Attribute prediction model training methods, attribute prediction methods and electronic devices

Abstract

This disclosure provides a method for training an attribute prediction model, an attribute prediction method, a model, an apparatus, a medium, and an electronic device, primarily relating to the field of transfer learning technology in artificial intelligence. The method includes acquiring source domain data, which includes behavioral sequences of sample objects in the source domain; training an initial neural network using the source domain data to obtain a source domain behavioral representation model; freezing the parameters of the source domain behavioral representation model and inserting fine-tuned network layers into the source domain behavioral representation model; acquiring target domain data, which includes behavioral sequences of sample objects in the source domain and attribute labels in the target domain; and training the source domain behavioral representation model with fine-tuned network layers using the target domain data to obtain an attribute prediction model for the target domain. This method can effectively transfer behavioral features from the source domain to the target domain, reducing the number of parameter adjustments and improving model training performance.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more specifically, to an attribute prediction model training method, an attribute prediction method, an attribute prediction model, an attribute prediction model training device, an attribute prediction device, a computer-readable medium, and an electronic device. Background Technology

[0002] With the development of computer and internet technology, viewing content or browsing information through online platforms has become an extremely important part of people's daily lives. For example, in areas such as short videos, news feeds, or image streams, users can typically complete the reading or viewing of content within tens of seconds. Therefore, within just a few hours or a day, hundreds of millions of users can generate tens of billions of user click / view behavior records. Based on this data, user preferences can be inferred, thereby continuing to produce and push potentially interesting short videos, news, images, and other content to users.

[0003] However, in areas such as ad feeds, the vast majority of users are new users with very few clicks or no clicks at all. This scenario is often called a cold start scenario, and the relevant users are called cold users. Due to the lack of user data, it is difficult to accurately push content to users in a cold start scenario. Therefore, how to predict the interests and other attributes of cold users in a cold start scenario is a problem that urgently needs to be solved. Summary of the Invention

[0004] The purpose of this disclosure is to provide an attribute prediction model training method, attribute prediction method, attribute prediction model, attribute prediction model training device, attribute prediction device, computer-readable medium, and electronic device, thereby overcoming, to at least a certain extent, the technical problems of difficulty in attribute data prediction existing in related technologies.

[0005] Other features and advantages of this disclosure will become apparent from the following detailed description, or may be learned in part from practice of this disclosure.

[0006] According to one aspect of the present disclosure, a method for training an attribute prediction model is provided. The method includes: acquiring source domain data, the source domain data including a sequence of behavior of sample objects in the source domain; training an initial neural network using the source domain data to obtain a source domain behavior representation model; freezing the parameters of the source domain behavior representation model and inserting a fine-tuned network layer into the source domain behavior representation model; acquiring target domain data, the target domain data including a sequence of behavior of sample objects in the source domain and attribute labels in the target domain; and training the source domain behavior representation model with the fine-tuned network layer using the target domain data to obtain an attribute prediction model for the target domain.

[0007] According to one aspect of the present disclosure, an attribute prediction model training apparatus is provided. The apparatus includes: a source domain data acquisition module configured to acquire source domain data, the source domain data including a sequence of behavior of sample objects in the source domain; a model pre-training module configured to train an initial neural network using the source domain data to obtain a source domain behavior representation model; a model adjustment module configured to freeze the parameters of the source domain behavior representation model and insert a fine-tuning network layer into the source domain behavior representation model; a target domain data acquisition module configured to acquire target domain data, the target domain data including a sequence of behavior of sample objects in the source domain and attribute labels in the target domain; and a model fine-tuning module configured to train the source domain behavior representation model with the fine-tuning network layer using the target domain data to obtain an attribute prediction model for the target domain.

[0008] In some embodiments of this disclosure, based on the above technical solutions, the model pre-training module includes: a node occlusion module, configured to select multiple target behavior nodes in the behavior sequence of the source domain data and replace the target behavior nodes with occlusion behavior nodes; a behavior prediction module, configured to input the behavior sequence with occlusion behavior nodes into an initial neural network to obtain behavior prediction information corresponding to the occlusion behavior nodes; a behavior error determination module, configured to determine the behavior prediction error based on the behavior prediction information and the target behavior nodes; and a pre-training parameter update module, configured to update the network parameters of the initial neural network using the behavior prediction error to obtain a source domain behavior representation model.

[0009] In some embodiments of this disclosure, based on the above technical solutions, the occlusion behavior node includes a specified behavior node and a random behavior node; the node occlusion module includes: a node classification module configured to determine a node classification ratio and classify the target behavior node into a specified node set, a random node set, and an original node set according to the node classification ratio; a specified replacement module configured to determine a specified behavior node and replace the target behavior node in the specified node set with the specified behavior node; and a random replacement module configured to determine a random behavior node and replace the target behavior node in the random node set with the random behavior node.

[0010] In some embodiments of this disclosure, based on the above technical solutions, the behavior prediction module includes: a pre-training input module configured to input a behavior sequence with occluded behavior nodes into an initial neural network; a pre-training mapping module configured to map the behavior sequence through an embedding layer in the initial neural network to obtain an embedding vector for each behavior node in the behavior sequence; and a pre-training convolution module configured to convolve the embedding vector through a convolution layer in the initial neural network to obtain behavior prediction information corresponding to the occluded behavior nodes.

[0011] In some embodiments of this disclosure, based on the above technical solutions, the pre-trained convolution module includes: a pre-trained convolution processing module configured to perform convolution processing on the embedding vector through the convolutional layer in the initial neural network to obtain convolution processing information; a pre-trained mapping processing module configured to perform mapping processing on the embedding vector and the convolution processing information through the residual connection branch in the initial neural network to obtain residual mapping information; and a pre-trained information overlay module configured to obtain behavior prediction information corresponding to the occlusion behavior node based on the convolution processing information and the residual mapping information.

[0012] In some embodiments of this disclosure, based on the above technical solutions, the convolution processing module includes: a dilated convolution module, configured to perform dilated convolution processing on the embedded vector through multiple dilated convolution layers with different dilation rates in the initial neural network to obtain convolution processing information.

[0013] In some embodiments of this disclosure, based on the above technical solutions, the model fine-tuning module includes: an attribute prediction module configured to input the behavior sequence in the target domain data into a source domain behavior representation model with a fine-tuned network layer to obtain attribute prediction information corresponding to the behavior sequence; an attribute error determination module configured to determine the attribute label associated with the behavior sequence and determine the attribute prediction error based on the attribute label and the attribute prediction information; and a fine-tuning parameter update module configured to update the parameters of the fine-tuned network layer using the attribute prediction error to obtain an attribute prediction model for the target domain.

[0014] In some embodiments of this disclosure, based on the above technical solutions, the attribute prediction module includes: a fine-tuning input module configured to input the behavior sequence in the target domain data into a source domain behavior representation model with a fine-tuning network layer; a fine-tuning convolution module configured to perform convolution processing on the behavior sequence through the convolutional layer in the source domain behavior representation model to obtain behavior representation information; and a fine-tuning mapping module configured to perform mapping processing on the behavior representation information through the fine-tuning network layer to obtain attribute prediction information corresponding to the behavior sequence.

[0015] In some embodiments of this disclosure, based on the above technical solutions, the fine-tuning network layer includes a convolutional branch and a residual branch; the fine-tuning mapping module includes: a fine-tuning convolutional processing module, configured to perform convolutional processing on the behavioral representation information through the convolutional branch to obtain convolutional prediction information; a fine-tuning mapping processing module, configured to perform mapping processing on the behavioral representation information through the residual branch to obtain residual prediction information; and a fine-tuning information overlay module, configured to determine attribute prediction information corresponding to the behavioral sequence based on the convolutional prediction information and the residual prediction information.

[0016] In some embodiments of this disclosure, based on the above technical solutions, the convolutional branch includes a dimensionality reduction convolutional layer, an activation layer, and an up-dimensional convolutional layer connected in sequence; the fine-tuning convolutional processing module includes: a dimensionality reduction convolutional module configured to perform convolution processing on the behavioral representation information through the dimensionality reduction convolutional layer to obtain low-dimensional prediction information with a dimension lower than that of the behavioral representation information; an activation module configured to perform mapping processing on the low-dimensional prediction information through the activation layer to obtain activation prediction information with nonlinear features; and an up-dimensional convolutional module configured to perform convolution processing on the activation prediction information through the up-dimensional convolutional layer to obtain convolutional prediction information with a dimension equal to that of the behavioral representation information.

[0017] In some embodiments of this disclosure, based on the above technical solutions, the attribute error determination module includes: an attribute label mapping module, configured to map the attribute labels according to the attribute information of the target domain to obtain label representation information; and a prediction similarity determination module, configured to determine the similarity between the label representation information and the attribute prediction information, and determine the similarity as the attribute prediction error.

[0018] In some embodiments of this disclosure, based on the above technical solutions, the model adjustment module includes: a parameter freezing module, configured to freeze the parameters of the source domain behavior representation model; and a fine-tuning layer insertion module, configured to insert multiple fine-tuning network layers at equal intervals between multiple convolutional layers in the source domain behavior representation model.

