CN115619463A

CN115619463A - Target user determination method and device, storage medium and electronic equipment

Info

Publication number: CN115619463A
Application number: CN202211295508.XA
Authority: CN
Inventors: 吴立帅
Original assignee: Beijing Jingdong Zhenshi Information Technology Co Ltd
Current assignee: Beijing Jingdong Zhenshi Information Technology Co Ltd
Priority date: 2022-10-21
Filing date: 2022-10-21
Publication date: 2023-01-17

Abstract

The disclosure provides a target user determination method, a target user determination device, a storage medium and an electronic device; relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring multi-dimensional feature data of a plurality of users, and generating user feature vectors and advertisement feature vectors according to the multi-dimensional feature data; inputting the user characteristic vector and the advertisement characteristic vector into a user characterization model to generate a first seed user vector and a first candidate user vector; inputting the first seed user vector and the first candidate user vector into a similarity pre-estimation model, generating a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector; and determining a target user in the candidate users according to the similarity. The method and the device can accurately and efficiently determine the similar population of the seed users, and further improve the accuracy of advertisement putting.

Description

Target user determination method and device, storage medium and electronic equipment

Technical Field

The present disclosure relates to the field of artificial intelligence technologies, and in particular, to a target user determination method, a target user determination device, a computer-readable storage medium, and an electronic device.

Background

With the development of network technology, advertisement delivery has become an important component in the network field. In an advertisement putting scene, there is often a demand for expanding the putting population.

For example, when offline advertisement delivery is performed, crowd selection is performed manually mainly based on historical behavior information of users and appeal of advertisers, and efficiency is low. Moreover, the manual crowd selection depends on historical behavior information of users (users) corresponding to item (material), so that a 'Martian' effect in advertisement delivery is easily caused. For example, a cold, but high quality, long-tailed content may be deemed to have a lower delivery priority and thus may not result in an effective delivery.

Therefore, in order to ensure effective delivery of the advertisement, it is necessary to provide an accurate and efficient method for targeting users.

It is to be noted that the information disclosed in the above background section is only for enhancement of understanding of the background of the present disclosure, and thus may include information that does not constitute prior art known to those of ordinary skill in the art.

Disclosure of Invention

The present disclosure is directed to a target user determination method, a target user determination apparatus, a computer-readable storage medium, and an electronic device, which overcome the problem of low efficiency and accuracy of similar population expansion due to limitations and disadvantages of the related art to some extent.

According to a first aspect of the present disclosure, there is provided a target user determination method, including:

acquiring multi-dimensional feature data of a plurality of users, and generating user feature vectors and advertisement feature vectors according to the multi-dimensional feature data, wherein the plurality of users comprise seed users and candidate users;

inputting the user feature vector and the advertisement feature vector into a user characterization model to generate a user characterization vector, wherein the user characterization vector comprises a first seed user vector and a first candidate user vector;

inputting the first seed user vector and the first candidate user vector into a similarity pre-estimation model, generating a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector;

and determining a target user in the candidate users according to the similarity.

In an exemplary embodiment of the present disclosure, the seed user includes one of a seed user who has historically delivered an advertisement and a seed user who is to be delivered an advertisement; before the obtaining the multidimensional feature data of the plurality of users, the method comprises the following steps:

acquiring a first advertisement characteristic vector of an advertisement to be delivered and a second advertisement characteristic vector of a historical advertisement delivered;

calculating the similarity between the first advertisement characteristic vector and the second advertisement characteristic vector, and determining a target advertisement similar to the advertisement to be delivered in the historical advertisements according to the similarity;

and constructing the seed user to be advertised based on the seed user targeted to be advertised.

In an exemplary embodiment of the present disclosure, the multi-dimensional feature data includes user feature data and advertisement feature data; the generating of the user feature vector and the advertisement feature vector according to the multi-dimensional feature data comprises:

vectorizing the user characteristic data to generate the user characteristic vector;

and vectorizing the advertisement characteristic data to generate the advertisement characteristic vector.

In an exemplary embodiment of the present disclosure, the user characteristic data includes at least one of discrete characteristic data and continuous characteristic data; the vectorizing the user feature data to generate the user feature vector includes:

counting the discrete characteristic data to obtain a plurality of first characteristic data;

encoding each first feature data to generate a plurality of first feature vectors;

normalizing each continuous feature data to generate a plurality of second feature vectors;

and splicing the plurality of first feature vectors and the plurality of second feature vectors to generate the user feature vector.

In an exemplary embodiment of the present disclosure, the user characterization model includes at least a first fully connected layer and a first attention layer; inputting the user feature vector and the advertisement feature vector into a user characterization model to generate a user characterization vector, including:

performing feature dimension reduction on the user feature vector and the advertisement feature vector by using the first full connection layer to obtain a user intermediate vector and an advertisement intermediate vector;

generating, by the first attention layer, the first seed user vector and the first candidate user vector based on the user intermediate vector and the advertisement intermediate vector.

In an exemplary embodiment of the present disclosure, the similarity pre-estimation model includes two network branches, each network branch includes at least a second fully-connected layer and a second attention layer; the inputting the first seed user vector and the first candidate user vector into a similarity prediction model to generate a second seed user vector and a second candidate user vector includes:

performing feature dimensionality reduction on the first seed user vector and the first candidate user vector by using the second full-connection layer to obtain a seed user intermediate vector and a candidate user intermediate vector;

generating, by the second attention layer, the second seed user vector and the second candidate user vector based on the seed user intermediate vector and the candidate user intermediate vector.

In an exemplary embodiment of the present disclosure, the method further comprises:

acquiring first training samples, wherein each first training sample is a user characteristic vector and an advertisement characteristic vector of a candidate user;

inputting the user characteristic vector and the advertisement characteristic vector of each candidate user into the user characterization model to obtain a corresponding user characterization vector and an advertisement characterization vector;

constructing a first objective function according to the user characterization vector and the advertisement characterization vector;

and iteratively updating the parameters of the user characterization model based on the first objective function, and finishing the training of the user characterization model when an iteration termination condition is met.

