Multimedia-oriented generative generalization cold start recommendation method

Abstract

The application discloses a kind of multimedia-oriented generative generalization cold start recommendation method, comprising:1. the interactive record of user and product is used to construct score matrix;2. input layer is constructed by one-hot encoding mode, and user, product is mapped to different embedding space by combining the multimedia features of product;3. the embedding of user, product is optimized by bayesian ranking loss function;4. construct generative neural network, including: prior neural network, encoder, decoder;5. construct uniformity enhanced conditional variational autoencoder, to make the latent space of original variational autoencoder more uniform and distinguishable;6. new product embedding generation method based on clustering obtains new product generative embedding;7. the degree of love of user to new product is predicted by the way of vector dot product.The application can make recommendations for new products using multimedia information of products, thereby alleviating the cold start problem in the recommendation system.

Description

Technical Field

[0001] This invention relates to the field of recommendation, specifically a generative recommendation method to alleviate the cold start problem in multimedia recommendation systems. Background Technology

[0002] Personalized recommendations have become an important means of helping users cope with information overload in online applications such as e-commerce and advertising. Learning accurate user and item representations is key to building effective recommendation systems. Among them, multimedia-based recommendation is an attractive technique that fully utilizes user-product interactions and rich multimedia features for representation learning. Generally, multimedia-based recommendation takes multimedia features and user IDs as inputs, feeds them into an embedding layer for user and product representation learning, and then models the interactions through inner product or neural networks. Although multimedia-based recommendation algorithms can effectively provide high-quality representations for recommendation needs, they have not yet been generalized to cold-start recommendation scenarios. In real-world scenarios, cold-start products appear rapidly over time, especially on news or short video recommendation platforms. Cold-start products only have multimedia features and lack historical interactions; therefore, generating high-quality product representations is a key challenge for cold-start product recommendation.

[0003] To overcome this problem, a common solution is to use multimedia features (e.g., images, text, knowledge graphs) as a bridge between pre-trained warm product representations and cold product representations.

[0004] To align the distributions of warm and cold representations, previous work attempted to narrow the gap between warm and multimedia feature representations by carefully designing alignment functions (e.g., mean squared error, mutual information maximization). The multimedia feature representation of the cold product was then used to preheat the cold product.

[0005] However, existing methods still have some limitations. Specifically, (1) the alignment function needs to be carefully designed, and it is difficult to guarantee the consistency of the cold and warm representation distributions through the function. In addition, cold representations are generated from multimedia features, while warm representations are learned from historical interactions and content, giving warm products more information. The method of bringing the two distributions closer together may reduce the representational power of warm products. (2) Aligning each discrete sample pair one-to-one may expose the model to some noise. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention breaks away from the recommendation model based on alignment functions and proposes a generative generalized cold-start recommendation method oriented towards multimedia. This method aims to utilize the multimedia information of products to make recommendations for new products, thereby alleviating the cold-start problem in recommendation systems.

[0007] The present invention adopts the following technical solution to solve the technical problem:

[0008] The present invention provides a generative generalized cold-start recommendation method for multimedia, characterized by the following steps:

[0009] Step 1: Construct a rating matrix using user and product interaction records:

[0010] Suppose there are M users and N products, let r mi Represents the m-th user u m For the i-th product v i The interaction situation is then denoted as R = {r}. mi} M×N If the m-th user u m For the i-th product v i If there is an interaction record, then let r mi =1, otherwise, let r mi =0;

[0011] Step 2: Construct the input layer using one-hot encoding and, in conjunction with the product's multimedia features, map users and products to different embedding spaces:

[0012] The multimedia feature matrix C = {c1, ..., c2} of the product is constructed using one-hot encoding. i， ...c N The user's embedding representation matrix U = {u1, ..., u2} m ,...u M The initial representation matrix of the product is V″ = {v″1, ..., v″}. i ,…..v″ N The product's representation matrix V = {v1, ..., v2} i ,...v N}; where c i v represents the i-th product i Multimedia feature vectors; u m Represents the m-th user u m The representation vector; v″ i v represents the i-th product i The initial representation vector; v i Let v represent the representation vector of the i-th product after incorporating multimedia features, and v i =v″ i +c i ;

