Based on the assignment flag, the subsequent live broadcast segment sequence is divided into a set of residual best-selling product segments and a set of segments for current products. Based on media behavior data, an original live media trend vector is constructed. A mask extraction operation is performed on the set of residual segments of popular content to obtain the shell trend migration part. The shell trend migration part is extracted from the original live media trend vector to generate the shell cut-off net media trend. The local integral value of the shell cut-off net media trend at the end time window is extracted as a scalar parameter characterizing the real-time net traffic intensity. For the set of residual segments of popular products, a blocking instruction is issued to trigger the resource extraction mechanism to release the available advertising budget. The available advertising budget is then injected into the set of segments of the current product. A real-time bidding bid is generated using a positive correlation bidding incentive function based on a scalar parameter representing the net intensity of real-time traffic.

[0006] Furthermore, multimodal data is collected to calculate the comprehensive carrying capacity of the site, including: Construct a semantic term set that includes the official names of sold-out best-selling products, informal colloquial terms, and abbreviations with frequencies exceeding a preset threshold; Feature extraction is performed on multimodal data based on semantic term sets, and the quantities of verbal relics, interactive relics, displayed relics, and alternative relics are quantified and generated respectively. The total number of relics, including those broadcast by voice, those interacted with, those displayed, and those replaced by others, are assigned weight constants and linearly weighted to achieve a comprehensive relic carrying capacity.

[0007] Furthermore, the amount of substitute relics and the coefficients of the weighting constants are generated, including: The text is scanned using rule templates that represent substitution relationships. The matching degree of product category level features is compared between the product identifier of the currently available product and the product attributes of the sold-out best-selling product. The underlying similarity is double-checked by combining the current product price with the preset historical price fluctuation range of the sold-out best-selling product. After the underlying similarity double check passes, the substitution trace quantity is obtained by normalization calculation with the total number of sentences in which the product entity is mentioned as the denominator. The weight constant coefficients are updated by back gradient descent using a machine learning model. The comprehensive heritage carrying capacity is input into the prediction layer containing the activation function to output the model's predicted conversion rate. The cross-entropy loss function is used to calculate the error between the model's predicted conversion rate and the true label of the negative sample fragment. Based on the error, the weight constant coefficients are updated by back gradient descent.

[0008] Furthermore, the calculation of the interpretable amount and the amount of relic absorption includes: Extract the explicit reference quantity, product card click quantity, shopping cart entry quantity, and page mounting persistence quantity of each segment within the current product segment set for the currently available products. Perform segment-specific normalization on the explicit reference quantity, product card click quantity, shopping cart entry quantity, and page mounting persistence quantity. Then, perform a weighted summation of the normalized explicit reference quantity, product card click quantity, shopping cart entry quantity, and page mounting persistence quantity to obtain the acceptable explanatory quantity. Calculate the product of the comprehensive relic carrying capacity of each fragment in the set of residual fragments of the hit product and the fragment position weight. Sum the products to obtain the relic absorption amount. The fragment position weight is a monotonically decreasing function based on the fragment index.

[0009] Furthermore, the target fitness function is constructed by fusing the receptive explanatory power and the relic absorption power to obtain the optimal fragment attribution scheme vector, including: The target fitness function is calculated by dividing twice the product of the acceptable interpretability and the relic absorption by the sum of the acceptable interpretability, the relic absorption, and the preset minimum positive number. The discrete binary particle swarm optimization algorithm is executed to iteratively optimize the velocity vector based on the target fitness function. The velocity vector is mapped to a probability distribution through the sigmoid activation function. A random number comparison mechanism is introduced to complete the discrete state flip of the ownership flag bit to zero or one, and the optimal fragment ownership scheme vector is output.

[0010] Furthermore, an original live streaming media trend vector is constructed based on media behavior data, including: The media popularity values ​​of each segment in the subsequent live broadcast segment sequence are arranged sequentially to form the original live broadcast media trend vector. The media popularity value is a weighted fusion value of the core media behavior indicators. When calculating core media behavior indicators, the preset percentile of the corresponding core media behavior indicator within the preset sliding window duration before the segment is extracted as the normalization denominator, and the actual observed absolute value within the segment is extracted as the numerator for normalization calculation. When the preset percentile is lower than the minimum baseline threshold, the minimum baseline threshold is forcibly used as the normalized denominator.

[0011] Furthermore, a mask extraction operation is performed based on the set of residual fragments from viral content, including: The media popularity value corresponding to the element belonging to the set of residual fragments of popular products is retained, and the value of the corresponding position of the element belonging to the set of fragments of the current product is set to zero, thus obtaining the shell trend migration part; A preset attenuation factor is applied to the shell trend migration part, and the shell trend migration part after applying the attenuation factor is subtracted from the original live media trend vector to achieve the stripping of the shell trend migration part, generating a shell-cut net media trend.

[0012] Furthermore, an instruction to block display was issued, including: Issuing a flag instruction to prohibit rendering, dynamically modifying the Uniform Resource Locator (URL) mapping parameters of the ad slot component in the current segment, or actively reducing the recall priority matching score of the corresponding promoted product in the bidding recall chain; Generating real-time bidding bids using a positively correlated bidding incentive function based on scalar parameters characterizing the net intensity of instantaneous traffic includes: The product of the scalar parameter representing the net intensity of real-time traffic and the preset bidding incentive coefficient is multiplied by one, and then multiplied by the base bidding bid in the original advertising resource configuration data to generate the real-time bidding bid. The preset bidding incentive coefficient is used to construct a quantitative mapping relationship with the preset target conversion cost or the real-time budget consumption rate of the current advertising plan.

[0013] Furthermore, when the set of current product segments output by the migration stripping optimization algorithm is empty, or when the scalar parameter representing the net intensity of real-time traffic is lower than the preset cold start threshold for multiple consecutive sampling periods, the cold start and circuit breaker fallback mechanism is triggered, the price increase redirection operation is suspended and a budget sleep command is sent to the bidding engine until a new valid segment is detected.

