A new media content diversion and conversion method

Abstract

The application belongs to the technical field of data processing, and particularly relates to a new media content diversion and conversion method, which is applied to a cross-platform content promotion scene including at least a first platform, a second platform and a third platform, and the method comprises the following steps: a content understanding processing step: performing a unified processing procedure on the content in each platform to generate a unified content feature representing the content; a cross-domain prediction modeling step: using a three-task network model to predict and output a target user behavior of the content, the target user behavior including a click probability, a conversion probability and a refund probability; a delivery allocation optimization step: generating a delivery scheme with the dimensions of platform, category, creative, audience group and time period. The application effectively improves the cross-platform consistency of content promotion through unified modeling and optimization strategies, solves the problems of model migration difficulty, target conflict and low resource allocation efficiency in the prior art, and has significant commercial value and engineering application effect.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method for attracting and converting new media content. Background Technology

[0002] Currently, multimedia content distribution and personalized recommendation technologies are widely used in internet information platforms, e-commerce platforms, and social media platforms. With the increasing popularity of short videos, mixed text and image feeds, and multi-channel promotion, users' demands for content presentation quality and platform distribution accuracy continue to rise. In existing technologies, mainstream platforms generally adopt content understanding methods based on the fusion of visual and textual features. These methods use deep learning models to extract image and textual semantic features to predict metrics such as click-through rates, conversion rates, or interaction rates. These technologies typically utilize convolutional neural networks to process image input and attention-based language models to process text input, then concatenate the two types of features and input them into a multi-task prediction network. However, this general approach still faces significant challenges in cross-platform and multi-domain data fusion and model consistency.

[0003] First, existing technologies suffer from data inconsistency in aligning text and image content. Significant differences in content upload specifications, image compression algorithms, video sampling strategies, and text structures across different platforms lead to feature shifts in the same content extracted from different platforms. Traditional methods often rely on independently trained visual and text encoders, lacking a globally consistent alignment mechanism. This makes it difficult to maintain consistency in unified content features when the model is applied across multiple platforms, resulting in discrepancies in recommendation results and ad placement allocation. Especially in commercial advertising, the prediction results for the same ad creative can vary considerably across platforms, severely impacting the efficiency and fairness of advertiser budget allocation. Second, existing multi-task prediction models are typically optimized for single-platform data and lack cross-domain transfer capabilities. Traditional multi-task learning frameworks (such as joint click-through and conversion prediction models) are often trained only on single-platform data. When the model is transferred to other platforms, its performance significantly degrades due to differences in user behavior patterns and content distribution. Some studies have attempted to share some parameters across different platforms using transfer learning or distillation techniques; however, in practical applications, differences in data distribution between platforms cause "domain drift" in the shared components, affecting model stability. As a result, the model needs to be retrained on the new platform, increasing training costs and slowing down deployment. Summary of the Invention

[0004] Therefore, the main objective of this invention is to provide a new media content traffic acquisition and conversion method applicable to multi-platform content promotion scenarios. By unifying content features, predicting cross-domain behavior, and optimizing multi-objective delivery, a complete system is constructed, encompassing content display, user conversion, and delivery within a closed loop. First, by integrating visual and textual feature extraction and alignment training, content from different platforms achieves consistent expression within a unified semantic space, thus eliminating feature offset issues between platforms. Then, a three-task network model with path consistency constraints is employed to jointly model and predict according to the logical order of clicks, conversions, and refunds, ensuring that the output matches the actual user behavior chain. Furthermore, a domain adaptation and transfer mechanism enables model sharing and rapid convergence among the first, second, and third platforms, reducing cross-platform training costs. Finally, in the delivery allocation stage, an optimal delivery plan is generated under budget, frequency, and category constraints using multi-objective integer programming, achieving a balance between maximizing conversions, minimizing refunds, and increasing clicks.

[0005] The technical solution adopted in this invention is as follows:

[0006] A new media content traffic acquisition and conversion method is applied to cross-platform content promotion scenarios that include at least a first platform, a second platform, and a third platform. The method includes the following steps: a content understanding and processing step: performing a unified processing flow on the content in each platform to generate unified content features representing the content; a cross-domain prediction and modeling step: based on the unified content features, using a three-task network model, through jointly executing alignment, freezing, and transfer training processes, predicting and outputting the target user behavior of the content, the target user behavior including click probability, conversion probability, and refund probability; and a delivery allocation optimization step: based on the unified content features and the predicted click probability, conversion probability, and refund probability, under budget, frequency, and category constraints, generating a delivery plan based on platform, category, creative, audience segmentation, and time period.

[0007] Furthermore, the content understanding processing steps specifically include: performing image and video frame processing to generate image semantic information, wherein the image and video frame processing includes: sampling multiple frames of images from the video content at a preset frequency and obtaining their cover images, scaling and standardizing each of the images; and performing forward inference on each of the standardized images using a Transformer-based visual model and a convolutional neural network-based model, respectively, to generate the image semantic vector and candidate image semantic vectors.

[0008] Furthermore, the image and video frame processing further includes: concatenating the image semantic vector of the same image with the candidate image semantic vector, and mapping them through at least one fully connected layer to obtain a single-frame fused image semantic vector of the image; and aggregating the single-frame fused image semantic vectors of multiple images corresponding to the content in chronological order to obtain the final fused image semantic vector of the content.

[0009] Furthermore, the content understanding processing step also includes performing text encoding: encoding the title and description text of the content into a token sequence using a sub-word segmenter; and inputting the token sequence into a Transformer encoder model, and taking the first and second vectors of the sequence output by the model as the text semantic vector of the content.

[0010] Furthermore, the content understanding processing step further includes performing image-text alignment training, which includes: forming positive sample pairs by combining the semantic vector of the final fused image with the semantic vector of the text containing the same content; within the same training batch, selecting text semantic vectors from different content texts to form a first difficult negative sample pair by using a similarity-based retrieval method, and selecting final fused image semantic vectors from different content images to form a second difficult negative sample pair; and performing similarity-close training on the positive sample pairs and similarity-separation training on the first and second difficult negative sample pairs by using a contrastive learning loss function.

[0011] Furthermore, the unified content feature is generated by splicing the final fused screen semantic vector, the text semantic vector, and the platform identifier, category identifier, and entry identifier in a preset order.

[0012] Furthermore, the three-task network model includes a shared coding subnet and three task subheadings; the shared coding subnet is composed of multiple fully connected maps and nonlinear activation functions stacked sequentially; the three task subheadings have the same structure, and each task subheading contains multiple fully connected maps, a nonlinear activation function, and a probability mapping unit.

[0013] Furthermore, the alignment phase in the training process includes: cyclically training the three task sub-heads in a preset order of click, conversion, and refund, and after each cycle, recalculating the threshold on the corresponding task sub-head for training samples that violate the preset path relationship, wherein the preset path relationship is that the conversion event must occur after the click event, and the refund event must occur after the conversion event; the freezing phase in the training process includes: fixing the parameters of the shared encoding subnet, and performing supervised fine-tuning of the output of the three task sub-heads based on historical display records, click records, order records, and refund records.

[0014] Furthermore, the transfer joint phase in the training process includes: performing three model transfers cyclically between the first platform, the second platform, and the third platform; wherein each model transfer includes: on the target platform, concatenating a domain adaptation pre-unit for the first layer of the shared coding subnet; and on the target platform, reusing the three task subheadings of the source platform as the initial state and performing supervised training.

