Information Processing Method, Apparatus, Computer Device, and Storage Medium

Abstract

The present application relates to an information processing method, apparatus, computer device, and storage medium. The method includes: obtaining sample user data, which includes a first number of sample objects; extracting first sample object features of the sample objects, first positive sample features of positive samples of the sample objects, and first negative sample features of negative samples of the sample objects through a feature extraction model, optimizing the feature extraction model to obtain a trained feature extraction model; extracting second sample object features of the first number of sample objects through the trained feature extraction model; performing multi-interest extraction on the second sample object features through a multi-interest extraction model to obtain a second number of sample interest features, optimizing to obtain a trained multi-interest extraction model, and finally obtaining a multi-interest feature extraction model. By using this method, the sample object features can have distinctiveness, improving the accuracy of the multi-interest feature extraction model in extracting the interest features of a target user and enhancing the effect of interest recommendation to the target user.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular, to an information processing method, apparatus, computer device, computer-readable storage medium, and computer program product. Background Art

[0002] With the development of artificial intelligence technology, relevant recommendations can be made based on user interests in various fields such as games, information, and videos. In a large-scale recommendation system, it is necessary to model user interests both in the item recall stage and the post-processing stage.

[0003] In traditional technologies, generally, average pooling is performed on the vectors of the user's historical click sequences, and a vector with the same dimension is output on both the user side and the item side and an inner product is calculated to fit and predict the user's click interest in the item. However, the accuracy of this method of prediction is not high, and the types of recalled items are too single, resulting in poor effects of interest recommendation for users. Summary of the Invention

[0004] Based on this, in order to solve the above technical problems, it is necessary to provide an information processing method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve the effect of interest recommendation to users.

[0005] An information processing method, the method comprising:

[0006] Obtain sample user data, where the sample user data includes sample behavior data, and the sample behavior data includes a first number of sample objects;

[0007] Extract a first sample object feature of the sample object, a first positive sample feature of the positive sample of the sample object, and a first negative sample feature of the negative sample of the sample object through a feature extraction model, and optimize the feature extraction model with the goal that the sample object feature is similar to the first positive sample feature and the sample object feature is far from the first negative sample feature until the training end condition is reached, to obtain a trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object;

[0008] Extract a second sample object feature of the first number of sample objects through the trained feature extraction model;

[0009] Perform multi-interest extraction on the first number of the second sample object features through a multi-interest extraction model to obtain a second number of sample interest features, and optimize the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached, to obtain a trained multi-interest extraction model;

[0010] Obtain a multi-interest feature extraction model through the trained feature extraction model and the trained multi-interest extraction model.

[0011] An information processing method, the method comprising:

[0012] Obtain target data of a target user, the target data including historical behavior data and attribute data, and the historical behavior data including a first number of behavior objects;

[0013] Perform multi-interest feature extraction on the target data through a multi-interest feature extraction model to obtain target features of the target user; the multi-interest feature extraction includes: extracting behavior object features of the first number of the behavior objects through a feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of the behavior object features through a multi-interest extraction model in the multi-interest feature extraction model to obtain the second number of first interest features, and extracting attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of the first interest features with the attribute features respectively to obtain the target features of the target user, and the target features include the second number of second interest features; the multi-interest feature extraction model is obtained by using the method as described above;

[0014] Determine a target recommendation object to be recommended to the user according to the target features.

[0015] An information processing device, the device comprising:

[0016] A sample data acquisition module, configured to acquire sample user data, the sample user data including sample behavior data, and the sample behavior data including a first number of sample objects;

[0017] A feature extraction model training module, configured to extract first sample object features of the sample objects, first positive sample features of positive samples of the sample objects, and first negative sample features of negative samples of the sample objects through a feature extraction model, and optimize the feature extraction model with the goal that the sample object features are close to the first positive sample features and the sample object features are far from the first negative sample features until the training end condition is reached, to obtain a trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object;

[0018] A sample feature extraction module, configured to extract second sample object features of the first number of sample objects through the trained feature extraction model;

[0019] A multi-interest extraction model training module, configured to perform multi-interest extraction on the first number of the second sample object features through a multi-interest extraction model to obtain a second number of sample interest features, and optimize the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until an optimization end condition is reached, so as to obtain a trained multi-interest extraction model;

[0020] A model determination module, configured to obtain a multi-interest feature extraction model through the trained feature extraction model and the trained multi-interest extraction model.

[0021] An information processing device, the device includes:

[0022] A target data acquisition module, configured to acquire target data of a target user, the target data includes historical behavior data and attribute data, and the historical behavior data includes a first number of behavior objects;

[0023] A target feature extraction module, configured to perform multi-interest feature extraction on the target data through a multi-interest feature extraction model to obtain target features of the target user; the multi-interest feature extraction includes: extracting behavior object features of the first number of the behavior objects through the feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of the behavior object features through the multi-interest extraction model in the multi-interest feature extraction model to obtain the second number of first interest features, and extracting attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of the first interest features with the attribute features respectively to obtain the target features of the target user, the target features include the second number of second interest features; the multi-interest feature extraction model is obtained by using the method as described above;

[0024] A recommended object determination module, configured to determine a target recommended object to be recommended to the target user according to the target features.

[0025] A computer device, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the above method are implemented.

[0026] A computer-readable storage medium, storing a computer program, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0027] A computer program product includes a computer program which, when executed by a processor, implements the steps of the above-mentioned method.

[0028] In the above information processing method, device, computer equipment, computer-readable storage medium and computer program product, by introducing positive samples and negative samples of sample objects during the training process of the feature extraction model of the multi-interest feature extraction model and then training the feature extraction model, the distance between similar sample object features can be shortened and the distance between dissimilar sample object features can be lengthened through the trained feature extraction model, thereby making the discrimination of sample object features higher. By using the discriminative sample object features during the training process of the multi-interest extraction model of the multi-interest feature extraction model, the accuracy of the extracted sample interest features can be made higher. Thus, the accuracy of extracting the interest features of the target user by the subsequent multi-interest feature extraction model is improved. Furthermore, the effect of interest recommendation to the target user based on the interest features of the target user is improved. Description of the Drawings

[0029] Figure 1 It is an application environment diagram of the information processing method in an embodiment;

[0030] Figure 2 It is a flowchart of the information processing method in an embodiment;

[0031] Figure 3 It is a schematic diagram of the feature extraction model in an embodiment;

[0032] Figure 4 It is a schematic diagram of the multi-interest extraction model in an embodiment;

[0033] Figure 5 It is a flowchart of the information processing method in another embodiment;

[0034] Figure 6 It is a schematic diagram of the multi-interest degree feature extraction model in an embodiment;

[0035] Figure 7 It is a schematic diagram of the interest degree prediction model in an embodiment;

[0036] Figure 8 It is a schematic diagram of the traditional information processing method in a specific embodiment;

[0037] Figure 9 It is a schematic diagram of the information processing method in a specific embodiment;

[0038] Figure 10 It is a structural block diagram of the information processing device in an embodiment;

[0039] Figure 11It is a structural block diagram of an information processing device in another embodiment;

[0040] Figure 12 It is an internal structure diagram of a computer device in one embodiment;

[0041] Figure 13 It is an internal structure diagram of a computer device in another embodiment. Detailed implementation manners

[0042] In order to make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0043] It should be noted first that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) of the users involved in the present application, for example, sample user data, sample behavior data, target data, historical behavior data, attribute data, etc., are all information and data authorized by the users or fully authorized by all parties.

[0044] In one embodiment, the information processing method provided by the present application can be applied to an application environment as Figure 1 shown. This application environment may involve both the terminal 102 and the server 104. In some embodiments, it may also involve the terminal 106 at the same time. Among them, the terminal 102 and the terminal 106 can communicate with the server 104 through the network respectively. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated in the server 104, or can be placed on the cloud or other network servers.

[0045] Specifically, the server 104 can obtain sample user data through the terminal 102 and / or the terminal 106. The sample user data includes sample behavior data. The sample behavior data includes a first number of sample objects. The server 104 can extract the first sample object features of the sample objects, the first positive sample features of the positive samples of the sample objects, and the first negative sample features of the negative samples of the sample objects through a feature extraction model, and take the sample object features as being close to the first positive sample features and the sample object features as being far from the first negative sample features as the goal, optimize the feature extraction model until the training end condition is reached, and obtain the trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object.

