Recommended methods, apparatuses, media, devices, and products

Abstract

A recommendation method, apparatus, medium, device, and product, relating to the field of artificial intelligence technology, are disclosed. The method includes: receiving user feature information and object feature information; inputting the user feature information and object feature information into a second machine learning model to obtain a recommendation result; the second machine learning model is trained by: training a first machine learning model based on a training dataset to obtain multiple training signals; and training the second machine learning model based on the training signals and the training dataset. The second machine learning model includes a main task network, an auxiliary task network, and a backbone network. The main task network is used to optimize the backbone network based on label data, and the auxiliary task network is used to optimize the backbone network based on label data and training signals. By introducing the main task network and auxiliary task network into the second machine learning model, the impact of the training of the first machine learning model on the normal use of the second machine learning model can be reduced.

Description

Technical Field

[0001] The technical solution relates to the field of artificial intelligence technology, specifically to a recommended method, apparatus, medium, equipment, and product. Background Technology

[0002] Knowledge distillation (KD) is a model compression technique that achieves model lightweighting by transferring knowledge from a large model to a smaller model. The large model can also be called a general model or teacher model, while the small model can be called a vertical model or student model. Currently, in the process of knowledge distillation, the student model is easily influenced by the teacher model, resulting in a low knowledge transfer rate. Summary of the Invention

[0003] This summary section is provided to briefly introduce the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0004] Firstly, a recommendation method is provided, including: The system receives user feature information and object feature information, inputs the user feature information and object feature information into a second machine learning model, and obtains recommendation results. The second machine learning model is trained in the following manner: The first machine learning model is trained based on the training dataset to obtain multiple training signals. The training dataset includes training data and label data. The second machine learning model is trained based on the training signal and the training dataset. The second machine learning model includes a main task network, an auxiliary task network, and a backbone network. The main task network is used to optimize the backbone network based on the label data, and the auxiliary task network is used to optimize the backbone network based on the label data and the training signal.

[0005] Secondly, a recommended device is provided, comprising: The recommendation module is configured to receive user feature information and object feature information, input the user feature information and the object feature information into a second machine learning model, and obtain recommendation results; the second machine learning model is trained in the following way: The training module is configured to train a first machine learning model based on a training dataset to obtain multiple training signals, wherein the training dataset includes training data and label data; and to train a second machine learning model based on the training signals and the training dataset, wherein the second machine learning model includes a main task network, an auxiliary task network, and a backbone network, wherein the main task network is used to optimize the backbone network based on the label data, and the auxiliary task network is used to optimize the backbone network based on the label data and the training signals.

[0006] Thirdly, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processing device, implements the steps of the method described in the first aspect.

[0007] Fourthly, an electronic device is provided, comprising: A storage device on which computer programs are stored; A processing device for executing the computer program in the storage device to implement the steps of the method described in the first aspect.

[0008] Fifthly, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0009] Through the above technical solution, upon receiving user feature information and object feature information, the pre-trained second machine learning model can obtain accurate recommendation results. The training process of the second machine learning model includes: training the first machine learning model based on the training dataset to obtain multiple training signals; and combining the training signals output by the first machine learning model with the training dataset to achieve efficient training of the second machine learning model. The second machine learning model includes a main task network, an auxiliary task network, and a backbone network. The main task network is used to optimize the backbone network based on label data, and the auxiliary task network is used to optimize the backbone network based on label data and training signals. Since the second machine learning model adopts a decoupled architecture, when training the second machine learning model using training signals and training datasets, on the one hand, the second machine learning model can predict the conversion probability of online services based on the main task network and optimize the backbone network in reverse; on the other hand, the second machine learning model can continuously optimize the backbone network based on the auxiliary task network, thus improving the knowledge transfer rate.

[0010] Other features and advantages of the technical solution will be described in detail in the following detailed implementation section. Attached Figure Description

[0011] The above and other features, advantages, and aspects of the technical solution will become more apparent when taken in conjunction with the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale. In the drawings: Figure 1 This is a flowchart illustrating a recommended method according to an exemplary embodiment.

[0012] Figure 2 This is an example diagram illustrating the overall framework of streaming online knowledge distillation in a recommendation method according to an exemplary embodiment.

[0013] Figure 3 This is an example diagram illustrating batch processing and stream processing in a recommended method according to an exemplary embodiment.

[0014] Figure 4 This is an example diagram illustrating the data access of a first machine learning model and a second machine learning model in a batch processing and stream processing process according to an exemplary embodiment of a recommended method.

[0015] Figure 5 This is a block diagram illustrating a recommended apparatus according to an exemplary embodiment.

[0016] Figure 6 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment. Detailed Implementation

[0017] The technical solution will now be described in more detail with reference to the accompanying drawings. Although certain scenarios are shown in the drawings, it should be understood that the technical solution can be implemented in various forms and should not be construed as limited to the scenarios described herein. Rather, these scenarios are provided to provide a more thorough and complete understanding of the technical solution. It should be understood that the accompanying drawings and the scenarios described are for illustrative purposes only and are not intended to limit the scope of protection of the technical solution.

[0018] It should be understood that the steps described in the method implementation may be performed in different orders and / or in parallel. Furthermore, the method implementation may include additional steps and / or omit the steps shown. The scope of the technical solution is not limited in this respect.

[0019] The term "comprising" and its variations as used herein can be open-ended, meaning "including but not limited to". The term "based on" can mean "at least partially based on". The term "one case" means "at least one case"; the term "another case" means "at least one additional case"; the term "some cases" means "at least some cases". Definitions of other terms will be given in the following description.

