Bidding winning rate prediction method and system based on deep learning

Abstract

The invention relates to the field of data processing, and relates to a bidding winning rate prediction method and system based on deep learning. The method comprises the following steps: inputting an original price and a data set into a bidding winning rate prediction model to obtain a winning rate; the bidding winning rate prediction model comprises an embedded layer, a feature crossing module, a first splicing module, a second splicing module, a mu network, a sigma network, a winning rate prediction module and a plurality of expert networks. By adopting the method provided by the invention, an expected bid winning probability value with relatively high accuracy can be finally output, and the uncertainty of the probability value can be reflected, so that a more robust basis is provided for a downstream bid decision, and particularly, stronger robustness is shown in a cold start scene, a data missing scene or an environment dramatic change scene.

Description

Technical Field

[0001] The present invention relates to the field of data processing. More specifically, the present invention relates to a bidding win rate prediction method and system based on deep learning. Background Art

[0002] In programmatic advertising systems, real-time bidding (RTB) has become the mainstream method for advertising transactions. Every time a user visits a webpage or opens an app, the ad system completes an auction for ad space within milliseconds. Advertisers bid based on the context of the impression opportunity, and the system decides whether to display the ad based on the bid amount and relevant strategies. In this process, the predicted probability of winning the bid (win rate) at the current bid price becomes a key decision-making indicator, directly affecting advertisers' budget efficiency and advertising platforms' profitability.

[0003] However, real-world applications present multiple challenges. First, the dynamic nature of the advertising delivery environment makes traditional static modeling methods difficult to adapt to rapidly changing bidding behavior. Second, advertising scenarios contain numerous high-dimensional, sparse features (such as ad IDs, user IDs, and contextual information). The representation and interaction modeling of these features are complex, making it difficult to accurately learn their impact on bid success rates. Furthermore, most existing methods use point estimates to output fixed win rates, lacking modeling for prediction uncertainty and prone to bias in edge cases.

[0004] Currently, logistic regression (LR) models are commonly used to predict win rates in programmatic advertising bidding. This method uses one-hot encoding of structured features and then estimates the probability of winning by training linear weights. Logistic regression models offer advantages such as simple structure, strong interpretability, and high computational efficiency, making them widely used for online deployment in large-scale advertising systems. However, while logistic regression offers high computational efficiency, it is limited by its inherent structure and can only capture linear relationships between features, making it difficult to express complex nonlinear interactions. In actual bidding scenarios, the influence between features is often nonlinear and dynamic, limiting the predictive power of logistic regression and making it difficult to meet the demand for high-precision predictions.

[0005] Furthermore, with the development of deep learning technology, neural network-based modeling methods (such as Wide&Deep and DeepFM) are increasingly being applied to win rate prediction tasks. These methods effectively represent high-dimensional, sparse features through embedding layers and automatically model nonlinear interactions between features using multi-layer neural networks. They have achieved promising results across multiple advertising platforms.

[0006] While neural network-based approaches improve the model's expressiveness, most only output a fixed point estimate of the win rate, ignoring the randomness and uncertainty inherent in the bidding environment. This leads to unstable performance in the face of abnormal samples, cold starts, or volatile market conditions. Furthermore, their relatively fixed model structure makes it difficult to adapt to dynamic changes in ad traffic, impacting the system's robustness and generalization capabilities.

[0007] In summary, the existing methods for predicting the probability of winning an advertisement under the current bid have technical problems of poor accuracy and robustness.

[0008] To solve the above problems, there is an urgent need for a deep learning model that has dynamic embedding capabilities, can handle complex feature relationships, and can output distributed prediction results, so as to improve the accuracy and robustness of bidding win rate predictions. Summary of the Invention

[0009] In order to solve the technical problem in the above-mentioned prior art that the method for predicting the winning probability of an advertisement under the current bid has poor accuracy and robustness, the present invention provides solutions in the following aspects.

