CN112559777A

CN112559777A - Content item delivery method and device, computer equipment and storage medium

Info

Publication number: CN112559777A
Application number: CN201910909517.5A
Authority: CN
Inventors: 盛名扬; 梅晓茸; 刘志山
Original assignee: Beijing Dajia Internet Information Technology Co Ltd
Current assignee: Beijing Dajia Internet Information Technology Co Ltd
Priority date: 2019-09-25
Filing date: 2019-09-25
Publication date: 2021-03-26
Anticipated expiration: 2039-09-25

Abstract

The disclosure relates to a content item delivery method, a content item delivery device, computer equipment and a storage medium, and belongs to the technical field of multimedia. According to the method, the user vectors are obtained through the content item obtaining request of the terminal, the plurality of content item vectors are obtained from the candidate feature library, different candidate content items are formed by different content materials in different combination modes, the traversing combination of the machine can improve the breadth and the depth of the content material combination, based on the user vectors, the plurality of content item vectors and the sequencing model, the content items corresponding to the second feature vectors are sequenced, the content items sequenced in the front target number are output, the content items in the front target number are delivered to the terminal, the content items formed in different combination modes can be delivered to the terminals of different users in a targeted mode, and the delivery effect of the content items is greatly improved.

Description

Content item delivery method and device, computer equipment and storage medium

Technical Field

The present disclosure relates to the field of multimedia technologies, and in particular, to a content item delivery method and apparatus, a computer device, and a storage medium.

Background

With the development of multimedia technology, advertisers can cooperate with platform operators to launch content items such as advertisements on a specified platform, so that a user terminal can browse the advertisements based on the specified platform, for example, the specified platform can be a website, an application client, a television column, and the like.

At present, before an advertiser submits an advertisement to be delivered to a platform operator, in order to achieve a better delivery effect, a plurality of advertisement materials such as videos, covers, documents and the like are usually made, and the advertisement materials such as the videos, the covers, the documents and the like are manually combined to form a plurality of combined advertisements. Furthermore, a putting experiment is executed on line in advance, advertisements with good putting effect are selected, and therefore the selected advertisements are put on a specified platform.

In the above process, for the same product, after the advertiser selects the advertisement through the delivery experiment, each user terminal receives the same advertisement, and because the preferences of different users for the advertisement are usually different, the delivery effect of the content items such as the advertisement is poor.

Disclosure of Invention

The present disclosure provides a content item delivery method, apparatus, computer device, and storage medium, to at least solve the problem of poor content item delivery effect in the related art. The technical scheme of the disclosure is as follows:

according to a first aspect of embodiments of the present disclosure, there is provided a content item delivery method, including:

acquiring a user vector of a user corresponding to a terminal according to a content item acquisition request of the terminal, wherein the user vector is used for representing personal characteristics of the user;

obtaining a plurality of content item vectors from a candidate feature library, wherein one content item vector is used for representing the content feature of one content item, the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold, the candidate feature library comprises the content item vectors of a plurality of candidate content items, and different candidate content items are formed by different content materials in different combination modes;

based on the user vector, the content item vectors and a sequencing model, sequencing the content items corresponding to the content item vectors, and outputting the content items with sequencing in the front target number;

and delivering the content items of the front target number to the terminal.

In one possible implementation, the sorting the content items corresponding to the content item vectors based on the user vector, the content item vectors, and a sorting model, and outputting a top-target number of content items according to a sorting order includes:

obtaining user features corresponding to the user vectors and a plurality of content item features corresponding to the plurality of content item vectors;

and inputting the user characteristics and the plurality of content item characteristics into the sequencing model, sequencing the content items corresponding to the content item vectors through the sequencing model, and outputting the content items with the sequencing quantity at the front target.

In one possible implementation, the sorting the content items corresponding to the content item vectors, and outputting the content items with the top target number of sorting includes:

when the sequencing model is a click rate estimation model, sequencing the content items corresponding to the content item vectors according to the sequence of click rate from large to small through the click rate estimation model, and outputting the content items with the click rate in the number of the front targets.

and when the sequencing model is a conversion rate estimation model, sequencing the content items corresponding to the content item vectors according to the sequence of the conversion rates from large to small through the conversion rate estimation model, and outputting the content items with the conversion rates at the front target quantity.

In one possible implementation, the obtaining, from the candidate feature library, a plurality of content item vectors includes:

acquiring a candidate data set corresponding to the user, wherein the candidate data set comprises a plurality of content materials;

obtaining the plurality of content item vectors from a candidate feature library corresponding to the candidate data set.

In one possible implementation, before obtaining the plurality of content item vectors from the candidate feature library, the method further includes:

performing iterative training on the initial model based on the plurality of sample user information and the plurality of sample content materials;

and when a training stopping condition is met, obtaining the sequencing model and content item vectors corresponding to the content items, wherein the content items are formed by the sample content materials in various combination modes.

In one possible embodiment, the iteratively training the initial model based on the plurality of sample user information and the plurality of sample content materials comprises:

inputting the plurality of sample user information into a first deep network, and performing weighting processing on the plurality of sample user information through the first deep network to obtain a plurality of sample user vectors, wherein the plurality of sample user vectors correspond to the plurality of sample user information;

inputting the sample content materials into a second deep network, and performing weighting processing on the sample content materials through the second deep network to obtain a plurality of sample content item vectors, wherein different sample content item vectors are used for representing content items formed by the sample content materials of different types in different combination modes;

and obtaining a loss function value of the training based on the sample user vectors and the sample content item vectors, and when the training stopping condition is not met, iteratively executing the operation executed in the training process.