[0015] Furthermore, the optimization steps for ad placement allocation include: constructing a multi-objective integer programming model, wherein a non-negative integer decision variable is defined for each candidate ad placement unit generated by the Cartesian product of platform, category, creative, audience segment, and time period; setting the following constraints for the model: constraining the total cost of all candidate ad placement units to be less than or equal to the preset budget; constraining the number of ad placements for each audience segment in each time period to be less than or equal to the frequency limit; and constraining the cost proportion of each category in the overall ad placement plan within a preset range; and setting a multi-objective function for the model to be solved in priority order, with the priorities being: maximizing the estimated total conversions, minimizing the estimated total refunds, and maximizing the estimated total clicks.

[0016] By adopting the above technical solutions, this invention achieves the following beneficial effects: By introducing an alignment training mechanism that integrates images and text in the content understanding stage, the semantic features of the same content across different platforms can form a stable mapping in a unified vector space, thereby achieving consistent cross-platform expression. This feature unification not only reduces the feature drift problem caused by differences in data standards in traditional multi-model solutions, but also significantly improves the accuracy of content matching and recall, establishing a shared feature foundation for cross-platform promotion. In the behavior prediction stage, this invention adopts a three-task network model. Through fixed task order training and path consistency checks, it ensures that the logical dependencies between the three behaviors of click, conversion, and refund conform to reality, making the prediction results interpretable and stable. Compared with existing methods that only independently model each indicator, this invention achieves a causal closed loop at the behavior level, fundamentally avoiding abnormal predictions such as "conversion without click" or "refund without conversion." Furthermore, in the transfer and joint stage, through the concatenation and parameter reuse of domain adaptation pre-units, efficient transfer between the first, second, and third platforms is achieved, enabling the model to converge quickly and maintain consistent performance across different platforms, significantly reducing cross-domain training costs. Ultimately, in the campaign allocation optimization stage, the multi-objective integer programming method proposed in this invention replaces the traditional linear weighted summation strategy. It solves the problem according to the priority order of maximizing conversion, minimizing refunds, and maximizing clicks, avoiding biases caused by manual weight setting and ensuring the global optimality of the result. This planning model simultaneously considers constraints such as budget, frequency, and category proportion, achieving a reasonable balance of campaign resources across multiple dimensions. In summary, this invention, through the combination of unified feature representation, path consistency constraints, and multi-objective discrete optimization, constructs a complete closed-loop system that is cross-platform scalable, interpretable, reproducible, and commercially viable, significantly improving the accuracy of content promotion, campaign revenue, and system stability. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a new media content traffic acquisition and conversion method provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram of the three-task network model architecture provided in an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of a multi-frame video feature fusion mechanism provided in an embodiment of the present invention. Detailed Implementation

[0020] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0021] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0022] Example 1: Reference Figure 1 This implementation method revolves around the first platform, the second platform, and the third platform, and sequentially executes the content understanding and processing step, the cross-domain prediction modeling step, and the delivery allocation optimization step to generate a delivery plan based on the dimensions of platform, category, creative, audience segmentation, and time period.

[0023] First, each video clip was sampled continuously for 12 frames at a fixed frequency of 2 frames per second, and one cover image was obtained, for a total of 13 images. Each image was processed as follows: the image was scaled to 256 pixels along its shortest side, and then cropped from the center to 224×224 pixels. The pixel values ​​were normalized to the range of 0 to 1. A fixed mean was subtracted from each channel and divided by a fixed standard deviation. The mean values ​​for the red channel were 0.485, the green channel 0.456, and the blue channel 0.406; the standard deviations for the red channel, green channel, and blue channel were 0.229, 0.224, and 0.225, respectively. These specific values ​​ensured that the brightness and color distribution of materials from different acquisition devices and platforms fell within a stable range, facilitating consistent response scales for subsequent models across the three platforms and reducing the interference of color drift on the same content across different upload devices on downstream judgments.

[0024] For each preprocessed image, the following deterministic process is performed: the image is divided into 196 image blocks of 16×16 pixels, and flattened into a vector sequence of length 768 pixels in a left-to-right, top-to-bottom order. A classification label vector is added to the first position of the sequence, and fixed positional information is added to each position. The resulting sequence is then processed through a 12-layer encoding structure, each layer containing multi-head attention computation, residual connection and layer normalization, forward transformation, residual connection and layer normalization. The classification label vector is taken as the image's semantic vector. Sequential image block processing preserves spatial adjacency relationships and builds a semantic representation from local texture to global composition layer by layer in the 12-layer stack; using the classification label as output facilitates comparison of different image contents in the same dimensional space and makes alignment with text semantic vectors easier.

[0025] For the same image, convolutional forward inference is performed simultaneously: the image is passed through several convolutional layers, batch normalization layers, and non-linear activation layers, with skip connections at several points to preserve the original information flow. Global average pooling is then performed on the last feature map to obtain candidate image semantic vectors. Convolutional extraction strengthens local stable features such as edges, textures, and shapes, complementing Transformer-based global modeling and improving cross-platform consistency. The image semantic vector and candidate image semantic vectors are concatenated, and the concatenated vector is passed through two fully connected layers and non-linear activation layers to obtain the single-frame fused image semantic vector. The arithmetic mean of 13 images in chronological order is calculated to obtain the final fused image semantic vector for the content. Using concatenation and remapping, the visual semantics from two different sources can be unified within a fixed dimension, avoiding scale inconsistencies caused by model structure differences across the three platforms. Arithmetic averaging across multiple frames can suppress motion blur and exposure anomalies in individual frames, ensuring stable reproduction of the final fused image semantic vector for the same content across the three platforms.

[0026] The title and description text are merged and decomposed into a tokenized sequence using a sub-segmenter. The maximum length is limited to 77 characters; any exceeding this length is truncated at the end, and placeholders are added at the end if the length is insufficient. This fixed length ensures sequence alignment during batch processing and reduces length fluctuations caused by different languages ​​and input methods. The tokenized sequence is mapped to a vector sequence, with fixed positional information encoded, and passed through a 12-layer encoding structure. Each layer includes multi-head attention, residual connections and layer normalization, forward transformation, residual connections, and layer normalization. The first and last vectors of the sequence are used as the text semantic vector. Through this 12-layer stacking, the subject-verb-object relationship, brand name, and functional description in the text are aggregated into the first and last vectors for easy alignment with the visual side. For each piece of content, the final merged visual semantic vector and text semantic vector form a positive sample pair. Within the same training batch, a BM25-based retrieval method is used to select the top 5 text semantic vectors with the highest similarity from other content texts for the final fused image semantic vector, forming the first hard-negative sample pair. Simultaneously, the top 5 final fused image semantic vectors with the highest similarity from other content images are selected for the same text semantic vector, forming the second hard-negative sample pair. Using hard-negative samples can approximate real-world interference scenarios, such as similar titles but different images, or similar images but different descriptions. This explicitly differentiates similar interference across the three platforms, ensuring that unified content features are not misled by similar noise during cross-platform migration.

[0027] In each batch, a similarity-increasing training round consisting only of positive sample pairs is performed first, bringing images and text with the same content closer together in the vector space. This is followed by a similarity-dispersing training round consisting only of weak negative sample pairs, separating similar but different images and text. Each batch contains 256 pieces of content, completing three rounds of traversal. This sequence first establishes stable anchor points for the same content, then removes the most easily confused adjacent interference, ultimately ensuring that identical content falls into the same neighborhood across the three platforms while different content is separated from each other. The final fused image semantic vector, text semantic vector, one-hot encoding of platform identifiers, one-hot encoding of class identifiers, and one-hot encoding of entry identifiers are concatenated in a fixed order to obtain unified content features. This unified order is maintained across the three platforms, ensuring that the input structure received by the subsequent three-task network model is completely consistent, facilitating stable operation during the freeze and transfer joint phases.

