A deep learning-based multi-channel advertisement intelligent distribution control system and method

Abstract

The application discloses a kind of multi-channel advertisement intelligent distribution control system and method based on deep learning, comprising the following steps: collecting multi-source advertisement delivery data, structured input dataset including user behavior sequence, advertisement content sequence and channel attribute sequence are constructed;Structured input dataset is encoded to feature vector, and three kinds of feature vectors are generated;Three kinds of feature vectors are input into ternary gate fusion structure, and fusion feature representation is generated;Fusion feature representation is input into improved PLE model, and click rate prediction value, conversion rate prediction value and budget cost prediction value are output;Strategy scoring function is constructed based on the above three prediction results, and the distribution score of candidate advertisement and channel combination is generated;According to distribution score ranking, the highest score combination is obtained, distribution instruction is generated and output to multi-channel advertisement system.The application improves the precision and cost control ability of advertisement distribution strategy under multi-target delivery, and is suitable for large-scale online advertisement delivery scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and advertising technology, and in particular to a multi-channel intelligent advertising distribution control system and method based on deep learning. Background Technology

[0002] With the rapid development of digital marketing and the continuous expansion of the multi-platform advertising ecosystem, advertising strategies are showing a trend towards diversification, real-time updates, and precision. To improve advertising efficiency and conversion rates, more and more platforms are introducing AI-based recommendation and control mechanisms to model and predict multi-source data such as user behavior, ad content, and channel attributes, thus assisting in intelligent ad delivery. However, most existing ad distribution methods employ static rules or single-goal-oriented models, optimizing only a single point based on user interests or conversion intentions. They fail to fully integrate user behavior sequences, ad content structure, and channel attribute characteristics, making it difficult to comprehensively model complex multi-objective task relationships.

[0003] Furthermore, traditional multi-task learning methods often use a uniform path modeling approach when dealing with the strategic differences brought about by different creative lifecycles, ignoring the performance differences between new and historical creatives in the delivery strategy, resulting in unreasonable allocation of strategy weights. At the same time, in the context of multi-channel delivery, the budget and frequency control strategies of different platforms are often simply spliced ​​into the model input, lacking targeted structural modeling, which limits the ability to coordinate and optimize targets such as budget consumption, click-through rate, and conversion rate.

[0004] Therefore, how to provide a multi-channel intelligent advertising distribution control system and method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-channel advertising intelligent distribution control system and method based on deep learning. This invention constructs a structured input dataset using user behavior sequences, advertising content sequences, and channel attribute sequences, integrates a ternary gating mechanism and an improved PLE model to generate multi-objective prediction results, and achieves intelligent matching and distribution control of advertising-channel combinations by introducing a task-specific expert structure and strategy scoring function that separates the paths of new and old creative materials. It has the advantages of balancing conversion rate, click-through rate and budget cost optimization, adapting to differences in creative material lifecycles, and improving the accuracy of multi-platform distribution.

[0006] A method for intelligent multi-channel advertising distribution control based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Collect multi-source advertising data and construct a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences;

[0008] The structured input dataset is encoded using feature vectorization to generate user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors.

[0009] Input the user behavior feature vector, the advertising content feature vector, and the channel attribute feature vector into the ternary gating fusion structure to generate a fused feature representation;

[0010] The fused feature representation is input into the improved PLE model, which outputs the predicted click-through rate, predicted conversion rate, and predicted budget cost, respectively.

[0011] A strategy scoring function is constructed based on the predicted click-through rate, conversion rate, and budget cost to generate distribution scores for candidate ads and channel combinations.

[0012] The candidate ads and channel combinations are sorted according to the distribution score, and the candidate ads and channel combinations with the highest scores are selected to generate distribution instructions.

[0013] Optionally, the step of collecting multi-source advertising delivery data and constructing a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences specifically includes:

[0014] Collect user behavior data, including user exposure records, click behavior, conversion behavior, dwell time and behavior timestamps on various advertising channels;

[0015] Collect advertising content data, including the text content, image or video content, creation time of the creative, advertising target type, and category tags to which the creative belongs;

[0016] Collect channel attribute data, including ad display platform type, ad placement identifier, ad time period, budget control strategy and frequency control parameters;

[0017] The system performs field standardization, time series alignment, and feature labeling on user behavior data, advertising content data, and channel attribute data, resulting in a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences.

