A ctr recommendation method based on feature interaction and score integration

Abstract

The application discloses a CTR recommendation method based on feature interaction and score integration. The steps are as follows: 1. All features are divided into four categories: Q_U_I features, user behavior features, domain-related features and domain ID features; 2. The scene interaction refinement module is used to interact the Q_U_I features and the domain-related features to obtain combined features. The combined features are spliced with the user historical behavior features, the Q_U_I features and the domain-related features to finally obtain the input features of the backbone; 3. The expert projection module is used to extract differentiated public features; 4. The gating mechanism is used to weight and add the public features and the specific features to obtain the advanced fusion features of each view. Then the advanced fusion features are input into the corresponding MLP to obtain the scores of each view, and then the gating mechanism is used to integrate the scores, and the final prediction result is obtained through the activation function sigmoid; 5. The loss function is used for optimization. The application can capture the inherent correlation information in the feature space and the label space, and improve the correctness of multi-domain CTR prediction.

Description

Technical Field

[0001] The present invention relates to CTR recommendation, and in particular to a multi-domain CTR recommendation method based on feature interaction and score integration. Background Art

[0002] Click-throughrate (CTR) prediction is one of the most common tasks in the recommendation field. However, manually predicting user preferences is extremely difficult. Therefore, using deep learning techniques for CTR prediction can benefit e-commerce platforms in presenting more personalized recommendations to users. Networks such as MMOE, CGC, and PLE have been specifically proposed for various recommendation tasks. These methods have achieved impressive performance on many large datasets, demonstrating the effectiveness of deep learning techniques in the recommendation field. Fundamentally, CTR prediction can be viewed as taking user and item characteristics as input, and the network outputs the probability of the user clicking on the item, which implicitly reflects the user's preference for the item. However, translating this to multi-domain recommendation is not easy. The diverse user behavior in cross-border e-commerce scenarios, coupled with the sparsity and uneven distribution of data, makes multi-domain CTR prediction a challenging task.

[0003] Traditional approaches involve jointly building a single model across all domains, or independently building a model for each domain. However, directly applying traditional methods to multi-domain recommendations presents several challenges: 1) Due to geographical and cultural differences, users in different domains may have significant differences in their shopping preferences. Joint training can eliminate the unique information of each domain. 2) The amount of available data across domains is extremely unbalanced, with many domains containing only a small amount of data. Independent training ignores the potential correlations between domains and cannot effectively train domains with limited data.

[0004] This paper proposes a new model structure that can achieve more accurate multi-domain CTR prediction. Summary of the Invention

[0005] This paper provides a CTR recommendation model based on feature interaction and score integration. First, through feature interaction, the combined features between local and global scenes are learned, which explicitly enhances the feature representation in multi-domain learning. The features are then concatenated and input into a common expert and a specific expert network to generate common and specific feature representations. A gating mechanism is used to integrate the common and specific features to obtain high-level fused features for each view. Finally, these are input into the corresponding MLP to obtain the scores for each view, and then a gating mechanism is used to integrate the scores for each view to obtain the final CTR prediction score. The entire model includes a scene interaction refinement module, an expert projection module, and a multi-view score integration module.

[0006] A CTR recommendation method based on feature interaction and score integration, the steps are as follows:

[0007] Step (1) divides the features in the dataset into four types: historical behavior features, Q_U_I features, domain-related features, and domain ID features.

[0008] Step (2) distinguishes the local features with different distributions in different fields through the designed scene interaction refinement module.

[0009] Step (3) extracts differentiated common features through the designed expert projection module, and introduces adaptive residual connections to accelerate learning and achieve more flexible projection.

[0010] Step (4) uses the designed multi-view score integration module to first extract high-level fusion features, and then outputs the results through score integration.

[0011] Step (5) improves the loss function to effectively avoid overfitting.

