Delivery processing method, related device and medium
Through the sparse gate control mechanism, some expert models are activated, which solves the problem of increased computing volume and processing load in delivery processing, and achieves the capacity improvement and accuracy maintenance of the delivery network.
Patent Information
- Application Number
- CN202310591415.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-23
- Publication Date
- 2024-11-26
AI Technical Summary
When the prior art increases the expert network capacity used for delivery processing, the calculation amount and processing load have significantly increased, affecting the capacity expansion and delivery accuracy.
The sparse gating mechanism is adopted to activate only the second number of expert models corresponding to each task, and the output is weighted and summed through the gated nodes, reducing the calculation amount and maintaining the capacity constant, while satisfying the differences between multiple tasks.
With smaller computing and processing load, the capacity of the expert network used for delivery processing is improved, the delivery accuracy is maintained, and the differences between multitasking are not deteriorated.
Smart Images

Figure CN119025741A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of big data, and in particular to a delivery processing method, related devices and media. Background Art
[0002] In the field of big data, it is often necessary to deliver information and video content to objects. How to more accurately find the objects that need these contents for the required tasks (for example, more likes, more comments, etc.) is a problem. At present, the industry widely uses PLE models for delivery. The PLE model assigns a dedicated expert model to each task, and establishes several shared expert models. Different implicit task representations are extracted through different expert models, and high-order outputs of different expert models are weighted and combined through gating nodes to effectively balance the sharing and mutual exclusion between multiple tasks. Although the delivery effect is improved by dedicated expert models, the computational workload and processing load are significantly increased, affecting the expansion of the expert network used for delivery processing. Summary of the invention
[0003] The disclosed embodiments provide a delivery processing method, related devices and media, which can improve the capacity of the expert network used in delivery processing with a small amount of calculation and processing load.
[0004] According to one aspect of the present disclosure, a delivery processing method is provided, including:
[0005] Quantizing a plurality of delivery basic features to obtain a plurality of first feature vectors, wherein the delivery basic features are used to predict task completion probabilities of a target object with respect to a plurality of tasks for the delivered content;
[0006] Cascading a plurality of the first feature vectors into a first cascade feature vector;
[0007] Inputting the first cascade feature vector into a first number of expert models;
[0008] Selecting a second number of the expert models from the first number of the expert models through the gating nodes corresponding to each of the tasks, and performing weighted summation on outputs of the selected second number of the expert models to obtain a second feature vector corresponding to the task, wherein the second number is smaller than the first number;
[0009] The second feature vectors corresponding to the multiple tasks are input into the prediction models corresponding to the multiple tasks to obtain the predicted task completion probabilities of the target object for the to-be-delivered content with respect to the multiple tasks.
[0010] According to one aspect of the present disclosure, there is provided a delivery processing device, comprising:
[0011] A vectorization unit, used for vectorizing a plurality of delivery basic features to obtain a plurality of first feature vectors, wherein the delivery basic features are used to predict task completion probabilities of a target object with respect to a plurality of tasks for the delivered content;
[0012] A cascading unit, used for cascading a plurality of the first feature vectors into a first cascade feature vector;
[0013] A first input unit, used for inputting the first cascade feature vector into a first number of expert models;
[0014] a first selection unit, configured to select a second number of the expert models from the first number of the expert models through a gating node corresponding to each of the tasks, and perform a weighted sum on outputs of the selected second number of the expert models to obtain a second feature vector corresponding to the task, wherein the second number is smaller than the first number;
[0015] The second input unit is used to input the second feature vectors corresponding to the multiple tasks into the prediction model corresponding to the multiple tasks to obtain the predicted task completion probability of the target object for the content to be delivered with respect to the multiple tasks.
[0016] Optionally, the first selection unit is specifically configured to:
[0017] Generate a weight vector based on the first cascade feature vector and the first matrix, wherein the number of columns of the first matrix is the first number, the number of rows is the dimension of the first cascade feature vector, and the weight vector indicates the weights of the first number of the expert models;
[0018] A second number of the expert models are selected from the first number of the expert models based on the weight vector.
[0019] Optionally, the first selection unit is specifically configured to:
[0020] Multiplying the first matrix by the first concatenated feature vector to obtain an original weight vector;
[0021] Multiplying a second matrix by the first concatenated feature vector to obtain a weight adjustment basis vector, wherein the number of columns of the second matrix is the first number, and the number of rows is the dimension of the first concatenated feature vector;
[0022] Generate a weight adjustment value vector based on the random disturbance vector and the weight adjustment basis vector;
[0023] The weight vector is generated based on the original weight vector and the weight adjustment value vector.
[0024] Optionally, the first selection unit is specifically configured to:
[0025] Obtaining a random number set, wherein the random numbers in the random number set conform to a normal distribution;
[0026] Randomly extracting a first number of the random numbers from the random number set to form the random perturbation vector;
[0027] The random disturbance vector is point-multiplied with the weight adjustment basis vector to obtain the weight adjustment value vector.
[0028] Optionally, the first selection unit is specifically configured to:
[0029] Determine the first second number of elements in the weight vector whose values are sorted from largest to smallest;
[0030] The expert models corresponding to the first second number of elements are used as the selected second number of expert models.
[0031] Optionally, the first selection unit is specifically configured to:
[0032] Modify the other elements in the weight vector that are not the first second number of elements in the descending order to negative infinity;
[0033] Performing normalized exponential processing on the weight vector to obtain first weights of a second number of the expert models;
[0034] The outputs of the second number of the expert models selected are weighted summed according to the first weights.
[0035] Optionally, the expert model includes a kernel parameter matrix and a bias vector, the number of columns of the kernel parameter matrix is the dimension of the output of the expert model, the number of rows is the dimension of the first cascaded feature vector, and the dimension of the bias vector is equal to the dimension of the output of the expert model;
[0036] The first selection unit is specifically used for:
[0037] Obtaining a three-dimensional kernel parameter tensor, wherein the three-dimensional kernel parameter tensor represents the kernel parameter matrices of a first number of the expert models;
[0038] Obtaining a bias vector matrix, wherein the bias vector matrix represents the bias vectors of a first number of the expert models;
[0039] Obtaining, from the three-dimensional kernel parameter tensor, the kernel parameter matrix corresponding to a second number of the first weights;
[0040] Obtaining, from the bias vector matrix, the bias vectors corresponding to a second number of the first weights;
[0041] The weighted sum is determined based on a second number of the first weights, the kernel parameter matrices corresponding to the second number of the first weights, and the bias vectors corresponding to the second number of the first weights.
[0042] Optionally, the first selection unit is specifically configured to:
[0043] Multiplying the kernel parameter matrix corresponding to each of the expert models in the second number of the expert models by the first cascade feature vector input into the expert model to obtain an intermediate vector;
[0044] Adding the intermediate vector to the bias vector corresponding to the expert model to obtain a result vector;
[0045] Applying an activation function to the result vector to obtain an output of the expert model;
[0046] The weighted sum is obtained based on outputs of a second number of the expert models and the first weights corresponding to the expert models.
[0047] Optionally, the second number is determined by:
[0048] Get the task type of the task;
[0049] Obtaining prediction accuracy requirements for the task;
[0050] The second number is determined based on the task type and the prediction accuracy requirement.
[0051] Optionally, the determining the second number based on the task type and the prediction accuracy requirement includes:
[0052] Based on the task type, determining a first score;
[0053] determining a second score based on the prediction accuracy requirement;
[0054] determining a total score based on the first score and the second score;
[0055] Based on the total score, the second number is determined.
[0056] Optionally, the first selection unit is specifically configured to:
[0057] Get the task type of the task;
[0058] From the first number of the expert models, a third number of the expert models matching the task type;
[0059] A second number of the expert models is selected from the third number of the expert models, wherein the first number is greater than the third number, and the third number is greater than the second number.
[0060] Optionally, the vectorization unit is specifically used for:
[0061] Acquire object features of the target object;
[0062] Acquire scene features of the current scene where the target object is located;
[0063] Filtering out a filtered content set from the to-be-filtered content set, and taking the features of each filtered content in the filtered content set as features of the to-be-delivered content in turn;
[0064] Determining the object feature, the scene feature, and the feature of the content to be delivered as a plurality of delivery basic features;
[0065] The determined multiple delivery basic feature vectors are quantized to obtain multiple first feature vectors.
[0066] Optionally, the vectorization unit is specifically used for:
[0067] If the delivery basic feature is a character feature, input the character feature into an embedding layer to obtain the first feature vector corresponding to the character feature;
[0068] If the delivery basic feature is a numerical feature, numerical mapping is performed on the numerical feature to obtain the first feature vector corresponding to the numerical feature;
[0069] If the delivery basic feature is a variable-length sequence feature, the variable-length sequence feature is pooled to obtain the first feature vector corresponding to the variable-length sequence feature.
[0070] Optionally, the delivery processing device further includes:
[0071] A first acquisition unit, used to acquire the focus task of the target object;
[0072] A second acquisition unit, configured to acquire a task completion probability for the task of interest from the task completion probabilities for the plurality of tasks;
[0073] A sorting unit, configured to sort the to-be-delivered content based on the task completion probability of the concerned task;
[0074] A delivery unit is used to deliver the content to be delivered to the target object according to the ranking.
[0075] Optionally, the plurality of delivery basic features are selected from a training sample in a training sample set, the training sample having a plurality of training labels for the plurality of tasks, the training labels indicating whether the task is completed for the training sample;
[0076] The first number of the expert models, the gating nodes, and the prediction model are jointly trained by the following process:
[0077] For a single task, determining an error subfunction of the task based on the training labels and the task completion probability of the task;
[0078] Based on the error sub-function, determining the error function of the training sample;
[0079] Based on the error function, a first number of the expert models, the gating nodes, and the prediction model are jointly trained.
[0080] Optionally, determining the error function of the training sample based on the error sub-function includes:
[0081] Obtaining task weights of a plurality of the tasks;
[0082] Based on the task weights, a weighted sum is performed on the error sub-functions of a plurality of the tasks to obtain the error function.
[0083] According to one aspect of the present disclosure, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the above-mentioned delivery processing method when executing the computer program.
[0084] According to one aspect of the present disclosure, a computer-readable storage medium is provided, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the delivery processing method as described above is implemented.
[0085] According to one aspect of the present disclosure, a computer program product is provided. The computer program product includes a computer program. The computer program is read and executed by a processor of a computer device, so that the computer device executes the delivery processing method as described above.
[0086] In the disclosed embodiment, no distinction is made between shared expert models and task-specific expert models, but a sparse gating mechanism is adopted, that is, for each task, only the second number of expert models are activated among the first number of expert models. In this way, no matter how large the capacity of the expert network is expanded with the addition of tasks, even when it is expanded to include a large number of expert models, only the second number of expert models are activated, and the overall computational load remains basically constant, thereby being able to increase the capacity of the expert network used for delivery processing under the condition of a smaller computational load and processing load. At the same time, the second number of expert models are selected from the first number of expert models for activation, and the expert models activated between different tasks may overlap, partially overlap, or not overlap at all, so that the differences between multiple tasks can be met and the delivery accuracy will not be deteriorated.
[0087] Other features and advantages of the present disclosure will be described in the following description, and partly become apparent from the description, or understood by practicing the present disclosure. The purpose and other advantages of the present disclosure can be realized and obtained by the structures particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0088] The accompanying drawings are used to provide further understanding of the technical solution of the present disclosure and constitute a part of the specification. Together with the embodiments of the present disclosure, they are used to explain the technical solution of the present disclosure and do not constitute a limitation on the technical solution of the present disclosure.
[0089] Figure 1 is a system architecture diagram of the delivery processing method of the embodiment of the present disclosure;
[0090] Figure 2A -B is a schematic diagram of an interface of an embodiment of the present disclosure applied to a video delivery scenario;
[0091] Figure 3 is a flow chart of a delivery processing method according to an embodiment of the present disclosure;
[0092] Figure 4 It is a general implementation structure diagram of the delivery processing method of the embodiment of the present disclosure;
[0093] Figure 5 yes Figure 3 A flow chart of generating a first feature vector based on the acquired delivery basic features in step 310;
[0094] Figure 6 yes Figure 5 Get a schematic diagram of the basic features of the delivery;
[0095] Figure 7 yes Figure 5 A flowchart of obtaining the filtered content set in step 530;
[0096] Figure 8 yes Figure 7 A flow chart of obtaining the filtered content set in step 730;
[0097] Fig. 9A -C is a schematic diagram of generating the first feature vector for different types of delivery basic features;
[0098] Fig.10 yes Figure 3 Schematic diagram of generating the first cascade feature vector in ;
[0099] Fig.11 yes Figure 3 A flowchart of selecting a second number of expert models from the first number of expert models in step 340;
[0100] Fig.12 yes Fig.11 A flow chart of generating a weight vector in step 1110;
[0101] Fig.13 yes Fig.12 A flow chart of generating a weight adjustment value vector in step 1230;
[0102] Fig.14 yes Fig.11 A flowchart of selecting a second number of expert models in step 1120;
[0103] Fig.15 yes Figure 3 Another flow chart of selecting a second number of expert models from the first number of expert models in step 340;
[0104] Fig.16 is a flow chart of generating the second eigenvector in FIG340;
[0105] Fig.17 yes Fig.16 A flowchart of determining a weighted sum of a second number of expert models in step 1630;
[0106] Fig.18A It is a schematic diagram of extracting a kernel parameter matrix from a three-dimensional kernel parameter tensor;
[0107] Fig.18B is a schematic diagram of extracting a bias vector from the bias vector matrix;
[0108] Fig.19 yes Fig.17 A flowchart of determining a weighted sum of outputs of a second number of expert models in step 1750;
[0109] Fig. 20is a flow chart for determining a second number;
[0110] Fig.21 yes Fig. 20 A specific flow chart of determining the second number in step 2030;
[0111] Fig. 22 is Figure 3 Then add a flow chart of the process of determining the order of delivery of the content to be delivered;
[0112] Fig.23 yes Fig. 22 A flow chart of sorting the content to be delivered in step 2230;
[0113] Fig.24 is a flow chart for jointly training a first number of expert models, gating nodes, and prediction models;
[0114] Fig.25 yes Fig.24 A flow chart of determining the error function in step 2420;
[0115] Fig.26 It is an implementation detail diagram of the delivery processing method of the embodiment of the present disclosure.
