Information processing method, device, equipment and storage medium

Abstract

Embodiments of the present disclosure relate to information processing methods, devices, equipment, and storage media. The information processing method includes obtaining attribute information related to a user and / or a target object, the attribute information indicating at least one attribute value of the user and / or the target object; extracting multiple single-granularity feature representations of the attribute information at multiple granularities by performing bucketing operations on the attribute information according to multiple bucketing strategies corresponding to multiple granularities; generating a multi-granularity fused feature representation by merging multiple single-granularity feature representations; and determining the user's interest in the target object based on the multi-granularity fused feature representation. By extracting features of information at multiple granularities, the feature representations of the attribute information related to the user and / or the target object at different granularities can be fused, which can achieve a more accurate assessment of the user's interest in the target object.

Description

Technical Field

[0001] The present disclosure generally relates to the field of recommendation systems, and more particularly to information processing methods, devices, apparatuses, and computer-readable storage media. Background Art

[0002] In many application scenarios, it is necessary to understand the user's interest in a specific object. A typical application scenario is personalized recommendation. The recommendation system determines whether the user is interested in a certain object by analyzing the relevant information between the user and the specific object, and then provides personalized recommendations for the target object to the corresponding user. In the Internet platform, providing users with accurate personalized recommendations can improve the user experience. The application scenarios of the recommendation system include but are not limited to product recommendations, service recommendations, news recommendations, web page recommendations, radio station recommendations, music recommendations, etc. For example, for e-commerce websites, accurate personalized recommendations can improve the efficiency of users in selecting products, thereby increasing the website's revenue. Therefore, in order to provide users with more accurate personalized recommendations, it is expected to be able to more accurately evaluate the user's interest in a specific object. Summary of the invention

[0003] According to some embodiments of the present disclosure, a solution for information processing is provided, in particular, an information processing solution based on multi-granularity feature extraction.

[0004] In a first aspect of the present disclosure, an information processing method is provided. The method includes acquiring attribute information related to a user and / or a target object, the attribute information indicating at least one attribute value of the user and / or the target object; performing a bucketing operation on the attribute information according to multiple bucketing strategies corresponding to multiple granularities, to extract multiple single-granularity feature representations of the attribute information at multiple granularities; generating a multi-granularity fused feature representation by merging multiple single-granularity feature representations; and determining the user's interest in the target object based on the multi-granularity fused feature representation.

[0005] In a second aspect of the present disclosure, a device for information processing is provided. The device includes an attribute information acquisition module, configured to acquire attribute information related to a user and / or a target object, the attribute information indicating at least one attribute value of the user and / or the target object; a single-granularity feature representation extraction module, configured to extract multiple single-granularity feature representations of the attribute information at multiple granularities by performing a bucketing operation on the attribute information according to multiple bucketing strategies corresponding to the multiple granularities; a multi-granularity fusion feature representation generation module, configured to generate a multi-granularity fusion feature representation by merging multiple single-granularity feature representations; and an interest level determination module, configured to determine the user's interest level in the target object based on the multi-granularity fusion feature representation.

[0006] In a third aspect of the present disclosure, an electronic device is provided, comprising a memory and a processor, wherein the memory is used to store computer executable instructions, and the computer executable instructions are executed by the processor to implement the method according to the first aspect of the present disclosure.

[0007] In a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein when the computer-executable instructions are executed by a processor, the processor executes the method according to the first aspect of the present disclosure.

[0008] According to various embodiments of the present disclosure, by extracting features of information at multiple granularities, it is possible to fuse feature representations of attribute information related to the user and / or target object at different granularities. This can achieve a more accurate assessment of the user's interest in the target object, thereby further helping to improve the accuracy and stability of subsequent tasks (such as personalized recommendations). BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The features, advantages and other aspects of the various implementations of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Several implementations of the present disclosure are shown here in an exemplary and non-limiting manner. In the accompanying drawings:

[0010] Figure 1 A schematic diagram showing an exemplary environment in which embodiments of the present disclosure can be implemented;

[0011] Figure 2 A schematic diagram of an interest level prediction module for generating an interest level according to some embodiments of the present disclosure is shown;

[0012] Figure 3 Some embodiments of the present disclosure are shown Figure 2 An example structural block diagram of a multi-granularity fusion feature representation generation module;

[0013] Figure 4 The method for training according to some embodiments of the present disclosure is shown. Figure 2 An example structural block diagram of a model training system for a feature extraction model;

[0014] Figure 5 Some embodiments of the present disclosure are shown Figure 4 Example structure diagram of the forget gate;

[0015] Figure 6 Some embodiments of the present disclosure are shown Figure 2 Example structure diagram of the feature extraction model;

[0016] Figure 7 A flowchart of an information processing method according to some embodiments of the present disclosure is shown;

[0017] Figure 8 A schematic structural block diagram of an apparatus for information processing according to some embodiments of the present disclosure is shown; and

[0018] Fig. 9 A block diagram of a computing device is shown in which one or more embodiments of the present disclosure may be implemented. DETAILED DESCRIPTION

[0019] The preferred implementations of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the preferred implementations of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the implementations set forth herein. On the contrary, these implementations are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0020] As used herein, the term "including" and its variations mean open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "based at least in part on". The terms "an example implementation" and "an implementation" mean "at least one example implementation". The term "another implementation" means "at least one additional implementation". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0021] As used herein, the term "model" can learn the association between the corresponding input and output from the training data, so that after the training is completed, the corresponding output can be generated for a given input. The generation of the model can be based on machine learning technology. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs by using multiple layers of processing units. A neural network model is an example of a model based on deep learning. In this article, "model" may also be referred to as "machine learning model", "learning model", "machine learning network" or "learning network", and these terms are used interchangeably in this article.

[0022] Generally, machine learning can include three stages, namely the training stage, the testing stage, and the use stage (also called the reasoning stage). In the training stage, a given model can be trained using a large amount of training data, and iterated continuously until the model can obtain consistent reasoning from the training data that is similar to the reasoning that human intelligence can make. Through training, the model can be considered to be able to learn the association between input and output (also called input-to-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing stage, the test input is applied to the trained model to test whether the model can provide the correct output, thereby determining the performance of the model. In the use stage, the model can be used to process the actual input based on the parameter values ​​obtained through training to determine the corresponding output.

[0023] As mentioned above, it is desirable to be able to more accurately assess the user's interest in the target object, as this is conducive to providing more accurate personalized recommendations to the user. Due to the continuous development of machine learning and deep learning technologies, it has been proposed that various machine learning models can be used to predict the user's interest in the target object. The proposed machine learning-based recommendation system can be configured to process user and / or object-related attribute information using a pre-trained prediction model to predict the user's interest in the target object. In subsequent recommendation tasks, target objects with a higher degree of interest can be recommended to the user.