[0014] Furthermore, an advertising optimization system based on AI-predicted media data trends is proposed to implement any of the above-mentioned advertising optimization methods based on AI-predicted media data trends, including: The feature extraction module is used to segment the live broadcast session to obtain the subsequent live broadcast segment sequence after recognizing that the sold-out best-selling products have been removed from the main promotion sequence, collect multimodal data to calculate the comprehensive heritage carrying capacity, and generate a semantic heritage chain of sold-out best-selling products in chronological order. The optimization and partitioning module is used to perform a migration stripping optimization algorithm on the subsequent live broadcast segment sequence based on the semantic legacy chain of sold-out best-selling products. It integrates the amount of interpretable content and the amount of legacy absorption to construct a target fitness function to obtain the optimal segment ownership scheme vector containing the ownership flag. Based on the ownership flag, the subsequent live broadcast segment sequence is divided into a set of residual best-selling product segments and a set of segments to be inherited by the current product. The trend decoupling module is used to construct the original live media trend vector based on media behavior data, perform mask extraction operation in combination with the set of residual segments of popular content to obtain the shell trend migration part, peel off the shell trend migration part from the original live media trend vector to generate the shell cut-off net media trend, and extract the local integral value of the shell cut-off net media trend at the end time window as a scalar parameter characterizing the real-time net traffic intensity. The bidding re-bid module is used to issue a blocking instruction to the set of residual segments of popular products to trigger the resource extraction mechanism to release the available advertising budget. The available advertising budget is injected into the set of segments of the current product, and a real-time bidding bid is generated using a positive correlation bidding incentive function based on a scalar parameter representing the real-time net traffic intensity.

[0015] Compared with existing technologies, it has the following advantages: This solution proposes an AI-based method and system for optimizing advertising placement based on predicting media data trends. This effectively addresses the core issue of misclassifying media popularity in high-frequency switching scenarios of live-streaming e-commerce due to residual context from past best-selling products. Existing advertising systems generally rely on superficial timeline alignment logic, easily misattributing surges in traffic left over from sold-out products to the conversion potential of currently available products. To address these shortcomings, this solution constructs a comprehensive legacy carrying capacity and semantic legacy chain encompassing multimodal dimensions, transforming the abstract contextual residual phenomena within the live-streaming room into a quantifiable continuous time series. Building upon this, the solution introduces a transfer-stripping optimization algorithm, which collaboratively evaluates the explanatory power and legacy absorption capacity to accurately delineate the boundaries of traffic camp affiliation within the discrete solution space. This mechanism overcomes the limitations of traditional static attribution, achieving precise decoupling of the deep semantics of live-streaming popularity. This allows the advertising engine to accurately identify and eliminate false traffic peaks caused by historical best-selling products, providing an objective and clean data foundation for subsequent resource allocation.

[0016] Building upon the established traffic segmentation, this solution further employs masking and trend stripping to thoroughly eliminate residual contextual influences from the original live-stream media popularity, reconstructing a shell-like, cut-off net media trend that highly aligns with genuine transaction intentions. Based on the extracted scalar parameters representing the real-time net traffic intensity, the system can execute asymmetric ad resource retargeting along the control chain. It precisely issues blocking instructions to historical residual segments to freeze ineffective budget consumption, while simultaneously channeling the released available ad budget into the current product segment, and generating real-time dynamic bidding based on a positively correlated bidding incentive function. This resource scheduling mechanism fundamentally prevents the blind consumption of ad budgets on ineffective traffic, significantly improving the overall return on investment. Furthermore, combined with built-in local sliding window normalization and cold-start fallback logic, it effectively suppresses the risks of traffic dispersion and vicious cycles that easily occur in bidding networks, ensuring the convergence and stability of long-term campaigns and demonstrating significant commercial conversion efficiency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This application provides a method and system for optimizing advertising placement based on AI-based prediction of media data trends; The method specifically includes the following steps: Step 1: In live-stream e-commerce scenarios, media data trends are highly dynamic and easily influenced by multimodal contexts. When a popular product sells out and exits the main promotion sequence, due to the inertia of the host's verbal delivery, residual user interaction, and page display delays, this sold-out best-selling product often continues to dominate the multimodal context in subsequent live-stream segments. This creates a special phenomenon where the transaction object has changed, but the semantic center of the media remains on the historical best-selling product—the semantic shell of the sold-out best-selling product. To prevent the advertising system from misjudging this residual false popularity as a sustainable trend for currently available products, this embodiment constructs a structured semantic legacy chain of sold-out best-selling products, transforming the aforementioned abstract semantic residual phenomenon into an input basis for algorithmic computation.

[0021] After recognizing that the sold-out best-selling product hx has been removed from the main promotion sequence, the system segments the subsequent live broadcast sessions along the timeline to obtain a sequence of subsequent live broadcast segments. ,in This represents the j-th consecutive subsequent live stream segment, where num is the total number of segments. After segmentation, the system synchronously collects each subsequent live stream segment. Multimodal data within, including segmented anchor spoken text. Fragment of user interaction text Fragment page displays text And the corresponding currently available product identifier for this segment. .

[0022] It should be noted that, due to the objective physical time delay difference between the transmission and decoding of audio streams and the rendering of bullet screen text in live streaming scenarios, the system aligns and calibrates the audio stream timeline and the interactive stream timeline based on the timestamp protocol of the live streaming platform before performing segmentation and acquisition, thereby eliminating feature extraction deviations caused by time misalignment.

[0023] Specifically, the aforementioned sold-out best-selling items (hx) include historically popular items whose inventory status changed to unsellable during the current live stream session, whose featured product links were removed, or which were explicitly stated to be removed from the platform by the streamer. The system employs event-triggered logic segmentation combined with duration constraints. When business events such as changes in product links, changes in the shopping cart's top position, or switching of the main page's icon are detected, the segmentation operation is triggered.

