Personalized recommendation exposure deviation removing method based on multi-level exposure fusion

Abstract

The invention belongs to the field of recommendation, and relates to a personalized recommendation exposure deviation removing method based on multi-level exposure fusion, which comprises the following steps: acquiring users and commodities of an e-commerce platform, and constructing commodity data of multiple exposure levels according to commodity data of the e-commerce platform; inputting the commodity data of the multiple exposure levels into the trained generative adversarial network to obtain generation samples of the multiple exposure levels; the generated samples of the multiple exposure levels are input into the trained interest enhancement and multi-exposure fusion module, and fusion features of the multiple exposure levels and final fusion features are obtained; inputting the user and commodity data of the e-commerce platform and the final fusion feature into a matching degree calculation module to obtain a matching degree score of the user and the commodity; the interest enhancement and multi-exposure fusion module is combined with a progressive enhancement fusion mechanism to suppress artifacts and low-frequency signals occurring in the fusion process, the fusion weight is optimized, smooth transition of features among exposure levels is ensured, and a more accurate recommendation result is generated.

Description

Technical Field

[0001] The present invention belongs to the field of e-commerce big data recommendation, and particularly relates to a personalized recommendation de-exposure deviation method based on multi-level exposure fusion. Background Art

[0002] With the rapid development of Internet technology, recommendation systems have become a key bridge connecting users with a vast amount of information and are widely used in multiple fields such as e-commerce, social media, and video streaming. However, user behavior data is essentially observational rather than experimental data, which leads to various biases in the data (such as popularity bias, position bias, selection bias, and exposure bias), especially exposure bias. This makes traditional recommendation systems often limited by this, which restricts the range of recommended content that users can access, makes it difficult for the recommendation model to accurately understand and reflect the comprehensive preferences of users, and thus weakens the accuracy and practicality of recommendations. Therefore, effectively alleviating exposure bias and improving the comprehensive performance of recommendation systems have become the focus of common concern and an urgent problem to be solved in the academic and industrial circles at home and abroad.

[0003] Currently, researchers at home and abroad have carried out a large amount of research work on eliminating exposure bias in the field of recommendation systems. The existing research results mainly focus on the following two aspects: debiasing methods based on causal inference and data adjustment, and debiasing methods based on deep learning and generative modeling. Debiasing methods based on causal inference and data adjustment: These methods mainly use technical means such as inverse propensity score weighting to adjust the weights of samples to compensate for the bias caused by uneven exposure, so as to introduce a more fair data distribution in the model training process. In addition, it also includes data augmentation techniques, such as generating synthetic data or expanding existing datasets, to reduce exposure bias and improve the comprehensiveness and fairness of recommendations. Debiasing methods based on deep learning and generative modeling: These methods use deep learning models to capture the implicit patterns and relationships in user behavior data, so as to more precisely handle the bias problems in the data. Through the powerful representation ability of deep learning and generative modeling techniques, these methods can more accurately identify user preferences, reduce recommendation errors caused by exposure bias, and improve the overall performance of the system.

[0004] However, there are some problems with the above recommendation methods:

[0005] 1. In a recommendation system, there are significant differences in the exposure levels of different items. Generating a multi-level exposure recommendation sequence requires dealing with the problems of data sparsity and diversity.

[0006] 2. It is impossible to accurately capture the subtle changes in user interests at each exposure level and achieve multi-faceted collaboration.

[0007] 3. In multi-level exposure fusion, artifact effects may occur due to overemphasis on certain exposure levels, or low-frequency interesting signals may be masked by the prominent performance of high-frequency signals. Summary of the Invention

[0008] To solve the above-mentioned problems in the prior art, the present invention adopts a personalized recommendation method for removing exposure bias based on multi-level exposure fusion, which is characterized by including: obtaining user and commodity data of an e-commerce platform, constructing commodity data with multiple exposure levels according to the commodity data of the e-commerce platform, inputting the user and commodity data of the e-commerce platform and the commodity data with multiple exposure levels into a trained recommendation model to obtain the matching degree score between the user and the commodity, and obtaining a recommendation list according to the matching degree score between the user and the commodity; the recommendation model includes: a generative adversarial network, an interest enhancement and multi-exposure fusion module, and a matching degree calculation module.