[0024] Most current machine learning-based recommendation systems require a single granularity to be determined in advance for bucketing operations, and attribute information is processed based on the single granularity for training prediction models, and the trained prediction models are used to predict the user's interest in the target object. During the use of the model, the information to be processed is also processed at the corresponding granularity and used as model input. In practical applications, trained prediction models often have better prediction accuracy in the early stage of use. However, as time goes by, the distribution of users and target objects may continue to change, causing the accuracy of the prediction model to decay over time. In general, fine-grained bucketing can better capture the current data distribution, obtain better fitting ability, and achieve the current best prediction accuracy (i.e., the memory of the prediction model). However, fine bucketing granularity has the risk of overfitting, resulting in a large decay rate of prediction accuracy over time and low stability of the prediction model. The decay rate of prediction accuracy over time for coarse-grained bucketing is small, and a higher stability of the prediction model (i.e., the generalization of the prediction model) can be obtained, but the current prediction accuracy is low.

[0025] However, in existing recommendation systems, it is difficult to determine the optimal bucket granularity for a certain attribute information so that the prediction model has both good memory and generalization, resulting in poor prediction model performance. Therefore, it is expected to obtain and use a prediction model with both good memory and generalization.

[0026] The embodiment of the present disclosure proposes a scheme for information processing to generate a degree of interest. According to the scheme, a bucketing operation is performed on attribute information related to a user and / or a target object according to multiple bucketing strategies corresponding to multiple granularities, thereby extracting multiple single-granularity feature representations of the attribute information at multiple granularities. Multiple single-granularity feature representations are merged into a multi-granularity fused feature representation. Based on the multi-granularity fused feature representation, the user's degree of interest in the target object is determined.

[0027] Thus, the user's interest in the target object is determined according to multiple bucketing strategies corresponding to different granularities. In this way, in this solution, the feature representations of attribute information related to the user and / or target object at different granularities can be integrated, and the real-time changes in the distribution of attribute information can be adaptively processed, so that better memory and stability (generalization) of the prediction of the interest level can be obtained at the same time.

[0028] Figure 1 1 shows a schematic diagram of an exemplary environment 100 in which embodiments of the present disclosure can be implemented. Figure 1 As shown, the example environment 100 includes a computing device or system 108 (for ease of discussion, the computing device 108 is used uniformly below), which includes an interest level prediction module 110. The interest level prediction module 110 is configured to generate a user 104's interest level 112 for the target object 102 based on the attribute information 106 of the target object 102 and / or the user 104. The interest level 112 indicates to what extent the user 104 is interested in the target object 102. The possible interest level of the user 104 for the target object 102 can be divided into multiple levels, or can be in a preset value interval. The interest level prediction module 110 can be configured to predict the interest level 112 from the interest levels of multiple levels or within a preset value interval. This prediction result, i.e., the interest level 112, can be used in subsequent tasks related to the user 104 and / or the target object 102.

[0029] In some embodiments, the computing device 108 further includes a recommendation information providing module 114, which is configured to determine whether to recommend the target object 102 to the user 104 based on the interest level 112. Specifically, the recommendation information providing module 114 can compare the interest level 112 with a preset threshold, and in response to determining that the interest level 112 exceeds the preset threshold, provide the user 104 with recommendation information 116 about the target object 102.

[0030] According to an exemplary embodiment of the present disclosure, the computing device 108 performs a bucketing operation on the attribute information 106 of the target object 102 and / or the user 104 according to multiple bucketing strategies corresponding to multiple granularities, so as to implement data preprocessing. The attribute information 106 of the target object 102 and / or the user 104 may include specific attribute information that may affect the user 104's interest in the target object 102. The attribute information 106 indicates one or more attribute values ​​corresponding to one or more attributes of the user 104 and / or the target object 102, respectively.

[0031] For example, in the field of e-commerce, the target object 102 may be a commodity; the user 104 may be a consumer. The attribute information 106 may include, for example, the user's age, the user's gender, the region where the user is located, the frequency of the user's website visits, etc. The attribute information 106 may also additionally or alternatively include commodity prices, commodity types, commodity usage restrictions, etc. In some examples, the attribute information 106 may only relate to the user 104 or only relate to the target object 102.

[0032] The computing device 108 performs bucketing operations on the attribute information 106 of the target object 102 and / or the user 104 according to multiple bucketing strategies corresponding to the multiple granularities for information processing. When the interest level predicted by the interest level prediction module 110 exceeds a preset threshold, indicating that the user may be interested in the product, the recommendation information providing module 114 provides the user with recommendation information (e.g., links) about the product for the user to browse and purchase.

[0033] The computing device 108 can be a terminal device or a server device. The terminal device can be, for example, various portable or fixed terminals, such as, tablet computers, desktop computers, notebook computers, mobile phones, general computing devices, etc. The server device can be, for example, a centralized server, a distributed server, a mainframe, an edge computing device, or a cloud, etc. Note that although shown as separate devices, in some cases, the functions of the computing device 108 can be implemented by a single physical device. That is, a single physical device can be configured to perform the training and use of the model. The embodiments of the present disclosure are not limited in this respect.

[0034] Some example embodiments of the present disclosure will be described below with continued reference to the accompanying drawings.

[0035] Figure 2 FIG. 1 is a schematic diagram of an interest level prediction module 110 for determining interest level according to some embodiments of the present disclosure. Figure 1 To discuss the determination of interest, Figure 2 It is shown in Figure 1 The interest level prediction module 110 is configured to generate an interest level 112 of the user 104 for the target object 102 .

[0036] The interest level prediction module 110 for generating the interest level can be implemented in a single computing device or multiple computing devices. For example, the interest level prediction module 110 can be implemented by Figure 1 The interest level prediction module 110 is implemented by the computing device 108 shown. Alternatively, the interest level prediction module 110 can also be implemented by other entities besides the computing device 108. The interest level prediction module 110 can be implemented on a device different from the device for training the model. Of course, in some cases, model training and model use can also be implemented on the same device or the same group of devices. This depends on the needs of actual computing resource deployment.

[0037] The interest level prediction module 110 is configured to obtain attribute information 106 related to the user 104 and / or the target object 102, where the attribute information 106 indicates at least one attribute value of the user 104 and / or the target object 102. The interest level prediction module 110 performs a bucketing operation on the attribute information 106 according to a plurality of bucketing strategies 210 corresponding to a plurality of granularities, to extract a plurality of single-granularity feature representations q1, q2, ..., q K .