[0013] Step 3: Optimize user and product embeddings using the Bayesian ranking loss function:

[0014] Step 3.1: Calculate the m-th user u according to equation (1). m For the i-th product vi level of liking

[0015]

[0016] In equation (1), < and > denote the vector dot product;

[0017] Step 3.2: Construct the loss function L′ according to equation (2), and optimize the loss function L′ using the stochastic gradient descent method to minimize L′, thereby obtaining the optimal user embedding matrix U. * Optimal product embedding matrix V * :

[0018]

[0019] In equation (2), (u m v i ) represents r in the interaction matrix R mi =1 corresponds to the m-th user u m and the i-th product v i ;(u m v i′ ) represents r in the interaction matrix R mi′ The m-th user u corresponding to =0 m and the i′th product v i′ ; The m-th user u m For the i′-th product v i′ The degree of liking; δ is the sigmoid activation function;

[0020] Step 4: Construct a conditional variational autoencoder, including: a prior neural network, encoder, and decoder.

[0021] Step 4.1: The prior neural network is based on the i-th product v i Multimedia feature vector c i The i-th product v is obtained through equations (3) and (4) respectively. i latent variable z i The mean μ of the prior distribution i The variance σ of the prior distribution i :

[0022] μ i =c i ·W μ +b μ (3)

[0023] σ i =c i ·W σ +b σ(4)

[0024] In equations (3) and (4), W μ W σ These are the two weight parameters to be learned in the prior neural network, b μ b σ These are the two bias parameters to be learned;

[0025] Step 4.2: The encoder, based on the i-th product v i Multimedia feature vector c i And the product embedding of the i-th product in the optimal product embedding matrix V* The i-th product v is obtained through equations (5) and (6) respectively. i latent variable z i The mean μ of the posterior distribution i The variance σ of the posterior distribution and the posterior distribution i ′:

[0026]

[0027]

[0028] In equations (5) and (6), W μ′ W σ′ These are the two weight parameters to be learned in the encoder, b μ′ b σ′ These are the two bias parameters to be learned in the encoder, and [;] indicates the concatenation operation;

[0029] Step 4.3: Sample random variables ∈ from the standard normal distribution, and then randomly sample the i-th product v according to equation (7). i latent variable z i :

[0030] z i =μ i′ +σ i′ ⊙∈ (7)

[0031] In equation (7), ⊙ represents the Hadamard product;

[0032] Step 4.4: The decoder, based on the i-th product v i latent variable z i and multimedia feature vector c i The i-th product v is obtained through equation (8). i Reconstructing product embedding v i ′:

[0033] v i ′=f ψ ([z ic i (8)

[0034] In equation (8), f ψ (·) represents a multilayer perceptron;

[0035] Step 5: Construct a conditional variational autoencoder with enhanced uniformity:

[0036] Step 5.1: Construct the ELBO loss function based on the conditional variational autoencoder using equation (9):

[0037]

[0038] In equation (9), and v′ iy These respectively represent product embedding and reconstructing product embedding v i The value of d in the y-th dimension. v KL[·||·] represents the dimension; KL[·||·] represents the KL divergence between two distributions. Represents a normal distribution;

[0039] Step 5.2: Construct the uniformity enhancement loss function L according to equation (10). uni :

[0040]

[0041] In equation (10), Indicates the i1th product latent variables The mean of the prior distribution; Indicates the i2th product latent variables The mean of the prior distribution; Expressing expectations;

[0042] Step 5.3: Construct the overall loss function L according to equation (11):

[0043] L = L ELBO +αL uni (11)

[0044] In equation (11), α represents the coefficient of uniformity enhancement loss;

[0045] Step 5.4: Optimize the loss function L using stochastic gradient descent to minimize equation (11), thereby obtaining the conditional variational autoencoder with enhanced uniformity after training.