[0116] Fig. 27 is a module diagram of a delivery processing device according to an embodiment of the present disclosure;
[0117] Fig.28 According to the embodiment of the present disclosure Figure 3 The terminal structure diagram of the delivery processing method shown;
[0118] Fig.29 According to the embodiment of the present disclosure Figure 3 The server structure diagram of the delivery processing method shown. DETAILED DESCRIPTION
[0119] In order to make the purpose, technical solution and advantages of the present disclosure more clear, the present disclosure is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure and are not used to limit the present disclosure.
[0120] Before further describing the embodiments of the present disclosure in detail, the nouns and terms involved in the embodiments of the present disclosure are described. The nouns and terms involved in the embodiments of the present disclosure are subject to the following interpretations:
[0121] Deep neural network: It is a multi-layer unsupervised neural network. Deep neural network uses the output features of the previous layer as the input of the next layer for feature learning. After layer-by-layer feature mapping, the features of the existing spatial samples are mapped to another feature space, so as to learn to have better feature expression for the existing input. Deep neural network has multiple nonlinear mapping feature transformations and can fit highly complex functions.
[0122] Activation function: is a function added to a neural network to help the network learn complex patterns in the data. Similar to the neuron-based model in the human brain, the activation function ultimately determines what to send to the next neuron. Without the activation function, each layer of the neural network is just a linear transformation, and the superposition of multiple layers of input is still a linear transformation. The linear change has poor expression ability and may not be able to express what the neural network wants to output. The activation function can introduce nonlinear factors to clearly express the output content.
[0123] Gradient descent algorithm: A method for minimizing the loss function. It calculates the gradient of the loss function at the current position and updates the parameters in the direction of gradient descent to find the optimal solution that minimizes the value of the loss function.
[0124] System architecture and scenario description of the application of the embodiments of the present disclosure
[0125] Figure 1 1 is a system architecture diagram used in the delivery processing method according to an embodiment of the present disclosure, and includes a server 110, the Internet 120, and an object terminal 130, etc.
[0126] The server 110 refers to a computer system that can provide content delivery services to the target terminal 130. Compared with the target terminal 130, the server 110 has higher requirements in terms of stability, security, performance, etc. The server 110 can be a high-performance computer in a network platform, a cluster of multiple high-performance computers, a part of a high-performance computer (such as a virtual machine), a combination of parts of multiple high-performance computers (such as virtual machines), etc. The server 110 can also communicate with the Internet 120 in a wired or wireless manner to exchange data.
[0127] The object terminal 130 is a device used to deliver content so that the object can view the delivered content. It includes various forms such as desktop computers, laptops, PDAs (personal digital assistants), mobile phones, car terminals, home theater terminals, and dedicated terminals. In addition, it can be a single device or a collection of multiple devices. For example, multiple devices are connected through a local area network and work together using a common display device to form a terminal. The terminal can also communicate with the Internet 120 in a wired or wireless manner to exchange data.
[0128] The embodiments of the present disclosure can be applied in various scenarios, such as Figure 2A -B shows a scenario of viewing the delivered video content in an instant short message application, etc.
[0129] like Figure 2A The interface of the instant short message application in the object terminal 130 is shown in FIG. The instant short message application can be used to communicate instantly with the contacts stored in the application. After the "Short Message" option in the function bar below is triggered, the short message details with the contacts are displayed. The instant short message application can also be used to watch videos. After the object triggers the "Video" option in the function bar, the interface of the instant short message application is as follows: Figure 2B shown.
[0130] exist Figure 2B In the example, the instant short message application opens the video interface and plays the delivered video content. The interface also displays the number of likes, favorites, and comments on the current video content before the current time. The subject can also like, favorite, and comment on the current video content in the interface. If the subject wants to watch other videos, the next video can be switched through specific operations (such as sliding the screen up). The video content delivered to the instant short message application for the subject to view can be determined by the server 110, that is, the server 110 predicts the video content that the subject may be interested in and delivers the video content to the instant short message application. Therefore, it is very important to be able to quickly and correctly judge the subject's preferences and accurately deliver video content to the subject according to the preferences. The disclosed embodiment can achieve the effect of improving computing efficiency while ensuring the accuracy of the delivered video content.
[0131] General description of the disclosed embodiments
[0132] According to an embodiment of the present disclosure, a delivery processing method is provided.
[0133] The delivery processing method is a process of determining the content to be delivered that the target object is interested in, and delivering the content to be delivered to the object terminal 130 of the target object. The target object is the object that wants to view the content to be delivered. The embodiments of the present disclosure can be applied to different types of content to be delivered, such as pictures, videos, articles, applications, etc. If the target object opens the video recommendation software, the delivery processing method of the embodiment of the present disclosure is used to predict the video content that the target object may be interested in, and deliver the video content to the object terminal 130 of the target object through the video recommendation software. When determining the delivery content, it is necessary to judge whether the content to be delivered is the content that the target object is interested in based on the completion of the task of the target object on the delivered content. Task refers to the operation performed by the target object on the delivered content, such as whether to like the content to be delivered, whether to collect the content to be delivered, etc. The delivery order of the content to be delivered is determined based on the completion of the target object on the content to be delivered.
[0134] The delivery processing method of the embodiment of the present disclosure is executed on the server 110. After the execution is completed, the obtained delivery content is transmitted to the object terminal 130 through the Internet 120, and the object terminal 130 displays the delivery content to the target object.
[0135] like Figure 3 As shown, according to one embodiment of the present disclosure, the delivery processing method includes:
[0136] Step 310: vectorize multiple delivery basic features to obtain multiple first feature vectors;
[0137] Step 320: cascade the plurality of first feature vectors into a first cascade feature vector;
[0138] Step 330: input the first cascade feature vector into a first number of expert models;
[0139] Step 340: select a second number of expert models from the first number of expert models through the gating nodes corresponding to each task, and perform weighted sum on the outputs of the selected second number of expert models to obtain a second feature vector corresponding to the task;
[0140] Step 350: Input the second feature vectors corresponding to the multiple tasks into the prediction models corresponding to the multiple tasks to obtain the predicted task completion probabilities of the target object for the content to be delivered with respect to the multiple tasks.
[0141] The above steps 310 - 350 are described in general below.
[0142] Figure 4 What is shown is the overall structural diagram of the delivery processing method of the embodiment of the present disclosure.
[0143] Figure 4The bottom layer in is the delivery basic features. There are multiple delivery basic features, which are used to predict the target object's attitude towards the delivery content and the task completion probability of multiple tasks.
[0144] Step 310 is executed on the delivery basic feature to perform feature vectorization to obtain a first feature vector (not shown in the figure). Each delivery basic feature corresponds to a first feature vector.
[0145] Step 320 is executed for the first feature vectors corresponding to all the delivery basic features to perform feature cascading to form a first cascade feature vector (not shown in the figure).
[0146] In step 330, the first cascade feature vector is input into a first number of expert models. The first number of expert models are used to analyze and calculate each dimension of the first cascade feature vector so as to balance the sharing and mutual exclusion between multiple tasks.
[0147] In step 340, each task corresponds to a gating node, and the gating node is used to integrate the output of the expert model into a high-order vector for the task. For each task, a second number of expert models is activated from the first number of expert models. It should be noted that the second number of expert models activated by different tasks is the same, but the expert models can be different. For example, the second number is 2, and the two expert models activated by task A are expert model B and expert model C, respectively, and the two expert models activated by task B are expert model C and expert model D, respectively.
[0148] After determining the second number of expert models activated for each task, a weighted sum is calculated on the outputs of the second number of expert models through the gating nodes corresponding to the task to obtain a second feature vector corresponding to the task (not shown in the figure).
[0149] In step 350, the second feature vector output by the gating node corresponding to the task is input into the prediction model corresponding to the task. The prediction model is used to output the task completion probability of the target object for the corresponding task of the content to be delivered according to the second feature vector (not shown in the figure). Based on the task completion probability, the delivery order of the content to be delivered to the target object can be determined.
[0150] The above steps 310 - 350 are described in detail below.
[0151] Detailed description of step 310
[0152] In step 310, a plurality of delivery basic features are vectorized, and a corresponding first feature vector is obtained according to each delivery basic feature.
[0153] Before quantizing multiple delivery basic features, it is necessary to obtain delivery basic features. Therefore, in one embodiment, Figure 5 As shown, multiple delivery basic feature vectors are vectorized to obtain multiple first feature vectors, including:
[0154] Step 510: Obtain object features of the target object;
[0155] Step 520: Obtain scene features of the current scene where the target object is located;
[0156] Step 530: Filter out a filtered content set from the to-be-filtered content set, and use the features of each filtered content in the filtered content set in turn as features of the to-be-delivered content;
[0157] Step 540: determining the object features, the scene features, and the features of the content to be delivered as a plurality of delivery basic features;
[0158] Step 550: quantize the determined multiple delivery basic feature vectors to obtain multiple first feature vectors.
[0159] In step 510, the object feature refers to a feature related to the target object. The object feature can be a feature of the target object itself, such as the industry, education level, etc. of the target object. For the same feature, if the feature value is different, the target object may have different preferences for the delivered content. For example, the target objects in different industries may prefer different delivery content. Taking video content as an example, the target in the news industry may pay more attention to videos related to hot events, and the target in the catering industry may like videos related to health and wellness.
[0160] Note that when obtaining the target object's object characteristics, the target object's consent must be obtained in advance. In addition, the collection, use and processing of these object characteristics will comply with relevant laws, regulations and standards. When obtaining the target object's consent, the target object's separate permission or consent can be obtained through pop-ups or by jumping to a confirmation page.
[0161] The object characteristics may also be the object characteristics of the object group (such as a family or work group) to which the target object belongs. This is because, in some cases, the object's preference for delivered content is greatly affected by the object group to which it belongs. For example, the object itself likes to watch music-related video content, but because other members of the object's family group like to watch game-related video content, the object may also be interested in game-related video content. Therefore, in one embodiment, obtaining the object characteristics of the target object also includes obtaining the object characteristics of the object group to which the target object belongs. Considering the object characteristics of the object group to which the target object belongs can improve the accuracy of content delivery. Similar to the above, when obtaining the object characteristics of the object group (such as a family) to which the target object belongs, the consent of the object group members must also be sought in advance, so it will not be repeated.
[0162] In one embodiment, the object characteristics of the target object can be obtained through registration information. The registration information here can be the registration information of the target object on the object terminal 130, or the registration information in the software for delivering the delivery content, etc. When the target object uses the object terminal 130 for the first time, the object terminal 130 may require the target object to register, and registration requires filling in information such as industry as registration information. In this way, the corresponding object characteristics can be obtained from the object terminal 130 used by the target object. When the target object uses the software for delivering the content to be delivered for the first time, it is also necessary to register, and the object may be required to fill in information such as industry as registration information. Therefore, the object characteristics can also be obtained through the registration information in the software. The advantages of obtaining object characteristics through registration information are: convenient and fast, and high acquisition efficiency.
[0163] When obtaining the registration information of the target object, as mentioned above, the consent of the target object must be obtained in advance. To save space, it is not repeated here.
[0164] In another embodiment, the object characteristics of the target object can be obtained through the historical viewing log of the target object within a predetermined time period before the current time point. The historical viewing log refers to the relevant records of the target object viewing the delivery content within the predetermined time period before the current time point. The predetermined time period can be one hour, one day, one week, one month, etc. The predetermined time period can be determined according to the frequency of the target object viewing the delivery content. If the target object views the delivery content frequently, the current object characteristics of the target object can be obtained through the viewing records within a shorter time period; if the target object views the delivery content less frequently, it is necessary to use the viewing records within a longer time period to ensure that there is enough data to be used to obtain the object characteristics. The predetermined time period can also be determined according to the type of delivery content. For example, if the delivery content is video content, a longer time period can be used as the predetermined time period to obtain the long-term preference of the target object; if the delivery content is news content, since news is time-sensitive, a shorter time period needs to be used as the predetermined time period.
[0165] When obtaining the historical viewing log of the target object, as mentioned above, the consent of the target object must be obtained in advance. To save space, it is not repeated here.
[0166] The content recorded in the viewing log may include: viewing content, viewing time, content type, whether it is watched completely, whether it is liked, whether it is commented, etc. For example, Table 1 shows the historical viewing log of the target object within one day before the current time point.
[0167]
[0168] Table 1
[0169] According to the historical viewing logs of the target object within one day before the current time point shown in Table 1, the object features of the target object can be obtained through the historical viewing logs.
[0170] In one embodiment, obtaining the object features of the target object through the historical viewing log includes: inputting the historical viewing log into the object feature extraction model, and generating the object features by the object feature extraction model. The object feature extraction model can predict the object features of the target object based on the data in the historical viewing log. For example, according to the historical viewing log in Table 1, the object feature extraction model predicts the object features of the target object including: hobbies of entertainment, reading, pets, etc. The advantage of this embodiment is that the object features of the target object can be obtained more accurately and comprehensively through the object feature extraction model, thereby improving the accuracy of obtaining the object features of the target object.
[0171] The advantage of the embodiment of obtaining object characteristics through the historical viewing log of the target object within a predetermined time period before the current time point is that the target object's interest level in the content that has been watched can be accurately obtained through the historical viewing log, and the target object's preference characteristics can be predicted based on the interest level, thereby improving the accuracy of obtaining the target object's object characteristics.
[0172] In step 520, the scene feature refers to the feature of the current scene in which the target object is viewing the delivered content. The same target object may also want to view different delivered content in different scenes. For example, in a work scene, the target object may want to view work-related delivered content more. If the delivered content is of interest to the target object, the target object may perform a collection task on the delivered content so that it can be viewed at any time when needed; when at home, the target object may want to view the delivered content related to his or her own preferences more. The target object is more likely to perform a task of liking the delivered content of interest. Therefore, the scene feature can be used to determine the task completion probability of the target object performing different tasks on the delivered content. The scene features include: the characteristics of the geographical area where the target object is currently located, the characteristics of the current time period, the characteristics of the current input day type (weekday, weekend, or holiday), etc.
[0173] When obtaining the scene features of the target object, the target object's consent must be obtained in advance, which is similar to the above and will not be repeated here.
[0174] The characteristics of the current geographical area refer to the characteristics of the geographical area where the target object is currently located. This geographical area can be an administrative geographical area, such as City A, City B, etc., or it can be a geographical area divided by entities on the map, such as the XX shopping mall area, the XX hospital area, etc. For administrative geographical areas, the delivery content that the target object is interested in has an impact. For example, if the target object is in City A, the delivery content can be tourist attractions or food culture related to City A. For geographical areas divided by entities, the delivery content that the target object is interested in also has an impact. For example, if the target object is in the XX shopping mall area, the delivery content can be activity information or special projects in the mall; if the target object is in the XX hospital area, the delivery content can be content related to health and wellness.