[0038] More specifically, the interest level prediction module 110 is configured to obtain attribute information 106 related to the user 104 and / or the target object 102, wherein the attribute information 106 indicates F attribute values ​​of the user 104 and / or the target object 102, where F is a positive integer. Each attribute value indicates a specific value of an attribute.

[0039] Through the bucketing operation, the computing device 108 can convert the initial attribute information into discrete data in digital form to achieve data preprocessing. The number of multiple granularities used for the bucketing strategy and the specific granularity division can be set as needed. Figure 2As shown, the interest level prediction module 110 is configured with K bucketing strategies 210-1, 210-2, ... 210-K corresponding to the K granularities, where K is an integer greater than 1. For ease of discussion, the bucketing strategies 210-1, 210-2, ... 210-K may be collectively or individually referred to as bucketing strategies 210. The interest level prediction module 110 performs bucketing operations on the attribute information 106 according to the K bucketing strategies, respectively, to convert the initial attribute information 106 into discrete data in digital form.

[0040] As an example, the computing device 108 may use three bucketing strategies (i.e., K=3) to process the attribute information 106, and the three bucketing strategies correspond to three granularities: coarse granularity, medium granularity, and fine granularity. Taking the commodity price as an example, the bucketing strategy corresponding to the coarse granularity is: the corresponding bucket number is 1 for not more than 500 yuan; the corresponding bucket number is 2 for more than 500 yuan and not more than 1,000 yuan; the corresponding bucket number is 3 for more than 1,000 yuan and not more than 3,000 yuan; ... The bucketing strategy corresponding to the medium granularity is: the corresponding bucket number is 1 for not more than 100 yuan; the corresponding bucket number is 2 for more than 100 yuan and not more than 200 yuan; the corresponding bucket number is 3 for more than 200 yuan and not more than 500 yuan; ... The bucketing strategy corresponding to the fine granularity is: the corresponding bucket number is 1 for the amount not higher than 10 yuan; the corresponding bucket number is 2 for the amount higher than 10 yuan and not higher than 20 yuan; the corresponding bucket number is 3 for the amount higher than 20 yuan and not higher than 50 yuan; the corresponding bucket number is 4 for the amount higher than 50 yuan and not higher than 100 yuan; the corresponding bucket number is 5 for the amount higher than 100 yuan and not higher than 150 yuan; ... It should be understood that the above bucketing strategy is only exemplary, and other bucketing strategies may also be adopted, and the computing device may have two or more than three bucketing strategies corresponding to different granularities.

[0041] Assume that the commodity price of the target object 102 is 125 yuan, so the attribute information 106 also indicates that the attribute value of "commodity price" is 125 yuan. In the three bucketing strategies corresponding to coarse granularity, medium granularity and fine granularity, the bucket numbers corresponding to the attribute value (i.e., 125 yuan) are 1, 2 and 5 respectively. The interest level prediction module 110 converts the attribute values ​​of the commodity price in the attribute information 106 into discrete data 1, 2 and 5 respectively by performing bucketing operations according to the bucketing strategies corresponding to coarse granularity, medium granularity and fine granularity.

[0042] It should be understood that if the attribute information 106 indicates multiple attribute values, the bucketing operation can be performed at the same number of granularities for each attribute value. For different attributes, the bucketing strategies corresponding to each granularity can be set accordingly. For example, if the attribute information 106 also indicates the age of the user 104, similar to the price of the product, three bucketing strategies for age (e.g., coarse granularity, medium granularity, and fine granularity) can be set to convert the specific age value of the user 104 into discrete data.

[0043] After the bucketing operation of the attribute information 106, the interest level prediction module 110 can generate K embedding representations X1, X2, . . . , X corresponding to the K granularities by performing an embedding operation on the discrete data in digital form. K ,in D is a preset dimension. Specifically, the discrete data corresponding to each attribute value is converted into a D-dimensional vector (also called an "embedding vector"). When F is greater than 1, the F D-dimensional embedding vectors form an F×D-dimensional embedding matrix. For ease of discussion, the single D-dimensional embedding vector when F is equal to 1 and the F×D-dimensional embedding matrix when F is greater than 1 can be collectively or individually referred to as an embedding representation and denoted by X. k To represent, where k∈{1,...,K}.

[0044] The interest level prediction module 110 uses K feature extraction models 220-1, 220-2, ..., 220-K corresponding to the K granularities to extract K embedding representations X1, X2, ..., X K Extract K corresponding single-granularity features q1, q2, ..., q K For ease of discussion, feature extraction models 220 - 1 , 220 - 2 , . . . 220 -K may be collectively or individually referred to as feature extraction models 220 .

[0045] The interest level prediction module 110 includes a multi-granularity fusion feature representation generation module 230, which is configured to generate a multi-granularity fusion feature representation by merging multiple single-granularity feature representations q1, q2, . . . , q K The interest level prediction module 110 is configured to determine the interest level 112 of the user 104 for the target object 102 using the interest level determination model 240 based on the multi-granularity fused feature representation q.

[0046] The interest level determination model 240 may be configured as a machine learning model or a deep learning model. By training the interest level determination model 240 with training data, the interest level determination model 240 may learn how to predict the accurate interest level from the multi-granularity fusion feature representation. The interest level determination model 240 may be trained using a variety of model training techniques. In some embodiments, the interest level determination model 240 may be trained by designing a loss function. In the training phase, the interest level determination model 240 may be trained by minimizing a predetermined interest level loss function L C To train the interest level determination model 240. The interest level loss function L C The interest level is determined based on the training data of the interest level determination model 240. For example, the training data may include the predicted interest level pred corresponding to the training attribute information for the interest level determination model 240. n and label y n To determine. Interest level loss function L C For example, it can be as follows:

[0047]

[0048] Where N is the number of samples in the training data set, that is, each sample indicates the corresponding training attribute information; pred n represents the predicted interest level based on the nth training attribute information; and the label y n Indicates the actual interest level of the training user in the training target object associated with the training attribute information.

[0049] In order to reduce the prediction interest level pred n As close as possible to label y n The interest level prediction module 110 based on the loss function may need to update the parameter value so that the interest level loss function L c Gets smaller and smaller until the convergence condition is reached. For example, in the interest level loss function L based on formula (1) C To meet the training objectives, in order to reduce the loss, it is necessary to adjust the embedding vector corresponding to the discrete data converted from the attribute information, the parameters of the feature extraction model 220, and the parameters of the multi-granularity fusion feature representation generation module 230. The convergence condition of the model training may be, for example, minimization of the loss of the loss function, such as being equal to zero or other acceptable values. In the initial stage of training, the embedding vector corresponding to the discrete data converted from the attribute information, the parameters of the feature extraction model 220, and the parameters of the multi-granularity fusion feature representation generation module 230 may have initial parameter values, which may be determined through the pre-training process of the model or in other ways.