[0010] In a first aspect, the present invention provides a bidding win rate prediction method based on deep learning, comprising: inputting an original price and a data set into a bidding win rate prediction model to obtain a win rate; the bidding win rate prediction model comprises an embedding layer, a feature cross module, a first splicing module, a second splicing module, a μ network, a σ network, a win rate prediction module, and multiple expert networks; Among them, the embedding layer is used to convert each discrete category feature in the dataset into a continuous dense embedding vector; the feature intersection module is used to capture the second-order cross information between each feature of the dense embedding vector, thereby obtaining the second-order feature intersection result; the first splicing module is used to connect the dense embedding vector and the second-order feature intersection result in the feature dimension to form a combined feature vector; the expert network is used to learn the representation of different feature combinations based on the combined feature vector, thereby obtaining a feature combination representation vector; the second splicing module is used to splice the feature combination representation vectors output by each expert network in the feature dimension, thereby obtaining a shared feature tensor; The μ network is used to predict the mean μ of the logarithm of the bid based on the shared feature vector; the σ network is used to predict the standard deviation σ of the logarithm of the bid based on the shared feature vector; the win rate prediction module is used to calculate the Z-score standardized price based on the original price, the mean μ of the logarithm of the bid, and the standard deviation σ of the logarithm of the bid, and calculate the win rate based on the standardized price; the standardized price is positively correlated with the mean μ of the logarithm of the bid, and negatively correlated with the standard deviation σ of the logarithm of the bid; the win rate is positively correlated with the standardized price; the data set includes multiple bidding data, and the characteristic attributes of the bidding data include row identification features, bid prices, and missing slot flags, and the row identification features include advertising IDs or user IDs.

[0011] Its beneficial effects are as follows: the bidding win rate prediction model of the present invention introduces a multi-expert network structure and a normal distribution modeling mechanism from a macro-architecture to simultaneously output the expected value (μ) and the standard deviation (σ), realizing a distributed prediction of the winning rate, thereby effectively characterizing the randomness and volatility in the bidding environment. It innovatively combines the normal distribution modeling mechanism with a dual-channel deep neural network (i.e., the μ network and the σ network) to achieve a joint prediction of the expected value and standard deviation of the programmatic advertising bidding win rate, breaking through the bottleneck of the traditional point estimation model's insufficient characterization of bidding uncertainty. By introducing dynamic embedding technology, the difficult problem of online learning and efficient updating of massive sparse features is solved, significantly improving the adaptability and training efficiency of the bidding win rate prediction model in real-time bidding scenarios.

[0012] Furthermore, the model architecture seamlessly integrates feature interaction, deep representation learning, and uncertainty estimation through an "embedding layer → NFM interaction modeling → multi-expert network → dual-task output" approach. In the output layer, the two parameters μ and σ are modeled to construct a probability density function, enabling continuous modeling of win rates.

[0013] Therefore, the method of the present invention can not only output an expected winning probability value with higher accuracy, but also reflect the uncertainty of the probability value, thereby providing a more robust basis for downstream bidding decisions, especially showing stronger robustness in cold start, data missing or environmental drastic changes scenarios.

[0014] Preferably, the calculation expression of the second-order feature intersection result is: ; Where, Represents the second-order feature cross result, which is a vector. represents the embedding vector of the i-th slot, represents the vector obtained by squaring each element of the embedding vector of the i-th slot, Represents a vector The vector obtained by squaring each element of ; Represents the vector obtained by summing the embedding vectors of each slot.

[0015] Preferably, the expert network includes five fully connected layers, and the calculation expression of the combined feature tensor is: ; Where, is the activation function; 、 、 、 and Represent the weight matrices of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer, respectively. 、 、 、 and Represent the bias vectors of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer respectively; 、 、 、 and Represent the output tensors of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer respectively; Indicates that the number of rows is 512 and the number of columns is The matrix, represents a matrix with 256 rows and 512 columns, represents a matrix with 256 rows and 256 columns. represents a matrix with 128 rows and 256 columns, represents a matrix with 128 rows and 128 columns. represents the combined eigenvector.

[0016] Preferably, the structure of the μ network is a three-layer fully connected network, and the calculation expression of the mean μ of the logarithm of the bid is:

[0017] in, represents the shared feature tensor, 、 and are the weight matrices of the first fully connected network layer of the μ network, the second fully connected network layer of the μ network, and the third fully connected network layer of the μ network, 、 and are the bias vectors of the first fully connected network layer of the μ network, the second fully connected network layer of the μ network, and the third fully connected network layer of the μ network, respectively. is the activation function.

