Model training method, live broadcast room recommendation method and device, equipment and storage medium

Abstract

The invention provides a model training method, a live broadcast room recommendation method and device, equipment and a storage medium, and relates to the technical field of computers, in particular to the technical field of artificial intelligence, neural network models and model training. The specific implementation scheme is as follows: constructing a prediction model comprising a feature embedding layer and a multi-task network layer; executing first-stage training by using the source scene sample and the target scene sample, and updating parameters of the feature embedding layer and the multi-task network layer; after the first-stage training is completed, executing second-stage training by using the target scene sample; and in the second-stage training, freezing the parameters of the feature embedding layer, and only updating the parameters of the multi-task network layer to obtain a prediction model adaptive to the target scene. According to the technical scheme, decoupling of underlying feature expression learning and upper-layer business logic learning is achieved, and universality and stability of underlying features are improved.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to the fields of artificial intelligence, neural network models, and model training technology. Background Technology

[0002] With the rapid development of internet information technology, deep learning-based recommender systems have been widely applied to various online service platforms such as e-commerce, content distribution, and social networks, becoming a key technology connecting users with massive amounts of information. Recommendation models typically utilize multi-dimensional data such as users' historical behavior sequences, user profiles, and item attributes. Through an embedding layer, sparse categorical features are mapped into low-dimensional dense vectors, and then deep neural networks are used to predict key metrics such as click-through rate (CTR) and conversion rate (CVR).

[0003] In a real-world application ecosystem, a mature platform typically encompasses multiple business scenarios. For example, a platform may simultaneously contain a primary scenario (source domain) with massive traffic and rich user behavior, as well as long-tail scenarios (target domain) that are in their initial stages or serve specific niche needs. Due to the uneven distribution of user traffic, the amount of data accumulated in different scenarios often varies significantly. Summary of the Invention

[0004] This disclosure provides a model training method, a live streaming recommendation method, an apparatus, a device, and a storage medium.

[0005] According to one aspect of this disclosure, a model training method is provided, comprising: Construct a prediction model that includes a feature embedding layer and a multi-task network layer; The first stage of training is performed using source scene samples and target scene samples to update the parameters of the feature embedding layer and the multi-task network layer. After the first phase of training is completed, the second phase of training is performed using the target scene samples; In the second phase of training, the parameters of the feature embedding layer are frozen, and only the parameters of the multi-task network layer are updated to obtain a prediction model adapted to the target scenario.

[0006] According to another aspect of this disclosure, a model training apparatus is provided, comprising: Build modules are used to construct prediction models that include feature embedding layers and multi-task network layers; The first update module is used to perform the first stage of training using source scene samples and target scene samples, and update the parameters of the feature embedding layer and the multi-task network layer. The execution module is used to perform the second phase of training using target scene samples after the first phase of training is completed. The second update module is used to freeze the parameters of the feature embedding layer during the second stage of training and update only the parameters of the multi-task network layer to obtain a prediction model adapted to the target scenario.

[0007] According to another aspect of this disclosure, a live streaming room recommendation method is provided, comprising: In response to a user's browsing request in a target scenario, the system acquires the user's real-time interactive behavior characteristics and the multimodal characteristics of the candidate live streaming rooms; wherein, the multimodal characteristics include at least one of video image features, audio speech features, and text features obtained by parsing the live streaming data of the candidate live streaming rooms. The prediction model adapted to the target scene obtained by the training method provided in any embodiment of this disclosure is used to obtain the predicted probability of user behavior for candidate live streaming rooms based on real-time interactive behavior characteristics and multimodal characteristics. Based on the predicted probability of behavior, a target live stream is selected from the candidate live streams for recommendation.

[0008] According to another aspect of this disclosure, a live streaming recommendation device is provided, comprising: The feature acquisition module is used to respond to the user's browsing request in the target scenario and acquire the user's real-time interactive behavior features and the multimodal features of the candidate live streaming rooms; wherein, the multimodal features include at least one of video image features, audio speech features and text features obtained by parsing the live streaming data of the candidate live streaming rooms; The prediction module is used to obtain a prediction model adapted to the target scene using the model training method provided in any embodiment of the present disclosure, and to obtain the predicted probability of user behavior for the candidate live room based on real-time interactive behavior characteristics and multimodal characteristics. The recommendation module is used to select target live streaming rooms from candidate live streaming rooms for recommendation based on the probability of behavior prediction.

[0009] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0010] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0011] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0012] According to the technical solution disclosed herein, based on the construction of a stable underlying feature representation using multi-source data, the decoupling of underlying feature representation learning and upper-level business logic learning is achieved, thereby improving the universality and stability of the underlying features.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a schematic flowchart of a model training method provided according to an embodiment of the present disclosure; Figure 2 This is an implementation flowchart of a cross-scenario few-shot learning method provided according to an embodiment of the present disclosure; Figure 3 This is a flowchart illustrating a live streaming room recommendation method according to another embodiment of this disclosure; Figure 4 This is a schematic diagram of the structure of a model training device according to an embodiment of the present disclosure; Figure 5 This is a schematic diagram of the structure of a live streaming recommendation device according to another embodiment of the present disclosure; Figure 6 This is a block diagram of an electronic device used to implement the model training method or live streaming recommendation method of the embodiments of this disclosure. Detailed Implementation

[0015] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0016] To improve recommendation performance in data-sparse target scenarios (small sample scenarios), cross-domain recommendation and multi-task learning (MTL) have become common technical approaches in the industry. Cross-domain recommendation aims to transfer rich knowledge learned in the source domain to the target domain to alleviate the cold start problem in the target domain; multi-task learning attempts to improve the model's generalization ability by jointly training multiple related tasks and leveraging the correlation between tasks. Related techniques typically employ parameter sharing (such as sharing underlying embeddings or sharing parts of network layers) to try to mine common features between different scenarios, thereby achieving data resource complementarity.