[0050] The prediction model based on fine-grained bucketing operation has better memory, but there is a risk of overfitting; while the prediction model based on coarse-grained bucketing operation has better generalization, but the fitting ability is weaker due to information loss. By determining the user's interest in the target object based on multi-granularity fusion feature representation, it is possible to fuse the feature representations of attribute information related to the user and / or target object at different granularities, so that the prediction model has both the generalization of coarse-grained bucketing operation and the memory of fine-grained bucketing operation.

[0051] However, the inventors noticed that for different attribute information distributions, the importance of bucketing operations of different granularities is different. In order to adaptively adjust the prediction model's attention to memorization and generalization according to different attribute information distributions, the present disclosure proposes a multi-granularity fusion model based on a self-attention mechanism. The multi-granularity fusion model is used to model and adjust each pair of single-granularity feature representations. k1 ,q k2 >the relative importance between them, where k1, k2∈{1,...,K}.

[0052] The following will be combined Figure 3 To describe the multi-granularity fusion model based on the self-attention mechanism according to an embodiment of the present disclosure. Figure 3 Some embodiments of the present disclosure are shown Figure 2 FIG. 2 is an example structural block diagram of a multi-granularity fusion feature representation generation module 230 .

[0053] like Figure 3 As shown, the multi-granularity fusion feature representation generation module 230 is based on K single-granularity feature representations q1, q2, ..., q K Determine each pair of single-granular feature representations k1 ,q k2 > The relative importance B between them. The multi-granularity fusion feature representation generation module 230 is configured to merge multiple single-granularity feature representations q1, q2, ..., q through the merging operation module 310 based on the determined relative importance B. K , to generate a multi-granularity fusion feature representation q.

[0054] In one embodiment, assuming A two-layer feed-forward network with ReLU activation function can be used to represent the single-granularity feature q corresponding to the k-th granularity k Projection to the relative importance vector B k , for example as follows:

[0055] B k =W (2) (σ(W (1) q k +b​​(1) )+b (2) ) Formula (2)

[0056] Among them B k The lth dimension is the single-granularity feature representation q k With q l The relative importance of k, l∈{1,...,K}, W (1) , W (2) 、b (1) and b (2) are all learnable parameters.

[0057] The merge operation module operation 310 may perform the merge, for example, in the following manner:

[0058] Q = concat([q1, ..., q K ]) T Formula (3)

[0059] B = concat([B1, ..., B K ]) Formula (4)

[0060] Y=Softmax(B)Q Formula (5)

[0061] q=flatten(Y) Formula (6)

[0062] Among them, O is obtained by merging K single-granularity feature representations, B is obtained by merging K relative importance vectors, and the multi-granularity fusion feature representation q is obtained based on the softmax function and the flatten function.

[0063] By using a multi-granularity fusion model based on the self-attention mechanism, the degree of attention paid by the prediction model to memorization and generalization can be adaptively adjusted according to the changes in the distribution of attribute information.

[0064] Due to the limitation of online deployment on the overall complexity of the model, some embodiments of the present disclosure use a fully connected network to learn the relative importance. Compared with the standard Transformer, the present disclosure reduces the learnable parameters without losing any performance.

[0065] The multi-granularity fusion model aims to fuse the feature representations of attribute information at different granularities, thereby achieving fusion complementarity. However, if the feature representations of attribute information at different granularities are similar, the fusion effect will be greatly reduced. Therefore, it is expected that there is a large difference between the feature representations of attribute information at different granularities. More specifically, it is expected that the sum of the differences between multiple single-granularity feature representations is greater than a predetermined difference threshold. In other words, it is expected that the K corresponding single-granularity feature representations q1, q2, ..., q extracted by the feature extraction model 220 areK The sum of the differences between is greater than a predetermined difference threshold. In order to ensure that the feature extraction model 220 can obtain such a differentiated single-granularity feature representation, an embodiment of the present disclosure provides a model training system for training the feature extraction model 220 by minimizing a predetermined common loss function.

[0066] Figure 4 The method for training according to some embodiments of the present disclosure is shown. Figure 2 For ease of discussion, refer to Figure 2 To discuss the determination of interest, Figure 4 It is shown in Figure 2 A feature extraction model 220 is configured to extract K corresponding single-granularity feature representations from K embedding representations for training, wherein the K embedding representations for training are determined based on training attribute information in a training dataset.

[0067] The model training system 400 can be implemented on a single computing device or multiple computing devices. The model training system 400 can be implemented on a device different from the device (e.g., device 108) using the feature extraction model 220, for example, the model training system 400 can be different from the computing device 108. Of course, in some cases, model training and model use can also be implemented on the same device or the same group of devices. This depends on the actual needs of computing resource deployment.

[0068] In the training phase, the model training system 400 learns the common feature representation based on the K embedding representations. In one embodiment, the model training system 400 is configured to use the DNN structure 410 to learn the common feature representation z L , for example as follows:

[0069] z0=concat([flatten(X1) T , ..., flatten(X K ) T ]) T Formula (7)

[0070] z L =f L (f L-1 (...f1(z0))) Formula (8)

[0071] where embedding represents X1, X2, ..., X K is determined based on the training attribute information in the training data set; z0 is obtained by merging the embedding representations of all K granularities as the joint input of the DNN structure 410; f i(i∈{1, ..., L}) is a fully connected network, and L is the number of layers of the DNN structure. Thus, a common feature representation between multiple single-granularity feature representations can be obtained to determine the differences in multiple feature representations for multiple feature extraction models.

[0072] The model training system 400 is configured to represent z based on the obtained public feature L , using K forget gates 420-1, 420-2, ..., 420-K corresponding to the K granularities, to represent q1, q2, ..., q according to the K single-granularity features. K Determine K corresponding features representing differences e1, e2, ..., e K For the convenience of discussion, forget gates 420 - 1 , 420 - 2 , . . . 420 -K may be collectively or individually referred to as forget gate 420 .

[0073] Considering that the public feature representation contained in each single-granularity feature representation has personalized differences, the present disclosure provides a forget gate 420 based on weighted public feature representation. Figure 5 Let's describe the forget gate 420.