In one possible implementation, the second deep network includes a target sub-network and a plurality of deep sub-networks, one deep sub-network corresponding to one type of sample content material;

the inputting the plurality of sample content materials into a second deep network, and performing weighting processing on the plurality of sample content materials through the second deep network to obtain a plurality of sample content item vectors includes:

respectively inputting different types of sample content materials into corresponding depth sub-networks, and respectively performing weighting processing on the sample content materials of each type through each depth sub-network to obtain the characteristic vectors of the sample content materials of each type;

and inputting the feature vectors of different types of sample content materials into the target sub-network based on different combination modes, performing weighting processing on the different types of sample content materials through the target sub-network, and outputting the plurality of sample content item vectors.

In one possible embodiment, the method further comprises:

and adjusting parameters of the first deep network at each interval of target time length based on the behavior information of the user collected in the target time length.

In a possible implementation manner, the obtaining, according to a content item obtaining request of a terminal, a user vector of a user corresponding to the terminal includes:

acquiring behavior information and material information of the user based on the content item acquisition request;

and acquiring a user vector corresponding to the behavior information and the data information based on the first deep network with the maximum timestamp.

According to a second aspect of embodiments of the present disclosure, there is provided a content item delivery apparatus comprising:

a first obtaining unit, configured to execute obtaining a user vector of a user corresponding to a terminal according to a content item obtaining request of the terminal, wherein the user vector is used for representing personal characteristics of the user;

a second obtaining unit configured to perform obtaining a plurality of content item vectors from a candidate feature library, wherein one content item vector is used for representing content features of one content item, the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold, the candidate feature library comprises content item vectors of a plurality of candidate content items, and different candidate content items are formed by different content materials in different combination manners;

a sorting unit configured to perform sorting of content items corresponding to respective content item vectors based on the user vector, the plurality of content item vectors, and a sorting model, and output a content item sorted in a top target number;

a delivery unit configured to perform delivery of the pre-target number of content items to the terminal.

In one possible embodiment, the sorting unit is configured to perform:

In one possible implementation, the second obtaining unit is configured to perform:

In one possible embodiment, the apparatus further comprises:

a training unit configured to perform iterative training of the initial model based on the plurality of sample user information and the plurality of sample content materials;

a deriving unit configured to perform deriving the ranking model and content item vectors corresponding to respective content items formed by the plurality of sample content materials in various combinations when a stop training condition is satisfied.

In one possible embodiment, the training unit comprises:

a first weighting subunit, configured to perform input of the plurality of sample user information into a first deep network, and perform weighting processing on the plurality of sample user information through the first deep network to obtain a plurality of sample user vectors, where the plurality of sample user vectors correspond to the plurality of sample user information;

a second weighting subunit configured to perform input of the plurality of sample content materials into a second deep network, and perform weighting processing on the plurality of sample content materials through the second deep network to obtain a plurality of sample content item vectors, where different sample content item vectors are used to represent respective content items formed by different combinations of the plurality of different types of sample content materials;

and the obtaining iteration subunit is configured to perform the operation executed in the training process based on the plurality of sample user vectors and the plurality of sample content item vectors to obtain the loss function value of the training, and when the training stopping condition is not met, the obtaining iteration subunit is configured to perform the operation executed in the training process in an iteration mode.

the second weighting subunit is configured to perform:

In one possible embodiment, the apparatus is further configured to perform:

In one possible implementation, the first obtaining unit is configured to perform:

According to a third aspect of embodiments of the present disclosure, there is provided a computer device comprising:

one or more processors;

one or more memories for storing the one or more processor-executable instructions;

wherein the one or more processors are configured to perform:

and delivering the content items of the front target number to the terminal.

According to a fourth aspect of embodiments of the present disclosure, there is provided a storage medium having at least one instruction which, when executed by one or more processors of a computer device, enables the computer device to perform a content item delivery method, the method comprising:

and delivering the content items of the front target number to the terminal.

According to a fifth aspect of embodiments of the present disclosure, there is provided a computer program product comprising one or more instructions which, when executed by one or more processors of a computer device, enable the computer device to perform a method of content item delivery, the method comprising:

and delivering the content items of the front target number to the terminal.

The technical scheme provided by the embodiment of the disclosure at least brings the following beneficial effects:

acquiring a user vector of a user corresponding to a terminal through a content item acquisition request of the terminal, thereby acquiring a plurality of content item vectors from a candidate feature library, wherein the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold value, the candidate feature library comprises the content item vectors of a plurality of candidate content items, different candidate content items are formed by different content materials in different combination modes, the traversing combination of a machine can improve the breadth and the depth of the content material combination, content items corresponding to each second feature vector are sorted based on the user vector, the plurality of content item vectors and a sorting model, content items in a front target number are output and sorted, the content items in the front target number are delivered to the terminal, thereby the content items formed in different combination modes can be delivered in a targeted manner to the terminals of different users, the delivery effect of the content items is greatly 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 and are not to be construed as limiting the disclosure.

Fig. 1 is a flow chart illustrating a method of content item delivery, according to an example embodiment.

FIG. 2 is a flow diagram illustrating a ranking model training method according to an exemplary embodiment.

Fig. 3 is a schematic training diagram of a ranking model according to an embodiment of the present disclosure.

Fig. 4 is a flow chart illustrating a method of content item delivery, according to an example embodiment.

Fig. 5 is a schematic diagram of a content item delivery method provided by an embodiment of the present disclosure.

Fig. 6 is a block diagram illustrating a logical structure of a content item delivery apparatus, according to an example embodiment.

Fig. 7 is a schematic structural diagram of a computer device according to an embodiment of the present disclosure.

Detailed Description

In order to make the technical solutions of the present disclosure better understood by those of ordinary skill in the art, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

It should be noted that the terms "first," "second," and the like in the description and claims of the present disclosure and in the above-described drawings are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the disclosure described herein are capable of operation in sequences other than those illustrated or otherwise described herein. The implementations described in the exemplary embodiments below are not intended to represent all implementations consistent with the present disclosure. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure, as detailed in the appended claims.