[0046] Server 104 extracts the second sample object features of the first number of sample objects through the trained feature extraction model; performs multi-interest extraction on the first number of second sample object features through the multi-interest extraction model to obtain the second number of sample interest features, and optimizes the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached to obtain the trained multi-interest extraction model; obtains the multi-interest feature extraction model on Server 104 through the trained feature extraction model and the trained multi-interest extraction model.

[0047] Thus, taking the user of terminal 102 as the target user, for example, when Server 104 makes interest recommendations to the target user, Server 104 can obtain the target data of the target user through terminal 102. The target data includes historical behavior data and attribute data, and the historical behavior data includes the first number of behavior objects.

[0048] Server 104 performs multi-interest feature extraction on the target data through the obtained multi-interest feature extraction model to obtain the target features of the target user; the multi-interest feature extraction includes: extracting the behavior object features of the first number of behavior objects through the feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of behavior object features through the multi-interest extraction model in the multi-interest feature extraction model to obtain the second number of first interest features, and extracting the attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of first interest features with the attribute features respectively to obtain the target features of the target user. The target features include the second number of second interest features; Server 104 determines the target recommended objects to be recommended to the target user according to the target features. Furthermore, the target recommended objects can be pushed to terminal 102 to implement interest recommendation for the target user.

[0049] In some other embodiments, when the data processing capabilities of terminal 102 and / or terminal 106 meet the requirements, after training and obtaining the multi-interest feature extraction model in Server 104, Server 104 can also send the multi-interest feature extraction model to terminal 102 and / or terminal 106. Thus, relevant data processing of the target data of the target user is performed in terminal 102 and / or terminal 106, and finally, the target recommended objects to be recommended to the target user are determined to implement interest recommendation for the target user.

[0050] Among them, the terminal 102 and the terminal 106 can be, but are not limited to, various mobile phones, computers, intelligent voice interaction devices, intelligent household appliances, vehicle-mounted terminals, portable wearable devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers. The embodiments involved in this application can be applied to various scenarios such as cloud technology and artificial intelligence. For example, the data storage system 104 can be a cloud storage system to store the sample user data of each user and the multi-interest feature extraction model obtained through training.

[0051] In one embodiment, as Figure 2 shown, an information processing method is provided. Taking the method applied to Figure 1 the server 104 in it as an example for illustration, it includes the following steps:

[0052] Step S202, obtain sample user data, where the sample user data includes sample behavior data, and the sample behavior data includes a first number of sample objects.

[0053] In one embodiment, the sample user data refers to the data of sample users that are used as training samples during the training process of the model. The sample user data includes sample behavior data, and the sample behavior data includes, but is not limited to, the access data of the sample user within a preset duration on various web pages and interfaces. Specifically, it can include data such as clicks, searches, and browsing. Among them, the preset duration can be set according to actual technical needs such as model training accuracy and model training speed, and is not limited here. In some embodiments, the sample user data further includes sample attribute data, and the sample attribute data includes, but is not limited to, the name, gender, age, city where the sample user is located, and the city level of the city where the sample user is located. It should be noted that the above-mentioned sample user data, sample behavior data, sample attribute data, etc. are all information and data that have been authorized by the user or fully authorized by all parties.

[0054] Among them, the various items included in the sample behavior data such as clicks, searches, and browsing are called sample objects, and the types of items include, but are not limited to, text, pictures, etc. Since the amount of data in the sample behavior data is large, there are multiple numbers of sample objects included. The number of sample objects in the obtained sample behavior data is determined as the first number, that is, the sample behavior data includes a first number of sample objects. The first number can be set according to actual technical needs and is not limited here. In one embodiment, the first number can be represented as N, and the first number of sample objects can be represented as sample object 1, sample object 2... sample object N.

[0055] Step S204, extracting the first sample object feature of the sample object, the first positive sample feature of the positive sample of the sample object, and the first negative sample feature of the negative sample of the sample object through the feature extraction model, and optimizing the feature extraction model with the goal that the sample object feature is close to the first positive sample feature and the sample object feature is far away from the first negative sample feature, until the training end condition is reached, and a trained feature extraction model is obtained; wherein the positive sample is the sample object, and the negative sample is another randomly selected sample object.

[0056] In one embodiment, in order to analyze and quantify the interests of sample users and further realize subsequent interest prediction and recommendation, it is necessary to first perform feature extraction analysis on the acquired sample user data and convert the sample user data into quantifiable feature vectors, which can be quantified using a feature extraction model. The model structure and model type of the feature extraction model can be set and selected according to actual technical needs.

[0057] In order to make the features of sample objects more distinguishable, the feature extraction model uses a contrastive learning model. Specifically, using a common feature extraction model, the distribution of features of sample objects is relatively uniform, and the idea of ​​contrastive learning is to shorten the distance between similar samples and widen the distance between dissimilar samples. Similar samples are called positive samples, and dissimilar samples are called negative samples. The goal of contrastive learning is to learn a good semantic representation space from samples. Contrastive learning can be unsupervised or self-supervised learning, that is, the feature extraction model of this embodiment can make the distinction between each feature higher. The contrastive learning method of the feature extraction model can also be called contrastive learning based on sample object features.

[0058] Among them, in order to simplify the network structure and suppress overfitting at the same time, the feature extraction model can adopt a fully connected network with a simpler structure after regularization processing. The type of regularization processing can be set according to actual technical needs. In one embodiment, Dropout is used, that is, the feature extraction model is a fully connected network with Dropout. Dropout is a method of randomly deleting neurons during the learning process. During training, neurons in the hidden layer are randomly selected and deleted. The deleted neurons no longer transmit signals. During testing, all neuron signals are transmitted. For the output of each neuron, it is necessary to multiply it by the deletion ratio during training before outputting it. Therefore, when the same sample object is input into the feature extraction model twice, due to the existence of Dropout, the characteristics of the sample object obtained are not exactly the same.

[0059] Since it is necessary to perform contrastive learning based on the positive and negative samples of the sample object to reduce the distance between similar samples and increase the distance between dissimilar samples, it is necessary to construct the positive and negative samples of the sample object and then perform the training of the subsequent feature extraction model. Among them, the positive sample is set as the sample object itself, that is, the sample object is input into the feature extraction model twice. The feature of the sample object is called the first sample object feature, and the feature of the positive sample of the sample object is called the first positive sample feature. The negative sample is another randomly selected sample object, and the feature of the negative sample of the sample object is called the first negative sample feature.

[0060] It can be understood that in order to improve the efficiency of information processing, and during the training process of the feature extraction model, it may have ended training without using all the input sample objects and their positive and negative samples. Therefore, it is not necessarily necessary to construct the positive and negative samples of the sample object for each sample object. When each sample object needs to be input, it is sufficient to construct the positive and negative samples of the sample object and input them.

[0061] In one embodiment, the sample object, the positive sample of the sample object, and the negative sample of the sample object are respectively input into the feature extraction model, and the corresponding features can be obtained, and then the subsequent training of the feature extraction model can be performed. Taking the extraction of the first sample object feature of the sample object by the feature extraction model as an example, it specifically may include steps S2041 to step S2042.

[0062] Step S2041, input the sample object into the feature extraction model to obtain the initial sample object feature, and the feature extraction model is a fully connected network after regularization processing.

[0063] Among them, the feature extraction model is a fully connected network after regularization processing, that is, the feature extraction model is a fully connected network with Dropout. Input the sample object into the feature extraction model, and the obtained feature of the sample object is called the initial sample object feature. Similarly, input the positive sample of the sample object into the feature extraction model, and the obtained feature of the sample object is called the initial positive sample feature. Input the negative sample of the sample object into the feature extraction model, and the obtained feature of the sample object is called the initial negative sample feature.

[0064] Since the types of sample objects are diverse, the dimensions of the obtained initial sample object features may be different. Similarly, the dimensions of the obtained initial positive sample features and initial negative sample features may also be different. Therefore, it is necessary to unify their dimensions for subsequent processing.

[0065] Step S2042, convert the initial sample object feature into a feature of a predetermined dimension vector to obtain the first sample object feature.

