Training method of recommendation model, creative recommendation method, device, equipment and medium

Abstract

The application discloses a kind of training method of recommendation model, creative recommendation method, device, equipment and medium.The method comprises: inputting ranking training set into initial commodity ranking model, obtaining the predicted commodity ranking result of output, obtaining the shared training feature corresponding to shared training set extracted by initial commodity ranking model, inputting creative training set and shared training feature into initial creative recommendation model, obtaining the predicted creative recommendation result of output, according to predicted commodity ranking result, initial commodity ranking model is iteratively trained to obtain the target commodity ranking model of training completion, and according to predicted creative recommendation result, initial creative recommendation model is iteratively trained to obtain the target creative recommendation model of training completion;Wherein, ranking training set true contains shared training set, and the model complexity of initial commodity ranking model is higher than the model complexity of initial creative recommendation model, improve the creative recommendation effect of creative recommendation model.

Description

Technical Field

[0001] This invention relates to the field of Internet technology, and in particular to a method for training a recommendation model, a creative recommendation method, an apparatus, a device, and a medium. Background Technology

[0002] In e-commerce, user behavior is influenced by the creative presentation of products. Presenting suitable creative content to different users has a significant impact on user behavior, platform traffic, and product sales.

[0003] In the process of realizing this invention, at least the following technical problems were found in the prior art:

[0004] Because creative recommendation models process massive amounts of data and need to meet the millisecond-level response requirements of e-commerce platforms, they cannot employ rich and diverse feature systems, resulting in poor creative recommendation performance. Summary of the Invention

[0005] This invention provides a training method for a recommendation model, a creative recommendation method, an apparatus, a device, and a medium to address the problem of poor expressive power in creative recommendation models and improve their creative recommendation performance.

[0006] According to an embodiment of the present invention, a method for training a recommendation model is provided, the method comprising:

[0007] The ranking training set is input into the initial product ranking model, and the creative training set is input into the initial creative recommendation model; wherein, the initial product ranking model includes a ranking feature network and a ranking output network, and the initial creative recommendation model includes a creative feature network and a creative output network;

[0008] The ranking feature network extracts features from the ranking training set to obtain ranking training features, and inputs the ranking training features into the ranking output network to obtain the output predicted product ranking result.

[0009] The shared training features corresponding to the shared training set are obtained from the ranked training features, and the creative training features are extracted from the creative training set through the creative feature network to obtain creative training features. The shared training features and the creative training features are then input into the creative output network to obtain the output predicted creative recommendation results.

[0010] Based on the predicted product ranking results, the initial product ranking model is iteratively trained to obtain the trained target product ranking model, and based on the predicted creative recommendation results, the initial creative recommendation model is iteratively trained to obtain the trained target creative recommendation model.

[0011] The ranking training set truly includes the shared training set, which contains user features and / or product features. The model complexity of the initial product ranking model is higher than that of the initial creative recommendation model.

[0012] According to another embodiment of the present invention, a creative recommendation method is provided, the method comprising:

[0013] The creative application set and the target application set are input into a pre-trained target creative recommendation model; wherein, the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network;

[0014] Creative application features are obtained by extracting features from the creative application set through the creative feature network, and target application features are obtained by extracting features from the target application set through the target feature network. The creative application features and the target application features are then input into the creative output network to obtain the output target creative recommendation results.

[0015] Based on the target creative recommendation results, display the creative product page;

[0016] The target application set includes feature values ​​corresponding to shared features in the shared training set. The shared features are user features or product features. The target creative recommendation model is obtained by training the recommendation model according to any embodiment of the present invention. The target feature network is the feature network in the target product ranking model used to extract shared training features from the shared training set.

[0017] According to another embodiment of the present invention, a training apparatus for a recommendation model is provided, the apparatus comprising:

[0018] The sorting training set input module is used to input the sorting training set into the initial product sorting model and the creative training set into the initial creative recommendation model; wherein, the initial product sorting model includes a sorting feature network and a sorting output network, and the initial creative recommendation model includes a creative feature network and a creative output network;

[0019] The predictive product ranking result output module is used to extract features from the ranking training set through the ranking feature network to obtain ranking training features, and input the ranking training features into the ranking output network to obtain the output predicted product ranking result.

[0020] The predictive creative recommendation result output module is used to obtain the shared training features corresponding to the shared training set in the ranking training features, extract creative training features from the creative training set through the creative feature network, and input the shared training features and the creative training features into the creative output network to obtain the output predictive creative recommendation result.

[0021] The creative recommendation model training module is used to iteratively train the initial product ranking model to obtain a trained target product ranking model based on the predicted product ranking results, and to iteratively train the initial creative recommendation model to obtain a trained target creative recommendation model based on the predicted creative recommendation results.

[0022] The ranking training set truly includes the shared training set, which contains user features and / or product features. The model complexity of the initial product ranking model is higher than that of the initial creative recommendation model.

[0023] According to another embodiment of the present invention, a creative recommendation device is provided, the device comprising:

[0024] The creative application set input module is used to input the creative application set and the target application set into the pre-trained target creative recommendation model; wherein, the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network;

[0025] The target creative recommendation result output module is used to extract creative application features from the creative application set through the creative feature network, extract target application features from the target application set through the target feature network, and input the creative application features and the target application features into the creative output network to obtain the output target creative recommendation result.

[0026] The creative product page display module is used to display the creative product page based on the target creative recommendation results;

[0027] The target application set includes feature values ​​corresponding to shared features in the shared training set. The shared features are user features or product features. The target creative recommendation model is obtained by training the recommendation model according to any embodiment of the present invention. The target feature network is the feature network in the target product ranking model used to extract shared training features from the shared training set.

[0028] According to another embodiment of the present invention, an electronic device is provided, the electronic device comprising:

[0029] At least one processor; and

[0030] A memory communicatively connected to the at least one processor; wherein,

[0031] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the training method or creative recommendation method of the recommendation model according to any embodiment of the present invention.

[0032] According to another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement the training method or creative recommendation method of the recommendation model described in any embodiment of the present invention.