[0009] The training process of the recommendation model includes:

[0010] S1. Obtain user and commodity data of an e-commerce platform, and construct commodity data with multiple exposure levels according to the commodity data of the e-commerce platform;

[0011] S2. Train the generative adversarial network according to the commodity data with multiple exposure levels to obtain a trained generative adversarial network;

[0012] S3. Input the commodity data with multiple exposure levels into the trained generative adversarial network to obtain generated samples with multiple exposure levels;

[0013] S4. Train the interest enhancement and multi-exposure fusion module according to the generated samples with multiple exposure levels to obtain a trained interest enhancement and multi-exposure fusion module;

[0014] S5. Input the generated samples with multiple exposure levels into the trained interest enhancement and multi-exposure fusion module to obtain fusion features with multiple exposure levels and final fusion features;

[0015] S6. Train the matching degree calculation module according to the user and commodity data of the e-commerce platform and the final fusion features to obtain a trained matching degree calculation module.

[0016] Beneficial Effects:

[0017] 1. The present invention uses a generative adversarial network to generate samples that meet specific exposure conditions, and ensures the consistency between the generated samples and the distribution of real samples, thereby ensuring the precise regulation and consistency of the sample distribution at different exposure levels, and solving the problems of data sparsity and diversity. 2. The interest enhancement and multi-exposure fusion module of the present invention combines a progressive reinforcement fusion mechanism to suppress artifacts and low-frequency signals that may appear during the fusion process, optimize the fusion weight distribution, and ensure the smooth transition of features between different exposure levels, thereby generating more accurate recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 FIG. is a framework diagram of a personalized recommendation de-exposure deviation method based on multi-level exposure fusion provided by an embodiment of the present invention;

[0019] Figure 2 FIG. is a diagram of a multi-exposure sample generation model provided by an embodiment of the present invention;

[0020] Figure 3 FIG. is a diagram of a multi-exposure sample fusion and hierarchical progressive optimization merging model provided by an embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] As Figure 1 described, the present invention adopts a personalized recommendation de-exposure deviation method based on multi-level exposure fusion, which specifically includes the following steps: obtaining user and commodity data of an e-commerce platform, constructing commodity data of multiple exposure levels according to the commodity data of the e-commerce platform, inputting the user and commodity data of the e-commerce platform and the commodity data of multiple exposure levels into a trained recommendation model to obtain a matching score between the user and the commodity, and generating a personalized recommendation list according to the matching score between the user and the commodity; the recommendation model includes: a generative adversarial network, an interest enhancement and multi-exposure fusion module, and a matching degree calculation module.

[0023] The training process of the recommendation model includes:

[0024] S1. Obtain user and commodity data of an e-commerce platform, and construct commodity data of multiple exposure levels according to the commodity data of the e-commerce platform The commodity data of each exposure level includes: the exposure frequency Frq of the commodity i , the click-through rate CTR of the commodity iand the exposure residence time Dur of the product i ;

[0025] The user data of the e-commerce platform includes: discrete data such as user ID and gender, and user historical behavior data (such as clicks and purchases); the product data of the e-commerce platform includes: discrete attributes such as product categories and brands, and continuous data such as the number of product exposures, click-through rate, and exposure residence time.

[0026] Constructing product data with multiple exposure levels based on the product data of the e-commerce platform includes:

[0027] Dividing products into multiple exposure levels to obtain a set of products at multiple exposure levels; specifically, the division of exposure levels is based on indicators such as the number of product exposures, click-through rate, and exposure residence time, and the division of each level can be based on the following rules: low exposure level: products with fewer exposures, lower click-through rate, and shorter residence time; medium exposure level: products with medium exposures, medium click-through rate, and medium residence time; high exposure level: products with more exposures, higher click-through rate, and longer residence time.

[0028] Calculating product data at multiple exposure levels based on the set of products at multiple exposure levels

[0029] Specifically, the number of product exposures Frq i is:

[0030]

[0031] where Frq i represents the number of exposures of the product data at the i-th exposure level, Ex j represents the number of exposures of product j, L i represents the set of products at the i-th exposure level, |L i | is the number of products at the i-th exposure level; Dur i represents the exposure residence time of the product data at the i-th exposure level, CTR i represents the click-through rate of the product data at the i-th exposure level, and the calculation formulas of Dur i and CTR i are similar to that of Frq i .