[0018] Preferably, the structure of the σ network is a three-layer fully connected network, and the calculation expression of the standard deviation σ of the logarithm of the bid is: ; Where, represents the standard deviation of the raw value of the logarithm of the bid price, 、 and They are The first layer of the network is a fully connected network, The second layer of the network is fully connected and The weight matrix of the third fully connected layer of the network, 、 and They are The first layer of the network is a fully connected network, The second layer of the network is fully connected and The bias vector of the third fully connected layer of the network, is the activation function, and clip represents the numerical truncation function.

[0019] Preferably, the price calculation expression after Z-score normalization is: ; Where, represents the normalized price, p is the original advertising bid, A positive number to prevent The value of is 0, and log is the natural logarithm function.

[0020] Preferably, the expression for calculating the winning rate based on the standardized price is:

[0021] Where, Indicates the winning rate, represents the error function.

[0022] Preferably, converting each discrete category feature in the data set into a continuous dense embedding vector includes: By looking up the embedding table corresponding to each discrete feature, each discrete feature is mapped into a vector; Reduce the dimension of each slot's vector to obtain the scalar of each slot; The scalars of all slots are concatenated in sequence to obtain a dense embedding vector.

[0023] Preferably, when training the bidding win rate prediction model, a cross entropy loss function is used to calculate the loss, and an Adam optimizer is used to optimize the parameters of the bidding win rate prediction model; the expression of the loss function is: ; in, Expressed as the cross entropy loss of a single sample, Indicates the actual click label (1 means click, 0 means no click), It is expressed as the click probability predicted by the model, that is, the winning rate winr, is the natural logarithm function.

[0024] In a second aspect, the present invention provides a bidding win rate prediction system based on deep learning, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the bidding win rate prediction method based on deep learning of the present invention is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematically illustrates a bidding win rate prediction model structure according to an embodiment of the present invention; Figure 2 It is a schematic diagram showing the structure of a bidding win rate prediction system based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0027] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0028] Example of a bidding win rate prediction method based on deep learning: like Figure 1 As shown, the bidding win rate prediction method based on deep learning of the present invention includes: inputting the original price and data set into the bidding win rate prediction model to obtain the win rate; the bidding win rate prediction model includes an embedding layer, a feature cross module, a first splicing module, a second splicing module, a μ network, a σ network, a win rate prediction module and multiple expert networks; Among them, the embedding layer is used to convert each discrete category feature in the data set into a continuous dense embedding vector; the feature intersection module is used to capture the second-order cross information between each feature of the dense embedding vector, thereby obtaining the second-order feature intersection result; the first splicing module is used to connect the dense embedding vector and the second-order feature intersection result in the feature dimension to form a combined feature vector; the expert network is used to learn the representation of different feature combinations based on the combined feature vector, thereby obtaining a feature combination representation vector; the second splicing module is used to splice the feature combination representation vectors output by each expert network in the feature dimension, thereby obtaining a shared feature tensor.

[0029] The μ network is used to predict the mean μ of the logarithm of the bid based on the shared feature vector; the σ network is used to predict the standard deviation σ of the logarithm of the bid based on the shared feature vector; the win rate prediction module is used to calculate the Z-score standardized price based on the original price, the mean μ of the logarithm of the bid, and the standard deviation σ of the logarithm of the bid, and calculate the win rate based on the standardized price; the standardized price is positively correlated with the mean μ of the logarithm of the bid, and negatively correlated with the standard deviation σ of the logarithm of the bid; the win rate is positively correlated with the standardized price; the data set includes multiple bidding data, and the characteristic attributes of the bidding data include row identification features, bid prices, and missing slot flags, and the row identification features include advertising IDs or user IDs.

[0030] In this embodiment, the number of expert networks may be set to 3, and in other embodiments, the number may also be set to other appropriate numbers.

[0031] In this example, the shape of the dense embedding vector generated by the embedding layer is [batch_size, num_slots * embed_dim]. Here, batch_size refers to the number of samples in a batch, num_slots refers to the number of discrete feature slots in each sample, and embed_dim refers to the vector dimension corresponding to each discrete feature after embedding.

[0032] The dense embedding vectors generated by the embedding layer are continuously updated through backpropagation during model training, helping to capture the potential relationship between each feature. Dynamic embedding is used to embed sparse features, specifically: In training mode:

[0033] In inference mode:

[0034] Among them, Ei represents the embedding vector of the i-th discrete category feature; U(−10,10) means that each vector element is initialized to a random value between [-10,10] using random sampling; Ei=0 means initializing the zero vector.