[0024] Furthermore, to ensure that each segment has complete information thickness and to prevent slicing anomalies in extreme cases, the system presets a minimum duration threshold parameter tm and a maximum forced slicing duration parameter. If the time span between the current triggering event and the previous breakpoint is less than the minimum duration threshold parameter tm, the system ignores the current triggering event and postpones it until the next triggering event that meets the duration constraint occurs before cutting off. If the time elapsed since the previous breakpoint reaches the maximum forced breakpoint duration parameter... If no triggering event occurs during this period, the system will forcibly execute the segmentation based on the current timestamp. The preferred range for the minimum duration threshold parameter tm is ten to fifteen seconds, which is set based on the average experience time taken by the host in e-commerce live streaming to complete a minimum script loop including asking questions, displaying information, and guiding clicks. This joint segmentation method with a fallback mechanism avoids fragmented segments that disrupt the host's continuous logic and eliminates the risk of feature sparsity and memory overflow caused by prolonged periods without event triggering.

[0025] After the multimodal data is ready, the system extracts and constructs a dedicated semantic term set vw for the sold-out best-selling product hx. The elements of this set vw not only include the official name, specifications, and brand names from the standard product master data, but also jointly extract informal verbal abbreviations used by users in the current live broadcast session's historical time period, as well as abbreviations frequently used by users in the bullet comments.

[0026] Specifically, the system traces historical conversation logs and uses word frequency statistics combined with a named entity recognition model to identify non-standard pronouns generated by the broadcaster in real time during the live stream that appear more frequently than a preset verbal threshold, and adds them to the set. For abbreviations in user comments, the system extracts the top 5% of words in the word frequency statistics or those with an absolute frequency greater than a preset threshold as the criteria for determining frequently used abbreviations. By mapping a single product identifier to a divergent semantic recall network with objective statistical boundaries, the system significantly improves the recall rate of entity words in colloquial, high-frequency live streaming scenarios.

[0027] Based on the constructed semantic term set vw, the system matches and extracts the collected multimodal raw data, quantifying the relic carrying capacity in four dimensions, namely, the amount of relic broadcast. Interactive Relics The number of relics on display and the amount of replacement remains .

[0028] Specifically, the system reads the text of the anchor's spoken words in segments. Perform word segmentation and pronoun anaphora resolution, count the number of entities that match the semantic term set vw, and divide them into segments. The total number of words in the broadcast, including product entities, is used as the denominator for segment-level normalization calculation to obtain the verbal broadcast trace quantity. For fragmented user interaction text After text cleaning and intent recognition, the system counts the number of bullet comments containing elements from the semantic term set vw and whose intent is classified as inquiring about inventory, restocking, or links, and strictly categorizes them by fragment. The total number of bullet comments is used as the denominator for normalization calculation to obtain the number of interactive relics. Simultaneously, the system utilizes an optical character recognition model to scan fragment pages and display text. The number of text bounding boxes containing the "sold out" or "out of stock" product identifiers or prompts is counted and extracted, and then quantified into the number of displayed relics using the total number of text bounding boxes identified in the current screen as the denominator. .

[0029] It should be noted that in all the normalized calculations above that use the absolute amount of business (such as the total number of words broadcast, the total number of bullet comments, the total number of clicks, etc.) as the denominator, if a denominator of zero is detected, the corresponding feature indicator will be directly assigned a value of zero to ensure the continuity and security of the underlying mathematical calculations.

[0030] It should be noted that the text segmentation, pronoun anaphora resolution, and intent recognition operations are all performed based on a natural language processing model that has been fine-tuned and pre-trained in an e-commerce vertical domain corpus. Limiting the parsing to the vertical domain ensures that the model can accurately identify proper nouns with strong live-stream e-commerce transaction attributes, fully guaranteeing the accuracy of the underlying feature extraction.

[0031] For the amount of substitute remains The system uses rule templates representing substitution relationships to scan text, and retrieves the currently available product identifiers after a template is matched. The underlying similarity of the product attributes of the sold-out best-selling product hx is double-checked. The rule template above adopts dependency syntax matching rules with a topology structure of 'current entity word + alternative state word + historical entity word', where the alternative state word is mapped to a preset feature word library containing 'alternate', 'upgraded version', etc., representing transitional relationships. The aforementioned underlying similarity double check specifically refers to determining whether the product category-level features of the two products match and belong to the same preset subcategory mapping dictionary, and whether the price of the current product falls within the fluctuation range of ±15% of the historical price of the sold-out best-selling product. When the dependency syntax regular expression match is successful and the double check passes, the transitional expression is recorded, and normalization is performed using the total number of sentences in which the product entity is mentioned in the fragment as the denominator to obtain the substitution trace quantity. The verification mechanism with objective floating boundaries effectively filters out accidental triggers when the streamer engages in cross-category casual conversation, while also leaving sufficient room for interpretation of equivalent infringement to prevent competitors from circumventing the rules by slightly adjusting category names.

[0032] In obtaining a single fragment After analyzing the various dimensional features within the fragment, the system uses a linear weighting method to reduce and fuse the multi-source heterogeneous modal residual features, and calculates the comprehensive relic carrying capacity of this fragment for the sold-out best-selling product hx. The weighted fusion formula used is as follows: In the formula, to The system assigns weight constant coefficients to the four corresponding feature dimensions in the overall residual calculation, and the sum of these weight constants equals one. Considering that verbal announcements and interactive comments are the core real-time dynamic variables driving the development of live streaming services, the system... and Assigning relatively high coefficient values ​​to visual presentation and logical substitution as auxiliary reinforcement variables, then... and Assign a relatively low coefficient value.