[0066] Among them, any implementable method can be used to transform the initial sample object features into features of a predetermined-dimensional vector. In one embodiment, a vectorization processing layer, i.e., an Embedding layer, can be used to encode the initial sample object features and transform them into Embedding features of a predetermined-dimensional vector. The predetermined dimension can be set according to actual technical needs, and then the first sample object features are obtained. That is, the first sample object features are low-dimensional dense vectors. That is, the feature extraction model in this embodiment is composed of a fully connected network with Dropout and an Embedding layer. Similarly, similar processing is performed on the initial positive sample features and the initial negative sample features to obtain the first positive sample features and the first negative sample features respectively. The specific process is not elaborated here.

[0067] After extracting the first sample object features of the sample object, the first positive sample features of the positive sample of the sample object, and the first negative sample features of the negative sample of the sample object through the feature extraction model, it is necessary to train the feature extraction model, that is, perform unsupervised or self-supervised learning to obtain the trained feature extraction model for subsequent recommendation for the target user.

[0068] Specifically, with the goal that the sample object features are close to the first positive sample features and the sample object features are far from the first negative sample features, the feature extraction model is optimized until the training end condition is reached to obtain the trained feature extraction model.

[0069] Among them, the training end condition can be set according to actual technical needs. In one embodiment, it can be set as the loss convergence of the test set. In another embodiment, the training end condition can also be set as reaching a preset number of iterations, etc. The loss function of the trained feature extraction model in one embodiment is expressed as follows:

[0070]

[0071] Among them, h i represents the product of the first sample object features of sample object i and the first positive sample features of the positive sample of sample object i, and h' i represents the expanded and replicated h i , represents the first negative sample features of the negative sample of sample object i, and τ is a hyperparameter.

[0072] Step S206, extract the second sample object features of the first number of sample objects through the trained feature extraction model.

[0073] In one embodiment, after obtaining the trained feature extraction model, the first number of sample objects need to be input into the trained feature extraction model again to obtain the second sample object features of the first number of sample objects. Among them, the second sample object features are discriminative features, that is, the distance between similar samples is shortened, and the distance between dissimilar samples is lengthened.

[0074] Among them, after passing through the Embedding layer of the feature extraction model, the second sample object features of the obtained first number of sample objects can also be serialized, that is, the first number of discriminative second sample object features are converted into corresponding vector sequences. In one embodiment, the vector sequence is represented as:

[0075]

[0076] Among them, I u represents the sample object, represents the second sample object feature, and E U represents the vector sequence.

[0077] For ease of understanding, this embodiment will describe the above calculation process in combination with the accompanying drawings. As Figure 3 shown is a schematic diagram of the feature extraction model. Among them, the feature extraction model uses a fully connected network after regularization processing, that is, a fully connected network with Dropout, to construct positive and negative samples of the sample object. The sample object, the positive sample of the sample object, and the negative sample of the sample object are input into the feature extraction model, and the first sample object feature, the first positive sample feature, and the first negative sample feature are respectively extracted. With the goal that the sample object feature is close to the first positive sample feature and the sample object feature is far from the first negative sample feature, the feature extraction model is optimized until the training end condition is reached, and the trained feature extraction model is obtained.

[0078] After obtaining the trained feature extraction model, each sample object is input into the model again to obtain the discriminative second sample object features of each sample object, and the second sample object features are serialized, that is, the second sample object features are all Embedding features of a preset dimension.

[0079] Step S208, perform multi-interest extraction on the first number of second sample object features through a multi-interest extraction model to obtain the second number of sample interest features, and optimize the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached, and obtain the trained multi-interest extraction model.

[0080] In one embodiment, a multi-interest vector of a sample user can be extracted from the features of the sample user to fully characterize the user's interests. Specifically, a multi-interest extraction model can be used for extraction. The multi-interest extraction model in one embodiment can select the Multi-Interest Extractor Layer model, which is mainly based on dynamic routing, and its model structure and model type can be set according to actual technical needs. In this embodiment, the multi-interest extraction model can be a capsule network. For a capsule network, generally, a group of vectors is input, an affine transformation is performed on this group of vectors, and a weighted sum is carried out, and then it is processed through a non-linear mapping function to obtain the output of another group of vectors.

[0081] Among them, it is necessary to first construct a capsule network. The number of layers of the capsule network can be set to two or three layers, which can be specifically set according to actual technical needs. In one embodiment, taking the capsule network set to two layers as an example, the first layer and the second layer are the low-level capsule and the high-level capsule respectively. The core goal of dynamic routing is to calculate the high-level capsule based on the low-level capsule through an iterative method. Specifically, the input of the capsule network is the output of the trained feature extraction model, that is, the input is the first number of second sample object features.

[0082] In one embodiment, the input second sample object features are represented as The number of input second sample object features is the first number, denoted by m here, and l represents the l-th layer. The output of the capsule network is the sample interest feature, and the sample interest feature is represented as First, it is necessary to calculate the routing weights between the two layers of the capsule network, and the routing weights are represented as b ij , and the calculation formula is:

[0083]

[0084] Among them, S ij represents a learnable fully connected matrix, which can be adaptively adjusted according to the loss of the model during the model training process. Each input second sample object feature corresponds to a separate learnable fully connected matrix. represents the second sample object features after being processed by the fully connected matrix.

[0085] After determining the routing weights, according to the routing weights b ij perform a weighted sum on to calculate the candidate output vector of the high-level capsule of the capsule network The calculation formula is expressed as:

[0086]

[0087] Wherein, w ij represents the routing weight after normalization between two layers, and the calculation formula is:

[0088]

[0089] After determining the candidate output vector of the capsule network it is processed using a non-linear mapping function to obtain the output of the high-level capsule The calculation formula is:

[0090]

[0091] Wherein, the non-linear mapping function can be set according to actual technical needs. In this embodiment, the squash function is adopted.

[0092] It should be noted that the above calculation process is an iterative calculation process of the capsule network. It is also necessary to optimize the multi-interest extraction model based on the similarity between the output sample interest feature and the corresponding second sample object feature, and adjust the weight corresponding to the second sample object feature. Specifically, the weight corresponding to the second sample object in the next iteration process can be updated on the basis of the weight corresponding to the second sample object in the previous iteration process, and the weight corresponding to the similar second sample object is amplified until the optimization end condition is reached, and the trained multi-interest extraction model is obtained.

[0093] Wherein, the optimization end condition can be set according to actual technical needs. In one embodiment, it can be set as the similarity threshold between the sample interest feature and the corresponding second sample object feature, or it can also be set as the numerical convergence of the high-level capsule. When the trained multi-interest extraction model is obtained, the second number of sample interest features corresponding to the first number of second sample object features input can be obtained. In one embodiment, the second number corresponding to the sample interest feature is represented as K, wherein, the second number is less than the first number.

[0094] For the convenience of understanding, the above calculation process will be described in conjunction with the accompanying drawings in this embodiment. As Figure 4The figure shows a schematic diagram of a multi-interest extraction model. Here, the multi-interest extraction model is taken as a two-layer capsule network for illustration. The feature vectors i1 and i2 are input into the capsule network, and the corresponding learnable fully connected matrices w1 and w2 are used to operate on the input feature vectors i1 and i2 respectively to obtain the feature vectors u1 and u2 after being processed by the fully connected matrices. The calculation formula is expressed as:

[0095] u1 = w1 × i1

[0096] u2 = w2 × i2

[0097] In the first iteration of the model, the randomly initialized routing weights corresponding to the feature vectors u1 and u2 after being processed by the fully connected matrices are set as c0 and d0. Weighted summation is performed according to the routing weights, and then the candidate output vector s1 corresponding to the low-level capsule is determined. The calculation formula is expressed as:

[0098] s1 = c0u1 + d0u2

[0099] The candidate output vector s1 is processed by using the non-linear mapping squash function to obtain the output v of the high-level capsule. The calculation formula is expressed as:

[0100]

[0101] In the second iteration of the model, the routing weights will be updated based on the routing weights in the first iteration. The updated routing weights are expressed as:

[0102] c1 = c0 + u1v

[0103] In the third iteration of the model, the routing weights will be updated based on the routing weights in the second iteration. The updated routing weights are expressed as:

[0104] c2 = c1 + u1v

[0105] This model will perform multiple iterations, and use dynamic routing to extract the interests of sample users until the optimization end condition is reached, and the trained multi-interest extraction model is obtained.