[0046] Step 6: Using a clustering-based new product embedding generation method, obtain the generated embedding of the new product:

[0047] Step 6.1: Based on the i-th product v i Multimedia feature embedding c i The k-means algorithm is used to embed the optimal product into the matrix V. * Clustered into K clusters;

[0048] Step 6.2: Calculate the preference center representation of the k-th cluster according to equations (12) and (13) respectively. k and the content center indicates p k :

[0049]

[0050]

[0051] In equations (12) and (13), N k S is the number of products in the k-th cluster. k It is the set of products belonging to the k-th cluster;

[0052] Step 6.3: Based on the multimedia features of the new product to be recommended, c new Using equation (14), we find the cluster k corresponding to the most similar cluster center. new :

[0053]

[0054] Step 6.4, let Indicates cluster k new The preference center vector will The input is fed into the trained encoder to generate the latent variable z of the new product according to equations (15) and (16). new mean μ new ′ and variance σ new ′:

[0055] μ new ′=[p k c new ]·W μ ′+b μ (15)

[0056] σ new ′=[p k c new ]·W σ ′+b σ ′ (16)

[0057] Step 6.5: Obtain the latent variable z of the new product by sampling according to equation (17). new :

[0058] znew =μ new ′+σ new ′⊙∈ (17)

[0059] Step 6.6: The potential variables z of the new product new and multimedia feature representation c new The input is fed into the trained decoder, thereby generating the embedded representation v′ of the new product according to equation (18). new :

[0060] v′ new =f ψ ([z new c new (18)

[0061] Step 7: Predict the m-th user u according to equation (19) m The degree of liking for new products

[0062]

[0063] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the generative generalization cold start recommendation method, and the processor is configured to execute the program stored in the memory.

[0064] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the generative generalized cold start recommendation method.

[0065] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0066] 1. This invention addresses the problem of recommending new multimedia projects by proposing a generative cold-start project recommendation framework based on a conditional variational autoencoder. It reconstructs and models the distribution of product embeddings by reconstructing the optimal product embeddings, and generates new product embeddings from the distribution based on multimedia features. Compared with current alignment-based methods, this invention is less affected by noisy data and can fully utilize the multimedia features of new products to generate high-quality new product representations.

[0067] 2. This invention designs a uniformity enhancement optimization objective. By optimizing the uniformity of the mean values ​​of the distributions corresponding to different latent variables in the conditional variational autoencoder, the overlapping of different latent variable distributions is reduced, so that the sampled latent variables belong to a single distribution rather than multiple distributions. This ensures that the latent space of the conditional variational autoencoder is more distinguishable and informative, resulting in a more accurate representation of the generated new product.

[0068] 3. This invention designs a clustering-aware method to obtain fuzzy pre-trained representations of items. It uses the multimedia embeddings of products for clustering, calculates product embedding centers and multimedia feature centers respectively, and selects appropriate product embedding centers to input into the model based on the multimedia features of new products, providing more information to the model to better generate item representations.

[0069] 4. This invention directly models the conditional distribution of warm representation based on multimedia features. By modeling the distribution, the warm representation can be obtained from the multimedia features during the testing phase without considering the availability of historical interaction data. This generative method fundamentally solves the previous limitations. Attached Figure Description

[0070] Figure 1 This is a flowchart of a generative multimedia cold start recommendation method according to the present invention. Detailed Implementation

[0071] In this embodiment, a method for generating corrective samples to improve data fairness in a recommender system includes: constructing a rating matrix using user and product interaction records; obtaining the optimal product embedding by combining the product's multimedia features; reconstructing the optimal embedding based on a uniformity-enhanced conditional variational autoencoder; and obtaining the generated embedding of the new product using a clustering-based new product embedding generation method. The overall process is as follows: Figure 1 As shown, specifically, it is done according to the following steps:

[0072] Step 1: Construct a rating matrix using user and product interaction records:

[0073] Suppose there are M users and N products, let r mi Represents the m-th user u m For the i-th product v i The interaction situation is then denoted as R = {r}. mi} M×N If the m-th user u m For the i-th product v i If there is an interaction record, then let r mi =1, otherwise, let r mi =0; The data sources for interaction relationships include various implicit feedback data such as clicks, favorites, and purchases within the system.