[0012] Furthermore, the four feature types described in step (1) are specifically divided as follows:

[0013] 1. Historical behavior features: This represents a collection of user historical behaviors, including sequences and sequence-related features. The specific formula is as follows, where |G| represents the number of historical behavior features:

[0014] G={g1, g2, ..., g |G|} (Formula 1)

[0015] 2. Q_U_I features: This represents a collection of query, user, and product features. These features are distributed inconsistently across domains, resulting in domain differences. The specific formula is as follows, where |L| represents the number of Q_U_I features:

[0016] L = {l1, l2, ..., l |L|} (Formula 2)

[0017] 3. Domain-related features: This represents a collection of features related to the domain, including product exposure features, product click features, and domain-high correlation features. The specific formula is as follows, where |S| represents the number of domain-related features:

[0018] S={s1,s2,...,s |S|} (Formula 3)

[0019] 4. Domain ID feature: This represents a collection of domain ID information. Each feature is an embedding feature generated by a one-hot vector. The specific formula is as follows, where |R| represents the number of domains:

[0020] R={r1,r2,...,r |R|} (Formula 4)

[0021] The representation of a dataset consisting of |R| domains is as follows:

[0022] D={D1,D2,...,D |R|} (Formula 5)

[0023]

[0024] Among them, D i represents the sample data of the i-th field, |D i | represents the number of samples in the i-th field.

[0025] represents the Dth |Dth in the i-th field i | features of the samples, represents the Dth |Dth in the i-th field i |CTR labels of samples.

[0026] Therefore, the set of 4 types of features of a sample is expressed as follows:

[0027]

[0028] in, represents the sample features of the jth sample in the i-th field, They represent the historical behavior characteristics, Q_U_I characteristics, and domain-related characteristics of the jth sample in the i-th field, r i Represents the embedding vector of the i-th domain id.

[0029] The scene interaction refinement module described in step (2) is as follows:

[0030] The Q_U_I features of different domains are independent, so their distributions are not exactly the same in different domains. These differences reflect the preferences of users in different domains. We explicitly construct a <Q_U_I feature, domain-dependent feature> combination feature to force the fusion of domain-dependent features and Q_U_I features to represent the feature interaction between local and global scenes.

[0031] Because the Q_U_I feature dimension is generally large, its dimension is first reduced. Then, the feature interaction operation between the Q_U_I feature and the domain-related feature is performed to obtain the combined feature. The specific formula is as follows:

[0032]

[0033]

[0034] Among them, σ represents the nonlinear activation function, ω l 、b l Represent the weights and biases of the fully connected layer, T represents the transpose of the matrix, represents the outer product operation, L r Represents the Q_U_I features after dimensionality reduction.

[0035] Among them, s represents the nonlinear activation function, ω l 、b l Represent the weights and biases of the fully connected layer, T represents the transpose of the matrix, represents the outer product operation, L r Represents the Q_U_I features after dimensionality reduction.

[0036] Next, this combined feature M is expanded into a one-dimensional vector, and the global differential feature C is extracted using a fully connected layer. The specific formula is as follows:

[0037]

[0038] Among them, ω c 、b c They represent the weights and biases of the fully connected layer of the combined feature M, respectively. Flatten() represents a function that expands multiple dimensions into a one-dimensional vector, such as reshape.

[0039] Finally, the concat operation is used to combine the G, L, S, and C features as the input of the backbone network. The specific formula is as follows:

[0040] I = concat(G|L|S|C) (Formula 11)

[0041] Where I∈R m I is the input of the backbone network, R m is the dimension of the input, that is, the sum of the G, L, S, and C feature dimensions.

[0042] The expert projection module described in step (3) is specifically as follows:

[0043] In multi-domain learning, multi-expert networks cannot model the difference in common features as in theory, because in the multi-domain scenario, the data is extremely imbalanced, and the expert gradient information for modeling the domain with large amount of data is larger, which will learn a better representation than other experts. This limits the expression ability of most experts. Therefore, an expert projection module is proposed to reduce the difficulty of expert learning and improve the overall expression ability of the expert.

[0044] The method is to let other experts learn the projection of the parameters of the leading expert instead of directly learning the parameters of the expert, and explicitly model the mapping offset caused by the difference, as follows:

[0045]

[0046] wherein W base ∈R m×n represents the basic weight provided by the leading expert, and represent the weight and bias of the i-th projection, respectively. represents the weight of the i-th projection expert obtained by projection. The leading expert refers to the expert who provides the basic weight and parameter projection. The leading expert refers to the expert who models the domain with large amount of data.

[0047] Because this way will add one more term than without projection, resulting in a decrease in learning speed and gradient, in order to solve this problem, a residual connection is introduced, which also helps to control the initialization parameters. The specific formula is as follows:

[0048]

[0049] wherein α i represents the i-th adaptive parameter value, and has no boundary limit. ⊙ represents the element-wise product operation.