[0019] In some embodiments of this disclosure, based on the above technical solutions, the behavior sequence of the sample object in the source domain is the content click behavior sequence of the sample object on the content in the source domain.

[0020] According to one aspect of the present disclosure, an attribute prediction method is provided, the method comprising: acquiring a behavior sequence of a test object in a source domain, and determining a target domain corresponding to the source domain; inputting the behavior sequence into a pre-trained attribute prediction model for the target domain to obtain attribute prediction information of the test object; wherein the attribute prediction model is trained by the attribute prediction model training method as described in the above technical solutions.

[0021] According to one aspect of the present disclosure, an attribute prediction apparatus is provided, the apparatus comprising: a data acquisition module configured to acquire a behavioral sequence of a test object in a source domain and determine a target domain corresponding to the source domain; and a model prediction module configured to input the behavioral sequence into a pre-trained attribute prediction model for the target domain to obtain attribute prediction information of the test object; wherein the attribute prediction model is trained by the attribute prediction model training method as described in the above technical solutions.

[0022] According to one aspect of the present disclosure, an attribute prediction model is provided, comprising: an input layer for inputting a behavioral sequence of a test object in a source domain; a convolutional network layer for performing convolution processing on the behavioral sequence to obtain behavioral representation information; a fine-tuning network layer located between multiple convolutional network layers for performing mapping processing on the behavioral representation information to obtain attribute prediction information corresponding to the behavioral sequence; and an output layer for outputting attribute labels of the test object in a target domain based on the attribute prediction information.

[0023] In some embodiments of this disclosure, based on the above technical solutions, the fine-tuning network layer includes a convolutional branch and a residual branch, wherein the convolutional branch includes a dimensionality-reducing convolutional layer, an activation layer, and an dimensionality-increasing convolutional layer connected in sequence.

[0024] In some embodiments of this disclosure, based on the above technical solutions, the number of convolution kernels in the dimensionality reduction convolutional layer is less than the number of input channels in the fine-tuning network layer, and the number of convolution kernels in the dimensionality increase convolutional layer is equal to the number of input channels in the fine-tuning network layer.

[0025] In some embodiments of this disclosure, based on the above technical solutions, the activation function of the activation layer is a nonlinear activation function based on Gaussian error linear units.

[0026] In some embodiments of this disclosure, based on the above technical solutions, the convolutional network layer is a dilated convolutional network layer.

[0027] In some embodiments of this disclosure, based on the above technical solutions, the attribute prediction model further includes residual connection branches distributed among multiple convolutional network layers.

[0028] According to one aspect of the present disclosure, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements the methods described in the above technical solutions.

[0029] According to one aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the methods described above by executing the executable instructions.

[0030] In the technical solution provided in this disclosure, a source domain behavior representation model is pre-trained using unsupervised learning with source domain data, and then the parameters of the source domain behavior representation model are fine-tuned by inserting fine-tuning network layers using target domain data. This effectively transfers behavioral features from the source domain to the target domain. By freezing the parameters of the source domain behavior representation model and inserting fine-tuning network layers, the learning results from the pre-training stage can be retained during the fine-tuning stage, while significantly reducing the number of parameter adjustments and improving the training speed and effectiveness of the model.

[0031] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0032] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings:

[0033] Figure 1 A schematic diagram of an exemplary system architecture applying the technical solutions disclosed herein is shown.

[0034] Figure 2 A flowchart illustrating the steps of an attribute prediction model training method in some embodiments of this disclosure is shown.

[0035] Figure 3 The structural framework of the attribute prediction model in some embodiments of this disclosure is illustrated schematically.

[0036] Figure 4 A flowchart illustrating the steps of training a source domain behavior representation model is shown in some embodiments of this disclosure.

[0037] Figure 5 A flowchart illustrating the steps of replacing a shading behavior node is shown schematically in some embodiments of this disclosure.

[0038] Figure 6 The flowchart illustrating the steps of obtaining behavior prediction information based on occlusion behavior nodes in some embodiments of this disclosure is shown in the illustration.

[0039] Figure 7 The flowchart illustrating the steps of convolution processing based on residual connections in some embodiments of this disclosure is shown schematically.

[0040] Figure 8 This schematically illustrates a network architecture for a source domain behavior representation model during the pre-training phase.

[0041] Figure 9 A flowchart illustrating the steps of training an attribute prediction model during the fine-tuning phase in some embodiments of this disclosure is shown schematically.

[0042] Figure 10 A network architecture for an attribute prediction model during the fine-tuning phase is illustrated schematically.

[0043] Figure 11 A flowchart illustrating the steps for obtaining attribute prediction information in some embodiments of this disclosure is shown schematically.

[0044] Figure 12 A flowchart illustrating the steps of determining attribute prediction error based on attribute prediction information in some embodiments of this disclosure is shown.

[0045] Figure 13 The schematic diagram illustrates the model architecture for mapping attribute tags.

[0046] Figure 14 The overall network architecture during the fine-tuning phase is illustrated schematically.

[0047] Figure 15 The network architecture of the fine-tuned network layer in some embodiments of this disclosure is illustrated schematically.

[0048] Figure 16 The flowchart illustrating the steps of obtaining attribute prediction information by fine-tuning network layers in some embodiments of this disclosure is shown in the illustration.

[0049] Figure 17 A flowchart illustrating the steps of obtaining convolutional prediction information through convolutional branches in some embodiments of this disclosure is shown schematically.

[0050] Figure 18 The attribute prediction method in some embodiments of this disclosure is illustrated schematically.

[0051] Figure 19 A schematic block diagram of an attribute prediction model training apparatus according to some embodiments of the present disclosure is shown.

[0052] Figure 20A structural block diagram of an attribute prediction apparatus in some embodiments of the present disclosure is shown schematically.

[0053] Figure 21 A schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present disclosure is shown. Detailed Implementation

[0054] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be more thorough and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art.

[0055] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0056] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0057] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0058] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0059] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0060] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0061] The technical solution disclosed herein mainly relates to machine learning technology in artificial intelligence, specifically to a transfer learning technology.

[0062] The main principle of this disclosed technical solution is that by learning behavioral sequence data in non-cold start scenarios, it can guide content recommendation for cold start users. At the same time, it can also predict unknown attribute information in cold start scenarios, such as gender, age, life status, parenting status, hobbies, etc., relatively accurately based on behavioral sequence data in non-cold start scenarios.

[0063] Specifically, this disclosure provides a transfer learning algorithm based on learning user representations. In the pre-training phase, the algorithm models a neural network sequence model using unsupervised learning based on behavioral sequence data. After the pre-trained model converges, the relevant model parameters are frozen. During downstream task learning, i.e., the fine-tuning phase of the algorithm, this disclosure provides a fine-tuning network layer, which is inserted into the pre-trained model. Supervised learning is then used to adjust the parameters of the fine-tuning network layer to meet the needs of different types of downstream tasks.

[0064] Figure 1 A schematic diagram of an exemplary system architecture applying the technical solutions disclosed herein is shown.

[0065] like Figure 1As shown, system architecture 100 may include client 110, network 120, and server 130. Client 110 may include various terminal devices such as smartphones, tablets, laptops, and desktop computers. Server 130 may include various server devices such as network servers, application servers, and database servers. Network 120 may be a communication medium of various connection types capable of providing a communication link between client 110 and server 130, such as a wired communication link, wireless communication link, etc.

[0066] Depending on the implementation requirements, the system architecture in this embodiment can have any number of clients, networks, and servers. For example, server 130 can be a server group composed of multiple server devices. Furthermore, the attribute prediction model training method and attribute prediction method in this embodiment can be applied to client 110 or server 130, and this disclosure does not impose any special limitations on them.

[0067] Taking the application of the attribute prediction model training method in the server 130 of this embodiment as an example, the client 110 can collect behavioral sequence data in the source domain. For example, it can use a short video viewing program installed on the client 110 to collect the behavioral sequence of users clicking to view short videos. At the same time, the client 110 can also collect attribute data in the target domain. For example, it can collect user behavior information such as clicking on advertisements, or it can collect attribute data such as the user's gender and age. In the source domain, each user usually has relatively rich click behavior data, including browsing history, rating, likes, and forwarding behavior data. However, the relevant users in the target domain are either completely new users with no data or cold users with only a small amount of click behavior data.

[0068] Client 110 sends the collected source and target domain data to server 130 via network 120. During the pre-training phase, server 130 uses the data from the source domain as training samples to pre-train a source domain behavior representation model that characterizes user behavior habits in the source domain. After the source domain behavior representation model is trained, it enters the fine-tuning phase. In the fine-tuning phase, the fine-tuned network layer is first inserted into the trained representation model. Then, the source domain behavior representation model with the inserted fine-tuned network layer is further trained using correlated data from the source and target domains. During the fine-tuning phase, the network parameters of the source domain behavior representation model need to be frozen, and only the relevant parameters of the fine-tuned network layer are adjusted. This ensures that only a small number of network parameters need to be adjusted to achieve the goal of predicting the attribute data of the target domain.