[0175] The current geographical area feature can be obtained from the positioning information of the object terminal 130 of the target object. The object terminal 130 of the target object has a positioning device such as GPS and Beidou, and the positioning device can continuously obtain the positioning information of the object terminal 130. The geographical area where the target object is currently located can be determined based on the positioning information as the current input geographical area feature. For example, according to the GPS positioning information obtained by the positioning device, it is mapped onto an electronic map and it is found that it belongs to China, and the current input geographical area feature "China" is obtained.
[0176] The current time period feature refers to the time period within a day when the target object views the delivered content, for example, 9:00-11:00, 11:00-13:00, 13:00-18:00, 18:00-22:00, etc. Since the target object has different activities in different time periods of the day, the purpose of viewing the recommended content may also be different. Therefore, the delivered content will be affected by the current time period feature. For example, from 9:00 to 11:00, the target object may be working, so the target object is likely to want to view work-related content when viewing the delivered content; from 18:00 to 22:00, the target object is often resting, and at this time, it is likely that the target object wants to view content related to his or her interests.
[0177] The current input day type feature refers to the feature of the type of day when the target object views the delivery content. It is divided into weekdays, weekends, and holidays. The type of delivery content that the target object wants to view may be different on different types of days. For example, on weekdays (Monday to Friday), the target object is often at work, and the delivery content that the target object wants to view may be work-related; on weekends (Saturday and Sunday), when the target object does not need to work, the delivery content that the target object wants to view may be related to his or her hobbies; on holidays, the target object may want to view delivery content related to holiday culture.
[0178] The current time period feature and the current input day type feature can be obtained from the system time of the object terminal 130 of the target object. The system time generally includes information of year, month, day, day of the week, hour, minute, and second. The current input time period feature can be determined from the information of hour, minute, and second. For example, if the current time is 13:52:26, it is determined that it belongs to 13:00-18:00, that is, the current input time period feature is 13:00-18:00. If the current date is Tuesday, May 9, 2023, it is determined that the current input day type feature is a working day.
[0179] In step 530, a filtered content set is selected from the content set to be filtered, and the features of each filtered content in the filtered content set are used as the features of the content to be delivered in turn. The content set to be filtered refers to a content set that has not been filtered and contains various types of content. The content set to be filtered contains a lot of content that the target object is not interested in. Therefore, the content in the content set to be filtered needs to be filtered to form a filtered content set. The filtered content set contains multiple content to be delivered that the target object may be interested in.
[0180] If you want to predict the target object's task completion probability for multiple tasks of the content to be delivered, you need to take out each content to be delivered in the filtered content set in turn and obtain the features of the content to be delivered. The features of the content to be delivered are used to represent the characteristics of multiple aspects of the content to be delivered, such as content type, content length, content release time, content keywords, etc.
[0181] In one embodiment, Figure 7 As shown, the filtered content set is filtered out from the to-be-filtered content set, including:
[0182] Step 710: Obtain a first tag of the target object;
[0183] Step 720: Obtain a second tag of the content to be filtered in the set of content to be filtered;
[0184] Step 730: Filter out a filtered content set from the to-be-filtered content set based on the matching degree between the first tag and the second tag.
[0185] In step 710, the first tag indicates the type of content that the target object is interested in.
[0186] In one embodiment, the first tag of the target object may be obtained through the historical viewing log. The historical viewing log has been described in detail in the above embodiment and will not be described again here.
[0187] Note that when obtaining the first tag of the target object, the target object's consent must be obtained in advance. In addition, the collection, use and processing of the first tag will comply with relevant laws, regulations and standards. When obtaining the target object's consent, the target object's separate permission or separate consent can be obtained through a pop-up window or by jumping to a confirmation page.
[0188] Obtaining the first label of the target object through the historical viewing log includes: determining the target object's preference type for the delivered content through task statistics of the watched content of each content type in the historical viewing log; and using the preference as the first label of the target object. Since the historical viewing log records the completion of the target object's tasks for each watched content (for example, whether to like, whether to collect, whether to comment, etc.), it is possible to infer whether the target object is interested in the watched content through the historical viewing log. For example, in the historical viewing log shown in Table 1, the target object likes and collects the delivered content A, so it can be inferred that the target object is interested in the delivered content A. However, the target object did not perform any operation on the delivered content B, so it can be inferred that the target object is not interested in the delivered content B. For example, Table 2 shows the task statistics for the watched content of each content type in the historical viewing log shown in Table 1.
[0189]
[0190]
[0191] Table 2
[0192] As shown in Table 2, the target object liked 25 times, collected 14 times, and commented on 12 times for the content of type "Entertainment".
[0193] Note that when obtaining the viewing content of each content type in the historical viewing log, the consent of the target object must be obtained.
[0194] In one embodiment, by counting the task completion of each content type in the historical viewing log, the first label of the target object is determined, including: calculating the task completion average of the content type, ranking the task completion average of all content types, and determining the first label of the target object. For example, according to the task completion statistics of each content type in Table 2, the task completion average of entertainment type content is (25+14+12) / 3=17; the task completion average of game type content is (5+0+2) / 3=2.33. After determining the task completion average of each content type, the content types are ranked in order from high to low according to the task completion average, and the first predetermined number of content types in the ranking are selected as the first label of the target object. For example, after calculating the task completion average of each content type in Table 2, the content types are ranked according to the task completion average, and the top 3 content types in the ranking are selected as the first label of the target object. The first label of the target object obtained by ranking the task completion average of each content type in Table 2 is entertainment, reading, and pets. The advantage of using the average result to determine the target object's preference type for the delivered content is that it fairly reflects the impact of each task on the target object's preference and improves the fairness of determining the first label.
[0195] In another embodiment, the first label of the target object is determined by counting the tasks of the viewing content of each content type in the historical viewing log, including: setting a weight for each task completion number; and determining the first label of the target object based on the weighted average ranking of the task completion of all content types. The reason for setting weights for different task completion numbers is that different task completion situations indicate that the target object has different levels of interest in the delivery content. For example, the target object may be more interested in the delivery content that is liked than in the delivery content that is watched in full; and the target object may be more interested in the delivery content that is commented on than in the delivery content that is liked. Therefore, the target object's interest in different types of delivery content can be represented by setting weights. For example, the weight of "Like" is 0.3; the weight of "Number of Favorites" is 0.5; and the weight of "Number of Comments" is 0.2. The weighted average of task completion of content types is calculated based on the weights and the number of task completions. For example, in Table 2, for entertainment-type content, the weighted average value of task completion is (25*0.3+14*0.5+12*0.2) / 1=16.9; for game-type content, the weighted average value of task completion is (5*0.3+0*0.5+2*0.2) / 1=1.9. After determining the weighted average value of task completion for each content type, the content types are ranked in descending order according to the weighted average value of task completion, and the first predetermined number of content types in the ranking are selected as the first label of the target object. For example, the top three content types in Table 2 are selected as the first label of the target object, so the first label of the target object for the delivered content is entertainment, reading, and pets. The advantage of using the weighted average result to determine the target object's preference type for the delivered content is that different weights can be set for different task completion times according to actual conditions, which improves the flexibility of determining the first label.
[0196] The advantage of determining the target object's preference by statistically analyzing the viewing tasks of each content type in the historical viewing log, and thus determining the first label of the target object, is that the target object's preference for the delivered content can be quickly obtained, thereby improving the efficiency of obtaining the target object's first label through the historical viewing log.
[0197] In step 720, the second tag indicates the type of content to be screened.
[0198] In one embodiment, obtaining the second label of the content to be screened includes: inputting the content to be screened into a label prediction model; and the label prediction model outputs the second label of the content to be screened. For example, inputting a content to be screened into the label prediction model, the second label of the content to be screened is obtained as: entertainment, movie.
[0199] When obtaining the second tag of the content to be screened, the consent of the publisher of the content to be screened must be obtained.
[0200] The advantage of outputting the second label through the label prediction model is that the accuracy of obtaining the second label is improved.
[0201] In step 730, the matching degree between the first tag and the second tag represents the matching degree between the to-be-screened content and the target object, and the to-be-screened content whose matching degree meets the predetermined condition is added to the screened content set.
[0202] In one embodiment, the degree of match between the first label and the second label is the proportion of the overlapping labels of the second label and the first label in all the first labels. For example, the first labels of the target object include: "student", "music", "popular", and "funny", and the second labels of the content A to be screened include: "music", "popular", and "folk songs". The overlapping labels of the first label and the second label are "music" and "popular", and the overlapping labels account for 50% of all the first labels. Therefore, the degree of match between the content A to be screened and the target object is 50%. The overlapping label of the second label of the content B to be screened and the first label is "student", and the overlapping labels account for 25% of all the first labels. Therefore, the degree of match between the content B to be screened and the target object is 25%. Similarly, the degree of match between the content C to be screened and the target object is 50%; the degree of match between the content D to be screened and the target object is 75%; and the degree of match between the content E to be screened and the target object is 0. The advantage of determining the degree of match by the proportion of overlapping labels in all the first labels is that the degree of match between the content to be screened and the target object can be quickly determined by overlapping labels, thereby improving the efficiency of obtaining the filtered content set.
[0203] Add the to-be-screened content whose matching degree meets the predetermined condition to the filtered content set. In one embodiment, the predetermined condition refers to adding the to-be-screened content whose matching degree with the target object is greater than or equal to the preset matching degree threshold to the filtered content set. For example, the preset matching degree threshold is 50%, that is, the to-be-screened content whose matching degree is greater than or equal to 50% is added to the filtered content set. In this embodiment, the advantage of determining the filtered content set based on the matching degree threshold is that it ensures that the to-be-delivered content in the filtered content set is content with a high matching degree with the target object, thereby improving the accuracy of determining the delivered content from the filtered content set.
[0204] In another embodiment, adding the to-be-screened contents whose matching degree meets the predetermined conditions to the set of screened contents includes: ranking the matching degree of the to-be-screened contents; setting a fixed number, and adding the first fixed number of the to-be-screened contents to the set of screened contents according to the order of the matching degree ranking from high to low. For example, if the fixed number is set to 10, the to-be-screened contents with the top 10 matching degrees with the target object are selected from the set of to-be-screened contents to be added to the set of screened contents. When the matching degree of the first label of the target object and the second label of the to-be-screened contents are both low, if the to-be-screened contents are screened according to the matching degree threshold, the number of contents in the set of screened contents may not meet the requirements for delivery to the target object. Therefore, through this embodiment, it can be ensured that the number of contents to be delivered in the set of screened contents meets the requirements for delivery to the target object. By analyzing the reaction of the target object to the contents in the set of screened contents, a more accurate first label of the target object is obtained, so as to improve the matching degree of the target object and the to-be-screened contents in subsequent screening, thereby obtaining contents with a higher matching degree. Therefore, in this embodiment, the advantage of determining the filtered content set based on the matching ranking is that it provides data support for obtaining a more accurate first label of the target object, thereby gradually improving the accuracy of determining the delivery content of the content set to be filtered.
[0205] In one embodiment, Figure 8 As shown, based on the matching degree between the first tag and the second tag, a filtered content set is filtered out from the to-be-filtered content set, including:
[0206] Step 810: based on the matching degree, select the matching content having the first tag and the second tag matching from the to-be-screened content set;
[0207] Step 820: Obtain multiple delivery indicators for each matching content;
[0208] Step 830: Based on multiple delivery indicators, filtered contents are filtered out from the matching contents to form a filtered content set.
[0209] In step 810, the process of obtaining matching content based on the matching degree between the first tag and the second tag is the same as the process of filtering the content to be filtered according to the predetermined conditions in step 730 in the above embodiment, and will not be repeated here.
[0210] In step 820, the delivery index of the matched content refers to an index used to evaluate the delivery effect of the content. The delivery effect refers to the feedback of the object viewing the content on the delivery of the content, such as whether the content has been watched to the end, whether the content has been liked, etc. The delivery index is to count the delivery effect of the content into specific values in order to evaluate the content. The delivery index can be the number of likes, comments, favorites, playback volume, completion rate, and recommendation volume of the content.
[0211] For example, the obtained matching content includes matching content A, matching content B, and matching content C. The delivery indicators of each matching content are obtained respectively, and the delivery indicators of matching content A include: number of likes (250), number of comments (1650), number of playbacks (1235), completion rate (72.5%), etc.; the delivery indicators of matching content B include: number of likes (9980), number of comments (6630), number of playbacks (456810), completion rate (46.6%), etc.; the delivery indicators of matching content C include: number of likes (3460), number of comments (1890), number of playbacks (36578), completion rate (57.5%), etc.
[0212] In step 930, the filtered content is filtered from the matching content according to the delivery index. Since the delivery index is a numerical representation of the delivery effect of the content, the delivery effect obtained after the filtered content determined based on the delivery index is delivered to the target object may be better. For example, based on the above example, according to the delivery index of the three matching contents, the final filtered content set includes matching content B and matching content C.
[0213] In one embodiment, based on multiple delivery indicators, filtered content is filtered out from the matching content to form a filtered content set, including: determining a sub-score of each delivery indicator; determining a matching content delivery score based on the sub-scores of all delivery indicators; and filtering out the filtered content from the matching content based on the matching content delivery score.
[0214] In one embodiment, the sub-score of each delivery indicator is determined by a delivery indicator and sub-score comparison table. Each delivery indicator has a corresponding delivery indicator and sub-score comparison table. The score of the delivery indicator can be determined by the comparison table.
[0215] For example, here is an example of a table comparing the number of likes / comments and sub-scores:
[0216] Likes / Comments Likes / Comments Sub-score More than 10,000 times 100 5000-10000 times 80 1000-5000 times 60 500-1000 times 40 200-500 times 20 Less than 200 times 10
[0217] Table 3
[0218] Based on the above example, the number of likes for the matching content A is 250. By looking up Table 3, the corresponding sub-score of the number of likes is 20 points; the number of comments is 1650. By looking up Table 3, the corresponding sub-score of the number of comments is 60 points.
[0219] The following is an example of a comparison table of playback indicators and sub-scores:
[0220] Views Play Quantum Fraction More than 500,000 times 100 100000-500000 times 80 50000-100000 times 60 10000-50000 times 40 5000-10000 times 30 1000-5000 times 20 Less than 1000 times 10
[0221] Table 4
[0222] Based on the above example, the playback volume of the matching content A is 3235 times. By looking up Table 4, the corresponding playback quantum score is 20 points; the playback volume of the matching content B is 456810 times. By looking up Table 4, the corresponding playback volume score is 80 points.