[0032] S2. Training the generative adversarial network according to the product data at multiple exposure levels to obtain a trained generative adversarial network;

[0033] As Figure 2 shown, the generative adversarial network includes a feature encoding module, a generator G, and a discriminator D; the training process of the generative adversarial network includes:

[0034] S21. Input the product data of multiple exposure levels into the feature encoding module to obtain the exposure encoding vector E, which can be specifically expressed as:

[0035] E = {e1, e2, …, e n},

[0036] e i = (Norm(Frqi), Norm(Dur i ), Norm(CTR i ))

[0037] where e i represents the conditional information of the product data of the i-th exposure level, Norm represents the normalization process to ensure a similar numerical range, and n is the number of exposure levels.

[0038] S22. Add random noise Z to the exposure encoding vector E to obtain the combined input feature vector Z input ;

[0039]

[0040] where is the combined input feature of the i-th exposure level;

[0041] S23. Input the combined input feature vector Z input into the generator G to obtain the generated sample set X gen , which can be expressed as:

[0042] X gen = G(Z input )

[0043]

[0044] where is the generated sample of the i-th exposure level.

[0045] Furthermore, the generator G is a structure composed of multiple neural network layers, which calculates layer by layer by receiving the combined input features to generate samples that meet specific exposure conditions.

[0046] The generator includes: an input layer, a hidden layer, and an output layer; the specific internal process of the generator can be expressed as follows:

[0047] At the input layer, for the combined input feature vector Z inputPerform non - linear mapping, usually through fully - connected layers and activation functions for transformation to extract initial features; in the hidden layer, it is usually composed of multiple convolutional layers or fully - connected layers to further process the initial features and extract features; in the output layer, the final set of generated samples is output, and the output sample set is X gen Meet the current exposure condition E.

[0048] S24. Perform interpolation sampling between the generated samples at multiple exposure levels and the commodity data at multiple exposure levels to obtain an interpolation sample set

[0049]

[0050] Among them, is the interpolation sample of the i - th exposure level.

[0051] S25. Input the generated samples, commodity data, and interpolation samples at multiple exposure levels into the discriminator D respectively to obtain corresponding discrimination results;

[0052] The specific internal steps of the discriminator are as follows:

[0053] The discriminator receives the generated sample the interpolation sample and the real commodity data at multiple exposure levels as inputs to distinguish the differences between the generated sample the interpolation sample and the real commodity data at different exposure levels The discriminator extracts features from the input samples through multiple convolutional layers, calculates the feature representations layer by layer, and finally converts them into probability values p ∈ [0, 1]. The conversion basis is:

[0054] p = σ(W out ·h L + b out )

[0055] Among them, W out and b out are the weights and biases of the output layer, h L represents the final feature representation of the output layer, and σ is the sigmoid activation function that compresses the output value into the range of [0, 1]. To sum up, the discrimination result D(X) of the discriminator D for the input sample can be summarized as:

[0056] D(X) = σ(W out ·f(W L-1 ·...f(W1·X + b1)…+ b L-1 )+ b out )

[0057] Among them, X is the input sample, and W l and b l are the weights and biases of the l-th layer, L represents the number of layers, and f() is a linear transformation;

[0058] The discriminator maps the input sample to the final probability value D(X) through a series of feature extraction layers, thereby indicating the possibility of whether it is a real sample.

[0059] S26. Calculate the total loss function value according to the discrimination results of the generated samples, commodity data, and interpolation samples at multiple exposure levels, and update the parameters of the generative adversarial network according to the total loss function value. When the total loss function value is minimized, the trained generative adversarial network is obtained.

[0060] Introduce the Wasserstein distance to measure the probability difference between the distribution of real samples and the distribution of generated samples, make the generated samples closer to the real sample distribution, and improve their authenticity. The difference loss function L D is expressed as:

[0061]

[0062] Among them, P real is the sample distribution of real commodity data at multiple exposure levels, and P real represents the distribution of generated samples.

[0063] To enhance the training stability of the generative adversarial network, a gradient penalty term is introduced at the same time to make the discriminator satisfy 1-Lipschitz continuity; the gradient penalty term L lip is defined as:

[0064]

[0065] Among them, λ is the weight of the gradient penalty term, is the interpolation sample sampled between the real commodity data X real at multiple exposure levels and the generated samples , and is the distribution of the interpolation sample.