[0069] Generally, the attribute data and interests of the same user in the source and target domains often exhibit significant similarity or correlation. Therefore, by utilizing transfer learning techniques to learn potential user interests and preferences from the source domain, the problem of missing data in the target domain can be effectively addressed, leading to accurate attribute data prediction.

[0070] The attribute prediction model provided in this disclosure can be applied to various application scenarios. For example, using user representations and tag representations finely tuned through transfer learning, it can be used in cold start scenarios of recommendation systems to recommend content to new users, especially suitable for scenarios with low user click behavior, such as advertising recommendations. The finely tuned user representation can be used for finding similar users, for matchmaking in scenarios such as dating websites. The finely tuned user representation and attribute tag representation can be used for predicting profiles of other users, including gender, age, life status (single / married), etc. Additionally, it can be used to detect the physical and mental health of adolescents, and whether they have criminal tendencies, etc.

[0071] The following detailed description, in conjunction with specific embodiments, details the attribute prediction model training method, attribute prediction method, attribute prediction model, attribute prediction model training device, attribute prediction device, computer-readable medium, and electronic device provided in this disclosure.

[0072] Figure 2 A flowchart illustrating the steps of an attribute prediction model training method in some embodiments of this disclosure is shown schematically. Figure 2 As shown, the method mainly includes the following steps:

[0073] Step S210. Obtain source domain data, which includes the behavioral sequence of the sample object in the source domain.

[0074] The sample objects can be users in one or more application domains, such as users who install and use video clients, news clients, or other applications on mobile phones, computers, or other terminal devices. One way to form a behavioral sequence is to arrange the behavioral nodes performed by the sample objects in the source domain in chronological order. The behavioral sequence of the sample objects in the source domain can be a sequence of clicks on content within the source domain; for example, this sequence could be recorded data from a user's clicks to watch videos or view news within a week.

[0075] Step S220. Train the initial neural network using source domain data to obtain a source domain behavior representation model.

[0076] Using the source domain data obtained in step S210, the initial neural network can be iteratively trained to continuously update the network parameters. When the expected training effect is met, such as network convergence or reaching the preset number of iterations, the source domain behavior representation model can be obtained. The source domain data is unlabeled data; therefore, the network training in this step uses unsupervised learning to mine the intrinsic information of the sample objects' behavioral sequences in the source domain, to represent the deep behavioral features of the sample objects in the source domain. The initial neural network used to construct the source domain behavior representation model in this step can be a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), or a Transformer model based on an encoder-decoder architecture. However, considering that behavioral sequence data in real-world scenarios is usually very long (for example, a user may watch thousands of short videos within a week), if RNN modeling is used, gradient explosion or vanishing problems are likely to occur. Furthermore, the inherent sequence dependencies of RNNs make it difficult to fully utilize graphics processing unit (GPU) resources for acceleration. If a Transformer model is used, the quadratic complexity becomes very time-consuming when the sequence is long because each element in the self-attention mechanism needs to calculate its similarity with other elements. Therefore, a CNN model is preferred for the initial neural network in this step. Of course, RNN, Transformer, or other models can also be used in other implementations, and this disclosure does not impose any special limitations on them.

[0077] Step S230. Freeze the parameters of the source domain behavior representation model and insert a fine-tuning network layer into the source domain behavior representation model.

[0078] Fine-tuning methods in transfer learning are often greatly influenced by the specific task and data. The final effect of fine-tuning depends on the operator's experience level. For example, whether the fine-tuning stage freezes all parameters of the pre-trained network, fine-tunes a portion of the network parameters, or fine-tunes all network parameters will produce different training results. Generally speaking, freezing all parameters can easily lead to inconsistencies between the source and target domains; fine-tuning all parameters can easily lead to forgetting pre-trained network parameters, which can easily cause overfitting when there are insufficient samples in the target domain. Fine-tuning a portion of the network is frequently used in practice, but it is highly dependent on parameter tuning and experience. Based on the above considerations, this step freezes all network parameters of the pre-trained source domain behavior representation model and inserts a fine-tuned network layer into the source domain behavior representation model for subsequent parameter fine-tuning. In some optional implementations, an initial neural network with a fine-tuned network layer can be pre-constructed. When training the source domain behavior representation model in step S220, the fine-tuned network layer is kept off, and then turned on in this step to achieve the effect of inserting a fine-tuned network layer. Additionally, when a convolutional neural network is used for the source domain behavior representation model, this step can insert multiple fine-tuning network layers at equal intervals between the multiple convolutional layers in the source domain behavior representation model. For example, a fine-tuning network layer can be inserted every two convolutional layers. The evenly distributed multiple fine-tuning network layers are beneficial for improving the training speed and effectiveness of the model.

[0079] Step S240. Obtain target domain data, which includes the behavioral sequence of the sample object in the source domain and the attribute labels in the target domain.

[0080] Target domain data serves as sample data in the fine-tuning phase, used to transfer the behavioral characteristics of sample objects from the source domain to the target domain. For example, the source domain could be video viewing, and the target domain could be various fields such as ad recommendation, news recommendation, music recommendation, or attribute prediction. Unlike source domain data, the target domain data obtained in this step is labeled data. In other words, the target domain data obtained in this step includes not only the behavioral sequences of sample objects in the source domain but also the attribute labels of sample objects in the target domain. When the target domain is ad recommendation, attribute labels can be labels corresponding to different products or different types of ads that the user clicks to watch. When the target domain is attribute prediction, attribute labels can be categorized labels for different aspects of the user, such as age, gender, and personality. For example, labels such as "teenager," "middle-aged," and "elderly" can be used for age, and labels such as "male" and "female" can be used for gender.

[0081] Step S250. Train the source domain behavior representation model with fine-tuned network layers using target domain data to obtain an attribute prediction model for the target domain.

[0082] Using the target domain data obtained in step S240, this step will continue training the source domain behavior representation model with fine-tuned network layers. During the training process in this step, the network parameters of the source domain behavior representation model have been frozen, so these network parameters will not be updated. However, through iterative training, the network parameters of the fine-tuned network layers will be continuously adjusted and optimized until the expected training effect is met, ultimately resulting in an attribute prediction model that predicts object attributes in the target domain.

[0083] In the attribute prediction model training method provided in this disclosure, a source domain behavior representation model is pre-trained using unsupervised learning with source domain data. Then, the parameters of the source domain behavior representation model are fine-tuned by inserting a fine-tuning network layer with target domain data. This effectively transfers behavioral features from the source domain to the target domain. By freezing the parameters of the source domain behavior representation model and inserting a fine-tuning network layer, the learning results from the pre-training stage can be retained during the fine-tuning stage, while significantly reducing the number of parameter adjustments and improving the training speed and effectiveness of the model.

[0084] Figure 3 The structural framework of the attribute prediction model in some embodiments of this disclosure is illustrated schematically. For example... Figure 3 As shown, the attribute prediction model trained in this embodiment of the present disclosure may mainly include: an input layer 310, a convolutional network layer 320, a fine-tuning network layer 330, and an output layer 340.

[0085] The input layer 310 is used to input the behavior sequence of the object under test in the source domain. The input layer 310 can map each behavior node in the behavior sequence to an embedding vector through the embedding matrix (i.e., as an embedding layer).

[0086] Convolutional network layer 320 is used to perform convolutional processing on the behavioral sequence to obtain behavioral representation information. Convolutional network layer 320 can be a dilated convolutional network layer with a certain dilation rate. Furthermore, adjacent convolutional network layers can have different dilation rates. Using dilated convolution can exponentially increase the receptive field of each layer.

[0087] The fine-tuning network layer 330 is located among multiple convolutional network layers 320 and is used to map behavioral representation information to obtain attribute prediction information corresponding to behavioral sequences. The fine-tuning network layer 330 can be distributed at equal intervals among multiple convolutional network layers 320, for example, one fine-tuning network layer 330 can be configured every two convolutional network layers 320.

[0088] Output layer 340 is used to output the attribute label of the test object in the target domain based on the attribute prediction information. Output layer 340 can use Softmax or other functions to calculate the classification probability of different attribute labels in the target domain, and take the classification result with the highest probability as the attribute label of the test object in the target domain.

[0089] In some optional implementations, the attribute prediction model 300 may further include residual connection branches distributed among multiple convolutional network layers 320. These residual connection branches allow information from shallower networks to be directly passed to subsequent deeper networks, changing the network's learning objective to learning the residuals between input and output data. This avoids the vanishing gradient phenomenon caused by continuously increasing network depth, thereby improving network training performance.

[0090] The details of the attribute prediction model training method in each embodiment of this disclosure are explained below with reference to the above model examples.

[0091] Figure 4 A flowchart illustrating the steps of training a source domain behavior representation model is shown schematically in some embodiments of this disclosure. For example... Figure 4 As shown, based on the above embodiments, step S220, training the initial neural network using source domain data to obtain a source domain behavior representation model, may include the following steps:

[0092] Step S410. Select multiple target behavior nodes in the behavior sequence of the source domain data and replace the target behavior nodes with occlusion behavior nodes.