[0074] Step 2: Construct the input layer using one-hot encoding and, in conjunction with the product's multimedia features, map users and products to different embedding spaces:

[0075] The multimedia feature matrix C = {c1, ..., c2} of the product is constructed using one-hot encoding. i , ...c NThe user's embedding representation matrix U = {u1, ..., u2} m ,...u M The initial representation matrix of the product is V″ = {v″1, ..., v″}. i ,...v″ N The product's representation matrix V = {v1, ..., v2} i ,...v N}; where c i v represents the i-th product i Multimedia feature vectors; u m Represents the m-th user u m The representation vector; v″ i v represents the i-th product i The initial representation vector; v i Let v represent the representation vector of the i-th product after incorporating multimedia features, and v i =v″ i +c i Product representation vector v i It also includes collaborative filtering information from interaction records and content information from multimedia features.

[0076] Step 3: Optimize user and product embeddings using the Bayesian ranking loss function:

[0077] Step 3.1: Calculate the m-th user u according to equation (1). m For the i-th product v i level of liking

[0078]

[0079] In equation (1), < and > denote the vector dot product;

[0080] Step 3.2: Construct the loss function L′ according to equation (2), and optimize the loss function L′ using the stochastic gradient descent method to minimize L′, thereby obtaining the optimal user embedding matrix U. * Optimal product embedding matrix V * :

[0081]

[0082] In equation (2), (u m v i ) represents r in the interaction matrix R mi =1 corresponds to the m-th user u m and the i-th product v i ;(u m v i′ ) represents r in the interaction matrix Rmi′ The m-th user u corresponding to =0 m and the i′th product v i ′; The m-th user u m For the i′-th product v i′ The degree of liking; δ is the sigmoid activation function.

[0083] Step 4: Construct a conditional variational autoencoder, including: a prior neural network, encoder, and decoder.

[0084] Suppose the optimal product embedding of the i-th product is determined by a latent variable z. i and its multimedia feature vector c i Together, we decided to use a conditional variational autoencoder to model the conditional distribution p(V). i ′|z i c i This allows the model to be based on the latent variable z of the new product. new and multimedia feature vector c new Generate product embeddings for it.

[0085] Step 4.1: Assume the prior distribution of the latent variable zi is... in, Let μ represent a normal distribution, and ~ represent a normal distribution. Therefore, the mean μ can be parameterized using a prior neural network. i and variance σ i Represent z i The prior distribution. The prior neural network is based on the i-th product v. i Multimedia feature vector c i The i-th product v is obtained through equations (3) and (4) respectively. i latent variable z i The mean μ of the prior distribution i The variance σ of the prior distribution i :

[0086] μ i =c i ·W μ +b μ (3)

[0087] σ i =c i ·W σ +b σ (4)

[0088] In equations (3) and (4), W μ W σ These are two weight parameters to be learned in the prior neural network, b μ bσ These are the two bias parameters to be learned.

[0089] Step 4.2, Assume the latent variable z i The posterior distribution is Therefore, the mean μ′ can be parameterized using the encoder. i and variance σ i ′ represents z i The prior distribution. The encoder is based on the i-th product v. i Multimedia feature vector c i and the optimal product embedding matrix V * Product embedding of the i-th product The i-th product v is obtained through equations (5) and (6) respectively. i latent variable z i The mean μ of the posterior distribution i The variance σ of the posterior distribution and the posterior distribution i ′:

[0090]

[0091]

[0092] In equations (5) and (6), W μ′ W σ′ These are two weight parameters to be learned in the encoder, b μ′ b σ′ These are the two bias parameters to be learned in the encoder, and [;] indicates the concatenation operation.