[0028] A shared encoding subnet and three task subheadings are established. The shared encoding subnet consists of four layers of fully connected mappings and nonlinear activations, each followed by a normalization and dropout unit to improve robustness across three platforms. The three task subheadings have a consistent structure, each containing two layers of fully connected mappings and nonlinear activations, and using a standard probability mapping function at the end to constrain the output to the range of 0 to 1, corresponding to click probability, conversion probability, and refund probability. The unified structure facilitates the fixed-order training during the alignment phase and the sequential calibration during the freeze phase, ensuring that the training order is executed according to rules.

[0029] Alignment Phase: Batches of 256 content items are constructed using data from the current sampling period. The training order is fixed as click, conversion, refund, repeated 10 times. In each round, the unified content features are first encoded through a shared encoding subnet, and then fed into the corresponding task subheadings to calculate the output. This order follows the natural path of seeing a click, from click to conversion, and from conversion to refund, gradually fixing the baseline for early behavior judgment. Subsequent behaviors are then learned on this baseline, reducing the negative interference of subsequent behaviors on early behaviors. After each round, path consistency checks are performed on all samples from that round. The rules are: conversion events can only occur after click events, and refund events can only occur after conversion events. If a sample with a click probability less than a fixed threshold but a conversion probability greater than a fixed threshold is detected, the threshold for that sample is recalculated in the conversion task subheading, lowering the conversion probability to the boundary consistent with the click probability; similarly, if a sample with a conversion probability less than a fixed threshold but a refund probability greater than a fixed threshold is detected, the threshold for that sample is recalculated in the refund task subheading, lowering the refund probability to the boundary consistent with the conversion probability. The fixed threshold is 0.5. This approach directly eliminates impossible events in the behavior path, avoids gradient signals that do not conform to the real funnel relationship during training, and ensures that the three-task network model maintains the same causal order of behavior on the three platforms.

[0030] Freezing Phase: After completing the alignment phase, the parameters of the shared coding subnet are fixed, allowing only three task subheadings to be updated. Supervised fine-tuning of the three task subheadings is performed using historical display records, click records, order records, and refund records. This approach locks the highly expressive but stable coding portion to prevent drift due to data differences across platforms, while allowing the mapping portion directly related to the output to quickly align with the records. Calibration is performed in a fixed order: click, conversion, refund. If a display record is positive and the click probability is below 0.5, the click probability of that sample is set to 0.5; if a click record is positive and the conversion probability is below 0.5, the conversion probability of that sample is set to 0.5; if an order record is positive and the refund probability is below 0.5, the refund probability of that sample is set to 0.5; if a record is negative and the corresponding probability is above 0.5, the probability is set to 0.5. Sequential calibration ensures that the three outputs are consistent with historical records under the same threshold rules, avoiding offsets caused by differences in record standards across platforms.

[0031] Migration and Joint Phase: Three migrations are performed sequentially: from platform 1 to platform 2, from platform 2 to platform 3, and from platform 3 to platform 1. During each migration, a first-layer concatenated domain adaptation pre-unit is used in the shared coding subnet of the target platform. This pre-unit sequentially performs input normalization, linear dimension alignment, and non-linear activation, ensuring that the unified content features of the target platform are mapped to the same distribution scale as the source platform before entering the shared coding subnet, reducing misjudgments caused by platform-specific distribution differences. The target platform reuses the three task subheadings from the source platform as the initial state, performs supervised training using the target platform's display, click, order, and refund records, and repeats the sequential calibration from the freeze phase. This process maps and migrates stable behaviors from the source platform to the target platform, and then aligns them using the target platform's records, achieving near-consistent performance with the source platform within a few rounds, typically converging within 3 to 5 rounds. For each combination of content with platform, category, and entry point, forward inference with the shared coding subnet and the three task subheadings is performed to obtain the click probability, conversion probability, and refund probability. All three probabilities are uniformly retained to three decimal places and stored for use in the campaign allocation optimization step. Using a uniform number of decimal places facilitates equivalent comparisons and sorting across the three platform campaign systems, avoiding inconsistencies caused by differences in floating-point precision.

[0032] Candidate ad units are generated by performing a Cartesian product of platform, category, creative, audience segment, and time period. For example, with 3 platforms, 12 categories, 50 creatives, 20 audience segments, and 24 time periods, there are 864,000 candidate ad units. Each candidate ad unit is associated with click probability, conversion probability, refund probability, cost per unit, frequency cap, and category identifier. The following constraints are set: Budget constraint: The total cost of all candidate ad units does not exceed a given budget, e.g., 100,000. Frequency constraint: The number of ad units for each audience segment in each time period does not exceed a given cap, e.g., 3. Category constraint: The cost percentage of each category in the overall campaign falls within a preset range, e.g., a minimum of 2% and a maximum of 15% for each category. The budget constraint ensures controllable total cost; the frequency constraint avoids over-reaching the same audience and causing negative reactions; the category constraint ensures broad coverage to avoid diminishing marginal returns due to over-concentration on a single category.

[0033] A multi-objective integer programming model is used to solve the problem. A multiple objective function is employed, solved in priority order: maximizing the estimated total conversions, minimizing the estimated total refunds, and maximizing the estimated total clicks. Specifically, under the constraint of satisfying the conditions, the solution set with the largest estimated total conversions is found first. Then, the solution set with the smallest estimated total refunds is found within this set. Finally, the solution set with the largest estimated total clicks is found within this set. This order reflects actual business preferences: ensuring purchases first, controlling refunds second, and finally increasing views and clicks third. Unlike simple weighted summation, priority-based solutions do not rely on manual weight settings, avoiding discrepancies caused by inconsistent weight interpretations across the three platforms.

[0034] For example, with a budget of 100,000, a frequency cap of 3, and a category percentage of 2% to 15%, an integer programming solver is used with a solution time limit of 60 seconds. After the solution is completed, approximately 420 to 560 non-zero ad units are output, covering 3 platforms, approximately 10 categories, approximately 35 creatives, and approximately 18 audience segments. Compared to a one-size-fits-all approach that does not differentiate between platforms, under the same budget, the estimated total conversions increase by approximately 12% to 18%, the estimated total refunds decrease by approximately 8% to 11%, and the total clicks increase by approximately 9% to 14%. This improvement comes from three aspects: unified content features ensure consistent semantic focus for the same creative across the three platforms; the three-task network model guarantees path consistency and distribution alignment during the freeze and migration joint phases; and the priority-based solution order ensures that limited resources are first used for ad units with high conversions and low refunds.

[0035] Within the common frame rate range of 5 to 30, sampling 2 frames per second can cover major shot changes without excessively increasing computational overhead. Taking 12 frames plus one cover image can simultaneously cover the main video content and the platform's initial display, reducing deviations caused by differences between the cover and the main text. 224×224 pixels have wide availability in existing large-scale pre-training systems, allowing for direct reuse of mature visual models. Fixed mean and standard deviation cover the data distribution across multiple device models and lighting scenarios, normalizing the materials across the three platforms to a consistent brightness and color range, thereby improving the stability of unified content features across the three platforms. Transformer-based visual models are better at global relationships and long-distance dependencies, while convolutional models are more robust to local details such as edges and textures. Directly taking one of these will result in distortion in certain scenarios, such as pure text images or complex texture scenes. After concatenation and remapping, a fixed-dimensional single-frame fused image semantic vector is formed, allowing both types of cues to simultaneously influence subsequent alignment training and the three-task network model without introducing complex weighting. First, bringing images and text pairs with the same content closer together can create a stable anchor point. Then, separating similar interfering elements can prevent the anchor point from shifting during the process. Performing two rounds of alignment reinforcement in this order can significantly reduce misjudgments across the three platforms caused by similar title writing or similar visual layout.