[0074] Figure 5 Some embodiments of the present disclosure are shown Figure 4 For the kth granularity among all K granularities, the forget gate 420 is configured to represent the single granularity feature q for training. k and the common feature representation z L , using MLP k Operation 510 generates a single-granularity feature representation q k and the common feature representation z L The corresponding weight and and Used to determine the single-granularity feature representation q from the dimension level k and the common feature representation z L In one embodiment, and Generated by a single-layer MLP with tanh activation function, for example as follows:

[0075]

[0076]

[0077] Where W k and b kis a learnable parameter. The forget gate 420 can obtain the weighted single-granularity feature representation and the weighted common feature representation through operations 520 and 530, respectively, and then obtain the feature representation difference through operation 540. For example, the feature representation difference e for the k-th granularity k as follows:

[0078]

[0079] where ⊙ represents bitwise multiplication.

[0080] return Figure 4 , the discriminator 430 is used to determine which feature extraction model the input feature representation difference comes from. In one embodiment, the discriminator 430 is configured to calculate the probability that the input feature representation difference comes from each feature extraction model of the K feature extraction models, determine the maximum probability for the K feature extraction models, and determine the feature extraction model corresponding to the maximum probability as the maximum probability feature extraction model of the feature representation difference.

[0081] The model training system 400 is configured to obtain a common loss function L based on the output of the discriminator 430. D , for example as follows:

[0082]

[0083] Where N is the number of samples in the training data set, that is, there are N training attribute information used to train the feature extraction model 220; nk represents the difference in feature representation for the kth granularity based on the nth training attribute information; D(e nk ) represents the feature representation difference e determined based on the calculation of the discriminator 430 nk The maximum probability of the K feature extraction model; c dn represents the maximum probability D(e nk ) corresponds to which feature extraction model. nk When it comes from the kth feature extraction model, that is, when the judgment of the discriminator 430 is correct, c dn is k, is 1; when the discriminator 430 determines e nk When the discriminator 430 makes a wrong judgment, c dn is not equal to k, is 0.

[0084] In order to achieve a large difference between the multiple single-granularity feature representations extracted by the multiple feature extraction models, the model training system 400 may need to update the parameter values ​​so that the common loss function L DGets smaller and smaller until the convergence condition is reached. For example, in the public loss function L based on formula (12) D To meet the training objectives, in order to reduce the loss, it is necessary to adjust the embedding vector corresponding to the discrete data converted from the attribute information, the parameters of the feature extraction model 220, the parameters of the DNN structure 410, and the parameters of the forget gate 420. The convergence condition of the model training can be, for example, the minimization of the loss of the loss function, such as equal to zero or equal to other acceptable values. In the initial stage of training, the embedding vector corresponding to the discrete data converted from the attribute information, the parameters of the feature extraction model 220, the parameters of the DNN structure 410, and the parameters of the forget gate 420 can have initial parameter values, which can be determined through the pre-training process of the model or in other ways.

[0085] Since multiple single-granularity feature representations are obtained based on the same training attribute information, and each feature extraction model tends to prioritize the most direct knowledge and patterns, multiple single-granularity features usually contain some common information. Providing these vectors directly to the discriminator may increase the difficulty of model optimization. By providing feature representation differences to the discriminator, the difficulty of model learning can be reduced.

[0086] By enhancing the differences between multiple single-granularity feature representations, the embodiments of the present disclosure can achieve complementarity between feature representations of attribute information at different granularities, thereby ensuring the effectiveness of the final multi-granularity fusion.

[0087] In some embodiments, the loss function L can be used based on the interest level. C And the common loss function L D A model overall loss function L is defined for jointly training the interest level determination model 240 and the model training system 400. The model overall loss function L is defined as follows:

[0088] L=L C +λL D Formula (13)

[0089] Where λ is a hyperparameter used to indicate the common loss function L D Relative to the interest level loss function L C the relative importance of .

[0090] Since there are multiple feature extraction models 220 in the embodiment of the present disclosure, which perform feature extraction in parallel for multiple granularities, this may increase the computational complexity of the model and increase the response time of the online service. Therefore, it is desirable to reduce the computational complexity of the feature extraction model.

[0091] In the existing xDeepFM, each hidden layer vector in the cross network is the input of layer 0 (i.e., x(0) ). In addition, the bitwise operation of the existing deep cross network (DCN) leads to the calculation of crosses within the same embedding vector. However, this part of the calculation cannot introduce any cross information between different attribute values ​​of the attribute information, thus limiting the computational performance of the feature extraction model.

[0092] The embodiment of the present disclosure provides a feature extraction model, which is configured to use a vector-wise cross network (VDCN) to extract a single-granularity feature representation from an embedded representation based on attribute information. Compared with the existing cross network, the vector-wise cross network can effectively reduce the computational complexity. Figure 6 A VDCN-based feature extraction model according to an embodiment of the present disclosure is described.

[0093] Figure 6 Some embodiments of the present disclosure are shown Figure 2 For ease of discussion, refer to Figure 2 To discuss the extraction of single-granularity feature representation, Figure 6 It is shown in Figure 2 The feature extraction model 220 is configured to extract a single-granularity feature representation from an embedding representation based on attribute information using a VDCN for a given granularity among multiple granularities.

[0094] In one embodiment, the attribute information 106 indicates F attribute values ​​of the user and / or the target object, where F is greater than 1. For the kth granularity among the K granularities, the interest level prediction module 110 is configured to generate F embedding vectors v1, v2, ..., v corresponding to the F attribute values, respectively, by performing a bucketing operation on the attribute information. F , where F embedding vectors v1, v2, ..., v F have the same number of dimensions D, where D is greater than 1. F embedding vectors v1, v2, ..., v F The embedding matrix X is composed of F×D dimensions k , for providing to the feature extraction model 220.

[0095] The feature extraction model 220 is configured to embed the received F×D dimensional matrix X k As the input of the cross part of VDCN to generate cross feature representation. The output of the lth layer of VDCN can be expressed as X (l) ,in l∈{0, 1, ..., L′}, where L is the total number of layers in the cross section of VDCN. The output of each layer of VDCN is as follows:

[0096]

[0097] where f1, f2∈{1, ..., F}, d∈{1, ..., D}, is the weight parameter of the lth layer.

[0098] Specifically, the feature extraction model 220 is configured to extract the received F×D dimensional embedding matrix X k , generate D F-dimensional column vectors x1, x2, ..., x D , D F-dimensional column vectors x1, x2, ..., x D By F embedding vectors v1, v2, ..., v F The feature extraction model 220 is configured to use the cross network 610 to extract D cross feature representations x of the D column vectors. c1 、x c2 ,......,x cD .

