Interest feature data acquisition method and device, electronic equipment and storage medium

Abstract

The present disclosure relates to an interest feature data acquisition method and device, electronic equipment and storage medium, and belongs to the field of computers. The method comprises: acquiring first interest feature data and second interest feature data, the first interest feature data being used to represent the interest features of a target account in a first time period, the second interest feature data being used to represent the interest features of the target account in a second time period, the first time period being longer than the second time period; determining an interest tendency weight using the first interest feature data and the second interest feature data, the interest tendency weight being used to represent the degree of each interest tendency of the target account; determining third interest feature data based on the interest tendency weight and pre-constructed interest space data, the interest space data being used to represent the inherent characteristics of each interest in a system interest set, and the third interest feature data being used to represent the overall interest of the target account. The present disclosure enriches interest feature data and is conducive to the diversity of interest features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, and particularly relates to an interest feature data acquisition method and device, electronic equipment and storage medium. BACKGROUND

[0002] In a recommendation system, whether the training of a recommendation model or the inference of the recommendation model needs to input interest feature data, which directly affects the performance of the recommendation model.

[0003] The existing interest feature data acquisition method is usually based on historical sequence data or interest labels of display records, which are often single and cannot fully represent the interest features, directly affecting the performance of the recommendation model. SUMMARY

[0004] The present application provides an interest feature data acquisition method to at least solve the problem of single interest feature data in the related art.

[0005] According to a first aspect of the embodiments disclosed in the present application, an interest feature data acquisition method is provided, comprising:

[0006] acquiring first interest feature data and second interest feature data, wherein the first interest feature data is used to represent the interest features of a target account in a first time period, the second interest feature data is used to represent the interest features of the target account in a second time period, the first time period is greater than the second time period,

[0007] determining an interest tendency weight by using the first interest feature data and the second interest feature data, the interest tendency weight is used to represent the degree of each interest tendency of the target account,

[0008] determining third interest feature data based on the interest tendency weight and interest space data, wherein the interest space data is used to represent the inherent characteristics of each interest in an interest set, and the third interest feature data is used to represent the overall interest of the target account.

[0009] As a possible implementation, the determination of the interest tendency weight by using the first interest feature data and the second interest feature data comprises:

[0010] determining the interest tendency weight according to the account attribute feature data of the target account and the historical interest sequence feature information,

[0011] wherein,

[0012] the account attribute feature data belongs to the first interest feature data,

[0013] the historical interest sequence feature information belongs to the second interest feature data.

[0014] As a possible implementation, the interest tendency weight is determined according to the account attribute feature data of the target account and the historical interest sequence feature information, and includes:

[0015] The account attribute feature data and the historical interest sequence feature information are input into a first neural network model to obtain an interest tendency weight vector of the target account, wherein an element value at each position in the interest tendency weight vector represents an interest tendency weight corresponding to each position.

[0016] As a possible implementation, the method further includes:

[0017] The target interest attribute feature data and the interest embedding feature data are obtained, wherein the target interest attribute feature data is used to represent the attribute feature of the target interest; and the interest embedding feature data is used to represent at least one of the basic feature of the target account, the basic feature of the target interest and the context feature.

[0018] The third interest feature data, the first interest feature data, the second interest feature data, the target interest attribute feature data and the interest embedding feature data are fused in the vector dimension to obtain an input feature vector.

[0019] As a possible implementation, the interest space data is pre-constructed in the following manner:

[0020] An interest base vector is set for each interest to represent the feature of the interest itself,

[0021] Each interest base vector is taken as a first dimension vector, and all interest base vectors are taken as a second dimension vector to obtain an interest space vector composed of the first dimension vector and the second dimension vector, wherein the dimension number of the first dimension vector is the dimension number of the interest base vector, the dimension number of the second dimension vector is the total number of the interest base vectors, and the interest space vector is the interest space data.

[0022] The vector values in the interest space vector are randomly and uniformly initialized.

[0023] As a possible implementation, the third interest feature data is determined based on the interest tendency weight and the interest attribute feature data, and includes:

[0024] For each interest of the target account, the interest base vector of the each interest of the target account and the interest tendency weight of the each interest are fused to obtain a fused interest base vector,

[0025] The average value of all fused interest base vectors is calculated to obtain a third interest feature vector of the target account, wherein the third interest feature vector is the third interest feature data.

[0026] As a possible implementation, the method further comprises:

[0027] inputting the obtained first interest feature data and second interest feature data, and third interest feature data as input data into the interest recommendation model,

[0028] performing interest recommendation for the target account according to the prediction result output by the interest recommendation model.

[0029] As a possible implementation, the method is implemented through an interest recommendation model, and a training method of the interest recommendation model comprises:

[0030] obtaining first interest feature sample data and second interest feature sample data;

[0031] determining sample interest tendency weights by using the first interest feature sample data and the second interest feature sample data;

[0032] determining third interest feature sample data based on the sample interest tendency weights and sample interest space data;

[0033] inputting the obtained first interest feature sample data, second interest feature sample data and third interest feature sample data as input sample data into the interest recommendation model;

[0034] determining a target loss function value according to a prediction result output by the interest recommendation model,

[0035] updating model parameters of the interest recommendation model according to the target loss function value until a training end condition is met.

[0036] As a possible implementation, the method further comprises:

[0037] obtaining target interest attribute feature sample data and interest embedding feature sample data;

[0038] performing vector dimension fusion on the third interest feature sample data, the first interest feature sample data, the second interest feature sample data, the target interest attribute feature sample data and the interest embedding feature sample data to obtain an input sample feature vector of the interest recommendation model.

[0039] As a possible implementation, the determination of the sample interest tendency weights by using the first interest feature sample data and the second interest feature sample data comprises:

[0040] determining the sample interest tendency weights according to account attribute feature sample data and historical interest sequence feature sample information of a sample account,

[0041] wherein,

[0042] The account attribute feature sample data belongs to first interest feature sample data,

[0043] The historical interest sequence feature sample information belongs to second interest feature sample data.

[0044] As a possible implementation, the sample interest tendency weight is determined according to the account attribute feature sample data and the historical interest sequence feature sample information of the sample account, including:

[0045] The account attribute feature sample data and the historical interest sequence feature sample information are input into the first neural network model to obtain a sample interest tendency weight vector of the sample account, wherein an element value at each position in the sample interest tendency weight vector represents a sample interest tendency weight corresponding to each position.

[0046] As a possible implementation, the sample interest space data is pre-constructed in the following manner:

[0047] A sample interest base vector for representing the characteristics of each interest is set,

[0048] Each sample interest base vector is taken as a first sample dimension vector, and all sample interest base vectors are taken as a second sample dimension vector to obtain a sample interest space vector composed of the first sample dimension vector and the second sample dimension vector, wherein the dimension of the first sample dimension vector is the dimension of the sample interest base vector, the dimension of the second sample dimension vector is the total number of sample interest base vectors, and the sample interest space vector is the sample interest space data.