[0018] Optionally, the step of performing feature vectorization encoding on the structured input dataset to generate user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors specifically includes:

[0019] For each record in the user behavior sequence, field-level embedding mapping is performed. For exposure records, click behavior, and conversion behavior, lookup table embedding is used. For dwell time, normalized numerical encoding is used. Behavior timestamps are converted into time intervals and introduced into position vector encoding. All field vectors are concatenated in chronological order to form a behavior embedding matrix. The behavior embedding matrix is ​​input into a single-layer bidirectional long short-term memory network to obtain the forward and reverse hidden states respectively. After concatenation, average pooling is performed to output the user behavior feature vector.

[0020] The text content in the ad content sequence is input into the BERT model to obtain the text semantic vector. The image or video content is input into the ResNet network to obtain the visual feature vector. The creative creation time, ad target type and creative category label are respectively input into the independent lookup table embedding structure to generate the structure field embedding vector. The text semantic vector, visual feature vector and structure field embedding vector are concatenated and input into the fusion fully connected network. The fusion fully connected network consists of a linear transformation layer and a ReLU activation function, which unifies the dimensions of each modality vector and completes information interaction, and outputs the ad content feature vector.

[0021] The ad display platform type, ad slot identifier, and ad placement time period are input into a lookup table embedding structure to generate a semantic embedding vector. All field values ​​in the budget control strategy and frequency control parameters are normalized and concatenated to form an ad placement strategy value vector. The semantic embedding vector and the ad placement strategy value vector are concatenated to form a complete channel attribute representation vector. This vector is then input into a two-layer feedforward neural network. The first layer uses a 128-dimensional linear mapping, ReLU activation, and Dropout operation, while the second layer uses a 64-dimensional linear mapping and ReLU activation. The output is a channel attribute feature vector with a unified dimension.

[0022] Optionally, the three-element gating fusion structure specifically includes a user behavior channel, an advertising content channel, and a channel attribute channel. The gating weight vectors are calculated separately and then weighted and fused. Finally, a fused feature representation is generated through a fully connected network.

[0023] Optionally, the step of inputting the fused feature representation into the improved PLE model and outputting the click-through rate prediction, conversion rate prediction, and budget cost prediction respectively includes:

[0024] An improved PLE model is constructed, which includes a shared expert network, a task-specific expert network that separates new and old materials, a task gating allocation structure, and a policy-aware fusion enhancement structure.

[0025] The fused feature representations are input into a shared expert network, which consists of multiple feedforward neural networks with different structures. Each feedforward neural network consists of two linear mapping layers and a ReLU activation function. The hidden dimensions, parameter initialization methods, and regularization configurations of the two linear mapping layers differ. The network outputs multiple shared expert feature representations, which together form a set of general feature representations.

[0026] The task-specific expert network is set up with a new and old creative path separation structure. The path selection is controlled by the new and old identifiers parsed from the creative creation time field in the ad content sequence, and outputs a set of new creative specific feature representations and a set of old creative specific feature representations.

[0027] The task gating allocation structure is constructed using a gating weight allocation method. The fused feature representation is input to the linear transformation layer and mapped to the task-aware representation vector. The task-aware representation vector is concatenated with the old and new labels and input to a gating scoring layer with Softmax normalization. The output is a click-through rate gating vector, a conversion rate gating vector, and a budget cost gating vector. These are applied to the general feature representation set, the new material-specific feature representation set, and the old material-specific feature representation set through a weighted fusion method, respectively, to generate feature representations for the three types of tasks.

[0028] The strategy-aware fusion enhancement structure includes a set of strategy-aware vector construction units and a splicing fusion network. The strategy-aware vector construction unit takes the freshness index of the advertising creative, the display frequency statistics, and the historical behavior confidence vector as input. Through standardization and a fully connected layer, it generates a strategy compensation vector. The freshness index is calculated from the difference between the creative creation time and the current delivery time.

[0029] The concatenation and fusion network concatenates the policy compensation vector with the feature representation of each task and inputs it into a two-layer feedforward neural network, outputting the fused and enhanced click-through rate feature representation, conversion rate feature representation, and budget cost feature representation;

[0030] The enhanced click-through rate (CTR), conversion rate (CTR), and budget cost (MCC) feature representations are input into three independent linear output layers to generate CTR predictions, conversion rate predictions, and MCC predictions, respectively.

[0031] Optionally, the task-specific expert network is configured with a new and old material path separation structure, specifically including:

[0032] Based on the creative creation time field in the ad content sequence, a new / old creative identifier variable is constructed to determine whether the creative to which each sample belongs is a new creative.