[0020] It should be noted that the concepts of "first" and "second" mentioned here are only used to distinguish different devices, modules or units, and are not used to limit the order of the functions performed by these devices, modules or units or their interdependencies.

[0021] It should be noted that the terms "one" and "more" used here are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] The names of messages or information exchanged between the multiple devices in the implementation are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0023] It is understandable that before using the technical solutions provided here, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in accordance with relevant laws and regulations, and their authorization should be obtained through appropriate means.

[0024] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations described herein.

[0025] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0026] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of the technical solution. Other methods that comply with relevant laws and regulations may also be applied to the implementation of the technical solution.

[0027] At the same time, it is understood that the data involved in the technical solution (including but not limited to the data itself, the acquisition or use of the data) should comply with the requirements of relevant laws, regulations and related provisions.

[0028] In related technologies, knowledge distillation is mainly used to transfer knowledge from a high-performance first machine learning model (teacher model) to a lightweight second machine learning model (student model). However, knowledge distillation technology has theoretical and engineering limitations when dealing with large-scale first machine learning models and dynamic industrial environments. For example, when the first machine learning model is updated, the distillation pipeline experiences delays, the signal of the first machine learning model becomes unstable, or the signal of the first machine learning model is missing, the performance of the second machine learning model and online services will be affected, leading to a poor user experience. Furthermore, the occurrence of these problems will also cause a decrease in the distillation transfer rate.

[0029] To address the aforementioned issues, a recommendation method, apparatus, medium, device, and product are proposed. The recommendation method can be applied to the Streaming Online Distillation (SODA) framework, which can be a decoupled large-scale knowledge distillation framework designed for industrial recommendation systems. This decoupled large-scale knowledge distillation framework can include a decoupling tower architecture, which includes a main task network, an auxiliary task network, and a backbone network. During knowledge distillation, introducing this decoupling tower structure can isolate unnecessary risks arising from the updating of the first machine learning model, ensuring normal online service while guaranteeing the continuous updating and optimization of the second machine learning model.

[0030] Figure 1 This is a flowchart illustrating a recommended method according to an exemplary embodiment. For example... Figure 1 As shown, the technical solution provides a recommendation method, which can be executed by a recommendation device, which can be implemented by software and / or hardware. Figure 1 As shown, the method may include the following steps.

[0031] In step S110, user feature information and object feature information are received, and the user feature information and object feature information are input into the second machine learning model to obtain recommendation results; the second machine learning model is trained in the following way.

[0032] In one scenario, user and object feature information can be received and input into a second machine learning model to obtain recommendation results. The user and object feature information can be obtained by feature extraction from the input data. In other words, user and object feature information can be obtained by feature extraction from the input data.

[0033] For example, user feature information can be obtained by extracting the user's personal information, and object feature information can be obtained by extracting the attribute features of the item. Furthermore, the feature extraction operation can be implemented by a second machine learning model or performed by other feature extraction models; no explicit limitation is made here, and the choice can be made according to the actual situation.

[0034] In the technical solution, the recommended method can be applied to a streaming online knowledge distillation framework, the structure of which can be as follows: Figure 2 As shown, based on Figure 2 It can be seen that the streaming online knowledge distillation framework may involve a first machine learning model 210, a storage module 220, and a second machine learning model 230. For example, the first machine learning model 210 can be a teacher model, and the second machine learning model 230 can be a student model.

[0035] The core objective of the streaming online knowledge distillation framework is to decouple model size from upper bound inference cost by independently maximizing two factors: teacher gain and distillation transfer rate. In other words, teacher gain can be maximized by optimizing the first machine learning model 210, while distillation transfer rate can be maximized by optimizing the second machine learning model 230; these two optimizations can be performed independently.

[0036] The aforementioned teacher gain can refer to the original performance advantage of the optimized first machine learning model relative to the baseline second machine learning model; the distillation transfer rate can be the proportion of the performance advantage of the first machine learning model that is successfully transferred to the distilled second machine learning model. By optimizing these two factors, the overall knowledge distillation gain can be maximized.

[0037] The storage module 220 can be used to receive multiple training signals sent by the first machine learning model 210 and store these training signals. When training the second machine learning model 230, the pre-stored training signals can be read from the storage module 220.

[0038] In another scenario, after obtaining the latest second machine learning model, it can be used for item recommendation. Specifically, the second machine learning model generates a conversion probability score for each candidate item based on user and object characteristics. This conversion probability score is then used to generate recommendation results during the ranking process.

[0039] In other words, after obtaining user and item association data, feature extraction can be performed on these two datasets to obtain user and item feature information. Based on this, the user and item feature information is input into a trained second machine learning model to predict the conversion probability of each candidate item, resulting in a conversion probability score. Then, based on the conversion probability scores, multiple candidate items are ranked to obtain a recommendation result, which is then sent to the user.

[0040] In step S120, the first machine learning model is trained based on the training dataset to obtain multiple training signals.

[0041] To maximize the overall knowledge distillation gain, the scale of the first machine learning model can be comprehensively expanded (scaled) across multiple dimensions. In maximizing this gain, the application scenario (application system) of the knowledge distillation can be determined first. Based on this, the parameter expansion dimensions corresponding to that application scenario can be obtained, and the parameters of the first machine learning model can be expanded according to these dimensions. Different application scenarios for knowledge distillation may result in different parameter expansion dimensions.