[0049] The vector values in the sample interest space vector are randomly and uniformly initialized.

[0050] As a possible implementation, the third interest feature sample data is determined based on the sample interest tendency weight and the interest attribute feature sample data, including:

[0051] For each interest of the sample account, a sample interest base vector of the interest of the sample account and a sample interest tendency weight of the interest are fused to obtain a sample fusion interest base vector,

[0052] An average value of all sample fusion interest base vectors is calculated to obtain a third sample interest feature vector of the sample account, wherein the third sample interest feature vector is the third interest feature sample data.

[0053] As a possible implementation, the model parameters of the interest recommendation model are updated according to the target loss function value, including:

[0054] The model parameters in the interest recommendation model and the first neural network model are updated according to the target loss function value, and the sample interest space vector remains unchanged.

[0055] According to a second aspect of the embodiments of the present disclosure, an interest feature data acquisition device is provided, comprising:

[0056] An acquisition module configured to acquire first interest feature data and second interest feature data of a target account, wherein the first interest feature data is used to represent interest features of the target account in a first time period, the second interest feature data is used to represent interest features of the target account in a second time period, the first time period is greater than the second time period,

[0057] An interest tendency weight determination module configured to determine an interest tendency weight by using the first interest feature data and the second interest feature data, the interest tendency weight is used to represent a degree of each interest tendency of the target account,

[0058] An interest feature determination module configured to determine third interest feature data based on the interest tendency weight and interest space data, wherein the interest space data is used to represent intrinsic features of each interest in an interest set, the third interest feature data is used to represent overall interest of the target account.

[0059] As a possible implementation, the interest tendency weight determination module is configured to determine the interest tendency weight according to account attribute feature data of the target account and historical interest sequence feature information,

[0060] wherein,

[0061] The account attribute feature data belongs to the first interest feature data,

[0062] The historical interest sequence feature information belongs to the second interest feature data.

[0063] As a possible implementation, the acquisition module is further configured to acquire target interest attribute feature data and interest embedding feature data, wherein the target interest attribute feature data is used to represent attribute features of a target interest; the interest embedding feature data is used to represent at least one of basic features of the target account, basic features of the target interest and context features;

[0064] The third interest feature data, the first interest feature data, the second interest feature data, the target interest attribute feature data and the interest embedding feature data are fused in vector dimension to obtain an input feature vector.

[0065] As a possible implementation, the interest tendency weight determination module is configured to:

[0066] The account attribute feature data and the historical interest sequence feature information are input into the first neural network model to obtain an interest tendency weight vector of the target account, wherein an element value at each position in the interest tendency weight vector represents an interest tendency weight corresponding to each position.

[0067] As a possible implementation, the apparatus further comprises:

[0068] The interest space construction module is configured to pre-set, for each interest, an interest base vector for representing a feature of the interest itself,

[0069] Each interest base vector is taken as a first dimension vector, and all interest base vectors are taken as a second dimension vector to obtain an interest space vector composed of the first dimension vector and the second dimension vector, wherein a dimension of the first dimension vector is a dimension of the interest base vector, a dimension of the second dimension vector is a third number of the interest base vector, and the interest space vector is interest space data.

[0070] The vector values in the interest space vector are randomly and uniformly initialized.

[0071] As a possible implementation, the interest feature determination module is configured to:

[0072] For each interest of the target account, the interest base vector of the each interest of the target account and the interest tendency weight of the each interest are fused to obtain a fused interest base vector.

[0073] An average value of all fused interest base vectors is calculated to obtain a third interest feature vector of the target account, wherein the third interest feature vector is third interest feature data.

[0074] As a possible implementation, the apparatus further comprises:

[0075] The recommendation module is configured to input the obtained first interest feature data and second interest feature data and the third interest feature data as input data into an interest recommendation model, and perform interest recommendation for the target account according to a prediction result output by the interest recommendation model.

[0076] As a possible implementation, the apparatus is implemented through an interest recommendation model, and a training apparatus of the interest recommendation model comprises:

[0077] The sample data acquisition module is configured to acquire first interest feature sample data and second interest feature sample data of a sample account,

[0078] The sample interest tendency weight determination module is configured to determine sample interest tendency weights by using the first interest feature sample data and the second interest feature sample data.

[0079] The sample interest feature determination module is configured to determine third interest feature sample data based on the sample interest tendency weight and the sample interest space data,

[0080] The sample recommendation module is configured to input the obtained first interest feature sample data, the second interest feature sample data and the third interest feature sample data as input sample data into the interest recommendation model,

[0081] The training module is configured to determine a target loss function value according to a prediction result output by the interest recommendation model corresponding to the input sample data, update model parameters of the interest recommendation model according to the target loss function value, and stop until a training end condition is met.

[0082] As a possible implementation, the sample acquisition module is further configured to:

[0083] acquire target interest attribute feature sample data and interest embedding feature sample data;

[0084] vector dimension fusion is performed on the third interest feature sample data, the first interest feature sample data, the second interest feature sample data, the target interest attribute feature sample data and the interest embedding feature sample data to obtain an input sample feature vector of the interest recommendation model.

[0085] As a possible implementation, the sample interest tendency weight determination module is configured to:

[0086] determine the sample interest tendency weight according to account attribute feature sample data of a sample account and historical interest sequence feature sample information,

[0087] wherein,

[0088] the account attribute feature sample data belongs to the first interest feature sample data,

[0089] the historical interest sequence feature sample information belongs to the second interest feature sample data.

[0090] As a possible implementation, the sample interest tendency weight determination module is configured to:

[0091] input the account attribute feature sample data and the historical interest sequence feature sample information into a first neural network model to obtain a sample interest tendency weight vector of the sample account, wherein an element value at each position in the sample interest tendency weight vector represents a sample interest tendency weight corresponding to each position.

[0092] As a possible implementation, the training device further comprises:

[0093] The sample interest space construction module is configured to pre-set, for each interest, a sample interest base vector for representing characteristics of the interest itself,

[0094] Each sample interest base vector is taken as a first sample dimension vector, and all sample interest base vectors are taken as a second sample dimension vector, to obtain a sample interest space vector composed of the first sample dimension vector and the second sample dimension vector, where the dimension of the first sample dimension vector is the dimension of the sample interest base vector, the dimension of the second sample dimension vector is the total number of sample interest base vectors, and the sample interest space vector is sample interest space data,

[0095] The vector values in the interest space vector are randomly and uniformly initialized.

[0096] As a possible implementation, the sample interest feature determination module is configured to:

[0097] For each interest of the sample account, the sample interest base vector of the each interest of the sample account and the sample interest tendency weight of the each interest are fused to obtain a sample fused interest base vector,

[0098] An average value of all sample fused interest base vectors is calculated to obtain a third sample interest feature vector of the sample account, where the third sample interest feature vector is third interest feature sample data.