[0093] The source domain data includes the behavioral sequences of sample objects within the source domain, and these behavioral sequences are composed of a series of behavioral nodes. This step first selects multiple target behavioral nodes from the behavioral sequences, and then replaces the selected target behavioral nodes with occlusion behavioral nodes, which is equivalent to using occlusion behavioral nodes to occlude some nodes in the behavioral sequences. The target behavioral nodes can be randomly assigned. Preferably, 20-30% of the behavioral nodes in a behavioral sequence can be randomly selected as target behavioral nodes.

[0094] Step S420. Input the behavior sequence with occluded behavior nodes into the initial neural network to obtain behavior prediction information corresponding to the occluded behavior nodes.

[0095] By occluding some behavior nodes in the behavior sequence, the initial neural network can learn the contextual relationships between each behavior node and use the unoccluded behavior nodes to predict the occluded content of the occluded behavior nodes, thus obtaining the behavior prediction information corresponding to the occluded behavior nodes.

[0096] Step S430. Determine the behavior prediction error based on the behavior prediction information and the target behavior node.

[0097] Step S420 yields the behavior prediction information corresponding to each occluded behavior node. This step compares this behavior prediction information with the corresponding target behavior node to determine the behavior prediction error. The behavior prediction error represents the reliability of the neural network's prediction of the occluded target behavior node. The lower the behavior prediction error, the closer the behavior prediction information obtained by the neural network is to the true target behavior node, and the stronger the prediction ability of the neural network.

[0098] Step S440. Update the network parameters of the initial neural network using the behavior prediction error to obtain the source domain behavior representation model.

[0099] Based on the behavior prediction error obtained in step S430, backpropagation can be performed in the initial neural network to complete one round of network parameter updates. Alternating between forward and backward propagation allows for continuous adjustment and optimization of the network parameters, ultimately resulting in a source domain behavior representation model that meets the expected training effect.

[0100] In some optional implementations, the occlusion behavior nodes used to occlude target behavior nodes in a behavior sequence can include two types: specified behavior nodes and random behavior nodes. Based on this, different types of occlusion behavior nodes can be used to occlude target behavior nodes. Figure 5 A flowchart illustrating the steps of replacing occlusion behavior nodes is shown schematically in some embodiments of this disclosure. For example... Figure 5 As shown, based on the above embodiments, replacing the target behavior node with an occlusion behavior node in step S410 may include the following steps:

[0101] Step S510. Determine the node classification ratio, and classify the target behavior nodes into the specified node set, the random node set, and the original node set according to the node classification ratio.

[0102] Multiple target behavior nodes can be identified in the behavior sequence. These target behavior nodes can be randomly divided into three parts: a designated node set, a random node set, and an original node set. The number of nodes in each of the designated, random, and original node sets should satisfy a node classification ratio. For example, this step can classify the target behavior nodes according to an 8:1:1 node classification ratio, where 80% of the target behavior nodes can be classified into the designated node set, 10% into the random node set, and the remaining 10% into the original node set.

[0103] Step S520. Determine the specified behavior node and replace the target behavior node in the specified node set with the specified behavior node.

[0104] For a target behavior node in a specified set of behavior nodes, replace it with a pre-defined behavior node. For example, if a behavior sequence contains 256 behavior nodes numbered 0-255, this step can identify the behavior node numbered 0 (or 255) as the specified behavior node. All target behavior nodes in the specified set of behavior nodes will be replaced with the behavior node numbered 0, that is, the behavior node numbered 0 is used to occlude all target behavior nodes in the specified set of nodes.

[0105] Step S530. Determine the random behavior nodes and replace the target behavior nodes in the random node set with random behavior nodes.

[0106] For each target behavior node in the set of random nodes, a randomly determined random behavior node is used to replace it. For example, if a behavior sequence includes 256 behavior nodes numbered 0-255, this step can randomly select one behavior node from these 256 (e.g., behavior node numbered 129) as the random behavior node, and then replace the target behavior node in the set of random behavior nodes with this random behavior node. Since the random behavior node is randomly selected, each target behavior node in the set of random nodes may be replaced with the same random behavior node, or it may be replaced with different random behavior nodes.

[0107] For the target behavior nodes in the original node set, no further occlusion operation is performed on them; this is equivalent to using the original behavior nodes to occlude themselves. However, these target behavior nodes will still be predicted in the subsequent neural network prediction process.

[0108] In this embodiment of the disclosure, by randomly occluding or not occluding a certain proportion of target behavior nodes, a certain amount of noise can be introduced into the behavior sequence with occluded behavior nodes, thereby improving the robustness of the trained model.

[0109] Figure 6 A flowchart illustrating the steps for obtaining behavior prediction information based on occlusion behavior nodes in some embodiments of this disclosure is shown. Figure 6 As shown, based on the above embodiments, step S420, inputting the behavior sequence with occluded behavior nodes into the initial neural network to obtain behavior prediction information corresponding to the occluded behavior nodes, may include the following steps:

[0110] Step S610. Input the behavior sequence with occluded behavior nodes into the initial neural network.

[0111] The initial neural network architecture typically includes an embedding layer and multiple convolutional layers, with the last convolutional layer potentially connected to a softmax layer.

[0112] Step S620. The behavior sequence is mapped through the embedding layer in the initial neural network to obtain the embedding vector of each behavior node in the behavior sequence.

[0113] The embedding layer is mainly used to map high-dimensional one-hot encoding to a low-dimensional embedding matrix, where each row of the embedding matrix represents the embedding vector of a row node.

[0114] Step S630. Perform convolution processing on the embedded vector through the convolutional layer in the initial neural network to obtain the behavior prediction information corresponding to the occlusion behavior node.

[0115] The multiple convolutional layers connected in sequence in the initial neural network can perform convolution processing on the embedding vector layer by layer, thereby obtaining the behavior prediction information corresponding to the occlusion behavior node.

[0116] Generally, the deeper the network, the more information from different levels can be extracted, and the more combinations of information from different levels are possible. However, as the number of network layers increases, problems such as vanishing gradients and exploding gradients are prone to occur, leading to a degradation in network performance. To address this issue, embodiments of this disclosure can introduce shortcut connections between the convolutional layers of the initial neural network, thereby introducing shallow information into deeper computations. Figure 7 A flowchart illustrating the steps of convolution processing based on residual connections in some embodiments of this disclosure is shown schematically. Figure 7 As shown, based on the above embodiments, step S630, performing convolution processing on the embedded vector through the convolutional layer in the initial neural network to obtain the behavior prediction information corresponding to the occlusion behavior node, may include the following steps:

[0117] Step S710. Perform convolution processing on the embedded vector through the convolutional layers in the initial neural network to obtain convolution processing information.

[0118] Each convolutional layer can use kernels of different sizes and numbers to convolve the embedded vector, obtaining corresponding convolutional processing information. For two adjacent convolutional layers, the convolutional processing information output by the previous layer serves as the input to the next layer, continuously extracting deeper information from the embedded vector through layer-by-layer convolution. In some optional implementations, the convolutional layers in the initial neural network can be dilated convolutional layers with a certain dilation rate. Accordingly, this step can use multiple dilated convolutional layers with different dilation rates in the initial neural network to perform dilated convolution processing on the embedded vector to obtain convolutional processing information.

[0119] Step S720. Map the embedding vector and convolution processing information through the residual connection branch in the initial neural network to obtain residual mapping information.

[0120] Residual connection branches distributed across multiple convolutional layers in the initial neural network can map the input information to obtain residual mapping information. A residual connection branch connecting an embedding layer and a convolutional layer maps the embedding vector and then passes the residual mapping information to the corresponding convolutional layer. A residual connection branch connecting two convolutional layers maps the convolutional processing information from the previous convolutional layer and then passes the residual mapping information to the next convolutional layer. Generally, residual connection branches directly perform an identity mapping on the input information.

[0121] Step S730. Obtain the behavior prediction information corresponding to the occlusion behavior node based on the convolution processing information and residual mapping information.

[0122] The convolutional processing information obtained from the convolutional layers and the residual mapping information obtained from the residual connection branches can be stacked element-wise, i.e., element-wise multiplication is performed. After the above steps of convolution, mapping, and stacking are performed by multiple convolutional layers and multiple residual connection branches, the behavior prediction information corresponding to the occlusion behavior node will finally be obtained.

[0123] The above embodiments describe the relevant methods and steps for pre-training source domain behavior representation models based on initial neural networks. Figure 8 This schematically illustrates a network architecture for a source domain behavior representation model during the pre-training phase. For example... Figure 8 As shown, the neural network includes an embedding layer 810 arranged from bottom to top, eight sequentially connected convolutional layers 820, and a softmax layer 830. Additionally, residual connection branches 840 are distributed between each convolutional layer 820.

[0124] During the model's pre-training phase, the behavior sequence input to embedding layer 810 is a sequence with occluded behavior nodes. As shown in the lower part of the figure, the complete behavior sequence consists of 17 behavior nodes numbered 0-16. When inputting it into embedding layer 810, behavior nodes numbered 2, 7, 9, 12, 14, and 16 are occluded (exampleed as black squares in the figure). After forward propagation layer by layer, the softmax layer outputs the prediction information corresponding to each occluded behavior node. These predictions are then compared with the real label data to calculate the cross-entropy loss function, yielding the prediction error. Based on this prediction error, backward propagation layer by layer is performed to adjust and optimize the network parameters at each layer.