[0106] Step S210, through the trained feature extraction model and the trained multi-interest extraction model, a multi-interest feature extraction model is obtained.

[0107] Among them, by combining the trained feature extraction model and the trained multi-interest extraction model, the multi-interest feature extraction model is jointly determined, which can be used to extract the interest features of target users in the future.

[0108] In the above information processing method, by obtaining sample user data, the sample user data includes sample behavior data, and the sample behavior data includes a first number of sample objects; extracting the first sample object feature of the sample object, the first positive sample feature of the positive sample of the sample object, and the first negative sample feature of the negative sample of the sample object through a feature extraction model, and aiming at making the sample object feature close to the first positive sample feature and far from the first negative sample feature, optimizing the feature extraction model until the training end condition is reached, and obtaining the trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object; extracting the second sample object feature of the first number of sample objects through the trained feature extraction model; performing multi-interest extraction on the first number of second sample object features through a multi-interest extraction model to obtain a second number of sample interest features, and optimizing the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached, and obtaining the trained multi-interest extraction model; obtaining a multi-interest feature extraction model through the trained feature extraction model and the trained multi-interest extraction model. By adopting the method of the above embodiment, by introducing the positive sample and negative sample of the sample object in the training process of the feature extraction model of the multi-interest feature extraction model and then training the feature extraction model, the distance between similar sample object features can be shortened and the distance between dissimilar sample object features can be lengthened through the trained feature extraction model, so that the discrimination degree of the sample object features is higher. By using the sample object features with discrimination degree in the training process of the multi-interest extraction model of the multi-interest feature extraction model for training, the accuracy of the extracted sample interest features can be higher, thereby improving the accuracy of the subsequent multi-interest feature extraction model in extracting the interest features of the target user, and further improving the effect of interest recommendation to the target user according to the interest features of the target user.

[0109] After training to obtain the multi-interest feature extraction model, the multi-interest feature extraction model can be used to perform interest recommendation for the target user.

[0110] In one of the embodiments, as Figure 5 shown, an information processing method is provided. Taking the example that this method is applied to Figure 1 the server 104 as an example, it includes the following steps:

[0111] Step S502, obtaining the target data of the target user, where the target data includes historical behavior data and attribute data, and the historical behavior data includes a first number of behavior objects.

[0112] The data of the target user is referred to as target data, and the target data includes historical behavior data. The historical behavior data includes, but is not limited to, the access data of the target user within a preset historical duration on various web pages and interfaces. Specifically, it may include data such as clicks, searches, and browsing. Among them, the preset historical duration can be set according to actual technical needs and is not limited here. The target data also includes attribute data, and the attribute data includes, but is not limited to, the name, gender, age, city where the target user is located, and the city level of the city where the target user is located, etc. It should be noted that the above-mentioned target data, historical behavior data, attribute data, etc. of the target user are all information and data authorized by the user or fully authorized by all parties.

[0113] Among them, various items included in the target behavior data such as clicks, searches, and browsing are referred to as behavior objects, and the types of items include, but are not limited to, text, pictures, etc. Since the amount of data in the target behavior data is large, there are multiple numbers of behavior objects included. The number of target objects in the obtained target behavior data is determined as the first number, that is, the historical behavior data includes the first number of behavior objects. The first number can be set according to actual technical needs and is not limited here. In one embodiment, the first number can be represented as N, and the first number of behavior objects are represented as behavior object 1, behavior object 2... behavior object N.

[0114] Step S504, perform multi-interest feature extraction on the target data through a multi-interest feature extraction model to obtain the target features of the target user; the multi-interest feature extraction includes: extracting the behavior object features of the first number of behavior objects through the feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of behavior object features through the multi-interest extraction model in the multi-interest feature extraction model to obtain the second number of first interest features, and extracting the attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of first interest features with the attribute features respectively to obtain the target features of the target user. The target features include the second number of second interest features; the multi-interest feature extraction model is obtained by using the method of steps S202 to S210.

[0115] In one embodiment, in order to quantify and characterize the user's interests, perform multi-interest feature extraction on the target data through a multi-interest feature extraction model, and refer to the features extracted by the multi-interest feature extraction model as target features, and the target features include the second number of second interest features. Among them, such as Figure 6The figure shows a schematic diagram of a multi-interest feature extraction model. The multi-interest feature extraction model includes a trained feature extraction model and a trained multi-interest extraction model. When using the multi-interest feature extraction model, first, the features corresponding to the first number of behavioral objects in the input are extracted through the trained feature extraction model. The extracted features are called behavioral object features, and the number of behavioral object features is the first number, denoted as N. Then, the first number of behavioral object features are input into the trained multi-interest extraction model for multi-interest extraction, and the obtained features are called first interest features. The number of first interest features is the second number, denoted as K, where the second number is less than the first number.

[0116] In one embodiment, in order to better combine the attribute data of the target user and the first interest features obtained by the multi-interest extraction model and improve the accuracy of subsequent prediction recall, after obtaining the second number of first interest features, it further includes: extracting the attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of first interest features with the attribute features respectively to obtain the target features of the target user, where the target features include the second number of second interest features, and the target features can be used to represent the user's interest in a certain aspect.

[0117] Specifically, the feature fusion method can be carried out in any implementable way. Taking the splicing method as an example, fusing the second number of first interest features with the attribute features respectively to obtain the target features of the target user includes: splicing the second number of first interest features with the attribute features respectively to obtain the target features of the target user.

[0118] Among them, when splicing the second number of first interest features with the attribute features respectively, the positions of the attribute features and the first interest features can be set according to the actual situation. In one embodiment, the attribute features can be spliced behind the first interest features.

[0119] In one embodiment, when splicing features, the dimension of the first interest features can be made the same as the dimension of the attribute features to improve the accuracy of the obtained target features after splicing. Referring to the foregoing embodiment, after feature extraction by the trained feature extraction model in the multi-interest feature extraction model, the obtained first number of behavioral object features are Embedding features. Therefore, the attribute data also needs to be subjected to corresponding Embedding processing, and the obtained attribute features are also Embedding features to ensure the consistency of the dimensions of the obtained first interest features and the attribute features.

[0120] Among them, after concatenating the second number of first interest features with the attribute features respectively, the concatenated feature vector also needs to be processed through two fully-connected layers to finally obtain the target feature of the target user. Among them, the activation functions of the two fully-connected layers can be set according to actual technical needs. In one embodiment, the activation function is set to the rectified linear unit (Relu) function, which can make the finally obtained target feature of the target user more representative, improve the generalization ability of the network, and alleviate overfitting.

[0121] Step S506: Determine the target recommendation object to be recommended to the target user according to the target feature.

[0122] In one embodiment, the object finally recommended to the user is called the target recommendation object, and the target recommendation object may include one or more.

[0123] Specifically, after obtaining the target feature, when determining the target recommendation object to be recommended to the target user according to the target feature, different methods can be used. One of them can be the recall method, that is, according to the target feature, recall multiple items with similar semantics to the target feature, and use the recalled multiple items as the objects to be recommended, and determine the final target recommendation object from each object to be recommended. Another one can be the prediction method, that is, according to the target feature, predict the interest degree of the target user for one or more determined objects to be recommended. The interest degree can represent the probability that the target user clicks and browses the object to be recommended. Furthermore, determine the final target recommendation object from the objects to be recommended. The following will give examples of these two methods respectively. It should be understood that in other embodiments, after obtaining the target feature, other methods can also be used to determine the target recommendation object.

[0124] In one embodiment, if the recall method is used to determine the target recommendation object, at this time, the above step of determining the target recommendation object to be recommended to the target user according to the target feature may include steps S5061 to S5062.

[0125] Step S5061: Perform similarity matching between each second interest feature and the feature of the object to be recommended for each object to be recommended, and obtain the object to be recommended that each second interest feature matches.

[0126] In one embodiment, the similarity matching may be performed according to the cosine similarity between features. Specifically, for each second interest feature, calculate the cosine similarity between the second interest feature and the feature of each object to be recommended. The closer the value of the cosine similarity is to 1, the closer the included angle is to 0 degrees, that is, the more similar the second interest feature is to the feature of the object to be recommended. Among them, obtaining the recommended objects matched by each second interest feature, that is, determining the feature of the object to be recommended that is most similar to each second interest feature. Among them, the number of objects to be recommended matched by each second interest feature is not limited, and can be one or more than one.