[0033] It should be noted that the weighting constant coefficient to It is not limited to static constants. The system can update these four weight coefficients using a machine learning model with backward gradient descent, based on the live-streaming characteristics of different product categories and utilizing historical mis-targeting logs. The true label data in the historical mis-targeting logs is constructed by comparing media performance within the monitored period; specifically, it is defined as negative sample segments where the click-through rate is normal but the actual conversion rate is significantly lower than the category baseline within a preset time window after ad exposure. The system inputs the fused relic carrying capacity into a prediction layer containing an activation function to output the model's predicted conversion rate. It then uses a cross-entropy loss function to calculate the error between this predicted conversion rate and the true label, thereby guiding the backward gradient update of the weights. This enables the method to possess cross-category adaptive generalization and error correction capabilities.

[0034] Subsequently, the system sequentially concatenates the comprehensive legacy carrying capacity of each segment according to the temporal order of the subsequent live broadcast segment sequence sq, generating a structured linked list that fully maps the historical influence of sold-out best-selling items, namely the semantic legacy chain ch of sold-out best-selling items. The serialization formula used is as follows: Through the above calculation sequence, discrete and multidimensional semantic residual representations are aggregated into a continuous one-dimensional time series structure, filling the blind spot of multimodal semantic calculation in business scenarios and providing an accurate and dimensionally unified input feature matrix for subsequent global optimization algorithms.

[0035] Step Two: After obtaining the structured semantic legacy chain ch of sold-out best-selling items in Step One, the system needs to make a decision on the media data trend attribution of the subsequent live stream segment sequence sq. Conventional advertising attribution usually defaults to all media behaviors occurring on the timeline to the currently listed product. This static attribution method can easily lead to serious data attribution misalignment when there is strong semantic interference left by the sold-out best-selling product hx. To clarify the true attribution of media data trends, this embodiment uses the semantic legacy chain ch output in the previous steps to perform a migration stripping optimization algorithm in the subsequent live stream segment sequence sq to determine the segment's affiliation.

[0036] The system first mathematically transforms the segment attribution problem into a binary classification search optimization process within a discrete space. Then, it considers num consecutive segments within a subsequent live stream segment sequence sq. The system constructs a set of attribution vectors of length num, which are used as candidate solutions for the optimization algorithm in the iteration space, denoted as . In this vector, For the j-th segment The attribution flag is strictly limited to a Boolean value of 0 or 1. When At that time, characterize the segment It belongs to the residual group that is still affected by the residual distribution of semantic features of sold-out best-selling items; when At that time, characterize the segment It has now been transferred to the actual receiving camp of the currently sold product CX.

[0037] For any set of candidate solutions zv, in order to quantify the rationality of the solution at the business level, the system uses segment-level media behavior data. Two independent orthogonal evaluation indices were extracted and calculated: the capacity for interpretation (pe) and the amount of relic absorption (ae).

[0038] The significance of the explanatory power pe lies in evaluating the classification of the candidate solution zv into the current product acceptance camp (i.e., This set of fragments represents the extent to which the system encompasses the actual transaction capacity of the currently available product (cx). The system calculates the weighted sum of four core business metrics for each fragment within this set, using the following formula: In the formula, For explicit reference quantity. For product card clicks, For the number of items added to the shopping cart, This refers to the persistence of page mounting. To ensure strict dimensional consistency among different segments during weighted calculations and to prevent baseline drift, the system performs segment-specific normalization on all four indicators mentioned above.

[0039] Specifically, With fragments The numerator is the number of words the anchor mentions about the currently available product CX, and the denominator is the total number of words spoken in that segment. The result is calculated by using the number of clicks on the card of the currently available product CX within the segment as the numerator and the total number of clicks on all product cards within the segment as the denominator; The result is calculated by using the number of times a specific product within the segment is added to the cart and converted into a total number of times the product is added to the cart within the segment as the denominator; The result is calculated using the duration the product occupies the main viewport as the numerator and the absolute duration of the segment as the denominator. Parameters to These are the corresponding weight constants, and their sum equals one. Considering that in a real live-streaming e-commerce conversion funnel, product card clicks and shopping cart entries represent direct transaction intentions, the system... and Give higher and The coefficient values ​​are used to ensure that the calculated explanatory power pe has an objective commercial conversion representation capability.

[0040] Regarding the relic absorption amount ae, its significance lies in evaluating whether the candidate scheme zv is classified as a "hot remnant" camp (i.e. The set of fragments represents the degree of coverage of the semantic legacy chain ch of the previously output sold-out best-selling items. The specific calculation formula is as follows: In the formula, To directly retrieve the corresponding fragment from the Relic Chain ch The overall carrying capacity of the relics. In the preliminary steps, It has been quantized into a standard scalar within the closed interval [0,1], and therefore can be directly mathematically coordinated with the normalized acceptable interpretable quantity pe at the same order of magnitude. The fragment position weights are determined by the objective physical law that the radiant energy of semantic relics naturally decays over time. Configured as a monotonically decreasing function based on fragment index j. In a preferred embodiment, the function exhibits an exponential decay formula. ,in This is a preset decay constant. The aforementioned decay mechanism effectively prevents the optimization algorithm from forcibly classifying fragments that are extremely far from the sold-out event and have weak relic characteristics into the residual category.

[0041] To ensure the optimization algorithm evolves in the correct gradient direction, the system fuses the two orthogonal evaluation metrics mentioned above, constructing a home quality metric qx for candidate solution zv as the target fitness function. The specific formula is as follows: In the formula, the parameter ep is a preset minimum positive number, and its preferred range is configured in the range of to The underlying function of introducing this parameter is to smooth the calculation process, eliminating the division-to-zero overflow anomaly caused by the numerator and denominator simultaneously approaching zero, thus ensuring the computational stability of the fitness function. The attribution quality index qx adopts a harmonic mean-based co-excitation structure. Its mathematical logic is that index qx only reaches its peak when candidate scheme zv can accurately allocate fragments with high conversion intentions to the receptive explanatory power pe, and also highly encapsulate fragments with high relic concentrations within the relic absorption power ae, thereby achieving an effective balance in camp demarcation.