[0223] Here is an example of a table comparing the completion rate metric and sub-scores:
[0224] Completion rate Completion Rate Sub-Score 80%-100% 100 60%-80% 80 40%-60% 60 20%-40% 40 10%-20% 20 0-10% 10
[0225] Table 5
[0226] Based on the above example, the completion rate of matching content A is 72.5%. By looking up Table 5, the corresponding completion rate sub-score is 80 points; the completion rate of matching content B is 46.6%. By looking up Table 5, the corresponding completion rate sub-score is 60 points.
[0227] The advantages of using a comparison table to determine the sub-scores of the delivery indicator are: easy to find and low processing overhead.
[0228] After the sub-scores of all delivery indicators are obtained, the matching content delivery score of the matching content is determined according to the sub-scores.
[0229] In one embodiment, the matching content delivery score can be the average of the sub-scores of each delivery indicator. For example, if the likes sub-score of matching content A is 20, the comments sub-score is 60, the play quantum score is 20, and the completion rate sub-score is 80, then the matching content delivery score of matching content A is (20+60+20+80) / 4=45. The advantage of using the average to calculate the matching content delivery score is that it equally reflects the impact of different delivery indicators on the matching content.
[0230] In another embodiment, the matching content delivery score can be a weighted average of the delivery indicator sub-scores. In this embodiment, it is necessary to set a weight for each delivery indicator sub-score. Based on the above example, the weight of the likes sub-score is 0.6, the weight of the comments sub-score is 0.6, the weight of the play quantum score is 0.4, and the weight of the completion rate sub-score is 0.7. Then the matching content delivery score of the matching content is (20*0.6+60*0.6+20*0.4+80*0.7) / (0.6+0.6+0.4+0.7)=48.7. The advantage of using a weighted average to calculate the matching content delivery score is that different delivery indicators have different effects on the matching content. By setting different weights for different delivery indicator sub-scores, the flexibility of screening the screened content can be improved.
[0231] After determining the matching content delivery scores of all matching contents, the filtered contents are filtered according to the matching content delivery scores and the predetermined conditions. The predetermined conditions for filtering may be:
[0232] Set a delivery score threshold, and use the matching content with a delivery score greater than the delivery score threshold as the content to be selected;
[0233] Alternatively, the matching contents are sorted from high to low according to the matching content delivery scores, and the first predetermined number of matching contents are used as the contents to be screened.
[0234] The method of filtering the filtered content according to the matching content delivery score is similar to the method of filtering the deleted content according to the matching degree of the first tag and the second tag in the aforementioned embodiment, and will not be repeated here.
[0235] The advantage of determining the filtered content based on the matching content delivery score of each matching content is that the calculation speed is fast, which improves the efficiency of determining the filtered content.
[0236] The advantage of the embodiment of steps 910-930 is that the to-be-screened content is further screened based on the delivery index, so that the content in the screened content set is more likely to be of interest to the target object, further improving the accuracy of delivery.
[0237] The advantage of the embodiment of steps 710-730 is that the delivery content is preliminarily screened according to the label of the target object, ensuring that the content to be delivered for multi-task prediction in the subsequent process is highly matched with the target object, saving computing resources and improving the matching degree between the content to be delivered and the target object.
[0238] In step 540, the features of the content to be delivered are combined with the object features obtained in step 510 and the scene features obtained in step 520 to serve as multiple delivery basic features. Figure 6As shown in the schematic diagram of the basic features of delivery, the object features include: object feature A, object feature B, and object feature C; the scene features include: scene feature A, scene feature B, and scene feature C; the filtered content set selected from the to-be-screened content set includes: to-be-delivered content A and to-be-delivered content B, wherein the to-be-delivered content A includes to-be-delivered content feature A1 and to-be-delivered content feature A2, and the to-be-delivered content B includes to-be-delivered content feature B1 and to-be-delivered content feature B2. The to-be-delivered content features of the to-be-delivered content in the filtered content set are sequentially used together with the object features and the scene features as the basic features of delivery. The delivery basic feature A includes object feature A, object feature B, object feature C, scene feature A, scene feature B, scene feature C, to-be-delivered content feature A1, and to-be-delivered content feature A2, which are used to predict the target object's task completion probability for multiple tasks of the to-be-delivered content A. The basic delivery feature B includes object feature A, object feature B, object feature C, scene feature A, scene feature B, scene feature C, feature B1 of the content to be delivered, and feature B2 of the content to be delivered, which are used to predict the task completion probability of the target object for multiple tasks of the content B to be delivered.
[0239] In step 550, the delivery basic feature vector is quantized to obtain a corresponding first feature vector.
[0240] Since the delivery basic features may include multiple types of features, such as character type or numeric type, it is necessary to vectorize the delivery basic features into a form that can be learned by the machine.
[0241] The basic features for delivery may be character-based features, such as industry and education level in object features, geographic area features in scene features, and content type in features of content to be delivered.
[0242] If the basic feature is a character feature, the character feature can be input into the embedding layer to obtain the first feature vector corresponding to the character feature.
[0243] Since character features are sequences of characters, each character is considered a discrete symbol that cannot be recognized by machines. The role of the embedding layer is to convert discrete characters into vectors that can be recognized by machines.
[0244] Input the character feature into the embedding layer, and the embedding layer will output the corresponding unique first feature vector for the character feature. For example, Fig. 9AAs shown, after the character feature "efficient" is input into the embedding layer, the first feature vector is [1,0,1,0,...,1,1,0]; after the character feature "doctor" is input into the embedding layer, the first feature vector is [0,0,1,0,...,0,1,0]; after the character feature "Shanghai" is input into the embedding layer, the first feature vector is [1,1,0,1,...,1,0,0].
[0245] The delivery basic feature may also be a numerical feature, for example, the content delivery amount in the feature of the content to be delivered.
[0246] If the basic feature is a numerical feature, the numerical feature is numerically mapped to obtain a first feature vector corresponding to the numerical feature.
[0247] The process of value mapping can be to convert the value into its corresponding binary form. Fig. 9B The numerical mapping conversion shown in the figure. The first eigenvector obtained by the numerical feature "27" after numerical conversion is 00011011; the first eigenvector obtained by the numerical feature "560" after numerical conversion is 1000110000; the first eigenvector obtained by the numerical feature "99" after numerical conversion is 01100011.
[0248] The basic delivery feature can also be an indefinite length sequence feature. An indefinite length sequence feature refers to a string feature consisting of a large number of characters of an uncertain number. For example, the object signature in the object feature is an object feature that is filled in and formed by the object itself.
[0249] If the basic feature is a variable-length sequence feature, the variable-length sequence feature is pooled to obtain a first feature vector corresponding to the variable-length sequence feature.
[0250] Pooling is the process of reducing the dimensionality of a high-dimensional vector. The specific process of pooling is to integrate several features in a preset neighborhood of a high-dimensional vector generated based on indefinite length sequence features into a new feature. Fig. 9CThe pooling process shown is shown in the figure. The indefinite length sequence feature is the object signature in the object feature. The object signature is "Reading makes me progress,...". According to the indefinite length sequence feature, the corresponding high-dimensional feature vector can be obtained, and the high-dimensional feature vector is [0,1,1,0,0,1,0,0,0,1,1,1,0,0,0,1,0,1,0,0,0,1,1,1]. The value of the preset neighborhood is set to 3, that is, every 3 feature quantities in the high-dimensional feature vector are taken as a neighborhood, and the 3 feature quantities are integrated into 1 feature quantity. The integration method can be to take the average value of the feature quantity in the neighborhood as the feature quantity of the neighborhood. For example, in the above high-dimensional feature vector, the first neighborhood is taken as [0,1,1], and the average value of the feature quantity in this neighborhood is 0.67. Therefore, the feature value of this neighborhood is 0.67. The second neighborhood is [0,0,1], and the average value of the feature quantity in this neighborhood is 0.33. Therefore, the feature value of this neighborhood is 0.33. The third neighborhood is [0,0,0], and the average value of the feature quantity in this neighborhood is 0. Therefore, the feature value of this neighborhood is 0. By analogy, the feature quantities of the subsequent neighborhoods are 1, 0, 0.67, 0, and 1. Therefore, the first feature vector obtained after pooling of the indefinite length sequence feature is [0.67, 0.33, 0, 1, 0, 0.67, 0, 1].
[0251] Integrating several features in a preset neighborhood into a new feature can also be done by taking the maximum value of the feature quantity in the neighborhood as the feature value of a neighborhood. For example, in the above high-dimensional feature vector, the value of the preset neighborhood is also set to 3, and the maximum value in the neighborhood is taken as the feature value of the neighborhood. Then, the first feature vector obtained after the indefinite length sequence feature is pooled is [1,1,0,1,0,1,0,1].
[0252] The advantage of using different vectorization methods for different types of delivery basic features is that the features can be vectorized according to the characteristics of different types of delivery basic features, thereby ensuring the accuracy of the feature vectorization results.
[0253] The embodiment of steps 510-550 combines the object characteristics of the target object, the scene characteristics of the scene, and the characteristics of the content to be delivered to form the basic delivery characteristics, so that the constructed basic delivery characteristics cover all aspects that affect the probability of the target object completing the task of the content to be delivered, thereby improving the comprehensiveness of obtaining the basic delivery characteristics.
[0254] Detailed description of step 320 and step 330
[0255] In step 320, a plurality of first feature vectors are concatenated into a first concatenated feature vector.
[0256] Cascading multiple first eigenvectors refers to the process of concatenating the first eigenvectors in the order of the output first eigenvectors to form a vector. Fig.10 In the first eigenvector cascading process shown in FIG. 1 , the first eigenvectors include: [1,0,1,1,0,1,1,0], [0,0,1,0,1,0,1,0], ..., [0,0,0,1,1,0,1,1], ..., [0,1,1,1,1,0,0,1,0,1], .... The first cascaded eigenvector obtained by cascading multiple first eigenvectors is: [1,0,1,1,0,1,1,0,0,0,1,0,1,0,1,0, ..., 0,0,0,1,1,0,1,1, ..., 0,1,1,1,1,0,0,1,0,1, ...].
[0257] In step 330, the first cascade feature vector is input into a first number of expert models. The first number is the total number of expert models used for multi-target prediction. The first number of expert models is determined by the task volume. Because, when the task volume increases, more dimensions need to be analyzed and calculated for the first cascade feature vector. Therefore, when the task volume increases, the first number of expert models will also increase.
[0258] The first cascade feature vectors are sequentially input into the expert model.
[0259] Detailed description of step 340
[0260] In step 340, a second number of expert models is selected from the first number of expert models through the gating nodes corresponding to the task. It should be noted that the second number must be smaller than the first number.
[0261] The reason for selecting the second number of expert models from the first number of expert models is that when the number of expert models increases with the increase of tasks, if the outputs of all the first number of expert models are input into the gating nodes corresponding to the tasks, the amount of calculation and the processing load will increase significantly. Therefore, the second number smaller than the first number is set to screen the first number of expert networks to ensure the calculation efficiency.
[0262] The second number of expert models selected for different tasks may overlap, partially overlap, or not overlap at all.
[0263] In one embodiment, Fig.11 As shown, selecting a second number of expert models from the first number of expert models includes:
[0264] Step 1110: Generate a weight vector based on the first cascaded eigenvector and the first matrix;
[0265] Step 1120: Select a second number of expert models from the first number of expert models based on the weight vector.
[0266] In step 1110, the first matrix is a trainable parameter matrix. The number of columns of the first matrix is the first number of expert models, and the number of rows is the dimension of the first cascade feature vector. The dimension of the first cascade feature vector is the sum of the dimensions of all first feature vectors. For example, if the first number of expert models is 5 and the dimension of the first cascade feature vector is 10, then the number of columns of the first matrix is 5 and the number of rows is 10.
[0267] The first matrix is trainable. That is, after a round of expert model screening, the first matrix can be updated based on the predicted results of the task completion probability, so that when the next expert model screening is performed, the second number of expert models activated by the task changes. If the first matrix is fixed, then the results of the second number of expert models selected each time may be the same, and the result of the task completion probability obtained will depend on the selected expert model, resulting in inaccurate delivery processing results.
[0268] The weight vector indicates the weights of the first number of expert models. The dimension of the weight vector is the first number, that is, each value in the weight vector represents the weight of an expert model.
[0269] Based on the first cascaded eigenvector and the first matrix, the calculation process of generating the weight vector is shown in Formula 1:
[0270]
[0271] In formula 1, represents the first matrix; X represents the first cascade eigenvector; H t (X) represents the weight vector. For example, the first number is 4, the dimension of the first concatenated feature vector is 3, then the number of columns of the first matrix is 4 and the number of rows is 3. Assume that the first matrix is The first cascade feature vector is [1,1,0]. The four values in the weight vector represent the weights of the four expert models.
[0272] In one embodiment, Fig.12 As shown, based on the first cascaded feature vector and the first matrix, a weight vector is generated, including:
[0273] Step 1210: multiply the first matrix by the first cascaded eigenvector to obtain an original weight vector;
[0274] Step 1220: multiply the second matrix by the first cascade eigenvector to obtain a weight adjustment basis vector;
[0275] Step 1230: Generate a weight adjustment value vector based on the random disturbance vector and the weight adjustment basis vector;
[0276] Step 1240: Generate a weight vector based on the original weight vector and the weight adjustment value vector.
[0277] In step 1210, the process of obtaining the original weight vector is the same as the process of calculating the weight vector in the above-mentioned embodiment. The calculation process is shown in formula 1 and will not be repeated here.
[0278] In step 1220, the second matrix is similar to the first matrix and is also a trainable parameter matrix. The number of columns of the second matrix is the first number, and the number of rows is the dimension of the first cascaded feature vector. In other words, the second matrix has the same dimension as the first matrix.
[0279] The second matrix is also trainable. After a round of expert model screening, the second matrix can be updated based on the predicted results of the task completion probability, so that the selected expert model changes when the next expert model screening is performed.
[0280] The weight adjustment basis vector generated based on the second matrix is used to perturb the original weight vector to prevent the generated weight vector from being overly dependent on certain eigenvalues.
[0281] The calculation process for generating weight-adjusted basis vectors is shown in Formula 2:
[0282]
[0283] In formula 2, represents the second matrix, Represents the weight adjustment basis vector. The dimension of the weight adjustment basis vector is the same as the dimension of the original weight vector.