[0066] Then the total loss function of the generative adversarial network is:

[0067]

[0068] Among them, λ1 and λ2 are the weights of the loss functions L D , L lip respectively.

[0069] S3. Input the commodity data at multiple exposure levels into the trained generative adversarial network to obtain the generated samples at multiple exposure levels;

[0070] S4. Train the interest enhancement and multi-exposure fusion module with generated samples of multiple exposure levels to obtain a trained interest enhancement and multi-exposure fusion module;

[0071] The interest enhancement and multi-exposure fusion module includes: a feature extraction module and an enhanced multi-head attention module; the training process of interest enhancement and multi-exposure fusion specifically includes:

[0072] S41. Input the generated samples of multiple exposure levels into the feature extraction module to extract the feature vectors F of each exposure level (i) ;

[0073] S42. Input the feature vectors F of each exposure level (i) into the enhanced multi-head attention module to obtain the fusion features of each exposure level

[0074] As Figure 3 shown, the enhanced multi-head attention module includes multiple multi-head attention modules, and each multi-head attention module corresponds to one exposure level; the processing of the feature vector F of the corresponding exposure level by the multi-head attention module includes: (i) :

[0075] S421. Generate the query, key, and value matrices of the feature vector F (i) in each attention head:

[0076] Q (i) = W Q F (i)

[0077] K (i) = W K F (i)

[0078] V (i) = W V F (i)

[0079] where Q (i) , K (i) , V (i) are the corresponding query, key, and value matrices, and W Q , W K , W V are the weight matrices corresponding to the query, key, and value matrices. This step ensures that independent query, key, and value matrices are generated for each head and are ready for attention calculation. Specifically, the input feature vector F (i) is divided into multiple subspaces, and each subspace corresponds to one attention head. For example, assume there are h attention heads, and the dimension of each head is where d is the dimension of the feature vector F (i) Each head focuses on a different feature subspace to capture feature details.

[0080] S422. Calculate the feature vector F according to the query matrix in the attention head (i) The shared matrix between different attention heads

[0081] Calculate the cross-attention of each pair of different heads a and b (a≠b) to ensure information sharing between the heads. The specific sharing process is as follows:

[0082] For the a-th head, calculate the attention score between it and the b-th head:

[0083]

[0084] where represents the attention sharing matrix from head a to head b at the i-th exposure level, are the query matrices of head a and head b. By calculating the shared matrix between different heads the correlation information between the heads can be obtained.

[0085] To prevent the values in the shared matrix from being too large and affecting the calculation stability, it is necessary to perform normalization before its application to ensure that the shared information will not be over-amplified in the calculation. The normalization formula for the shared matrix is as follows:

[0086]

[0087] where k and l are the indices of the attention heads, is the normalized shared matrix at the i-th exposure level.

[0088] Through the normalized shared matrix, the weight ratio balance of each head is ensured, and the fairness of multi-head information sharing is achieved.

[0089] S423. Dynamically update the feature vector F according to the shared matrix to obtain the updated attention weight matrix (i) in the attention weight matrix of each attention head to obtain the updated attention weight matrix

[0090]

[0091] where α is a regulation parameter used to control the balance between self-features and head-shared information, is a learnable attention weight matrix. This weighted average strategy ensures that each head retains its original characteristics while integrating the information of other heads.

[0092] S424. Calculate the output features of each attention head according to the query matrix, key matrix, value matrix, and updated attention weight matrix in each attention head:

[0093]

[0094] where d k is the scaling factor, is the weighted output of the m-th head at the i-th exposure level, and normalization is ensured through the softmax operation. are the query, key, and value matrices of the m-th head respectively.

[0095] S425. Combine the output features of each attention head to obtain the fused features;

[0096] Specifically, concatenate the output features of each head into a total output matrix and integrate it through a fully connected layer to obtain the final fused feature representation at this level

[0097]

[0098] where W O is the output layer weight, which is used to integrate the outputs of each head in the fully connected layer to obtain the fused multi-level exposure sample features. To further enhance the feature expression, non-linear activation and regularization processing are added to prevent overfitting during the feature fusion process and improve the generalization ability of the model.

[0099] S43. Use the progressive reinforcement fusion mechanism to perform weighted fusion on the fused features of each exposure level to obtain the final fused feature H final ;

[0100] In order to suppress the out-of-range artifacts and low-frequency halo problems that may occur during the fusion process, the progressive reinforcement fusion mechanism is used during the fusion process.