[0125] like Figure 8 As shown in the figure, convolutional layer 820 in this neural network is a dilated convolutional layer with a certain dilation rate. The dilation rates of each convolutional layer shown in the figure are {1, 2, 4, 8, 1, 2, 4, 8} from bottom to top. Using dilated convolution can increase the receptive field exponentially. As exemplified by the solid lines in the figure, the receptive fields of the first to fourth convolutional layers are {3, 7, 15, 31} respectively.

[0126] In real-world applications involving long sequences, such as a user watching 1000 short videos a day, the neural network can be continuously deepened to extract deeper information. For example, it can be repeatedly... Figure 8 In a similar neural network architecture, the dilation values ​​of dilated convolutional layers can be set as follows: {1, 2, 4, …,128, 1, 2, 4, …,128, 1, 2, 4, …,128}. As the number of layers increases, the larger the dilation value, the larger the receptive field of the convolutional layer.

[0127] In the pre-training phase of the source domain behavior representation model based on source domain data, unsupervised learning is used throughout the training. After the model converges, the final softmax layer can be removed, and a fine-tuning network layer can be inserted. Then, the fine-tuning phase based on target domain data begins. In the fine-tuning phase, supervised learning can be performed using the behavioral sequences of sample objects in the target domain data and their associated attribute labels in the target domain, thereby training an attribute prediction model for the target domain.

[0128] Figure 9 A flowchart illustrating the steps of training an attribute prediction model during the fine-tuning phase in some embodiments of this disclosure is shown. Figure 9As shown, based on the above embodiments, step S250, training the source domain behavior representation model with fine-tuned network layers using target domain data to obtain an attribute prediction model for the target domain, may include the following steps:

[0129] Step S910. Input the behavior sequence in the target domain data into the source domain behavior representation model with fine-tuned network layers to obtain attribute prediction information corresponding to the behavior sequence.

[0130] Similar to the forward propagation process of the source domain behavior representation model during the pre-training phase, after inputting the behavior sequence from the target domain data into the source domain behavior representation model with fine-tuning network layers, the source domain behavior representation model and the fine-tuning network layers inserted therein can extract information from the behavior sequence layer by layer, and finally obtain the attribute prediction information corresponding to the behavior sequence output by the model.

[0131] Step S920. Determine the attribute labels associated with the behavior sequence, and determine the attribute prediction error based on the attribute labels and attribute prediction information.

[0132] In the target domain data, the behavioral sequence of each sample object is labeled and associated with attribute labels. The attribute prediction error can be determined based on the difference between the attribute labels and the attribute prediction information. The lower the attribute prediction error, the closer the attribute prediction information is to the true attribute labels, and the stronger the model's attribute prediction ability in the target domain.

[0133] Step S930. Update the parameters of the fine-tuned network layer using the attribute prediction error to obtain the attribute prediction model for the target domain.

[0134] Based on the attribute prediction error obtained in step S920, backpropagation can be performed in the source domain behavior representation model with fine-tuned network layers to complete one round of updating the network parameters in the model. Since the network parameters of the source domain behavior representation model have been frozen, only the network parameters in the fine-tuned network layers will be updated during backpropagation. By alternating between forward and backpropagation, the network parameters in the fine-tuned network layers can be continuously adjusted and optimized, ultimately obtaining an attribute prediction model that meets the expected training effect.

[0135] Figure 10 A network architecture for an attribute prediction model during the fine-tuning phase is illustrated schematically. For example... Figure 10 As shown, the basic architecture of the attribute prediction model is... Figure 8Based on the source domain behavior representation model shown, a fine-tuned network layer 1010 is inserted every two convolutional layers. After inputting the behavior sequence in the target domain data into the source domain behavior representation model with the fine-tuned network layer shown in the figure and performing a bottom-up forward propagation, attribute prediction information can be obtained from the output of a fine-tuned network layer located at the top of the figure. This attribute prediction information can be information obtained after pooling processing, such as average pooling or max pooling.

[0136] Figure 11 A flowchart illustrating the steps for obtaining attribute prediction information in some embodiments of this disclosure is shown schematically. For example... Figure 11 As shown, based on the above embodiments, step S910, inputting the behavior sequence from the target domain data into the source domain behavior representation model with fine-tuned network layers to obtain attribute prediction information corresponding to the behavior sequence, may include the following steps:

[0137] Step S1110. Input the behavioral sequence from the target domain data into the source domain behavioral representation model with fine-tuned network layers;

[0138] Step S1120. Perform convolution processing on the behavior sequence through the convolutional layer in the source domain behavior representation model to obtain behavior representation information;

[0139] Step S1130. The behavioral representation information is mapped by fine-tuning the network layers to obtain attribute prediction information corresponding to the behavioral sequence.

[0140] Since the network parameters of the source domain behavior representation model are frozen, the learning results for behavior representation during the pre-training phase are retained. Therefore, the convolutional layers in the source domain behavior representation model can obtain behavior representation information by performing convolution processing on the behavior sequences in the target domain data. The fine-tuning network layers are used to transfer the behavior representation information corresponding to the source domain to the target domain, and obtain the attribute prediction information of the behavior sequence in the target domain through mapping processing.

[0141] Figure 12 A flowchart illustrating the steps of determining attribute prediction error based on attribute prediction information in some embodiments of this disclosure is shown. For example... Figure 12 As shown, based on the above embodiments, the step S920 of determining the attribute prediction error according to the attribute label and attribute prediction information may include the following steps:

[0142] Step S1210. Map the attribute labels according to the attribute information of the target domain to obtain label representation information.

[0143] Figure 13 This schematically illustrates the model architecture for mapping attribute labels. For example... Figure 13As shown, for attribute labels 1310 in the target domain data, an embedding matrix 1320 can be used to map them to obtain label representation information corresponding to the behavioral sequences in the target domain data. For target domains with different attribute information, embedding matrices with different structures need to be constructed. For example, if the target domain is ad recommendation, the length of the embedding matrix is ​​the number of ads; if the target domain is gender prediction, the length of the embedding matrix is ​​two. In this embodiment, the target domain can be an attribute prediction scenario, such as predicting a user's age, gender, personality, life status, violent tendencies, success potential, preferences, etc. Attribute labels corresponding to life status can include, for example, high school, university, single, married, dating, elderly, etc.; attribute labels corresponding to personality can include, for example, optimistic, negative, shy, etc. Furthermore, the target domain in this embodiment can also be various content recommendation scenarios, such as ad recommendations, news recommendations, music recommendations, dating and matchmaking recommendations, long and short video recommendations, etc.

[0144] Step S1220. Determine the similarity between the label representation information and the attribute prediction information, and define the similarity as the attribute prediction error.

[0145] After obtaining the label representation information in step S1210, this step compares the label representation information with the corresponding attribute prediction information, calculates the similarity between the two, and determines the calculated similarity as the attribute prediction error. For example, this step can determine the attribute prediction error by calculating the vector dot product or the cosine similarity. Figure 14 The overall network architecture during the fine-tuning phase is illustrated schematically. For example... Figure 14 As shown, during the fine-tuning phase of training, attribute prediction model 1410 (corresponding to...) Figure 10 The network architecture shown can output attribute prediction information based on behavioral sequences in the target domain data. This attribute prediction information is user representation information obtained through average pooling. Meanwhile, the label mapping model 1420 (corresponding to...) Figure 13 The model architecture shown can output label representation information based on the attribute labels in the target domain data. This step calculates the inner product of the label representation information and the attribute prediction information to obtain the attribute prediction error. In the label mapping model 1420, different target domains correspond to different label categories. For example, Figure 14The document lists examples for various tag categories, including violent tendencies, life status, success potential, gender, personality, and preferences. For instance, "life status" could include tags such as high school, university, single, married, dating, elderly, etc.; "personality" could include tags such as optimistic, negative, shy, etc.; and "preferences" (i.e., content recommendation scenarios) could include tags such as ads recommended to inactive users, news recommendations, music recommendations, dating / matchmaking recommendations, and short / long video recommendations, etc.

[0146] Figure 15 The network architecture of fine-tuned network layers in some embodiments of this disclosure is illustrated schematically. For example... Figure 15 As shown, the fine-tuning network layer mainly includes convolutional branch 1510 and residual branch 1520. Based on this, Figure 16 A flowchart illustrating the steps of obtaining attribute prediction information by fine-tuning network layers in some embodiments of this disclosure is shown. Figure 16 As shown, step S1130, mapping the behavioral representation information by fine-tuning the network layers to obtain attribute prediction information corresponding to the behavioral sequence, may include the following steps:

[0147] Step S1610. Perform convolution processing on the behavioral representation information through convolution branches to obtain convolutional prediction information.

[0148] Step S1620. Map the behavioral representation information through residual branching to obtain residual prediction information.

[0149] Step S1630. Determine the attribute prediction information corresponding to the behavior sequence based on the convolutional prediction information and the residual prediction information.

[0150] The residual branch 1520 can be used to directly map the behavioral representation information of the input fine-tuning network layer to the output of the fine-tuning network layer. The mapped residual prediction information is superimposed with the convolutional prediction information obtained by convolution processing through the convolution branch 1510, thereby obtaining the attribute prediction information corresponding to the behavioral sequence.