[0033] Set a predefined time threshold. When the creation time of the material is less than or equal to the current time, set the material new / old identifier variable to 1 to indicate a new material; otherwise, set it to 0 to indicate an old material.

[0034] In parallel, new material expert subnetworks and old material expert subnetworks are constructed. Each subnetwork consists of two feedforward neural networks with different numbers of neurons, which receive fused feature representations and output sets of new material-specific feature representations and sets of old material-specific feature representations, respectively.

[0035] Optionally, the step of constructing a strategy scoring function based on predicted click-through rate, predicted conversion rate, and predicted budget cost to generate distribution scores for candidate ads and channel combinations specifically includes:

[0036] Input the predicted click-through rate, conversion rate, and budget cost into the strategy scoring function to generate a distribution score;

[0037] Based on the distribution score, and combining the ad content sequence and the channel attribute sequence, a matching combination of each ad and its corresponding available channels is constructed, and a corresponding distribution score is assigned to each combination.

[0038] A multi-channel advertising intelligent distribution control system based on deep learning according to an embodiment of the present invention includes the following modules:

[0039] The data acquisition and preprocessing module is used to collect multi-source advertising data and perform field standardization, time series alignment, and feature labeling.

[0040] The feature vector encoding module is used to perform feature vector encoding on user behavior sequences, advertising content sequences, and channel attribute sequences respectively.

[0041] The three-element gating fusion structure module is used to input user behavior feature vectors, advertising content feature vectors and channel attribute feature vectors into the three-element gating fusion structure and output the fused feature representation.

[0042] The improved PLE model prediction module is used to input the fused feature representation into the improved PLE model and output the predicted click-through rate, conversion rate, and budget cost respectively.

[0043] The strategy scoring generation module is used to construct a strategy scoring function based on the predicted click-through rate, conversion rate, and budget cost, and generate a distribution score under multi-objective optimization.

[0044] The optimization and instruction generation module is used to optimize the distribution of candidate advertisements and channel combinations based on their distribution scores, select the combination with the highest score, and generate distribution instructions.

[0045] The beneficial effects of this invention are:

[0046] (1) This invention introduces a joint structure of fusion feature representation and improved PLE model to achieve multi-objective prediction of click-through rate, conversion rate and budget cost. It solves the problem that the existing advertising system is dominated by a single optimization objective and it is difficult to fully take into account the advertising effect and cost efficiency. It significantly improves the overall performance of strategy decision-making.

[0047] (2) By constructing a new and old material path separation structure in the task-specific expert network and dynamically selecting expert sub-networks for feature extraction based on material creation time, the system can achieve differentiated modeling according to the material life cycle stage, which enhances the adaptability to the cold start of new materials and the performance degradation of old materials, and improves the material utilization efficiency and deployment continuity.

[0048] (3) In the process of strategy scoring, the predicted click-through rate, the predicted conversion rate and the predicted budget cost are integrated to construct a strategy scoring function, and candidate ads and channel combinations are constructed based on the scoring value. This effectively enhances the model's ability to make multi-channel distribution decisions and improves the accuracy of distribution strategies and cross-platform deployment effects. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a multi-channel advertising intelligent distribution control method based on deep learning proposed in this invention;

[0051] Figure 2 This is a framework diagram of the improved PLE model in a deep learning-based intelligent distribution control method for multi-channel advertising proposed in this invention.

[0052] Figure 3 This is a structural diagram of a multi-channel advertising intelligent distribution control system based on deep learning proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1-2 A deep learning-based intelligent distribution control method for multi-channel advertising includes the following steps:

[0055] Step 1: Collect multi-source advertising data and construct a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences;

[0056] Step 2: Perform feature vectorization encoding on the structured input dataset to generate user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors;

[0057] Step 3: Input the user behavior feature vector, the advertising content feature vector, and the channel attribute feature vector into the ternary gating fusion structure to generate the fused feature representation;

[0058] Step 4: Input the fused feature representation into the improved PLE model, and output the predicted click-through rate, predicted conversion rate, and predicted budget cost respectively;

[0059] Step 5: Construct a strategy scoring function based on the predicted click-through rate, conversion rate, and budget cost to generate distribution scores for candidate ads and channel combinations;

[0060] Step 6: Sort the candidate ads and channel combinations according to the distribution score, select the candidate ads and channel combinations with the highest scores, and generate distribution instructions.