[0099] In other words, the feature extraction model 220 uses a set of D parallel sub-cross networks to transform D F-dimensional column vectors x1, x2, ..., x D are input into independent cross networks to obtain cross feature representations for each dimension. For example, the dth column vector x d is input to the dth sub-cross network, and the output is as follows:

[0100]

[0101] The column vectors x are d The output vector, weight parameters and bias parameters of the lth layer of the dth sub-cross network are used as input. For each layer of each sub-cross network, the total number of learnable parameters is F×2; therefore, for a set of D parallel sub-cross networks, the total number of learnable parameters for each layer is D×F×2. Compared with existing DCNs, VDCN does not increase the number of learnable parameters, and therefore does not increase the computational complexity.

[0102] In this way, by learning the interactions in each dimension independently, the output of the cross network is no longer restricted to x (0) In addition, VDCN enables avoiding self-intersection within the same embedding vector, so that the calculation focuses more on the feature intersection of interest and improves computational efficiency.

[0103] The feature extraction model 220 is further configured to extract a joint feature representation h of the F embedding vectors using the fully connected network 620, for example as follows:

[0104] h (0) =concat(v1, ..., v F ) Formula (16)

[0105] h (l+1) = relu(W (l) h (l) +b (l) ) Formula (17)

[0106] where h (l) , W (l) and b (l) They are respectively the output vector, weight parameter and bias parameter of the lth layer of the fully connected network 620.

[0107] The single-granularity feature representation merging generation module 630 combines D cross feature representations x c1 、x c2 ,......,x cD and the joint feature representation h to generate a single-granularity feature representation q at the k-granularity k For example, the single-granularity feature representation merging generation module 630 generates D cross feature representations x c1 、x c2 ,......,x cD The output matrix composed of is flattened into a vector x c , and the single-granularity feature representation q is obtained by the following operation k :

[0108]

[0109] Therefore, compared with the traditional bit-wise DCN, by utilizing the vector-level cross network, the embodiments of the present disclosure can more effectively extract the feature cross of the embedded vector.

[0110] A flowchart of a method and a block diagram of a possible device according to some embodiments of the present disclosure are given below.

[0111] Figure 7 1 shows a flow chart of an information processing method 700 according to some embodiments of the present disclosure. The information processing method 700 is used to determine the user's interest in a target object. In some embodiments, the method 700 may be implemented in Figure 1 For ease of discussion, the method 700 is described from the perspective of the interest level prediction module 110.

[0112] In box 710, the interest level prediction module 110 obtains attribute information related to the user and / or the target object, and the attribute information indicates at least one attribute value of the user and / or the target object. In box 710, the interest level prediction module 110 performs a bucketing operation on the attribute information according to multiple bucketing strategies corresponding to the multiple granularities to extract multiple single-granularity feature representations of the attribute information at multiple granularities. In box 710, the interest level prediction module 110 generates a multi-granularity fused feature representation by merging multiple single-granularity feature representations. In box 710, the interest level prediction module 110 determines the user's interest level in the target object based on the multi-granularity fused feature representation.

[0113] In some embodiments, generating a multi-granularity fused feature representation includes: determining a relative importance between each pair of single-granularity feature representations in a plurality of single-granularity feature representations; and merging the plurality of single-granularity feature representations based on the determined relative importance to generate a multi-granularity fused feature representation.

[0114] In some embodiments, extracting multiple single-granularity feature representations includes: performing bucketing operations on attribute information according to multiple bucketing strategies to generate multiple embedding representations corresponding to multiple granularities; and extracting multiple single-granularity feature representations from multiple embedding representations using multiple feature extraction models corresponding to the multiple granularities, so that the sum of the differences between the multiple single-granularity feature representations is greater than a predetermined difference threshold.

[0115] In some embodiments, multiple feature extraction models are trained by minimizing a predetermined common loss function, wherein the common loss function is determined based on multiple feature representation differences between multiple training single-granularity feature representations for the multiple feature extraction models and multiple weighted common feature representations, and the multiple weighted common feature representations are determined by utilizing multiple weights corresponding to each of the multiple granularities to perform common feature representations, and the common feature representations are determined based on multiple embedding representations.

[0116] In some embodiments, the attribute information indicates multiple attribute values ​​of a user and / or a target object; wherein extracting multiple single-granularity feature representations includes: for a given granularity among multiple granularities, generating multiple embedding vectors corresponding to the multiple attribute values ​​respectively by performing a bucketing operation on the attribute information, wherein the multiple embedding vectors have the same number of dimensions; generating multiple column vectors composed of vector elements of corresponding dimensions in multiple dimensions in the multiple embedding vectors; using a deep cross network (DCN) to extract multiple cross feature representations of each of the multiple column vectors; and generating a single-granularity feature representation at a given granularity by at least merging multiple cross feature representations.

[0117] In some embodiments, generating a single-granularity feature representation at a given granularity includes: extracting a joint feature representation of multiple embedding vectors; and generating a single-granularity feature representation at the given granularity by merging multiple cross feature representations and the joint feature representation.

[0118] In some embodiments, determining a user's interest level in a target object includes: determining the user's interest level in the target object using an interest level determination model, wherein the interest level determination model is trained by minimizing a predetermined interest level loss function, and the interest level loss function is determined during training based on a training interest level and a label corresponding to training attribute information for the interest level determination model, and the label indicates the actual interest level of the training user associated with the training attribute information in the training target object.

[0119] In some embodiments, the method 700 may further include comparing the interest level with a preset threshold; and if it is determined that the interest level exceeds the preset threshold, providing recommendation information about the target object to the user through the recommendation information providing module 114 .

[0120] Figure 8 The schematic structural block diagram of the apparatus 800 for information processing according to some embodiments of the present disclosure is shown. The apparatus 800 for information processing may be implemented as or included in the computing device 108.

[0121] As shown in the figure, the device 800 for information processing includes an attribute information acquisition module 810, which is configured to acquire attribute information related to the user and / or the target object, and the attribute information indicates at least one attribute value of the user and / or the target object. The device 800 for information processing also includes a single-granularity feature representation extraction module 820, which is configured to extract multiple single-granularity feature representations of the attribute information at multiple granularities by performing bucket operations on the attribute information according to multiple bucket strategies corresponding to the multiple granularities. The device 800 for information processing also includes a multi-granularity fusion feature representation generation module 830, which is configured to generate a multi-granularity fusion feature representation by merging multiple single-granularity feature representations. The device 800 for information processing also includes an interest level determination module 840, which is configured to determine the user's interest level in the target object based on the multi-granularity fusion feature representation.

[0122] In one embodiment, the multi-granularity fused feature representation generation module 830 includes: a relative importance determination module, configured to determine the relative importance between each pair of single-granularity feature representations in multiple single-granularity feature representations; and a multi-granularity fused feature representation generation module, configured to merge multiple single-granularity feature representations based on the determined relative importance to generate a multi-granularity fused feature representation.