[0035] In this embodiment, for a batch of samples, it is assumed that the embedding vector obtained by the embedding layer has a shape of [batch_size, total_embed_dim], and the second-order cross information obtained by the feature cross module is a feature cross vector with a shape of [batch_size, embed_dim]. They correspond to the same batch of samples. During the first splicing module, the embedding vector and cross vector of each sample are concatenated along the feature dimension to form a new tensor with a shape of [batch_size, total_embed_dim + embed_dim]. This tensor combines the original features and cross features and serves as the input for subsequent models to help capture more complex feature relationships. total_embed_dim represents the total dimension of the dense embedding vector.

[0036] In this embodiment, the data set is input into the bidding win rate prediction model as follows: first, the data in the data set is stored in the input dictionary, and then the input dictionary is preprocessed to obtain a nested dictionary containing multiple slot IDs, and finally the nested dictionary is input into the embedding layer of the bidding win rate prediction model.

[0037] This input dictionary contains all the features required for model training and prediction. Each feature is in the form of a tensor, which represents the set of values ​​of all attributes of this feature in the dataset.

[0038] The input dictionary is a built-in data structure that stores data as key-value pairs. Dictionaries are useful for quickly accessing data based on an identifier. Common operations include adding key-value pairs, removing items, and looking up the value corresponding to a key.

[0039] In this example, the row identity feature has a shape of [batch_size]. The bid price has a shape of [batch_size] and defaults to a tensor of zeros if not provided. The missing slot flag is a flag for missing features, identifying the missing slots. The Embedding layer extracts embeddings from these input features and performs further computations.

[0040] The win rate distribution parameters (μ, σ) output by this method can be used as state input in the reinforcement learning environment, helping the agent to weigh expected returns and risks in strategy learning, effectively improving its adaptive ability in complex and changing environments.

[0041] The bidding win rate prediction model of this invention introduces a multi-expert network structure and a normal distribution modeling mechanism from a macro-architecture perspective to simultaneously output the expected value (μ) and standard deviation (σ), achieving a distributed prediction of the win rate, thereby effectively characterizing the randomness and volatility in the bidding environment. It innovatively combines the normal distribution modeling mechanism with a dual-channel deep neural network to achieve a joint prediction of the expected value and standard deviation of programmatic advertising bidding win rates, breaking through the bottleneck of traditional point estimation models' inadequate characterization of bidding uncertainty. By introducing the dynamic embedding technology of TensorFlow Recommenders Addons, it solves the difficult problem of online learning and efficient updating of massive sparse features, significantly improving the adaptability and training efficiency of the bidding win rate prediction model in real-time bidding scenarios.

[0042] Furthermore, the model architecture seamlessly integrates feature interaction, deep representation learning, and uncertainty estimation through an "embedding layer → NFM interaction modeling → multi-expert network → dual-task output" approach. In the output layer, the two parameters μ and σ are modeled to construct a probability density function, enabling continuous modeling of win rates.

[0043] Therefore, the method of the present invention can not only output an expected winning probability value with higher accuracy, but also reflect the uncertainty of the probability value, thereby providing a more robust basis for downstream bidding decisions, especially showing stronger robustness in cold start, data missing or environmental drastic changes scenarios.

[0044] The method of the present invention has the ability to model uncertainty. In addition to predicting the winning rate of advertising bidding, it can be expanded to other fields such as insurance pricing, credit risk assessment, and real-time risk control, realizing continuous modeling and confidence output of result distribution, and enhancing the robustness of the decision-making system in data sparse or distribution drift environments.

[0045] In one embodiment, the calculation expression of the second-order feature intersection result is: ; Where, Represents the result of the second-order feature intersection, which is a vector with a shape of [batch_size, embed_dim]. Represents the embedding vector of the i-th slot, with a shape of [batch_size, embed_dim]; represents the vector obtained by squaring each element of the embedding vector of the i-th slot, Represents a vector The vector obtained by squaring each element of ; Represents the vector obtained by summing the embedding vectors of each slot. The embedding vector of the i-th slot refers to the vector composed of all attribute values ​​of the i-th feature attribute in the dataset.