[0036] In one alternative implementation, the sampling frequency for image and video frame processing can be changed from 2 frames per second to 4 frames per second, with the total number of images adjusted accordingly to 25; the rest of the process remains unchanged. This change is suitable for content with fast camera transitions. For text encoding, the maximum length of 77 can be replaced with 128, suitable for long descriptive content; to maintain batch processing efficiency, the fixed-length truncation and padding strategy is still maintained. In image-text alignment training, the selection of the most difficult samples can be changed from the top 5 to the top 10 to increase the alignment difficulty; when the batch size remains unchanged, the alignment reinforcement rounds can be adjusted from 3 rounds to 2 rounds to control the training duration. In the sequential calibration during the freeze phase, the fixed threshold of 0.5 can be replaced with 0.6 to raise the threshold for positive conversion and refund rates for high-risk categories, thereby focusing more on stable conversions in the placement allocation optimization step.

[0037] Example 2: For each piece of content, 12 frames are continuously sampled from the video at a fixed frequency of 2 frames per second, and one cover image is extracted, forming a sequence of 13 images. This setup covers the core frame for approximately 6 seconds, capturing the main event in short content with many camera cuts without significantly increasing subsequent computational latency. For each image, it is first scaled to 256 pixels along the shortest side, using bilinear interpolation to maintain the aspect ratio; then, it is center-cropped to 224×224 pixels. Fixed size and center cropping ensure consistent input structure across the first, second, and third platforms, avoiding feature drift caused by different upload resolutions and display cropping strategies. Pixel values ​​are linearly scaled to the range of 0 to 1; standardized by channel: red channel mean 0.485, green channel mean 0.456, blue channel mean 0.406; red channel standard deviation 0.229, green channel standard deviation 0.224, blue channel standard deviation 0.225. By using publicly available and commonly used channel statistics, mature visual pre-training systems can be directly reused, ensuring that the color distribution of materials across the three platforms falls within the same range, thereby reducing the impact of differences in lighting and equipment.

[0038] The 224×224 pixel image is divided into 196 non-overlapping 16×16 pixel blocks. These blocks are then flattened into a sequence of image block vectors of uniform length, arranged from top to bottom and left to right, with a classification label vector added at the beginning of the sequence. The sequence is organized in a fixed order to ensure consistent execution across all three platforms. Fixed positional information is encoded for each location. The sequence is then passed through a 12-layer encoding structure, with each layer sequentially performing multi-head attention computation, residual connection and layer normalization, forward transformation, residual connection and layer normalization. This process integrates local texture and global composition layer by layer, forming a stable and comparable representation of the main object, layout, and scene arrangement. The output vector at the classification label location is taken as the image's semantic vector. The classification label location is chosen because it continuously gathers information from all image blocks within the 12-layer stack, stably representing the global semantics of the entire image and facilitating alignment and comparison with a vector from the text side.

[0039] For the same image, a multi-level stacked structure of convolution, batch normalization, and non-linear activation is sequentially applied, with skip connections between several layers to preserve early edge and texture information. This process enhances locally invariant features such as contours and textures. Global average pooling is then performed on the final feature map to obtain candidate image semantic vectors. Global average pooling is chosen over max pooling because averaging is more robust to local noise and can maintain the integrity of subject features in images containing captions, watermarks, or partial occlusion. To ensure the reproducibility of experimental results, the following elements are fixed on the three platforms: image patch stride of 16 pixels, number of encoding layers of 12, cropping size of 224×224 pixels, channel statistics using the aforementioned constants, and convolutional structure levels and order remaining unchanged. All inference is performed deterministically, without any random augmentation or random discarding, thus ensuring consistent image semantic vectors and candidate image semantic vectors for the same input on different platforms. The above process unifies the image input of the same content across the three platforms to the same geometric scale and color distribution, and then obtains a dual expression of global semantics and local details through two complementary visual paths. The Transformer-based visual path aggregates information across image patches, making it suitable for judging scene relationships and layout structure; the convolutional neural network-based path enhances edges and textures, making it suitable for judging product shapes and text boundaries. Parallel inference of the two paths can significantly reduce the distortion probability of a single path in specific scenarios, making subsequent alignment and fusion more stable.

[0040] In one alternative implementation, the sampling frequency is increased from 2 frames per second to 4 frames per second, and the number of frames is increased from 12 frames to 24 frames per second, while the cover image remains at 1, for a total of 25 images. This setting is suitable for content with frequent camera transitions; if computational load needs to be maintained, a fixed downsampling interval of 2 can be used during subsequent aggregation, retaining only 12 frames and 1 cover image for calculation. Center cropping is replaced with adaptive cropping that keeps the subject centered: first, the subject area is located using a target detector with minimal cost, and then proportional scaling and cropping are performed based on the candidate box closest to the image center. This setting can improve subject sharpness in screen recordings where the subject is significantly offset. In extremely low-light or high-saturation scenes, color jitter correction with fixed parameters can be added before normalization, with brightness gain set to a fixed value between 1.1 and 1.2, and saturation suppression set to a fixed value between 0.9 and 1.0, to reduce differences in automatic exposure from the device.

[0041] For the same image, the semantic vectors are concatenated in a fixed order: first the image semantic vector, then the candidate image semantic vectors, forming a fixed-length vector. This fixed order ensures complete consistency in the stitching results across the three platforms, preventing incomparable shifts in subsequent mapping stages due to order differences. The concatenated result is then passed sequentially through two fully connected layers and a linear rectified activation function, outputting the single-frame fused image semantic vector for that image. The two-layer mapping compresses the information from the two visual paths to a uniform scale and direction without relying on learnable weight allocation; the linear rectified activation function suppresses anomalous pixels and local noise, which helps maintain numerical stability in motion-blurred frames.

[0042] The semantic vectors of the fused images from 13 single frames are aligned in a fixed order: the cover image is placed first, followed by the next 12 frames arranged chronologically. This order ensures that the cover image, as the first image seen by the user, is consistently included at the beginning of the aggregation, avoiding randomness introduced by differences in first-frame selection across different platforms. An element-wise arithmetic average is performed on the 13 vectors to obtain the final fused semantic vector. The arithmetic average assigns the same fixed contribution across all frames, without relying on any learned weights, thus ensuring repeatability and portability across the three platforms. Simultaneously, the averaging operation naturally smooths out jitter and occasional occlusion in single frames, reducing the impact of single-frame anomalies on the overall expression. The final fused semantic vector is normalized using a fixed vector norm, ensuring its length falls within a uniform scale. This fixed scale approach facilitates the use of a unified threshold and a uniform number of decimal places when aligning images and comparing across the three platforms. Performing concatenation followed by nonlinear mapping unifies the complementary information of global relationships and local details into the same vector space, avoiding channel scale inconsistencies before temporal aggregation. Placing the cover image at the beginning of the sequence and averaging it with subsequent frames can establish a stable compromise between the first screen, which has the greatest impact on user decision-making, and subsequent main screens. This ensures that the semantic vector of the final merged screen reflects both the appeal of the first screen and the thematic consistency of continued viewing.

[0043] In one alternative implementation, for content with rapid motion, a fixed-ratio truncated averaging can be used: first, discard the first and last two frames in chronological order, and then average only the vectors from the middle nine frames and the cover image (a total of 10 vectors). This setting reduces the interference of invalid frames on the expression in content with long opening transitions and ending whitespace. For static text and image content, the contribution of the cover image can be enhanced: the vector of the cover image is duplicated twice and averaged with the vectors from the remaining 12 frames (a total of 14 vectors), thereby emphasizing the text and key images on the first screen. The number of duplicates is fixed and independent of training, and the implementation process remains consistent across the three platforms.