[0123] In one embodiment, the single-granularity feature representation extraction module 820 includes: an embedding representation generation module, configured to perform bucketing operations on attribute information according to multiple bucketing strategies to generate multiple embedding representations corresponding to multiple granularities; and an extraction module based on a feature extraction model, configured to use multiple feature extraction models corresponding to multiple granularities to extract multiple single-granularity feature representations from multiple embedding representations, so that the sum of the differences between the multiple single-granularity feature representations is greater than a predetermined difference threshold.

[0124] In one embodiment, multiple feature extraction models are trained by minimizing a predetermined common loss function, wherein the common loss function is determined based on multiple feature representation differences between multiple training single-granularity feature representations and multiple weighted common feature representations for the multiple feature extraction models. In one embodiment, multiple weighted common feature representations are determined by using multiple weights corresponding to each of the multiple granularities to determine the common feature representation, and the common feature representation is determined based on multiple embedding representations.

[0125] In one embodiment, the attribute information indicates multiple attribute values ​​of the user and / or the target object; wherein the single-granularity feature representation extraction module 820 includes: an embedding vector generation module, configured to generate multiple embedding vectors corresponding to multiple attribute values ​​respectively for a given granularity among multiple granularities by performing a bucketing operation on the attribute information, wherein the multiple embedding vectors have the same number of dimensions; a column vector generation module, configured to generate multiple column vectors composed of vector elements of corresponding dimensions in multiple dimensions in the multiple embedding vectors for a given granularity among multiple granularities; a cross feature representation extraction module, configured to extract multiple cross feature representations of each of the multiple column vectors for a given granularity among multiple granularities using a deep cross network (DCN); and a single-granularity feature representation generation module, configured to generate a single-granularity feature representation at a given granularity for a given granularity among multiple granularities, at least by merging multiple cross feature representations.

[0126] In one embodiment, the single-granularity feature representation generation module includes: a joint feature representation extraction module configured to extract a joint feature representation of multiple embedding vectors; and a single-granularity feature representation merging generation module configured to generate a single-granularity feature representation at a given granularity by merging multiple cross feature representations and joint feature representations.

[0127] In one embodiment, the interest level determination module is further configured to determine the user's interest level in the target object using an interest level determination model. In one embodiment, the interest level determination model is determined based on the training interest level corresponding to the training attribute information and an associated label, the label indicating the actual interest level of the training user associated with the training attribute information in the training target object.

[0128] In one embodiment, the apparatus 800 for information processing further includes a comparison module configured to compare the interest level with a preset threshold; and a recommendation information providing module configured to provide the user with recommendation information about the target object in response to determining that the interest level exceeds the preset threshold.

[0129] Fig. 9 900 is a block diagram of a computing device / server 900 in which one or more embodiments of the present disclosure may be implemented. It should be understood that Fig. 9 The computing device / server 900 shown is merely exemplary and should not constitute any limitation on the functionality and scope of the embodiments described herein. Fig. 9 The computing device / server 900 shown may be used to implement Figure 1 The computing device 108 or Figure 8 The device 800 of the embodiment of the present invention may be used to partially implement one or more modules of the computing device 108 or the device 800. The computing device / server 900 may also be used to implement Figure 4 system 400 or a portion of system 400.

[0130] like Fig. 9 As shown, computing device / server 900 is in the form of a general computing device. Components of computing device / server 900 may include, but are not limited to, one or more processors or processing units 910, memory 920, storage device 930, one or more communication units 940, one or more input devices 950, and one or more output devices 960. Processing unit 910 may be an actual or virtual processor and is capable of performing various processes according to a program stored in memory 920. In a multi-processor system, multiple processing units execute computer executable instructions in parallel to increase the parallel processing capabilities of computing device / server 900.

[0131] The computing device / server 900 typically includes a plurality of computer storage media. Such media may be any available media accessible to the computing device / server 900, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 920 may be a volatile memory (e.g., registers, caches, random access memory (RAM)), a non-volatile memory (e.g., a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), flash memory), or some combination thereof. The storage device 930 may be a removable or non-removable medium, and may include a machine-readable medium, such as a flash drive, a disk, or any other medium, which may be capable of being used to store information and / or data (e.g., training data for training) and may be accessed within the computing device / server 900.

[0132] The computing device / server 900 may further include additional removable / non-removable, volatile / non-volatile storage media. Fig. 9 As shown in , a disk drive for reading or writing from a removable, non-volatile disk (e.g., a "floppy disk") and an optical drive for reading or writing from a removable, non-volatile optical disk may be provided. In these cases, each drive may be connected to the bus (not shown) by one or more data media interfaces. Memory 920 may include a computer program product 925 having one or more program modules that are configured to perform various methods or actions of various embodiments of the present disclosure.

[0133] The communication unit 940 enables communication with other computing devices via a communication medium. Additionally, the functions of the components of the computing device / server 900 can be implemented in a single computing cluster or multiple computing machines that can communicate via a communication connection. Therefore, the computing device / server 900 can operate in a networked environment using a logical connection with one or more other servers, a network personal computer (PC), or another network node.

[0134] Input device 950 may be one or more input devices, such as a mouse, keyboard, trackball, etc. Output device 960 may be one or more output devices, such as a display, speaker, printer, etc. Computing device / server 900 may also communicate with one or more external devices (not shown) as needed, such as storage devices, display devices, etc., through communication unit 940, with one or more devices that allow users to interact with computing device / server 900, or with any device that allows computing device / server 900 to communicate with one or more other computing devices (e.g., network card, modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0135] According to an exemplary implementation of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is also provided, which is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method described above.

[0136] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices, equipment, and computer program products implemented according to the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0137] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0138] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0139] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple implementations of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some implementations as replacements, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions.

[0140] The above descriptions of various implementations of the present disclosure are exemplary, non-exhaustive, and not limited to the disclosed implementations. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The selection of terms used herein is intended to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the various implementations disclosed herein.