[0093] Step 4.3, because directly from Sampling latent variables from a distribution results in non-differentiable computations, making it impossible to update model parameters using gradient backpropagation. Therefore, resampling techniques are used to obtain the latent variable z. i Specifically, this is implemented by sampling random variables ∈ from a standard normal distribution, and then randomly sampling the i-th product v according to equation (7). i latent variable z i :

[0094] z i =μ i′ +σ i′ ⊙∈ (7)

[0095] In equation (7), ⊙ represents the Hadamard product.

[0096] Step 4.4: The decoder determines the i-th product v. i latent variable z i and multimedia feature vector c i The i-th product v is obtained through equation (8).i Reconstructing product embedding v i ′:

[0097] v i ′=f ψ ([z i c i (8)

[0098] In equation (8), f ψ (·) represents a multilayer perceptron.

[0099] Step 5: Construct a uniformity-enhanced conditional variational autoencoder. By optimizing the uniformity of the means of the distributions corresponding to different latent variables in the conditional variational autoencoder, the overlap between different latent variable distributions is reduced, ensuring that the sampled latent variables belong to a single distribution rather than multiple distributions. This ensures that the latent space of the conditional variational autoencoder is more distinguishable and informative, resulting in a more accurate representation of the generated new product.

[0100] Step 5.1: Construct the ELBO loss function based on the conditional variational autoencoder using equation (9):

[0101]

[0102] In equation (9), and v′ iy These respectively represent product embedding and reconstructing product embedding v i The value of d in the y-th dimension. v The dimension is represented by ; KL[·||·] represents the KL divergence between two distributions.

[0103] Step 5.2: Construct the uniformity enhancement loss function L according to equation (10). uni :

[0104]

[0105] In equation (10), Indicates the i1th product latent variables The mean of the prior distribution; Indicates the i2th product latent variables The mean of the prior distribution; It expresses expectation.

[0106] Step 5.3: Construct the overall loss function L according to equation (11):

[0107] L = L ELBO +αL uni (11)

[0108] In equation (11), α represents the coefficient of uniformity enhancement loss.

[0109] Step 5.4: Optimize the loss function L using stochastic gradient descent to minimize equation (11), thereby obtaining the conditional variational autoencoder with enhanced uniformity after training.

[0110] Step 6: Using a clustering-based new product embedding generation method, obtain the generated embedding of the new product:

[0111] Step 6.1: Based on the i-th product v i Multimedia feature embedding c i The k-means algorithm is used to embed the optimal product into the matrix V. * The products are clustered into K clusters, and the products in each cluster have similar multimedia feature embeddings.

[0112] Step 6.2: Calculate the preference center representation of the k-th cluster according to equations (12) and (13) respectively. k and the content center indicates p k :

[0113]

[0114]

[0115] In equations (12) and (13), N k S is the number of products in the k-th cluster. k It is the set of products belonging to the k-th cluster.

[0116] Step 6.3: Based on the multimedia features of the new product to be recommended, c new Using equation (14), we find the cluster k corresponding to the most similar cluster center. new :

[0117]

[0118] Step 6.4, let Indicates cluster k new The preference center vector is used to obtain cluster preference centers based on the similarity of multimedia features. This can be viewed as a fuzzy embedding of the new product. Utilizing this fuzzy embedding to obtain more accurate latent variables helps the model generate more accurate new product embeddings. The input is fed into the trained encoder to generate the latent variable z of the new product according to equations (15) and (16). new mean μ new ′ and variance σ new ′:

[0119] μnew ′=[p k c new ]·W μ ′+b μ (15)

[0120] σ new ′=[p k c new ]·W σ ′+b σ ′ (16)

[0121] Step 6.5: Obtain the latent variable z of the new product by sampling according to equation (17). new :

[0122] z new =μ new ′+σ new ′⊙∈ (17)