[0042] After establishing the vector structure and fitness function, the system invokes the Discrete Binary Particle Swarm Optimization (BSO) algorithm to perform iterative optimization in the global solution space. Upon algorithm startup, the system randomly initializes a particle population consisting of multiple candidate schemes zv and substitutes the position and state of each generation of particles into the home quality index qx formula to calculate the fitness. To ensure the algorithm's engineering reproducibility and convergence stability, the system predefines certain hyperparameter benchmarks: a population size of 20 to 50 particles, a maximum number of iterations of 30 to 100, an inertia weight that decreases linearly with the number of iterations, and both the individual learning factor and the global learning factor are preferably set to a constant of 2.0.

[0043] During the iterative update phase, the algorithm updates the velocity vector based on the individual's historical best position and the population's global best position. It then uses the Sigmoid activation function to map continuous velocity values ​​to a probability distribution falling within the interval (0,1), and introduces a random number comparison mechanism to assign each segment to a flag. The algorithm flips to a discrete state of 0 or 1. The optimization algorithm terminates when the number of iterations reaches a preset maximum threshold, or when the variance of the global optimal attribution quality index qx is less than a preset convergence threshold over several consecutive generations. The system extracts the globally optimal particle in the convergent state and outputs its solution vector as the optimal fragment attribution scheme vector bv.

[0044] Finally, the system performs camp cutting based on the solution space mapping relationship of the optimal fragment attribution scheme vector bv: the attribution flag bits in the vector are removed. The corresponding fragment Aggregate and output as a set of residual fragments of popular content (sh); assign the attribution flags to the vectors. The corresponding fragment Aggregation outputs the current product segment set sc. By introducing a rigorous collaborative evaluation formula and discrete optimization model, this step effectively overcomes the quantitative misalignment caused by traditional static time-series attribution, achieving accurate extraction of true transaction intentions and residual semantic features, and providing a classification boundary with a definite mathematical basis for the subsequent system to reconstruct a pure net media trend.

[0045] Step 3: After obtaining the optimal segment attribution vector bv through the migration stripping optimization algorithm in Step 2, and effectively delineating the set of residual best-selling segments sh and the set of segments sc that the current product inherits, the system needs to further map the discrete attribution camps to the continuous media data trend. In the live streaming bidding network, the original media exposure performance often presents a highly mixed state, that is, the seemingly high traffic simultaneously carries the real appeal of the currently available product cx and the residual artificially high popularity of the sold-out best-selling product hx. In order to provide bidding anchors with a real conversion basis for subsequent advertising resource retargeting, this embodiment performs decoupling and shell cutting operations on the original media trend based on the camp division results.

[0046] The system first bases its data on media behavior of each segment. A one-dimensional original live media trend vector, denoted as ot, is constructed on the time series. This original live media trend vector ot is formed by all consecutive segments in the subsequent live segment sequence sq. Media overall popularity value Arranged sequentially, its vector structure is represented as follows .

[0047] For a single vector element, i.e., a fragment Media overall popularity value The system calculates the weighted average of five core media behavior indicators, as shown in the following formula: In the formula, The normalized value for the number of viewers. For normalized values ​​of comment interaction, Normalized value for product card clicks. To input the normalized value into the shopping cart, Normalized click value for ad entry points. Parameter to These are the weight constants corresponding to each indicator, and their sum equals one. The system automatically retrieves the preset weight mapping table and performs dynamic allocation by parsing the optimization target field (such as Object_Type) returned by the advertising delivery interface. When parsing the business configuration that prioritizes conversion rate, the system automatically assigns a click weight constant to the product card. Weighting constants for shopping cart entry A higher coefficient value; when parsing the business configuration that prioritizes shallow exposure, the system automatically assigns a weighted constant to the number of viewers. Weighting constant for comment interaction A relatively high coefficient value.

[0048] To ensure the weighted comparability of the five heterogeneous media indicators under the same dimension, and to eliminate the interference of local outlier peaks (such as instantaneous traffic fluctuations caused by sudden welfare distributions) on the calculation benchmark, the system abandons the conventional approach of using global historical extreme values ​​as the denominator, and instead configures a local sliding window smoothing mechanism. Specifically, the system uses the current segment... The 95th percentile of the corresponding indicator within the pre-set sliding window duration is used as the normalized denominator, with the segment... The actual observed absolute value within the range is used as the numerator to compress the index data into the closed interval [0,1].

[0049] It is understandable that the selection of the 95th percentile mentioned above is only a preferred example, and the system can also dynamically adjust this quantile range based on the historical variance of the live broadcast room (e.g., dynamically fluctuating between 90% and 99%). Meanwhile, the system has a minimum baseline threshold configured at the underlying level. When the percentile calculated within the aforementioned sliding window is lower than this minimum baseline threshold, the system forcibly uses this minimum baseline threshold as the denominator to prevent oversensitivity of features and gradient divergence risks due to an insufficient data base in the early stages of the live broadcast session.

[0050] After constructing the original live media trend vector ot, the system extracts the trend components belonging to the semantic shell's domain from the set of viral residual fragments sh, i.e., the shell trend migration part, denoted as dt. In the underlying vector operation logic, this shell trend migration part dt is a mask extraction vector with the same dimension as the original live media trend vector ot. The specific rule is: for all elements belonging to the set of viral residual fragments sh, their corresponding media comprehensive popularity value is retained. For all elements belonging to the current product segment set sc, set the value at the corresponding position to zero.

[0051] After mask extraction, the system performs a trend stripping operation, subtracting the shell trend migration portion dt from the original live media trend vector ot to generate the shell-cut net media trend, denoted as nt. The corresponding vector subtraction formula is as follows: The significance of this subtraction formula lies in separating the true from the false in terms of bidding popularity. Compared to the blind smoothing based on empirical thresholds in conventional data processing, the stripping action in this step is based on rigorous semantic attribution and global optimization. Since the non-zero elements in the dt vector are strictly equivalent to the corresponding elements in the ot vector, this vector subtraction operation mathematically ensures the non-negativity of the calculation result, reliably reducing the false trend peaks caused by the residual semantics of sold-out best-selling items (hx) to zero.