The user information to which the present disclosure relates may be information authorized by the user or sufficiently authorized by each party.

The content items to which the present disclosure relates are formed from a combination of different types of content material, the types of content material may include at least one of a video, a document, a cover, or a content item location. For example, the content item may be a video-type advertisement or a teletext-type advertisement.

The user vector related to the present disclosure is used to represent a personal feature of a user, for example, the user vector may be in the form of an embedding (embedding) vector, and at this time, the user vector may be referred to as "user embedding", where the user vector refers to a vector representation of an irregular user feature in an embedding space, that is, the irregular user feature is mapped to the embedding space, and a feature matrix with a fixed size may be obtained, where the feature matrix is the user vector.

The content item vector according to the present disclosure is used to represent content features of a content item, for example, the content item vector may also be in the form of an embedded vector, and at this time, the content item vector may be referred to as "content item embedding", the content item vector refers to a vector representation of irregular content item features in an embedding space, that is, irregular content item features are mapped to the embedding space, and a feature matrix with a fixed size may be obtained, and this feature matrix is the content item vector.

It should be noted that, in the present disclosure, each content material that can be selected during the delivery and each sample content material that is used during the training may be the same content material, and in order to distinguish the training scenario from the delivery scenario, the content material is referred to as "sample content material" in the training scenario, and is referred to as "content material" in the delivery scenario.

Fig. 1 is a flowchart illustrating a content item delivery method according to an exemplary embodiment, and referring to fig. 1, the embodiment is applied to a computer device, and details are described below by taking the computer device as an example:

in step 101, the server obtains a user vector of a user corresponding to the terminal according to a content item obtaining request of the terminal, wherein the user vector is used for representing personal characteristics of the user.

In step 102, the server obtains a plurality of content item vectors from a candidate feature library, wherein one content item vector is used for representing the content feature of one content item, the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold, the candidate feature library comprises the content item vectors of a plurality of candidate content items, and different candidate content items are formed by different content materials in different combinations.

In step 103, the server sorts the content items corresponding to each content item vector based on the user vector, the plurality of content item vectors and the sorting model, and outputs the content items sorted in the top target number.

In step 104, the server delivers the previous target number of content items to the terminal.

The method provided by the embodiment of the disclosure obtains a user vector of a user corresponding to a terminal through a content item obtaining request of the terminal, thereby obtaining a plurality of content item vectors from a candidate feature library, wherein the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold, the candidate feature library comprises content item vectors of a plurality of candidate content items, different candidate content items are formed by different content materials in different combination modes, the traversal combination of a machine can improve the breadth and depth of the content material combination, based on the user vector, the plurality of content item vectors and a ranking model, ranking content items corresponding to respective second feature vectors, outputting content items ranked in a front target number, delivering the front target number of content items to the terminal, thereby being capable of targeting terminals of different users, the content items formed by different combination modes are put in a targeted manner, so that the putting effect of the content items is greatly improved.

In one possible implementation, sorting the content items corresponding to the respective content item vectors based on the user vector, the plurality of content item vectors and a sorting model, and outputting a top-targeted number of content items for sorting comprises:

acquiring a user characteristic corresponding to the user vector and a plurality of content item characteristics corresponding to the plurality of content item vectors;

inputting the user characteristic and the plurality of content item characteristics into the sequencing model, sequencing the content items corresponding to each content item vector through the sequencing model, and outputting the content items with sequencing positioned at the front target number.

In one possible implementation, sorting the content items corresponding to the respective content item vectors, and outputting the content items sorted by the top target number includes:

when the sequencing model is a conversion rate estimation model, sequencing the content items corresponding to the content item vectors according to the sequence of conversion rates from large to small through the conversion rate estimation model, and outputting the content items with the conversion rates at the front target number.

In one possible implementation, obtaining a plurality of content item vectors from the candidate feature library comprises:

the plurality of content item vectors are obtained from a candidate feature library corresponding to the candidate data set.

and when the training stopping condition is met, obtaining the sequencing model and content item vectors corresponding to the content items, wherein the content items are formed by the sample content materials in various combination modes.

In one possible embodiment, iteratively training the initial model based on the plurality of sample user information and the plurality of sample content materials comprises:

inputting the plurality of sample content materials into a second deep network, and performing weighting processing on the plurality of sample content materials through the second deep network to obtain a plurality of sample content item vectors, wherein the weighting processing comprises:

based on different combination modes, the feature vectors of different types of sample content materials are input into the target sub-network, the different types of sample content materials are weighted through the target sub-network, and the plurality of sample content item vectors are output.

In one possible embodiment, the method further comprises:

In one possible implementation manner, the obtaining, according to a content item obtaining request of a terminal, a user vector of a user corresponding to the terminal includes:

acquiring behavior information and data information of the user based on the content item acquisition request;

All the above optional technical solutions may be combined arbitrarily to form the optional embodiments of the present disclosure, and are not described herein again.

Fig. 2 is a flowchart illustrating a ranking model training method according to an exemplary embodiment, and referring to fig. 2, before describing a content item delivery method provided by the embodiment of the present disclosure, a ranking model training method applied to a computer device is described, which is described below by taking the computer device as a server as an example.

The ranking model provided by the embodiment of the disclosure is obtained by iterative training of an initial model, where the initial model may include a first deep network and a second deep network, where the second deep network may include a target subnetwork and a plurality of deep subnetworks, and one deep subnetwork corresponds to one type of sample content material, and on this basis, taking a user vector as user embedding and a content item vector as content item embedding as examples, a server may perform the following training steps:

in step 201, the server inputs a plurality of sample user information into a first deep network, and performs weighting processing on the plurality of sample user information through the first deep network to obtain a plurality of sample user vectors, where the plurality of sample user vectors correspond to the plurality of sample user information.