[0046] In one embodiment, the expert network includes five fully connected layers, and the calculation expression of the combined feature tensor is: ; Where, is the activation function; 、 、 、 and Represent the weight matrices of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer, respectively. 、 、 、 and Represent the bias vectors of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer respectively; 、 、 、 and Represent the output tensors of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer respectively; Indicates that the number of rows is 512 and the number of columns is The matrix, represents a matrix with 256 rows and 512 columns, represents a matrix with 256 rows and 256 columns. represents a matrix with 128 rows and 256 columns, represents a matrix with 128 rows and 128 columns. Represents the combined feature vector. Dropout represents the neural network regularization.

[0047] Dropout is only applied to the outputs of the first three layers. It is a common neural network regularization technique used to prevent overfitting. Its core idea is to randomly "drop out" some neurons in the network during training, temporarily setting their outputs to zero. This allows the network to learn using a different subnetwork each time it is trained. This reduces interdependence between neurons and enhances the model's generalization ability.

[0048] In one embodiment, it is assumed that there are three expert networks, and the feature combination representation vectors output by the three expert networks are respectively , then the second splicing module splices the feature combination representation vectors output by each expert network on the feature dimension to obtain the shared feature tensor for:

[0049] That is, the output vectors of the three experts are concatenated in the feature dimension to obtain a tensor with a shape of [batch_size,384].

[0050] In one embodiment, the structure of the μ network is a three-layer fully connected network, and the calculation expression of the mean μ of the logarithm of the bid is:

[0051] in, represents the shared feature tensor, 、 and are the weight matrices of the first fully connected network layer of the μ network, the second fully connected network layer of the μ network, and the third fully connected network layer of the μ network, 、 and are the bias vectors of the first fully connected network layer of the μ network, the second fully connected network layer of the μ network, and the third fully connected network layer of the μ network, respectively. is the activation function. The shape is [64, d], connecting the input z and the 64 neurons of the first layer of the fully connected network; The shape is [32, 64], connecting the 64 neurons of the first layer of the fully connected network and the 32 neurons of the second layer of the fully connected network; The shape is [1, 32], connecting the 32 neurons of the second layer of the fully connected network and the output μ; 、 and The shapes are

[64] ,

[32] and [1] respectively. The shape is [batch_size, 1]. The shape of [32, 64] means The shape is 32 rows and 64 columns. Represents dot product.

[0052] The μ network is used to predict the log median μ, a key parameter required for subsequent win rate calculations. ReLU activation is used to ensure that the output of μ is positive, and clipping is used to ensure that the value of μ does not exceed 100%.

[0053] In one embodiment, the structure of the σ network is a three-layer fully connected network, and the calculation expression of the standard deviation σ of the logarithm of the bid is: ; Where, represents the standard deviation of the raw value of the logarithm of the bid price, 、 and They are The first layer of the network is a fully connected network, The second layer of the network is fully connected and The weight matrix of the third fully connected layer of the network, 、 and They are The first layer of the network is a fully connected network, The second layer of the network is fully connected and The bias vector of the third fully connected layer of the network, is the activation function. clip represents the numerical truncation function. Represents dot product.

[0054] In this embodiment, The shape is [64, d], connecting the input z and 64 neurons; The shape is [32, 64], connecting the 64 neurons in the first layer and the 32 neurons in the second layer; The shape of is [1, 32], connecting the 32 neurons in the second layer and the output σ. 、 and The shapes of are

[64] ,

[32] and [1] respectively. Clip is used to limit the output range of σ to ensure that it does not exceed the given minimum (ε) and maximum (M). The shape is [batch_size, 1].

[0055] The σ network is used to predict the logarithmic standard deviation σ, which is a key parameter required for subsequent win rate calculations. ReLU activation is used to ensure that the output of σ is positive, and clipping is used to ensure that the values ​​of μ and σ do not exceed 100%.

[0056] In one embodiment, the price calculation expression after Z-score normalization is: ; Where, represents the normalized price, p is the original advertising bid, A positive number to prevent The value of is 0, and log is the natural logarithm function.

[0057] In this embodiment, It is a very small positive number used to prevent the occurrence of log(0) and numerical errors; Logarithmic transformation of bids is often used to bring prices into a manageable range. μ can be used to represent the central trend of prices. It can be used to characterize the uncertainty or fluctuation range of price distribution. It can reflect the degree of deviation of the current price from μ.

[0058] In this embodiment, The shape is [batch_size].

[0059] In one embodiment, the expression for calculating the winning rate based on the normalized price is:

[0060] Where, Indicates the winning rate, represents the error function.