[0099] As a possible implementation, the training module is configured to:

[0100] According to the target loss function value, the model parameters in the interest recommendation model and the first neural network model are updated, and the sample interest space vector is kept unchanged.

[0101] According to a third aspect of the embodiments of the present disclosure, an electronic device is provided, including:

[0102] a processor;

[0103] a memory for storing executable instructions of the processor;

[0104] The processor is configured to execute the executable instructions to implement the interest feature data acquisition method.

[0105] According to a fourth aspect of the embodiments of the present disclosure, a computer readable storage medium is provided, when at least one instruction in the computer readable storage medium is executed by a processor of an electronic device, the electronic device can implement the interest feature data acquisition method.

[0106] According to a fifth aspect of the embodiments of the present disclosure, a computer program product is provided, including a computer program, which is executed by a processor to implement the interest feature data acquisition method.

[0107] The technical solutions provided by the embodiments disclosed in the present application bring at least the following beneficial effects:

[0108] The embodiments of the present application pre-construct a global interest feature space, use the first interest feature data that remembers long-term interest and the second interest feature data that reflects short-term interest, capture the tendency weight value of the target account on each interest, and obtain the third interest feature data of the target account, thereby enriching the interest feature data and being beneficial to retaining the diversity of interest features. The third interest feature data can help the recommendation system better infer the preferences of the target account, thereby enhancing the recommendation effect of the recommendation system. Meanwhile, the implementation of the embodiments of the present application is simple and effective, and can improve the estimation effect of the recommendation system without bringing a large amount of time consumption.

[0109] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0110] The accompanying drawings incorporated in the specification and forming a part thereof illustrate embodiments consistent with the present disclosure and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0111] Figure 1 is a flowchart of an interest feature data acquisition method according to an exemplary embodiment;

[0112] Figure 2 is a flowchart of an interest feature data acquisition method according to an exemplary embodiment;

[0113] Figure 3 is a flowchart of an interest recommendation method using the acquired interest feature data according to an exemplary embodiment;

[0114] Figure 4 is a flowchart of an interest recommendation model training method according to an exemplary embodiment;

[0115] Figure 5 is a schematic diagram of an interest feature data acquisition device according to an exemplary embodiment;

[0116] Figure 6 is a schematic diagram of an interest recommendation device according to an exemplary embodiment;

[0117] Figure 7 is a schematic diagram of an interest recommendation model training device according to an exemplary embodiment;

[0118] Figure 8 is a schematic diagram of an electronic device according to an example embodiment. DETAILED DESCRIPTION

[0119] In order to make the ordinary person skilled in the art better understand the technical solutions disclosed in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings.

[0120] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all the embodiments consistent with the present application. Rather, they are only examples of devices and methods consistent with some aspects of the present application as detailed in the appended claims.

[0121] In existing recommendation systems, the method of constructing interest space is generally:

[0122] Method one, based on explicit record of interest label, to construct interest space. This method usually defines all interest labels in advance, and when it involves a certain interest (such as clicking on a certain video), the defined interest label is recorded. Accumulated interest labels are used as interest feature data.

[0123] Method two is similar to method one, which also requires explicit recording of interest labels, but will vectorize the recorded interest labels, that is, use a vector to represent a certain interest, and finally use interest vector clusters as interest feature data.

[0124] Method three, through historical sequence data to construct the interest space of the target account. This method is more extensive, generally using attention (Attention) neural network structure to process historical sequence data, so as to obtain interest feature vector.

[0125] Among the above three main methods, methods one and two are based on explicit record of interest label, which is limited by artificial definition of interest label, which will limit the diversity of interest and the size of interest space, resulting in that many implicit interest features cannot be expressed and obtained.

[0126] For the third mode, the general self-attention mode is to construct the relationship between interests itself, for example, to construct the relationship between videos, ignoring the association relationship between the target account and the interest; and the target-attention mode takes the target account side feature as a query, although the relationship between the target account and the interest is constructed, but only the interest tendency degree on the target account and the historical sequence data is constructed. Considering the performance problem, the dimension of the historical sequence feature vector is generally required to meet the set length threshold, if the dimension of the historical sequence feature vector is greater than the set length threshold, the target-attention can only construct the short-term interest space, and also cannot form the interest vector pool. For the long short-term memory network (LSTM) and the current memory gate network, the structure is relatively complex, time-consuming or requires a large amount of memory, which is difficult to adapt to the general recommendation model.

[0127] Therefore, the embodiment of the present application proposes an interest feature data acquisition method to construct an interest space that is conducive to improving the effect of the recommendation system.

[0128] Referring to Figure 1 , as shown in the drawings, Figure 1 is a flowchart of an interest feature data acquisition method according to an exemplary embodiment. The method comprises:

[0129] Step 101, acquiring first interest feature data and second interest feature data, wherein the first interest feature data is used to represent the interest features of the target account in the first time period, and the second interest feature data is used to represent the interest features of the target account in the second time period, and the first time period is greater than the second time period;

[0130] Step 102, determining the interest tendency weight by using the first interest feature data and the second interest feature data, wherein the interest tendency weight is used to represent the degree of each interest tendency of the target account,

[0131] Step 103, determining the third interest feature data based on the interest tendency weight and the pre-constructed interest space data, wherein the interest space data is used to represent the self-features of each interest in the interest set, and the third interest feature data is used to represent the overall interest of the target account.

[0132] The embodiment of the present application uses the first interest feature data which remembers the long-term interest features and the second interest feature data which reflects the short-term interest features, combines the pre-constructed interest space data, and mines the third interest feature data, which retains the diversity of the interest features and enriches the interest feature data.

[0133] For the convenience of understanding the embodiments of the present application, the following will be described by taking the recommendation of video as an interest as an example, and it should be understood that the present application is not limited to the feature data of video as an interest, and any interest that is concerned can be applicable.

[0134] The recommendation system often uses a neural network model for prediction, and the input of the neural network model can include account attribute data uid, video attribute data pid, and basic features of the target account, basic features and context features of the video, which are usually defined as embedding features, denoted as emb i In addition, there is historical sequence data, such as a set of watched historical videos, denoted as ActionList=[a1, a2,..., a K ], where uid, pid, emb i , and video a i are obtained by querying the vector table to obtain the vector corresponding to each feature, which is called embedding.

[0135] The recommendation system generally obtains new historical sequence feature information from the historical sequence data [a1, a2,..., a K ] through an Attention neural network structure or a SumPooling operation, denoted as O a , and combines these features together as the input of the recommendation system, denoted as input=[uid, O a , emb1, emb2,..., emb N , pid], and obtains the final prediction result pred through the recommendation system, and the final prediction result pred is represented as:

[0136] pred=δ(F(input, W1));

[0137] Wherein, F function represents a neural network model, W1 is a trainable parameter in the neural network, and δ(x) is an activation function, such as a sigmiod function.