Optionally, at least one of profile information or behavior information of the user may be included in each sample user information, for example, the profile information may include at least one of a name, a gender, an age, a nickname, a occupation, or a geographic location of the user, and the behavior information may include at least one of historical browsing behavior information or historical consumption behavior information of the user.

In the above process, the sample user information may be stored in the local database, or may be obtained by downloading from the cloud database, and of course, the sample user information may also be sent to the server by the terminal. In some embodiments, the server may store profile information submitted when the user registers the account in a local database, and acquire behavior information of the user from the terminal in real time.

In step 201, after obtaining a plurality of sample user information, the server may perform feature extraction on the plurality of sample user information to obtain a plurality of sample user features, input the plurality of sample user features into a first Deep network (DNN) of the initial model, where the first Deep network may include a plurality of hidden layers, the hidden layers are connected in series, perform weighting processing on the sample user features through the plurality of hidden layers, and output a sample user embedding (embedding) vector by a last hidden layer, where each sample user vector corresponds to one sample user information.

In step 202, the server inputs different types of sample content materials into the depth sub-networks corresponding to the types in the second depth network, and performs weighting processing on the sample content materials of the types through the depth sub-networks to obtain feature vectors of the sample content materials of the types.

In the above process, the type of the sample content material may include at least one of a video, a document, a cover, or a content item position, and different types of sample content materials may be combined in different combination manners to form different content items, for example, for the same video material, the same cover material, and the same content item position material, the three may be combined with different document materials to obtain different content items.

Optionally, the content items may be video-type or image-text-type, that is, each content item does not necessarily include all types of sample content materials, and in the case of the content item being an advertisement, the video-type advertisement may include a video, a file, a cover, an advertisement position, a sticker, and an advertisement, and the image-text-type advertisement may not include a video but carry a picture, a file, an advertisement position, a sticker, and an advertisement.

It should be noted that, when obtaining each sample content material, the advertiser may submit the content material to be delivered to the administrative user, and the administrative user enters each content material submitted by different advertisers into the material library, which may include a plurality of data sets, and each content material of each data set in the material library is used as the sample content material during training, where the material library may be stored locally or in the cloud.

On the basis, the advertiser does not need to perform online putting experiments, content items are optimized with low efficiency, but only content materials to be put can be submitted to the management user, the management user inputs the content materials into a data set of a material library, and a ranking model is trained on the basis of the material library, so that the content items most matched with the users can be intelligently selected by a machine in the subsequent putting process (namely, the material combination mode with the highest matching degree with the users), the labor cost for combining the content materials is greatly saved, the putting effect of the content items is improved, and the specific putting process is detailed in the next embodiment.

In the process, the server sets a depth sub-network for each type of sample content material in the second depth network, so that feature extraction is performed on each sample content material to obtain material features of each sample content material, the material features of the same type of sample content material are input into the same depth sub-network, the material features of different types of sample content materials are input into different depth sub-networks, and each depth sub-network performs weighting processing on each input material feature to obtain a feature vector of each sample content material.

Fig. 3 is a schematic diagram of training a ranking model provided in an embodiment of the present disclosure, referring to fig. 3, a server may set a video DNN, a cover DNN, a document DNN, and a content item location DNN in a second deep network, after extracting features of all video materials, the server inputs material features of each video material into the video DNN, performs weighting processing on the material features of each video material through the video DNN, and can output feature vectors of each video material through forward propagation. In the ranking model shown in fig. 3, since the left side is a first deep network for obtaining user vectors and the right side is a second deep network for obtaining content item vectors, the ranking model can be referred to as a "double tower model" visually, and after each deep sub-network outputs the feature vectors of each sample content material, the feature vectors of each sample content material can be input into a target sub-network, and the following step 203 is performed.

In step 203, the server inputs the feature vectors of the different types of sample content materials into the target subnetwork based on the different combination modes, and performs weighting processing on the different types of sample content materials through the target subnetwork to output the plurality of sample content item vectors.

In step 203, after obtaining the feature vectors of the sample content materials of the respective types, the server can form different sample content items after arranging and combining the sample content materials of the different types, so that the server can input the feature vectors of the sample content materials of the different types into the target sub-network based on different combination modes, and calculate the embedded vector of each sample content item through the target sub-network.

In the process, the machine can traverse all the combination modes of all the sample content materials, so that the embedding vectors of all the sample content items formed by all the combination modes are calculated, and compared with the situation that all the optional combinations of the content materials are difficult to exhaust when the combination is carried out manually, the traversing combination based on the machine in the embodiment of the disclosure can greatly improve the combination width and depth of the content materials, is favorable for excavating the combination modes with better delivery effect, and further improves the delivery effect of the content items.

In step 202-.

In the process, different depth sub-networks are configured in the second depth network for different types of sample content materials, and the depth sub-networks are relatively independent, so that any depth sub-network can be customized and adjusted, the generalization capability, flexibility and portability of the second depth network are greatly improved, and further, the feature vectors of the sample content materials of various types are fused by the target sub-network, so that the embedded vectors of the sample content items are calculated, and the method can be better adapted to large data scenes such as a massive material library.

In the above-described case, one possible implementation is provided in which the server configures a deep sub-network for each type of sample content material, in some embodiments, the server may target different types of sample content material, and may also not be limited to deep networks, other network structures besides the deep network may be correspondingly provided, for example, CNN (Convolutional Neural network) is used to extract feature vectors of video materials, LSTM (Long Short-Term Memory) is used to extract feature vectors of document materials, VGG (Visual Geometry Group) is used to extract feature vectors of cover materials, DNN is used to extract feature vectors of content item location materials, therefore, the feature vectors of different types of sample content materials can be extracted more flexibly, and the accuracy of each sub-network is improved.