In an exemplary embodiment of the present disclosure, the method further comprises: the method further comprises the following steps:

acquiring second training samples, wherein each second training sample is a vector pair consisting of a seed user vector and a target user vector;

determining the similarity of the seed user vector and the target user vector through the similarity pre-estimation model;

constructing a second objective function according to the similarity of the seed user vector and the target user vector;

and iteratively updating the parameters of the similarity estimation model based on the second objective function, and finishing the training of the similarity estimation model when an iteration termination condition is met.

According to a second aspect of the present disclosure, there is provided a target user determination apparatus comprising:

the system comprises a feature vector generation module, a feature vector generation module and a feature vector generation module, wherein the feature vector generation module is used for acquiring multi-dimensional feature data of a plurality of users and generating user feature vectors and advertisement feature vectors according to the multi-dimensional feature data, and the plurality of users comprise seed users and candidate users;

a feature vector generation module, configured to input the user feature vector and the advertisement feature vector into a user feature model, and generate a user feature vector, where the user feature vector includes a first seed user vector and a first candidate user vector;

the similarity calculation module is used for inputting the first seed user vector and the first candidate user vector into a similarity prediction model, generating a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector;

and the target user determining module is used for determining a target user in the candidate users according to the similarity.

According to a third aspect of the present disclosure, there is provided a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the method of any one of the above.

According to a fourth aspect of the present disclosure, there is provided an electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the method of any one of the above via execution of the executable instructions.

Exemplary embodiments of the present disclosure may have some or all of the following benefits:

in the target user determination method provided by the exemplary embodiment of the present disclosure, multi-dimensional feature data of a plurality of users is obtained, and a user feature vector and an advertisement feature vector are generated according to the multi-dimensional feature data; inputting the user characteristic vector and the advertisement characteristic vector into a user characterization model to generate a first seed user vector and a first candidate user vector; inputting the first seed user vector and the first candidate user vector into a similarity pre-estimation model, generating a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector; and determining a target user in the candidate users according to the similarity. On one hand, different feature domains are sufficiently learned and combined through the user characterization model to generate a user characterization vector, and the user characterization vector can more accurately reflect the potential interest of the user, so that the similar population of the seed user can be accurately determined. Moreover, the similarity between the seed user and the candidate user can be calculated more accurately by utilizing the seed user vector and the candidate user vector output by the similarity estimation model, so that the accuracy of similar population expansion is further improved, and the accuracy of advertisement putting is further improved; on the other hand, the target user can be determined through the user representation model and the similarity estimation model without manual selection, and the efficiency of similar population expansion is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and together with the description, serve to explain the principles of the disclosure. It should be apparent that the drawings in the following description are merely examples of the disclosure and that other drawings may be derived by those skilled in the art without inventive effort.

Fig. 1 illustrates a schematic diagram of an exemplary system architecture to which the target user determination method and apparatus in embodiments of the present disclosure may be applied;

FIG. 2 schematically illustrates a flow chart of a target user determination method in accordance with an embodiment of the disclosure;

FIG. 3 schematically illustrates an architectural diagram of a user characterization model in accordance with an embodiment of the present disclosure;

FIG. 4 is a schematic diagram illustrating an architecture of a similarity prediction model according to an embodiment of the present disclosure;

FIG. 5 schematically illustrates a flow diagram of offline ad targeting in accordance with an embodiment of the present disclosure;

fig. 6 schematically shows a block diagram of a target user determination apparatus according to an embodiment of the present disclosure;

FIG. 7 schematically illustrates a structural schematic diagram of a computer system suitable for use with an electronic device implementing an embodiment of the disclosure.

Detailed Description

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in many different forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of embodiments of the disclosure. One skilled in the relevant art will recognize, however, that the subject matter of the present disclosure can be practiced without one or more of the specific details, or with other methods, components, devices, steps, and the like. In other instances, well-known technical solutions have not been shown or described in detail to avoid obscuring aspects of the present disclosure.

Furthermore, the drawings are merely schematic illustrations of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus their repetitive description will be omitted. Some of the block diagrams shown in the figures are functional entities and do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor devices and/or microcontroller devices.

Fig. 1 is a schematic diagram illustrating an exemplary system architecture to which a target user determination method and apparatus according to an embodiment of the present disclosure may be applied.

As shown in fig. 1, the system architecture 100 may include one or more of

terminal devices

101, 102, 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the

terminal devices

101, 102, 103 and the server 105. Network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, to name a few. The

terminal devices

101, 102, 103 may be various electronic devices having a display screen, including but not limited to desktop computers, portable computers, smart phones, tablet computers, and the like. It should be understood that the number of terminal devices, networks, and servers in fig. 1 is merely illustrative. There may be any number of terminal devices, networks, and servers, as desired for implementation. For example, server 105 may be a server cluster comprised of multiple servers, and the like.

The target user determination method provided by the embodiment of the present disclosure may be executed by the server 105, and accordingly, the target user determination apparatus may be disposed in the server 105, and the server may send the user feature vector, the advertisement feature vector, the user characterization vector, the target user, and the like to the terminal device, and the terminal device may display the user feature vector, the advertisement feature vector, the user characterization vector, the target user, and the like to a worker. However, it is easily understood by those skilled in the art that the target user determination method provided in the embodiment of the present disclosure may also be executed by the

terminal devices

101, 102, and 103, and accordingly, the target user determination apparatus may also be disposed in the

terminal devices

101, 102, and 103, for example, after being executed by the terminal devices, the user feature vector, the advertisement feature vector, the user characterization vector, the target user, and the like may be directly displayed on a display screen of the terminal device to be displayed to a worker, which is not particularly limited in this exemplary embodiment.