[0101] Specifically, obtain the fusion weight

[0102] at the current iteration round t, i initialize the initial fusion weight β

[0103]

[0104] Among them, is the initial weight, ensuring that each feature participates in the fusion evenly.

[0105] According to the progressive reinforcement fusion mechanism, in each iteration, the weights of each layer are gradually adjusted through the update formula to optimize the fusion effect.

[0106] The update formula can be expressed as:

[0107]

[0108] Among them, η is the learning rate, is the fusion loss function. By gradually optimizing ensure the smooth transition of the feature weights of each layer.

[0109] According to the fusion weight weight the fusion features of all exposure levels to perform weighted fusion to obtain the preliminary fusion feature H f ′ inal ; the weighting formula is as follows:

[0110]

[0111] Among them, is the latest current fusion weight coefficient, reflecting the weight of each layer of features in the final result, and ensuring that the features of each layer are reasonably fused in the recommended samples.

[0112] In the progressive reinforcement fusion mechanism, an artifact detection mechanism is introduced to calculate the difference ΔH between the fusion results of different layers fused :

[0113]

[0114] When the difference value exceeds the set threshold, it is regarded as a potential artifact and adjusted to obtain the feature H″ final , and the process can be expressed as:

[0115] H″ final = H′ final - γ·ΔH fused

[0116] Among them, γ is the artifact suppression coefficient, which adjusts the fusion result and suppresses the low-frequency halo phenomenon.

[0117] Perform progressive smoothing on the optimized fusion feature H″ final to further eliminate the errors caused by excessive weight changes, and obtain:

[0118] H final= σ·H″ final +(1 - σ)·(W map ·H fused )

[0119] where σ is the smoothing factor that controls the smoothness of the final output feature, H fused is the fused feature forming a matrix, and W map is a linear transformation matrix used to project the feature dimension of H fused to the same dimension as the optimized fused feature.

[0120] S44. Calculate the fused loss function value based on the fused features of multiple exposure levels The final fused feature H final and update the parameters of the interest enhancement and multi - exposure fusion module according to the fused loss function value. When the fused loss function value is minimized, the trained interest enhancement and multi - exposure fusion module is obtained.

[0121] The smoothing loss function is used to constrain the change of the fusion weights to ensure a smooth transition of the weights. The formula is as follows:

[0122]

[0123] where t is the current iteration number, is the fusion weight of the i - th exposure level in the j - th iteration, is the fusion weight of the i - th exposure level in the (j - 1)-th iteration.

[0124] The feature difference loss function is used to constrain the difference between features of different exposure levels to avoid the artifact effect caused by over - emphasizing certain levels. The formula is as follows:

[0125]

[0126] is the fused feature of the i - th exposure level, represents the square of the L2 norm.

[0127] The low - frequency signal protection loss function is used to protect the low - frequency signal from being masked by the high - frequency signal. The formula is as follows:

[0128]

[0129] where, is the low - frequency feature of the i - th exposure level, which is obtained by low - pass filtering the fused feature extracted.

[0130] The complete loss function can be expressed as:

[0131]

[0132] where is the j-th loss function in the set R, and R is the set of loss functions, are the learnable weight parameters for each loss function:

[0133]

[0134] S5. Input the generated samples of multiple exposure levels into the trained interest enhancement and multi-exposure fusion module to obtain the final fused features;

[0135] S6. Train the matching degree calculation module based on the final fused features to obtain the trained matching degree calculation module.

[0136] The matching degree calculation module includes: a user feature extraction module, a commodity feature extraction module, and a calculation module; the training process of the matching degree calculation module includes:

[0137] Input the user data of the e-commerce platform into the user feature extraction module to obtain the feature vector U of each user u u ; input the commodity data of the e-commerce platform into the commodity feature extraction module to obtain the feature vector J of each commodity j j ;

[0138] Specifically, the extraction of the user feature extraction module includes: mapping discrete data such as user ID and gender through an embedding layer into dense vectors to obtain static features; inputting the user's historical behavior data into a bidirectional LSTM network to extract temporal features; concatenating the static features and the temporal features, and inputting the concatenated features into a fully connected layer to obtain the feature vector of the user.