[0061] The formula is based on the cumulative distribution function (CDF) of the normal distribution. The resulting winr is a value between 0 and 1, indicating the probability that the ad will win.

[0062] In this embodiment, The shape is [batch_size].

[0063] Error Function (erf) is an important special function in mathematics and statistics, often used in probability theory, statistical distribution and signal processing. Its formula is defined as: .

[0064] In one embodiment, converting each discrete category feature in the dataset into a continuous dense embedding vector includes: S101, mapping each discrete feature into a vector by searching the embedding table corresponding to each discrete feature; In this embodiment, the discrete feature is a user ID or an advertisement ID.

[0065] S102, reducing the dimension of the vector of each slot to obtain a scalar of each slot; A slot may correspond to multiple IDs, such as the most recently clicked ads in a sequence of ad clicks, each with a vector; these vectors form a set. Dimensionality reduction for this slot's vector means compressing or aggregating the corresponding vectors into a single vector.

[0066] There are many ways to reduce dimensionality, such as averaging or linear transformation.

[0067] S103. Concatenate the scalars of all slots in sequence to obtain a dense embedding vector.

[0068] In one embodiment, a cross entropy loss function is used to calculate the loss when training the bidding win rate prediction model, and an Adam optimizer is used to optimize the parameters of the bidding win rate prediction model. The expression of the loss function is: ; in, Expressed as the cross entropy loss of a single sample, Indicates the actual click label (1 means click, 0 means no click), It is expressed as the click probability predicted by the model, that is, the winning rate winr, is the natural logarithm function.

[0069] The Adam optimizer is a commonly used first-order gradient-based adaptive optimization algorithm widely used for training deep neural networks. It combines the principles of momentum and RMSProp to adaptively adjust the learning rate of each parameter, thereby accelerating convergence and improving model performance.

[0070] Example of a bidding win rate prediction system based on deep learning: The present invention also provides a bidding success rate prediction system based on deep learning. Figure 2 As shown, the bidding win rate prediction system based on deep learning includes a processor and a memory, and the memory stores computer program instructions. When the computer program instructions are executed by the processor, a bidding win rate prediction method based on deep learning described in the above embodiment is implemented.

[0071] The deep learning-based bidding win rate prediction system also includes other components familiar to those skilled in the art, such as a communication bus and a communication interface. Their settings and functions are known in the art and will not be described in detail here.

[0072] In the present invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, the computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store the required information and can be accessed by an application, module, or both. Any such computer storage medium can be part of, accessible to, or connectable to a device. Any application or module described in the present invention can be implemented using computer-readable / executable instructions that can be stored or otherwise maintained by such a computer-readable medium.

[0073] In the description of this specification, "multiple" and "several" mean at least two, such as two, three or more, etc., unless otherwise clearly defined.

[0074] While several embodiments of the present invention have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous modifications, variations, and alternatives will occur to those skilled in the art without departing from the concept and spirit of the present invention. It should be understood that various alternatives to the embodiments of the present invention described herein may be employed in practicing the present invention.

Claims

1. A bidding win rate prediction method based on deep learning, characterized in that: include: Inputting the original price and data set into a bidding win rate prediction model to obtain the win rate; the bidding win rate prediction model includes an embedding layer, a feature cross module, a first splicing module, a second splicing module, a μ network, a σ network, a win rate prediction module, and multiple expert networks; Among them, the embedding layer is used to convert each discrete category feature in the dataset into a continuous dense embedding vector; the feature intersection module is used to capture the second-order cross information between each feature of the dense embedding vector, thereby obtaining the second-order feature intersection result; the first splicing module is used to connect the dense embedding vector and the second-order feature intersection result in the feature dimension to form a combined feature vector; the expert network is used to learn the representation of different feature combinations based on the combined feature vector, thereby obtaining a feature combination representation vector; the second splicing module is used to splice the feature combination representation vectors output by each expert network in the feature dimension, thereby obtaining a shared feature tensor; The μ network is used to predict the mean μ of the logarithm of the bid based on the shared feature vector; the σ network is used to predict the standard deviation σ of the logarithm of the bid based on the shared feature vector; the win rate prediction module is used to calculate the Z-score standardized price based on the original price, the mean μ of the logarithm of the bid, and the standard deviation σ of the logarithm of the bid, and calculate the win rate based on the standardized price; the standardized price is positively correlated with the mean μ of the logarithm of the bid, and negatively correlated with the standard deviation σ of the logarithm of the bid; the win rate is positively correlated with the standardized price; the data set includes multiple bidding data, and the characteristic attributes of the bidding data include row identification features, bid prices, and missing slot flags, and the row identification features include advertising IDs or user IDs.