The technical solution of the embodiment of the present disclosure is explained in detail below:

in the exemplary embodiment of the present disclosure, offline advertisement placement in the logistics industry is taken as an example for explanation. At present, when off-line advertisement delivery is carried out, people group selection is carried out manually mainly based on historical behavior information of users and appeal of advertisers, and efficiency is low. Moreover, the manual crowd selection depends on historical behavior information of users (users) corresponding to item (material), and a Martian effect in advertisement putting is easily caused. For example, a cold, but high quality, long-tailed content may be deemed to have a lower delivery priority and thus may not result in an effective delivery.

Based on one or more of the above problems, the present exemplary embodiment provides a target user determination method, which establishes a Look-impact model based on an Attention mechanism, and can implement accurate expansion of offline logistics advertisement crowd. The look-elevation technology can expand crowd similar to seed users through big data analysis and machine learning according to a small number of seed users, namely, automatic expansion is carried out according to the common attributes of the seed crowd, so that the coverage of potential users is expanded, and marketing/advertising effects are improved.

It is to be noted that the method and approach for obtaining multidimensional feature data of a user in the exemplary embodiments of the present disclosure is compliant. Referring to fig. 2, the target user determination method may include the following steps S210 to S240:

s210, obtaining multi-dimensional feature data of a plurality of users, and generating user feature vectors and advertisement feature vectors according to the multi-dimensional feature data, wherein the plurality of users comprise seed users and candidate users;

step S220, inputting the user characteristic vector and the advertisement characteristic vector into a user characterization model to generate a user characterization vector, wherein the user characterization vector comprises a first seed user vector and a first candidate user vector;

step S230, inputting the first seed user vector and the first candidate user vector into a similarity prediction model to generate a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector;

and S240, determining a target user in the candidate users according to the similarity.

In the target user determination method provided by the exemplary embodiment of the present disclosure, multi-dimensional feature data of a plurality of users is obtained, and a user feature vector and an advertisement feature vector are generated according to the multi-dimensional feature data; inputting the user characteristic vector and the advertisement characteristic vector into a user characterization model to generate a first seed user vector and a first candidate user vector; inputting the first seed user vector and the first candidate user vector into a similarity pre-estimation model, generating a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector; and determining a target user in the candidate users according to the similarity. On one hand, different feature domains are sufficiently learned and combined through the user characterization model to generate a user characterization vector, and the user characterization vector can more accurately reflect the potential interest of the user, so that the similar population of the seed user can be accurately determined. Moreover, the similarity of the seed users and the candidate users can be more accurately calculated by using the seed user vector and the candidate user vector output by the similarity estimation model, so that the accuracy of similar population expansion is further improved, and the accuracy of advertisement putting is further improved; on the other hand, the target user can be determined through the user representation model and the similarity estimation model without manual selection, and the efficiency of similar population expansion is improved.

The above steps of the present exemplary embodiment will be described in more detail below.

In step S210, multi-dimensional feature data of a plurality of users is obtained, and a user feature vector and an advertisement feature vector are generated according to the multi-dimensional feature data, where the plurality of users include a seed user and a candidate user.

Offline advertisement placement in the logistics industry refers to distributing advertisement materials to warehouses so that advertisements can be placed to users through the warehouses. The off-line advertisement putting mode can be advertisement along with package, outer package advertisement and the like, and the putting process is realized by relying on the logistics package of the order task in the warehouse. For example, the advertisement leaflet may be attached to a package, and the user receives the package and browses and clicks on the product through the link in the advertisement leaflet, which may be regarded as successful delivery of the offline advertisement. When the Look-elevation model is used for determining the advertisement putting population, the seed user can be obtained, and the population similar to the seed user is expanded. The seed user may be one user or a group of users. For example, a seed user may be 100 million users that click/share/purchase a certain item. The candidate users may be a plurality of users in the pool who have not placed an advertisement for the item, such as 2000 million users between the ages of 15 and 50.

The multidimensional characteristic data of the user can comprise user characteristic data and advertisement characteristic data, the user characteristic data can comprise characteristic data of multiple dimensions such as attribute characteristic data and behavior characteristic data, and the advertisement characteristic data can refer to material information in a user logistics order. The attribute feature data may include portrait information of the user, such as age, gender, academic calendar, occupation, and the like, and may also include information of the user, such as a common address, an interest preference, and the like. The behavior feature data can be behavior features of browsing, clicking, consulting, buying, collecting and the like of the user. The advertisement characteristic data may include SKU (stock keeping unit), category of the goods, and the like.

After the multi-dimensional feature data of the seed user and the multi-dimensional feature data of the candidate user are obtained, vectorization processing can be performed on the feature data of each dimension to obtain a user feature vector and an advertisement feature vector. For example, the vectorization processing may be performed on the user feature data to generate a user feature vector, and the vectorization processing may be performed on the advertisement feature data to generate an advertisement feature vector. For example, the user may be characterized by using the portrait information of the user and a plurality of behavior characteristics of the user, and the advertisement may be characterized by using the material information and the product SKU, the class to which the product belongs, and the like included in the plurality of behavior characteristics of the user.

Taking the user characteristic data as an example, the user characteristic data may further include at least one of discrete characteristic data and continuous characteristic data. The behavior characteristic data of the user such as browsing, clicking, consulting, purchasing, collecting and the like is continuous characteristic data, and the age, sex, academic calendar, commodity category and the like of the user are discrete characteristic data. Illustratively, the discrete feature values may be counted, the frequency of occurrence of each feature data is counted to serve as first feature data, and the first feature data is encoded to obtain a first feature vector. For example, the first feature data is encoded as one-hot (one-hot encoded) features. For the continuous feature data, normalization processing may be performed on the continuous feature data to generate a second feature vector. The normalized continuous features can convert features of different magnitudes into features of the same magnitude, so that the influence of different magnitudes on the model is avoided. For example, each dimension feature can be linearly mapped to a target range, and if mapped to [0,1] or [ -1,1], it can also be normalized by standard deviation, which is not limited by this disclosure.