[0044] A byte-pair encoding sub-word segmenter is used, with a vocabulary size of 30,000. This size can cover common product names, functional terms, and brand words, while maintaining a short sequence length to reduce computational latency across three platforms. Titles and descriptions are merged into a single string, uniformly converted to lowercase, and redundant spaces and invisible characters are removed, while common punctuation is retained. This cleaning aims to ensure consistent input standards across different platforms, avoiding meaningless segmentation due to input method differences. The sub-word segmenter encodes the string into a marker sequence; a start marker is added to the beginning of the sequence, and an end marker is added to the end; the maximum length is set to 77, with any exceeding this length truncated at the end, and any insufficient length padded with placeholder markers until the length reaches 77. This fixed length facilitates alignment during batch processing, ensuring identical sequence shapes across all three platforms.

[0045] Each marker is mapped to a fixed-length vector using a lookup table; a fixed positional information encoding is added to each position, enabling the model to distinguish structural positions such as the beginning, body, and end. Fixed positional encoding facilitates the reproduction of the same positional sensitivity across the three platforms. The sequence is sequentially processed through a 12-layer encoding structure, with each layer performing multi-head attention computation, residual connections and layer normalization, forward transformation, residual connections and layer normalization in sequence. Layer-by-layer stacking allows information from local word groups to be aggregated at higher phrase and sentence levels, thus stably recognizing information such as efficacy descriptions, usage scenarios, and limitations without relying on external dictionaries. The output vector at the first position of the sequence is taken as the text semantic vector. The first position is chosen because the starting marker continuously interacts with all positions throughout the sequence's computation, ultimately aggregating the overall semantics of the title and explanatory text; simultaneously, this approach avoids the uncertainty caused by selecting positions in texts of different lengths, allowing the three platforms to directly align to vectors at the same position. The sub-word segmenter has a natural advantage in product names and compound word scenarios with many out-of-vocabulary words, breaking long words down into stable sub-fragments, thus ensuring consistent recognition across different platforms and seller styles. The fixed length and fixed first position strategy ensures that the text semantic vectors remain comparable under different input lengths and different punctuation distributions, making it easier to align the text and images with the semantic vectors of the final fused image.

[0046] In one alternative implementation, when the explanatory text is generally long, the maximum length can be increased from 77 to 128. To control latency, when the explanatory text exceeds 256 characters, a fixed-rule summary can be performed first, such as retaining the first 64 characters and the most frequently occurring key phrases. When the content is mainly short headings, the end marker can be omitted, and only the start marker and the actual marker sequence are used; this approach can slightly reduce the computational load while maintaining high relevance to the first and last outputs in short text scenarios. When there is multilingual mixed input, the vocabulary size can be increased to 50,000, and non-Latin characters can be retained during the cleaning stage to cover more linguistic environments; the three platforms share the same vocabulary file to ensure completely consistent encoding results.

[0047] Example 3: For each piece of content, the final fused image semantic vector and the text semantic vector of that content are paired to form a positive sample pair according to a one-to-one correspondence. This approach ensures that the image and text of the same content have stable pairing anchors from the beginning of training, avoiding misleading by similar content in the early stages. Within the same training batch, two types of retrieval are performed on each piece of content: using the final fused image semantic vector of that content as the query, similar items are retrieved from the text semantic vectors of other content using a retrieval method based on term frequency and inverse document frequency. The top 5 items are selected from high to low similarity and form the first hard-to-negative sample pair with the final fused image semantic vector of that content. Using the text semantic vector of that content as the query, similar items are retrieved from the final fused image semantic vectors of other content. The top 5 items are selected from high to low similarity and form the second hard-to-negative sample pair with the text semantic vector of that content. The reason for selecting highly similar non-homologous pairs as hard-to-negative samples is that these pairs are most likely to cause confusion in online recommendations, such as similar titles but different images, or similar image compositions but different descriptions. Clearly distinguishing these pairs during the training phase can significantly reduce the mismatch rate when the three platforms are launched.

[0048] Each training batch contains a fixed 256 entries. Two rounds of alignment enhancement are performed on each batch in the following order: First round: Similarity narrowing: Only positive sample pairs are used. For each pair, the similarity between the image and text is calculated, increasing the similarity of homologous pairs after forward propagation and backward update. The result of this round is saved as the "narrowed alignment state". The purpose of narrowing is to establish a clear vector neighborhood of "same content", forming the reference center for subsequent separation operations. Second round: Similarity separation: Only the first and second most difficult negative sample pairs are used. For each pair, the similarity is calculated, decreasing the similarity of non-homologous pairs after forward propagation and backward update. The core of this round is to push the most similar but not identical pairs out of the neighborhood of the "narrowed alignment state", thus forming an operable safe interval in the vector space and avoiding mistaking similar interference for homologous matching during online retrieval. Each batch undergoes these two rounds sequentially for one complete enhancement. The same dataset is traversed three times. In practice, three iterations can stably establish homogeneous aggregation without introducing the risk of overfitting to certain batches of samples, facilitating consistent reproducibility across the first, second, and third platforms. Cosine similarity is used as the metric. This metric, which uses the directional consistency of vectors as its primary basis, is naturally suitable for measuring whether images and text describe the same object or scene. If, within five consecutive evaluation intervals, the average similarity increase of positive sample pairs on the validation set is less than 0.002, and the average similarity decrease of hard negative sample pairs is less than 0.002, then the current stage is considered converged, and the training proceeds to the next stage or training stops. This dual-threshold convergence criterion is simple to implement and can reproduce consistent stopping points across the three platforms. Unit-length normalization is performed on the final fused image semantic vector and text semantic vector to avoid the impact of numerical scale differences between different samples on similarity calculation. The normalization uses the same implementation across the three platforms to ensure that the numerical paths of the alignment training are completely consistent.

[0049] First, positive sample pairs are brought closer together, fixing homologous images and text in the same neighborhood; then, negative sample pairs are separated, pushing the most easily confused non-homologous pairs out of the neighborhood boundary. This sequence brings two direct effects when the system is deployed on the first, second, and third platforms: more stable retrieval and recall phases. When faced with content with highly similar titles or compositions, the system is more likely to return the correct homologous pair, rather than similar but inconsistent content. Unified content features are more comparable. Because homologous neighborhoods and interference intervals are clearly defined, the unified content features across the three platforms have a consistent correspondence during cross-platform computation, facilitating the reuse of the three-task network model across multiple platforms.

[0050] In one alternative implementation, the first 5 items can be adjusted to the first 10 to enhance the difficulty of differentiation; to control the training time, the number of traversals can be reduced from 3 to 2. When there is a lot of text noise, fixed-rule denoising can be performed on the text before retrieval to remove duplicate punctuation and terms from the marketing slogan thesaurus, and then the retrieval can be performed to reduce interference from irrelevant similarity. In a memory-constrained environment, the batch size can be adjusted from 256 to 128; to maintain the proportion of difficult-to-bear samples, the total number of difficult-to-bear samples for each item can still be fixed at 10.