[0042] As an alternative approach, recommendation systems can be expanded from three parameter dimensions: the number of dense parameters, sequence length, and the amount of training data. The number of dense parameters can be the total number of parameters in the fully connected network of the model structure. Increasing the hidden layer dimension, stacking network layers, and adding attention heads can increase the parameter size, thereby improving the model's feature representation and non-linear fitting capabilities. Sequence length can be the length of the user behavior sequence, which is the length of the user's historical behavior data received and processed by the model in the recommendation scenario. For example, extending the sequence length can expand the model's perceptual scope, enabling it to capture users' longer-term and more complex dynamic interests and preferences. The amount of training data can be the size of the sample data used to train the first machine learning model. Expanding the number of samples, enriching feature dimensions, or increasing the diversity of data scenarios can increase the amount of training data.

[0043] For example, in a recommender system, the gain of the first machine learning model can be improved by increasing the number of dense parameters, increasing the sequence length, or increasing the amount of training data. To improve the capacity and expressive power of the first machine learning model, the number of dense layer parameters can be increased, for example, by using a first architecture to expand the dense parameters. This first architecture can be a core deep learning architecture for recommendation ranking in industrial-grade intelligent recommender systems. It can utilize efficient token mixing operations as a lightweight alternative to the standard Transformer attention mechanism. Based on this design, the computational complexity can still be controlled during the expansion process. For example, the first architecture can expand the number of dense parameters of the first machine learning model from 1 billion to 7 billion, thus improving the expressive power of the first machine learning model.

[0044] Optionally, since longer user behavior sequences can provide richer historical information about users' long-term interests, the sequence length (user behavior sequence length) can be extended. To effectively utilize ultra-long sequences, a second architecture can be adopted. This second architecture can be a deep learning architecture for encoding ultra-long user behavior sequence features in industrial-grade intelligent recommendation systems. This second architecture can overcome the quadratic complexity bottleneck of the standard Transformer through strategies such as sequence compression and hybrid attention mechanisms. For example, the second architecture can support efficient end-to-end modeling of sequences up to 20,000 characters in length. Here, the efficient design of the second architecture complements that of the first architecture.

[0045] Optionally, the amount of training data can be expanded based on the optimal scaling law of computing power, which guides the training and optimization of machine learning models, especially large-scale industrial recommendation models. In other words, to ensure a balanced scaling between the number of model parameters and the amount of training data, the amount of training data for the first machine learning model can be increased. Validation confirms that the performance gains obtained by the first machine learning model through data scaling can be effectively transferred to the second machine learning model through distillation. This not only allows the second machine learning model to avoid the cost of processing a larger dataset but also allows it to benefit from more data, thus further improving the overall cost-effectiveness of the streaming online knowledge distillation framework.

[0046] In summary, before training the first machine learning model based on the training dataset, at least one of the following configuration operations can be performed on the first machine learning model: expanding the number of dense parameters, expanding the sequence length, and expanding the amount of training data. The specific number of parameters to be expanded is not explicitly limited here and can be selected according to the actual situation.

[0047] It's important to note that the knowledge gained from modifying the first machine learning model can be successfully transferred to the second. Therefore, the second machine learning model can enjoy the benefits of data gain without incurring the computational cost of processing large amounts of data. In other words, the amount of training data required to train the second machine learning model in a recommender system may be far less than that required to train the first machine learning model. This means the second machine learning model can capture most of the capabilities of a large-scale first machine learning model with less data. Furthermore, the training of the second machine learning model can still be based on the original small dataset, demonstrating that the benefits gained by the first machine learning model through data expansion can also be transferred.

[0048] As an alternative approach, for computer vision scenarios, the gain of the first machine learning model can be achieved by expanding data dimensions such as image resolution, multi-scale input, and point cloud density; for speech recognition scenarios, the gain can be achieved by expanding data dimensions such as speech sampling rate, speech duration, and bandwidth. It is evident that the expansion dimensions used to gain the first machine learning model differ for different scenarios. The following embodiments will use a recommendation system as an example to illustrate how knowledge distillation is specifically implemented.

[0049] Since the performance gains of the first machine learning model are transferred to the efficient second machine learning model, the high cost gains generated by the first machine learning model can be amortized, which makes it possible to have models with a large number of parameters and long sequence lengths without incurring a proportional online inference overhead.

[0050] In some implementations, the training dataset may include training data and label data, wherein the training data may include multiple sample data, which may include user feature information and object feature information.

[0051] For example, training data may include a user set U and an item set I. The user set U may include key information corresponding to multiple users, and the item set I includes key information for multiple items. Here, items can be commodities or promotional information related to commodities, etc. A sample consisting of a user and an item can be called a user-item pair (u, i). In the training data, one user can correspond to multiple items. For example, a first sample can be generated based on user A and item a. Similarly, a second sample can be generated based on user B and item b.

[0052] Furthermore, the same item can correspond to multiple users. For example, the first sample mentioned above can be generated based on item 'a' and user A. Similarly, a third sample can be generated based on item 'a' and user B. The specific method for generating training / sample data is not explicitly limited here and can be chosen according to the actual situation.

[0053] After obtaining the training data, feature extraction can be performed on the training data to obtain feature vector information. This feature vector information may include the aforementioned user feature information and object feature information. Optionally, the feature vector information may also include context information and historical interaction sequence information, etc.

[0054] Each of the above sample data points can correspond to a tag data point. This tag data can be the actual label of the sample data, used to characterize the user's conversion rate for a particular product. In other words, the tag data can represent the user's actions regarding the target product. These actions can include clicking, purchasing, adding to favorites, or adding to cart, etc.

[0055] Continuing with the examples above, user A and item a have a corresponding real label of 1, indicating that user A clicked on item A; user B and item b have a corresponding real label of 0, indicating that user B did not click on item A. These examples are for illustrative purposes only and are not intended to limit the actual implementation; the specific implementation should be based on the actual situation.