In the exemplary embodiment of the present disclosure, a pipeline mechanism may be used to serialize the vectorization process of each dimension feature data, so as to implement concatenation of multiple first feature vectors and multiple second feature vectors, thereby generating feature vectors of each user. It can be seen that the feature vector of each user is composed of normalized continuous feature data and encoded discrete feature data. The Pipeline mechanism is a batch processing technology, and can improve data processing efficiency.

Similarly, when the advertisement feature data is vectorized, all advertisement feature values of the user may be counted, for example, an enumerated mode or average of each advertisement feature is used as the processed feature data. The purchasing power of the user can be described by counting the data of average collection, purchase adding, order placing unit price and the like of the user, and the loyalty of the user is measured by counting the average browsing time and the order placing times of the user in a certain time period. Similarly, the category features in the advertisement feature data can be coded, the numerical features in the advertisement feature data are normalized, and the advertisement feature vectors are obtained through summarization by a pipeline mechanism. The process of generating advertisement feature vectors is similar to the process of generating user feature vectors, and is not described in detail here.

In order to increase the diversity of advertisement characteristics, chinese word segmentation can be performed on characteristic data such as commodity information and commodity SKU (stock keeping unit) browsed, purchased and collected by a user, and the characteristic data is segmented into range attributes with finer granularity to form new characteristics. For example, word segmentation may be performed based on a dictionary, or based on statistics, or based on rules, which is not specifically limited in this disclosure. For example, by segmenting the item SKU, the color characteristics of the item can be obtained.

Importantly, in example embodiments of the present disclosure, the determination of target users may include both crowd expansion and submarine mining scenarios. Crowd extension refers to the extension of historical crowd in the old advertisement putting scene. The hidden customer mining means mining users in a user pool under a new advertisement putting scene to determine potential customers. Correspondingly, the seed user can be one of the seed user who has historically delivered the advertisement and the seed user who is to be delivered the advertisement. In an old advertisement putting scene, seed user groups of different grades of commodities can be established according to historical advertisement information successfully put by an advertiser. For example, multiple seed user groups may be established for different brands of mobile advertisements. In the new advertisement putting scene, no historical putting information can be referred to, so that the similarity recall of advertisement materials can be carried out, and the seed user group for carrying out advertisement putting on similar commodities is used as the seed user group for putting the advertisement.

For example, a first advertisement feature vector of the advertisement to be delivered and a second advertisement feature vector of the advertisement to be delivered in history can be obtained, and the similarity between the first advertisement feature vector and the second advertisement feature vector can be calculated, so as to determine the target advertisement to be delivered similar to the advertisement to be delivered in history according to the similarity. Finally, the seed users to be advertised can be constructed based on the seed users targeted for advertising. For example, a first advertisement feature vector may be obtained according to features of multiple domains, such as a product SKU and a product class, of the product in the advertisement to be delivered, and a second advertisement feature vector may be obtained according to features of multiple domains, such as a product SKU and a product class, of the product in the historical advertisement to be delivered. And calculating the distance between the first advertisement characteristic vector and the second advertisement characteristic vector to obtain the similarity between the two vectors. For example, a cosine distance, a Euclidean distance, etc. between the first advertisement feature vector and the second advertisement feature vector may be calculated. By calculating the similarity between the advertisement to be delivered and a plurality of historical delivered advertisements, the historical delivered advertisement with the maximum similarity to the advertisement to be delivered can be used as the target delivered advertisement. It can be understood that the seed user who targets to place the advertisement has a higher probability of being interested in the commodity in the advertisement to be placed, and therefore, the seed user who targets to place the advertisement can be used as the seed user who waits to place the advertisement, so that the seed user can be used for realizing the hidden passenger mining in a new advertisement placement scene.

In the example, by accumulating successful seed user groups for historical advertisement delivery and using the seed user groups for knowledge guidance in a new advertisement delivery scene, the defects of manual experience can be made up, and accurate mining of the hidden passengers can be realized.

In step S220, the user feature vector and the advertisement feature vector are input into a user characterization model, and a user characterization vector is generated, where the user characterization vector includes a first seed user vector and a first candidate user vector.

Since the users include seed users and candidate users, the user feature vectors generated according to the multidimensional feature data of each user may include the seed user feature vectors and the candidate user feature vectors. In an example embodiment, the user characterization model may be an improved YouTube DNN (depth semantics) model that includes at least a first fully-connected layer and a first attention layer. For example, the first full-connection layer may be used to perform feature dimension reduction on the user feature vector and the advertisement feature vector, so as to obtain a user intermediate vector and an advertisement intermediate vector. Further, based on the user intermediate vector and the advertisement intermediate vector, a first seed user vector and a first candidate user vector may be generated by the first attention layer.

Referring to fig. 3, a schematic structural diagram of the YouTube DNN model fused with the Attention mechanism is shown. The model includes 3 hidden layers of DNN structures, namely an input layer 310, a hidden layer 320, and an output layer 330. The input layer 310 is configured to process the input features, for example, the input user features and the advertisement features may be respectively subjected to a splicing process or an average pooling process, and the processed user features and the processed advertisement features may be input into the full connection layer 1 (301) for spatial transformation. The hidden layer 320 in turn comprises a fully connected layer 1 (301), a fully connected layer 2 (303) and a fully connected layer 3 (304), each fully connected layer connecting an activation function, such as a ReLU (Rectified Linear Unit) activation function, and a first attention layer (302) is added between the fully connected layer 1 (301) and the fully connected layer 2 (303), the first attention layer (302) being able to fully learn to combine different feature domains. The output layer 330 may output one user feature vector and one advertisement feature vector correspondingly. It is to be appreciated that in order to obtain local feature saliency using the attention layer, it is common to add the attention layer to a shallow fully-connected layer. In other examples, the attention layer may be added to the deep fully-connected layer, such as the first attention layer (302) may be added between the fully-connected layer 2 (303) and the fully-connected layer 3 (304), which is not limited by this disclosure.