[0050] Each projection expert interacts with the input feature I similar to the real expert, where I ∈ R m , and then adds an additional bias to obtain the differentiated common feature The specific formula is as follows:

[0051]

[0052] The multi-view score integration module of step (4) is as follows:

[0053] After extracting the differentiated common feature, the domain-specific feature Each view has a specific expert that is dedicated to extracting discriminative features. Each view has a gating network with the following formula:

[0054]

[0055]

[0056]

[0057] Among them, r represents the domain ID feature corresponding to the sample, represents the specific features extracted from the j-th view, MLP represents the multi-layer perceptron, N comm Indicates the number of projection experts in the expert projection module. and denote the weights and biases of the gating network of the jth view, respectively, where k∈[0, N comm ]. Indicates the gate value of the j-th view for the k-th feature. When k=0, it indicates a specific feature, k∈[1, N comm ], it indicates a common feature. j represents the high-level fusion features of the j-th view, j∈[1, N view ].

[0058] Finally, the high-level fusion features of each view are passed through the corresponding MLP to obtain multiple scores, and then these scores are integrated through the gating mechanism, and finally the final CTR prediction result is obtained through the sigmoid function. The specific formula is as follows:

[0059]

[0060]

[0061]

[0062] Among them, R j ∈R 1 Indicates the score, and are the weights and biases of the gating network, j∈[1, N view ],N view Indicates the number of views (number of domains), that is, N view =|R|.

[0063] The loss function of step (5) is specifically designed as follows:

[0064] By adjusting the loss function, our model can be trained end-to-end. The specific formula is as follows:

[0065]

[0066] The first term represents the loss calculation formula, which can be cross-entropy, mean square error, etc. The second term is a regularization term, λ j is a hyperparameter of the j-th view, which is used to control the ability of regularization.

[0067] Beneficial effects of the present invention:

[0068] The present invention predicts multi-domain CTR tasks based on feature interaction and score integration. In order to clearly enhance the feature representation in multi-domain learning, the scene interaction refinement module is used to obtain the combined features of Q_U_I features and domain-related features, so that it represents the feature interaction between the local scene and the global scene of the embedded features. In order to alleviate the significant diversity of features in different scenarios, the uneven distribution and sparsity of data, the expert projection module is used to generate differentiated common feature representations. Finally, in order to enhance the fitting ability of the sparsity of the label data distribution, the multi-view score integration module is used to adaptively combine the output scores of the multi-branch network. This network can capture the inherent correlation information in the feature space and the label space, and improve the accuracy of multi-domain CTR prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a specific flow diagram of the method of the present invention.

[0070] Figure 2 It is a schematic diagram of the entire model framework in the method of the present invention. DETAILED DESCRIPTION

[0071] The present invention will be further described in detail below with reference to the accompanying drawings:

[0072] like Figure 1 As shown in Figure 2, the model based on feature interaction and score integration includes the following steps:

[0073] Attachment Figure 2 The overall model framework diagram of the present invention is shown. Figure 2 To explain the details of each module separately, the specific implementation steps are as follows:

[0074] Step (1) divides the features into four parts, mainly historical behavior features, Q_U_I features, domain-related features, and domain ID features, represented by G, L, S, and r respectively. Each sample set can be expressed as: Among them, r is represented by the embedding feature of the one-hot vector.

[0075] Step (2) designed a scene interaction module, first using a layer of full connection to reduce the dimension of the high-dimensional Q_U_I feature, using the reduced dimension feature and the domain-related feature to perform an outer product operation to obtain a combined feature, then expanding the combined feature into a one-dimensional vector, and extracting the global differential feature C through a layer of full connection. The G, L, S, C features are concat-operated to obtain I∈R m As the input of the backbone later.

[0076] Step (3) designed the expert projection module, set each field as a specific expert, and three projection experts. First create a basic weight W base ∈R m×n , creating a weight and bias for each projection: and The weight of the projection expert is obtained by multiplying the projection weight and the basic weight matrix and adding the projection bias. Then add a residual connection α i ⊙W base Improve learning speed and help control initialization parameters. Finally, multiply it with the input feature I and add the bias of the projection expert. Get differentiated common features

[0077] Step (4) Design a multi-view score integration module, design a view for each domain, and after obtaining the common features, it is necessary to extract the specific domain features. Each view uses an MLP, which takes the concatenated features I as input and outputs the specific features of the corresponding domain. Then, a gating mechanism is used to perform weighted summation on the three common features and each specific feature to obtain the high-level fusion feature H of the j-th view. j Finally, the high-level fusion features of each view are passed through their respective MLPs to obtain multiple scores for the corresponding view, which are then weighted and summed through a gating mechanism, and the sigmoid function is used to obtain the final multi-domain CTR prediction result.