In step 204, the server obtains the loss function value of the current training based on the plurality of sample user vectors and the plurality of sample content item vectors, and when the training stopping condition is not satisfied, iteratively executes the operation executed in the training process.

In the above process, the server may input the plurality of sample user vectors into the activation layer, perform nonlinear processing on each sample user vector through an activation function in the activation layer, input each sample user vector after the nonlinear processing into the normalization layer, perform normalization processing on each sample user vector based on an exponential normalization (softmax) function in the normalization layer to obtain a plurality of sample user feature matrices, further, input the plurality of sample content item vectors into the activation layer and the normalization layer in sequence, perform a step similar to the above process, and obtain a plurality of sample content item feature matrices. The activation function may be a function of tanh, ReLU, sigmoid, or the like.

Further, the server may calculate similarity between any sample user feature matrix and any sample content item feature matrix, perform an operation performed by calculating the similarity on each sample user feature matrix and each sample content item feature matrix, obtain a plurality of similarities between the plurality of users and the plurality of sample content items, and obtain a loss function value of the current training based on the plurality of similarities. Optionally, the similarity may be an inverse of an euclidean distance, or may be an inverse of a cosine distance, and the like, and the measurement index of the similarity is not specifically limited in the embodiment of the present disclosure.

It should be noted that the initial model may have different optimization objectives. For example, when the initial model is a Click Through Rate (CTR) estimation model, the loss function may be defined as a mean square error between the predicted Click Through Rate and the actual Click Through Rate, and accordingly, the optimization objective may be that the mean square error between the predicted Click Through Rate and the actual Click Through Rate is smaller than a second objective threshold; for another example, when the initial model is a ConVersion Rate (CVR) prediction model, the loss function may be defined as a mean square error between the predicted ConVersion Rate and the actual ConVersion Rate, and accordingly, the optimization objective may be that the mean square error between the predicted ConVersion Rate and the actual ConVersion Rate is smaller than a third objective threshold.

The second target threshold or the third target threshold is any value greater than or equal to 0, the second target threshold and the third target threshold may be the same or different, and the specific value of the second target threshold or the third target threshold is not limited in the embodiment of the present disclosure.

Based on the above example, a higher similarity means a better match between the user and the sample content item, when the initial model is the click-through rate estimation model, a better match between the sample content item and the user means a higher probability of the user clicking on the sample content item, and when the initial model is the conversion rate estimation model, a better match between the sample content item and the user means a higher probability of the user purchasing the product indicated by the sample content item.

In the above step 201-.

In step 205, when the stop training condition is satisfied, the server obtains the ranking model and content item vectors corresponding to the respective content items, wherein the respective content items are formed by the plurality of sample content materials in various combinations.

In the above process, the training stopping condition may be that the loss function value is smaller than the second target threshold or the third target threshold, and of course, the training stopping condition may also be that the number of iterations reaches the target number, and the content of the training stopping condition is not specifically limited in the embodiments of the present disclosure.

In the above process, because the initial model adjusts the parameters of each hidden layer in the first deep network or the second deep network based on the back propagation algorithm during iterative training, the content item vectors of each content item in each iterative process can also change, and when the training is completed, a trained ranking model can be obtained, and the content item vectors of each content item can also be used as a by-product of the training and stored in the feature library of the server.

In some embodiments, since usually the advertiser does not perform frequent update and update on the sample content material after submitting the sample content material, the server directly stores each content item vector in the feature library, which is equivalent to constructing a feature library for offline retrieval of content item vectors for calling of a subsequent delivery process. However, the behavior information of the user is continuously updated over time, so that the server can perform the real-time optimization on the first deep network by performing the following step 206.

In step 206, at each target duration, the server adjusts parameters of the first deep network based on the behavior information of the user collected in the target duration.

Any value greater than 0 may be used for the target duration, for example, the target duration may be 5 minutes, and the value of the target duration is not specifically limited in the embodiments of the present disclosure.

In the above process, the server is equivalent to target duration every interval, and performs one iterative training on the first deep network again based on the behavior information of the user collected in the target duration, and this training process of online updating may be referred to as "fine tuning" (fine tuning), so that the behavior information of the user can be analyzed in real time, and the problem of accuracy degradation of the ranking model due to hysteresis of the behavior information is avoided.

For example, every 5 minutes, the server adjusts parameters of the first deep network according to historical browsing behavior information and historical consumption behavior information of each user within the 5 minutes, and since the 5-minute interval is short and the behavior information of the user within the 5 minutes is unlikely to be greatly different from the behavior information of the last 5 minutes, such a target duration can avoid redundant computation amount caused by frequent training on the basis of ensuring model accuracy, and still achieve the effect similar to "real-time updating" of the first deep network.

According to the method provided by the embodiment of the disclosure, iterative training is performed on the initial model based on the information of a plurality of sample users and a plurality of sample content materials until a training stopping condition is met, the sequencing model and content item vectors corresponding to each content item are obtained, each content item vector is stored in the feature library, and the feature library for offline retrieval of the content item vector is constructed for calling a subsequent delivery process.

Furthermore, the machine can traverse all possible arrangement and combination modes when combining different types of sample content materials, so that the labor cost caused by manual combination can be saved, the combination breadth and depth of the content materials are greatly improved, the combination mode with a better delivery effect is favorably mined, the feature library for storing content item vectors is enriched, and the delivery effect of the content items is improved.

Furthermore, for each sub-network in the second deep network, a completely independent sub-network can be adopted for iterative training, so that the network structure of each sub-network can be flexibly configured according to the delivery requirement, the portability of the whole sequencing model is improved, and the iterative expansion of the sequencing model is facilitated.

Further, parameter adjustment is carried out on the first deep network at intervals of the target duration based on the behavior information of the user acquired in the target duration, so that the behavior information of the user can be analyzed in real time, and the problem that the accuracy of the sequencing model is reduced due to the delay of the behavior information is solved.