Illustratively, the user feature vectors and advertisement feature vectors obtained by the Pipeline mechanism may be entered into the YouTube DNN model as shown in fig. 3. Specifically, the user feature vectors and the advertisement feature vectors are respectively averaged and pooled and then transferred to the full link layer 1 (301), and the user feature vectors and the advertisement feature vectors averaged and pooled by the full link layer 1 (301) can be used for feature dimension reduction to obtain user intermediate vectors and advertisement intermediate vectors. Then, the user intermediate vector and the advertisement intermediate vector are input into a first attention layer (302), the composition of the advertisement intermediate vector can be captured through the first attention layer (302), and the advertisement intermediate vector similar to the user intermediate vector is weighted, so that the expression capacity of the advertisement intermediate vector can be improved. Correspondingly, after the two multi-dimensional vectors are output from the first attention layer (302), the two multi-dimensional vectors can be subjected to space reconversion sequentially through the full connection layer 2 (303) and the full connection layer 3 (304), so that a user characterization vector and an advertisement characterization vector are obtained. Wherein the user characterization vector may include a first seed user vector and a first candidate user vector.

In this example, an Attention mechanism is fused in the YouTube DNN model, which may enable the depth model to possess the ability to automatically focus on key features in a plethora of information. And the output of the last hidden layer of the model is used as a user representation vector, and low-order features of different dimensions can be aggregated in the space, so that the user representation vector can accurately reflect the potential interest of a user, and the subsequent crowd extension can be accurately realized.

In step S230, the first seed user vector and the first candidate user vector are input into a similarity prediction model to generate a second seed user vector and a second candidate user vector, and a similarity between the second seed user vector and the second candidate user vector is calculated.

Referring to fig. 4, a schematic structural diagram of the similarity prediction model is shown. The similarity pre-estimation model is of a double-tower structure, namely comprises two network branches, and each network branch at least comprises a second full-connection layer (401) and a second attention layer (402) which are respectively used for processing the seed user characterization vectors and the candidate user characterization vectors. For example, feature dimensionality reduction may be performed on the first seed user vector and the first candidate user vector by using each second fully-connected layer to obtain a seed user intermediate vector and a candidate user intermediate vector. Further, a second seed user vector and a second candidate user vector may be generated by the second attention layer based on the seed user intermediate vector and the candidate user intermediate vector. And finally, determining a target user in the candidate users by calculating the similarity between the second seed user vector and the second candidate user vector, and performing offline advertisement delivery on the target user.

For example, a first sub-user vector may be input into a first branch of the similarity prediction model, and spatially translated by a second fully-connected layer (401) in the first branch and an activation function connected to the second fully-connected layer (401). The reduced-dimension first seed user vector may then be input into a second attention layer (402) to obtain a second seed user vector. The activation function may be a prellu (Parametric reconstructed Linear Unit) activation function. Similarly, the first candidate user vector may be input into the second branch of the similarity prediction model to perform the same processing, so as to obtain a second candidate user vector, and the specific processing process is not described in detail herein. It can be understood that the vector features which are more valuable than those in the current candidate user can be selected from the seed user characterization vectors through the attention mechanism, so that the expression capability of the seed user characterization vectors can be improved, and the similarity between the seed user and the candidate user can be calculated more accurately.

After the second seed user vector and the second candidate user vector are obtained, the similarity between the two vectors can be calculated, so as to determine a candidate user similar to the seed user according to the similarity. For example, a cosine distance between two vectors may be calculated, and a euclidean distance, a mahalanobis distance, etc. between two vectors may also be calculated to obtain a similarity score of the two vectors. It should be noted that the similarity score between each candidate user and the current seed user may be normalized to 0 to 1, so that the candidate users may be conveniently screened according to the normalized similarity score to determine a plurality of target users similar to the seed user. In other examples, the similarity between each candidate user and a plurality of seed users may be calculated, the plurality of similarity scores are weighted and averaged, the calculation result is used as the final similarity score of each candidate user, and the target user among the candidate users is determined according to the similarity scores.

In an exemplary embodiment, before predicting the target user online, the user characterization model and the similarity prediction model may be trained in advance. For the user characterization model, a YouTube DNN model fused with the Attention mechanism is taken as an example. For example, first training samples may be obtained, where each of the first training samples is a user feature vector and an advertisement feature vector of a candidate user. The user characteristic vector and the advertisement characteristic vector of each candidate user can be input into the user characterization model to obtain the corresponding user characterization vector and advertisement characterization vector, and a first objective function is constructed according to the user characterization vector and the advertisement characterization vector. And iteratively updating the parameters of the user characterization model based on the first objective function, and finishing the training of the model when an iteration termination condition is met. The first objective function may be an NCE (Noise contrast Estimation) loss function. It can be understood that the samples corresponding to the completion of the commodity link skip by the user in the training samples are positive samples, the samples which are not touched or are not finished with the commodity link skip after being touched are negative samples, and other samples in the intermediate link are invalid data.