[0033] According to another embodiment of the present invention, a computer program product is provided, including a computer program that, when executed by a processor, implements the training method or creative recommendation method of the recommendation model described in any embodiment of the present invention.

[0034] The technical solution of this invention solves the problem of poor expressive power of the creative recommendation model by jointly training the creative recommendation model with a product ranking model that has a higher model complexity than the creative recommendation model. Furthermore, the creative recommendation model uses shared training features extracted from the product ranking model that correspond to the shared training set. While the shared training set still adopts a simple feature system, the recognizability of the shared training set in the creative recommendation model is enhanced, thereby improving the creative recommendation effect of the creative recommendation model.

[0035] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 A flowchart illustrating a training method for a recommendation model provided in one embodiment of the present invention;

[0038] Figure 2 A flowchart illustrating another method for training a recommendation model provided in an embodiment of the present invention;

[0039] Figure 3A flowchart illustrating another method for training a recommendation model provided in an embodiment of the present invention;

[0040] Figure 4 A flowchart illustrating a creative recommendation method provided in one embodiment of the present invention;

[0041] Figure 5 A flowchart illustrating a specific example of a creative recommendation method provided in an embodiment of the present invention;

[0042] Figure 6 A schematic diagram of the structure of a training device for a recommendation model provided in one embodiment of the present invention;

[0043] Figure 7 This is a schematic diagram of the structure of a creative recommendation device provided in one embodiment of the present invention;

[0044] Figure 8 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0046] It should be noted that the terms "first," "second," "initial," "target," "reference," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0047] Figure 1This is a flowchart illustrating a training method for a recommendation model according to an embodiment of the present invention. This embodiment is applicable to training a network model for creative recommendations. The method can be executed by a training device for the recommendation model, which can be implemented in hardware and / or software and can be configured in a terminal device. Figure 1 As shown, the method includes:

[0048] S110. Input the sorting training set into the initial product sorting model, and input the creative training set into the initial creative recommendation model.

[0049] Specifically, the ranking training set represents the set of parameter features used to train the product ranking model. In one specific embodiment, the ranking training set includes product feature sets corresponding to multiple products. For example, the product feature set includes, but is not limited to, at least one product feature such as product identifier, product value attribute value, product type, product browsing history features, product review history features, store information, brand information, and product keywords.

[0050] In another specific embodiment, the ranking training set also includes a user feature set. For example, the user feature set includes, but is not limited to, at least one user feature such as user identifier, user profile features, statistical features of user behavior, color preference, and promotional information preference, wherein user behavior includes, but is not limited to, click behavior, browsing behavior, add-to-cart behavior, favorite behavior, and purchase behavior, etc.

[0051] In one specific embodiment, the number of features in the product feature set and user feature set in the ranking training set is on the order of hundreds. The advantage of this setting is that it ensures the feature system of the product ranking model matches the model complexity, thereby guaranteeing the product ranking performance of the product ranking model.

[0052] In another specific embodiment, the ranking training set also includes a context feature set and / or a subject feature set of the entity to which each product belongs. For example, the context feature set includes, but is not limited to, the product's recommendation time, the e-commerce platform to which the product belongs, the user's terminal device information, and the user's geographical location information, etc., while the subject feature set includes, but is not limited to, the entity's registration address, shipping address, value attribute information, and entity level information, etc.

[0053] The training set for ranking is not limited here; it can be customized according to actual needs. For example, the training set for ranking can be obtained from an online log system.

[0054] Specifically, the creative training set represents the set of creative-related features used to train the creative recommendation model. The creative training set contains multiple creative feature sets corresponding to different creatives. For example, the creative feature set includes, but is not limited to, at least one creative feature such as creative identifier, creative tag, layout information, image information, copywriting information, and promotional information, but is not limited to the example scenario.

[0055] In this embodiment, the initial product ranking model includes a ranking feature network and a ranking output network, and the initial creative recommendation model includes a creative feature network and a creative output network. The model complexity of the initial product ranking model is higher than that of the initial creative recommendation model.

[0056] For example, the higher model complexity of the initial product ranking model compared to the initial creative recommendation model can be reflected in the higher backpropagation computation cost and / or the greater number of model parameters in the initial product ranking model. Here, backpropagation computation cost represents the computational power required for the network model to perform one backpropagation, and the number of model parameters indicates the number of model parameters that the network model needs to learn.

[0057] S120. Through the ranking feature network, the ranking training set is used to extract features to obtain ranking training features, and the ranking training features are input into the ranking output network to obtain the output predicted product ranking result.

[0058] For example, the initial product recommendation model consists of at least one neural network, including but not limited to Multilayer Perceptron (MLP), attention mechanism network, convolutional neural network, recurrent neural network, and Deep & Cross Network (DCN). The MLP consists of an input layer, multiple hidden layers, and an output layer, where each layer consists of multiple neurons, and each neuron is connected to all neurons in the layer above it.

[0059] Specifically, the ranking training features include feature vectors corresponding to each parameter feature in the ranking training set. The predicted product ranking result includes product weights for at least two products, where each product weight represents the probability value of a product corresponding to a preset user behavior. For example, preset user behaviors include, but are not limited to, click behavior, add-to-cart behavior, purchase behavior, or favorite behavior.

[0060] In one specific embodiment, the product weight is the product click-through rate, which represents the probability that a user will click on the product when faced with it.

[0061] S130. Obtain the shared training features corresponding to the shared training set from the ranked training features, and extract creative training features from the creative training set through the creative feature network. Then, input the shared training features and creative training features into the creative output network to obtain the output predicted creative recommendation results.

[0062] Specifically, the shared training set represents the set of parameters and features used to train the initial product ranking model and the initial creative recommendation model, respectively. In this embodiment, the ranking training set truly includes the shared training set, which contains user features and / or product features.

[0063] In one specific embodiment, the number of features in the shared training set is on the order of ten. The advantage of this setting is that it ensures the feature set of the creative recommendation model matches its model complexity, preventing the model from being unable to fully capture the complex content contained within the shared training features due to an overly rich shared training set. This avoids underfitting issues and thus guarantees the effectiveness of the creative recommendation model.