[0127] Step S5062, post-process the objects to be recommended matched by each second interest feature to obtain the target recommended objects to be recommended to the target user.

[0128] In one embodiment, after obtaining multiple objects to be recommended respectively matched by each second interest vector, it is necessary to post-process the objects to be recommended to obtain the target recommended objects to be recommended to the target user. Among them, the post-processing methods include but are not limited to screening, sorting, etc., and can be specifically selected according to actual technical needs. Combining the actual situation of interest recommendation, the number of target recommended objects finally determined to be recommended to the target user can be one or more, which is not limited here.

[0129] Based on the above-mentioned method, the objects to be recommended corresponding to the most similar preset number of object features to be recommended can be recalled, that is, the obtained objects to be recommended include multiple ones.

[0130] Among them, the preset number can be set according to actual technical needs. In one embodiment, it can be set to N. The method of determining the most similar preset number of object features to be recommended and recalling the objects to be recommended corresponding to the most similar preset number of object features to be recommended can choose any implementable method. In one embodiment, it can be implemented by the vector retrieval method, and the vector retrieval method can be selected according to actual technical needs. For example, the Faiss retrieval method can be used. Faiss is a library for clustering and similarity search, which can provide efficient similarity search and clustering for dense vectors, support the search of vectors at the billion level, and is the most mature approximate nearest neighbor search library at present. It includes various algorithms for searching vector sets of any size, as well as support codes for algorithm evaluation and parameter adjustment, and its function includes similarity search.

[0131] In one embodiment, if the target recommended object is determined by prediction, at this time, the above-mentioned determining the target recommended object to be recommended to the target user according to the target feature may further include step S5063 to step S5064.

[0132] Step S5063: Calculate the interest degree of the target user in the object to be recommended by performing an operation on the second number of second interest features and the features of the object to be recommended of the object to be recommended. The operation includes: performing a weighted sum on the second number of second interest features and the features of the object to be recommended of the object to be recommended to obtain a target weighted feature; and performing a dot product operation on the target weighted feature and the features of the object to be recommended of the object to be recommended to obtain the interest degree of the target user in the object to be recommended.

[0133] In one embodiment, for a determined object to be recommended, the second number of second interest features in the target features of the target user and the features of the object to be recommended of the object to be recommended can be calculated to obtain the interest degree of the target user in the object to be recommended, and this interest degree is used to represent the interest of the target user in the recommended object. Specifically, calculating the interest degree of the target user in the object to be recommended by performing an operation on the second number of second interest features and the features of the object to be recommended of the object to be recommended includes: calculating the interest of the target user in the object to be recommended by performing an operation on each target feature and the features of the object to be recommended of the object to be recommended through an interest degree prediction model.

[0134] Among them, since a target user has multiple second interest features, and different second interest features may correspond to different interests, the interest degree prediction model in this embodiment can adopt a model including an attention prediction model. The attention prediction model is the Label-aware Attention Layer, which adopts the Attention method to determine the content that the target user is more interested in.

[0135] In one embodiment, the second number of second interest features and the features of the object to be recommended of the object to be recommended are calculated through an interest degree prediction model. Among them, the operation includes: performing a weighted sum on the second number of second interest features and the features of the object to be recommended of the object to be recommended to obtain a target weighted feature; and performing a dot product operation on the target weighted feature and the features of the object to be recommended of the object to be recommended to obtain the interest degree of the target user in the object to be recommended.

[0136] Specifically, the Label-aware Attention Layer is used to calculate the correlation between the second number of second interest features and the features of the object to be recommended of the object to be recommended, and different weights are assigned to the second interest features. Among them, the target weighted feature is represented as The operation process of the above weighted sum can be expressed as:

[0137]

[0138] Among them, represents the features of the object to be recommended, V uIt represents the second interest feature, p is a hyperparameter, and softmax represents the normalization process.

[0139] After determining the target weighted feature Then, perform a dot product operation with the feature of the object to be recommended to obtain the interest degree of the target user in the object to be recommended.

[0140] In one embodiment, the determination method of the interest degree prediction model includes steps S5064 to S5066.

[0141] Step S5064: Use the attention prediction model to perform weighted summation on the sample target feature and the sample object feature to obtain the sample target weighted feature; the sample target feature is obtained by performing multi-interest feature extraction on the sample user data through the multi-interest feature extraction model, and the sample target feature includes the second number of sample second interest features.

[0142] Among them, the attention prediction model is the Label-aware Attention Layer. During the training process of the interest degree prediction model, it performs weighted summation on the sample target feature and the determined sample object feature to obtain the sample target weighted feature. The specific calculation method of the weighted summation is the same as that in the above embodiment and will not be elaborated here.

[0143] It should be noted that during the training of the interest degree prediction model, the multi-interest feature extraction model is already a pre-trained model. Therefore, the sample target feature can be the obtained public dataset, or the feature obtained after the sample user data is input into the multi-interest feature extraction model, that is, the sample target feature can be obtained by performing multi-interest feature extraction on the sample user data through the multi-interest feature extraction model, and the sample target feature includes the second number of sample second interest features.

[0144] Step S5065: Use the feature mapping model to extract the sample user mapping feature of the sample target weighted feature, the second positive sample feature of the positive sample of the sample target weighted feature, and the second negative sample feature of the negative sample of the sample target weighted feature, and optimize the feature mapping model with the goal that the sample user mapping feature is close to the second positive sample feature and the sample object feature is far from the second negative sample feature until the training end condition is reached, and obtain the trained feature mapping model.

[0145] In one embodiment, in order to improve the distinction between user vectors of different target users and improve the accuracy of interest determination, the feature mapping model is similar to the feature extraction model. The feature mapping model uses a contrastive learning model, and the model structure of the feature mapping model can be the same as the model structure of the feature extraction model, that is, a fully connected network with regularization processing. The type of regularization processing can be set according to actual technical needs. In one embodiment, Dropout is used, that is, the feature mapping model is a fully connected network with Dropout. The contrastive learning method of the feature mapping model can also be called contrastive learning based on target weighted features.

[0146] Among them, contrastive learning needs to construct positive samples and negative samples of sample target weighted features, the positive sample is set to the sample target weighted feature itself, the negative sample is set to another randomly selected sample target weighted feature, the extracted features of the sample target weighted feature are called sample mapping features, the features of the positive samples of the sample target weighted feature are called second positive sample features, and the features of the negative samples of the sample target weighted feature are called second negative sample features. Specifically, the sample user mapping features of the sample target weighted feature, the second positive sample features of the positive samples of the sample target weighted feature, and the second negative sample features of the negative samples of the sample target weighted feature are extracted through the feature mapping model.

[0147] Specifically, the goal of contrastive learning is to shorten the distance between similar sample target weighted features and to increase the distance between dissimilar sample target weighted features. That is, the feature mapping model is optimized with the goal of making the sample user mapping feature close to the second positive sample feature and the sample object feature far from the second negative sample feature until the training end condition is reached, thereby obtaining the trained feature mapping model.

[0148] The training end condition can be set according to actual technical needs. In one embodiment, it can be set to the loss convergence of the test set. In another embodiment, the training end condition can also be set to reaching a preset number of iterations. After reaching the training end condition, the loss function of the trained feature mapping model is expressed as follows:

[0149]

[0150] in, h i Indicates the product of the sample user mapping feature and the second positive sample feature, h′ i h for expansion and replication i , represents the second negative sample feature, and τ is a hyperparameter.

[0151] Step S5066, determining the interest prediction model through the attention prediction model and the trained feature mapping model.

[0152] Among them, the interest prediction model determined by the attention prediction model and the trained feature mapping model includes the attention prediction model and the trained feature mapping model.

[0153] Based on the above-described embodiments, in the process of training the interest prediction model, the feature mapping model can be combined to perform feature mapping on the target weighted features of each user, and the loss of the trained feature mapping model is used as the auxiliary loss of the attention prediction model. That is, the interest prediction model includes the attention prediction model and the trained feature mapping model, which improves the accuracy of interest determination.