[0061] In this embodiment, the step of collecting multi-source advertising data and constructing a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences specifically includes:

[0062] Collect user behavior data, including user exposure records, click behavior, conversion behavior, dwell time and behavior timestamps on various advertising channels;

[0063] Collect advertising content data, including the text content, image or video content, creation time of the creative, advertising target type, and category tags to which the creative belongs;

[0064] Collect channel attribute data, including ad display platform type, ad placement identifier, ad time period, budget control strategy and frequency control parameters;

[0065] The system performs field standardization, time series alignment, and feature labeling on user behavior data, advertising content data, and channel attribute data, resulting in a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences.

[0066] This implementation method constructs a structured input dataset with a clear structure and consistent time sequence by performing field standardization, time series alignment, and feature labeling operations on the collected user behavior data, advertising content data, and channel attribute data. This provides a unified input foundation for subsequent feature vectorization encoding and model learning.

[0067] In this embodiment, the step of performing feature vectorization encoding on the structured input dataset to generate user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors specifically includes:

[0068] For each record in the user behavior sequence, field-level embedding mapping is performed. For exposure records, click behavior, and conversion behavior, lookup table embedding is used. For dwell time, normalized numerical encoding is used. Behavior timestamps are converted into time intervals and introduced into position vector encoding. All field vectors are concatenated in chronological order to form a behavior embedding matrix. The behavior embedding matrix is ​​input into a single-layer bidirectional long short-term memory network to obtain the forward and reverse hidden states respectively. After concatenation, average pooling is performed to output the user behavior feature vector.

[0069] The text content in the ad content sequence is input into the BERT model to obtain the text semantic vector. The image or video content is input into the ResNet network to obtain the visual feature vector. The creative creation time, ad target type and creative category label are respectively input into the independent lookup table embedding structure to generate the structure field embedding vector. The text semantic vector, visual feature vector and structure field embedding vector are concatenated and input into the fusion fully connected network. The fusion fully connected network consists of a linear transformation layer and a ReLU activation function, which unifies the dimensions of each modality vector and completes information interaction, and outputs the ad content feature vector.

[0070] The ad display platform type, ad slot identifier, and ad placement time period are input into a lookup table embedding structure to generate a semantic embedding vector. All field values ​​in the budget control strategy and frequency control parameters are normalized and concatenated to form an ad placement strategy value vector. The semantic embedding vector and the ad placement strategy value vector are concatenated to form a complete channel attribute representation vector. This vector is then input into a two-layer feedforward neural network. The first layer uses a 128-dimensional linear mapping, ReLU activation, and Dropout operation, while the second layer uses a 64-dimensional linear mapping and ReLU activation. The output is a channel attribute feature vector with a unified dimension.

[0071] This implementation enhances the contextual awareness of behavioral sequences by combining field-level embedding, temporal location encoding, and bidirectional long short-term memory networks in user behavior modeling; it integrates BERT semantic encoding, ResNet visual feature extraction, and structural field embedding in advertising content modeling, and achieves unified cross-modal information expression by fusing fully connected networks; and it combines lookup table embedding and normalized numerical encoding in channel attribute modeling, and introduces a two-layer feedforward neural network with differentiated structures to enhance the expression accuracy and nonlinear modeling capability of channel control features.

[0072] In this embodiment, the three-element gating fusion structure specifically includes a user behavior channel, an advertising content channel, and a channel attribute channel. The gating weight vectors are calculated separately and then weighted and fused. Finally, a fused feature representation is generated through a fully connected network.

[0073] This implementation improves the interaction efficiency and expression consistency between multi-source features by designing a ternary gating fusion structure, providing a structurally stable fusion feature representation for subsequent prediction modeling.

[0074] In this embodiment, the step of inputting the fused feature representation into the improved PLE model and outputting the click-through rate prediction, conversion rate prediction, and budget cost prediction respectively includes:

[0075] An improved PLE model is constructed, which includes a shared expert network, a task-specific expert network that separates new and old materials, a task gating allocation structure, and a policy-aware fusion enhancement structure.

[0076] The fused feature representations are input into a shared expert network, which consists of multiple feedforward neural networks with different structures. Each feedforward neural network consists of two linear mapping layers and a ReLU activation function. The hidden dimensions, parameter initialization methods, and regularization configurations of the two linear mapping layers differ. The network outputs multiple shared expert feature representations, which together form a set of general feature representations.

[0077] The task-specific expert network is set up with a new and old creative path separation structure. The path selection is controlled by the new and old identifiers parsed from the creative creation time field in the ad content sequence, and outputs a set of new creative specific feature representations and a set of old creative specific feature representations.