2. The bidding win rate prediction method based on deep learning according to claim 1, characterized in that: The calculation expression of the second-order feature intersection result is: ； Where, Represents the second-order feature cross result, which is a vector. represents the embedding vector of the i-th slot, represents the vector obtained by squaring each element of the embedding vector of the i-th slot, Represents a vector The vector obtained by squaring each element of ; Represents the vector obtained by summing the embedding vectors of each slot.

3. The bidding win rate prediction method based on deep learning according to claim 1, characterized in that: The expert network includes five fully connected layers, and the calculation expression of the combined feature tensor is: ； Where, is the activation function; 、 、 、 and Represent the weight matrices of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer, respectively. 、 、 、 and Represent the bias vectors of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer respectively; 、 、 、 and Represent the output tensors of the first fully connected layer, the second fully connected layer, the third fully connected layer, the fourth fully connected layer, and the fifth fully connected layer respectively; Indicates that the number of rows is 512 and the number of columns is The matrix, represents a matrix with 256 rows and 512 columns, represents a matrix with 256 rows and 256 columns. represents a matrix with 128 rows and 256 columns, represents a matrix with 128 rows and 128 columns. represents the combined eigenvector.

4. The bidding win rate prediction method based on deep learning according to claim 1, characterized in that: The structure of the μ network is a three-layer fully connected network. The calculation expression of the mean μ of the logarithm of the bid is: in, represents the shared feature tensor, 、 and are the weight matrices of the first fully connected network layer of the μ network, the second fully connected network layer of the μ network, and the third fully connected network layer of the μ network, 、 and are the bias vectors of the first fully connected network layer of the μ network, the second fully connected network layer of the μ network, and the third fully connected network layer of the μ network, respectively. is the activation function.

5. The bidding win rate prediction method based on deep learning according to claim 1, characterized in that: The structure of the σ network is a three-layer fully connected network. The calculation expression of the standard deviation σ of the logarithm of the bid is: ； Where, represents the standard deviation of the raw value of the logarithm of the bid price, 、 and They are The first layer of the network is a fully connected network, The second layer of the network is fully connected and The weight matrix of the third fully connected layer of the network, 、 and They are The first layer of the network is a fully connected network, The second layer of the network is fully connected and The bias vector of the third fully connected layer of the network, is the activation function, and clip represents the numerical truncation function.

6. The bidding win rate prediction method based on deep learning according to claim 1, characterized in that: The price calculation expression after Z-score standardization is: ； Where, represents the normalized price, p is the original advertising bid, A positive number to prevent The value of is 0, and log is the natural logarithm function.

7. The bidding win rate prediction method based on deep learning according to claim 6, characterized in that: The expression for calculating the winning rate based on the standardized price is: Where, Indicates the winning rate, represents the error function.

8. The bidding win rate prediction method based on deep learning according to claim 1, characterized in that: Converting each discrete category feature in the dataset into a continuous dense embedding vector includes: By looking up the embedding table corresponding to each discrete feature, each discrete feature is mapped into a vector; Reduce the dimension of each slot's vector to obtain the scalar of each slot; The scalars of all slots are concatenated in sequence to obtain a dense embedding vector.

9. The method for predicting bidding success rate based on deep learning according to any one of claims 1 to 8, wherein: When training the bidding win rate prediction model, the cross entropy loss function is used to calculate the loss, and the Adam optimizer is used to optimize the parameters of the bidding win rate prediction model. The expression of the loss function is: ； in, Expressed as the cross entropy loss of a single sample, Indicates the actual click label (1 means click, 0 means no click), It is expressed as the click probability predicted by the model, that is, the winning rate winr, is the natural logarithm function.

10. A bidding win rate prediction system based on deep learning, comprising a processor and a memory, wherein the memory stores computer program instructions, characterized in that: When the computer program instructions are executed by the processor, the bidding win rate prediction method based on deep learning as described in any one of claims 1 to 9 is implemented.

Images