[0154] In one embodiment, after the training of the feature mapping model in the interest prediction model is completed, when using the trained interest prediction model to calculate the second number of second interest features and the to-be-recommended object features of the to-be-recommended object, specifically: perform weighted summation on the second number of second interest features and the to-be-recommended object features of the to-be-recommended object to obtain the target weighted feature; input the obtained target weighted feature into the trained feature mapping model for processing, and perform a dot product operation on the target weighted feature obtained after the processing of the feature mapping model and the to-be-recommended object features of the to-be-recommended object to obtain the interest of the target user in the to-be-recommended object.

[0155] For ease of understanding, the above calculation process will be described in conjunction with the accompanying drawings in this embodiment. As Figure 7 shown is a schematic diagram of the interest prediction model, where the interest prediction model includes an attention prediction model and a trained feature mapping model. The target features of the target user include the second number of interest features 1, 2... K. By performing operations on each interest feature of the user and the to-be-recommended object features X of the to-be-recommended object through the attention prediction model, the weights α1, α2... α k corresponding to each interest feature of the user are determined, and then the normalized weights α′1, α′2... α′ k are determined. According to the interest features and their corresponding normalized weights, perform weighted summation on each interest feature and the to-be-recommended object features to obtain the target weighted feature U. Use the trained feature mapping model to process the target weighted feature U to obtain the processed target weighted feature U, and perform a dot product operation on the processed target weighted feature U and the to-be-recommended object features X of the to-be-recommended object to obtain the interest P of the target user in the to-be-recommended object.

[0156] Among them, when training the feature mapping model, by constructing positive and negative samples of the target weighted feature U of the sample user, the user mapping feature U1 of the target weighted feature U, the user mapping feature U' corresponding to the positive sample of the target weighted feature, and the target weighted feature Y corresponding to the negative sample of the target weighted feature are extracted through the feature mapping model, and the feature mapping model is optimized with the goal that the user mapping feature U1 is close to the target weighted feature U' and the user mapping feature U1 is far from the target weighted feature Y until the training end condition is reached, and the trained feature mapping model is obtained.

[0157] Step S5064, according to the interest degrees of each object to be recommended, screen and determine the target recommended object to be recommended to the target user from each object to be recommended.

[0158] In one embodiment, the interest degree can represent the probability that the target user clicks and browses the object to be recommended. Therefore, according to the interest degrees of each object to be recommended, the target recommended object to be recommended to the target user can be screened and determined from each object to be recommended. Among them, any implementable method can be used for the screening process. Considering the actual situation, the target recommended object recommended to the target user can be one or more.

[0159] Using the information processing method of the above embodiment, through the multi-interest feature extraction module, multiple distinguishable interest features of the target object can be extracted. Thus, when determining the target recommended object to be recommended to the target user through multiple interest features, the type of the recommended object can be prevented from being too single, the effect of interest recommendation for the target user can be improved, and further the user experience can be improved.

[0160] In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0161] Taking the interest recommendation to the target user as an example, in the traditional technology, the deep recall model DSSM (Deep Structured Semantic Models) is used to perform interest recommendation on the target user. Its principle is to perform average pooling on the vectors of the user's historical click sequence, and an inner product is performed on the vectors with the same dimension output on the user side and the item side respectively to fit the user's click interest. Specifically, as Figure 8The figure shows a schematic diagram of a traditional information processing method. Here, Q represents the historical behavior data of the target user, XQ represents the features of the historical behavior data, D represents the object to be recommended, XD represents the features of the object to be recommended, W1, W2, W3, and W4 represent the weights in the process of processing the above features, YQ represents the finally obtained interest features of the user, and YD represents the features of the object to be recommended of the finally obtained object to be recommended. The dot product of YQ and YD is calculated to fit the click interest of the target user in the object to be recommended, that is, the degree of interest. However, in this way, there is only one user-side vector, and there is a problem that the types of objects to be recommended recalled are too single.

[0162] In a specific embodiment, the information processing method provided in this embodiment mainly includes a model training stage and a model running stage. Among them, a multi-interest feature extraction model is trained through the model training stage, and the multi-interest feature extraction model is run through the model running stage, so as to realize interest recommendation to the target user.

[0163] As Figure 9 The figure shows a schematic diagram of the information processing method. Taking the training of the multi-interest feature extraction model through the model training stage as an example, the steps are as follows:

[0164] Obtain sample user data, which includes sample behavior data. The sample behavior data includes a first number of sample objects; among them, the sample behavior data includes data on sample user clicks, searches, and browsing, and the sample object is the item corresponding to the sample behavior data.

[0165] By means of contrastive learning, positive and negative samples of the sample object are constructed. Among them, the positive sample is the sample object, that is, the sample object is input into the feature extraction model twice, and the negative sample is another randomly selected sample object.

[0166] Through the feature extraction model, extract the first sample object feature A of the sample object, the first positive sample feature A' of the positive sample of the sample object, and the first negative sample feature B of the negative sample of the sample object. Among them, the feature extraction model is a fully connected network with Dropout.

[0167] With the goal that the sample object feature A is close to the first positive sample feature A' and the sample object feature A is far from the first negative sample feature B, optimize the feature extraction model until the training end condition is reached, and obtain the trained feature extraction model; among them, the training end condition can be set according to actual technical needs. In one embodiment, it is set to the convergence of the loss of the test set; the loss function of the trained feature extraction model is expressed as follows:

[0168]

[0169] Among them, h i represents the product of the first sample object feature of sample object i and the first positive sample feature of the positive sample of sample object i, h′ i represents the extended and replicated h i , represents the first negative sample feature of the negative sample of sample object i, and τ is a hyperparameter.

[0170] After obtaining the trained feature extraction model, input the first number of sample objects into the trained feature extraction model again, and extract the second sample object features of the first number of sample objects through the trained feature extraction model; among them, the feature extraction model also includes an Embedding layer to convert the second sample object features into corresponding vector sequences, and the vector sequences are expressed as:

[0171]

[0172] where, I u represents the sample object, represents the second sample object feature, E U represents the vector sequence.

[0173] Construct a multi-interest extraction model, where the multi-interest extraction model is the Multi-Interest ExtractorLayer, and in one embodiment, a capsule network is selected. Input the first number of second sample object features into the multi-interest extraction model, and perform multi-interest extraction on the first number of second sample object features through the multi-interest extraction model to obtain the second number of sample interest features; based on the similarity between the second number of sample interest features and the corresponding second sample object features, optimize the multi-interest extraction model until the optimization end condition is reached to obtain the trained multi-interest extraction model.

[0174] In one embodiment, taking the first iteration of the multi-interest extraction model as an example, represent the input m groups of second sample object features as l represents the l-th layer, and represent the sample interest features output by the capsule network as Calculate the routing weight b between two layers of the capsule network ij , where the calculation formula is:

[0175]

[0176] where, S ij represents the learnable fully connected matrix, represents the second sample object feature processed by the fully connected matrix.

[0177] According to the routing weight b ij for Perform weighted summation to calculate the candidate output vector of the high-level capsule in the capsule network The calculation formula is expressed as:

[0178]

[0179] where w ij represents the routing weight after normalization between two layers, and the calculation formula is:

[0180]

[0181] After determining the candidate output vector of the capsule network perform processing on it using the non-linear mapping squash function to obtain the output of the high-level capsule The calculation formula is:

[0182]

[0183] Through multiple iterative calculations of the multi-interest extraction model, optimize the multi-interest extraction model. When the optimization end condition is reached, obtain the trained multi-interest extraction model. The multi-interest feature extraction model includes the trained feature extraction model and the trained multi-interest extraction model.

[0184] Taking the interest recommendation for the target user through the multi-interest feature extraction model in the model operation stage as an example, the steps are as follows:

[0185] Obtain the target data of the target user. The target data includes historical behavior data and attribute data. The historical behavior data includes the first number of behavior objects; among them, the behavior object is the item clicked by the target user, Figure 9 which are represented as item 1, item 2... item N in the text. The attribute data includes gender, age, the city where the user is located, and the city level of the city where the user is located.

[0186] Perform multi-interest feature extraction on the target data through the multi-interest feature extraction model to obtain the target features of the target user. The target features include the second number of second interest features.