For example, the first training sample may be scaled into a training set, a validation set, and a test set, such as a scale of 7:1:2. the training set is used for training model parameters, the verification set is used for observing performance changes in the model training process, under-fitting or over-fitting is avoided, and the test set is used for evaluating the model effect. When the parameters of the user characterization model are optimized, taking the YouTube DNN model fused with the Attention mechanism as an example, the parameters may be initialized, the maximum iteration number, the selection of the activation function, the selection of the optimizer, the early stop condition, the learning rate, and the like. When the parameters of the model are trained, the training is terminated when all the parameters tend to converge or a certain iteration number is met. In the process of adjusting model parameters, model characteristics are also adjusted, evaluation can be carried out according to the similarity between the characteristics, the characteristic dimensionality is reduced as much as possible, and overfitting is avoided.

For the similarity prediction model, when the model is trained, second training samples can be obtained, and each second training sample is a vector pair formed by a seed user vector and a target user vector. Determining the similarity of the seed user vector and the target user vector through a similarity pre-estimation model, constructing a second target function according to the similarity of the seed user vector and the target user vector, iteratively updating parameters of the similarity pre-estimation model based on the second target function, and finishing the training of the model when an iteration termination condition is met. The similarity estimation model is a multi-classification task, and the category of the seed user group corresponding to each candidate user needs to be predicted, so that the second objective function can be a plurality of groups of cross entropy loss functions. Similarly, the second training sample may be represented as 7:1: the ratio of 2 is divided into a training set, a validation set and a test set. The prediction result of the model is the probability that each candidate user belongs to the seed user group. And when all parameters in the model tend to converge or satisfy a certain iteration number, the training is terminated.

After the user characterization model and the similarity estimation model are trained, the model performance of each model can be evaluated by using a test set. For example, the evaluation can be performed by using indices such as AUC (Area Under ROC Curve enclosed with coordinate axes), accuracy, recall, prec @ k, and F1 value. Where AUC is a common metric in ranking model evaluation, the ROC curve is also called receiver operating characteristic curve. Taking the evaluation of positive samples as an example: AUC is the positive and negative sample pair in the traversal data, where the probability that the positive sample predictor is greater than the negative sample.

Pre is the accuracy of the model, where TP represents the data that successfully predicts positive samples as positive, FP represents the data that mispredicts negative samples as positive, and TP + FP represents all the data predicted as positive samples, i.e.:

rec is the recall ratio of the model, where TP represents data that successfully predicts positive samples as positive, FN represents data that mispredicts positive samples as negative, and TP + FN represents all positive sample data, i.e.:

f1 is a harmonic mean of model accuracy and recall, where Pre is the accuracy of the model and Rec is the recall of the model, i.e.:

prec @ k is the proportion of recommended k-top advertising categories that users actually reach.

In the example, an attention mechanism is introduced into the user characterization model and the similarity pre-estimation model, and the attention mechanism enables the same group of seed user vectors to provide different weights when interest prediction is performed on different candidate users, so that the expression capability of the seed user vectors is improved, and the accuracy of model prediction is further improved.

In step S240, a target user of the candidate users is determined according to the similarity.

For example, all candidate users may be sorted in descending order according to the similarity score, and a certain number (e.g., 20) of the previous candidate users may be selected as the target users. All candidate users may also be sorted in an ascending order according to the similarity score, and a certain number (e.g., 20) of the candidate users may be selected as target users. A similarity threshold may also be preset, and the candidate users meeting the similarity threshold are taken as target users, for example, the similarity threshold may be set to 0.5, or may also be set to 0.7, which is not specifically limited by the present disclosure. For example, a candidate user with a similarity score greater than a similarity threshold may be the target user.

After the target users are determined, the advertisement quantitative delivery can be realized according to the number of the target users. Cost threshold control delivery can also be achieved by combining appeal and similarity distribution of advertisers. For example, when the similarity threshold is set to 0.5, only 1/3 of the target users have similarity scores greater than 0.7, and at this time, only the target users having similarity scores greater than 0.7 may be subjected to advertisement delivery to adjust the lower limit of the advertisement delivery cost. In the advertisement putting process, advertisements can be directionally put in packages of target users or media of communities where the advertisements are located by combining advertisement media forms (such as bill advertisements, self-service cabinets and the like) needing to be displayed, so that accurate crowd selection and hidden passenger mining are realized, and target user groups are reached.

In the exemplary embodiment of the present disclosure, based on the successful delivery experience in the early stage of the advertisement, in combination with a large amount of historical behavior information and the like in the user logistics scene, the accurate behavior characterization of the seed user can be obtained by using the YouTube DNN model fused with the Attention mechanism. In addition, when online real-time prediction is carried out, similarity scores between behavior representations of all candidate users in the user pool and the clustered seed user clustering centers can be calculated, and target extended crowds can be accurately selected according to the scores. For example, methods such as K-means clustering, mean shift clustering, density-based clustering, and the like may be performed on the seed user to achieve the purpose of reducing the amount of model calculation, which is not limited in this embodiment.

In an example embodiment, referring to fig. 5, targeted placement of advertisements may be accomplished according to steps S501 through S505.

Step S501, seed user groups are established: and establishing seed user groups of different third-level commodities based on the successful delivery experience in the early stage of the advertisement. For the old advertisement putting scene, the seed user group can be directly obtained. For a new advertisement putting scene, recalling the similarity of the advertisement materials according to i2i (item-item similarity is calculated), and using seed users of similar products for hidden customer excavation of the current advertisement;

step S502, generating a seed user characterization vector and a candidate user characterization vector: learning high-order representation of user features through a YouTube DNN model fused with an Attention mechanism to obtain a seed user characterization vector and a candidate user characterization vector;

s503, similarity estimation: taking the seed user characterization vector and the candidate user characterization vector as the input of a basic model (a full connection layer and a PReLU activation function), and calculating the cosine similarity between the two vectors after vector space reconversion is carried out by an attention layer to obtain a similarity pre-estimation score;

step S504, determining target population: after the similarity pre-estimation scores of the candidate users and the seed users are obtained, a plurality of candidate users meeting the similarity threshold value can be determined to serve as a target group pool for current advertisement delivery according to fund or delivery quantity requirements of an advertiser;

step S505, offline advertisement putting: each user corresponds to one or more types of offline advertisement reaching media, and under the constraint of advertiser capital, the actual advertisement delivery crowd with maximized accumulated similarity and the corresponding offline reaching media can be obtained by solving the multidimensional planning problem, so that the gain maximization of the currently delivered advertisement is ensured.