[0056] After obtaining the training dataset, it can be input into the first machine learning model to continuously optimize its parameters using multiple training data points and corresponding label data. During this process, the first machine learning model can generate corresponding output data based on the training data. This output data can be the prediction data output by the first machine learning model in response to the training data; this prediction data can also be called the training signal or log odds (P). T ).

[0057] The training signal can be the raw score predicted by the first machine learning model, and it can be a carrier of the core knowledge of the first machine learning model. For example, the training signal can include complex information such as user association information and product association logic learned by the first machine learning model from massive amounts of data. Here, the training signal can also be called a distilled signal.

[0058] In some implementations, after receiving the training dataset, features can be extracted from the training dataset to obtain a first feature vector. This first feature vector may include numerical features, identifier-based features, cross features, and sequence features. Based on this, the first feature vector is input into a scaled and upgraded feature encoding layer to obtain feature vector information. Here, the scaled and upgraded feature encoding layer is obtained by expanding at least one of the following: the number of dense parameters, the sequence length, and the amount of training data. On this basis, the task tower (task network) of the first machine learning model can perform prediction operations based on the feature vector information output by the feature encoding layer to obtain teacher prediction scores (Teacher logits). These teacher prediction scores are the aforementioned training signals. Subsequently, the first machine learning model can send its output prediction scores to a corresponding storage device for storage.

[0059] For example, the first machine learning model f T You can input data x ui The first machine learning model f is optimized with the true label y∈[0,1] as the target. During the forward propagation of training, T The calculated logarithmic odds P T It can be written as a training signal to an external storage medium, such as a parameter server or a cache medium.

[0060] It should be noted that a black-box distillation strategy can be used to achieve knowledge transfer between the first and second machine learning models. That is, the second machine learning model can be supervised by the training signal finally output by the first machine learning model. The black-box distillation strategy can not only guarantee distillation performance, but also achieve the optimal balance between effectiveness and efficiency.

[0061] In step S130, a second machine learning model is trained based on the training signal and the training dataset.

[0062] As described above, in improving the overall gain of knowledge transfer, one can maximize teacher gain by optimizing the first machine learning model, and another can maximize distillation transfer rate by optimizing the second machine learning model. The optimization of the first machine learning model has been described in detail in the above embodiments and will not be repeated here. This section will detail how to optimize the second machine learning model to maximize the distillation transfer rate.

[0063] In some implementations, the second machine learning model may include a main task network, an auxiliary task network, and a backbone network. The main task network can be used to optimize the backbone network based on labeled data, and the auxiliary task network can be used to optimize the backbone network based on labeled data and training signals. Therefore, the second machine learning model f... S Such as Figure 2 The decoupling tower design shown allows for the training of both the main task network and the auxiliary task network using only a portion of the training data. In other words, the second machine learning model can be split into two independent networks (towers), and these two independent networks can share a common feature processing backbone network. The independent networks are the aforementioned main task network and auxiliary task network.

[0064] Here, the main task network 231, also known as the main task tower, is primarily responsible for predicting the final conversion rate (CVR) during online service, and its optimization is mainly based on labeled data (real labels). The auxiliary task network 232, also known as the auxiliary tower, is mainly used to optimize the process by combining real labels and the training signal (log-odds signal) of the first machine learning model during distillation. The parameters of the backbone network 233 can be updated jointly by the gradients of the main task network and the auxiliary task network, thus ensuring that the knowledge learned by the second machine learning model benefits from both direct task supervision and distilled knowledge. Here, the gradient can be the gradient of the cross-entropy loss, which has higher gradient efficiency and can avoid the gradient vanishing problem.

[0065] It should be noted that, based on Figure 2 The overall architecture shown can decouple the training of the first and second machine learning models during the knowledge distillation process. This decoupling design ensures that a centrally trained first machine learning model can provide supervision signals (training signals) for multiple independent second machine learning models, thereby significantly reducing the training cost of each second machine learning model. At the same time, the framework can also be well adapted to the streaming data update characteristics inherent in recommendation systems.

[0066] In summary, to maximize the distillation mobility, the structure of the second machine learning model was improved by proposing a decoupling tower structure. This tower structure splits the network into a main task network and an auxiliary task network. The main task network is used for stable online service, while the auxiliary task network is used for receiving and processing training signals. Based on this decoupling tower structure, a key risk isolation mechanism can be established between the core recommendation task and the distillation process.

[0067] For example, the second machine learning model can be a lightweight second machine learning model f S The lightweight second machine learning model fS We can use the input data x ui The main task network and backbone network are trained with the real label y; simultaneously, the second machine learning model f S It can combine the real label y and the training signal P obtained from external storage. T Optimize its auxiliary towers and backbone network.

[0068] In some implementations, during the training of a second machine learning model based on training signals and a training dataset, features can first be extracted from the training dataset to obtain a first feature vector. This first feature vector is then input into a backbone network (feature encoding layer) to obtain a second feature vector. Here, the backbone network is used to encode the first feature vector, compressing the original high-dimensional features into a low-dimensional, efficient shared feature representation.

[0069] Based on this, the main task network optimizes the backbone network, whereby the optimization of the backbone network by the main task network is based on the second feature vector and label data. For example, a first prediction result can be obtained based on the second feature vector output by the backbone network, and a first loss information can be obtained based on the first prediction result and label data. Based on this first loss information, the parameters of the backbone network are optimized.

[0070] As described above, the main task network is primarily used to generate the final conversion rate prediction results when providing online services. It can use standard task loss, i.e., it is trained with labeled data (real labels) as the target, and the loss parameters of the main task network are isolated from the distillation process.

[0071] Optionally, after obtaining the second feature vector, the main task network can be controlled to optimize the backbone network, and the auxiliary task network can be controlled to optimize the backbone network. The optimization of the backbone network by the auxiliary task network is based on the second feature vector, label data, and training signal.