[0138] In the recommendation system, the prediction result of the target account to each event of the candidate set video needs to be predicted for recommendation, such as predicting the like prediction probability value of the target account u to the candidate set video [p1, p2,..., p n ], and the subsequent recommendation system can sort and select the top-K videos according to the prediction probability value and send them to the target account. Whether the prediction value of the "like" event is accurate will affect the effect of the whole recommendation system. The events can also include, for example, "forward", "click", and "effective watching". The recommendation system often uses a neural network model to predict such events of the video.

[0139] See Figure 2 As shown, Figure 2 This is a flowchart illustrating a method for acquiring interest feature data according to an exemplary embodiment. The method includes:

[0140] Step 201: Pre-construct interest space data to characterize the features of each interest in the interest set.

[0141] As an example, for each interest in the set of interests, an interest basis vector is set to represent the characteristics of that interest itself. Each interest basis vector is used as the first dimension vector, and all interest basis vectors are used as the second dimension vector, resulting in an interest space vector composed of the first and second dimension vectors. Here, the dimension of the first dimension vector is the same as the dimension of the interest basis vectors, and the dimension of the second dimension vector is the same as the third dimension of the interest basis vectors. The interest space vector represents the interest space data.

[0142] For example, by treating each interest basis vector as a row vector and all interest basis vectors as column vectors, we obtain the interest space vector. The dimension of the row vector is the same as the dimension of the interest basis vectors, and the dimension of the column vector is the total number of interest basis vectors for the target account.

[0143] From a mathematical perspective, we can define M D-dimensional interest space vectors to represent the characteristics of the M interests themselves. The interest space vector H is denoted as: H = [h1, h2, ..., h...]. M ], H∈R M×D ,

[0144] Each D dimension h i The vector is used to represent the features of the interest itself, denoted as h. i ∈R D Vector h i As the basis vectors of the interest space vector, the first dimension is D-dimensional, and the column vectors are M-dimensional. M and D can be modified and adjusted as needed, thus adjusting the size of the interest space. The interest space vector refers to the global interest set of the recommendation system and does not differ from different target accounts.

[0145] The vector values ​​in the interest space vector are generated using a random and uniform initialization method to make the values ​​of each vector in the interest space vector similar.

[0146] For example, interests can be categorized by video type, with each type representing a different interest, such as suspense dramas as interest 1, period dramas as interest 2, and so on. Interest 1 is represented by interest basis vector h1, interest 2 by interest basis vector h2, and so on. Each interest basis vector contains information in D dimensions. All interest basis vectors in the interest set constitute the interest space vector.

[0147] In step 202, first interest feature data and second interest feature data are obtained, and the first interest feature data and the second interest feature data are used to determine an interest tendency weight,

[0148] As an example, the first interest feature data is used to represent the interest features of the target account in a first time period, the second interest feature data is used to represent the interest features of the target account in a second time period, the first time period is greater than the second time period, and the interest tendency weight is used to represent the degree of each interest tendency of the target account.

[0149] In the input data of the input recommendation system, the historical sequence feature information O a The short-term interest information is provided, the account attribute characteristics uid are provided, the long-term interest information is provided, and the degree of each interest tendency of the long-term interest information is obtained through the function FuncG used to obtain the degree of each interest tendency. Based on this, the interest tendency weight used to represent the degree of each interest tendency of the target account can be determined based on the account attribute characteristic data and the historical interest sequence feature information,

[0150] Wherein,

[0151] The first interest feature data is the account attribute characteristic data,

[0152] The second interest feature data is the historical interest sequence feature information.

[0153] For example, based on the account attribute characteristic data, the function FuncG([uid, O a ], W2) is used to obtain the interest tendency weight used to represent the degree of each interest tendency of the target account, the element value in the interest tendency weight vector is the interest tendency weight, and the interest tendency weight vector is denoted as G, G ∈ R M ; the function FuncG([uid, O a ], W2) is implemented by a first neural network model, for example, a multi-layer neural network such as an MLP structure. Wherein W2 is the parameter of the neural network, uid is the account attribute characteristic, O a is the historical sequence feature information obtained by an Attention structure or a SumPooling operation, [uid, O a ] is the input of the first neural network model, and the interest tendency weight vector G is the output of the first neural network model. At this time, G = [g1, g2,..., g M ] each element g i ∈ R 1 represents the degree of interest tendency of the interest basis vector h i in the interest space vector, as a multi-interest coordinate.

[0154] In step 203, third interest feature data, i.e., total interest feature data, for representing the overall interest of the first interest feature data is determined based on the interest tendency weight and the interest attribute feature data.

[0155] For each interest of any target account, the product of the interest base vector of the interest of the target account and the interest tendency weight of the interest is calculated to obtain the interest fusion interest base vector of the interest,

[0156] The average of all fusion interest base vectors is calculated to obtain the third interest feature vector of the target account, wherein the third interest feature vector is the third interest feature data.

[0157] The calculation of the third interest feature data is mathematically expressed as:

[0158]

[0159] Wherein, v represents the third interest feature vector of the overall interest of the target account uid, the dimension of the third interest feature vector is the same as that of the interest base vector, both are D dimensions.

[0160] For example, according to the third interest feature vector of the target account u, the average of all interest tendencies of the target account u can be determined.

[0161] The embodiment can mine the short-term and medium-term and long-term multi-interest tendencies of each target account, and through the aggregation of multiple interest preferences, it can effectively help the recommendation model to obtain more accurate prediction results. The obtained third interest feature data is beneficial to mine the diversity of interests and obtain multi-interest information, which can be used for assisting recall, rough sorting, fine sorting and other processing, and is beneficial to help mine the objects of each interest direction of the target account, including but not limited to video, picture, commodity, electronic resource and other interest contents. The embodiment is simple and effective, and can be conveniently applied to different recommendation systems, which is beneficial to improve the recommendation effect of the model without bringing new time consumption and memory consumption. The third interest feature vector of the target account extracted by the embodiment can be conveniently used in any custom model structure to improve the overall performance effect of the model, and the training method of the original model does not need to be changed, for example, the target loss function, optimizer and the like of the training do not need to be changed.

[0162] Referring to Figure 3 shown in FIG. 1, Figure 3 is a flowchart of an interest recommendation method using the obtained interest feature data according to an exemplary embodiment. The interest recommendation method includes:

[0163] In step 301, input data of an interest recommendation model is obtained;

[0164] As an example, the third interest feature data of the target account u is obtained according to steps 201 to 203. The third interest feature data can be obtained online during the recommendation process or obtained in advance offline with the recommendation process.

[0165] The third interest feature data, first interest feature data, second interest feature data, target interest attribute feature data, and interest embedding feature data are merged along a vector dimension and used as the input feature vector for the interest recommendation model.

[0166] The first interest feature data is account attribute feature data.

[0167] The second interest feature data is historical interest sequence feature information.