Claims

1. An information processing method, include: Acquire attribute information related to a user and / or a target object, wherein the attribute information indicates at least one attribute value of the user and / or the target object; By performing a bucketing operation on the attribute information according to a plurality of bucketing strategies corresponding to the plurality of granularities, a plurality of single-granularity feature representations of the attribute information at the plurality of granularities are extracted; Generating a multi-granularity fused feature representation by merging the multiple single-granularity feature representations; as well as Determining the user's interest in the target object based on the multi-granularity fusion feature representation; The generating of the multi-granularity fusion feature representation comprises: Determining the relative importance between each pair of single-granularity feature representations in the plurality of single-granularity feature representations; and Merging the plurality of single-granularity feature representations based on the determined relative importance to generate the multi-granularity fused feature representation; Extracting the plurality of single-granularity feature representations includes: Performing a bucketing operation on the attribute information according to the multiple bucketing strategies to generate multiple embedding representations corresponding to the multiple granularities respectively; and A plurality of feature extraction models corresponding to the plurality of granularities are used to extract a plurality of single-granularity feature representations from the plurality of embedded representations, so that a sum of differences between the plurality of single-granularity feature representations is greater than a predetermined difference threshold.

2. The method of claim 1 , wherein the plurality of feature extraction models are trained by minimizing a predetermined common loss function, wherein the common loss function is determined based on a plurality of feature representation differences between a plurality of training single-granularity feature representations for the plurality of feature extraction models and a plurality of weighted common feature representations, and The plurality of weighted common feature representations are determined by using a plurality of weights corresponding to the plurality of granularities to perform common feature representation, and the common feature representation is determined based on the plurality of embedded representations.

3. The method according to claim 1, wherein the attribute information indicates multiple attribute values ​​of the user and / or the target object ； The extracting of the plurality of single-granularity feature representations comprises: For a given particle size among the plurality of particle sizes, Generating a plurality of embedding vectors corresponding to the plurality of attribute values ​​respectively by performing a bucketing operation on the attribute information, wherein the plurality of embedding vectors have the same number of dimensions; Generate a plurality of column vectors consisting of vector elements of corresponding dimensions in a plurality of dimensions in the plurality of embedding vectors; Utilizing a deep cross network (DCN) to extract multiple cross feature representations of each of the multiple column vectors; as well as A single-granularity feature representation at the given granularity is generated at least by merging the multiple cross-feature representations.

4. The method according to claim 3, wherein a single-granularity feature representation is generated at the given granularity include: extracting a joint feature representation of the plurality of embedding vectors; as well as A single-granularity feature representation at the given granularity is generated by merging the multiple cross feature representations and the joint feature representation.

5. The method according to claim 1, wherein determining the user's interest in the target object include: Determining the user's interest in the target object using an interest level determination model, The interest level determination model is determined based on the training interest level corresponding to the training attribute information and an associated label, wherein the label indicates the actual interest level of the training user associated with the training attribute information in the training target object.

6. The method according to claim 1, further comprising: include: comparing the interest level with a preset threshold; as well as If it is determined that the interest level exceeds the preset threshold, recommendation information about the target object is provided to the user.

7. A device for information processing, include: An attribute information acquisition module, configured to acquire attribute information related to a user and / or a target object, wherein the attribute information indicates at least one attribute value of the user and / or the target object; A single-granularity feature representation extraction module is configured to extract multiple single-granularity feature representations of the attribute information at the multiple granularities by performing a bucketing operation on the attribute information according to multiple bucketing strategies corresponding to the multiple granularities respectively; A multi-granularity fusion feature representation generation module is configured to: generate a multi-granularity fusion feature representation by merging the multiple single-granularity feature representations; as well as an interest level determination module, configured to determine the user's interest level in the target object based on the multi-granularity fusion feature representation; The multi-granularity fusion feature representation generation module includes: A relative importance determination module, configured to determine the relative importance between each pair of single-granularity feature representations in the plurality of single-granularity feature representations; a multi-granularity fused feature representation generating module, configured to merge the plurality of single-granularity feature representations based on the determined relative importance to generate the multi-granularity fused feature representation; The single-granularity feature representation extraction module comprises: an embedding representation generating module, configured to perform a bucketing operation on the attribute information according to the multiple bucketing strategies to generate multiple embedding representations corresponding to the multiple granularities respectively; and The extraction module based on the feature extraction model is configured to extract multiple single-granularity feature representations from the multiple embedded representations using multiple feature extraction models corresponding to the multiple granularities, so that the sum of the differences between the multiple single-granularity feature representations is greater than a predetermined difference threshold.

8. The apparatus of claim 7, wherein the plurality of feature extraction models are trained by minimizing a predetermined common loss function, wherein the common loss function is determined based on a plurality of feature representation differences between a plurality of training single-granularity feature representations for the plurality of feature extraction models and a plurality of weighted common feature representations, and The plurality of weighted common feature representations are determined by using a plurality of weights corresponding to the plurality of granularities to perform common feature representation, and the common feature representation is determined based on the plurality of embedded representations.

9. The apparatus according to claim 7, wherein the attribute information indicates a plurality of attribute values ​​of the user and / or the target object. ； The single-granularity feature representation extraction module comprises: an embedding vector generating module, configured to generate, for a given granularity among the plurality of granularities, a plurality of embedding vectors respectively corresponding to the plurality of attribute values ​​by performing a bucketing operation on the attribute information, wherein the plurality of embedding vectors have the same number of dimensions; a column vector generation module configured to generate, for a given granularity among the plurality of granularities, a plurality of column vectors consisting of vector elements of corresponding dimensions among the plurality of dimensions in the plurality of embedded vectors; a cross feature representation extraction module configured to extract, for a given granularity among the plurality of granularities, a plurality of cross feature representations of each of the plurality of column vectors using a deep cross network (DCN); and The single-granularity feature representation generating module is configured to generate a single-granularity feature representation at a given granularity among the multiple granularities by at least merging the multiple cross feature representations.

10. The apparatus according to claim 9, wherein the single-granularity feature representation generation module include: a joint feature representation extraction module configured to extract a joint feature representation of the plurality of embedding vectors; as well as The single-granularity feature representation merging generation module is configured to generate a single-granularity feature representation at the given granularity by merging the multiple cross feature representations and the joint feature representation.

11. The apparatus according to claim 7, wherein the interest level determination module is further configured to determine the user's interest level in the target object using an interest level determination model, The interest level determination model is determined based on the training interest level corresponding to the training attribute information and an associated label, wherein the label indicates the actual interest level of the training user associated with the training attribute information in the training target object.

12. The device according to claim 7, further comprising: include: a comparison module, configured to compare the interest level with a preset threshold; as well as The recommendation information providing module is configured to provide the user with recommendation information about the target object in response to determining that the interest level exceeds the preset threshold.

13. An electronic device, include: processor; as well as A memory for storing computer executable instructions, wherein when the computer executable instructions are executed by the processor, the memory is configured to implement the method according to any one of claims 1 to 6.

14. A computer-readable storage medium having computer-executable instructions stored thereon, wherein when the computer-executable instructions are executed by a processor, the processor is caused to perform the method according to any one of claims 1 to 6.

Images