[0284] In step 1230, the random perturbation vector increases the uncertainty of the perturbation term. The random perturbation vector is a non-trainable vector with a dimension of the first number. That is, the random perturbation vector is not affected by the results of each round of training and is randomly generated.
[0285] In one embodiment, Fig.13 As shown, based on the random perturbation vector and the weight adjustment basis vector, a weight adjustment value vector is generated, including:
[0286] Step 1310: Obtain a random number set, where the random numbers in the random number set conform to a normal distribution;
[0287] Step 1320: randomly extract a first number of random numbers from the random number set to form a random perturbation vector;
[0288] Step 1330: perform dot multiplication of the random disturbance vector and the weight adjustment basis vector to obtain a weight adjustment value vector.
[0289] In step 1310, a third number of random numbers within a predetermined range is obtained based on a normal distribution, and the generated random numbers are combined into a random number set, wherein the third number is greater than the first number. For example, the third number is set to 10, the predetermined range is set to [0,5], and the generated random number set may be: [1.8, 2.2, 2.5, 3.7, 2.5, 2.4, 2.9, 0.5, 3.2, 4.5].
[0290] In step 1320, a first number of random numbers is randomly selected from the random number set. For example, the first number is 5, and five random numbers are randomly selected from the random number set as random perturbation vectors. The random perturbation vectors may be: [1.8, 2.2, 2.5, 3.2, 4.5].
[0291] In step 1330, the dimensions of the random perturbation vector and the weight adjustment basis vector are both the first number, so the random perturbation vector and the weight adjustment vector can be dot-multiplied to obtain a weight adjustment value vector with the first number of dimensions. Each value in the weight adjustment value vector is a perturbation factor for determining the weight of an expert model.
[0292] The process of calculating the weight adjustment value vector is shown in Formula 3:
[0293]
[0294] In Formula 3, α represents the random perturbation vector, represents the weight-adjusted basis vector, Represents the weight adjustment value vector. For example, if the random perturbation vector is [1.8, 2.2, 2.5, 3.2, 4.5] and the weight adjustment basis vector is [0, 2, 1, 1, 0], then the weight adjustment value vector is [1.8*0, 2.2*2, 2.5*1, 3.2*1, 4.5*0] = [0, 4.4, 2.5, 3.2, 0].
[0295] In steps 1310-1330, a random disturbance vector is generated based on a normal distribution, which improves the stability of the random numbers in the random disturbance vector, thereby improving the stability of the generated weight adjustment value vector.
[0296] In step 1240, a weight vector is generated based on the original weight vector and the weight adjustment value vector. The order of the values in the weight vector corresponds to the arrangement order of the expert models, and the values in the weight vector represent the weights of the expert models at the corresponding positions. For example, the arrangement order of the five expert models is arranged in the order of AE, and the weight vector is [1, 2.7, 1.9, 3.4, 0.8]. Then the weight of expert model A is 1, the weight of expert model B is 2.7, the weight of expert model C is 1.9, and so on.
[0297] In one embodiment, the weight vector is obtained by adding the original weight vector and the weight adjustment value vector. Since the original weight vector and the weight adjustment value vector are both vectors with a first number of dimensions, the obtained weight vector is also a vector with a first number of dimensions.
[0298] Based on the above embodiment, the calculation process of generating the weight vector is shown in Formula 4:
[0299]
[0300] In Formula 4, the first half is the original weight vector generated based on the first matrix and the first cascade eigenvector, and the second half is the weight adjustment value vector generated based on the random perturbation vector and the weight adjustment basis vector. The two parts are added together to obtain the weight vectors of the first number of expert networks. For example, the original weight vector is: [1,2,1,0,1], and the weight adjustment value vector is [0,4.4,2.5,3.2,0], then the weight vector is: [1+0,2+4.4,1+2.4,0+3.2,1+0] = [1,6.4,3.4,3.2,1].
[0301] The advantage of generating the weight vector by summing is that the original weight vector and the weight adjustment value vector have the same influence on the generated weight vector, thereby improving the fairness of the generated weight vector.
[0302] In another embodiment, the weight vector is obtained by calculating a weighted sum of the original weight vector and the weight adjustment value vector.
[0303] In this embodiment, weights are set for the original weight vector and the weight adjustment value vector. For example, the weight of the original weight vector is 0.6, and the weight of the weight adjustment value vector is 0.4. Based on the example in the above embodiment, the final generated weight vector is: 0.6*[1,2,1,0,1]+0.4*[0,4.4,2.5,3.2,0]=[0.6*1+0.4*0.4,0.6*2+0.4*4.4,0.6*1+0.4*2.5,0.6*0+0.4*3.2,0.6*1+0.4*0]=[0.76,2.96,1.6,1.28,0.6].
[0304] The advantage of generating the weight vector by weighted sum is that the weights can be flexibly set for the original weight vector and the weight adjustment value vector, thereby improving the flexibility of generating the weight vector.
[0305] In the embodiment of steps 1210-1240, interference factors are added in the process of generating weight vectors to prevent the generated weight vectors from being overly dependent on certain eigenvalues. Therefore, the advantage of the embodiment of steps 1210-1240 is that the robustness and accuracy of generating weight vectors are improved.
[0306] In step 1120, a second number of expert models are selected from the first number of expert models according to the weight vector.
[0307] In one embodiment, Fig.14 As shown, based on the weight vector, a second number of expert models are selected from the first number of expert models, including:
[0308] Step 1410: determine the first second number of elements in the weight vector whose values are sorted from largest to smallest;
[0309] Step 1420: Use the expert models corresponding to the first second number of elements as the selected second number of expert models.
[0310] Since each element in the weight vector corresponds to the weight of an expert model, by sorting the elements in the weight vector from large to small, the weights of different expert models can be sorted from large to small.
[0311] For example, the weight vector is [0.76, 2.96, 1.6, 1.28, 0.6], and the elements correspond to expert model A, expert model B, expert model C, expert model D, and expert model E. The elements in the weight vector are sorted from large to small, and the result is [2.96, 1.6, 1.28, 0.76, 0.6]. If the second number is 3, then the expert models corresponding to the first three elements in the sorted weight vector are: expert model B, expert model C, and expert model D.
[0312] In this embodiment, the expert models are selected in the order of their weights, so that the selected expert models have a greater impact on the prediction of the task completion probability, thereby improving the accuracy of screening the second number of expert models.
[0313] In the embodiment of steps 1110-1120, based on the weights of different expert models, the second number of expert models are selected from the first number of expert models to ensure that the expert model selected for the task is the most influential. Therefore, the advantage of the embodiment of steps 1110-1120 is that the reliability of determining the second number of expert models is improved.
[0314] In another embodiment, Fig.15 As shown, selecting a second number of expert models from the first number of expert models includes:
[0315] Step 1510: Get the task type of the task;
[0316] Step 1520: Select a third number of expert models that match the task type from the first number of expert models;
[0317] Step 1530: Select a second number of expert models from the third number of expert models.
[0318] Task types are divided according to the impact of the task on the delivery results of the content to be delivered. For example, whether to like, comment, and other tasks can be completed is very important for whether the content to be delivered is delivered to the target object. Because through this type of task, you can intuitively see whether the target object is interested in the content to be delivered. Therefore, this type of task is a first-class task. Tasks such as whether to complete the viewing and whether to share with other objects have a lower impact on whether the content to be delivered is delivered to the target object. Because some content to be delivered, even if it is not watched in full or not shared with other objects, the target object is still interested. Therefore, this type of task is a second-class task. Similarly, tasks can be divided into multiple categories based on the impact of whether it is delivered to the target object.
[0319] Since different expert models can handle different types of tasks. For example, the calculation of expert model A is more accurate and has a better modeling effect for a type of task, then the task type corresponding to expert model A is a type of task. Therefore, before determining the second number of expert models required for the task, the first number of expert models can be screened to select the third number of expert models that match the task. The first number of expert models includes expert models corresponding to multiple tasks. Therefore, the first number is greater than the third number.
[0320] In step 1530, the method of selecting the second number of expert models from the third number of expert models is the same as the method of selecting the second number of expert models from the first number of expert models in the aforementioned embodiment. The third number is greater than the second number. First, the weight vectors of the third number of expert models are obtained; secondly, based on the weight vectors, the second number of expert models are selected from the third number of expert models. The specific process has been described in detail in the aforementioned embodiment and will not be repeated here.
[0321] In this embodiment, an expert model matching the task is first obtained, and then a second number of expert models are screened again from the matching expert models, thereby avoiding the problem of low computational efficiency caused by directly screening the second number of expert models when the first number is too large. Furthermore, screening based on the matching expert models can ensure that the screened second number of expert models are all strongly related to the task. Therefore, the advantage of this embodiment is that the efficiency and accuracy of obtaining the second number of expert models are improved.
[0322] In step 340, after the second number of expert models are selected, the gating node corresponding to the task needs to perform a weighted sum on the outputs of the second number of selected expert models to obtain a second feature vector corresponding to the task.
[0323] Based on the aforementioned embodiment of selecting a second number of expert models based on the weight vector, as Fig.16 As shown, in one embodiment, a weighted sum is performed on the outputs of the selected second number of expert models, including:
[0324] Step 1610: modify the other elements in the weight vector that are not the first second number of elements sorted from largest to smallest to negative infinity;
[0325] Step 1620: performing normalized exponential processing on the weight vector to obtain first weights of a second number of expert models;
[0326] Step 1630: Perform a weighted sum on the outputs of the second number of expert models selected according to the first weight.
[0327] In step 1410, the elements in the weight vector are sorted from large to small. The expert model corresponding to the first second number of elements is selected according to the sorting result. In 1610, all elements after the first second number of elements in the sorting result are modified to negative infinity. The advantage of modifying to negative infinity is that it clearly shows which expert models corresponding to the elements in the weight vector should be activated, and saves the amount of calculation in the subsequent process.
[0328] Therefore, the result obtained after modifying the elements in the weight vector is as shown in Formula 5:
[0329]
[0330] In Formula 5, H t (X) i Indicates H t The i-th element in (X) corresponds to the i-th expert model among all expert models. k represents the second number. If H t (X) i The value belongs to H t The first k elements in (X) are sorted, so keep H t (X)H t (X) i Otherwise, H t (X) i becomes negative infinity.
[0331] The weight vector is subjected to normalized exponential processing. The purpose of normalized exponential processing is to limit the weights of the second number of expert models to the range of [0,1] so that different expert models are comparable. The normalized exponential processing uses the softmax function. The weight vector is input to the softmax function, where the second number of elements in the weight vector that are not modified to negative infinity will obtain the first weight of the corresponding expert model after passing through the softmax function.
[0332] The process of calculating the first weights of the second number of expert models is shown in Formula 6:
[0333] g t (X) = softmax(H t (X),k) (Formula 6).
[0334] In Formula 6, k represents the second number, g t (X) represents the first weight of the second number of expert models. The part in the softmax function represents that the first k elements in the weight vector are normalized exponentially.
[0335] After obtaining the first weights of the second number of expert models, a weighted sum is calculated for the outputs of the second number of expert models.
[0336] In one embodiment, the expert model includes a kernel parameter matrix and a bias vector. Fig.17 As shown, the outputs of the second number of expert models selected are weighted and summed according to the first weight, including:
[0337] Step 1710: Obtain a three-dimensional kernel parameter tensor, where the three-dimensional kernel parameter tensor represents kernel parameter matrices of a first number of expert models;
[0338] Step 1720: Obtain a bias vector matrix, where the bias vector matrix represents bias vectors of a first number of expert models;
[0339] Step 1730: Obtain a kernel parameter matrix corresponding to a second number of first weights from the three-dimensional kernel parameter tensor;
[0340] Step 1740: Obtain bias vectors corresponding to a second number of first weights from the bias vector matrix;
[0341] Step 1750: Determine a weighted sum based on the second number of first weights, the kernel parameter matrix corresponding to the second number of first weights, and the bias vector corresponding to the second number of first weights.
[0342] The three-dimensional kernel parameter tensor is a three-dimensional tensor composed of the kernel parameter matrices of the first number of expert models. Fig.18A As shown in the three-dimensional kernel parameter tensor, each layer represents a kernel parameter matrix of an expert model. The number of columns of the kernel parameter matrix is the dimension of the output of the expert model, and the number of rows is the dimension of the first cascade feature vector. In the expert model, the kernel parameter matrix is used to perform high-order mapping on the first cascade feature vector.
[0343] The bias vector matrix is composed of the bias vectors of the first number of expert models, and the dimension of the bias vector is equal to the dimension of the output of the expert model. Fig.18B As shown in the bias vector matrix, each row represents a bias vector matrix of an expert model. The bias vector is used in the expert model to perform linear addition on the convolved data, so that the data value has a horizontal offset to increase the nonlinear ability of the expert network and avoid useful data being filtered out.
[0344] The kernel parameter matrix and the bias vector matrix store the parameters of the corresponding expert models. The three-dimensional kernel parameter tensor stores the kernel parameter matrices of all expert models together; the bias vector matrix stores the bias vectors of all expert models together.
[0345] After determining the second number of expert models according to the weight vector, the kernel parameter matrix and bias vector required for the corresponding expert model can be directly determined in the three-dimensional kernel parameter tensor and the bias vector matrix by the index of the first weight corresponding to the expert model in the weight vector. For example, the second number is 3, and the indexes of the top three elements in the weight vector are 1, 4, and 9, respectively. Therefore, the kernel parameter matrix with indexes of 1, 4, and 9 is obtained from the three-dimensional kernel parameter tensor, and the bias vector with index values of 1, 4, and 9 is obtained from the bias vector matrix.
[0346] After obtaining the kernel parameter matrix and the bias vector corresponding to the expert model, the output of the expert model for the first cascade feature vector can be obtained. The gating node corresponding to the task determines the weighted sum based on the output of the expert model and the first weight.
[0347] In one embodiment, Fig.19 As shown, determining the weighted sum based on the second number of first weights, the kernel parameter matrix corresponding to the second number of first weights, and the bias vector corresponding to the second number of first weights includes:
[0348] Step 1910: multiply the kernel parameter matrix corresponding to each expert model in the second number of expert models by the first cascade feature vector of the input expert model to obtain an intermediate vector;
[0349] Step 1920: Add the intermediate vector to the bias vector corresponding to the expert model to obtain a result vector;
[0350] Step 1930: Apply an activation function to the result vector to obtain the output of the expert model;
[0351] Step 1940: Obtain a weighted sum based on the outputs of the second number of expert models and the first weights corresponding to the expert models.