[0151] Continue to refer to Figure 15 The convolutional branch 1510 may further include a dimension-reducing convolutional layer 1511, an activation layer 1512, and an increased-dimensional convolutional layer 1513 connected in sequence. The activation function of the activation layer 1512 is preferably a non-linear activation function based on Gaussian Error Linear Units (GELU). Alternatively, in some other embodiments, the activation function of the activation layer 1512 may also be a function such as Sigmoid or ReLU; this disclosure does not impose specific limitations on this.

[0152] In this embodiment, the fine-tuning network layer can construct a residual block based on a bottleneck structure. Specifically, the number of convolutional kernels in the dimensionality reduction convolutional layer 1511 can be configured to be less than the number of input channels in the fine-tuning network layer, while the number of convolutional kernels in the dimensionality increase convolutional layer 1513 can be configured to be equal to the number of input channels in the fine-tuning network layer.

[0153] For example, if the fine-tuning network layer has 4k input channels, then the number of feature maps input to the fine-tuning network layer is also 4k. The dimensionality reduction convolutional layer 1511 can use k 1×1 convolutional kernels to transform the dimensionality space of the input data, producing a dimensionality reduction effect. Correspondingly, the dimensionality increase convolutional layer 1513 can use 4k 1×1 convolutional kernels to further increase the dimensionality, ensuring that the input and output of the fine-tuning network layer have the same dimensionality space.

[0154] Based on this, Figure 17 A flowchart illustrating the steps of obtaining convolutional prediction information through convolutional branches in some embodiments of this disclosure is shown schematically. For example... Figure 17 As shown, step S1610, performing convolution processing on the behavioral representation information through convolution branches to obtain convolutional prediction information, may include the following steps:

[0155] Step S1710. Perform convolution processing on the behavioral representation information through a dimensionality reduction convolutional layer to obtain low-dimensional prediction information with a dimension lower than that of the behavioral representation information.

[0156] Step S1720. The low-dimensional prediction information is mapped through the activation layer to obtain activation prediction information with nonlinear characteristics.

[0157] Step S1730. Perform convolution processing on the activation prediction information through a dimension-upgrading convolutional layer to obtain convolutional prediction information with a dimension equal to that of the behavioral representation information.

[0158] The design of fine-tuning network layers based on bottleneck structures can significantly reduce the number of parameters that need to be adjusted during the fine-tuning stage, thereby reducing the network size during the fine-tuning stage. This makes it easier to deploy fine-tuning models on mobile platforms, and it can also effectively avoid overfitting problems in terms of performance.

[0159] Figure 15 The fine-tuned network layer shown can be represented by the following formula:

[0160] ;

[0161]

[0162] in, x For the input vector, y For the output vector,F The residual mapping function to be learned. W For the weights of the convolution terms, b This is a bias term.

[0163] In a fine-tuned network layer, the total number of parameters for the convolutional term weights W is: Bias term b The total number of parameters is .

[0164] The following example uses an attribute prediction model with 6 fine-tuning network layers and 12 convolutional network layers to estimate the number of tuning parameters.

[0165] The model information related to the convolutional network layers includes: the dilatation rates of the 12 convolutional network layers are {1, 2, 4, 8, 16, 32, 1, 2, 4, 8, 16, 32}, the number of input channels is 256 (i.e., 4k=256), the total amount of input data is 1 million, the average length of the action sequences is 500, the kernel size is 3, and the total parameters are 10 after ignoring the bias term and layer normalization parameters and removing the softmax layer. 6 ×256+500×256×256×3×12≈1.4 billion.

[0166] The number of parameters in the fine-tuning network layer is 256 × 64 × 2 × 6 = 190,000, which accounts for only 0.01% of the total number of parameters. In other words, by inserting a fine-tuning network layer, only 0.01% of the parameters need to be re-optimized during the fine-tuning stage to achieve the same effect as optimizing the total number of parameters.

[0167] The attribute prediction model training methods provided in the above embodiments can be used to train attribute prediction models for various target domains, and attribute prediction can be performed on the target domain based on these models. Figure 18 The illustrations schematically depict attribute prediction methods in some embodiments of this disclosure, such as... Figure 18 As shown, the method mainly includes the following steps:

[0168] Step S1810. Obtain the behavior sequence of the object under test in the source domain and determine the target domain corresponding to the source domain;

[0169] Step S1820. Input the behavior sequence into a pre-trained attribute prediction model for the target domain to obtain attribute prediction information of the object to be tested.

[0170] Experimental verification shows that the model prediction method provided in this disclosure has high accuracy in cold-start user recommendation and profile prediction scenarios. Cold-start users lack necessary click behavior, and in actual business, they are usually retrieved using methods such as popularity. Therefore, this disclosure uses the popularity method as a comparison baseline, referred to as MP. Simultaneously, this disclosure also employs another baseline, which uses user click behavior from the source domain scenario as input data and employs the same fine-tuning network layers. However, during the fine-tuning stage, all parameters in the attribute prediction model are optimized; that is, the parameters of the pre-trained source domain behavior representation model and the parameters of the fine-tuned network layers inserted therein are simultaneously optimized. This baseline can be called TFRec_Nopretrain. The baseline corresponding to the attribute prediction model trained in the above embodiments of this disclosure can be called TFRec_Adapter.

[0171] The specific settings for experimental verification of this disclosure are as follows:

[0172] Experimental setup:

[0173] In the experiments, all models used Adam as the optimizer, with a batch size of 256 and a learning rate of 0.0005 during the fine-tuning phase. 90% of the training data was used as the training set, and the remainder as the test set. During evaluation, a user-clicked item was used as a positive sample, and 99 negative samples were randomly selected. An attribute prediction model was used to score these 100 samples, and the top-5 samples (the five highest-scoring samples) were used to determine whether they contained any positive samples, represented by the parameter Hit@5. The experimental hardware environment consisted of a Tesla P40 GPU and Tensorflow version 1.7.0.

[0174] Application Scenario 1:

[0175] In this application scenario, the source domain represents click-and-view behavior in a browser's information feed, and the target domain represents click-and-view behavior in a content-driven social media platform's information feed. Users in the target domain are all cold-start users, with no more than five viewing actions per user, and most users exhibiting only one viewing action. The maximum length of a click-and-view sequence for users in the source domain is 100. The total content volume of the source domain data is 640,000, and the total content volume of the target domain data is 17,880.

[0176] The experimental results for this application scenario are shown in Table 1.

[0177] Table 1. Results of the content-based social media platform information feed click experiment

[0178]

[0179] Application Scenario 2:

[0180] In this application scenario, the source domain represents click-and-view behavior from a browser's information feed, and the target domain represents ad click-and-view behavior from an advertising platform. Specifically, all users in the target domain exhibit no more than 5 view behaviors, with most users having fewer than 3 view behaviors. The maximum length of the source domain user click behavior sequence is 50. The total content of the source domain data is 11350, and the total content of the target domain data is 163.

[0181] The experimental results for this application scenario are shown in Table 2.

[0182] Table 2. Results of the Ad Click-Through Experiment on the Advertising Platform

[0183]

[0184] Results Analysis: Both application scenarios 1 and 2 are recommendation system scenarios. It can be seen that after pre-training and fine-tuning using the transfer learning framework provided in this embodiment, TFRec_Adapter significantly outperforms the popularity method in the Hit@5 metric. Furthermore, although TFRec_Nopretrain and TFRec_Adapter have the same fine-tuned network structure, TFRec_Adapter's results are significantly better than TFRec_Nopretrain. This confirms that the attribute prediction model training method used in this embodiment can not only significantly reduce the amount of parameter tuning but also achieve better network training results than full parameter optimization.

[0185] Application Scenario 3:

[0186] In addition to content recommendation, the attribute prediction model provided in this disclosure can also be used to predict user profile information. Tables 3 to 5 provide the experimental results for predicting user profile information such as gender, life status, and age.

[0187] Table 3. Results of the gender prediction experiment

[0188]

[0189] Table 4 Results of the Life Status Prediction Experiment

[0190]

[0191] Table 5. Results of the age prediction experiment (each ten years is considered a separate age group).

[0192]

[0193] As can be seen from the experimental results in Tables 3 to 5, the attribute prediction model provided in this embodiment of the present disclosure also has better predictive ability in predicting user profile information.

[0194] It should be noted that although the steps of the method in this disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0195] The following describes an apparatus embodiment of this disclosure, which can be used to execute the attribute prediction model training method and attribute prediction method in the above embodiments of this disclosure. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the method embodiments described above.

[0196] Figure 19 A schematic block diagram of an attribute prediction model training apparatus according to some embodiments of this disclosure is shown. For example... Figure 19 As shown, the attribute prediction model training device 1900 mainly includes: a source domain data acquisition module 1910, configured to acquire source domain data, which includes the behavioral sequences of sample objects in the source domain; a model pre-training module 1920, configured to train an initial neural network using the source domain data to obtain a source domain behavioral representation model; a model adjustment module 1930, configured to freeze the parameters of the source domain behavioral representation model and insert fine-tuning network layers into the source domain behavioral representation model; a target domain data acquisition module 1940, configured to acquire target domain data, which includes the behavioral sequences of sample objects in the source domain and attribute labels in the target domain; and a model fine-tuning module 1950, configured to train the source domain behavioral representation model with fine-tuning network layers using the target domain data to obtain an attribute prediction model for the target domain.