[0168] Step 302: Input the input feature data into the interest recommendation model, and make interest recommendations based on the prediction results of the interest recommendation model.

[0169] As an example, the input feature data is denoted as:

[0170] input = [v, uid, O] a ,emb1,emb2,...,emb N ,pid]

[0171] Where v represents the third interest feature data, uid represents the account attribute feature data, and O a For historical interest sequence feature information, emb i Embedded feature data for interests, where pid is the target interest attribute feature data.

[0172] The interest recommendation model can be a conventional second neural network model F(x, W3), such as an MLP model. This model yields a prediction result `pred` for events related to the target interest (e.g., liking). The prediction result `pred` can be expressed mathematically as:

[0173] pred=δ(F(input,W3))

[0174] Where W3 represents the model parameters of the interest recommendation model.

[0175] See Figure 4 As shown, Figure 4 This is a flowchart illustrating a training method for an interest recommendation model according to an exemplary embodiment. The training method includes:

[0176] Step 401: Obtain sample data for the interest recommendation model.

[0177] As an example, the first interest feature sample data and the second interest feature sample data are obtained, wherein the first interest feature sample data is used to represent the interest feature of the sample account in a first time period, and the second interest feature sample data is used to represent the interest feature of the sample account in a second time period, and the first time period is greater than the second time period. The third interest feature sample data is obtained according to steps 201-203. In this way, the long-term sample information of the sample account is memorized by using each account attribute feature sample data uid (the first interest feature sample data), the short-term interest is reflected by the historical interest sequence sample data (the second interest feature sample data), the preference degree of different degrees of interest is obtained, and the third sample interest feature vector of the sample account is obtained by a simple summary method.

[0178] The sample data of the recommendation system is sorted out from the log. For example, when the sample account u watches the video p and takes the event operation of “liking” or “not liking”, a sample data can be formed: the account attribute feature data (gender, age, etc.) of the sample account u, the feature data (video content label, duration, etc.) of the video p, and the event (whether to like).

[0179] The historical interest sequence sample data, the interest embedding feature sample data, and the target interest sample data can also be obtained from the log.

[0180] In step 402, the sample data is input into the interest recommendation model,

[0181] As an example, the third interest feature sample data, the first interest feature sample data, the second interest feature sample data, the target interest attribute feature sample data, and the interest embedding feature sample data are combined in vector dimension, as the input sample feature vector of the interest recommendation model.

[0182] In step 403, the target loss function value is determined according to the prediction result output by the interest recommendation model corresponding to the input sample data,

[0183] As an example, the target loss function can be cross entropy, etc.

[0184] In step 404, the model parameters of the interest recommendation model are updated according to the target loss function value until the training end condition is met.

[0185] In the phase of updating the model parameters of the interest recommendation model by back propagation, the parameters W2 and W3 in the interest recommendation model F(x, W3) and the sample interest tendency function FuncG([uid, O a ], W2) (the first neural network model) are updated by back propagation; the sample interest space vector H ∈ R M×DThe M sample interest basis vectors are kept unchanged after initialization, so that the M sample interest basis vectors have distinguishability, and the diversity of the sample interest space is maintained.

[0186] As a possible implementation, the sample interest space vector is pre-constructed in the following manner:

[0187] For each interest, a sample interest basis vector is set to represent the characteristics of the interest itself,

[0188] Each sample interest basis vector is taken as a first sample dimension vector, and all sample interest basis vectors are taken as a second sample dimension vector, to obtain a sample interest space vector composed of the first sample dimension vector and the second sample dimension vector, wherein the dimension of the first sample dimension vector is the dimension of the sample interest basis vector, the dimension of the second sample dimension vector is the total number of sample interest basis vectors, and the sample interest space vector is sample interest space data.

[0189] The vector values in the sample interest space vector are randomly and uniformly initialized.

[0190] Referring to Figure 5 , Figure 5 is a schematic diagram of an interest feature data acquisition device according to an exemplary embodiment. The device comprises:

[0191] An acquisition module 51 is configured to acquire first interest feature data and second interest feature data, wherein the first interest feature data is used to represent the interest characteristics of a target account in a first time period, the second interest feature data is used to represent the interest characteristics of the target account in a second time period, and the first time period is greater than the second time period,

[0192] An interest tendency weight determination module 52 is configured to determine an interest tendency weight using the first interest feature data and the second interest feature data,

[0193] An interest feature determination module 53 is configured to determine third interest feature data based on the interest tendency weight and pre-constructed interest space data.

[0194] Wherein,

[0195] The interest tendency weight determination module 52 is configured to determine the interest tendency weight according to the account attribute feature data of the target account and the historical interest sequence feature information,

[0196] The acquisition module 51 is further configured to acquire target interest attribute feature data and interest embedding feature data,

[0197] The third interest feature data, the first interest feature data, the second interest feature data, the target interest attribute feature data, and the interest embedding feature data are merged in vector dimensions as input feature vectors of the interest recommendation model.

[0198] The interest tendency weight 52 determination module is configured to:

[0199] The account attribute feature data and the historical interest sequence feature information are input into a first neural network model for obtaining each interest tendency degree, and the interest tendency weight vector of the target account is obtained through the first neural network model.

[0200] The device further includes:

[0201] The interest space construction module 54 is configured to pre-set, for each interest, an interest base vector for representing the characteristics of the interest itself,

[0202] Each interest base vector is taken as a first dimension vector, and all interest base vectors are taken as a second dimension vector, to obtain an interest space vector composed of the first dimension vector and the second dimension vector, wherein the dimension of the first dimension vector is the dimension of the interest base vector, the dimension of the second dimension vector is the third number of the interest base vector, and the interest space vector is interest space data.

[0203] The vector values in the interest space vector are randomly and uniformly initialized.

[0204] The interest feature determination module 53 is configured to:

[0205] For each interest of any target account, the product of the interest base vector of the interest of the target account and the interest tendency weight of the interest is calculated to obtain a fused interest base vector.

[0206] The average value of all fused interest base vectors is calculated to obtain a third interest feature vector of the target account, wherein the third interest feature vector is the third interest feature data.

[0207] Referring to Figure 6 as shown, Figure 6 is a schematic diagram of an interest recommendation device according to an exemplary embodiment.

[0208] The device includes:

[0209] The obtaining module 51 is configured to obtain first interest feature data and second interest feature data, wherein the first interest feature data is used to represent the interest features of a target account in a first time period, the second interest feature data is used to represent the interest features of the target account in a second time period, the first time period is greater than the second time period,

[0210] The interest tendency weight determination module 52 is configured to determine an interest tendency weight by using the first interest feature data and the second interest feature data, the interest tendency weight being used to represent a degree of each interest tendency of the target account,

[0211] The interest feature determination module 53 is configured to determine third interest feature data based on the interest tendency weight and pre-constructed interest space data, the interest space data being used to represent intrinsic features of each interest, and the third interest feature data being used to represent overall interests of the target account,

[0212] The recommendation module 55 is configured to input the obtained interest feature data as input data into an interest recommendation model, and perform interest recommendation according to a prediction result output by the interest recommendation model.