[0352] In step 1910, for one of the expert models in the second number of expert models, a kernel parameter matrix corresponding to the expert model is multiplied by a first cascade feature vector of the input expert model to obtain a high-order vector of the first cascade feature vector.
[0353] In step 1920, the dimension of the intermediate vector is the same as the dimension of the bias vector corresponding to the expert model, and the two vectors are added to obtain the result vector of the expert model.
[0354] In step 1930, an activation function is applied to the result vector to obtain the output of the expert model. The activation function may be a sigmoid activation function, a tanh activation function, a Relu activation function, etc.
[0355] Taking the Relu activation function as an example, based on steps 1910-1930, the process of obtaining the output of the expert model is shown in Formula 7:
[0356] e i (X) = ReLu(W i ·X+b i ) (Formula 7).
[0357] In Formula 7, W i represents the kernel parameter matrix of the expert network with index i; b i represents the bias vector of the expert network; X refers to the first cascade feature vector; ReLu is the activation function; e i (X) is the output of the expert network for the first cascade feature vector.
[0358] In step 1940, after obtaining the outputs of the second number of expert networks, the gated node calculates the weighted sum using the first weight of each expert network and the output of the expert network. The process of calculating the weighted sum in the gated node is shown in Formula 8:
[0359]
[0360] In formula 8, [g t (X)] i It means that the first weight of the i-th expert network is obtained from the first weight of the second number of expert models. In the gating node corresponding to the task, the output of the second number of expert networks is multiplied by the corresponding first weight to obtain the weighted sum, that is, the second feature vector corresponding to the task output by the gating node.
[0361] In this embodiment, the output of the expert model is obtained through the activation function, so that the output result introduces nonlinear factors, and the high-order modeling results of some complex types of features are clearer. Based on the output of the expert model and the output result of the gated node corresponding to the corresponding first weight calculation task, the influence of different expert models on the task is clearly shown.
[0362] The advantage of the embodiment of steps 1710-1750 is that the parameters of the first number of expert models are stored using the three-dimensional kernel parameter tensor and the bias vector matrix, thereby realizing unified storage of the parameters and facilitating data management. When calling, the parameters of the activated expert model are directly found through the index of the first weight in the weight vector, thereby avoiding the calculation of other unactivated expert models and improving the efficiency of determining the expert model parameters and the output of the expert model.
[0363] The advantage of steps 1610-1630 is that the first weights of the second number of expert models are normalized, so that different expert models are comparable, avoiding the problem of output result deviation due to excessive weight difference.
[0364] For each task, the second number of activated expert models needs to be preset and needs to be smaller than the first number. Fig. 20 As shown, the second number is determined by:
[0365] Step 2010: Get the task type of the task;
[0366] Step 220: Obtain the prediction accuracy requirement of the task;
[0367] Step 2030: Determine a second number based on the task type and prediction accuracy requirement.
[0368] The task type has been described in detail in the above embodiments and will not be described here again. Different task types indicate different degrees of influence of the task on the delivery result. Therefore, if the task is a type of task with a high degree of influence on the delivery result, the second number of expert models that need to be activated to obtain the task completion probability may be relatively high.
[0369] The prediction accuracy requirement refers to the accuracy requirement for the task completion probability. The accuracy requirement can be 0.1, 0.01, etc. An accuracy requirement of 0.1 means that the task completion probability is constrained to the tenth place; an accuracy requirement of 0.01 means that the task completion probability is constrained to the hundredth place, and so on. The higher the accuracy requirement, the more accurate the result of the task completion probability. Therefore, the higher the prediction accuracy requirement, the higher the second number of expert models that need to be activated may be.
[0370] In one embodiment, Fig.21 As shown, based on the task type and the prediction accuracy requirement, determining the second number includes:
[0371] Step 2110: Determine a first score based on the task type;
[0372] Step 2120: Determine a second score based on the prediction accuracy requirement;
[0373] Step 2130: Determine a total score based on the first score and the second score;
[0374] Step 2140: Determine a second number based on the total score.
[0375] In step 2110, the first score may be determined based on the task type by using a first comparison table. Table 6 shows an example of a first comparison table.
[0376] Task Type First Score A type of task 100 Second type of task 80 Three types of tasks 60 Four types of tasks 40
[0377] Table 6
[0378] Referring to Table 6, if the task type is a second-class task, the first score corresponding to the task type is 80.
[0379] In step 2120, the second score may be determined based on the prediction accuracy requirement by using a second comparison table. Table 7 shows an example of a second comparison table.
[0380] Prediction accuracy requirements Second score 0.0001 100 0.001 80 0.01 60 0.1 40
[0381] Table 7
[0382] Referring to Table 7, if the prediction accuracy requirement is 0.01, the second score corresponding to the prediction accuracy requirement is 60.
[0383] In step 2130, the total score may be determined based on the first score and the second score by an average or a weighted average.
[0384] In one embodiment, the total score is calculated using an average. For example, if the first score is 80 and the second score is 60, then the total score is (80+60) / 2=70. The advantage of using an average to calculate the total score is that the task type and the prediction accuracy requirement have the same impact on the calculation of the total score, which improves the fairness of determining the second number.
[0385] In another embodiment, the total score is calculated using a weighted average. In this embodiment, first, weights need to be set for the first score and the second score respectively. For example, the weight of the first score is 0.7, the weight of the second score is 0.3, the first score is 80, and the second score is 60, then the total score is 80*0.7+60*0.3=74. The advantage of using a weighted average to calculate the total score is that different weights can be flexibly set for the first score and the second score according to the needs of the actual application, thereby improving the flexibility of determining the second number.
[0386] In step 2140, the second number is determined according to the total score. The higher the total score, the more expert models are needed, and the higher the second number.
[0387] In one embodiment, the second number can be determined by a third comparison table between the total score and the second number. Table 8 shows an example of the third comparison table.
[0388] Total score Second number (ratio of the first number) 80-100 70% 60-80 50% 40-60 30% Below 40 10%
[0389] Table 8
[0390] The second number in the third comparison table can be a specific number or a ratio of the second number to the first number. Referring to Table 8, if the total score is 70, then the ratio of the second number to the first number is 50%. In other words, assuming that the first number is 10, then the second number should be 5.
[0391] The advantage of determining the second number through the third comparison table is that it is easy to query and has high processing efficiency.
[0392] The advantage of the embodiment of steps 2110-2140 is that the second number is determined based on the corresponding scores of the task type and the prediction accuracy requirement, and the first score and the second score can be adjusted according to the needs of the actual application, thereby adjusting the second number, thereby improving the flexibility of determining the second number.
[0393] The advantage of the embodiment of steps 2010-2030 is that the second number is determined according to the task type and the accuracy requirement for the task completion probability, so that the second number can meet the task requirements, achieving the effect that the delivery accuracy will not be deteriorated even if the expert model is not used.
[0394] Detailed description of step 350
[0395] In step 350, for a single task, the second feature vector output by the gating node corresponding to the task is input into the prediction model corresponding to the task, and the prediction model outputs the predicted target object for the content to be delivered and the task completion probability of the corresponding task.
[0396] For example, the task corresponding to the second expert network is: whether the target object will like the content to be delivered. Then, the third feature vector output by the second gated node in the second expert network is input into the prediction model corresponding to the task, and the prediction model will output the task completion probability for this task. For example, the task completion probability is 0.85, which means that the target object has an 85% chance of liking the content to be delivered.
[0397] The prediction model corresponding to each task will output the task completion probability of the target object for the corresponding task.
[0398] The process of the prediction model obtaining the task completion probability of the corresponding task is shown in Formula 9:
[0399]
[0400] In Formula 9, represents the second feature vector of the gating node output corresponding to the task; p t represents the task completion probability obtained by the prediction model; θ t Indicates the prediction model corresponding to the task, which can be a deep neural network model.
[0401] The process of determining the content to be delivered to the target audience
[0402] After the task completion probability of the target object for each task is obtained, the content to be delivered to the target object is determined based on the task completion probabilities of different tasks.
[0403] In one embodiment, Fig. 22 As shown, after inputting the second feature vectors corresponding to the multiple tasks into the prediction models corresponding to the multiple tasks to obtain the predicted task completion probabilities of the target objects for the content to be delivered with respect to the multiple tasks, the delivery processing method further includes:
[0404] Step 2210: Obtain the focus tasks of the target object;
[0405] Step 2220: Obtain the task completion probability of the task of interest from the task completion probabilities of the plurality of tasks;
[0406] Step 2230: sort the content to be delivered based on the task completion probability of the concerned task;
[0407] Step 2240: Deliver the content to be delivered to the target object according to the ranking.
[0408] The target object's attention task refers to the task that can represent the target object's interest in the delivered content among multiple tasks. For example, multiple tasks include: whether to like, whether to comment, and whether to forward. If the target object likes to comment on the content of interest when viewing the delivered content, then the target object's attention task is "whether to comment".
[0409] In one embodiment, obtaining the focus tasks of the target object includes: obtaining the historical viewing logs of the target object within a predetermined time period before the current time; and determining the focus tasks of the target object based on the historical viewing logs.
[0410] When obtaining the target object's attention task, the target object's consent must be obtained in advance. This is similar to the previous content, so I will not repeat it.
[0411] In the above embodiment, the object characteristics of the target object are obtained through the historical viewing log. The historical viewing log refers to the relevant records of the delivery content viewed by the target object in a predetermined time period before the current time point, including the task execution status for each delivery content. The historical viewing log has been introduced in detail in the above embodiment and will not be repeated here. When obtaining the historical viewing log of the target object, the consent of the target object must be obtained in advance.
[0412] In one embodiment, determining the focus task of the target object based on the historical viewing log includes: counting the number of executions of each task in the historical viewing log; and taking the task with the highest number of executions as the focus task of the target object.
[0413] Take the historical viewing log in Table 1 as an example, in which the tasks performed on the delivered content include: whether to watch and like, whether to collect, and whether to comment. Get the number of executions of each task in the historical viewing log, that is, the number of times the value corresponding to each task is "yes". In Table 1, the number of "whether to like" is 4 times, the number of "whether to collect" is 2 times, and the number of "whether to comment" is 2 times. Therefore, it can be inferred from Table 1 that the operation that the target object may perform on the content of interest is "like". Therefore, the target object's focus task is "whether to like".
[0414] In this embodiment, the focus task is determined by counting the number of times the target object executes each task, which ensures that the focus task is the action most frequently executed by the target object on the delivered content, thereby improving the accuracy of obtaining the focus task.
[0415] The advantage of determining the attention task of the target object based on the historical viewing log is that the attention task is determined by analyzing the short-term object behavior of the target object, which improves the matching degree between the attention task and the target object, thereby improving the accuracy of obtaining the attention task.
[0416] After step 350, the prediction model corresponding to each task will output the task completion probability of the target object for the task. The task completion probability of the target object's attention task is determined from the task completion probabilities of multiple tasks. For example, for content A to be delivered, content B to be delivered, and content C to be delivered, the task completion probabilities of task A, task B, task C, and task D are obtained through the prediction models corresponding to multiple tasks respectively. The target object's attention task is task B, so the task completion probability of task B of content A to be delivered, content B to be delivered, and content C to be delivered is obtained. After obtaining the task completion probability of the attention task of each content to be delivered, the task completion probability is sorted. The task completion probability of task B of content A to be delivered is 0.56, the task completion probability of task B of content B to be delivered is 0.37, and the task completion probability of task B of content C to be delivered is 0.73. Then, the three contents to be delivered are sorted in order from high to low according to the task completion probability of task B, and the sorting result is: content C to be delivered, content A to be delivered, content B to be delivered.
[0417] The order in which the content to be delivered is sorted by the probability of completing the task is the order in which the content to be delivered is delivered to the target object. In the above example, the content C to be delivered is delivered to the target object first, because the target object has the highest probability of executing task B on the content C to be delivered, which means that the target object is most likely to be interested in the content C among all the content to be delivered.
[0418] The advantage of the embodiment of steps 2210-2240 is that the delivery order of the content to be delivered is determined by obtaining the focus tasks of the target object, so that the priority delivered content to be delivered is more likely to be of interest to the target object, thereby improving the matching degree between the delivered content and the target object.
[0419] In another embodiment, Fig.23 As shown, based on the task completion probability of the concerned task, the content to be delivered is sorted, including:
[0420] Step 2310: If there are multiple tasks of interest, determine the overall task completion probability based on the task completion probabilities of the multiple tasks of interest;
[0421] Step 2320: Sort the content to be delivered based on the overall task completion probability.
[0422] The target object may perform multiple tasks for the content that it is interested in. For example, if the target object is interested in a delivery content, it may like the delivery content and make comments on it.
[0423] If there are multiple tasks of interest, then the task completion probability of each task of interest is obtained, and then the overall task completion probability of the content to be delivered is determined based on the task completion probability of each task of interest.
[0424] For example, the task completion probabilities of multiple focus tasks of multiple contents to be delivered are shown in Table 9:
[0425]
[0426]
[0427] Table 9
[0428] As shown in Table 9, the target object's attention tasks include attention content A and attention content B.
[0429] The overall task completion probability can be determined by calculating the average of the task completion probabilities of multiple tasks of interest. For example, according to Table 1, the overall task completion probability of content A to be delivered is (0.45+0.56) / 2=0.505. The overall task completion probability of content B to be delivered is (0.39+0.78) / 2=0.585. The overall task completion probability of content C to be delivered is (0.69+0.44) / 2=0.565. Therefore, based on the overall task completion probability, the result of sorting the content to be delivered is: content B to be delivered, content C to be delivered, and content A to be delivered.
[0430] The advantage of calculating the overall task completion probability by averaging is that the task completion probabilities of different focus tasks have the same impact on determining the overall task completion probability, which improves the fairness of determining the overall task completion probability.
[0431] The overall task completion probability can also be determined by calculating the weighted average of the task completion probabilities of multiple tasks of interest. In this embodiment, weights must first be set for different tasks of interest. For example, in the above example, the weight set for task A is 0.6, and the weight set for task B is 0.4. Then, the overall task completion probability of content A to be delivered is 0.45*0.6+0.56*0.4=0.494; the overall task completion probability of content B to be delivered is 0.39*0.6+0.78*0.4=0.546; the overall task completion probability of content C to be delivered is 0.69*0.6+0.44*0.4=0.59. Therefore, based on the overall task completion probability, the result of sorting the content to be delivered is: content C to be delivered, content B to be delivered, and content A to be delivered.