[0064] For example, the shared training set includes user characteristics such as user identifiers and user profile features, and / or product characteristics such as product identifiers, brand information, and product information. This embodiment does not limit the selection of features for the shared training set; specific settings can be customized according to actual needs.

[0065] In one specific embodiment, the initial creative recommendation model is a multilayer perceptron. In one embodiment, if both the initial creative recommendation model and the initial product ranking model contain only multilayer perceptrons, then the number of neurons and / or the number of layers used in each layer of the initial creative recommendation model is less than that in the initial product ranking model.

[0066] In one specific embodiment, the predicted creative recommendation result includes creative weights corresponding to each product and at least one creative, whereby the creative weight represents the probability value of the product under that creative corresponding to a preset user behavior.

[0067] In one specific embodiment, the creative weight is the creative click-through rate, which represents the probability that a user will click on a product under that creative.

[0068] For example, assuming the ranking training set contains N products and the creative training set contains M creatives, then the predicted creative recommendation result contains M×N creative weights.

[0069] S140. Based on the predicted product ranking results, the initial product ranking model is iteratively trained to obtain the trained target product ranking model, and based on the predicted creative recommendation results, the initial creative recommendation model is iteratively trained to obtain the trained target creative recommendation model.

[0070] In one specific embodiment, the initial product ranking model is iteratively trained to obtain the trained target product ranking model based on the predicted product ranking results, including: determining the ranking loss function value based on the predicted product ranking results and the actual product ranking results; and iteratively adjusting the model parameters of the initial product ranking model based on the ranking loss function value to obtain the trained target product ranking model.

[0071] Specifically, the real product ranking results include a behavior occurrence label for each product and a preset user behavior. The behavior occurrence label indicates whether the preset user behavior actually occurred when the user was looking at the product.

[0072] For example, the loss functions corresponding to the ranking loss function value include, but are not limited to, the squared loss function, the logarithmic loss function, the exponential loss function, the mean squared error loss function, the logistic regression loss function, the Huber loss function, the cross-entropy loss function, and the Kullback-Leibler divergence loss function, etc.

[0073] In one specific embodiment, the loss function corresponding to the ranking loss function value is the cross-entropy loss function. For example, the ranking loss function value CE... sort Satisfy the following formula:

[0074]

[0075] Where N represents the amount of data for the products in the sorted training set, y i p represents the label of the action that occurred corresponding to the i-th product in the actual product ranking result. i This represents the weight of the i-th product in the predicted product ranking results.

[0076] In one specific embodiment, the initial creative recommendation model is iteratively trained to obtain the trained target creative recommendation model based on the predicted creative recommendation results, including: determining the creative loss function value based on the predicted creative recommendation results and the actual creative recommendation results; and iteratively adjusting the model parameters of the initial creative recommendation model based on the creative loss function value to obtain the trained target creative recommendation model.

[0077] Specifically, the real creative recommendation results include the behavior occurrence tags corresponding to the products under each creative and the preset user behavior. The behavior occurrence tags indicate whether the preset user behavior actually occurred when the user was faced with the products under that creative.

[0078] For example, the loss functions corresponding to the creative loss function value include, but are not limited to, the squared loss function, the logarithmic loss function, the exponential loss function, the mean squared error loss function, the logistic regression loss function, the Huber loss function, the cross-entropy loss function, and the Kullback-Leibler divergence loss function, etc.

[0079] The technical solution of this embodiment solves the problem of poor expressive power of the creative recommendation model by jointly training the creative recommendation model with the product ranking model, which has a higher model complexity. Furthermore, the creative recommendation model uses the shared training features extracted by the product ranking model that correspond to the shared training set. While the shared training set still adopts a simple feature system, the recognizability of the shared training set in the creative recommendation model is enhanced, thereby improving the creative recommendation effect of the creative recommendation model.

[0080] Figure 2 This is a flowchart illustrating another training method for a recommendation model provided in one embodiment of the present invention. This embodiment further refines the step of "iteratively training the initial creative recommendation model to obtain a trained target creative recommendation model based on the predicted creative recommendation results" in the above embodiment. In this embodiment, iteratively training the initial creative recommendation model to obtain a trained target creative recommendation model based on the predicted creative recommendation results includes: determining the product creative recommendation results based on the predicted product ranking results and the predicted creative recommendation results; adjusting the model parameters of the initial creative recommendation model based on the product creative recommendation results until the iteration termination condition is met, thus obtaining the trained target creative recommendation model. Figure 2 As shown, the method includes:

[0081] S210. Input the sorting training set into the initial product sorting model, and input the creative training set into the initial creative recommendation model.

[0082] S220. Through the ranking feature network, the ranking training set is used to extract features to obtain ranking training features, and the ranking training features are input into the ranking output network to obtain the output predicted product ranking result.

[0083] S230. Obtain the shared training features corresponding to the shared training set from the ranked training features, and extract creative training features from the creative training set through the creative feature network. Then, input the shared training features and creative training features into the creative output network to obtain the output predicted creative recommendation results.

[0084] S240. Based on the predicted product ranking results, iteratively train the initial product ranking model to obtain the trained target product ranking model.

[0085] S210-S240 in this embodiment are the same as those in the above embodiments. Figure 1 The S110-S140 shown are the same or similar, and will not be described again in this embodiment.

[0086] S250. Based on the predicted product ranking results and the predicted creative recommendation results, determine the product creative recommendation results.

[0087] In this embodiment, the predicted creative recommendation result includes the gain weight generated by each creative in the creative training set on the product weight of each product, and the product creative recommendation result includes the creative weight corresponding to each product and at least one creative.

[0088] Specifically, the gain weight represents the probability value corresponding to the creative and the preset user behavior. In one specific embodiment, the gain weight is the click gain rate, which represents the increase in the probability of a user clicking on a product without the creative, or the probability of a user clicking on the creative.

[0089] Specifically, creative weight represents the probability value of a product associated with a given creative and a preset user behavior. In one specific embodiment, creative weight is the creative click-through rate (CTR), which represents the likelihood of a user clicking on a product associated with that creative.