[0123] Step 6.6: The potential variables z of the new product new and multimedia feature representation c new The input is fed into the trained decoder, thereby generating the embedded representation v′ of the new product according to equation (18). new :

[0124] v′ new =f ψ ([z new c new (18)

[0125] Step 7: Predict the m-th user u according to equation (19) m The degree of liking for new products

[0126]

[0127] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0128] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0129] Example:

[0130] To verify the effectiveness of this method, this invention uses Amazon-Baby, a publicly available dataset commonly used in the field of multimedia recommender systems. The experiment was conducted under a cold-start setting, meaning that none of the products in the test set appeared in the training set, and the multimedia features of the products in the test set were unavailable during the training phase. This invention uses the widely adopted recall rate (Recall@K) and normalized discount gain (NDCG@K) as performance evaluation metrics for the recommender system. Recall@K measures the percentage of items recalled from the top-K rank list of items truly liked by users in the test data. NDCG@K further considers the hit position of the recalled items. Higher results for both metrics indicate better recommender system performance.

[0131] Table 1. Cold start performance results of the method of the present invention and the comparative method on Amazon Baby.

[0132]

[0133] As shown in Table 1, to illustrate the effectiveness of the present invention, several common cold start recommendation models were selected for comparison, including (1) content-based cold start methods, KNN, DUIF; (2) robustness-based cold start methods, DropoutNet, MTPR; (3) alignment-based cold start methods, Heater, CLCRec; and (4) generative cold start method GAR.

[0134] The present invention demonstrates better recommendation performance under the cold start setting of the Amazon-Baby dataset; the experimental results fully verify the effectiveness of the present invention in improving the performance of multimedia cold start recommendation.