[0078] The task gating allocation structure is constructed using a gating weight allocation method. The fused feature representation is input to the linear transformation layer and mapped to the task-aware representation vector. The task-aware representation vector is concatenated with the old and new labels and input to a gating scoring layer with Softmax normalization. The output is a click-through rate gating vector, a conversion rate gating vector, and a budget cost gating vector. These are applied to the general feature representation set, the new material-specific feature representation set, and the old material-specific feature representation set through a weighted fusion method, respectively, to generate feature representations for the three types of tasks.

[0079] The strategy-aware fusion enhancement structure includes a set of strategy-aware vector construction units and a splicing fusion network. The strategy-aware vector construction unit takes the freshness index of the advertising creative, the display frequency statistics, and the historical behavior confidence vector as input. Through standardization and a fully connected layer, it generates a strategy compensation vector. The freshness index is calculated from the difference between the creative creation time and the current delivery time.

[0080] The concatenation and fusion network concatenates the policy compensation vector with the feature representation of each task and inputs it into a two-layer feedforward neural network, outputting the fused and enhanced click-through rate feature representation, conversion rate feature representation, and budget cost feature representation;

[0081] The enhanced click-through rate (CTR), conversion rate (CTR), and budget cost (MCC) feature representations are input into three independent linear output layers to generate CTR predictions, conversion rate predictions, and MCC predictions, respectively.

[0082] This implementation constructs an improved PLE model, integrating feature representations into a task-specific expert network and a shared expert network. The task-specific expert network employs a path separation structure between new and old content to enhance adaptability to the heterogeneity of new and old content; the shared expert network adopts a unified structure and serves all tasks. By introducing a task gating structure, click-through rate prediction, conversion rate prediction, and budget cost prediction are generated separately after fusing the outputs of each expert, achieving multi-objective joint modeling and balanced performance improvement under scenarios with differences in content lifecycles.

[0083] In this embodiment, the task-specific expert network is configured as a new and old material path separation structure, specifically including:

[0084] Based on the creative creation time field in the ad content sequence, a new / old creative identifier variable is constructed to determine whether the creative to which each sample belongs is a new creative.

[0085] Set a predefined time threshold. When the creation time of the material is less than or equal to the current time, set the material new / old identifier variable to 1 to indicate a new material; otherwise, set it to 0 to indicate an old material.

[0086] In parallel, a new material expert subnetwork and an old material expert subnetwork are constructed. Each subnetwork consists of two feedforward neural networks with different numbers of neurons, which receive fused feature representations and output sets of new material-specific feature representations and sets of old material-specific feature representations, respectively.

[0087] This implementation introduces new and old material identifier variables, and determines the material lifecycle status based on the material creation time field in the ad content sequence and a predefined time threshold, accurately distinguishing between new and old materials. Furthermore, it constructs new material expert subnetworks and old material expert subnetworks in parallel, each containing two layers of feedforward neural networks with different structures. This enables materials with different lifecycles to obtain differentiated feature modeling capabilities in terms of structure, effectively improving the ability of fused features to represent the characteristics of new and old materials, thereby enhancing the model's adaptability and response to diverse ad materials in distribution strategies.

[0088] In this embodiment, the step of constructing a strategy scoring function based on the predicted click-through rate, predicted conversion rate, and predicted budget cost to generate distribution scores for candidate ads and channel combinations specifically includes:

[0089] The predicted click-through rate, conversion rate, and budget cost are input into the strategy scoring function to generate a distribution score. The strategy scoring Score function is constructed as follows:

[0090]

[0091] Where CTR represents the predicted click-through rate, CVR represents the predicted conversion rate, Cost represents the predicted budget cost, α, β, and γ are task adjustment weights, set according to the advertising optimization strategy, satisfying α>0, β>0, γ>0, and σ(·) is the Sigmoid normalization function;

[0092] Based on the distribution score, and combining the ad content sequence and the channel attribute sequence, a matching combination of each ad and its corresponding available channels is constructed, and a corresponding distribution score is assigned to each combination.

[0093] This implementation constructs a strategy scoring function, using the predicted click-through rate, conversion rate, and budget cost output by the improved PLE model as input factors. It employs a normalized weighted scoring mechanism combined with business weight coefficients to quantitatively evaluate the distribution potential of each ad and channel combination. Without introducing ranking operations, it dynamically generates scoring results based on each set of predicted values ​​and defines a set of candidate ad and channel combinations accordingly. This effectively achieves the coupling construction of distribution scoring and combination definition, providing a structured foundation for subsequent optimal decision-making and improving the scoring mechanism's adaptability to multi-objective balance and distribution efficiency.