[0432] The advantage of calculating the overall task completion probability by weighted average is that the weights of the task completion probabilities of different focus tasks can be adjusted according to the actual application situation, which improves the flexibility of calculating the overall task completion probability.
[0433] The advantage of the embodiment of steps 2310-2320 is that the delivery order of the content to be delivered is determined by obtaining multiple focus tasks of the target object, so that the final delivery order obtained does not only depend on the task completion probability of one task, thereby improving the accuracy of determining the delivery order of the content to be delivered.
[0434] Training process of the first number of expert models, gated nodes, and prediction models in the disclosed embodiment
[0435] In order to train the first number of expert models, gating nodes, and prediction models in the embodiment of the present disclosure, a training sample set is constructed. The training sample set includes multiple training samples. Each training sample contains multiple basic features for delivery and multiple training labels for multiple tasks. The training label indicates whether the task is completed for the training sample. For example, the multiple training labels in a training sample are {Task A: 1, Task B; 0, Task C: 0}, where a training label of 1 represents that the task is completed, and a training label of 0 represents that the task is not completed. That is to say, in this example, Task A is completed, and Task B and Task C are not completed.
[0436] A first number of expert models, gating nodes, and prediction models are trained based on the training sample set.
[0437] In one embodiment, Fig.24 As shown, the first number of expert models, gating nodes, and prediction models are jointly trained through the following process:
[0438] Step 2410: for a single task, based on the training labels and the task completion probability of the task, determine the error subfunction of the task;
[0439] Step 2420: Determine the error function of the training sample based on the error sub-function;
[0440] Step 2430: Based on the error function, jointly train a first number of expert models, gating nodes, and prediction models.
[0441] In this embodiment, first, basic delivery features in a training sample are obtained, and steps 310 - 350 are performed in sequence, and finally task completion probabilities of multiple tasks are obtained by the prediction model.
[0442] In step 2410, the error subfunction is calculated for a single task among the multiple tasks obtained by the prediction model. The error subfunction refers to the distance between the task completion probability and the task label for a single task. The closer the distance, the smaller the error and the better the prediction effect. First, determine the task completion probability of a single task; second, obtain the training label for the task in the training sample. If the training label indicates that the task is completed, then the training label is 1, otherwise, the training label is 0; finally, based on the task completion probability of a single task and the training label of the task, determine the error subfunction of the task.
[0443] The calculation process of determining the error sub-function is shown in formula 10:
[0444] L t =-q t ·log(p t )-(1-q t )·log(1-p t )(Formula 10).
[0445] In formula 10, q t represents the training label (0 or 1), p t represents the task completion probability of the task. For example, if the training label of task A is 1 and the task completion probability is 0.5, then the error subfunction L of the task is t =-1·log(0.5)=0.3.
[0446] In step 2420, after determining the error subfunction of a single task, the error function of the entire training sample is determined based on the error subfunctions of all tasks. The error function refers to the gap between the task completion probability of all tasks and the corresponding training labels. The smaller the gap, the better the prediction result of the model.
[0447] In one embodiment, determining the error function of the training sample based on the error sub-function includes: summing the error sub-functions of multiple tasks to obtain the error function.
[0448] In this embodiment, the calculation process of determining the error function is shown in Formula 11:
[0449]
[0450] In Formula 11, T represents the number of tasks, L t Represents the error subfunction of a single task. For example, the training sample includes three tasks, among which the error subfunction of task A is 0.3; the error subfunction of task B is 0.45; and the error subfunction of task C is 0.27. The error function of the training sample is 0.3+0.45+0.27=1.02.
[0451] The advantage of determining the error function by summing up the error sub-functions is that the error sub-function corresponding to each task contributes equally to determining the error function of the training sample, thereby improving the fairness of multiple tasks when training the model.
[0452] In another embodiment, Fig.25 As shown, based on the error sub-function, the error function of the training sample is determined, including:
[0453] Step 2510: Obtain task weights of multiple tasks;
[0454] Step 2520: Based on the task weights, perform weighted summation on the error sub-functions of multiple tasks to obtain an error function.
[0455] In this embodiment, the calculation process of determining the error function is shown in Formula 12:
[0456]
[0457] In Formula 12, T represents the number of tasks, W t represents the task weight of a single task, L t Represents the error subfunction of a single task. The task weights of multiple tasks can be determined according to the actual application scenario. For example, the training sample includes three tasks, the task weight of task A is 0.5, the task weight of task B is 0.3, and the task weight of task C is 0.2. The error subfunction of task A is 0.3; the error subfunction of task B is 0.45; and the error subfunction of task C is 0.27. Then, the error function of the training sample is 0.5*0.3+0.3*0.45+0.2*0.27=0.339.
[0458] In this embodiment, the advantage of weighting the error sub-functions and determining the error function is that different weights can be set for different tasks according to actual application scenarios, thereby improving the flexibility of the training model.
[0459] In step 2430, after the error function is determined, the first expert network, the second expert network, and the prediction model are trained based on the error function. The purpose of the training is to reduce the error function, and a reduced error function represents a more accurate prediction result. Therefore, the training process is to update the parameters in the first expert network, the second expert network, and the prediction model so that the updated model will reduce the error function generated in the end.
[0460] The first number of expert models, gating nodes, and prediction models may be trained based on the error function using a gradient descent method so that the parameters of the first number of expert models, gating nodes, and prediction models may be updated in the direction of the gradient descent of the error function.
[0461] The advantage of the embodiment of steps 2510-2530 is that the error function of the entire sample is determined by the error sub-function of a single task, so that the prediction result of each task has an impact on the training model, thereby improving the comprehensiveness and accuracy of the training model.
[0462] Detailed diagram of the implementation of the delivery processing method of the present disclosure embodiment
[0463] Refer to the following Fig.26 , the implementation details of the delivery processing method of the embodiment of the present disclosure are explained in detail and by way of example.
[0464] In step 2601, object features of a target object and scene features of a current scene in which the target object is located are obtained.
[0465] In step 2602, a filtered content set is obtained from the to-be-filtered content set, and the features of each filtered content in the filtered content set are used in turn as features of the to-be-delivered content.
[0466] In step 2603, the object features, the scene features, and the features of the content to be delivered are determined as a plurality of delivery basic features, and the determined plurality of basic features are vectorized to obtain a plurality of first feature vectors.
[0467] In step 2604, a plurality of first feature vectors are concatenated into a first concatenated feature vector.
[0468] In step 2605, the first cascade feature vector is input into a first number of expert models.
[0469] In step 2606, a weight vector is generated based on the first concatenated eigenvector and the first matrix.
[0470] In step 2607, the first second number of elements in the weight vector whose values are sorted from large to small are determined, and the expert models corresponding to the first two number of elements are used as the selected second number of expert models.
[0471] In step 2608, the weight vector is normalized to obtain a first weight.
[0472] In step 2609, the kernel parameter matrix and the bias vector corresponding to the second number of first weights are obtained.
[0473] In step 2610, a weighted sum is determined based on the second number of first weights, the kernel parameter matrix corresponding to the first weights, and the bias vector corresponding to the first weights to obtain a second eigenvector corresponding to the task.
[0474] In step 2611, the second feature vectors corresponding to the multiple tasks are input into the prediction models corresponding to the multiple tasks to obtain the predicted task completion probabilities of the target object for the content to be delivered with respect to the multiple tasks.
[0475] In step 2612, the contents to be delivered are sorted based on the task completion probabilities of the multiple tasks, and the contents to be delivered are delivered to the target objects according to the sorting.
[0476] Description of the apparatus and device of the present disclosure
[0477] It is to be understood that, although the steps in the above-mentioned flowcharts are sequentially displayed according to the characterization of arrows, these steps are not necessarily executed in sequence according to the order of arrow characterization. Unless there is a clear description in the present embodiment, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the above-mentioned flowcharts can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of the steps or stages in other steps.
[0478] It should be noted that in each specific implementation of the present application, when it comes to the need to perform relevant processing based on data related to the characteristics of the task content, such as task content attribute information or attribute information sets, the permission or consent of the task content will be obtained first, and the collection, use and processing of these data will comply with the relevant laws, regulations and standards of the relevant countries and regions. In addition, when the embodiment of the present application needs to obtain task content attribute information, it will obtain a separate permission or separate consent for the task content through a pop-up window or jump to a confirmation page. After clearly obtaining a separate permission or separate consent for the task content, the necessary task content-related data used to enable the normal operation of the embodiment of the present application will be obtained.
[0479] Fig. 27The structure diagram of the delivery processing device 2700 provided in the embodiment of the present disclosure. The delivery processing device 2700 includes:
[0480] A vectorization unit 2710 is used to vectorize a plurality of delivery basic features to obtain a plurality of first feature vectors, wherein the delivery basic features are used to predict the task completion probability of a target object with respect to a plurality of tasks for the delivery content;
[0481] A cascading unit 2720, configured to cascade the plurality of first feature vectors into a first cascade feature vector;
[0482] A first input unit 2730, configured to input a first cascade feature vector into a first number of expert models;
[0483] A first selection unit 2740 is used to select a second number of expert models from the first number of expert models through the gating nodes corresponding to each task, and perform weighted sum on outputs of the selected second number of expert models to obtain a second feature vector corresponding to the task, wherein the second number is smaller than the first number;
[0484] The second input unit 2750 is used to input the second feature vectors corresponding to the multiple tasks into the prediction models corresponding to the multiple tasks to obtain the predicted task completion probabilities of the target object for the content to be delivered with respect to the multiple tasks.
[0485] Optionally, the first selection unit 2740 is specifically configured to:
[0486] Generate a weight vector based on the first cascade feature vector and the first matrix, wherein the number of columns of the first matrix is the first number, the number of rows is the dimension of the first cascade feature vector, and the weight vector indicates the weights of the first number of expert models;
[0487] Based on the weight vector, a second number of expert models are selected from the first number of expert models.
[0488] Optionally, the first selection unit 2740 is specifically configured to:
[0489] Multiply the first matrix by the first concatenated eigenvector to obtain the original weight vector;
[0490] Multiplying the second matrix by the first concatenated eigenvector to obtain a weight adjustment basis vector, wherein the number of columns of the second matrix is the first number, and the number of rows is the dimension of the first concatenated eigenvector;
[0491] Generate a weight adjustment value vector based on the random perturbation vector and the weight adjustment basis vector;
[0492] A weight vector is generated based on the original weight vector and the weight adjustment value vector.
[0493] Optionally, the first selection unit 2740 is specifically configured to:
[0494] Get a random number set, the random numbers in the random number set conform to the normal distribution;
[0495] Randomly extracting a first number of random numbers from a random number set to form a random disturbance vector;
[0496] Multiply the random perturbation vector by the weight adjustment basis vector to obtain the weight adjustment value vector.
[0497] Optionally, the first selection unit 2740 is specifically configured to:
[0498] Determine the second first number of elements in the weight vector whose values are sorted from largest to smallest;
[0499] The expert models corresponding to the first second number of elements are used as the selected second number of expert models.
[0500] Optionally, the first selection unit 2740 is specifically configured to:
[0501] Modify the other elements in the weight vector that are not the first second number of elements sorted from largest to smallest to negative infinity;
[0502] Performing normalized exponential processing on the weight vector to obtain first weights of a second number of expert models;
[0503] The outputs of the selected second number of expert models are weighted summed according to the first weight.
[0504] Optionally, the expert model includes a kernel parameter matrix and a bias vector, the number of columns of the kernel parameter matrix is the dimension of the output of the expert model, the number of rows is the dimension of the first cascade feature vector, and the dimension of the bias vector is equal to the dimension of the output of the expert model;
[0505] The first selection unit 2740 is specifically used for:
[0506] Obtaining a three-dimensional kernel parameter tensor, wherein the three-dimensional kernel parameter tensor represents a kernel parameter matrix of a first number of expert models;
[0507] Obtaining a bias vector matrix, the bias vector matrix representing bias vectors of a first number of expert models;
[0508] Obtaining a kernel parameter matrix corresponding to a second number of first weights from the three-dimensional kernel parameter tensor;
[0509] Obtaining bias vectors corresponding to a second number of first weights from the bias vector matrix;
[0510] A weighted sum is determined based on the second number of first weights, a kernel parameter matrix corresponding to the second number of first weights, and a bias vector corresponding to the second number of first weights.
[0511] Optionally, the first selection unit 2740 is specifically configured to:
[0512] Multiplying the kernel parameter matrix corresponding to each expert model in the second number of expert models by the first cascade feature vector of the input expert model to obtain an intermediate vector;
[0513] Add the intermediate vector to the bias vector corresponding to the expert model to obtain the result vector;
[0514] Apply the activation function to the result vector to get the output of the expert model;
[0515] A weighted sum is obtained based on outputs of the second number of expert models and first weights corresponding to the expert models.
[0516] Optionally, the second number is determined by:
[0517] Get the task type of the task;
[0518] Obtain the prediction accuracy requirements of the task;
[0519] The second number is determined based on the task type and the prediction accuracy requirement.
[0520] Optionally, based on the task type and the prediction accuracy requirement, determining the second number includes:
[0521] Based on the task type, determining a first score;
[0522] determining a second score based on the prediction accuracy requirement;
[0523] determining a total score based on the first score and the second score;
[0524] Based on the total score, a second number is determined.
[0525] Optionally, the first selection unit 2740 is specifically configured to:
[0526] Get the task type of the task;
[0527] From the first number of expert models, a third number of expert models matching the task type;
[0528] A second number of expert models is selected from the third number of expert models, wherein the first number is greater than the third number, and the third number is greater than the second number.
[0529] Optionally, the vectorization unit 2710 is specifically used for:
[0530] Obtain object features of a target object;
[0531] Obtain scene features of the current scene where the target object is located;
[0532] Filter out a filtered content set from the to-be-filtered content set, and use the features of each filtered content in the filtered content set in turn as features of the to-be-delivered content;
[0533] Determine the object features, the scene features, and the features of the content to be delivered as a plurality of delivery basic features;
[0534] The determined multiple delivery basic feature vectors are quantized to obtain multiple first feature vectors.
[0535] Optionally, the vectorization unit 2710 is specifically used for:
[0536] If the basic feature is a character feature, the character feature is input into the embedding layer to obtain the first feature vector corresponding to the character feature;
[0537] If the basic feature is a numerical feature, the numerical feature is numerically mapped to obtain the first feature vector corresponding to the numerical feature;
[0538] If the basic feature is a variable-length sequence feature, the variable-length sequence feature is pooled to obtain a first feature vector corresponding to the variable-length sequence feature.