[0052] In a supplementary elastic stripping implementation, to accommodate bidding defense strategies of varying intensities and to guard against potential equivalent circumvention designs, the system may also apply a decay factor to the shell trend migration portion dt. (The range of values ​​is) The corresponding transformation formula is: This setup includes industrial scenarios where residual heat is stripped away proportionally, giving the trend restructuring process broader business compatibility.

[0053] The final output of the shell-cutting net media trend nt retains only the overall media popularity value truly generated and sustained by the currently available product cx in its vector sequence. To achieve closed-loop integration with lower-level actions, the system extracts the mean of non-zero elements or the locally normalized integral value of this net media trend vector nt within the final time window (such as a preset T sampling periods before the current segmentation point, preferably aligned to the single bidding delay period of the underlying bidding engine), and outputs it as a scalar parameter representing the real-time net traffic intensity to step four. If all elements within the final time window are zero, the scalar parameter ns is directly output as 0, thereby triggering the cold start and circuit breaker fallback mechanism of the lower-level advertising system. This scalar parameter effectively removes the semantic shell interference wrapped in time sequence, providing a direct and objective decision-making basis for the accurate retargeting of subsequent advertising bids and budgets.

[0054] Step Four: After obtaining the shell-segmented net media trend (nt) through the previous steps and clearly defining the camp boundaries between the set of residual best-selling segments (sh) and the set of segments currently inheriting the product (sc), the system executes the underlying ad resource redirection. In traditional ad delivery engines, ad bidding actions often simply follow the overall window's traffic fluctuation curve, leading to budgets being easily over-consumed by artificially high popularity during product switching periods. To correct the above business deviation, this embodiment, based on the net trend and camp attribution results, completes the dynamic redirection and flow of ad resources at the underlying control link.

[0055] The system's input base includes the set of features output from previous steps and the raw ad resource configuration data (ad) passed in from the ad engine. This raw ad resource configuration data (ad) includes at least the current plan's effective ad budget pool limit, base bid, referral entry page link, and the currently promoted product's mounting object. In this embodiment, the system breaks away from the global static execution mode of the raw ad resource configuration data (ad) and instead performs asymmetric technical intervention based on the fragment's affiliation to a specific camp.

[0056] For traffic slices belonging to the set of residual best-selling product fragments (sh), the system triggers a resource extraction mechanism in the bidding network. Since the semantic core of this traffic is still occupied by the residual characteristics of the sold-out best-selling product (hx), maintaining high-level traffic redirection will cause a severe misalignment between traffic and conversion targets. The system automatically identifies the resource configuration items currently attached to the aforementioned residual best-selling product fragments and issues underlying demotion and display blocking instructions. Specifically, the system blocks display calls by issuing a disable_render flag instruction to the underlying bidding engine, dynamically modifying the Uniform Resource Locator (URL) mapping parameters of the ad components within the current fragment, or proactively reducing the recall priority matching score of the corresponding promoted product in the bidding recall chain. Simultaneously, the bidding bid for the aforementioned fragments is forcibly lowered to a preset platform floor price threshold, and the available budget pool for the aforementioned fragments is frozen proportionally. Through this precise, interface-level demotion intervention, the system effectively prevents the ineffective consumption of advertising budget on false residual popularity.

[0057] Simultaneously, the system re-injects the released available advertising budget into the traffic slice corresponding to the current product segment set sc, performing resource enhancement and retargeting allocation. To quantify the intensity of the injection and enhancement, the system retrieves the scalar parameter ns, representing the real-time net traffic intensity, from the end of step three, and dynamically generates the real-time bid after retargeting based on this scalar parameter ns. The specific bid calculation formula (or positive correlation bid incentive function) is as follows: In the formula, rp is the actual bid after retargeting decision; bp is the base bid set in the original ad resource configuration data ad; ns is the real-time net traffic intensity scalar parameter extracted from the self-purified media trend nt, which has been normalized to the closed interval [0,1] in the previous steps; kf is the preset bidding incentive coefficient. To ensure the objectivity and calculability of the parameter setting, the system constructs a quantitative mapping relationship between the bidding incentive coefficient kf and the preset target conversion cost (Target CPA) or the real-time budget consumption rate of the current ad plan within the advertising system.

[0058] In a preferred configuration, the bidding incentive coefficient kf is inversely proportional to the target conversion cost; that is, the more stringent the target conversion cost requirement, the more convergent the price increase incentive. The mathematical logic of the above bidding formula is that the increase in real-time bidding is strictly positively correlated with the net acceptance strength of the currently sold product cx, which has been quantitatively verified, thereby ensuring that the premium action is anchored to the true conversion intention.

[0059] To prevent uncontrolled ad placement under extreme livestream conditions, the system incorporates cold start and circuit breaker mechanisms at the algorithm's core. When the set of current product segments (sc) output after step two optimization is empty, or when the calculated scalar parameter (ns) falls below the preset cold start threshold for multiple consecutive sampling periods, the system determines that the currently available product (cx) has not yet established an independent and effective conversion context in the livestream. When this protection mechanism is triggered, the system proactively suspends the price increase and retargeting operation and sends a budget dormancy command to the bidding engine until a new, effective segment is detected. This mechanism fills the logical gaps under specific business boundaries, preventing budget losses caused by indiscriminate blind placement.