[0139] The extraction of the commodity feature extraction module includes: converting discrete attributes such as commodity category and brand through an embedding layer into dense vectors to obtain attribute features; normalizing and concatenating continuous data such as the exposure times and click-through rates of the commodity to obtain statistical feature; concatenating the attribute encoding and the statistical features to obtain the feature vector of the commodity.

[0140] Input the feature vector of each user, the feature vector of each commodity, and the final fused features into the calculation module to obtain the matching degree score between the user and the commodity;

[0141] The matching degree score S u,j between the user and the commodity is as follows:

[0142]

[0143] Among them, W ma and b match are the learnable weight matrix and bias term, representing the concatenation operation of the user feature vector and the item feature vector; σ is the Sigmoid activation function that maps the matching degree score to the range of [0, 1];

[0144] Calculate the value of the loss function according to the matching degree score between the user and the item, update the parameters of the matching degree calculation module according to the value of the loss function, and when the value of the loss function is the smallest, obtain the trained matching degree calculation module.

[0145] Loss function is:

[0146]

[0147] y u,j is the true label (1 indicates that user u has a positive interaction with item j, 0 indicates no interaction), and D is the set of data of users and items on the e-commerce platform.

[0148] According to the matching degree score S u,j sort all candidate items to generate a sorted list of candidate items; select the top k items from the sorted list to generate a personalized recommendation list R u .

[0149] The above-mentioned embodiments further elaborate on the purpose, technical solutions, and advantages of the present invention. It should be understood that the above-mentioned embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A personalized recommendation de-exposure bias method based on multi-level exposure fusion, characterized in that, Including: Obtain user and product data of an e-commerce platform, construct product data with multiple exposure levels based on the product data of the e-commerce platform, and input the user and product data of the e-commerce platform and the product data with multiple exposure levels into a trained recommendation model to obtain the matching score between users and products; Obtain a recommendation list based on the matching score between users and products; The recommendation model includes: a generative adversarial network, an interest enhancement and multi-exposure fusion module, and a matching degree calculation module; The training process of the recommendation model includes: S1. Obtain user and product data of an e-commerce platform, and construct product data with multiple exposure levels based on the product data of the e-commerce platform; S2. Train the generative adversarial network according to the product data with multiple exposure levels to obtain a trained generative adversarial network; S3. Input the product data with multiple exposure levels into the trained generative adversarial network to obtain generated samples with multiple exposure levels; S4. Train the interest enhancement and multi-exposure fusion module according to the generated samples with multiple exposure levels to obtain a trained interest enhancement and multi-exposure fusion module; S5. Input the generated samples with multiple exposure levels into the trained interest enhancement and multi-exposure fusion module to obtain fused features with multiple exposure levels and final fused features; S6. Train the matching degree calculation module according to the user and product data of the e-commerce platform and the final fused features to obtain a trained matching degree calculation module.

2. The personalized recommendation de-exposure deviation method based on multi-level exposure fusion according to claim 1, characterized in that The generative adversarial network includes: a feature encoding module, a generator G, and a discriminator D; the training process of the generative adversarial network includes: S21. Input the product data with multiple exposure levels into the feature encoding module to obtain an exposure encoding vector E; S22. Add random noise Z to the exposure encoding vector E to obtain the combined input feature vector Z i npu t ; S23. Input the combined input feature vector Z input into the generator G to obtain the generated sample set X gen ; The generated sample set X gen includes generated samples with multiple exposure levels; S24. Perform interpolation sampling between the generated samples at multiple exposure levels and the product data at multiple exposure levels to obtain an interpolation sample set Interpolation sample set It includes interpolation samples at multiple exposure levels; S25. Input the generated samples with multiple exposure levels, product data, and interpolation samples into the discriminator D respectively to obtain corresponding discrimination results; S26. Calculate the total loss function value according to the discrimination results of the generated samples with multiple exposure levels, product data, and interpolation samples, update the parameters of the generative adversarial network according to the total loss function value, and when the total loss function value is the smallest, obtain a trained generative adversarial network.

3. A personalized recommendation de-exposure deviation method based on multi-level exposure fusion according to claim 2, characterized in that The total loss function of the generative adversarial network is: Among them, L D is the difference loss function between the discrimination results of commodity data at multiple exposure levels and the discrimination results of generated samples, and λ lip is the gradient penalty term based on the discrimination results of interpolation samples at multiple exposure levels. λ1 and λ2 are the weights of the loss function L D and the gradient penalty term L lip respectively.