[0213] The interest recommendation model is a machine learning model,

[0214] The obtaining module 51 is further configured to:

[0215] obtain target interest attribute feature data and interest embedding feature data;

[0216] merge the third interest feature data, the first interest feature data, the second interest feature data, the target interest attribute feature data, and the interest embedding feature data in a vector dimension, as an input feature vector of the interest recommendation model.

[0217] The interest tendency weight determination module 52 is configured to determine the interest tendency weight according to account attribute feature data and historical interest sequence feature information of the target account,

[0218] The device further comprises:

[0219] The interest space construction module 54 is configured to pre-set, for each interest, an interest base vector used to represent intrinsic features of the interest,

[0220] obtain an interest space vector composed of a first dimension vector and a second dimension vector by taking each interest base vector as the first dimension vector and taking all interest base vectors as the second dimension vector, wherein a dimension of the first dimension vector is a dimension of the interest base vector, a dimension of the second dimension vector is a third number of the interest base vectors, and the interest space vector is the interest space data,

[0221] randomly and uniformly initialize vector values in the interest space vector.

[0222] The interest tendency weight determination module 52 is configured to:

[0223] The account attribute feature data and the historical interest sequence feature information are input into a first neural network model for obtaining each interest tendency degree, and an interest tendency weight vector of the target account is obtained through the first neural network model, wherein an element value in the interest tendency weight vector is an interest tendency weight.

[0224] The interest feature determination module 53 is configured to:

[0225] For each interest of any target account, a product of an interest base vector of the interest of the target account and an interest tendency weight of the interest is calculated to obtain a fused interest base vector,

[0226] An average value of all the fused interest base vectors is calculated to obtain a third interest feature vector of the target account, wherein the third interest feature vector is third interest feature data.

[0227] Referring to Figure 7 illustrated, Figure 7 is a schematic diagram of an interest recommendation model training apparatus according to an example embodiment. The apparatus includes:

[0228] The sample acquisition module 71 is configured to acquire first interest feature sample data and second interest feature sample data, wherein the first interest feature sample data is used to represent interest features of a sample account in a first time period, the second interest feature sample data is used to represent interest features of the sample account in a second time period, and the first time period is greater than the second time period,

[0229] The sample interest tendency weight determination module 72 is configured to determine sample interest tendency weights by using the first interest feature sample data and the second interest feature sample data, wherein the sample interest tendency weights are used to represent sample interest tendency degrees of the sample account,

[0230] The sample interest feature determination module 73 is configured to determine third interest feature sample data based on the sample interest tendency weights and pre-constructed sample interest space data, wherein the sample interest space data is used to represent self-features of each interest, and the third interest feature sample data is used to represent overall interests of the sample account,

[0231] The sample recommendation module 75 is configured to input the acquired first interest feature sample data, second interest feature sample data, and third interest feature sample data as input sample data into the interest recommendation model,

[0232] The training module 76 is configured to determine a target loss function according to a prediction result output by the interest recommendation model corresponding to the input sample data, and update model parameters of the interest recommendation model according to the target loss function until a training end condition is met.

[0233] The sample obtaining module 71 is further configured to:

[0234] obtain target interest attribute feature sample data and interest embedding feature sample data;

[0235] perform vector dimension merging on the third interest feature sample data, the first interest feature sample data, the second interest feature sample data, the target interest attribute feature sample data, and the interest embedding feature sample data, as an input sample feature vector of the interest recommendation model.

[0236] The sample interest tendency weight determining module 72 is configured to:

[0237] determine a sample interest tendency weight according to account attribute feature sample data of a sample account and historical interest sequence feature sample information.

[0238] The sample interest tendency weight determining module 72 is configured to:

[0239] input the account attribute feature sample data and the historical interest sequence feature sample information into a first neural network model for obtaining a sample interest tendency degree, and obtain, through the first neural network model, a sample interest tendency weight vector of the sample account, wherein an element value in the sample interest tendency weight vector is a sample interest tendency weight.

[0240] The apparatus further includes:

[0241] The sample interest space constructing module 74 is configured to pre-set, for each interest, a sample interest base vector for representing a feature of the interest itself,

[0242] obtain, as a first sample dimension vector, each sample interest base vector, obtain, as a second sample dimension vector, all sample interest base vectors, and obtain a sample interest space vector composed of the first sample dimension vector and the second sample dimension vector, wherein a dimension of the first sample dimension vector is a dimension of the sample interest base vector, a dimension of the second sample dimension vector is a total number of the sample interest base vectors, and the sample interest space vector is sample interest space data.

[0243] randomly and uniformly initialize vector values in the sample interest space vector.

[0244] The sample interest feature determining module 73 is configured to:

[0245] for each interest of a sample account, calculate a product of a sample interest base vector of the interest of the sample account and a sample interest tendency weight of the interest, to obtain a sample fusion interest base vector,

[0246] The average of all sample fusion interest base vectors is calculated to obtain a third sample interest feature vector of the target account, where the third sample interest feature vector is third interest feature sample data.

[0247] The training module 76 is configured to:

[0248] According to the target loss function value, the model parameters in the interest recommendation model and the first neural network model are updated, and the sample interest space vector remains unchanged.

[0249] Referring to Figure 8 as shown, Figure 8 is a schematic diagram of an electronic device according to an exemplary embodiment. The electronic device includes a processor; a memory for storing executable instructions of the processor;

[0250] The processor is configured to execute the executable instructions to implement the method for obtaining any interest feature data.

[0251] In an exemplary embodiment, a computer-readable storage medium including at least one instruction is also provided, for example, a memory including at least one instruction, the at least one instruction being executable by a processor in a computer device to complete the method for obtaining any interest feature data in the above embodiments.

[0252] Optionally, the above computer-readable storage medium can be a non-transitory computer-readable storage medium, for example, the non-transitory computer-readable storage medium can include ROM (Read-Only Memory), RAM (Random-Access Memory), CD-ROM (Compact Disc Read-Only Memory), magnetic tape, floppy disk, and optical data storage device, etc.

[0253] In an exemplary embodiment, a computer program product is also provided, including one or more instructions executable by a processor of a computer device to complete the method for obtaining interest feature data of each of the above embodiments.

[0254] Other embodiments of the application disclosed will be readily apparent to those skilled in the art after considering the specification and practice of the application disclosed herein. The application disclosed is intended to cover any variations, uses or adaptive changes of the application disclosed herein that follow, in general, the principles of the application disclosed and include common knowledge or conventional technical means in the art that are not disclosed by the application disclosed. The specification and examples are only considered as exemplary, and the true scope and spirit of the application disclosed are indicated by the following claims.

[0255] It should be understood that the application disclosure is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the application disclosure is limited only by the appended claims.