[0060] In real-time bidding engines, increasing bids based on positive feedback such as clicks often leads to a vicious cycle where "higher bids result in more traffic, which in turn drives up bids again." To ensure the cybernetics stability of the retargeting loop, the local sliding window normalization mechanism introduced in step three of this solution plays a crucial role in balancing and coordinating the system. As the absolute values ​​of clicks and other metrics increase with the bid RP increase, the denominator of the normalized sliding window at the system's underlying level (such as the 95th percentile) also increases proportionally with the overall traffic. This mechanism performs real-time dynamic peak smoothing on the positive feedback data during the retargeting bidding process, suppressing the infinite expansion of the scalar parameter ns from the underlying algorithm architecture, thereby ensuring the mathematical convergence and long-term stability of the entire bidding feedback loop.

[0061] After the targeted extraction and enhanced calculation of the above resources, the system finally generates the final configuration result (rd) for ad resource retargeting. This final configuration result (rd) encapsulates the restructured budget allocation flow, real-time bidding queue, and updated product mounting mapping table in the form of structured logs, and then hands them over to the ad system's bidding engine for physical distribution and display.

[0062] Specifically, through a complete data processing loop from steps one to four, this solution successfully shifts the core decision anchor for real-time optimization of live-stream advertising from semantic residual heat, which has time lag and mixed interference, to the net media trend of current product's actual transactability. In the multimodal, high-concurrency e-commerce bidding environment, this processing method provides the system with high-purity, interference-resistant, and mathematically convergent decision data support, significantly improving the conversion efficiency and return on investment of advertising resources.

[0063] Furthermore, refer to Figure 2 As shown, an advertising optimization system based on AI-predicted media data trends is proposed to implement any of the above-mentioned advertising optimization methods based on AI-predicted media data trends, including: The feature extraction module is used to segment the live broadcast session to obtain the subsequent live broadcast segment sequence after recognizing that the sold-out best-selling products have been removed from the main promotion sequence, collect multimodal data to calculate the comprehensive heritage carrying capacity, and generate a semantic heritage chain of sold-out best-selling products in chronological order. The optimization and partitioning module is used to perform a migration stripping optimization algorithm on the subsequent live broadcast segment sequence based on the semantic legacy chain of sold-out best-selling products. It integrates the amount of interpretable content and the amount of legacy absorption to construct a target fitness function to obtain the optimal segment ownership scheme vector containing the ownership flag. Based on the ownership flag, the subsequent live broadcast segment sequence is divided into a set of residual best-selling product segments and a set of segments to be inherited by the current product. The trend decoupling module is used to construct the original live media trend vector based on media behavior data, perform mask extraction operation in combination with the set of residual segments of popular content to obtain the shell trend migration part, peel off the shell trend migration part from the original live media trend vector to generate the shell cut-off net media trend, and extract the local integral value of the shell cut-off net media trend at the end time window as a scalar parameter characterizing the real-time net traffic intensity. The bidding re-bid module is used to issue a blocking instruction to the set of residual segments of popular products to trigger the resource extraction mechanism to release the available advertising budget. The available advertising budget is injected into the set of segments of the current product, and a real-time bidding bid is generated using a positive correlation bidding incentive function based on a scalar parameter representing the real-time net traffic intensity.

[0064] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An advertising optimization method based on AI-predicted media data trends, characterized in that, include: After identifying that a sold-out best-selling product has been removed from the main promotion sequence, the live broadcast session is segmented to obtain the subsequent live broadcast segment sequence. Multimodal data is collected to calculate the comprehensive heritage carrying capacity, and the sold-out best-selling semantic heritage chain is generated by concatenating them in chronological order. Based on the semantic legacy chain of sold-out best-selling products, a migration stripping optimization algorithm is performed on the subsequent live broadcast segment sequence. The target fitness function is constructed by integrating the amount of interpretable content and the amount of legacy absorption to obtain the optimal segment assignment scheme vector containing the assignment flag. Based on the assignment flag, the subsequent live broadcast segment sequence is divided into a set of residual best-selling product segments and a set of segments for current products. Based on media behavior data, an original live media trend vector is constructed. A mask extraction operation is performed on the set of residual segments of popular content to obtain the shell trend migration part. The shell trend migration part is extracted from the original live media trend vector to generate the shell cut-off net media trend. The local integral value of the shell cut-off net media trend at the end time window is extracted as a scalar parameter characterizing the real-time net traffic intensity. For the set of residual segments of popular products, a blocking instruction is issued to trigger the resource extraction mechanism to release the available advertising budget. The available advertising budget is then injected into the set of segments of the current product. A real-time bidding bid is generated using a positive correlation bidding incentive function based on a scalar parameter representing the net intensity of real-time traffic.

2. The advertising placement optimization method based on AI-predicted media data trends according to claim 1, characterized in that, Collect multimodal data to calculate the comprehensive carrying capacity of the site, including: Construct a semantic term set that includes the official names of sold-out best-selling products, informal colloquial terms, and abbreviations with frequencies exceeding a preset threshold; Feature extraction is performed on multimodal data based on semantic term sets, and the quantities of verbal relics, interactive relics, displayed relics, and alternative relics are quantified and generated respectively. The total number of relics, including those broadcast by voice, those interacted with, those displayed, and those replaced by others, are assigned weight constants and linearly weighted to achieve a comprehensive relic carrying capacity.

3. The advertising placement optimization method based on AI-predicted media data trends according to claim 2, characterized in that, The generation of alternative relics and weight constant coefficients includes: The text is scanned using rule templates that represent substitution relationships. The matching degree of product category level features is compared between the product identifier of the currently available product and the product attributes of the sold-out best-selling product. The underlying similarity is double-checked by combining the current product price with the preset historical price fluctuation range of the sold-out best-selling product. After the underlying similarity double check passes, the substitution trace quantity is obtained by normalization calculation with the total number of sentences in which the product entity is mentioned as the denominator. The weight constant coefficients are updated by back gradient descent using a machine learning model. The comprehensive heritage carrying capacity is input into the prediction layer containing the activation function to output the model's predicted conversion rate. The cross-entropy loss function is used to calculate the error between the model's predicted conversion rate and the true label of the negative sample fragment. Based on the error, the weight constant coefficients are updated by back gradient descent.