4. A personalized recommendation de-exposure bias method based on multi-level exposure fusion according to claim 1, characterized in that The interest enhancement and multi-exposure fusion module includes: a feature extraction module and an enhanced multi-head attention module; the training process of the interest enhancement and multi-exposure fusion module includes: S41. Input the generation samples of multiple exposure levels into the feature extraction module to extract the feature vectors F of each exposure level (i) ; S42. Input the feature vector F of each exposure level (i) into the enhanced multi-head attention module to obtain the fused features of each exposure level S43. Use the progressive reinforcement fusion mechanism to fuse the fusion features of all exposure levels by weighted fusion to obtain the final fusion feature H final ; S44. According to the fusion features of multiple exposure levels The final fusion feature H final Calculate the fusion loss function value, update the parameters of the interest enhancement and multi-exposure fusion module according to the fusion loss function value. When the fusion loss function value is minimized, the trained interest enhancement and multi-exposure fusion module is obtained.

5. A personalized recommendation debiasing method based on multi-level exposure fusion according to claim 4, characterized in that, The enhanced multi-head attention module includes multiple multi-head attention modules, and each multi-head attention module corresponds to one exposure level; The processing of the feature vector corresponding to the exposure level by the multi-head attention module includes: S421. Generate query, key, and value matrices of the feature vector in each attention head; S422. Calculate the shared matrix of the feature vector between different attention heads according to the query matrix in the attention head; S423. Update the weight matrix of the feature vector in each attention head according to the shared matrix to obtain the updated weight matrix of each attention head; S424. Calculate the output feature of each attention head according to the query matrix, key matrix, value matrix, and updated weight matrix in each attention head; S425. Combine the output features of each attention head to obtain a fused feature.

6. A personalized recommendation debiasing method based on multi-level exposure fusion according to claim 4, characterized in that, The fusion features of all exposure levels Performing weighted fusion includes: obtaining the fusion weights of the current iteration round, and performing weighted fusion on the fusion features of all exposure levels to obtain the preliminary fusion feature H f ′ inal ; calculating the fusion feature difference ΔH based on the fusion features of all exposure levels f ′ used , combining the preliminary fusion feature H f ′ inal , the fusion feature difference ΔH fused and the fusion features of all exposure levels to obtain the final multi-level fusion feature H final .

7. A personalized recommendation de-exposure bias method based on multi-level exposure fusion according to claim 6, characterized in that According to the fusion features of all exposure levels Calculate the fusion feature difference ΔH fused Including: Where n is the number of exposure levels.

8. A personalized recommendation de-exposure bias method based on multi-level exposure fusion according to claim 6, characterized in that, Combine the preliminary fusion feature H′ final , the fusion feature difference ΔH fused and the fusion features of all exposure levels The combination includes: H final = σ · (H′ final - γ · ΔH fused ) + (1 - σ) · (W map · H fused ) where σ is the smoothing factor, γ is the artifact suppression coefficient, and W map is the linear transformation matrix, and H fused is the fused feature of all exposure levels concatenated into a matrix.

9. A personalized recommendation de-exposure bias method based on multi-level exposure fusion according to claim 4, characterized in that, Fusion loss function is as follows: Among them, is the j-th loss function in the set R, where R is the set of loss functions, is the smoothed loss function of the fusion weights before the current iteration round, is the difference loss function between the fusion features at different exposure levels, is the loss function between the fusion feature and its low-frequency feature, and α j is the weight of each loss function in the current iteration round.

10. A method for removing exposure bias in personalized recommendation based on multi-level exposure fusion according to claim 1, characterized in that, The matching degree calculation module includes: a user feature extraction module, a commodity feature extraction module, and a calculation module; the training process of the matching degree calculation module includes: inputting the user data of the e-commerce platform into the user feature extraction module to obtain the feature vector of each user; inputting the commodity data of the e-commerce platform into the commodity feature extraction module to obtain the feature vector of each commodity; inputting the feature vector of each user, the feature vector of each commodity, and the final fused feature into the calculation module to obtain the matching degree score between each user and each commodity; calculating the loss function value according to the matching degree score between each user and each commodity, updating the parameters of the matching degree calculation module according to the loss function value, and when the loss function value is the smallest, obtaining the trained matching degree calculation module.

Images