Claims

1. A method for acquiring interest feature data, characterized in that, include: Acquire first interest feature data and second interest feature data, wherein the first interest feature data is used to characterize the interest features of the target account in a first time period, and the second interest feature data is used to characterize the interest features of the target account in a second time period, wherein the first time period is longer than the second time period. Using the first interest feature data and the second interest feature data, interest tendency weights are determined, which are used to characterize the degree of each interest tendency of the target account. Based on the interest tendency weights and global interest space data, third interest feature data is determined, wherein the interest space data is used to characterize the characteristics of each interest in the interest set and is not different due to different target accounts, and the third interest feature data is used to characterize the overall interests of the target account. in, The determination of third interest feature data based on the interest tendency weights and global interest space data includes: For the target account, the intrinsic features of each interest in the interest space data are fused with the interest tendency weight of each interest to obtain fused interest feature data for each interest in the interest space data. The average value of all the fused interest feature data is calculated to obtain the third interest feature data of the target account.

2. The acquisition method according to claim 1, characterized in that, The step of determining the interest tendency weight using the first interest feature data and the second interest feature data includes: Based on the account attribute feature data and historical interest sequence feature information of the target account, the interest tendency weight is determined. in, The account attribute feature data belongs to the first interest feature data. The historical interest sequence feature information belongs to the second interest feature data.

3. The acquisition method according to claim 2, characterized in that, The step of determining the interest tendency weight based on the account attribute feature data and historical interest sequence feature information of the target account includes: The account attribute feature data and the historical interest sequence feature information are input into the first neural network model to obtain the interest tendency weight vector of the target account, wherein the element value at each position in the interest tendency weight vector represents the interest tendency weight corresponding to each position.

4. The acquisition method according to claim 1, characterized in that, The method also includes: Obtain target interest attribute feature data and interest embedding feature data, wherein the target interest attribute feature data is used to characterize the attribute features of the target interest; the interest embedding feature data is used to characterize at least one of the basic features of the target account, the basic features of the target interest, and the contextual features; The third interest feature data, the first interest feature data, the second interest feature data, the target interest attribute feature data, and the interest embedding feature data are fused along the vector dimension to obtain the input feature vector.

5. The acquisition method according to claim 1, characterized in that, The interest space data is pre-constructed in the following manner: For each interest, set an interest basis vector to represent the characteristics of that interest itself. Each of the interest basis vectors is used as a first-dimensional vector, and all the interest basis vectors are used as a second-dimensional vector to obtain an interest space vector composed of the first-dimensional vector and the second-dimensional vector. The dimension of the first-dimensional vector is the same as the dimension of the interest basis vectors, and the dimension of the second-dimensional vector is the total number of interest basis vectors. The interest space vector represents the interest space data. The vector values ​​in the interest space vector are initialized randomly and uniformly.

6. The acquisition method according to claim 5, characterized in that, The process of fusing the intrinsic characteristics of each interest in the interest space data with the interest tendency weight of each interest for the target account includes: For each interest of the target account, the interest basis vector of each interest is fused with the interest tendency weight of each interest to obtain a fused interest basis vector. The calculation of the average value of all the fused interest feature data includes: The average value of all the fused interest basis vectors is calculated to obtain the third interest feature vector of the target account, wherein the third interest feature vector is the third interest feature data.

7. The acquisition method according to claim 1, characterized in that, The method further includes: The acquired first interest feature data, second interest feature data, and third interest feature data are input into the interest recommendation model. Based on the prediction results output by the interest recommendation model, interest recommendations are made for the target account.

8. The acquisition method according to claim 7, characterized in that, The method is implemented through an interest recommendation model, and the training method of the interest recommendation model includes: Obtain sample data of the first interest feature and sample data of the second interest feature; Using the first interest feature sample data and the second interest feature sample data, determine the sample interest tendency weight; Based on the sample interest tendency weights and sample interest space data, the third interest feature sample data is determined; The acquired first interest feature sample data, second interest feature sample data, and third interest feature sample data are used as input sample data and input into the interest recommendation model; Based on the prediction results output by the interest recommendation model, the target loss function value is determined. The model parameters of the interest recommendation model are updated based on the target loss function value until the training termination condition is met.

9. The acquisition method according to claim 8, characterized in that, The training method further includes: Obtain target interest attribute feature sample data and interest embedding feature sample data; The third interest feature sample data, the first interest feature sample data, the second interest feature sample data, the target interest attribute feature sample data, and the interest embedding feature sample data are fused along the vector dimension to obtain the input sample feature vector of the interest recommendation model.

10. The acquisition method according to claim 8, characterized in that, The step of determining the sample interest tendency weight using the first interest feature sample data and the second interest feature sample data includes: Based on the sample account attribute feature data and historical interest sequence feature sample information, the sample interest tendency weights are determined. in, The account attribute feature sample data belongs to the first interest feature sample data. The historical interest sequence feature sample information belongs to the second interest feature sample data.

11. The acquisition method according to claim 10, characterized in that, Based on the account attribute feature sample data and historical interest sequence feature sample information of the sample accounts, the sample interest tendency weights are determined, including: The account attribute feature sample data and the historical interest sequence feature sample information are input into the first neural network model to obtain the sample interest tendency weight vector of the sample account, wherein the element value at each position in the sample interest tendency weight vector represents the sample interest tendency weight corresponding to each position.

12. The acquisition method according to claim 11, characterized in that, The sample interest space data is pre-constructed in the following manner: For each interest, a sample interest basis vector is set to characterize the features of that interest itself. Each of the sample interest basis vectors is used as a first sample dimension vector, and all of the sample interest basis vectors are used as a second sample dimension vector to obtain a sample interest space vector composed of the first sample dimension vector and the second sample dimension vector. The dimension of the first sample dimension vector is the same as the dimension of the sample interest basis vectors, and the dimension of the second sample dimension vector is the total number of sample interest basis vectors. The sample interest space vector is the sample interest space data. The vector values ​​in the sample interest space vector are initialized randomly and uniformly.

13. The acquisition method according to claim 12, characterized in that, The process of determining the third interest feature sample data based on the sample interest tendency weights and interest attribute feature sample data includes: For each interest of a sample account, the sample interest basis vector of that interest is fused with the sample interest tendency weight of that interest to obtain the sample fused interest basis vector. Calculate the average value of all the sample fusion interest basis vectors to obtain the third sample interest feature vector of the sample account, wherein the third sample interest feature vector is the third interest feature sample data.

14. The acquisition method according to claim 13, characterized in that, The step of updating the model parameters of the interest recommendation model based on the target loss function value includes: Based on the target loss function value, update the model parameters in the interest recommendation model and the first neural network model while keeping the sample interest space vector unchanged.