[0539] Optionally, the delivery processing device 2700 further includes:
[0540] A first acquisition unit (not shown), configured to acquire a focus task of a target object;
[0541] A second acquisition unit (not shown), configured to acquire a task completion probability for a focused task from task completion probabilities for a plurality of tasks;
[0542] A sorting unit (not shown), configured to sort the content to be delivered based on the task completion probability of the concerned task;
[0543] The delivery unit (not shown) is used to deliver the content to be delivered to the target object according to the ranking.
[0544] Optionally, the multiple delivery basic features are selected from a training sample in the training sample set, the training sample has multiple training labels for multiple tasks, and the training label indicates whether the task is completed for the training sample;
[0545] The first number of expert models, gating nodes, and prediction models are jointly trained through the following process:
[0546] For a single task, determine the error subfunction of the task based on the training labels and the task completion probability of the task;
[0547] Based on the error sub-function, determining the error function of the training sample;
[0548] Based on the error function, a first number of expert models, gating nodes, and a prediction model are jointly trained.
[0549] Optionally, determining the error function of the training sample based on the error sub-function includes:
[0550] Get the task weights of multiple tasks;
[0551] Based on the task weights, the error sub-functions of multiple tasks are weighted and summed to obtain the error function.
[0552] Reference Fig.28 , Fig.28 The structural block diagram of the terminal part for implementing the delivery processing method of the embodiment of the present disclosure is as follows: the terminal includes: a radio frequency (RF) circuit 2810, a memory 2815, an input unit 2830, a display unit 2840, a sensor 2850, an audio circuit 2860, a wireless fidelity (WiFi) module 2870, a processor 2880, and a power supply 2890. Those skilled in the art can understand that Fig.28 The terminal structure shown does not constitute a limitation on the mobile phone or computer, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.
[0553] The RF circuit 2810 may be used for receiving and sending signals during information transmission or calls. In particular, after receiving the downlink information from the base station, it is sent to the processor 2880 for processing; in addition, the designed uplink data is sent to the base station.
[0554] The memory 2815 may be used to store software programs and modules. The processor 2880 executes various functional applications and data processing of the content terminal by running the software programs and modules stored in the memory 2815 .
[0555] The input unit 2830 may be used to receive input digital or character information and generate key signal input related to the setting and function control of the content terminal. Specifically, the input unit 2830 may include a touch panel 2831 and other input devices 2832.
[0556] The display unit 2840 may be used to display input information or provided information and various menus of the content terminal. The display unit 2840 may include a display panel 2841.
[0557] The audio circuit 2860, the speaker 2861, and the microphone 2862 may provide an audio interface.
[0558] In this embodiment, the processor 2880 included in the terminal can execute the delivery processing method of the previous embodiment.
[0559] The terminals of the embodiments of the present disclosure include but are not limited to mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. The embodiments of the present invention can be applied to various scenarios, including but not limited to content recommendation, data screening, etc.
[0560] Fig.29 A block diagram of the structure of a portion of the server 110 for implementing the delivery processing method of the embodiment of the present disclosure. The server 110 may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPUs) 2922 (for example, one or more processors) and memories 2932, and one or more storage media 2930 (for example, one or more mass storage devices) storing application programs 2942 or data 2944. Among them, the memory 2932 and the storage medium 2930 can be short-term storage or permanent storage. The program stored in the storage medium 2930 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the server 110. Furthermore, the central processing unit 2922 can be configured to communicate with the storage medium 2930 and execute a series of instruction operations in the storage medium 2930 on the server 110.
[0561] The server 110 may also include one or more power supplies 2926, one or more wired or wireless network interfaces 2950, one or more input and output interfaces 2958, and / or one or more operating systems 2941, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.
[0562] The central processor 2922 in the server 110 can be used to execute the delivery processing method of the embodiment of the present disclosure.
[0563] The embodiments of the present disclosure also provide a computer-readable storage medium, which is used to store program codes, and the program codes are used to execute the delivery processing methods of the aforementioned embodiments.
[0564] The embodiment of the present disclosure also provides a computer program product, which includes a computer program. A processor of a computer device reads and executes the computer program, so that the computer device performs the above-mentioned delivery process.
[0565] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present disclosure and the above-mentioned drawings are used to distinguish similar contents, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "comprises" and "comprising" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0566] It should be understood that in the present disclosure, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated content, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the associated content before and after is in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.
[0567] It should be understood that in the description of the embodiments of the present disclosure, the meaning of multiple (or multiple items) is more than two, greater than, less than, exceed, etc. are understood to not include the number, and above, below, within, etc. are understood to include the number.
[0568] In the several embodiments provided in the present disclosure, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.
[0569] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0570] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0571] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, server 110, or network device, etc.) to perform all or part of the steps of each embodiment method of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), disk or optical disk and other media that can store program codes.
[0572] It should also be understood that the various implementations provided in the embodiments of the present disclosure can be combined arbitrarily to achieve different technical effects.
[0573] The above is a specific description of the implementation methods of the present disclosure, but the present disclosure is not limited to the above implementation methods. Technical personnel familiar with the art can also make various equivalent modifications or substitutions without violating the spirit of the present disclosure. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present disclosure.
Claims
1. A delivery processing method, It is characterized in that include: Quantizing a plurality of delivery basic features to obtain a plurality of first feature vectors, wherein the delivery basic features are used to predict task completion probabilities of a target object with respect to a plurality of tasks for the delivered content; Cascading a plurality of the first feature vectors into a first cascade feature vector; Inputting the first cascade feature vector into a first number of expert models; Selecting a second number of the expert models from the first number of the expert models through the gating nodes corresponding to each of the tasks, and performing weighted summation on outputs of the selected second number of the expert models to obtain a second feature vector corresponding to the task, wherein the second number is smaller than the first number; The second feature vectors corresponding to the multiple tasks are input into the prediction models corresponding to the multiple tasks to obtain the predicted task completion probabilities of the target object for the to-be-delivered content with respect to the multiple tasks.
2. The delivery processing method according to claim 1, It is characterized in that The step of selecting a second number of the expert models from the first number of the expert models comprises: Generate a weight vector based on the first cascade feature vector and the first matrix, wherein the number of columns of the first matrix is the first number, the number of rows is the dimension of the first cascade feature vector, and the weight vector indicates the weights of the first number of the expert models; A second number of the expert models are selected from the first number of the expert models based on the weight vector.
3. The delivery processing method according to claim 2, It is characterized in that The step of generating a weight vector based on the first cascaded feature vector and the first matrix includes: Multiplying the first matrix by the first concatenated feature vector to obtain an original weight vector; Multiplying a second matrix by the first concatenated feature vector to obtain a weight adjustment basis vector, wherein the number of columns of the second matrix is the first number, and the number of rows is the dimension of the first concatenated feature vector; Generate a weight adjustment value vector based on the random disturbance vector and the weight adjustment basis vector; The weight vector is generated based on the original weight vector and the weight adjustment value vector.
4. The delivery processing method according to claim 3, It is characterized in that The step of generating a weight adjustment value vector based on the random disturbance vector and the weight adjustment basis vector comprises: Obtaining a random number set, wherein the random numbers in the random number set conform to a normal distribution; Randomly extracting a first number of the random numbers from the random number set to form the random perturbation vector; The random disturbance vector is point-multiplied with the weight adjustment basis vector to obtain the weight adjustment value vector.
5. The delivery processing method according to claim 2, It is characterized in that The selecting a second number of the expert models from the first number of the expert models based on the weight vector comprises: Determine the first second number of elements in the weight vector whose values are sorted from largest to smallest; The expert models corresponding to the first second number of elements are used as the selected second number of expert models.
6. The delivery processing method according to claim 5, It is characterized in that The step of performing a weighted sum on the outputs of the selected second number of expert models comprises: Modify the other elements in the weight vector that are not the first second number of elements in the descending order to negative infinity; Performing normalized exponential processing on the weight vector to obtain first weights of a second number of the expert models; The outputs of the second number of the expert models selected are weighted summed according to the first weights.
7. The delivery processing method according to claim 6, It is characterized in that The expert model includes a kernel parameter matrix and a bias vector, the number of columns of the kernel parameter matrix is the dimension of the output of the expert model, the number of rows is the dimension of the first cascade feature vector, and the dimension of the bias vector is equal to the dimension of the output of the expert model; The step of performing a weighted sum on the outputs of the second number of the expert models selected according to the first weight comprises: Obtaining a three-dimensional kernel parameter tensor, wherein the three-dimensional kernel parameter tensor represents the kernel parameter matrices of a first number of the expert models; Obtaining a bias vector matrix, wherein the bias vector matrix represents the bias vectors of a first number of the expert models; Obtaining, from the three-dimensional kernel parameter tensor, the kernel parameter matrix corresponding to a second number of the first weights; Obtaining, from the bias vector matrix, the bias vectors corresponding to a second number of the first weights; The weighted sum is determined based on a second number of the first weights, the kernel parameter matrices corresponding to the second number of the first weights, and the bias vectors corresponding to the second number of the first weights.
8. The delivery processing method according to claim 7, It is characterized in that The determining the weighted sum based on the second number of the first weights, the kernel parameter matrix corresponding to the second number of the first weights, and the bias vector corresponding to the second number of the first weights comprises: Multiplying the kernel parameter matrix corresponding to each of the expert models in the second number of the expert models by the first cascade feature vector input into the expert model to obtain an intermediate vector; Adding the intermediate vector to the bias vector corresponding to the expert model to obtain a result vector; Applying an activation function to the result vector to obtain an output of the expert model; The weighted sum is obtained based on outputs of a second number of the expert models and the first weights corresponding to the expert models.
9. The delivery processing method according to claim 1, It is characterized in that The second number is determined by: Get the task type of the task; Obtaining the prediction accuracy requirement of the task; The second number is determined based on the task type and the prediction accuracy requirement.
10. The delivery processing method according to claim 9, It is characterized in that The determining the second number based on the task type and the prediction accuracy requirement includes: Based on the task type, determining a first score; determining a second score based on the prediction accuracy requirement; determining a total score based on the first score and the second score; Based on the total score, the second number is determined.
11. The delivery processing method according to claim 1, It is characterized in that The step of selecting a second number of the expert models from the first number of the expert models comprises: Get the task type of the task; From the first number of the expert models, a third number of the expert models matching the task type; A second number of the expert models is selected from the third number of the expert models, wherein the first number is greater than the third number, and the third number is greater than the second number.
12. The delivery processing method according to claim 1, It is characterized in that The step of vectorizing the plurality of delivery basic features to obtain a plurality of first feature vectors includes: Acquire object features of the target object; Acquire scene features of the current scene where the target object is located; Filtering out a filtered content set from the to-be-filtered content set, and taking the features of each filtered content in the filtered content set as features of the to-be-delivered content in turn; Determining the object feature, the scene feature, and the feature of the content to be delivered as a plurality of delivery basic features; The determined multiple delivery basic feature vectors are quantized to obtain multiple first feature vectors.
13. The delivery processing method according to claim 12, It is characterized in that The step of vectorizing the determined multiple delivery basic features to obtain multiple first feature vectors includes: If the delivery basic feature is a character feature, input the character feature into an embedding layer to obtain the first feature vector corresponding to the character feature; If the delivery basic feature is a numerical feature, numerical mapping is performed on the numerical feature to obtain the first feature vector corresponding to the numerical feature; If the delivery basic feature is a variable-length sequence feature, the variable-length sequence feature is pooled to obtain the first feature vector corresponding to the variable-length sequence feature.
14. The delivery processing method according to claim 1, It is characterized in that After inputting the second feature vectors corresponding to the plurality of tasks into the prediction models corresponding to the plurality of tasks to obtain the predicted task completion probabilities of the target objects for the to-be-delivered content with respect to the plurality of tasks, the delivery processing method further includes: Obtaining the focus task of the target object; Acquire the task completion probability of the task of interest from the task completion probabilities of the plurality of tasks; sorting the to-be-delivered contents based on the task completion probabilities of the concerned tasks; The content to be delivered is delivered to the target object according to the order.
15. The delivery processing method according to claim 1, It is characterized in that The plurality of delivery basic features are selected from a training sample in a training sample set, the training sample having a plurality of training labels for the plurality of tasks, the training labels indicating whether the tasks are completed for the training sample; The first number of the expert models, the gating nodes, and the prediction model are jointly trained by the following process: For a single task, determining an error subfunction of the task based on the training labels and the task completion probability of the task; Based on the error sub-function, determining the error function of the training sample; Based on the error function, a first number of the expert models, the gating nodes, and the prediction model are jointly trained.
16. The delivery processing method according to claim 15, It is characterized in that The step of determining the error function of the training sample based on the error sub-function includes: Obtaining task weights of a plurality of the tasks; Based on the task weights, a weighted sum is performed on the error sub-functions of a plurality of the tasks to obtain the error function.
17. A delivery processing device, It is characterized in that include: A vectorization unit, used for vectorizing a plurality of delivery basic features to obtain a plurality of first feature vectors, wherein the delivery basic features are used to predict task completion probabilities of a target object with respect to a plurality of tasks for the delivered content; A cascading unit, used for cascading a plurality of the first feature vectors into a first cascade feature vector; A first input unit, used for inputting the first cascade feature vector into a first number of expert models; a first selection unit, configured to select a second number of the expert models from the first number of the expert models through a gating node corresponding to each of the tasks, and perform a weighted sum on outputs of the selected second number of the expert models to obtain a second feature vector corresponding to the task, wherein the second number is smaller than the first number; The second input unit is used to input the second feature vectors corresponding to the multiple tasks into the prediction model corresponding to the multiple tasks to obtain the predicted task completion probability of the target object for the content to be delivered with respect to the multiple tasks.
18. An electronic device comprising a memory and a processor, wherein the memory stores a computer program. It is characterized in that When the processor executes the computer program, the delivery processing method according to any one of claims 1 to 16 is implemented.
19. A computer-readable storage medium storing a computer program. It is characterized in that When the computer program is executed by a processor, the delivery processing method according to any one of claims 1 to 16 is implemented.
20. A computer program product, comprising a computer program, wherein the computer program is read and executed by a processor of a computer device, so that the computer device executes the delivery processing method according to any one of claims 1 to 16.
Citation Information
Cited By
AI dynamic weight analysis-based short video projection flow accurate regulation and control system
CN120915985A