[0187] Among them, the multi-interest feature extraction includes: extracting the behavior object features of item 1, item 2... item N through the feature extraction model in the multi-interest feature extraction model to obtain N (the first number) of behavior object features. Taking item 1 as an example, input the information of item 1 into the fully connected network after regularization processing to extract the corresponding behavior object feature.

[0188] The multi-interest extraction model in the multi-interest feature extraction model performs multi-interest extraction on the first number of behavioral object features to obtain the second number of first interest features; that is, after multi-interest extraction, multi-interest feature vectors 1, feature vector 2... feature vector K are obtained.

[0189] The multi-interest feature extraction model extracts the attribute features of the attribute data, and fuses the second number of first interest features with the attribute features respectively to obtain the target features of the target user; in one embodiment, the feature fusion method is to splice the first interest feature and the attribute feature respectively, and after being processed by two fully connected layers, the target features of the target user are obtained. The target features include the second number of second interest features. In one embodiment, the activation function of the two fully connected layers is the Relu function.

[0190] According to the target features, determine the target recommended objects to be recommended to the target user; where the determination methods here include two types. One is to use the recall method to recall the recommended objects with similar semantics to the target features for each target feature respectively. The other is to use the prediction method to predict the interest degree of the target user for the recommended objects according to the target features. The following will explain the two methods respectively.

[0191] In one embodiment, if the recall method is adopted, the Faiss retrieval method can be used to perform similarity matching between each second interest feature and the recommended object features of each recommended object to obtain the recommended objects matched by each second interest feature; perform post-processing steps such as screening and sorting on the recommended objects matched by each second interest feature to obtain the target recommended objects to be recommended to the target user, where the target recommended objects can be one or more.

[0192] In one embodiment, if the prediction method is adopted, the interest degree prediction model can be used to calculate the second number of second interest features and the recommended object features of the recommended object to obtain the interest degree of the target user for the recommended object; where the interest degree prediction model includes an attention prediction model and a trained feature mapping model, and the attention prediction model is the Label-aware Attention Layer.

[0193] Specifically, the attention prediction model is used to perform weighted summation on the second number of second interest features and the recommended object features of the recommended object to obtain the target weighted features The above calculation process can be expressed as:

[0194]

[0195] Among them, represents the recommended object feature, V uIt represents the second interest feature, p is a hyperparameter, and softmax represents the normalization process.

[0196] The trained feature mapping model is used to process the target weighted feature to obtain the processed target weighted feature; and the dot product operation is performed between the processed target weighted feature and the feature of the object to be recommended of the object to be recommended respectively to obtain the interest degree of the target user in the object to be recommended. Among them, the determination method of the interest degree prediction model includes: performing weighted summation on the sample target feature and the sample object feature through the attention prediction model to obtain the sample target weighted feature; the sample target feature is obtained by extracting multi-interest features from the sample user data through the multi-interest feature extraction model, and the sample target feature includes the second number of sample second interest features; constructing positive and negative samples of the sample target feature, the positive sample is the sample target feature, and the negative sample is another randomly selected sample target weighted feature. The sample user mapping feature of the sample target weighted feature, the second positive sample feature of the positive sample of the sample target weighted feature, and the second negative sample feature of the negative sample of the sample target weighted feature are extracted through the feature mapping model, and the goal is that the sample user mapping feature is close to the second positive sample feature and the sample object feature is far from the second negative sample feature. Optimize the feature mapping model until the training end condition is reached to obtain the trained feature mapping model; through the attention prediction model and the trained feature mapping model, the interest degree prediction model can be obtained.

[0197] It should be understood that although the steps in the respective flowcharts involved in the above embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the respective flowcharts involved in the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or steps or stages in other steps.

[0198] Based on the same inventive concept, the present application also provides an information processing device for implementing the information processing method involved in the above embodiments. The implementation solutions provided by this device to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more of the following information processing device embodiments can refer to the limitations on the information processing method in the above text and will not be repeated here.

[0199] In one of the embodiments, as Figure 10As shown, an information processing device 10 is provided. This device can be a software module, a hardware module, or a combination of both to form a part of a computer device. Specifically, the device includes: a sample data acquisition module 1010, a feature extraction model training module 1020, a sample feature extraction module 1030, a multi-interest extraction model training module 1040, and a model determination module 1050, where:

[0200] The sample data acquisition module 1010 is used to acquire sample user data, and the sample user data includes sample behavior data, and the sample behavior data includes a first number of sample objects.

[0201] The feature extraction model training module 1020 is used to extract the first sample object features of the sample objects, the first positive sample features of the positive samples of the sample objects, and the first negative sample features of the negative samples of the sample objects through the feature extraction model, and optimize the feature extraction model with the goal that the sample object features are similar to the first positive sample features and the sample object features are far from the first negative sample features until the training end condition is reached, and obtain the trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object.

[0202] The sample feature extraction module 1030 is used to extract the second sample object features of the first number of sample objects through the trained feature extraction model.

[0203] The multi-interest extraction model training module 1040 is used to perform multi-interest extraction on the first number of the second sample object features through the multi-interest extraction model to obtain a second number of sample interest features, and optimize the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached, and obtain the trained multi-interest extraction model.

[0204] The model determination module 1050 is used to obtain a multi-interest feature extraction model through the trained feature extraction model and the trained multi-interest extraction model.

[0205] In one embodiment, the feature extraction model training module 1020 is used to input the sample object into the feature extraction model to obtain initial sample object features, and the feature extraction model is a fully connected network after regularization processing; convert the initial sample object features into features of a predetermined dimension vector to obtain the first sample object features.

[0206] In one embodiment, as Figure 11As shown, an information processing device 11 is provided. This device can be a software module, a hardware module, or a combination of both to form part of a computer device. Specifically, the device includes: a target data acquisition module 1110, a target feature extraction module 1120, and a recommended object determination module 1130, where:

[0207] The target data acquisition module 1110 is used to acquire the target data of the target user. The target data includes historical behavior data and attribute data, and the historical behavior data includes a first number of behavior objects.

[0208] The target feature extraction module 1120 is used to perform multi-interest feature extraction on the target data through a multi-interest feature extraction model to obtain the target features of the target user. The multi-interest feature extraction includes: extracting the behavior object features of the first number of the behavior objects through the feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of the behavior object features through the multi-interest extraction model in the multi-interest feature extraction model to obtain the second number of first interest features, and extracting the attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of the first interest features with the attribute features respectively to obtain the target features of the target user. The target features include the second number of second interest features.

[0209] The recommended object determination module 1130 is used to determine the target recommended objects to be recommended to the target user according to the target features.

[0210] In one embodiment, the information processing device 11 further includes the information processing device 10.

[0211] In one embodiment, the recommended object determination module 1130 is used to perform similarity matching between each of the second interest features and the to-be-recommended object features of each to-be-recommended object to obtain the to-be-recommended objects matched by each of the second interest features; perform post-processing on the to-be-recommended objects matched by each of the second interest features to obtain the target recommended objects to be recommended to the target user.

[0212] In one embodiment, the recommended object determination module 1130 is further configured to perform an operation on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object to obtain the degree of interest of the target user in the to-be-recommended object. The operation includes: performing a weighted sum on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object to obtain a target weighted feature; and performing a dot product operation on the target weighted feature and the to-be-recommended object features of the to-be-recommended object to obtain the degree of interest of the target user in the to-be-recommended object; and screening and determining a target recommended object to be recommended to the target user from each of the to-be-recommended objects according to the degree of interest of each of the to-be-recommended objects.

[0213] In one embodiment, the recommended object determination module 1130 is further configured to perform an operation on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object through an interest degree prediction model to obtain the degree of interest of the target user in the to-be-recommended object.

[0214] In one embodiment, the recommended object determination module 1130 further includes: an interest degree prediction model determination module.

[0215] In one embodiment, the interest degree prediction model determination module is configured to perform a weighted sum on the sample target feature and the sample object feature through an attention prediction model to obtain a sample target weighted feature; the sample target feature is obtained by performing multi-interest feature extraction on sample user data through the multi-interest feature extraction model, and the sample target feature includes the second number of sample second interest features; extracting the sample user mapping feature of the sample target weighted feature, the second positive sample feature of the positive sample of the sample target weighted feature, and the second negative sample feature of the negative sample of the sample target weighted feature through a feature mapping model, and taking the sample user mapping feature being close to the second positive sample feature and the sample object feature being far from the second negative sample feature as the goal to optimize the feature mapping model until the training end condition is reached to obtain a trained feature mapping model; processing the sample target weighted feature through the trained feature mapping model to obtain a processed sample target weighted feature; performing a dot product operation on each of the processed sample target weighted features and the sample object feature respectively to obtain corresponding sample interest degrees; and determining the interest degree prediction model through the attention prediction model and the trained feature mapping model.