Fig. 4 is a flowchart illustrating a content item delivery method according to an exemplary embodiment, and referring to fig. 4, the embodiment may be implemented based on the ranking model trained by the foregoing embodiment, which is applied to a computer device, and the following description takes the computer device as a server, a user vector as user embedding, and a content item vector as content item embedding as examples.

In step 401, the server acquires behavior information and profile information of the user based on a content item acquisition request of the terminal.

In the above process, when the user accesses the server through the terminal for the first time, the terminal may send an account registration request to the server, where the account registration request may carry information of information authorized by the user to the server, such as at least one of name, gender, age, nickname, occupation, or geographic location, and the server may store the information of the user in the database according to the account identifier. Further, when the terminal subsequently sends a content item acquisition request to the server each time, at least one of the historical browsing behavior information or the historical consumption behavior information of the user may be carried in the content item acquisition request.

Based on the above situation, when the server acquires the behavior information and the data information of the user, the server may analyze the content item acquisition request of the terminal to obtain the account identifier corresponding to the user, the historical browsing behavior information and the historical consumption behavior information of the user, on one hand, determine the historical browsing behavior information and the historical consumption behavior information as the behavior information of the user, on the other hand, query whether the data information corresponding to the index can be hit in the database by using the analyzed account identifier as the index, and when the data information can be hit, acquire the data information of the user.

In step 402, the server obtains a user vector corresponding to the behavior information and the profile information based on the first deep network with the largest timestamp.

In the foregoing embodiment, it has been described that the server may perform online update on the first deep network every target time, and optionally, when the server finishes updating the first deep network every time, the server may allocate a timestamp to the first deep network after the update, where the timestamp may adopt a physical clock or a logical clock, and the like, as long as unidirectional increment is maintained, and the assignment manner of the timestamp is not specifically limited in the embodiments of the present disclosure.

Based on the above situation, when the server acquires the user vector, the server needs to acquire the user vector based on the first deep network with the largest timestamp. Specifically, after the first deep network is updated (i.e., parameter adjustment) in one round, the first deep network with the largest timestamp can be obtained, and the user vector of each user can be obtained at the same time, so that the corresponding relationship between the user and the user vector can be established.

In

step

401 and 402, the server obtains a user vector of a user corresponding to the terminal according to the content item obtaining request of the terminal, where the user vector is used to represent the personal characteristics of the user. In some embodiments, the server may also not perform online update on the first deep network, and then the server may determine the user vector directly based on the correspondence between the user and the user vector, thereby simplifying the process of obtaining the user vector.

In step 403, the server obtains a candidate data set corresponding to the user, the candidate data set including a plurality of content material.

In the above process, the server may store the orientation tag in correspondence with each data set in storing each data set, so that when the profile information of the user is consistent with the orientation tag of any data set, the data set is determined to be a candidate data set corresponding to the user.

For example, if a certain data set carries a label "female" or "city a", when the profile information of the user includes "female" or "city a", the data set is determined as a candidate data set corresponding to the user, and when the profile information of the user includes "male" or "city a", the data set is determined as not a candidate data set corresponding to the user.

The process of selecting candidate data sets from the various data sets is equivalent to directing the user to a specific one or more candidate data sets based on a certain directing strategy, so that the calculation amount of the subsequent sorting process is reduced by the screening mode.

In step 404, the server obtains a plurality of content item vectors from a candidate feature library corresponding to the candidate data set, one content item vector being used to represent the content feature of one content item, and the similarity between the plurality of content item vectors and the user vector being higher than a first target threshold.

In the above process, the candidate feature library comprises a content item vector of a plurality of candidate content items, different candidate content items being formed by different content material in different combinations.

Based on the ranking model training process in the above embodiment, the server obtains content item vectors of candidate content items corresponding to each candidate data set through traversal combination of the machine, the candidate content items are formed by each content material (as sample content material during training) in the candidate data set in various combination manners, and after the training is completed, the content item vectors of each candidate content item are stored in the feature library, and in addition, a corresponding relationship between the data set and the feature library can be established.

Optionally, since the number of candidate data sets may be one or more, the number of candidate feature libraries may also be one or more accordingly, and the number of candidate data sets or candidate feature libraries is not particularly limited in the embodiments of the present disclosure.

In the above case, the server may determine, according to the correspondence between the data set and the feature library, a candidate feature library corresponding to the candidate data set, and based on a similar vector search, obtain, from content item vectors of candidate content items pre-stored in the candidate feature library, a plurality of content item vectors whose similarity with the user vector obtained in the above step 402 is higher than a first target threshold.

In step 403 and step 404, the server may obtain a plurality of content item vectors from the candidate feature library, wherein the first target threshold is any value greater than or equal to 0. That is, the server can quickly recall, from the massive data of the candidate feature library, content item vectors whose similarity with the user vector is higher than the first target threshold value through the similar vector retrieval, and in the recall process, the server can pick out the content item vectors of thousands of content items which are relatively matched (that is, have similarity higher than the first target threshold value) from a data set with millions or even tens of millions of data sets.

Fig. 5 is a schematic diagram of a content item delivery method provided by an embodiment of the present disclosure, and referring to fig. 5, in some embodiments, in addition to recalling based on similarity vector retrieval, the server may perform the above recall process of content item vectors based on an e & e (exploration and exploitation) algorithm or a random exploration recall algorithm, and integrate the content item vectors recalled by multiple recall ways, and perform step 405 below.

In step 405, the server sorts the content items corresponding to the content item vectors based on the user vector, the plurality of content item vectors and the sorting model, and outputs the content items sorted in the top target number.

In the above process, the server may obtain the user feature corresponding to the user vector and the plurality of content item features corresponding to the plurality of content item vectors, input the user feature and the plurality of content item features into the ranking model, rank the content items corresponding to the respective content item vectors by the ranking model, and output the content items ranked in the front target number.