[0094] refer to Figure 3 A multi-channel advertising intelligent distribution control system based on deep learning includes the following modules:

[0095] The data acquisition and preprocessing module is used to collect multi-source advertising data and perform field standardization, time series alignment, and feature labeling.

[0096] The feature vector encoding module is used to perform feature vector encoding on user behavior sequences, advertising content sequences, and channel attribute sequences respectively.

[0097] The three-element gating fusion structure module is used to input user behavior feature vectors, advertising content feature vectors and channel attribute feature vectors into the three-element gating fusion structure and output the fused feature representation.

[0098] The improved PLE model prediction module is used to input the fused feature representation into the improved PLE model and output the predicted values ​​of click-through rate, conversion rate, and budget cost, respectively.

[0099] The strategy scoring generation module is used to construct a strategy scoring function based on the predicted click-through rate, conversion rate, and budget cost, and generate a distribution score under multi-objective optimization.

[0100] The optimization and instruction generation module is used to optimize the distribution of candidate advertisements and channel combinations based on their distribution scores, select the combination with the highest score, and generate distribution instructions.

[0101] Example 1:

[0102] To verify the feasibility of this invention in practice, it was applied to a 30-day multi-channel advertising campaign conducted by an advertising agency in a certain location. This campaign promoted product advertisements across multiple short video platforms (Platform A, Platform B, and Platform C), encompassing multiple ad creatives, budget strategies, and combinations of channel attributes. The operator aimed to improve ad click-through rates and conversion rates while maintaining budget control, and to resolve issues such as the low initial weight of new ad creatives and conflicts with channel distribution strategies.

[0103] In this experiment, the collected historical advertising data was first processed, including user behavior data (browsing, clicks, conversions, etc.) from the past 90 days, advertising content data (material ID, type, creation time, cover image embedding vector, etc.), and channel attribute data (platform category, historical frequency of placement, budget limit, etc.). After standardization, time alignment, and label construction, a structured input dataset was generated. Subsequently, feature vectorization encoding was used to construct user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors, respectively.

[0104] Next, a ternary gated fusion structure is input, and the weights of the three types of features are dynamically adjusted using a gating mechanism to output a fused feature representation, which is then input into an improved PLE model. This model includes a shared expert network, a new creative expert sub-network, and an old creative expert sub-network. The new / old attribute of the creative is controlled by the creative creation time field, and path separation modeling is performed using structurally differentiated feedforward neural networks. Through gated fusion and multi-expert structure joint learning, predicted click-through rate, conversion rate, and budget cost are output, respectively. Based on the scoring results, candidate ad and channel combinations are automatically constructed, and distribution scores are generated. In the next step of generating distribution instructions, the scores are used to push the data to the multi-platform system for actual distribution. Table 1 shows some of the delivery score data generated by the system during the experiment:

[0105] Table 1 Scoring Data for Candidate Combinations of Multi-Channel Advertising Implementation

[0106] Ad ID channel Predicted click-through rate Conversion rate forecast Cost forecast Strategy Score AD001 Platform A 0.112 0.031 0.039 0.1364 AD002 Platform B 0.094 0.027 0.028 0.1206 AD003 Platform A 0.089 0.019 0.020 0.1068 AD004 C platform 0.132 0.034 0.045 0.1551 AD005 Platform B 0.076 0.022 0.018 0.0992

[0107] As shown in Table 1, AD004, due to its high conversion and click-through rates on Platform C, achieved the highest overall score (0.1551) despite its slightly higher cost, and was therefore prioritized in the final distribution. In contrast, AD005, while having the lowest cost, had weaker clicks and conversions, resulting in a lower score and automatic allocation in descending order. The scoring mechanism effectively achieved optimal control under the three-dimensional constraints of clicks, conversions, and budget.

[0108] To further evaluate the advantages of this invention in controlling the lifecycle of advertising materials, we compared and analyzed the actual distribution efficiency of new advertising materials and historical materials within 10 days before the campaign. The evaluation indicators included actual click-through rate, actual conversion rate, and average cost.

[0109] Table 2 Comparison of the effects of new and old materials.

[0110] Material Category Average actual click-through rate Average actual conversion rate Average cost per click (RMB) new material 0.107 0.031 0.42 old material 0.112 0.032 0.41

[0111] Analysis of Table 2 shows that after adopting the improved PLE new and old material path separation structure and delivery scoring mechanism of this invention, the click-through rate and conversion rate of the new materials are basically close to those of the old materials, and the delivery cost is comparable. This indicates that the system effectively alleviates the problems of low initial weight and cold start difficulty for new materials, and improves the delivery efficiency of new materials. In traditional systems, new materials are often marginalized due to their limited history, while this model, through the identification of the material creation time field and the distinguishable training of the expert structure, achieves improved delivery fairness under the awareness of the life cycle.