[0090] Specifically, for each product, the product weight is summed with the gain weight of each creative, and the sum is used as the creative weight of the product under each creative.

[0091] For example, suppose the creative training set contains 3 types of creative ideas, and the gain weights of the 3 types of creative ideas in the predicted creative recommendation results are 0.01, 0.02 and 0.03 respectively. If the weight of the product corresponding to product A in the predicted product ranking results is 0.7, then the creative weights of product A under the 3 types of creative ideas in the product creative recommendation results are 0.71, 0.72 and 0.73 respectively.

[0092] S260. Based on the product creative recommendation results, adjust the model parameters of the initial creative recommendation model until the iteration termination condition is met, and obtain the trained target creative recommendation model.

[0093] In one specific embodiment, the model parameters of the initial creative recommendation model are adjusted based on the product creative recommendation results, including: determining the creative loss function value based on the product creative recommendation results and the actual creative recommendation results; and adjusting the model parameters of the initial creative recommendation model based on the creative loss function value.

[0094] Specifically, the real creative recommendation results include the behavior occurrence tags corresponding to the products under each creative and the preset user behavior. The behavior occurrence tags indicate whether the preset user behavior actually occurred when the user was faced with the products under that creative.

[0095] For example, the loss functions corresponding to the creative loss function value include, but are not limited to, the squared loss function, the logarithmic loss function, the exponential loss function, the mean squared error loss function, the logistic regression loss function, the Huber loss function, the cross-entropy loss function, and the Kullback-Leibler divergence loss function, etc.

[0096] In another specific embodiment, the model parameters of the initial creative recommendation model are adjusted based on the product creative recommendation results, including: determining the creative loss function value based on the product creative recommendation results and the actual creative recommendation results; using the backpropagation algorithm, determining the parameter gradient corresponding to the model parameters of the initial creative recommendation model based on the creative loss function value, and adjusting the model parameters of the initial creative recommendation model based on the parameter gradient.

[0097] In this embodiment, during the adjustment of the model parameters of the initial creative recommendation model, the model parameters of the initial product ranking model are processed using a stopping gradient backpropagation function.

[0098] For example, the creativity loss function value CE creat Satisfy the following formula:

[0099] CE creat =E(x-(Q+stop-gradient(P)))

[0100] Where E represents the loss function, x represents the actual creative recommendation result, Q represents the predicted creative recommendation result, stop-gradient represents the stopping gradient backpropagation function, and P represents the predicted product ranking result.

[0101] Since the product recommendation results are determined based on the predicted product ranking and the predicted product recommendation results, the impact of the product recommendations on the preset user behavior will be backpropagated to the initial product ranking model through the parameter gradient, introducing parameter interference into the initial product ranking model and thus reducing the training effect of joint training. This embodiment sets a stopping gradient backpropagation function so that the parameter gradient only adjusts the model parameters of the initial product ranking model, and not the model parameters of the initial product ranking model, thereby ensuring the effectiveness of the product recommendation model.

[0102] For example, the termination conditions for iterative training include, but are not limited to, at least one of the following: the training sample in the current iterative training process is the last training sample, the number of iterations reaches the threshold, the creativity loss function value converges, and the ranking loss function value converges.

[0103] Creative recommendation models are modeling models based on the triple [user-product-creative], meaning that for a given user, only one creative can typically be displayed for the same product. First, the computational complexity of creative recommendation models is enormous. For example, when there are N products and M creatives, the computational complexity is M×N, M orders of magnitude higher than that of product recommendation models. Second, creative recommendation models suffer from sparse training samples, leading to underfitting. Third, the pre-defined distribution of user behavior among products is significantly higher than the distribution of the same product across different creatives, making it easy for creative recommendation models to learn the pre-defined distribution of user behavior among products, rather than the distribution of user behavior among creatives.

[0104] The technical solution in this embodiment, from the perspective of task separation, models the gain difference between creative ideas and preset user behaviors by setting a creative recommendation model. It approximately separates the traditional triple [user-product-creative] into two binary tuples: [user-product] applied to the product ranking model and [user-creative] applied to the creative recommendation model. This significantly reduces the task difficulty of the creative recommendation model, thereby reducing its computational complexity from the traditional M×N order of magnitude to M order of magnitude, improving the convergence speed of the creative recommendation model, and solving the underfitting problem that is prone to occur in creative recommendation models. Furthermore, since the creative recommendation model in this embodiment is based on the binary tuple [user-creative], it reduces the interference of behavioral distribution differences among products on the creative recommendation model, achieving lower computational complexity while further improving the creative recommendation effect.

[0105] Figure 3 This is a flowchart illustrating another training method for a recommendation model provided in one embodiment of the present invention. This embodiment further refines the selection of the "shared training set" in the above embodiments. Figure 3 As shown, the method includes:

[0106] S310. Input the sorting training set into the initial product sorting model, and input the creative training set into the initial creative recommendation model.

[0107] S320. Through the ranking feature network, the ranking training set is used to extract features to obtain ranking training features, and the ranking training features are input into the ranking output network to obtain the output predicted product ranking result.

[0108] S330. Obtain the shared training features corresponding to the shared training set from the ranked training features, and extract creative training features from the creative training set through the creative feature network. Then, input the shared training features and creative training features into the creative output network to obtain the output predicted creative recommendation results.

[0109] S340. Based on the predicted product ranking results, iteratively train the initial product ranking model to obtain the trained target product ranking model, and based on the predicted creative recommendation results, iteratively train the initial creative recommendation model to obtain the trained target creative recommendation model.

[0110] S310-S340 in this embodiment are the same as those in the above embodiments. Figure 1 S110-S140 shown are the same as or similar to those in the above embodiments. Figure 2 The S210-S260 shown are the same or similar, and will not be described again in this embodiment.

[0111] S350. Obtain the target creative recommendation model trained on at least two shared training sets respectively.

[0112] In this embodiment, the shared features in at least two shared training sets are different, and the shared features are user features or product features. For example, shared training set A contains user identifiers, user profile features, and product identifiers; shared training set B contains user identifiers, user profile features, and brand information; and shared training set C contains user identifiers, user profile features, product identifiers, and product information.