[0072] For example, a second prediction result can be obtained based on the second feature vector output by the backbone network, and a second loss information can be obtained based on the second prediction result and label data. Simultaneously, a third loss information can be obtained based on the second prediction result and training signal. On this basis, a fourth loss information is obtained by combining the second and third loss information, and the parameters of the backbone network are optimized based on the fourth loss information.

[0073] As described above, the auxiliary task network is mainly used in the knowledge distillation process. It can be trained using a combination loss. The auxiliary task network can accept the same input features (second feature vector) as the main task network, and through learning, its output is aligned with the output distribution of the first machine learning model.

[0074] Based on the above decoupled architecture, it can be ensured that even if the first machine learning model is updated, or the distillation pipeline experiences delays, or the signal of the first machine learning model is unstable or missing, the performance of the main task network of the second machine learning model and the online service will not be affected. This decoupled architecture design is crucial for ensuring the service level agreement in practical applications.

[0075] In the training of the second machine learning model for industrial recommendation, negative sampling is usually used to improve efficiency, while the first machine learning model is usually trained on full, unbiased data. The differential training of the two will result in a distribution shift in the signal used to train the first and second machine learning models in the label space.

[0076] To address the aforementioned issues, during the training of the second machine learning model based on the training signal and training dataset, the second machine learning model can first perform debiasing correction on the training signal to obtain a corrected signal. Based on this corrected signal and the training dataset, the second machine learning model is then trained. For example... Figure 2 As shown, after the second machine learning model 230 obtains the training signal from the storage module 220, it can perform debiasing processing on the training signal based on the debiasing module 234 to obtain the debiased teacher signal (training signal). Based on this, the auxiliary task network 232 can comprehensively optimize the backbone network according to the second feature vector of the backbone network, the label data, and the debiased training signal (teacher signal).

[0077] The aforementioned debiasing correction primarily adjusts the prediction results of the first machine learning model in reverse, based on the student's negative sampling rules, to adapt the teacher's output to the learning data of the second machine learning model, thereby eliminating the baseline bias between the first and second machine learning models. For example, after reading the training signal output by the first machine learning model, the debiasing module can obtain the negative sampling ratio of the second machine learning model. This negative sampling ratio can be obtained based on the rules for real-time reading of the current batch of training data from the second machine learning model. Based on this, a correction coefficient is calculated according to the negative sampling ratio, and the training signal output by the first machine learning model is debiased using this correction coefficient to obtain the corrected signal. In other words, the debiasing module can dynamically correct the training signal of the first machine learning model according to the data distribution of the second machine learning model to ensure the stability of the second machine learning model's training.

[0078] The bias correction module dynamically corrects the training signal of the first machine learning model before it is fed into the distillation process, based on the known negative sampling ratio used in the training batch of the second machine learning model. This correction aligns the training signal with the actual data distribution of the second machine learning model, resulting in more stable training and better final performance, thus ensuring a high overall knowledge distillation transfer rate to a certain extent.

[0079] In summary, to maximize the effectiveness of knowledge transfer, a robust second machine learning model with a decoupling tower structure and an integrated bias removal mechanism is introduced. This second machine learning model can employ an efficient black-box distillation algorithm.

[0080] As an alternative approach, to meet the dual requirements of fast convergence and real-time adaptability, the following can be adopted: Figure 3 The diagram illustrates a combination of batch and stream processing for knowledge distillation. This combination can be termed a hybrid batch-stream engineering pipeline, where the batch processing stage can be a batch processing stage (batch distillation stage), and the stream processing stage can be a stream processing stage (stream distillation stage). Therefore, the knowledge distillation process involved in the technical solution can include two stages: a batch processing stage (batch distillation stage) and a stream processing stage.

[0081] For example, in the initial stage of model training, the goal is to enable the second machine learning model to quickly converge to a high performance baseline. Therefore, in the batch processing stage, the first machine learning model can compute training signals from a large-scale static training corpus (training data) and write these signals to a high-capacity parameter server (PS). Based on this, the second machine learning model can efficiently read these pre-computed signals from the parameter server in large batches, using shuffling or other methods. This setup allows for rapid training of the second machine learning model using substantial computing resources, such as large batch sizes and multiple GPUs (Graphics Processing Units), thus achieving near-optimal performance in a very short time.

[0082] After initialization via batch distillation, the second machine learning model can enter the streaming stage (streaming distillation stage) to process continuously generated new user interaction data. The streaming stage has high real-time requirements. The first machine learning model can calculate training signals for the newly generated streaming data and write them to low-latency, high-speed storage media. At this point, the second machine learning model can continuously consume these newly generated signals and their corresponding real-world labels, enabling it to quickly adapt to changes in data distribution, user behavior, or item catalogs, thus maintaining its performance advantage with minimal latency.

[0083] This hybrid approach ensures that the second machine learning model is both high-performance and can respond quickly to dynamic online recommendation systems.

[0084] In one scenario, given that the training dataset is historical data, multiple training signals can be sent to the parameter server. In this case, the second machine learning model can read the training signals from the parameter server and train itself based on these signals. It's important to note that the second machine learning model does not begin training while the first machine learning model is batch-processing and outputting training signals. That is, once the first machine learning model obtains historical data and determines that the second machine learning model has not yet started training, it can send its training signals, obtained based on the historical data, to the parameter server. Here, historical data can also be referred to as existing data or historical stock data.

[0085] For example, historical data can be stored in a distributed file system (Hadoop Distributed File System, HDFS). That is, during the batch processing stage, the first machine learning model can read historical data from the distributed file system and use the historical training data as training data to train the first machine learning model. In other words, by reading the training data from the distributed file system, multiple training signals can be calculated and obtained. Based on this, the first machine learning model can send these training signals to the parameter server.