[0051] The unified content features are obtained by concatenating the following elements in a fixed order: Final fused image semantic vector; Text semantic vector; One-hot encoding of platform identifiers. The first, second, and third platforms correspond to different positions in the 3D vector, with each activation position being unique; One-hot encoding of category identifiers. Taking 12 commonly used categories as an example, the one-hot vector length is 12, with only one activated position; One-hot encoding of entry identifiers. Taking 6 common entry points as an example, the one-hot vector length is 6, with only one activated position. This fixed order requires complete consistency across the three platforms, and adjustments to the positions or concatenation order are not allowed. This ensures that the three-task network model receives the same input structure on any platform, resulting in consistent behavior during the freeze and transfer joint phases. In a common implementation, the final fused image semantic vector and text semantic vector can each be 768 dimensions, the platform identifier is 3-dimensional one-hot, the category identifier is 12-dimensional one-hot, and the entry identifier is 6-dimensional one-hot, resulting in a concatenated 1557-dimensional vector. This dimension can complete a forward computation in milliseconds in a general computing environment, meeting the latency requirements of online inference on three platforms. Before entering the three-task network model, the overall vector is normalized to its unit length. This process ensures that the unified content features of different content are encoded on the same scale, facilitating numerical comparability across different batches and platforms; it also prevents a portion of the vectors from dominating subsequent mappings due to their large numerical amplitude.

[0052] When a category or entry point is temporarily missing, instead of replacing it with a zero vector, the "unknown" position is used as the activation position to avoid meaningless all-zero fragments during training and inference. This strategy can uniformly handle abnormal inputs across three platforms, reducing unpredictable discrepancies. The final fused image semantic vector and text semantic vector are placed at the beginning, allowing the semantic backbone to be read by the three-task network model first; the platform identifier, category identifier, and entry point identifier are placed at the end, serving as conditional cues and helping the network to make lightweight distinctions between platform, category, and entry point differences without changing the semantic backbone. This structure is particularly advantageous in the transfer and union phase because platform-related information is explicitly restricted to a small fragment at the end, and only the consistency of the adapted fragment needs to be considered during transfer.

[0053] In one alternative implementation, when the number of categories expands to 20, the one-hot encoding length of the category identifier is changed to 20, while the remaining concatenation order remains unchanged. When connecting to a fourth platform, the one-hot encoding length of the platform identifier is changed to 4. In existing deployments on three platforms, only one bit needs to be added; no additional modifications to the backbone of the existing three-task network model are required. When the number of entry points is refined from 6 to 10, the one-hot encoding length of the entry point identifier is changed to 10. To maintain comparability of historical data, the positions of the original 6 types in the new encoding can be retained, and mappings can be added to the newly added positions.

[0054] The shared encoding subnet consists of four layers of fully connected mappings and linear rectified activations stacked sequentially, each followed by a normalization layer and a fixed proportion of dropout units. This structure maps unified content features to an intermediate representation used for behavior judgment. The four-layer depth strikes a balance between expressive power and latency in online scenarios, typically with a single forward computation of less than 10 milliseconds. The three task subheadings correspond to click probability, conversion probability, and refund probability, respectively. They have a consistent structure, each containing two layers of fully connected mappings and linear rectified activations, ending with a standard probability mapping unit, limiting the output to the range of 0 to 1. The unified structure ensures consistent numerical behavior during training and inference, facilitating execution along the same process on the first, second, and third platforms.

[0055] The input consists of uniform content features, with a batch size of 256. This size provides a balance between throughput and convergence speed on common acceleration hardware. Training is performed iteratively in a fixed order: click, conversion, and refund. Each sub-head has exclusive update rights within its round, while other sub-heads are computed forward but not updated. This fixed order follows the objective sequence of view-click-buy-redeem, helping to first determine the judgment boundaries of early behaviors before learning subsequent behaviors, reducing the disruption of earlier judgments by later behaviors. After each round, all samples are checked: if the click probability is below 0.5 but the conversion probability is above 0.5, the threshold for the conversion task sub-head of that sample is recalculated to adjust the conversion probability to no higher than 0.5; if the conversion probability is below 0.5 but the refund probability is above 0.5, the threshold for the refund task sub-head of that sample is recalculated to adjust the refund probability to no higher than 0.5. This rule directly eliminates illogical path combinations, ensuring the three-task network model always follows the causal order of behaviors during learning. For each piece of content and each combination of platform, category, and entry point, a unified content feature is input. The results are then processed sequentially through a shared encoding subnet and three task subheadings, outputting click probability, conversion probability, and refund probability. All three outputs are uniformly retained to three decimal places and written as strings for storage, to be read by subsequent campaign allocation optimization steps. Uniform decimal places avoid inconsistencies in floating-point representation across different platforms. After batch inference, a path consistency check is performed again on the three outputs for each piece of content. If combinations that do not conform to "conversion only occurs after click" or "refund only occurs after conversion" are found, a one-time correction is performed using the same threshold strategy as in the training phase to ensure consistency before deployment.

[0056] The shared encoding subnet focuses on compressing uniform content features into abstract expressions related to user behavior. Three task subheadings are responsible for determining clicks, conversions, and refunds, respectively. Training in a fixed order during the alignment phase, with path consistency checks after each round, significantly reduces mutual interference: the click task subheading stabilizes first, the conversion task subheading learns from stable click decisions, and the refund task subheading learns from stable conversion decisions. This forward-looking constraint order aligns with real user behavior paths, thus enabling easier generalization when deployed on the first, second, and third platforms.

[0057] Example 4: Display records, click records, order records, and refund records are collected from the first, second, and third platforms. After aligning the content identifiers with timestamps, four types of event tags are formed for the same content. Each content item contains at least one display record. If a click record exists, it is recorded as a positive click; if an order record exists, it is recorded as a positive conversion; if a refund record exists, it is recorded as a positive refund. Using unified content features as input, each training batch contains 256 content items, with samples from the three platforms evenly mixed in proportions, for example, the first, second, and third platforms account for approximately 34%, 33%, and 33%, respectively. To ensure reproducibility, the random seed is fixed at 42, and the batch order is fixed in ascending order. Each training cycle trains sequentially according to the click task subheader, conversion task subheader, and refund task subheader. The shared encoding subnet participates in forward computation and updates throughout the entire cycle. This order follows the causal sequence of seeing a click, click to conversion, and conversion to refund, stabilizing the early behavior judgments in the early stages of training. At the end of each training cycle, a path consistency check is performed on each sample from all batches within that cycle. The check follows two rules: if a click is negative and then becomes positive, the sample does not conform to the click-first relationship. If a click becomes negative and then a refund becomes positive, the sample does not conform to the conversion-first relationship.

[0058] For samples that do not conform to the relationship, the threshold is recalculated according to the following fixed strategy: when a click is negative but converts to positive, the conversion probability of the sample is directly set to 0.5. When a conversion is negative but a refund is positive, the refund probability of the sample is directly set to 0.5. Since 0.5 is a commonly used decision boundary for binary classification, setting the probability to the boundary value is equivalent to withdrawing the strong judgment of the subheading for that sample, thereby eliminating the result that conflicts with the preceding behavior. Using a fixed boundary instead of a learnable threshold can achieve consistent and repeatable correction results on the first, second, and third platforms. After completing several training cycles and passing the validation set stability test, all parameters of the shared encoding subnet are fixed, and only the three task subheadings are updated. This can lock in the established semantic compression capability, allowing the mapping directly related to the output to more quickly fit the recording caliber.

[0059] Training alignment sets are constructed for each of the three task subheadings: For the click task subheading, samples with both display and click records are used as positive samples, and samples with display records but no click records are used as negative samples. For the conversion task subheading, samples with both click and order records are used as positive samples, and samples with click records but no order records are used as negative samples. For the refund task subheading, samples with both order and refund records are used as positive samples, and samples with order records but no refund records are used as negative samples. This selection method ensures that each subheading is supervised only on its directly related records, avoiding interference from cross-subheadings.