Optionally, the user features may be user features extracted based on the user information in step 201 of the previous embodiment, and in a subsequent training process, the server maps the user features to the embedding space through the first deep network, so as to obtain the user vector, and therefore, the user features, the user vector, and the user have a corresponding relationship therebetween.

Similarly, the content item feature may be a feature vector of each content material constituting the content item in step 202 of the previous embodiment, and in the subsequent training process, the server maps the feature vector of each content material to the embedding space through the target subnetwork in the second deep network, so as to obtain a content item vector, and therefore, the content item feature, the content item vector, and the content item have a corresponding relationship therebetween.

In some embodiments, the server may store, in the candidate feature library, each content item vector in correspondence with each content item identifier, so that, after a plurality of content item vectors are recalled based on the above-mentioned plurality of recall manners, a plurality of content item identifiers corresponding to the plurality of content item vectors are obtained according to a mapping relationship between the content item vectors and the content item identifiers, and further, each content item feature is stored in correspondence with each content item identifier in the original feature library, so that a plurality of content item features corresponding to the plurality of content item identifiers may be queried from the original feature library according to the obtained plurality of content item identifiers.

Similarly, the server may also store each user vector and the user identifier correspondingly, so that after the user vector is obtained, the user identifier corresponding to the user vector is obtained according to the mapping relationship between the user vector and the user identifier, and the user feature corresponding to the user identifier is obtained.

In the above process, since the user vector and the content item vector are generally embedding vectors, and the degree of abstraction of the embedding vectors is high, some details are inevitably lost, and by acquiring the user feature and the content item feature and then sorting based on the user feature and the content item feature, the problem that the embedding vectors lose detail information can be avoided, so that the sorting of the content items is more accurate.

In the above process, due to the difference of the optimization targets, the ranking model can be divided into a click rate pre-estimation model or a conversion rate pre-estimation model, which will be described below. Wherein the target number may be any positive integer greater than or equal to 1.

In some embodiments, when the ranking model is a click-through rate prediction model, the server may obtain, through the click-through rate prediction model, the click-through rates of the users with respect to the content items, rank, according to the order of the click-through rates from large to small, the content items corresponding to the content item vectors, and output the content items whose click-through rates are located in the number of the previous targets.

In some embodiments, when the ranking model is a conversion rate prediction model, the server may obtain, through the conversion rate prediction model, a conversion rate of the user with respect to each content item, rank, according to a descending order of the conversion rate, the content items corresponding to each content item vector, and output the content items of which the conversion rates are located at the previous target number.

In some embodiments, the ranking model may not be the click-through rate prediction model or the conversion rate prediction model trained in the above embodiments, that is, the click-through rate prediction model or the conversion rate prediction model trained in the previous embodiment is only used to obtain each user vector and each content item vector, and is not put into ranking prediction in practical application.

In this case, the ranking model is not limited to the click rate prediction model or the conversion rate prediction model in the above two cases, and optionally, the ranking model may also be a machine learning model such as DNN, GBDT (Gradient Boosting Decision Tree), XGBoost (eXtreme Gradient Boosting), and the like.

In the above process, the server may generally select several content items ranked in the top from thousands of content items based on the ranking model, so as to further reduce the screening cost, and further achieve accurate evaluation of the click through rate or the arrival rate, so that the selected content items have a better delivery effect.

In step 406, the server delivers the previous target number of content items to the terminal.

In the above process, after the server determines the previous target number of content items, the server may package content materials of the previous target number of content items as content item delivery responses, and return the content item delivery responses to the terminal, so that when the terminal receives the content item delivery responses, the content item delivery responses are analyzed to obtain content materials corresponding to the content items, and the content materials of the content items are rendered based on a GPU (Graphics Processing Unit), so that the delivered content items may be displayed on a screen of the terminal.

Furthermore, as the machine can explore all possible combination modes in the content materials, on one hand, an advertiser does not need to perform manual combination and delivery experiments (which is equivalent to manual trial and error) on the content materials, and the labor cost in the delivery process of the content items is greatly saved; on the other hand, for the combination mode which has never been tried in some manual combinations, it is possible that some specific users will be more effectively delivered in such a combination mode, and therefore, the upper limit of the delivery effect of the content item is raised. For example, in some short video platforms, the combined depth of the content material explored by the machine can be increased by 1000 times compared to manual combining of the content material.

Further, since the offline delivery experiment in the related art is performed according to the preference of the popular users for content item selection, the preference of some popular users for content items is ignored, so that there is no difference in delivery of content items. In the embodiment of the present disclosure, because the primary screening is directly performed based on the similarity between the user vector and the content item vector, and the secondary screening is performed based on the ranking model, not only content items of different products can be personalized for different users, but also a content material combination that is most likely to be clicked (or converted) by the user can be screened based on the similarity when the content items of the same product are released, so that the releasing effect of thousands of people and thousands of faces is achieved when the content items are released.

For example, taking the content item as an advertisement, when advertising is performed on the same clothing, if a certain user likes a refreshing picture, the server may advertise the clothing with a refreshing cover for the user, and if a certain user likes a rich picture, the server may advertise the clothing with a rich cover for the user.

Fig. 6 is a block diagram illustrating a logical structure of a content item delivery apparatus, according to an example embodiment. Referring to fig. 6, the apparatus includes a first obtaining unit 601, a second obtaining unit 602, a sorting unit 603, and a delivering unit 604.