[0113] Specifically, for each shared training set, the target creative recommendation model is trained using S310-S340 in this embodiment based on the shared training set.

[0114] S360. For each target creative recommendation model, determine the model metric parameters of the target creative recommendation model based on the creative test set and the shared test set corresponding to the target creative recommendation model.

[0115] Specifically, the creative test set represents the set of creative-related parameter features used to test the target creative recommendation model, while the shared test set represents the set of user and / or product-related parameter features used to test the target creative recommendation model. The shared features in the shared test set are consistent with the shared features in the shared training set corresponding to the target creative recommendation model.

[0116] In one specific embodiment, determining the model metric parameters of the target creative recommendation model based on the creative test set and the shared test set corresponding to the target creative recommendation model includes: inputting the creative test set and the shared test set corresponding to the target creative recommendation model into the target creative recommendation model to obtain the output target creative recommendation result; obtaining the actual creative response results corresponding to at least two creatives in the creative test set; and determining the model metric parameters of the target creative recommendation model based on the actual creative response results and the target creative recommendation result.

[0117] In this embodiment, the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network. The target feature network is a feature network in the target product ranking model used to extract shared training features from the shared training set. Specifically, the creative test set and the shared test set corresponding to the target creative recommendation model are input into the target creative recommendation model. The creative feature network extracts features from the creative test set to obtain creative test features, and the target feature network extracts features from the shared test set to obtain shared test features. Finally, the creative test features and the shared test features are input into the creative output network to obtain the output target creative recommendation result.

[0118] Specifically, the actual creative response results include behavioral response tags corresponding to each product and at least one creative idea. The behavioral response tags indicate whether the user actually performs the preset user behavior when faced with the product under each creative idea.

[0119] Specifically, model metrics are used to describe the performance of the target creative recommendation model. Examples of model metrics include, but are not limited to, accuracy, recall, precision, F1 score, or AUC value, but are not limited to the example scenario.

[0120] S370. Based on at least two model metric parameters, filter at least two target creative recommendation models to obtain the final target creative recommendation model.

[0121] In one specific embodiment, at least two target creative recommendation models are screened based on at least two model metric parameters to obtain the final target creative recommendation model, including: selecting the target creative recommendation model with the largest model metric parameter as the final target creative recommendation model.

[0122] In another specific embodiment, the final target creative recommendation model is obtained by filtering at least two target creative recommendation models based on at least two model metric parameters. This includes: classifying at least two target creative recommendation models according to their shared training sets to obtain at least one group of creative recommendation models; for each group of creative recommendation models, determining incremental metric parameters based on the model metric parameters corresponding to the two target creative recommendation models in the group; selecting the creative recommendation model group with the largest incremental metric parameter as the target creative recommendation model group, and selecting the target creative recommendation model with the larger model metric parameter in the target creative recommendation model group as the final target creative recommendation model.

[0123] In this embodiment, the shared training sets corresponding to the two target creative recommendation models in the creative recommendation model group satisfy an inclusion relationship. For example, in a creative recommendation model group, assuming that the shared training set corresponding to target creative recommendation model A contains user identifiers, user profile features, and product identifiers, and if the shared training set corresponding to target creative recommendation model B contains user identifiers, user profile features, product identifiers, and brand information, then it means that the shared training set corresponding to target creative recommendation model B contains the shared training set corresponding to target creative recommendation model A, and the feature difference is 1. Here, the feature difference between the two shared training sets corresponding to each creative recommendation model group is not limited.

[0124] Specifically, the incremental metric parameter is the difference between the model metric parameter of the target creative recommendation model corresponding to the shared training set with more shared features and the model metric parameter of the target creative recommendation model corresponding to the shared training set with fewer shared features in the creative recommendation model group. The incremental metric parameter may be positive or negative.

[0125] Specifically, the incremental metric parameter represents the gain effect of differing shared features on the model metric parameters of the target creative recommendation model in the two shared training sets corresponding to the creative recommendation model group. When the incremental metric parameter is positive, it indicates a positive gain effect of the differing shared features on the model metric parameters of the target creative recommendation model; when the incremental metric parameter is negative, it indicates a negative gain effect of the differing shared features on the model metric parameters of the target creative recommendation model.

[0126] Taking the above example, the incremental metric parameter is the difference between the model metric parameter of the target creative recommendation model B and the model metric parameter of the target creative recommendation model A. The incremental metric parameter represents the gain effect of brand information on the model metric parameter of the target creative recommendation model.

[0127] The advantage of comparing gain metric parameters is that it enables the final target creative recommendation model to adapt to different feature environments, thereby improving the overfitting performance of the final target creative recommendation model.

[0128] The technical solution of this embodiment obtains target creative recommendation models trained on at least two shared training sets. For each target creative recommendation model, the model metric parameters of the target creative recommendation model are determined based on the creative test set and the shared test set corresponding to the target creative recommendation model. Based on at least two model metric parameters, the at least two target creative recommendation models are filtered to obtain the final target creative recommendation model. This solves the problem that the shared training set set set by human customization is limited by experience, improves the matching degree between the shared training set set and the creative recommendation model, and further improves the creative recommendation effect of the creative recommendation model.

[0129] Figure 4 This is a flowchart illustrating a creative recommendation method according to an embodiment of the present invention. This embodiment is applicable to situations where multiple creative ideas corresponding to a product are recommended. The method can be executed by a creative recommendation device, which can be implemented in hardware and / or software and can be configured in a terminal device. Figure 4 As shown, the method includes:

[0130] S410. Input the creative application set and the target application set into the pre-trained target creative recommendation model.

[0131] Specifically, the creative application set represents a dataset composed of creative feature sets of multiple candidate creatives. For example, the creative feature set includes, but is not limited to, at least one creative feature such as creative identifier, creative tag, layout information, image information, copywriting information, and promotional information, but is not limited to the example scenario.