[0078] Step (5) improves the loss function. The first term uses the traditional cross-entropy or meansquare error, where the true value is the label and the predicted value is the final CTR prediction result. A regularization term is added after the first term. This term is used to apply a cross-entropy loss to the score of each view, where the true value is the label value and the predicted value is the score of each view. Then a hyperparameter is multiplied to control the regularization ability to prevent the model from overfitting.

[0079] As shown in Table 1, a public data set AliCCP is selected, which is from the real traffic log of Taobao recommendation system. The data set contains click label and purchase label, and only click label is used in the present application for multi-domain CTR prediction task. Different business scenarios are regarded as domains, and the data set contains three domains, and the samples in the three domains are extremely imbalanced. The statistical information of the data set is shown in Table 1, and the data volume of domain1, domain2 and domain3 is extremely different, and the click rate is sparse.

[0080] Table 1

[0081]

[0082] The comparative model used in the experiment includes existing multi-task model and multi-domain model, and all the models take predicting CTR of multiple domains as the task. AUC in Table 2 is a common index in recommendation system, and AUC is defined as the area under the ROC curve. Its meaning is: randomly selecting a pair of positive and negative samples, AUC is the probability that the predicted positive sample is True is greater than the probability that the predicted negative sample is True. The higher the AUC, the better the performance of the model.

[0083] Table 2

[0084]

[0085] As can be seen from Table 2, the method of the present application based on the multi-domain CTR prediction model of feature interaction and score integration has achieved obvious effect in the multi-domain CTR recommendation system.

Claims

1. A CTR recommendation method based on feature interaction and score integration, characterized by It includes the following steps: Step (1) divides the features in the dataset into four types: historical behavior features, Q_U_I features, domain-related features, and domain ID features, which are represented by G, L, S, and r respectively. Each sample set is represented as: Q_U_I features represent the set of query, user, and product features, including query features, user features, and product features; Step (2): Use the designed scene interaction refinement module to distinguish local features with different distributions in different fields. First, use a fully connected layer to reduce the dimension of the Q_U_I feature with a higher dimension. Then, perform an outer product operation on the reduced feature and the domain-related feature to obtain a combined feature. Next, expand the combined feature into a one-dimensional vector, and extract the global differentiation feature C through a fully connected layer. Perform concat operation on G, L, S, and C features to obtain I∈R m As the input of the backbone later; Step (3) designed the expert projection module, set each field as a specific expert, three projection experts; first create a basic weight W base ∈R m×n , creating a weight and bias for each projection: and The weight of the projection expert is obtained by multiplying the projection weight and the basic weight matrix and adding the projection bias. Then add a residual connection α i ⊙W base Improve learning speed and help control initialization parameters; finally multiply with input feature I and add the bias of projection expert Get differentiated common features Step (4) Design a multi-view score integration module, design a view for each domain, and after obtaining the common features, it is necessary to extract the specific domain features. Each view uses an MLP, which takes the concatenated features I as input and outputs specific features of the corresponding field; Then, a gating mechanism is used to perform weighted summation on the three common features and each specific feature to obtain the high-level fusion feature H of the j-th view. j Finally, the high-level fusion features of each view are passed through their respective MLPs to obtain multiple scores for the corresponding view, which are then weighted and summed using a sigmoid function through a gating mechanism to obtain the final multi-domain CTR prediction result. Step (5): Improve the loss function to effectively avoid overfitting.

2. A CTR recommendation method based on feature interaction and score integration according to claim 1, characterized in that The specific classification of the 4 types of features described in Step (1) is as follows: Historical behavior feature: Represents the set of user historical behaviors, including sequences and sequence-related features. The specific formula is as follows, where |G| represents the number of historical behavior features: G={g 1, g2,…,g |G| } (Formula 1) Q_U_I feature: Represents the set of query, user, and commodity features, including query features, user features, and commodity features. The specific formula is as follows, where |L| represents the number of Q_U_I features: L={l 1, l 2, …,l |L| } (Formula 2) Domain-related feature: Represents the set of features related to the domain, including commodity exposure features, commodity click features, and features highly related to the domain. The specific formula is as follows, where |S| represents the number of domain-related features: S={s 1, s 2, …,s |S| } (Formula 3) Domain ID feature: Represents the set of ID information of the domain. Each feature is an embedding feature generated by a one-hot vector. The specific formula is as follows, where |R| represents the number of domains: R={r 1, r 2, …,r |R| } (Formula 4) The representation of the dataset composed of |R| domains is as follows: D={D1,D2,…,D |R| } (Formula 5) Among them, D i represents the sample data of the i-th field, |D i | represents the number of samples in the i-th field; represents the Dth |Dth field in the i-th field i | features of the samples, represents the Dth |Dth field in the i-th field i |CTR labels of samples; Therefore, the set representation of the 4 types of features of a sample is as follows: in, represents the sample features of the jth sample in the i-th field, They represent the historical behavior characteristics, Q_U_I characteristics, and domain-related characteristics of the jth sample in the i-th field, r i Represents the embedding vector of the i-th domain id.