[0216] For the specific limitations of the information processing device, reference can be made to the limitations on the information processing method in the foregoing text, which will not be elaborated here. Each module in the above information processing device can be implemented in whole or in part by software, hardware, or a combination thereof. Each of the above modules can be embedded in the processor in the computer device in hardware form or be independent of it, or can be stored in the memory in the computer device in software form, so as to facilitate the processor to call and execute the operations corresponding to each of the above modules.

[0217] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 12 shown. The computer device includes a processor, a memory, and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store information processing data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements an information processing method.

[0218] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 13 shown. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an information processing method. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the computer device housing, or an external keyboard, touchpad, or mouse, etc.

[0219] Those skilled in the art can understand that Figure 12 and Figure 13The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have a different component layout.

[0220] In one embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0221] In one embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0222] In one embodiment, a computer program product or a computer program is provided. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the steps in the above method embodiments.

[0223] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application, such as sample behavior data, target data, etc., are all information and data authorized by the user or fully authorized by all parties.

[0224] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0225] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0226] The above embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. An information processing method, characterized in that, The method includes: Obtaining sample user data, where the sample user data includes sample behavior data, and the sample behavior data includes a first number of sample objects; Extracting a first sample object feature of the sample object, a first positive sample feature of the positive sample of the sample object, and a first negative sample feature of the negative sample of the sample object through a feature extraction model, and optimizing the feature extraction model with the goal that the sample object feature is similar to the first positive sample feature and the sample object feature is far from the first negative sample feature until the training end condition is reached, to obtain a trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object; Extracting a second sample object feature of the first number of sample objects through the trained feature extraction model; Performing multi-interest extraction on the first number of the second sample object features through a multi-interest extraction model to obtain a second number of sample interest features, and optimizing the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached, to obtain a trained multi-interest extraction model; Obtaining a multi-interest feature extraction model through the trained feature extraction model and the trained multi-interest extraction model.

2. The method according to claim 1, characterized in that, The extracting a first sample object feature of the sample object through the feature extraction model includes: Inputting the sample object into the feature extraction model to obtain an initial sample object feature, where the feature extraction model is a fully connected network after regularization processing; Converting the initial sample object feature into a feature of a predetermined dimension vector to obtain the first sample object feature.

3. An information processing method, characterized in that, The method includes: Obtaining target data of a target user, where the target data includes historical behavior data and attribute data, and the historical behavior data includes a first number of behavior objects; Performing multi-interest feature extraction on the target data through a multi-interest feature extraction model to obtain a target feature of the target user; the multi-interest feature extraction includes: extracting behavior object features of the first number of behavior objects through the feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of behavior object features through the multi-interest extraction model in the multi-interest feature extraction model to obtain the second number of first interest features, and extracting attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of the first interest features with the attribute features respectively to obtain the target feature of the target user, where the target feature includes the second number of second interest features; the multi-interest feature extraction model is obtained by using the method described in claim 1; Determining a target recommendation object to be recommended to the target user according to the target feature.

4. The method according to claim 3, wherein The fusing the second number of the first interest features with the attribute features respectively to obtain the target feature of the target user includes: Concatenate each of the second number of the first interest features with the attribute feature to obtain the target feature of the target user.

5. The method according to claim 3, characterized in that Determining the target recommendation object to be recommended to the target user according to the target feature includes: Perform similarity matching between each of the second interest features and the to-be-recommended object features of each to-be-recommended object to obtain the to-be-recommended objects matched by each of the second interest features; Perform post-processing on the to-be-recommended objects matched by each of the second interest features to obtain the target recommendation object to be recommended to the target user.

6. The method according to claim 3, characterized in that, Determining the target recommendation object to be recommended to the target user according to the target feature further includes: Perform an operation on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object to obtain the interest degree of the target user in the to-be-recommended object. The operation includes: performing weighted summation on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object to obtain a target weighted feature; and performing a dot product operation on the target weighted feature and the to-be-recommended object features of the to-be-recommended object to obtain the interest degree of the target user in the to-be-recommended object; According to the interest degree of each to-be-recommended object, screen and determine the target recommendation object to be recommended to the target user from each to-be-recommended object.

7. The method according to claim 6, characterized in that, Performing an operation on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object to obtain the interest degree of the target user in the to-be-recommended object includes: Perform an operation on each of the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object through an interest degree prediction model to obtain the interest degree of the target user in the to-be-recommended object; The determination method of the interest degree prediction model includes: Perform weighted summation on the sample target feature and the sample object feature through an attention prediction model to obtain a sample target weighted feature; the sample target feature is obtained by performing multi-interest feature extraction on sample user data through the multi-interest feature extraction model, and the sample target feature includes the second number of sample second interest features; Extract the sample user mapping feature of the sample target weighted feature, the second positive sample feature of the positive sample of the sample target weighted feature, and the second negative sample feature of the negative sample of the sample target weighted feature through a feature mapping model, and optimize the feature mapping model with the goal that the sample user mapping feature is close to the second positive sample feature and the sample object feature is far from the second negative sample feature until the training end condition is reached to obtain the trained feature mapping model; Determine the interest degree prediction model through the attention prediction model and the trained feature mapping model.

8. The method according to claim 7, wherein Performing an operation on each of the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object through the interest degree prediction model to obtain the interest degree of the target user in the to-be-recommended object includes: Perform weighted summation on the second number of the second interest features and the to-be-recommended object features of the to-be-recommended object to obtain a target weighted feature; Input the obtained target weighted features into the trained feature mapping model for processing; Perform a dot product operation on the target weighted features obtained after processing by the feature mapping model and the features of the object to be recommended of the object to be recommended, to obtain the degree of interest of the target user in the object to be recommended.

9. An information processing apparatus, characterized in that, The device includes: A sample data acquisition module, configured to acquire sample user data, where the sample user data includes sample behavior data, and the sample behavior data includes a first number of sample objects; A feature extraction model training module, configured to extract the first sample object features of the sample objects, the first positive sample features of the positive samples of the sample objects, and the first negative sample features of the negative samples of the sample objects through a feature extraction model, and optimize the feature extraction model with the goal that the sample object features are similar to the first positive sample features and the sample object features are far from the first negative sample features until the training end condition is reached, to obtain a trained feature extraction model; the positive sample is the sample object, and the negative sample is another randomly selected sample object; A sample feature extraction module, configured to extract the second sample object features of the first number of sample objects through the trained feature extraction model; A multi-interest extraction model training module, configured to perform multi-interest extraction on the first number of the second sample object features through a multi-interest extraction model, to obtain a second number of sample interest features, and optimize the multi-interest extraction model based on the similarity between the second number of sample interest features and the corresponding second sample object features until the optimization end condition is reached, to obtain a trained multi-interest extraction model; A model determination module, configured to obtain a multi-interest feature extraction model through the trained feature extraction model and the trained multi-interest extraction model.

10. An information processing apparatus, characterized in that, The device includes: A target data acquisition module, configured to acquire target data of a target user, where the target data includes historical behavior data and attribute data, and the historical behavior data includes a first number of behavior objects; A target feature extraction module, configured to perform multi-interest feature extraction on the target data through a multi-interest feature extraction model, to obtain the target features of the target user; the multi-interest feature extraction includes: extracting the behavior object features of the first number of behavior objects through the feature extraction model in the multi-interest feature extraction model, performing multi-interest extraction on the first number of behavior object features through the multi-interest extraction model in the multi-interest feature extraction model, to obtain the second number of first interest features, and extracting the attribute features of the attribute data through the multi-interest feature extraction model, and fusing the second number of the first interest features with the attribute features respectively, to obtain the target features of the target user, where the target features include the second number of second interest features; the multi-interest feature extraction model is obtained by using the method described in claim 1; A recommended object determination module, configured to determine a target recommended object to be recommended to the target user according to the target features.

11. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

12. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images