[0112] Overall, within the 30-day experimental period, the method of this invention achieved an average click-through rate increase of 5.7%, an average conversion rate increase of 4.2%, and budget usage deviation rate controlled within ±7.3%, significantly outperforming traditional placement strategies based on manual rules or fixed channel weights. This embodiment demonstrates that the method of this invention has the following significant advantages in multi-channel placement tasks: it can achieve dynamic balance optimization between conversion rate, click-through rate, and budget control; it has the ability to separate new and old creative paths and adapt to cold starts; it supports automated, high-precision ad-channel strategy matching and scoring decisions, and has high practical application value and industry promotion potential.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-channel advertising intelligent distribution control method based on deep learning, characterized in that, Includes the following steps: Collect multi-source advertising data and construct a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences; The structured input dataset is encoded using feature vectorization to generate user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors. Input the user behavior feature vector, the advertising content feature vector, and the channel attribute feature vector into the ternary gating fusion structure to generate a fused feature representation; The fused feature representation is input into the improved PLE model, which outputs the predicted click-through rate, predicted conversion rate, and predicted budget cost, respectively. A strategy scoring function is constructed based on the predicted click-through rate, conversion rate, and budget cost to generate distribution scores for candidate ads and channel combinations. The candidate ads and channel combinations are sorted according to the distribution score, and the candidate ads and channel combinations with the highest scores are selected to generate distribution instructions; The improved PLE model includes a shared expert network, a task-specific expert network that separates new and old materials, a task gating allocation structure, and a policy-aware fusion enhancement structure. The fused feature representations are input into a shared expert network, which consists of multiple feedforward neural networks with different structures. Each feedforward neural network consists of two linear mapping layers and a ReLU activation function. The hidden dimensions, parameter initialization methods, and regularization configurations of the two linear mapping layers differ. The network outputs multiple shared expert feature representations, which together form a set of general feature representations. The task-specific expert network is set up with a new and old creative path separation structure. The path selection is controlled by the new and old identifiers parsed from the creative creation time field in the ad content sequence, and outputs a set of new creative specific feature representations and a set of old creative specific feature representations. The task gating allocation structure is constructed using a gating weight allocation method. The fused feature representation is input to the linear transformation layer and mapped to the task-aware representation vector. The task-aware representation vector is concatenated with the new and old labels and input to a gating scoring layer with Softmax normalization. The output is a click-through rate gating vector, a conversion rate gating vector, and a budget cost gating vector. These are applied to the general feature representation set, the new material-specific feature representation set, and the old material-specific feature representation set through a weighted fusion method, respectively, to generate feature representations for the three types of tasks. The strategy-aware fusion enhancement structure includes a set of strategy-aware vector construction units and a splicing fusion network. The strategy-aware vector construction unit takes the freshness index of the advertising creative, the display frequency statistics, and the historical behavior confidence vector as input. Through standardization and a fully connected layer, it generates a strategy compensation vector. The freshness index is calculated from the difference between the creative creation time and the current delivery time. The concatenation and fusion network concatenates the policy compensation vector with the feature representation of each task and inputs it into a two-layer feedforward neural network, outputting the fused and enhanced click-through rate feature representation, conversion rate feature representation, and budget cost feature representation; The enhanced click-through rate (CTR), conversion rate (CTR), and budget cost (MCC) feature representations are input into three independent linear output layers to generate CTR predictions, conversion rate predictions, and MCC predictions, respectively.

2. The multi-channel advertising intelligent distribution control method based on deep learning according to claim 1, characterized in that, The collection of multi-source advertising data, constructing a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences, specifically includes: Collect user behavior data, including user exposure records, click behavior, conversion behavior, dwell time and behavior timestamps on various advertising channels; Collect advertising content data, including the text content, image or video content, creation time of the creative, advertising target type, and category tags to which the creative belongs; Collect channel attribute data, including ad display platform type, ad placement identifier, ad time period, budget control strategy and frequency control parameters; The system performs field standardization, time series alignment, and feature labeling on user behavior data, advertising content data, and channel attribute data, resulting in a structured input dataset containing user behavior sequences, advertising content sequences, and channel attribute sequences.