[0132] In this embodiment, the target application set includes feature values ​​corresponding to shared features in the shared training set. These shared features are either user features or product features. For example, the target application set includes user features such as the user identifier and user profile features of the target user, and / or product features such as the product identifier, brand information, and product information of the target product. Here, the target user refers to the user who initiates the request in the deployment scenario of the creative recommendation model, and the target product refers to the product waiting to be displayed.

[0133] The target creative recommendation model in this embodiment is obtained by using the training method of the recommendation model provided in any of the above embodiments. The training method of the target recommendation model will not be described again here.

[0134] In this embodiment, the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network. The target feature network is a feature network in the target product ranking model used to extract shared training features from the shared training set.

[0135] S420. Creative application features are obtained by extracting features from the creative application set through the creative feature network, and target application features are obtained by extracting features from the target application set through the target feature network. The creative application features and target application features are then input into the creative output network to obtain the output target creative recommendation results.

[0136] Specifically, the creative application feature is represented by the feature vector corresponding to the creative features in the creative application set, and the target application feature is represented by the feature vector corresponding to the shared features in the target application set.

[0137] Specifically, the target creative recommendation results include the creative weights corresponding to the target product and at least one creative, with the creative weight representing the probability value of the target product under that creative corresponding to the preset user behavior.

[0138] S430. Based on the target creative recommendation results, display the creative product page.

[0139] Specifically, the creative with the highest creative weight in the target creative recommendation results is selected as the target creative, and the creative product page for the target product is displayed based on the target creative. For example, the target creative includes, but is not limited to, layout content, dynamic element content, promotional activity content, and multimedia content. The specific content of the creative is not limited here and can be customized according to actual needs.

[0140] Figure 5 The flowchart illustrates a specific example of a creative recommendation method provided in an embodiment of the present invention. Specifically, the creative recommendation method includes an offline training mode and an online service mode.

[0141] In offline training mode, online log data is collected from the online log system, and training data is filtered based on this data, specifically including a ranking training set and a creative training set. The ranking training set is used as input data for the product ranking model to obtain the output predicted product ranking results. For example, the ranking training set includes rich user features, product features, context features, and subject features. A small number of user features and / or product features from the ranking training set are used as a shared training set to obtain the shared training features extracted by the ranking feature network in the product ranking model, corresponding to the shared training set. The creative training set and the shared training features are used as input data for the creative recommendation model to obtain the output predicted creative recommendation results. Based on the predicted product ranking results and the predicted creative recommendation results, the product ranking model and the creative recommendation model are jointly trained to obtain the trained product ranking model and creative recommendation model. The trained creative recommendation model is then deployed online.

[0142] In online service mode, upon receiving a user request from a target user, the system retrieves the target product corresponding to the target user and multiple candidate creatives for that product. Using an online-deployed creative recommendation model, it selects the target creative that matches the target user and product from the candidate creatives and displays the creative product page generated based on the target creative. Based on the user behavior generated by the target user towards the creative product page, online log data is generated and added to the online log system.

[0143] The technical solution of this embodiment solves the problem of poor expressive power of the creative recommendation model by jointly training it with the product ranking model, which has a higher model complexity. The feature network used to extract shared training features from the product ranking model is deployed as the target feature network in the trained creative recommendation model. While the shared training set still uses a simple feature system, the recognizability of the shared training set in the creative recommendation model is enhanced, thereby improving the creative recommendation effect of the creative recommendation model.

[0144] It should be noted that the collection, use, storage, sharing and transfer of user personal information involved in the technical solution of the present invention all comply with the provisions of relevant laws and regulations, and require notification to users and obtaining their consent or authorization. When applicable, user personal information is subjected to de-identification and / or anonymization and / or encryption technical processing.

[0145] The following are embodiments of the training apparatus for the recommendation model provided in this invention. This apparatus and the training method for the recommendation model in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the training apparatus for the recommendation model, please refer to the content on the training method for the recommendation model in the above embodiments.

[0146] Figure 6 This is a schematic diagram of a training device for a recommendation model provided in one embodiment of the present invention. Figure 6 As shown, the device includes: a sorting training set input module 510, a predicted product sorting result output module 520, a predicted creative recommendation result output module 530, and a creative recommendation model training module 540.

[0147] The sorting training set input module 510 is used to input the sorting training set into the initial product sorting model and the creative training set into the initial creative recommendation model. The initial product sorting model includes a sorting feature network and a sorting output network, and the initial creative recommendation model includes a creative feature network and a creative output network.

[0148] The predicted product ranking result output module 520 is used to extract features from the ranking training set through the ranking feature network to obtain ranking training features, and input the ranking training features into the ranking output network to obtain the output predicted product ranking result.

[0149] The predictive creative recommendation result output module 530 is used to obtain the shared training features corresponding to the shared training set in the ranking training features, and to extract creative training features from the creative training set through the creative feature network. The shared training features and creative training features are then input into the creative output network to obtain the output predicted creative recommendation result.

[0150] The creative recommendation model training module 540 is used to iteratively train the initial product ranking model to obtain the trained target product ranking model based on the predicted product ranking results, and to iteratively train the initial creative recommendation model to obtain the trained target creative recommendation model based on the predicted creative recommendation results.

[0151] The ranking training set truly includes a shared training set, which contains user features and / or product features. The model complexity of the initial product ranking model is higher than that of the initial creative recommendation model.

[0152] The technical solution of this embodiment solves the problem of poor expressive power of the creative recommendation model by jointly training the creative recommendation model with the product ranking model, which has a higher model complexity. Furthermore, the creative recommendation model uses the shared training features extracted by the product ranking model that correspond to the shared training set. While the shared training set still adopts a simple feature system, the recognizability of the shared training set in the creative recommendation model is enhanced, thereby improving the creative recommendation effect of the creative recommendation model.

[0153] In one specific embodiment, the creative recommendation model training module 540 includes:

[0154] The product creative recommendation result determination unit is used to determine the product creative recommendation result based on the predicted product ranking result and the predicted creative recommendation result;

[0155] The target creative recommendation model determination unit is used to adjust the model parameters of the initial creative recommendation model based on the product creative recommendation results until the iteration termination condition is met, thus obtaining the trained target creative recommendation model.