Claims

1. A generative generalized cold-start recommendation method for multimedia, characterized in that, The procedure is as follows: Step 1: Construct a rating matrix using user and product interaction records: Suppose there are M users and N products, let r mi Represents the m-th user u m For the i-th product v i The interaction situation is then denoted as R = {r}. mi } M×N If the m-th user u m For the i-th product v i If there is an interaction record, then let r mi =1, otherwise, let r mi =0; Step 2: Construct the input layer using one-hot encoding and, in conjunction with the product's multimedia features, map users and products to different embedding spaces: The multimedia feature matrix C = {c1, ..., c2} of the product is constructed using one-hot encoding. i , ...c N The user's embedding representation matrix U = {u1, ..., u2} m ,...u M The initial representation matrix of the product is V″ = {v″1, ..., v″}. i ,...v″ N The product's representation matrix V = {v1, ..., v2} i ,...v N }; where c i v represents the i-th product i Multimedia feature vectors; u m Represents the m-th user u m The representation vector; v″ i v represents the i-th product i The initial representation vector; v i Let v represent the representation vector of the i-th product after incorporating multimedia features, and v i =v″ i +c i ; Step 3: Optimize user and product embeddings using the Bayesian ranking loss function: Step 3.1: Calculate the m-th user u according to equation (1). m For the i-th product v i level of liking In equation (1), < and > denote the vector dot product; Step 3.2: Construct the loss function L′ according to equation (2), and optimize the loss function L′ using the stochastic gradient descent method to minimize L′, thereby obtaining the optimal user embedding matrix U. * Optimal product embedding matrix V * : In equation (2), (u m v i ) represents r in the interaction matrix R mi =1 corresponds to the m-th user u m and the i-th product v i ;(u m v i′ ) represents r in the interaction matrix R mi′ The m-th user u corresponding to =0 m and the i′th product v i′ ; The m-th user u m For the i′-th product v i′ The degree of liking; δ is the sigmoid activation function; Step 4: Construct a conditional variational autoencoder, including: a prior neural network, encoder, and decoder. Step 4.1: The prior neural network is based on the i-th product v i Multimedia feature vector c i The i-th product v is obtained through equations (3) and (4) respectively. i latent variable z i The mean μ of the prior distribution i The variance σ of the prior distribution i : m i =c i ·W μ +b μ (3) s i =c i ·W σ +b σ (4) In equations (3) and (4), W μ W σ These are the two weight parameters to be learned in the prior neural network, b μ b σ These are the two bias parameters to be learned; Step 4.2: The encoder, based on the i-th product v i Multimedia feature vector c i and the optimal product embedding matrix V * Product embedding of the i-th product The i-th product v is obtained through equations (5) and (6) respectively. i latent variable z i The mean μ of the posterior distribution i The variance σ of the posterior distribution and the posterior distribution. i ′: In equations (5) and (6), W μ′ W σ′ These are the two weight parameters to be learned in the encoder, b μ′ b σ′ These are the two bias parameters to be learned in the encoder, and [;] indicates the concatenation operation; Step 4.3: Sample random variables ∈ from the standard normal distribution, and then randomly sample the i-th product v according to equation (7). i latent variable z i : z i =μ i ′+σ i ′⊙∈(7) In equation (7), ⊙ represents the Hadamard product; Step 4.4: The decoder, based on the i-th product v i latent variable z i and multimedia feature vector c i The i-th product v is obtained through equation (8). i Reconstructing product embedding v i ′: v i ′=f ψ ([z i ;c i ]) (8) In equation (8), f ψ (·) represents a multilayer perceptron; Step 5: Construct a conditional variational autoencoder with enhanced uniformity: Step 5.1: Construct the ELBO loss function based on the conditional variational autoencoder using equation (9): In equation (9), and v′ iy These respectively represent product embedding and reconstructing product embedding v i The value of d in the y-th dimension. v KL[·||·] represents the dimension; KL[·||·] represents the KL divergence between two distributions. Represents a normal distribution; Step 5.2: Construct the uniformity enhancement loss function L according to equation (10). uni : In equation (10), Indicates the i1th product latent variables The mean of the prior distribution; Indicates the i2th product latent variables The mean of the prior distribution; Expressing expectations; Step 5.3: Construct the overall loss function L according to equation (11): L=L ELBO +αL uni (11) In equation (11), α represents the coefficient of uniformity enhancement loss; Step 5.4: Optimize the loss function L using stochastic gradient descent to minimize equation (11), thereby obtaining the conditional variational autoencoder with enhanced uniformity after training. Step 6: Using a clustering-based new product embedding generation method, obtain the generated embedding of the new product: Step 6.1: Based on the i-th product v i Multimedia feature embedding c i The k-means algorithm is used to embed the optimal product into the matrix V. * Clustered into K clusters; Step 6.2: Calculate the preference center representation of the k-th cluster according to equations (12) and (13) respectively. k and the content center indicates p k : In equations (12) and (13), N k S is the number of products in the k-th cluster. k It is the set of products belonging to the k-th cluster; Step 6.3: Based on the multimedia features of the new product to be recommended, c new Using equation (14), we find the cluster k corresponding to the most similar cluster center. new : Step 6.4, let Indicates cluster k new The preference center vector will The input is fed into the trained encoder to generate the latent variable z of the new product according to equations (15) and (16). new mean μ new ′ and variance σ new ′: μ new ′=[p k ;c new ]·W μ′ +b μ′ (15) s new ′=[p k ;c new ]·W σ′ +b σ′ (16) Step 6.5: Obtain the latent variable z of the new product by sampling according to equation (17). new : z new =μ new ′+σ new ′⊙∈ (17) Step 6.6: The potential variables z of the new product new and multimedia feature representation c new The input is fed into the trained decoder, thereby generating the embedded representation v′ of the new product according to equation (18). new : v′ new =f ψ ([z new ;c new ]) (18) Step 7: Predict the m-th user u according to equation (19) m The degree of liking for new products 2. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the generative generalization cold start recommendation method of claim 1, wherein the processor is configured to execute the programs stored in the memory.

3. A computer-readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to perform the steps of the generative generalized cold start recommendation method of claim 1.

Images