It should be noted that although the various steps of the methods of the present disclosure are depicted in the drawings in a particular order, this does not require or imply that these steps must be performed in this particular order, or that all of the depicted steps must be performed, to achieve desirable results. Additionally or alternatively, certain steps may be omitted, multiple steps combined into one step execution, and/or one step broken down into multiple step executions, etc.

Further, in the present exemplary embodiment, a target user determination apparatus is also provided, and the apparatus may be applied to a server or a terminal device. Referring to fig. 6, the target user determination apparatus 600 may include a feature vector generation module 610, a characterization vector generation module 620, a similarity calculation module 630, and a target user determination module 640, wherein:

the feature vector generation module 610 is configured to obtain multi-dimensional feature data of a plurality of users, and generate a user feature vector and an advertisement feature vector according to the multi-dimensional feature data, where the plurality of users include a seed user and a candidate user;

a feature vector generation module 620, configured to input the user feature vector and the advertisement feature vector into a user feature model, and generate a user feature vector, where the user feature vector includes a first seed user vector and a first candidate user vector;

a similarity calculation module 630, configured to input the first seed user vector and the first candidate user vector into a similarity prediction model, generate a second seed user vector and a second candidate user vector, and calculate a similarity between the second seed user vector and the second candidate user vector;

a target user determining module 640, configured to determine a target user of the candidate users according to the similarity.

In an alternative embodiment, the seed user comprises one of a seed user who has historically placed an advertisement and a seed user who is to be placed an advertisement; the target user determination device 600 further includes:

the system comprises a feature vector acquisition module, a feature vector acquisition module and a feature vector acquisition module, wherein the feature vector acquisition module is used for acquiring a first advertisement feature vector of an advertisement to be delivered and a second advertisement feature vector of a historical advertisement delivered;

the target advertisement determining module is used for calculating the similarity between the first advertisement characteristic vector and the second advertisement characteristic vector and determining a target advertisement similar to the advertisement to be delivered in the historical advertisement delivery according to the similarity;

and the seed user determining module is used for constructing the seed user to be advertised based on the seed user targeted to be advertised.

In an alternative embodiment, the multi-dimensional feature data comprises user feature data and advertisement feature data; the feature vector generation module 610 includes:

the first vectorization module is used for vectorizing the user feature data to generate the user feature vector;

and the second vectorization module is used for vectorizing the advertisement characteristic data to generate the advertisement characteristic vector.

In an alternative embodiment, the user characteristic data comprises at least one of discrete characteristic data and continuous characteristic data; the first vector quantization module comprises:

the first characteristic data processing submodule is used for counting the discrete characteristic data to obtain a plurality of first characteristic data;

a first feature vector generation submodule for encoding each of the first feature data to generate a plurality of first feature vectors;

a second feature vector generation submodule for normalizing each of the continuous feature data to generate a plurality of second feature vectors;

and the first vector generation submodule is used for splicing the plurality of first feature vectors and the plurality of second feature vectors to generate the user feature vector.

In an alternative embodiment, the user characterization model includes at least a first fully connected layer and a first attention layer; the characterization vector generation module 620 includes:

the first feature dimension reduction submodule is used for performing feature dimension reduction on the user feature vector and the advertisement feature vector by using the first full connection layer to obtain a user intermediate vector and an advertisement intermediate vector;

a second vector generation sub-module for generating the first seed user vector and the first candidate user vector through the first attention layer based on the user intermediate vector and the advertisement intermediate vector.

In an optional embodiment, the similarity prediction model includes two network branches, each network branch includes at least a second full connection layer and a second attention layer; the similarity calculation module 630 includes:

the second feature dimension reduction sub-module is used for performing feature dimension reduction on the first seed user vector and the first candidate user vector by using the second full connection layer to obtain a seed user intermediate vector and a candidate user intermediate vector;

a third vector generation sub-module configured to generate the second seed user vector and the second candidate user vector through the second attention layer based on the seed user intermediate vector and the candidate user intermediate vector.

In an optional embodiment, the target user determining apparatus 600 further includes a first model training module, configured to obtain first training samples, where each of the first training samples is a user feature vector and an advertisement feature vector of a candidate user; inputting the user characteristic vector and the advertisement characteristic vector of each candidate user into the user characterization model to obtain a corresponding user characterization vector and an advertisement characterization vector; constructing a first objective function according to the user characterization vector and the advertisement characterization vector; and iteratively updating the parameters of the user characterization model based on the first objective function, and finishing the training of the user characterization model when an iteration termination condition is met.

In an optional implementation, the target user determination apparatus 600 further includes a second model training module, configured to obtain second training samples, where each of the second training samples is a vector pair consisting of a seed user vector and a target user vector; determining the similarity of the seed user vector and the target user vector through the similarity pre-estimation model; constructing a second objective function according to the similarity of the seed user vector and the target user vector; and iteratively updating the parameters of the similarity estimation model based on the second objective function, and finishing the training of the similarity estimation model when an iteration termination condition is met.

The specific details of each module in the target user determining apparatus have been described in detail in the corresponding target user determining method, and therefore are not described herein again.

Each module in the above device may be a general-purpose processor, including: a central processing unit, a network processor, etc.; but may also be a digital signal processor, an application specific integrated circuit, a field programmable gate array or other programmable logic device, discrete gate or transistor logic, discrete hardware components. The modules may also be implemented in software, firmware, etc. The processors in the above device may be independent processors or may be integrated together.