[0197] In some embodiments of this disclosure, based on the above embodiments, the model pre-training module includes: a node occlusion module, configured to select multiple target behavior nodes in the behavior sequence of the source domain data and replace the target behavior nodes with occlusion behavior nodes; a behavior prediction module, configured to input the behavior sequence with occlusion behavior nodes into an initial neural network to obtain behavior prediction information corresponding to the occlusion behavior nodes; a behavior error determination module, configured to determine the behavior prediction error based on the behavior prediction information and the target behavior nodes; and a pre-training parameter update module, configured to update the network parameters of the initial neural network using the behavior prediction error to obtain a source domain behavior representation model.

[0198] In some embodiments of this disclosure, based on the above embodiments, the occlusion behavior node includes a specified behavior node and a random behavior node; the node occlusion module includes: a node classification module, configured to determine a node classification ratio, and classify the target behavior node into a specified node set, a random node set, and an original node set according to the node classification ratio; a specified replacement module, configured to determine a specified behavior node, and replace the target behavior node in the specified node set with the specified behavior node; and a random replacement module, configured to determine a random behavior node, and replace the target behavior node in the random node set with a random behavior node.

[0199] In some embodiments of this disclosure, based on the above embodiments, the behavior prediction module includes: a pre-training input module configured to input a behavior sequence with occluded behavior nodes into an initial neural network; a pre-training mapping module configured to map the behavior sequence through an embedding layer in the initial neural network to obtain an embedding vector for each behavior node in the behavior sequence; and a pre-training convolution module configured to convolve the embedding vector through a convolution layer in the initial neural network to obtain behavior prediction information corresponding to the occluded behavior nodes.

[0200] In some embodiments of this disclosure, based on the above embodiments, the pre-trained convolution module includes: a pre-trained convolution processing module configured to perform convolution processing on the embedded vector through the convolutional layer in the initial neural network to obtain convolution processing information; a pre-trained mapping processing module configured to perform mapping processing on the embedded vector and the convolution processing information through the residual connection branch in the initial neural network to obtain residual mapping information; and a pre-trained information overlay module configured to obtain behavior prediction information corresponding to the occlusion behavior node based on the convolution processing information and the residual mapping information.

[0201] In some embodiments of this disclosure, based on the above embodiments, the convolution processing module includes: a dilated convolution module, configured to perform dilated convolution processing on the embedded vector through multiple dilated convolution layers with different dilation rates in the initial neural network to obtain convolution processing information.

[0202] In some embodiments of this disclosure, based on the above embodiments, the model fine-tuning module includes: an attribute prediction module, configured to input a behavior sequence from the target domain data into a source domain behavior representation model with a fine-tuned network layer to obtain attribute prediction information corresponding to the behavior sequence; an attribute error determination module, configured to determine the attribute label associated with the behavior sequence and determine the attribute prediction error based on the attribute label and the attribute prediction information; and a fine-tuning parameter update module, configured to update the parameters of the fine-tuned network layer using the attribute prediction error to obtain an attribute prediction model for the target domain.

[0203] In some embodiments of this disclosure, based on the above embodiments, the attribute prediction module includes: a fine-tuning input module, configured to input the behavior sequence in the target domain data into a source domain behavior representation model with a fine-tuning network layer; a fine-tuning convolution module, configured to perform convolution processing on the behavior sequence through the convolutional layer in the source domain behavior representation model to obtain behavior representation information; and a fine-tuning mapping module, configured to perform mapping processing on the behavior representation information through the fine-tuning network layer to obtain attribute prediction information corresponding to the behavior sequence.

[0204] In some embodiments of this disclosure, based on the above embodiments, the fine-tuning network layer includes a convolutional branch and a residual branch; the fine-tuning mapping module includes: a fine-tuning convolutional processing module, configured to perform convolutional processing on the behavior representation information through the convolutional branch to obtain convolutional prediction information; a fine-tuning mapping processing module, configured to perform mapping processing on the behavior representation information through the residual branch to obtain residual prediction information; and a fine-tuning information overlay module, configured to determine attribute prediction information corresponding to the behavior sequence based on the convolutional prediction information and the residual prediction information.

[0205] In some embodiments of this disclosure, based on the above embodiments, the convolutional branch includes a dimensionality-reducing convolutional layer, an activation layer, and an increasing-dimensional convolutional layer connected in sequence; the fine-tuning convolutional processing module includes: a dimensionality-reducing convolutional module configured to perform convolution processing on the behavioral representation information through the dimensionality-reducing convolutional layer to obtain low-dimensional prediction information with a dimension lower than that of the behavioral representation information; an activation module configured to perform mapping processing on the low-dimensional prediction information through the activation layer to obtain activation prediction information with nonlinear features; and an increasing-dimensional convolutional module configured to perform convolution processing on the activation prediction information through the increasing-dimensional convolutional layer to obtain convolutional prediction information with a dimension equal to that of the behavioral representation information.

[0206] In some embodiments of this disclosure, based on the above embodiments, the attribute error determination module includes: an attribute label mapping module, configured to map attribute labels according to the attribute information of the target domain to obtain label representation information; and a prediction similarity determination module, configured to determine the similarity between the label representation information and the attribute prediction information, and determine the similarity as the attribute prediction error.

[0207] In some embodiments of this disclosure, based on the above embodiments, the model adjustment module includes: a parameter freezing module, configured to freeze the parameters of the source domain behavior representation model; and a fine-tuning layer insertion module, configured to insert multiple fine-tuning network layers at equal intervals between multiple convolutional layers in the source domain behavior representation model.

[0208] In some embodiments of this disclosure, based on the above embodiments, the behavior sequence of the sample object in the source domain is the content click behavior sequence of the sample object on the content in the source domain.

[0209] Figure 20A structural block diagram of an attribute prediction apparatus according to some embodiments of this disclosure is shown schematically. For example... Figure 20 As shown, the attribute prediction device 2000 mainly includes: a data acquisition module 2010, configured to acquire the behavior sequence of the object to be tested in the source domain and determine the target domain corresponding to the source domain; and a model prediction module 2020, configured to input the behavior sequence into a pre-trained attribute prediction model for the target domain to obtain attribute prediction information of the object to be tested; wherein, the attribute prediction model is trained by the attribute prediction model training method as described in the above embodiments.

[0210] Figure 21 A schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present disclosure is shown.

[0211] It should be noted that, Figure 21 The computer system 2100 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0212] like Figure 21 As shown, the computer system 2100 includes a Central Processing Unit (CPU) 2101, which can perform various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) 2102 or programs loaded from storage section 2108 into Random Access Memory (RAM) 2103. The RAM 2103 also stores various programs and data required for system operation. The CPU 2101, ROM 2102, and RAM 2103 are interconnected via a bus 2104. An input / output (I / O) interface 2105 is also connected to the bus 2104.

[0213] The following components are connected to I / O interface 2105: an input section 2106 including a keyboard, mouse, etc.; an output section 2107 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 2108 including a hard disk, etc.; and a communication section 2109 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 2109 performs communication processing via a network such as the Internet. A drive 2110 is also connected to I / O interface 2105 as needed. Removable media 2111, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 2110 as needed so that computer programs read from them can be installed into storage section 2108 as needed.

[0214] In particular, according to embodiments of this disclosure, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 2109, and / or installed from removable medium 2111. When the computer program is executed by central processing unit (CPU) 2101, it performs various functions defined in the system of this application.

[0215] It should be noted that the computer-readable medium shown in the embodiments of this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0216] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0217] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0218] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this disclosure.

[0219] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein.

[0220] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for training an attribute prediction model, characterized in that, include: Obtain source domain data, which includes a sequence of actions of a sample object in the source domain, where the source domain represents an application, and the action sequence is a sequence of click actions of the sample object on content in the application, including at least one action data among browsing history, rating, likes, and shares; The source domain data is used to perform unsupervised learning training on the initial neural network to obtain a source domain behavior representation model, which is used to represent the behavioral habits of the sample object in the application. Freeze all network parameters of the source domain behavior representation model, and insert multiple fine-tuned network layers at equal intervals between multiple convolutional layers in the source domain behavior representation model; Obtain target domain data, which includes the behavior sequence and attribute tags in the target domain. The target domain is an advertising recommendation domain, news recommendation domain, music recommendation domain, dating recommendation domain, video recommendation domain, or attribute prediction domain. The source domain behavior representation model with a fine-tuned network layer is trained using the target domain data in a supervised learning manner, and the network parameters of the fine-tuned network layer are updated to obtain an attribute prediction model for the target domain. The attribute prediction model is used to predict the attributes of objects in the target domain, and the obtained attribute prediction information is used to recommend advertisements, news, music, dating and matchmaking services, videos, or predict the attribute tags of the objects. The step of using the source domain data to perform unsupervised learning training on the initial neural network to obtain a source domain behavior representation model includes: selecting multiple target behavior nodes in the behavior sequence and replacing the target behavior nodes with occluded behavior nodes; inputting the behavior sequence with occluded behavior nodes into the initial neural network to obtain behavior prediction information; determining the behavior prediction error based on the behavior prediction information and the target behavior nodes; and updating the network parameters of the initial neural network using the behavior prediction error to obtain the source domain behavior representation model. The step of using the target domain data to perform supervised learning training on the source domain behavior representation model with a fine-tuned network layer, updating the network parameters of the fine-tuned network layer, and obtaining an attribute prediction model for the target domain includes: inputting the behavior sequence from the target domain data into the source domain behavior representation model with a fine-tuned network layer to obtain attribute prediction information; determining the attribute label associated with the behavior sequence; determining the attribute prediction error based on the attribute label and the attribute prediction information; and updating the parameters of the fine-tuned network layer using the attribute prediction error to obtain an attribute prediction model for the target domain.