3. A CTR recommendation method based on feature interaction and score integration according to claim 2, characterized in that The scene interaction refinement module described in Step (2) is specifically as follows: 2-1. Explicitly construct a <Q_U_I feature, domain-related feature> combined feature to force the fusion of domain-related features and Q_U_I features, which is used to represent the feature interaction between the local scene and the global scene. First, reduce its dimension, and then perform the feature interaction operation on the Q_U_I feature and the domain-related feature to obtain the combined feature. The specific formula is as follows: Among them, σ represents the nonlinear activation function, ω l 、b l Represent the weights and biases of the fully connected layer, T represents the transpose of the matrix, represents the outer product operation, L r Represents the Q_U_I feature after dimensionality reduction; 2-2. Expand this combined feature M into a one-dimensional vector, and use a fully connected layer to extract the global differentiation feature C. The specific formula is as follows: Among them, ω c 、b c They represent the weights and biases of the fully connected layer of the combined feature M, and flatten() represents the function of expanding multiple dimensions into a one-dimensional vector; 2-3. Use the concat operation to combine the G, L, S, and C features as the input of the backbone network. The specific formula is as follows: I = concat(G|L|S|C) (Formula 11) Where I∈R m , I is the input of the backbone network, R m is the dimension of the input, that is, the sum of the G, L, S, and C feature dimensions.

4. A CTR recommendation method based on feature interaction and score integration according to claim 3, characterized in that The expert projection module described in Step (3) is specifically as follows: 3-1. Provide well-trained parameters by the leading expert, and let other experts learn the projection of the leading expert's parameters instead of directly learning the expert's parameters, which clarifies the mapping offset caused by the difference. The specific formula is as follows: Among them, W base ∈R m×n represents the basic weight provided by the leading expert, and Represent the weight and bias of the i-th projection respectively; represents the weight of the i-th projection expert obtained through projection; 3-2. Introduce a residual connection to control the initialization parameters. The specific formula is as follows: Among them, α i represents the i-th adaptive parameter value without any boundary restrictions; ⊙ represents the element-wise product operation; 3-3. Each projection expert interacts with the input feature I like a real expert, where I∈R m , and then add an additional bias Get differentiated common features The specific formula is as follows:

5. A CTR recommendation method based on feature interaction and score integration according to claim 4, characterized in that The multi-view score integration module described in Step (4) is specifically as follows: 4-1. Extracting differentiated common features Finally, domain-specific features need to be extracted Each view has a specific expert dedicated to extracting discriminative features; each view has a gating network with the following formula: Among them, r represents the domain ID feature corresponding to the sample, represents the specific features extracted from the j-th view, MLP represents the multi-layer perceptron, N comm represents the number of projection experts in the expert projection module; and denote the weight and bias of the gating network of the jth view, respectively, where k∈[0,N comm ]; Indicates the gate value of the j-th view for the k-th feature. When k=0, it indicates a specific feature, k∈[1,N comm When H j represents the high-level fusion features of the j-th view, j∈[1,N view ]; 4-2. The high-level fusion features of each view are passed through the corresponding MLP to obtain multiple scores, and then these scores are integrated through the gating mechanism, and finally the final CTR prediction result is obtained through the sigmoid function. The specific formula is as follows: Among them, R j ∈R 1 Indicates the score, and are the weights and biases of the gating network, j∈[1,N view ],N view Indicates the number of views, that is, N view =|R|.

6. A CTR recommendation method based on feature interaction and score integration according to claim 5, characterized in that The specific design of the loss function in Step (5) is as follows: Adjust the loss function to enable the model to be trained end-to-end. The specific formula is as follows: Among them, the first term represents the calculation formula of the loss, the second term is a regularization term, λ j is a hyperparameter of the j-th view, which is used to control the ability of regularization.

Images