4. The advertising placement optimization method based on AI-predicted media data trends according to claim 1, characterized in that, Calculate the amount of interpretation and the amount of relic absorption, including: Extract the explicit reference quantity, product card click quantity, shopping cart entry quantity, and page mounting persistence quantity of each segment within the current product segment set for the currently available products. Perform segment-specific normalization on the explicit reference quantity, product card click quantity, shopping cart entry quantity, and page mounting persistence quantity. Then, perform a weighted summation of the normalized explicit reference quantity, product card click quantity, shopping cart entry quantity, and page mounting persistence quantity to obtain the acceptable explanatory quantity. Calculate the product of the comprehensive relic carrying capacity of each fragment in the set of residual fragments of the hit product and the fragment position weight. Sum the products to obtain the relic absorption amount. The fragment position weight is a monotonically decreasing function based on the fragment index.

5. The advertising placement optimization method based on AI-predicted media data trends according to claim 4, characterized in that, The target fitness function is constructed by fusing the interpretable amount and the relic absorption amount to obtain the optimal fragment attribution scheme vector, including: The target fitness function is calculated by dividing twice the product of the acceptable interpretability and the relic absorption by the sum of the acceptable interpretability, the relic absorption, and the preset minimum positive number. The discrete binary particle swarm optimization algorithm is executed to iteratively optimize the velocity vector based on the target fitness function. The velocity vector is mapped to a probability distribution through the sigmoid activation function. A random number comparison mechanism is introduced to complete the discrete state flip of the ownership flag bit to zero or one, and the optimal fragment ownership scheme vector is output.

6. The advertising placement optimization method based on AI-predicted media data trends according to claim 1, characterized in that, Based on media behavior data, an original live streaming media trend vector is constructed, including: The media popularity values ​​of each segment in the subsequent live broadcast segment sequence are arranged sequentially to form the original live broadcast media trend vector. The media popularity value is a weighted fusion value of the core media behavior indicators. When calculating core media behavior indicators, the preset percentile of the corresponding core media behavior indicator within the preset sliding window duration before the segment is extracted as the normalization denominator, and the actual observed absolute value within the segment is extracted as the numerator for normalization calculation. When the preset percentile is lower than the minimum baseline threshold, the minimum baseline threshold is forcibly used as the normalized denominator.

7. The advertising placement optimization method based on AI-predicted media data trends according to claim 6, characterized in that, Perform mask extraction operations based on the set of residual fragments from viral content, including: The media popularity value corresponding to the element belonging to the set of residual fragments of popular products is retained, and the value of the corresponding position of the element belonging to the set of fragments of the current product is set to zero, thus obtaining the shell trend migration part; A preset attenuation factor is applied to the shell trend migration part, and the shell trend migration part after applying the attenuation factor is subtracted from the original live media trend vector to achieve the stripping of the shell trend migration part, generating a shell-cut net media trend.

8. The advertising placement optimization method based on AI-predicted media data trends according to claim 1, characterized in that, Issue a command to block display, including: Issuing a flag instruction to prohibit rendering, dynamically modifying the Uniform Resource Locator (URL) mapping parameters of the ad slot component in the current segment, or actively reducing the recall priority matching score of the corresponding promoted product in the bidding recall chain; Generating real-time bidding bids using a positively correlated bidding incentive function based on scalar parameters characterizing the net intensity of instantaneous traffic includes: The product of the scalar parameter representing the net intensity of real-time traffic and the preset bidding incentive coefficient is multiplied by one, and then multiplied by the base bidding bid in the original advertising resource configuration data to generate the real-time bidding bid. The preset bidding incentive coefficient is used to construct a quantitative mapping relationship with the preset target conversion cost or the real-time budget consumption rate of the current advertising plan.

9. The advertising placement optimization method based on AI-predicted media data trends according to claim 8, characterized in that, Also includes: When the set of current product segments output by the migration stripping optimization algorithm is empty, or when the scalar parameter representing the real-time net flow intensity is lower than the preset cold start threshold for multiple consecutive sampling periods, the cold start and circuit breaker fallback mechanism is triggered, the price increase redirection operation is suspended and a budget sleep command is sent to the bidding engine until a new valid segment is detected.

10. An advertising placement optimization system based on AI-predicted media data trends, used to implement the advertising placement optimization method based on AI-predicted media data trends as described in any one of claims 1-9, characterized in that, include: The feature extraction module is used to segment the live broadcast session to obtain the subsequent live broadcast segment sequence after recognizing that the sold-out best-selling products have been removed from the main promotion sequence, collect multimodal data to calculate the comprehensive heritage carrying capacity, and generate a semantic heritage chain of sold-out best-selling products in chronological order. The optimization and partitioning module is used to perform a migration stripping optimization algorithm on the subsequent live broadcast segment sequence based on the semantic legacy chain of sold-out best-selling products. It integrates the amount of interpretable content and the amount of legacy absorption to construct a target fitness function to obtain the optimal segment ownership scheme vector containing the ownership flag. Based on the ownership flag, the subsequent live broadcast segment sequence is divided into a set of residual best-selling product segments and a set of segments to be inherited by the current product. The trend decoupling module is used to construct the original live media trend vector based on media behavior data, perform mask extraction operation in combination with the set of residual segments of popular content to obtain the shell trend migration part, peel off the shell trend migration part from the original live media trend vector to generate the shell cut-off net media trend, and extract the local integral value of the shell cut-off net media trend at the end time window as a scalar parameter characterizing the real-time net traffic intensity. The bidding re-bid module is used to issue a blocking instruction to the set of residual segments of popular products to trigger the resource extraction mechanism to release the available advertising budget. The available advertising budget is injected into the set of segments of the current product, and a real-time bidding bid is generated using a positive correlation bidding incentive function based on a scalar parameter representing the real-time net traffic intensity.

Images