2. The attribute prediction model training method according to claim 1, characterized in that, The occlusion behavior node includes a specified behavior node and a random behavior node; replacing the target behavior node with the occlusion behavior node includes: Determine the node classification ratio, and classify the target behavior node into a specified node set, a random node set, and an original node set according to the node classification ratio; Determine a specified behavior node, and replace the target behavior node in the specified node set with the specified behavior node; Identify random behavior nodes and replace the target behavior nodes in the set of random nodes with the random behavior nodes.

3. The attribute prediction model training method according to claim 1, characterized in that, The step of inputting the behavior sequence with occluded behavior nodes into the initial neural network to obtain behavior prediction information includes: Input the behavior sequence with the occlusion behavior node into the initial neural network; The behavior sequence is mapped through the embedding layer in the initial neural network to obtain the embedding vector of each behavior node in the behavior sequence; The embedding vector is convolved by the convolutional layers in the initial neural network to obtain the behavior prediction information corresponding to the occlusion behavior node.

4. The attribute prediction model training method according to claim 3, characterized in that, The step of convolving the embedded vector through the convolutional layers in the initial neural network to obtain the behavior prediction information corresponding to the occlusion behavior node includes: The embedding vector is convolved through the convolutional layers in the initial neural network to obtain convolutional processing information. The embedding vector and the convolution processing information are mapped through the residual connection branch in the initial neural network to obtain residual mapping information; The behavior prediction information corresponding to the occlusion behavior node is obtained based on the convolution processing information and the residual mapping information.

5. The attribute prediction model training method according to claim 4, characterized in that, The step of performing convolution processing on the embedded vector through the convolutional layers in the initial neural network to obtain convolution processing information includes: The embedding vector is subjected to dilated convolution processing through multiple dilated convolutional layers with different dilation rates in the initial neural network to obtain the convolution processing information.

6. The attribute prediction model training method according to claim 1, characterized in that, The step of inputting the behavioral sequence from the target domain data into the source domain behavioral representation model with fine-tuned network layers to obtain attribute prediction information includes: The behavioral sequence is input into the source domain behavioral representation model with fine-tuned network layers; The behavior sequence is convolved by the convolutional layer in the source domain behavior representation model to obtain behavior representation information. The behavior representation information is mapped through the fine-tuned network layer to obtain the attribute prediction information corresponding to the behavior sequence.

7. The attribute prediction model training method according to claim 6, characterized in that, The fine-tuned network layer includes convolutional branches and residual branches; The step of mapping the behavioral representation information through the fine-tuned network layer to obtain the attribute prediction information corresponding to the behavioral sequence includes: The behavior representation information is processed by convolution through the convolution branch to obtain convolutional prediction information; The residual branch is used to map the behavioral representation information to obtain residual prediction information. The attribute prediction information corresponding to the behavior sequence is determined based on the convolution prediction information and the residual prediction information.

8. The attribute prediction model training method according to claim 7, characterized in that, The convolutional branch includes a sequentially connected dimensionality-reducing convolutional layer, an activation layer, and an updimensional convolutional layer; the process of performing convolutional processing on the behavioral representation information through the convolutional branch to obtain convolutional prediction information includes: The behavior representation information is convolved by the dimensionality reduction convolutional layer to obtain low-dimensional prediction information with a dimension lower than that of the behavior representation information. The low-dimensional prediction information is mapped through the activation layer to obtain activation prediction information with nonlinear characteristics. The activation prediction information is convolved by the up-dimensional convolutional layer to obtain the convolutional prediction information with a dimension equal to that of the behavioral representation information.

9. The attribute prediction model training method according to claim 1, characterized in that, The step of determining the attribute prediction error based on the attribute label and the attribute prediction information includes: The attribute labels are mapped based on the attribute information of the target domain to obtain label representation information; The similarity between the label representation information and the attribute prediction information is determined, and the similarity is defined as the attribute prediction error.

10. An attribute prediction method, characterized in that, include: Obtain the behavioral sequence of the object under test in the source domain, and determine the target domain corresponding to the source domain; The source domain represents an application, and the behavior sequence is a sequence of click behavior of the object under test on content in the application. The click behavior sequence includes at least one behavior data among browsing history, rating, likes and forwards. The behavioral sequence is input into a pre-trained attribute prediction model for the target domain to obtain the attribute prediction information of the object to be tested. The target domain is the advertising recommendation domain, news recommendation domain, music recommendation domain, dating recommendation domain, video recommendation domain, or attribute prediction domain. The attribute prediction model is trained by the attribute prediction model training method as described in any one of claims 1 to 9; the attribute prediction model is used to predict the attributes of objects in the target domain, and the obtained attribute prediction information is used to recommend advertisements, news, music, dating and matchmaking services, videos, or predict the attribute tags of the objects.

11. A training device for an attribute prediction model, characterized in that, The device includes: The source domain data acquisition module is configured to acquire source domain data, which includes a sequence of actions of a sample object in the source domain, where the source domain represents an application. The action sequence is a sequence of click actions of the sample object on content in the application, and the click action sequence includes at least one action data among browsing history, rating, likes, and forwards. The model pre-training module is configured to perform unsupervised learning training on an initial neural network using the source domain data to obtain a source domain behavior representation model, which is used to represent the behavioral habits of the sample object in the application. The model tuning module is configured to freeze all network parameters of the source domain behavior representation model and insert multiple fine-tuning network layers at equal intervals between multiple convolutional layers in the source domain behavior representation model. The target domain data acquisition module is configured to acquire target domain data, which includes the behavioral sequence of the sample object in the source domain and the attribute labels in the target domain. The target domain is an advertising recommendation domain, a news recommendation domain, a music recommendation domain, a dating recommendation domain, a video recommendation domain, or an attribute prediction domain. The model fine-tuning module is configured to use the target domain data to perform supervised learning training on the source domain behavior representation model with the fine-tuned network layer, update the network parameters of the fine-tuned network layer, and obtain an attribute prediction model for the target domain. The attribute prediction model is used to predict the attributes of objects in the target domain, and the obtained attribute prediction information is used to recommend advertisements, news, music, dating and matchmaking services, videos, or predict the attribute tags of the objects. The model pre-training module includes: a node occlusion module, configured to select multiple target behavior nodes in the behavior sequence and replace the target behavior nodes with occlusion behavior nodes; a behavior prediction module, configured to input the behavior sequence with occlusion behavior nodes into the initial neural network to obtain behavior prediction information; a behavior error determination module, configured to determine the behavior prediction error based on the behavior prediction information and the target behavior nodes; and a pre-training parameter update module, configured to update the network parameters of the initial neural network using the behavior prediction error to obtain the source domain behavior representation model. The model fine-tuning module includes: an attribute prediction module, configured to input the behavioral sequence in the target domain data into the source domain behavioral representation model with the fine-tuned network layer to obtain attribute prediction information; determine the attribute label associated with the behavioral sequence; determine the attribute prediction error based on the attribute label and the attribute prediction information; and update the parameters of the fine-tuned network layer using the attribute prediction error to obtain an attribute prediction model for the target domain.

12. An attribute prediction device, characterized in that, The device includes: The data acquisition module is configured to acquire the behavior sequence of the object under test in the source domain and determine the target domain corresponding to the source domain; the source domain represents an application, and the behavior sequence is the click behavior sequence of the object under test on the content in the application, the click behavior sequence including at least one behavior data among browsing history, rating, likes and forwards; The model prediction module is configured to input the behavior sequence into a pre-trained attribute prediction model for the target domain to obtain attribute prediction information of the object to be tested. The target domain is an advertising recommendation domain, a news recommendation domain, a music recommendation domain, a dating recommendation domain, a video recommendation domain, or an attribute prediction domain. The attribute prediction model is trained by the attribute prediction model training method as described in any one of claims 1 to 9; the attribute prediction model is used to predict the attributes of objects in the target domain, and the obtained attribute prediction information is used to recommend advertisements, news, music, dating and matchmaking services, videos, or predict the attribute tags of the objects.

13. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 9.

14. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the method as described in any one of claims 1 to 9.

15. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to perform the method as described in any one of claims 1 to 9 by executing the executable instructions.

Images