3. The multi-channel advertising intelligent distribution control method based on deep learning according to claim 1, characterized in that, The step of performing feature vectorization encoding on the structured input dataset to generate user behavior feature vectors, advertising content feature vectors, and channel attribute feature vectors specifically includes: For each record in the user behavior sequence, field-level embedding mapping is performed. For exposure records, click behavior, and conversion behavior, lookup table embedding is used. For dwell time, normalized numerical encoding is used. Behavior timestamps are converted into time intervals and introduced into position vector encoding. All field vectors are concatenated in chronological order to form a behavior embedding matrix. The behavior embedding matrix is ​​input into a single-layer bidirectional long short-term memory network to obtain the forward and reverse hidden states respectively. After concatenation, average pooling is performed to output the user behavior feature vector. The text content in the ad content sequence is input into the BERT model to obtain the text semantic vector. The image or video content is input into the ResNet network to obtain the visual feature vector. The creative creation time, ad target type and creative category label are respectively input into the independent lookup table embedding structure to generate the structure field embedding vector. The text semantic vector, visual feature vector and structure field embedding vector are concatenated and input into the fusion fully connected network. The fusion fully connected network consists of a linear transformation layer and a ReLU activation function, which unifies the dimensions of each modality vector and completes information interaction, and outputs the ad content feature vector. The ad display platform type, ad slot identifier, and ad placement time period are input into a lookup table embedding structure to generate a semantic embedding vector. All field values ​​in the budget control strategy and frequency control parameters are normalized and concatenated to form an ad placement strategy value vector. The semantic embedding vector and the ad placement strategy value vector are concatenated to form a complete channel attribute representation vector. This vector is then input into a two-layer feedforward neural network. The first layer uses a 128-dimensional linear mapping, ReLU activation, and Dropout operation, while the second layer uses a 64-dimensional linear mapping and ReLU activation. The output is a channel attribute feature vector with a unified dimension.

4. The multi-channel advertising intelligent distribution control method based on deep learning according to claim 1, characterized in that, The three-element gating fusion structure specifically includes a user behavior channel, an advertising content channel, and a channel attribute channel. The gating weight vectors are calculated separately and then weighted and fused. Finally, a fused feature representation is generated through a fully connected network.

5. The multi-channel advertising intelligent distribution control method based on deep learning according to claim 1, characterized in that, The task-specific expert network is configured with a new and old material path separation structure, specifically including: Based on the creative creation time field in the ad content sequence, a new / old creative identifier variable is constructed to determine whether the creative to which each sample belongs is a new creative. Set a predefined time threshold. When the creation time of the material is less than or equal to the current time, set the material new / old identifier variable to 1 to indicate a new material; otherwise, set it to 0 to indicate an old material. In parallel, new material expert subnetworks and old material expert subnetworks are constructed. Each subnetwork consists of two feedforward neural networks with different numbers of neurons, which receive fused feature representations and output sets of new material-specific feature representations and sets of old material-specific feature representations, respectively.

6. The multi-channel advertising intelligent distribution control method based on deep learning according to claim 1, characterized in that, The strategy scoring function, constructed based on predicted click-through rate, predicted conversion rate, and predicted budget cost, generates distribution scores for candidate ads and channel combinations, specifically including: Input the predicted click-through rate, conversion rate, and budget cost into the strategy scoring function to generate a distribution score; Based on the distribution score, and combining the ad content sequence and the channel attribute sequence, a matching combination of each ad and its corresponding available channels is constructed, and a corresponding distribution score is assigned to each combination.

7. A deep learning-based intelligent multi-channel advertising distribution control system, executing the deep learning-based intelligent multi-channel advertising distribution control method according to any one of claims 1 to 6, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect multi-source advertising data and perform field standardization, time series alignment, and feature labeling. The feature vector encoding module is used to perform feature vector encoding on user behavior sequences, advertising content sequences, and channel attribute sequences respectively. The three-element gating fusion structure module is used to input user behavior feature vectors, advertising content feature vectors and channel attribute feature vectors into the three-element gating fusion structure and output the fused feature representation. The improved PLE model prediction module is used to input the fused feature representation into the improved PLE model and output the predicted values ​​of click-through rate, conversion rate, and budget cost, respectively. The strategy scoring generation module is used to construct a strategy scoring function based on the predicted click-through rate, conversion rate, and budget cost, and generate a distribution score under multi-objective optimization. The optimization and instruction generation module is used to optimize the distribution of candidate advertisements and channel combinations based on their distribution scores, select the combination with the highest score, and generate distribution instructions.

Images