[0156] The predicted product ranking results include the product weights corresponding to at least two products, the predicted creative recommendation results include the gain weight generated by each creative in the creative training set on the product weight of each product, and the product creative recommendation results include the creative weights corresponding to each product and at least one creative.

[0157] In one specific embodiment, the target creative recommendation model determination unit is specifically used for:

[0158] The creative loss function value is determined based on the product creative recommendation results and the actual creative recommendation results;

[0159] The backpropagation algorithm is used to determine the parameter gradients corresponding to the model parameters of the initial creative recommendation model based on the creative loss function value, and the model parameters of the initial creative recommendation model are adjusted according to the parameter gradients.

[0160] In the process of adjusting the model parameters of the initial creative recommendation model, the model parameters of the initial product ranking model are processed using a stopping gradient backpropagation function.

[0161] In one specific embodiment, the device further includes:

[0162] The target creative recommendation model acquisition module is used to acquire target creative recommendation models trained on at least two shared training sets respectively; wherein, the shared features in the at least two shared training sets are different, and the shared features are user features or product features;

[0163] The model metric parameter determination module is used to determine the model metric parameters of each target creative recommendation model based on the creative test set and the shared test set corresponding to the target creative recommendation model.

[0164] The target creative recommendation model filtering module is used to filter at least two target creative recommendation models based on at least two model metric parameters to obtain the final target creative recommendation model.

[0165] In one specific embodiment, the model metric parameter determination module is specifically used for:

[0166] Input the creative test set and the shared test set corresponding to the target creative recommendation model into the target creative recommendation model to obtain the output target creative recommendation result;

[0167] Obtain the actual creative response results for at least two creatives in the creative test set;

[0168] Based on the actual creative response results and the target creative recommendation results, determine the model measurement parameters of the target creative recommendation model.

[0169] In one specific embodiment, the target creative recommendation model filtering module is specifically used for:

[0170] Based on the shared training sets corresponding to at least two target creative recommendation models, classify the at least two target creative recommendation models to obtain at least one group of creative recommendation models; wherein the shared training sets corresponding to the two target creative recommendation models in the creative recommendation model group satisfy the inclusion relationship.

[0171] For each creative recommendation model group, the incremental metric parameters are determined based on the model metric parameters corresponding to the two target creative recommendation models in the creative recommendation model group.

[0172] The creative recommendation model group corresponding to the largest incremental metric parameter is taken as the target creative recommendation model group, and the target creative recommendation model corresponding to the larger model metric parameter in the target creative recommendation model group is taken as the final target creative recommendation model.

[0173] The creative recommendation device provided in the embodiments of the present invention can execute the creative recommendation method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0174] Figure 7 This is a schematic diagram of the structure of a creative recommendation device provided in one embodiment of the present invention. Figure 7 As shown, the device includes: a creative application set input module 610, a target creative recommendation result output module 620, and a creative product page display module 630.

[0175] The creative application set input module 610 is used to input the creative application set and the target application set into the pre-trained target creative recommendation model; the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network.

[0176] The target creative recommendation result output module 620 is used by the creative application feature extraction module to extract creative application features from the creative application set through the creative feature network, extract target application features from the target application set through the target feature network, and input the creative application features and target application features into the creative output network to obtain the output target creative recommendation result.

[0177] The creative product page display module 630 is used to display the creative product page based on the target creative recommendation results;

[0178] The target application set includes feature values ​​corresponding to shared features in the shared training set. The shared features are user features or product features. The target creative recommendation model is obtained by training the recommendation model using any embodiment of the present invention. The target feature network is the feature network in the target product ranking model used to extract the shared training features from the shared training set.

[0179] The technical solution of this embodiment solves the problem of poor expressive power of the creative recommendation model by jointly training it with the product ranking model, which has a higher model complexity. The feature network used to extract shared training features from the product ranking model is deployed as the target feature network in the trained creative recommendation model. While the shared training set still uses a simple feature system, the recognizability of the shared training set in the creative recommendation model is enhanced, thereby improving the creative recommendation effect of the creative recommendation model.

[0180] The creative recommendation device provided in the embodiments of the present invention can execute the creative recommendation method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0181] Figure 8 This is a schematic diagram of an electronic device provided according to one embodiment of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0182] like Figure 8As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor 11. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0183] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information or data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0184] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the training method for the recommendation model or the creative recommendation method provided in the above embodiments.

[0185] In some embodiments, the training method or creative recommendation method of the recommendation model provided in the above embodiments can be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the training method or creative recommendation method of the recommendation model described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the training method or creative recommendation method of the recommendation model by any other suitable means (e.g., by means of firmware).

[0186] Various embodiments of the systems and techniques described above herein can be implemented in the following systems or combinations thereof: digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0187] Computer programs used for training methods or creative recommendation methods to implement the recommendation models of this invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0188] In the context of this application, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable storage medium. Examples of machine-readable storage media include, based on an electrical connection of at least one wire, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0189] To provide interaction with a user, the systems and techniques described herein can be implemented on a terminal device having: a display device for displaying information to the user (e.g., a cathode-ray tube (CRT) or liquid crystal display (LCD) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the terminal device. Other types of devices can also provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0190] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0191] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0192] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0193] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for training a recommendation model, characterized in that, include: The ranking training set is input into the initial product ranking model, and the creative training set is input into the initial creative recommendation model; wherein, the initial product ranking model includes a ranking feature network and a ranking output network, and the initial creative recommendation model includes a creative feature network and a creative output network; The ranking feature network extracts features from the ranking training set to obtain ranking training features, and inputs the ranking training features into the ranking output network to obtain the output predicted product ranking result. The shared training features corresponding to the shared training set are obtained from the ranked training features, and the creative training features are extracted from the creative training set through the creative feature network to obtain creative training features. The shared training features and the creative training features are then input into the creative output network to obtain the output predicted creative recommendation results. Based on the predicted product ranking results, the initial product ranking model is iteratively trained to obtain the trained target product ranking model, and based on the predicted creative recommendation results, the initial creative recommendation model is iteratively trained to obtain the trained target creative recommendation model. The ranking training set truly includes the shared training set, which contains user features and / or product features. The model complexity of the initial product ranking model is higher than that of the initial creative recommendation model.