[0060] After fine-tuning, a unified calibration is performed in the order of clicks, conversions, and refunds: if the display record is positive but the click probability is below 0.5, the click probability is set to 0.5. If the click record is positive but the conversion probability is below 0.5, the conversion probability is set to 0.5. If the order record is positive but the refund probability is below 0.5, the refund probability is set to 0.5. This order ensures that the three outputs are aligned to the recorded facts on a common boundary, facilitating consistent decision-making with the same threshold when deployed on three platforms. The alignment phase executes 10 training cycles, with each cycle covering all training batches. An evaluation is performed on the validation set every two cycles. If, in two consecutive evaluations, the path consistency violation rate of the validation set decreases by less than 0.2 percentage points, and the average accuracy of clicks, conversions, and refunds all increase by less than 0.2 percentage points, the alignment phase stops and enters the freeze phase; the freeze phase stops when the accuracy of the validation set shows no improvement for three consecutive times. This condition is executed with the same logic on all three platforms and can be stably reproduced. Fixed-sequence training and calibration eliminates impossible behavior combinations early in training, preventing gradients from contradictory labels. Freezing the shared encoding subnet and then fine-tuning the three task subheadings to closely match the recording caliber ensures the output quickly aligns with real-world business events. This combination reduces bias caused by cross-platform recording differences, maintaining consistent decision boundaries across the first, second, and third platforms during online inference.

[0061] In high-risk categories, the path consistency check and sequence calibration can be increased from 0.5 to 0.6 to raise the threshold for positive conversions and positive refunds, resulting in more robust ad allocation. In memory-constrained environments, the batch size can be reduced from 256 to 128; to maintain consistency across the three platforms, all three platforms should be adjusted simultaneously. When there is a significant difference in sample size among the three platforms, oversampling can be performed on smaller sample platforms within the batch to bring the proportions of the three platforms closer to a balanced range, for example, between 33% and 34%. Perform three migrations sequentially: from platform 1 to platform 2, from platform 2 to platform 3, and from platform 3 to platform 1. This sequence ensures that each pair of platforms undergoes a targeted migration, forming a closed loop by returning to the starting point on the third migration. Prepare display records, click records, order records, and refund records for the target platform for each migration, aligned by content identifiers to form the target platform migration training set. The recommended size is no less than 100,000 items to cover the main categories and entry points of the platform.

[0062] A pre-unit for domain adaptation is concatenated before the first layer of the shared coding subnet on the target platform. This position allows for the absorption of platform-specific distribution differences without altering the main body of the shared coding subnet. The pre-unit sequentially performs input normalization, linear mapping, and nonlinear activation. The mean and standard deviation of the input normalization are calculated offline on the target platform's sample set. Linear mapping restores the unified content features to a numerical range consistent with the source platform, and nonlinear activation improves the continuity of the mapping's representation. The linear mapping is initialized as a unit mapping, and nonlinear activation uses linear rectification to ensure consistency with the input distribution of the source platform in the initial migration phase, avoiding damage to the learned representation due to excessive transformation. The click task subheader, conversion task subheader, and refund task subheader from the source platform are directly copied to the target platform as initial subheaders. The parameters of the shared coding subnet are kept consistent with the source platform. Fixed-order training and sequential calibration are repeated on the target platform, i.e., clicks first, conversions in the middle, and refunds last. After each training cycle, path consistency checks and threshold recalculation are performed, with a default threshold of 0.5. Each migration executes 3 to 5 training cycles, and the decision to stop early is based on the accuracy and path consistency violation rate on the target platform's validation set. Training stops when the accuracy improvement is less than 0.2 percentage points and the violation rate decrease is less than 0.2 percentage points. After completing the transfer from the first platform to the second platform and from the second platform to the third platform, a transfer from the third platform to the first platform is performed. The same process is used to perform a regression calibration on the first platform to eliminate the cumulative offset that may have been caused by the previous two transfers. A pre-adaptor unit is adapted in the concatenated domain before the shared coding subnet to capture and absorb the platform-specific numerical and distribution differences, thereby preserving the unified content features learned from the source platform and their mapping relationship to behavior. Reusing the three task subheadings of the source platform as the initial state can significantly shorten the number of training rounds required to achieve stable output on the target platform. Closed-loop transfer keeps the three platforms in a mutually aligned and constrained state, preventing long-term drift of any one platform.

[0063] Candidate ad units are generated using the Cartesian product of platform, category, creative, audience segment, and time period. Example: With 3 platforms, 12 categories, 50 creatives, 20 audience segments, and 24 time periods, the number of candidate ad units is 864,000. Each candidate ad unit is associated with click probability, conversion probability, and refund probability. These three probabilities are derived from the inference results of the three-task network model under the corresponding platform, category, and entry point combinations, and are uniformly retained to three decimal places. Each candidate ad unit is also associated with cost per unit, frequency cap, and category identifier. A non-negative integer decision value is defined for each candidate ad unit, representing the number of ad placements for that unit. Integer form helps to correspond to actual purchasing units, facilitating distribution and execution. The total cost of all candidate ad units does not exceed the preset budget. For example, a budget of 100,000. Budget constraints ensure controllable expenditure. The number of ad placements for each audience segment in each time period does not exceed the frequency cap for that segment. For example, the cap is 3. This constraint limits repeated reach and reduces ad fatigue. The cost percentage of each category in the overall campaign falls within a preset range. For example, the minimum percentage for each category is 2%, and the maximum is 15%. This constraint ensures broad coverage and prevents excessive concentration in a single category from causing a rapid decline in marginal returns.

[0064] Under the premise of satisfying all constraints, maximize the estimated total conversions first. This objective takes precedence, reflecting a resource allocation tendency that prioritizes transactions. In the solution set where the first objective reaches its maximum value, select the solution with the smallest estimated total refunds. This allows for prioritizing the avoidance of combinations with high refund risks without sacrificing transaction volume. In the solution set that simultaneously satisfies the first and second objectives, select the solution with the largest estimated total clicks. This step, after ensuring purchases and controlling refunds, further enhances views and clicks to stimulate upstream interest. If the integer programming solver does not directly support multiple objectives, it can be executed in three steps: first, solve for the first objective and record the optimal value; then, solve for the second objective with the added equality constraints of the first objective; finally, solve for the third objective while simultaneously satisfying the equality constraints of the first two objectives. To avoid infeasibility caused by numerical rounding, a very small margin can be set for the equality constraints, such as a tolerance between 0 and 1.

[0065] An integer programming solver supporting integer variables and linear constraints is used, with a time limit of 60 seconds. When a feasible solution is obtained within the time limit, the optimal number of deliverables for all candidate ad units is read. Candidate ad units with 0 deliverables are filtered out, and the remaining units are aggregated into ad campaign entries by platform, category, creative, audience segment, and time period. Each ad campaign entry includes platform, category, creative, audience segment, time period, and number of deliverables. Before exporting ad campaign entries, a constraint consistency check is performed: the total budget does not exceed 100,000; the number of deliverables for any audience segment in any time period does not exceed 3; the cost percentage for any category is between 2% and 15%. If any violation occurs, the solution phase is returned to add constraints or the time limit is extended for resolving. Prioritization is not dependent on manually set weighting coefficients, avoiding the problem of weight interpretation deviations across the three platforms. By first maintaining conversion rates, then reducing refund rates, and finally increasing click rates, the ad campaign strategy is more stable under the same budget and consistent with actual business objectives. Setting the decision value to a non-negative integer allows the result to be directly mapped to the delivery unit on the platform side, reducing conversion loss during delivery.