[0086] Optionally, after training on historical data is completed, the first machine learning model can continue to calculate training signals for newly added training data (streaming incremental data). During this process, it can be determined whether the second machine learning model is in the batch processing stage. If it is determined that the second machine learning model is in the batch processing stage, the first machine learning model can send its acquired training signals to both the parameter server and the cache medium. In other words, when the training dataset is streaming incremental data, and the first machine learning model is in the streaming processing stage while the second machine learning model is in the batch processing stage, multiple training signals generated by the first machine learning model can be sent to both the parameter server and the cache medium simultaneously.

[0087] For example, streaming incremental data can be stored in a designated message queue, which can be a message middleware that can adopt a storage-compute separation architecture. The underlying data persistence of the designated message queue can rely on a distributed storage system.

[0088] Optionally, in response to the training dataset being streaming incremental data and the second machine learning model being in the streaming processing stage, multiple training signals can be sent to a cache medium for storage. The second machine learning model can then read the training signals from the cache medium and train itself based on them. In other words, when both the first and second machine learning models are in the streaming training stage, if new training data is added, the first machine learning model can write its generated training signals to the cache medium, while the second machine learning model can directly read the training signals from the cache medium.

[0089] To better illustrate the above process, the following is given: Figure 4 The example diagram shown is based on Figure 4 As can be seen, in the initial stage of training, the first machine learning model can read historical data (training data) from the distributed file system. This historical data can generate multiple training signals, which the first machine learning model can then send to the parameter server (high-speed server). Subsequently, the second machine learning model can directly read the training signals from this parameter server during training.

[0090] Because the training of the second machine learning model is delayed by a specified time compared to the first machine learning model, the second machine learning model may not have started training yet when the first machine learning model completes its initial training phase. As the second machine learning model begins training, new streaming incremental data will be generated. This streaming incremental data can serve as new historical data. Therefore, when the first machine learning model obtains the training signals corresponding to these streaming incremental data, it can send them to both the cache medium and the parameter server, facilitating subsequent reading of these training signals by the second machine learning model. Based on this, when both the second and first machine learning models are in the streaming processing phase, the cache medium can act as an access intermediary. That is, the first machine learning model can write its generated training signals to the cache medium, while the second machine learning model can directly read the training signals from the cache medium and train itself based on the read training signals.

[0091] The unified hybrid batch-streaming engineering pipeline described above ensures the real-time performance of knowledge distillation. In non-stationary environments, the training signals continuously generated by the first machine learning model can offset performance drift caused by data distribution shifts. Without the rapid initialization achieved through the batch processing stage, the performance improvement of the second machine learning model is extremely slow, requiring months to approach the performance level achieved by the hybrid method (batch + streaming) within days. Therefore, the batch distillation stage is crucial for achieving practical value realization efficiency, while the streaming stage is a necessary condition for ensuring the long-term performance stability of the model.

[0092] It should be noted that the total learning loss of the second machine learning model can be the sum of the loss information of the main task network and the weighted loss of the auxiliary task network. The loss information of the main task network is greater than that of the auxiliary task network. For example, the total learning loss L = A + w*B. Here, A can be the loss information of the main task network, B can be the loss information predicted by the auxiliary task network, and w can be the weight. For example, w can be 200, which can be used to effectively balance the scale between the loss of the main task network and the loss of the auxiliary task network, thereby maximizing the transfer rate.

[0093] Optionally, a trend perception score can be obtained by analyzing streaming incremental data. This score can be based on a combination of interaction speed, novelty score, and relative performance. A higher trend perception score indicates a stronger trend of emerging content in the streaming incremental data. Based on this, the weights w corresponding to the distillation loss function can be dynamically adjusted according to the trend perception score. For example, for data with high trend perception scores, the weights w of the distillation loss function can be reduced. This reduces the guidance role of the first machine learning model for specific new content, allowing the second machine learning model to rely more on real-time user feedback and its own understanding of new trends. Conversely, for data with low trend perception scores, the weights of the distillation loss function can be increased to maintain the guidance role of the strong knowledge signals from the first machine learning model.

[0094] By systematically scaling dimensions, the performance gain of the first machine learning model can be maximized. Furthermore, through robust second machine learning model design, efficient black-box distillation algorithms, and a batch-streaming engineering pipeline, distillation mobility can be maximized. In other words, regarding the distillation algorithm, employing a practical and efficient black-box strategy allows the second machine learning model to be directly supervised using the output training signal of the first machine learning model based on cross-entropy loss, thus ensuring the algorithm's versatility and low overhead. Additionally, at the engineering framework level, a hybrid batch and streaming distillation pipeline balances the convergence speed and real-time adaptability of the first and second machine learning models, ensuring that the second machine learning model can be quickly initialized and continuously adapted to changing data distributions. Moreover, the second machine learning model, obtained through knowledge distillation, can accurately score the conversion probability of each candidate item upon receiving user and object feature information, generating a score based on this conversion probability score, which in turn generates more accurate and user-relevant recommendations.

[0095] Figure 5 This is a schematic diagram illustrating the structure of a recommended device according to an exemplary embodiment. For example... Figure 5 As shown, the technical solution provides a recommendation device 500, which may include a recommendation module 510 and a training module 520.