Exemplary embodiments of the present disclosure also provide a computer-readable storage medium having stored thereon a program product capable of implementing the above-described method of the present specification. In some possible embodiments, various aspects of the disclosure may also be implemented in the form of a program product comprising program code for causing an electronic device to perform the steps according to various exemplary embodiments of the disclosure as described in the above-mentioned "exemplary methods" section of this specification, when the program product is run on the electronic device. The program product may employ a portable compact disc read only memory (CD-ROM) and include program code, and may be run on an electronic device, such as a personal computer. However, the program product of the present disclosure is not limited thereto, and in this document, a readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

A computer readable signal medium may include a propagated data signal with readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A readable signal medium may be any readable medium that is not a readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

Program code embodied on a readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

Program code for carrying out operations for aspects of the present disclosure may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any kind of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or may be connected to an external computing device (e.g., through the internet using an internet service provider).

The exemplary embodiment of the present disclosure also provides an electronic device capable of implementing the above method. An electronic device 700 according to such an exemplary embodiment of the present disclosure is described below with reference to fig. 7. The electronic device 700 shown in fig. 7 is only an example and should not bring any limitation to the functions and the scope of use of the embodiments of the present disclosure.

As shown in fig. 7, electronic device 700 may take the form of a general purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one memory unit 720, a bus 730 that connects the various system components (including the memory unit 720 and the processing unit 710), and a display unit 740.

The storage unit 720 stores program code that may be executed by the processing unit 710 to cause the processing unit 710 to perform steps according to various exemplary embodiments of the present disclosure described in the "exemplary methods" section above in this specification. For example, processing unit 710 may perform one or more of the method steps of any of fig. 2 and 5.

The storage unit 720 may include readable media in the form of volatile memory units, such as a random access memory unit (RAM) 721 and/or a cache memory unit 722, and may further include a read only memory unit (ROM) 723.

The memory unit 720 may also include programs/utilities 724 having a set (at least one) of program modules 725, such program modules 725 including, but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which, or some combination thereof, may comprise an implementation of a network environment.

Bus 730 may be any representation of one or more of several types of bus structures, including a memory unit bus or memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

The electronic device 700 may also communicate with one or more external devices 800 (e.g., keyboard, pointing device, bluetooth device, etc.), with one or more devices that enable a user to interact with the electronic device 700, and/or with any devices (e.g., router, modem, etc.) that enable the electronic device 700 to communicate with one or more other computing devices. Such communication may occur via an input/output (I/O) interface 750. Also, the electronic device 700 may communicate with one or more networks (e.g., a Local Area Network (LAN), a Wide Area Network (WAN), and/or a public network such as the Internet) via the network adapter 760. As shown, the network adapter 760 communicates with the other modules of the electronic device 700 via the bus 730. It should be appreciated that although not shown in the figures, other hardware and/or software modules may be used in conjunction with the electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.

Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein may be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure may be embodied in the form of a software product, which may be stored in a non-volatile storage medium (which may be a CD-ROM, a usb disk, a removable hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which may be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the exemplary embodiments of the present disclosure.

Furthermore, the above-described figures are merely schematic illustrations of processes included in methods according to exemplary embodiments of the present disclosure, and are not intended to be limiting. It will be readily understood that the processes shown in the above figures are not intended to indicate or limit the chronological order of the processes. In addition, it is also readily understood that these processes may be performed, for example, synchronously or asynchronously in multiple modules.

It should be noted that although in the above detailed description several modules or units of the device for action execution are mentioned, such a division is not mandatory. Indeed, the features and functionality of two or more modules or units described above may be embodied in one module or unit, according to embodiments of the present disclosure. Conversely, the features and functions of one module or unit described above may be further divided into embodiments by a plurality of modules or units.

It will be understood that the present disclosure is not limited to the precise arrangements that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for target user determination, comprising:

inputting the first seed user vector and the first candidate user vector into a similarity prediction model to generate a second seed user vector and a second candidate user vector, and calculating the similarity between the second seed user vector and the second candidate user vector;

2. The method of claim 1, wherein the seed users comprise one of seed users who have historically placed advertisements and seed users who are to be placed advertisements; before the obtaining of the multi-dimensional feature data of the plurality of users, the method includes:

acquiring a first advertisement characteristic vector of an advertisement to be delivered and a second advertisement characteristic vector of a historical advertisement to be delivered;

and constructing the seed users to be advertised based on the seed users targeted to be advertised.

3. The targeted user determination method of claim 1, wherein the multi-dimensional feature data comprises user feature data and advertisement feature data; the generating of the user feature vector and the advertisement feature vector according to the multi-dimensional feature data includes:

vectorizing the user feature data to generate the user feature vector;

4. The target user determination method of claim 3, wherein the user characteristic data comprises at least one of discrete characteristic data and continuous characteristic data; the vectorizing the user feature data to generate the user feature vector includes:

5. The target-user determination method of claim 1, wherein the user characterization model comprises at least a first fully connected layer and a first attention layer; inputting the user feature vector and the advertisement feature vector into a user characterization model to generate a user characterization vector, including:

6. The method according to claim 1, wherein the similarity pre-estimation model comprises two network branches, each network branch comprising at least a second full connection layer and a second attention layer; the inputting the first seed user vector and the first candidate user vector into a similarity pre-estimation model to generate a second seed user vector and a second candidate user vector, including:

7. The method of claim 1, further comprising:

8. The method of claim 1, further comprising:

9. A target user determination apparatus, comprising:

10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the method of any one of claims 1 to 8.

11. An electronic device, comprising:

a processor; and

a memory for storing executable instructions of the processor;

wherein the processor is configured to perform the method of any of claims 1-8 via execution of the executable instructions.