15. A device for acquiring interest feature data, characterized in that, include: The acquisition module is configured to acquire first interest feature data and second interest feature data of a target account, wherein the first interest feature data is used to characterize the interest features of the target account in a first time period, and the second interest feature data is used to characterize the interest features of the target account in a second time period, wherein the first time period is longer than the second time period. The interest tendency weight determination module is configured to determine interest tendency weights using the first interest feature data and the second interest feature data. These interest tendency weights characterize the degree of each interest tendency of the target account. The interest feature determination module is configured to determine third interest feature data based on the interest tendency weights and global interest space data, wherein the interest space data is used to characterize the features of each interest in the interest set and is not different due to different target accounts, and the third interest feature data is used to characterize the overall interests of the target account. in, The interest feature determination module is configured to, for the target account, fuse the intrinsic features of each interest in the interest space data with the interest tendency weight of each interest to obtain a fused interest vector, and calculate the average value of all the fused interest vectors to obtain the third interest feature vector of the target account.

16. The acquiring device according to claim 15, characterized in that, The interest tendency weight determination module is configured to determine the interest tendency weight based on the account attribute feature data and historical interest sequence feature information of the target account. in, The account attribute feature data belongs to the first interest feature data. The historical interest sequence feature information belongs to the second interest feature data.

17. The acquiring device according to claim 15, characterized in that, The acquisition module is further configured to acquire target interest attribute feature data and interest embedding feature data, wherein the target interest attribute feature data is used to characterize the attribute features of the target interest; the interest embedding feature data is used to characterize at least one of the basic features of the target account, the basic features of the target interest, and the contextual features; The third interest feature data, the first interest feature data, the second interest feature data, the target interest attribute feature data, and the interest embedding feature data are fused along the vector dimension to obtain the input feature vector.

18. The acquiring device according to claim 17, characterized in that, The interest tendency weight determination module is configured as follows: The account attribute feature data and historical interest sequence feature information are input into the first neural network model to obtain the interest tendency weight vector of the target account. The element value of each position in the interest tendency weight vector represents the interest tendency weight corresponding to each position.

19. The acquiring device according to claim 15, characterized in that, The device also includes: The interest space construction module is configured to pre-set interest basis vectors for each interest to characterize the features of that interest itself. Each of the interest basis vectors is used as a first-dimensional vector, and all the interest basis vectors are used as a second-dimensional vector to obtain an interest space vector composed of the first-dimensional vector and the second-dimensional vector. The dimension of the first-dimensional vector is the same as the dimension of the interest basis vectors, and the dimension of the second-dimensional vector is the same as the dimension of the interest basis vectors. The interest space vector represents the interest space data. The vector values ​​in the interest space vector are initialized randomly and uniformly.

20. The acquiring device according to claim 19, characterized in that, The interest feature determination module is configured as follows: For each interest of the target account, the interest basis vector of each interest is fused with the interest tendency weight of each interest to obtain a fused interest basis vector. The average value of all the fused interest basis vectors is calculated to obtain the third interest feature vector of the target account, wherein the third interest feature vector is the third interest feature data.

21. The acquiring device according to claim 15, characterized in that, The device further includes: The recommendation module is configured to input the acquired first interest feature data, second interest feature data, and third interest feature data into the interest recommendation model, and make interest recommendations for the target account based on the prediction results output by the interest recommendation model.

22. The acquiring device according to claim 21, characterized in that, The device is implemented through an interest recommendation model, and the training device for the interest recommendation model includes: The sample data acquisition module is configured to acquire sample data of the first interest feature and the second interest feature of the sample account. The sample interest tendency weight determination module is configured to determine the sample interest tendency weights using the first interest feature sample data and the second interest feature sample data. The sample interest feature determination module is configured to determine third interest feature sample data based on the sample interest tendency weights and sample interest space data. The sample recommendation module is configured to input the acquired first interest feature sample data, second interest feature sample data, and third interest feature sample data into the interest recommendation model. The training module is configured to determine the target loss function value based on the prediction result output by the interest recommendation model corresponding to the input sample data, and update the model parameters of the interest recommendation model based on the target loss function value until the training termination condition is met.

23. The acquiring device according to claim 22, characterized in that, The sample acquisition module is further configured as follows: Obtain target interest attribute feature sample data and interest embedding feature sample data; The third interest feature sample data, the first interest feature sample data, the second interest feature sample data, the target interest attribute feature sample data, and the interest embedding feature sample data are fused along the vector dimension to obtain the input sample feature vector of the interest recommendation model.

24. The acquiring device according to claim 22, characterized in that, The sample interest tendency weight determination module is configured as follows: The sample interest tendency weights are determined based on the account attribute feature sample data and historical interest sequence feature sample information of the sample accounts. in, The account attribute feature sample data belongs to the first interest feature sample data. The historical interest sequence feature sample information belongs to the second interest feature sample data.

25. The acquiring device according to claim 22, characterized in that, The sample interest tendency weight determination module is configured as follows: The account attribute feature sample data and historical interest sequence feature sample information are input into the first neural network model to obtain the sample interest tendency weight vector of the sample account, wherein the element value of each position in the sample interest tendency weight vector represents the sample interest tendency weight corresponding to each position.

26. The acquiring device according to claim 25, characterized in that, The training device also includes: The sample interest space construction module is configured to pre-set sample interest basis vectors for each interest to characterize the features of that interest itself. Each of the sample interest basis vectors is used as a first sample dimension vector, and all of the sample interest basis vectors are used as a second sample dimension vector to obtain a sample interest space vector composed of the first sample dimension vector and the second sample dimension vector. The dimension of the first sample dimension vector is the same as the dimension of the sample interest basis vectors, and the dimension of the second sample dimension vector is the total number of sample interest basis vectors. The sample interest space vector is the sample interest space data. The vector values ​​in the interest space vector are initialized randomly and uniformly.

27. The acquiring device according to claim 26, characterized in that, The sample interest feature determination module is configured as follows: For each interest of the sample account, the sample interest basis vector of that interest and the sample interest tendency weight of that interest are fused to obtain the sample fused interest basis vector. Calculate the average value of all the sample fusion interest basis vectors to obtain the third sample interest feature vector of the sample account, wherein the third sample interest feature vector is the third interest feature sample data.

28. The acquiring device according to claim 27, characterized in that, The training module is configured as follows: Based on the target loss function value, update the model parameters in the interest recommendation model and the first neural network model while keeping the sample interest space vector unchanged.

29. An electronic device, characterized in that, include: processor; Memory for storing the executable instructions of the processor; The processor is configured to execute the executable instructions to implement the method for acquiring interest feature data as described in any one of claims 1 to 14.

30. A computer-readable storage medium, characterized in that, When at least one instruction in the computer-readable storage medium is executed by the processor of the electronic device, the electronic device is able to implement the method for acquiring interest feature data as described in any one of claims 1 to 14.

31. A computer program product comprising a computer program that, when executed by a processor, implements the method for acquiring interest feature data as described in any one of claims 1 to 14.

Images