[0096] The recommendation module 510 is configured to receive user feature information and object feature information, input the user feature information and the object feature information into a second machine learning model, and obtain a recommendation result; the second machine learning model is trained in the following manner: The training module 520 is configured to train a first machine learning model based on a training dataset to obtain multiple training signals, wherein the training dataset includes training data and label data; and to train a second machine learning model based on the training signals and the training dataset, wherein the second machine learning model includes a main task network, an auxiliary task network and a backbone network, wherein the main task network is used to optimize the backbone network based on the label data, and the auxiliary task network is used to optimize the backbone network based on the label data and the training signals.

[0097] In some implementations, the training module 520 includes: The extraction submodule is configured to extract features from the training dataset to obtain a first feature vector. The encoding submodule is configured to perform feature encoding on the first feature vector based on the backbone network to obtain a second feature vector; The first optimization submodule is configured to control the main task network to optimize the backbone network, wherein the optimization of the backbone network by the main task network is based on the second feature vector and the label data; The second optimization submodule is configured to control the auxiliary task network to optimize the backbone network, wherein the optimization of the backbone network by the auxiliary task network is based on the second feature vector, the label data, and the training signal.

[0098] In some implementations, the first optimization submodule is further configured to obtain a first prediction result based on the second feature vector, obtain first loss information based on the first prediction result and the label data, and optimize the parameters of the backbone network based on the first loss information.

[0099] In some implementations, the second optimization submodule is further configured to obtain a second prediction result based on the second feature vector; obtain second loss information based on the second prediction result and the label data; obtain third loss information based on the second prediction result and the training signal; obtain fourth loss information by combining the second loss information and the third loss information; and optimize the parameters of the backbone network based on the fourth loss information.

[0100] In some embodiments, the recommended device 500 further includes: The sending module is configured to send the plurality of training signals to the parameter server in response to the training dataset being historical data. Training module 520 is also configured to read the training signal from the parameter server and train the second machine learning model based on the training signal.

[0101] In some implementations, the sending module is further configured to send the plurality of training signals to the parameter server and the cache medium in response to the training dataset being streaming incremental data and the second machine learning model being in the batch processing stage. Training module 520 is also configured to read the training signal from the parameter server and train the second machine learning model based on the training signal.

[0102] In some implementations, the sending module is further configured to send the plurality of training signals to the cache medium in response to the training dataset being streaming incremental data and the second machine learning model being in the streaming processing stage. Training module 520 is also configured to read the training signal from the cache medium and train the second machine learning model based on the training signal.

[0103] In some embodiments, the recommended device 500 further includes: The configuration module is configured to increase the number of dense parameters; increase the sequence length; and increase the amount of training data.

[0104] In some implementations, the training module 520 is further configured to perform debiasing correction on the training signal to obtain a corrected signal; and to train the second machine learning model based on the corrected signal and the training dataset.

[0105] Upon receiving user and object feature information, a pre-trained second machine learning model can obtain accurate recommendation results. The training process of the second machine learning model includes: training a first machine learning model based on a training dataset; obtaining multiple training signals; and then combining the training signals output by the first machine learning model with the training dataset to achieve efficient training of the second machine learning model. The second machine learning model includes a main task network, an auxiliary task network, and a backbone network. The main task network optimizes the backbone network based on label data, and the auxiliary task network optimizes the backbone network based on label data and training signals. Because the second machine learning model architecture adopts a decoupled architecture, when training the second machine learning model using training signals and training datasets, on the one hand, the second machine learning model can predict the conversion probability of online services based on the main task network and optimize the backbone network in reverse; on the other hand, the second machine learning model can continuously optimize the backbone network based on the auxiliary task network, thus improving the knowledge transfer rate. Furthermore, by introducing an innovative decoupled tower second machine learning model architecture and combining it with a bias correction mechanism, signal bias problems can be solved, ensuring service stability. Furthermore, by combining unified batch processing and flow engineering pipelines, key challenges in industrial distillation can be overcome, thereby improving the robustness, efficiency, and ultimate knowledge transfer rate of the distillation process.

[0106] The following is for reference. Figure 6 The diagram illustrates a structural schematic of an electronic device 600 suitable for implementing the above-described technical solution. The terminal device may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Personal Computers), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs (Televisions), desktop computers, etc. Figure 6 The electronic device shown is merely an example and should not be construed as limiting its functionality or scope of use.

[0107] like Figure 6As shown, electronic device 600 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage device 608 into random access memory (RAM) 603. The random access memory 603 also stores various programs and data required for the operation of electronic device 600. The processing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. An input / output (I / O) interface 605 is also connected to bus 604.

[0108] Typically, the following devices can be connected to the input / output interface 605: input devices 606 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 607 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 608 including, for example, magnetic tape, hard disk, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 An electronic device 600 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0109] In particular, depending on certain circumstances, the processes described in the flowchart above can be implemented as computer software programs. For example, a computer program product is provided, comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. This computer program can be downloaded and installed from a network via communication device 609, or installed from storage device 608, or installed from read-only memory 602. When the computer program is executed by processing device 601, it performs the functions defined in the above-described methods.

[0110] It should be noted that the aforementioned computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may 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, or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In one case, a computer-readable storage medium may 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 another case, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. The transmitted data signal 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: wires, optical fibers, RF (Radio Frequency), etc., or any suitable combination thereof.

[0111] In some implementations, the client or server can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include Local Area Networks (LANs), Wide Area Networks (WANs), the Internet (e.g., the Internet), and peer-to-peer networks (e.g., ad-hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0112] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0113] The aforementioned computer-readable medium carries one or more programs. When the electronic device executes the aforementioned one or more programs, the electronic device causes the following: it receives user feature information and object feature information, inputs the user feature information and the object feature information into a second machine learning model, and obtains a recommendation result; the second machine learning model is trained in the following manner: a first machine learning model is trained based on a training dataset to obtain multiple training signals, the training dataset including training data and label data; the second machine learning model is trained according to the training signals and the training dataset, the second machine learning model including a main task network, an auxiliary task network, and a backbone network, the main task network being used to optimize the backbone network according to the label data, and the auxiliary task network being used to optimize the backbone network according to the label data and the training signals.