A first obtaining unit 601 configured to perform obtaining a user vector of a user corresponding to a terminal according to a content item obtaining request of the terminal, wherein the user vector is used for representing personal characteristics of the user;

a second obtaining unit 602 configured to perform obtaining a plurality of content item vectors from a candidate feature library, wherein one content item vector is used for representing content features of one content item, the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold, the candidate feature library comprises content item vectors of a plurality of candidate content items, and different candidate content items are formed by different content materials in different combinations;

a sorting unit 603 configured to perform sorting of content items corresponding to the respective content item vectors based on the user vector, the plurality of content item vectors and a sorting model, and output a content item sorted in a top target number;

a delivery unit 604 configured to perform the delivery of the previous target number of content items to the terminal.

The device provided by the embodiment of the disclosure acquires a user vector of a user corresponding to a terminal through a content item acquisition request of the terminal, thereby acquiring a plurality of content item vectors from a candidate feature library, wherein the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold value, the candidate feature library comprises content item vectors of a plurality of candidate content items, different candidate content items are formed by different content materials in different combination modes, the traversal combination of a machine can improve the breadth and the depth of the content material combination, based on the user vector, the plurality of content item vectors and a ranking model, content items corresponding to respective second feature vectors are ranked, content items ranked at a front target number are output, the front target number of content items are delivered to the terminal, thereby being capable of targeting terminals of different users, the content items formed by different combination modes are put in a targeted manner, so that the putting effect of the content items is greatly improved.

In a possible implementation, the sorting unit 603 is configured to perform:

In one possible implementation, the second obtaining unit 602 is configured to perform:

In a possible embodiment, based on the apparatus composition of fig. 6, the apparatus further comprises:

and the obtaining unit is configured to obtain the sequencing model and content item vectors corresponding to the content items when the training stopping condition is met, wherein the content items are formed by the sample content materials in various combinations.

In a possible embodiment, based on the apparatus composition of fig. 6, the training unit comprises:

a second weighting subunit, configured to perform input of the plurality of sample content materials into a second deep network, and perform weighting processing on the plurality of sample content materials through the second deep network to obtain a plurality of sample content item vectors, where different sample content item vectors are used to represent content items formed by different combinations of the plurality of different types of sample content materials;

the second weighting subunit is configured to perform:

In one possible embodiment, the apparatus is further configured to perform:

In one possible implementation, the first obtaining unit 601 is configured to perform:

With regard to the apparatus in the above-described embodiment, the specific manner in which the respective units perform operations has been described in detail in the embodiment related to the content item delivery method, and will not be elaborated upon here.

Fig. 7 is a schematic structural diagram of a computer device 700, which may have a relatively large difference due to different configurations or performances, and may include one or more processors (CPUs) 701 and one or more memories 702, where the memory 702 stores at least one instruction, and the at least one instruction is loaded and executed by the processors 701 to implement the content item delivery method provided in the embodiments. Of course, the computer device 700 may also have components such as a wired or wireless network interface, a keyboard, and an input/output interface, so as to perform input/output, and the computer device 700 may also include other components for implementing device functions, which are not described herein again.

In an exemplary embodiment, there is also provided a storage medium comprising instructions, such as a memory comprising instructions, executable by a processor of a computer device to perform the content item delivery methods provided by the various embodiments described above. Alternatively, the storage medium may be a non-transitory computer readable storage medium, which may be, for example, a ROM (Read-Only Memory), a RAM (Random-Access Memory), a CD-ROM (Compact Disc Read-Only Memory), a magnetic tape, a floppy disk, an optical data storage device, and the like.

In an exemplary embodiment, there is also provided a computer program product comprising one or more instructions executable by a processor of a terminal to perform the method of content item delivery described above, the method comprising: acquiring a user vector of a user corresponding to a terminal according to a content item acquisition request of the terminal, wherein the user vector is used for representing personal characteristics of the user; obtaining a plurality of content item vectors from a candidate feature library, wherein one content item vector is used for representing the content feature of one content item, the similarity between the plurality of content item vectors and the user vector is higher than a first target threshold value, the candidate feature library comprises the content item vectors of a plurality of candidate content items, and different candidate content items are formed by different content materials in different combination modes; based on the user vector, the plurality of content item vectors and the sequencing model, sequencing the content items corresponding to the content item vectors, and outputting the content items with the sequencing quantity at the front target; delivering the previous target number of content items to the terminal. Optionally, the instructions may also be executable by a processor of the terminal to perform other steps involved in the exemplary embodiments described above.

Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the disclosure following, in general, the principles of the disclosure and including such departures from the present disclosure as come within known or customary practice within the art to which the disclosure pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the disclosure being indicated by the following claims.

It will be understood that the present disclosure is not limited to the precise arrangements 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 of content item delivery, comprising:

and delivering the content items of the front target number to the terminal.

2. The method of claim 1, wherein ranking the content items corresponding to each content item vector based on the user vector, the plurality of content item vectors, and a ranking model, and outputting a top-ranked target number of content items comprises:

3. The content item delivery method according to claim 1 or 2, wherein the sorting the content items corresponding to the content item vectors, and the outputting the content items sorted in the top target number comprises:

4. The content item delivery method according to claim 1 or 2, wherein the sorting the content items corresponding to the content item vectors, and the outputting the content items sorted in the top target number comprises:

5. The method of claim 1, wherein prior to obtaining the plurality of content item vectors from the candidate feature library, the method further comprises:

6. The method of content item delivery according to claim 5, wherein iteratively training the initial model based on the plurality of sample user information and the plurality of sample content material comprises:

7. The content item delivery method of claim 6, wherein the second deep network comprises a target sub-network and a plurality of deep sub-networks, one deep sub-network corresponding to one type of sample content material;

8. A content item delivery apparatus, comprising:

9. A computer device, comprising:

one or more processors;

wherein the one or more processors are configured to execute the instructions to implement the content item delivery method of any one of claim 1 to claim 7.

10. A storage medium, wherein at least one instruction in the storage medium, when executed by one or more processors of a computer device, enables the computer device to perform the content item delivery method of any one of claims 1 to 7.