2. The method according to claim 1, characterized in that, The step of iteratively training the initial creative recommendation model based on the predicted creative recommendation results to obtain the trained target creative recommendation model includes: Based on the predicted product ranking results and the predicted creative recommendation results, the product creative recommendation results are determined; Based on the product creative recommendation results, the model parameters of the initial creative recommendation model are adjusted until the iteration termination condition is met, thus obtaining the trained target creative recommendation model. The predicted product ranking result includes product weights corresponding to at least two products, the predicted creative recommendation result includes the gain weight generated by each creative in the creative training set on the product weight of each product, and the product creative recommendation result includes the creative weights corresponding to each product and at least one creative.

3. The method according to claim 2, characterized in that, The step of adjusting the model parameters of the initial creative recommendation model based on the product creative recommendation results includes: Based on the product creative recommendation results and the actual creative recommendation results, determine the creative loss function value; Using the backpropagation algorithm, the parameter gradients corresponding to the model parameters of the initial creative recommendation model are determined based on the creative loss function value, and the model parameters of the initial creative recommendation model are adjusted based on the parameter gradients. Specifically, during the adjustment of the model parameters of the initial creative recommendation model, the model parameters of the initial product ranking model are processed using a stopping gradient backpropagation function.

4. The method according to claim 1, characterized in that, The method further includes: Obtain a target creative recommendation model trained on at least two shared training sets respectively; wherein the shared features in the at least two shared training sets are different, and the shared features are user features or product features; For each target creative recommendation model, the model metric parameters of the target creative recommendation model are determined based on the creative test set and the shared test set corresponding to the target creative recommendation model. Based on at least two model metric parameters, at least two target creative recommendation models are selected to obtain the final target creative recommendation model.

5. The method according to claim 4, characterized in that, The step of determining the model metric parameters of the target creative recommendation model based on the creative test set and the shared test set corresponding to the target creative recommendation model includes: The creative test set and the shared test set corresponding to the target creative recommendation model are input into the target creative recommendation model to obtain the output target creative recommendation result; Obtain the actual creative response results corresponding to at least two creatives in the creative test set; Based on the actual creative response results and the target creative recommendation results, the model metric parameters of the target creative recommendation model are determined.

6. The method according to claim 4, characterized in that, The step of selecting the final target creative recommendation model by filtering at least two target creative recommendation models based on at least two model metric parameters includes: Based on the shared training sets corresponding to at least two target creative recommendation models, the at least two target creative recommendation models are classified to obtain at least one group of creative recommendation models; wherein, the shared training sets corresponding to the two target creative recommendation models in the creative recommendation model group satisfy an inclusion relationship. For each creative recommendation model group, the incremental metric parameters are determined based on the model metric parameters corresponding to the two target creative recommendation models in the creative recommendation model group. The creative recommendation model group corresponding to the largest incremental metric parameter is taken as the target creative recommendation model group, and the target creative recommendation model corresponding to the larger model metric parameter in the target creative recommendation model group is taken as the final target creative recommendation model.

7. A creative recommendation method, characterized in that, include: The creative application set and the target application set are input into a pre-trained target creative recommendation model; wherein, the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network; Creative application features are obtained by extracting features from the creative application set through the creative feature network, and target application features are obtained by extracting features from the target application set through the target feature network. The creative application features and the target application features are then input into the creative output network to obtain the output target creative recommendation results. Based on the target creative recommendation results, display the creative product page; The target application set includes feature values ​​corresponding to shared features in the shared training set. The shared features are user features or product features. The target creative recommendation model is obtained by training the recommendation model according to any one of claims 1-6. The target feature network is the feature network in the target product ranking model used to extract shared training features from the shared training set.

8. A training device for a recommendation model, characterized in that, include: The sorting training set input module is used to input the sorting training set into the initial product sorting model and the creative training set into the initial creative recommendation model; wherein, the initial product sorting model includes a sorting feature network and a sorting output network, and the initial creative recommendation model includes a creative feature network and a creative output network; The predictive product ranking result output module is used to extract features from the ranking training set through the ranking feature network to obtain ranking training features, and input the ranking training features into the ranking output network to obtain the output predicted product ranking result. The predictive creative recommendation result output module is used to obtain the shared training features corresponding to the shared training set in the ranking training features, extract creative training features from the creative training set through the creative feature network, and input the shared training features and the creative training features into the creative output network to obtain the output predictive creative recommendation result. The creative recommendation model training module is used to iteratively train the initial product ranking model to obtain a trained target product ranking model based on the predicted product ranking results, and to iteratively train the initial creative recommendation model to obtain a trained target creative recommendation model based on the predicted creative recommendation results. The ranking training set truly includes the shared training set, which contains user features and / or product features. The model complexity of the initial product ranking model is higher than that of the initial creative recommendation model.

9. A creative recommendation device, characterized in that, include: The creative application set input module is used to input the creative application set and the target application set into the pre-trained target creative recommendation model; wherein, the target creative recommendation model includes a creative feature network, a target feature network, and a creative output network; The target creative recommendation result output module is used by the creative application feature extraction module to extract creative application features from the creative application set through the creative feature network, extract target application features from the target application set through the target feature network, and input the creative application features and the target application features into the creative output network to obtain the output target creative recommendation result. The creative product page display module is used to display the creative product page based on the target creative recommendation results; The target application set includes feature values ​​corresponding to shared features in the shared training set. The shared features are user features or product features. The target creative recommendation model is obtained by training the recommendation model according to any one of claims 1-6. The target feature network is the feature network in the target product ranking model used to extract shared training features from the shared training set.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the training method of the recommendation model according to any one of claims 1-6 or the creative recommendation method according to claim 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the training method of the recommendation model according to any one of claims 1-6 or the creative recommendation method according to claim 7.

12. A computer program product comprising a computer program that, when executed by a processor, implements the training method of the recommendation model according to any one of claims 1-6 or the creative recommendation method according to claim 7.

Images