[0114] Computer program code for performing the above operations can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include, but are not limited to, object-oriented programming languages, as well as conventional procedural programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0115] The flowcharts and block diagrams in the accompanying figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products under various scenarios. In this respect, 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 the 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 figures. 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 the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can 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.

[0116] The modules mentioned above can be implemented in software or hardware. The name of a module does not necessarily limit the module itself; for example, a training module can also be described as "a module for training the second machine learning model based on the training signal and the training dataset".

[0117] The functions described above can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field-Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application-Specific Standard Parts (ASSPs), Systems on Chips (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0118] In this context, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0119] The above description is merely illustrative and explains the technical principles employed. Those skilled in the art should understand that the scope of the technical solution is not limited to specific combinations of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features provided herein that have similar functions.

[0120] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. Multitasking and parallel processing may be advantageous in certain environments. Similarly, although some specific implementation details are included in the above discussion, these should not be interpreted as limitations on the scope of the technical solution. Certain features described in the context of a single example can also be implemented in combination in a single example. Conversely, various features described in the context of a single example can also be implemented individually or in any suitable sub-combination in multiple examples.

[0121] Although the technical solution has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims. Regarding the aforementioned apparatus, the specific manner in which each module performs its operation has already been described in detail in the section concerning the method, and will not be elaborated upon here.

Claims

1. A recommendation method, comprising: Receive user feature information and object feature information, input the user feature information and object feature information into the second machine learning model, and obtain recommendation results; The second machine learning model was trained in the following way: The first machine learning model is trained based on the training dataset to obtain multiple training signals. The training dataset includes training data and label data. The second machine learning model is trained based on the training signal and the training dataset. The second machine learning model includes a main task network, an auxiliary task network, and a backbone network. The main task network is used to optimize the backbone network based on the label data, and the auxiliary task network is used to optimize the backbone network based on the label data and the training signal.

2. The recommendation method according to claim 1, wherein training the second machine learning model based on the training signal and the training dataset comprises: Feature extraction is performed on the training dataset to obtain a first feature vector; Based on the backbone network, the first feature vector is encoded to obtain the second feature vector; The main task network is controlled to optimize the backbone network, and the optimization of the backbone network by the main task network is based on the second feature vector and the label data; The auxiliary task network is controlled to optimize the backbone network, and the optimization of the backbone network by the auxiliary task network is based on the second feature vector, the label data, and the training signal.

3. The recommendation method according to claim 2, wherein controlling the main task network to optimize the backbone network includes: A first prediction result is obtained based on the second feature vector, and a first loss information is obtained based on the first prediction result and the label data; The parameters of the backbone network are optimized based on the first loss information.

4. The recommendation method according to claim 2, wherein controlling the main task network to optimize the backbone network includes: A second prediction result is obtained based on the second feature vector, and a second loss information is obtained based on the second prediction result and the label data; The third loss information is obtained based on the second prediction result and the training signal; The fourth loss information is obtained by combining the second loss information and the third loss information; The parameters of the backbone network are optimized based on the fourth loss information.

5. The recommended method according to claim 1, further comprising: In response to the fact that the training dataset is historical data, the plurality of training signals are sent to the parameter server; The step of training the second machine learning model based on the training signal and the training dataset includes: The training signal is read from the parameter server, and the second machine learning model is trained based on the training signal.

6. The recommended method according to claim 5, further comprising: In response to the fact that the training dataset is streaming incremental data and the second machine learning model is in the batch processing stage, the multiple training signals are sent to the parameter server and the cache medium. The step of training the second machine learning model based on the training signal and the training dataset includes: The training signal is read from the parameter server, and the second machine learning model is trained based on the training signal.

7. The recommended method according to claim 6, further comprising: In response to the training dataset being streaming incremental data and the second machine learning model being in the streaming processing stage, the plurality of training signals are sent to the cache medium; The step of training the second machine learning model based on the training signal and the training dataset includes: The training signal is read from the cache medium, and the second machine learning model is trained based on the training signal.

8. The recommendation method according to any one of claims 1 to 7, wherein before training the first machine learning model based on the training dataset, at least one of the following configuration operations is performed on the first machine learning model: Expand the number of dense parameters; Extend sequence length; Expand the amount of training data.

9. The recommendation method according to any one of claims 1 to 7, wherein training the second machine learning model based on the training signal and the training dataset comprises: The training signal is debiased to obtain the corrected signal; The second machine learning model is trained based on the corrected signal and the training dataset.

10. A recommended device, comprising: The recommendation module is configured to receive user feature information and object feature information, input the user feature information and the object feature information into a second machine learning model, and obtain recommendation results; the second machine learning model is trained in the following way: The training module is configured to train a first machine learning model based on a training dataset to obtain multiple training signals, wherein the training dataset includes training data and label data; and to train a second machine learning model based on the training signals and the training dataset, wherein the second machine learning model includes a main task network, an auxiliary task network, and a backbone network, wherein the main task network is used to optimize the backbone network based on the label data, and the auxiliary task network is used to optimize the backbone network based on the label data and the training signals.

11. A computer-readable medium having a computer program stored thereon, wherein, When executed by a processing device, the computer program performs the steps of the method according to any one of claims 1-9.

12. An electronic device, comprising: A storage device on which computer programs are stored; A processing device for executing the computer program in the storage device to implement the steps of the method according to any one of claims 1-9.

13. A computer program product comprising a computer program, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-9.

Images