[0066] Figure 2 The overall architecture of the three-task network model used in this invention is shown. For example... Figure 2As shown, the model employs a multi-task learning architecture, comprising a shared encoding subnet and three parallel task subheadings. At the model's input, unified content features serve as the input data for the entire network. These unified content features integrate video image semantic information, text semantic information, and structured features such as platform, category, and entry point, forming a high-dimensional feature vector. The unified content features are connected to the shared encoding subnet via arrows. The shared encoding subnet consists of multiple layers of fully connected mapping layers and non-linear activation functions stacked sequentially. As shown in the figure, the shared encoding subnet includes fully connected layer 1 with ReLU activation, fully connected layer 2 with ReLU activation, and fully connected layer N with ReLU activation, forming a deep neural network structure. This multi-layer stacking design enables the model to learn deep abstract representations of content features, providing a strong feature foundation for subsequent multi-task prediction. The parameters of the shared encoding subnet are shared among the three tasks, thereby achieving cross-task knowledge transfer and improved generalization capabilities. The output of the shared encoding subnet is simultaneously fed to the three parallel task subheadings via branch connections: click prediction subheading, conversion prediction subheading, and refund prediction subheading. As shown in the figure, these three task subheadings are structurally identical. Each subheading contains multiple fully connected layers, a ReLU nonlinear activation function, and a Sigmoid probability mapping unit. The click prediction subheading predicts the probability that a user will click on the content, the conversion prediction subheading predicts the probability that a user will make a purchase after clicking, and the refund prediction subheading predicts the probability that a user will request a refund after making a purchase. Each subheading outputs its corresponding probability value, which is then used in a multi-objective integer programming model to achieve optimal resource allocation.

[0067] Figure 3 This invention details the technical solution for multi-frame feature extraction and fusion of video content. This technical solution corresponds to the image and video frame processing flow and is a core component of the content understanding processing steps. For example... Figure 3 As shown, the processing flow begins with video input. The system samples the input video frames at a preset time frequency, for example, one frame every second, thus obtaining a video frame sequence of frame 1, frame 2, frame 3, up to frame N. The figure marks the timestamps corresponding to each frame; frame 1 corresponds to t=0 seconds, frame 2 to t=1 seconds, and so on, with frame N corresponding to t=N seconds. This uniform sampling strategy ensures comprehensive coverage of the video content while controlling computational complexity. For each sampled frame, the system employs a dual-model parallel processing strategy for feature extraction. Specifically, after scaling and normalization preprocessing, each frame is simultaneously input into a Transformer-based visual model and a convolutional neural network-based model for forward inference. The Transformer model can capture the global semantic information and long-range dependencies of the image, outputting a Transformer feature vector (labeled as...). , , to Convolutional neural network models, on the other hand, excel at extracting local texture and spatial structure features, outputting CNN feature vectors (labeled as...). , , to This dual-model complementary design makes the feature representation more comprehensive and robust. After obtaining the dual-model features of each frame, the system concatenates the Transformer feature vector and the CNN feature vector of the same frame to generate a single-frame fused feature. As shown in the figure, the fused feature of frame i... From this frame and It is pieced together according to dimensions, that is The concatenated feature vectors are then non-linearly mapped through at least one fully connected layer to learn the interaction between the two types of features. Finally, the system fuses the single-frame features from all frames. to The data is aggregated chronologically to generate the final fused semantic vector F of the entire video. As shown in the annotation, temporal aggregation can be implemented using methods such as average pooling or attention mechanisms. Average pooling is simple and efficient, while attention mechanisms can adaptively assign weights to different frames, highlighting the contribution of keyframes. The final output feature vector F fully represents the semantic information of the video content, providing high-quality input features for subsequent cross-domain prediction modeling.

[0068] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A method for attracting and converting traffic from new media content, applied to cross-platform content promotion scenarios including at least a first platform, a second platform, and a third platform, characterized in that, The method includes the following steps: Content understanding processing step: Perform a unified processing flow on the content in each platform to generate unified content features that represent the content; Cross-domain prediction modeling steps: Based on the unified content features, a three-task network model is used to predict and output the target user behavior of the content through a joint training process of alignment, freezing, and transfer. The target user behavior includes click probability, conversion probability, and refund probability. Delivery allocation optimization steps: Based on the unified content features and the predicted click probability, conversion probability, and refund probability, a delivery plan is generated under budget, frequency, and category constraints, with dimensions of platform, category, creative, audience segmentation, and time period. The content understanding processing steps specifically include: performing image and video frame processing to generate image semantic information. The image and video frame processing includes: sampling multiple frames of images from the video content at a preset frequency and obtaining their cover images; scaling and normalizing each image; and performing forward inference on each normalized image using a Transformer-based visual model and a convolutional neural network-based model to generate an image semantic vector and candidate image semantic vectors. The image and video frame processing further includes: concatenating the image semantic vector and candidate image semantic vectors of the same image, and mapping them through at least one fully connected layer to obtain the image. The content understanding processing steps include: obtaining a single-frame fused image semantic vector; and aggregating the single-frame fused image semantic vectors of multiple images corresponding to the content in chronological order to obtain the final fused image semantic vector of the content; the content understanding processing steps also include performing text encoding: encoding the title and description text of the content into a token sequence using a sub-word segmenter; and inputting the token sequence into a Transformer encoder model, and taking the first and second vectors of the sequence output by the model as the text semantic vector of the content; the unified content feature is generated by concatenating the final fused image semantic vector, the text semantic vector, and the platform identifier, category identifier, and entry identifier in a preset order using one-hot encoding.

2. The method according to claim 1, characterized in that, The content understanding processing step further includes performing image-text alignment training, which includes: forming positive sample pairs by combining the semantic vector of the final fused image with the semantic vector of the text containing the same content; within the same training batch, selecting text semantic vectors from different content texts to form a first difficult negative sample pair by using a similarity-based retrieval method, and selecting final fused image semantic vectors from different content images to form a second difficult negative sample pair; and performing similarity-close training on the positive sample pairs and similarity-separation training on the first and second difficult negative sample pairs by using a contrastive learning loss function.

3. The method according to claim 2, characterized in that, The three-task network model includes a shared coding subnet and three task subheadings; the shared coding subnet is composed of multiple fully connected maps and nonlinear activation functions stacked sequentially; the three task subheadings have the same structure, and each task subheading contains multiple fully connected maps, a nonlinear activation function and a probability mapping unit.

4. The method according to claim 3, characterized in that, The alignment phase in the training process includes: cyclically training the three task sub-headers in a preset order of click, conversion, and refund, and recalculating the threshold for training samples that violate the preset path relationship at the corresponding task sub-header after each cycle. The preset path relationship requires that the conversion event must occur after the click event, and the refund event must occur after the conversion event. The freeze phase in the training process includes: fixing the parameters of the shared encoding subnet, and performing supervised fine-tuning of the outputs of the three task sub-headers based on historical display records, click records, order records, and refund records.

5. The method according to claim 4, characterized in that, The transfer joint phase in the training process includes: performing three model transfers cyclically between the first platform, the second platform, and the third platform; wherein each model transfer includes: on the target platform, concatenating a domain adaptation pre-unit for the first layer of the shared coding subnet; and on the target platform, reusing the three task subheadings of the source platform as the initial state and performing supervised training.

6. The method according to claim 5, characterized in that, The optimization steps for ad placement allocation include: constructing a multi-objective integer programming model, wherein a non-negative integer decision variable is defined for each candidate ad placement unit generated by the Cartesian product of platform, category, creative, audience segment, and time period; setting the following constraints for the model: constraining the total cost of all candidate ad placement units to be less than or equal to the preset budget; constraining the number of ad placements for each audience segment in each time period to be less than or equal to the frequency limit; and constraining the cost proportion of each category in the overall ad placement plan within a preset range; and setting a multi-objective function for the model to be solved in priority order, with the priorities being: maximizing the estimated total conversions, minimizing the estimated total refunds, and maximizing the estimated total clicks.

Images