[0017] While existing cross-domain recommendation and multi-task learning techniques have alleviated the data sparsity problem to some extent, the following core technical bottlenecks still exist when facing joint modeling of primary and secondary scenarios with vastly different sample sizes: First, the feature sharing mechanism lacks precision and struggles to balance the commonalities and unique characteristics of different scenarios.

[0018] Existing cross-domain methods typically employ coarse-grained parameter sharing strategies (e.g., fully sharing the embedding layer). When there are differences in data distribution between the source and target domains, excessive sharing can lead to the source domain's feature distribution dominating the expression of the shared space, thereby masking the unique feature patterns of the target domain and causing a "negative transfer" phenomenon. Conversely, if completely independent parameter designs are used, the knowledge of the source domain cannot be effectively utilized, resulting in insufficient training of the target domain due to insufficient samples.

[0019] Second, there is a serious gradient-dominated problem in joint training.

[0020] In end-to-end joint training, because the number of samples in the source domain is much larger than that in the target domain, the loss function and its resulting gradient are often dominated by the source domain data. This causes the update direction of the model parameters to primarily serve the optimization objective of the source domain, while the optimization requirements of the target domain are overwhelmed or ignored, resulting in the model's prediction performance in the target scenario failing to meet expectations.

[0021] Third, fine-tuning all parameters in small sample scenarios can easily lead to overfitting.

[0022] In deep recommendation models, the feature embedding layer typically accounts for the majority of the parameters. When attempting to fine-tune a model with a large number of embedding parameters using a small number of samples from the target domain, it is highly susceptible to overfitting due to insufficient data to support the search in the high-dimensional parameter space. This results in the model performing well on the training set but exhibiting a significant decrease in its generalization ability on real-world lines.

[0023] This disclosure provides a model training method. Figure 1 This is a schematic flowchart of a model training method provided according to an embodiment of this disclosure. Figure 1 As shown, the method includes the following steps: S110. Construct a prediction model that includes a feature embedding layer and a multi-task network layer.

[0024] In this embodiment of the disclosure, the prediction model can be understood as a deep learning model used to estimate the probability of a user performing a specific behavior in a specific scenario, such as a click-through rate (CTR) model or a conversion rate (CVR) model. The feature embedding layer refers to a network layer that maps high-dimensional sparse features (such as user ID, product ID) to low-dimensional dense vectors. The multi-task network layer refers to a neural network structure located above the feature embedding layer, used to process feature vectors and output multiple prediction results; it can be a multilayer perceptron (MLP), deep neural network (DNN), or other structures.

[0025] For example, in an e-commerce live streaming recommendation system, this prediction model could be a conversion rate estimation model used to predict whether users will purchase products in an immersive live streaming room.

[0026] S120. Perform the first stage of training using source scene samples and target scene samples, updating the parameters of the feature embedding layer and the multi-task network layer. The source scene samples include interactive behavior features of the target object and multimodal features representing scene content. The target scene samples also include interactive behavior features of the target object and multimodal features representing scene content. Multimodal features include at least one of the following: video image features, audio speech features, and text features obtained by parsing the live stream data of the candidate live rooms.

[0027] In this disclosure, "scenario" broadly refers to an interactive environment within an application or service platform that has an independent interactive interface, specific business logic, or specific data distribution characteristics. Different scenarios typically correspond to different user behavior patterns or intent stages. For example, a scenario can be different channels in a recommendation system, different display formats (such as list pages and detail pages), or different business modules (such as browsing areas and transaction areas).

[0028] The source scenario can be understood as a scenario with relatively abundant data and a large accumulation of user behavior data. Its characteristics are usually manifested in a large traffic base, high user interaction frequency, and rich sample accumulation, which can fully train the underlying feature representation. However, it may not be as accurate as the target scenario for certain specific conversion goals. For example, the live broadcast card display scenario in the information feed.

[0029] The target scenario can be understood as a scenario with relatively sparse data volume that serves as the primary optimization target. Its characteristics typically include relatively low traffic, belonging to long-tail businesses or in the cold start phase, and a scarce number of samples (small sample size), making it difficult to independently support the full convergence of large-scale deep models. For example, a full-screen immersive live streaming scenario where a user clicks on a card to enter.

[0030] The first training phase, known as the Update phase, aims to leverage the massive amounts of data from the source scenario to aid in learning feature representations. Specifically, it records business attribute data associated with the interactive behaviors (such as clicks, views, and dwell times) generated by the target object (e.g., a user) in the source / target scenario (e.g., a live stream). In one example, this could involve recording the target object's behavioral characteristics (e.g., user ID, historical click sequence, active time period), live stream characteristics (e.g., live stream ID, streamer ID, store level), and product characteristics (e.g., product ID, category ID, price range) to obtain the interactive behavior features.

[0031] For example, the system simultaneously reads click log data from the live streaming scene (source scene) and interaction log data from the immersive live streaming scene (target scene), mixes them, and inputs them into the model as training samples. The system then uses the backpropagation algorithm to simultaneously adjust the vector values ​​of the embedding layer and the weights of the upper network.

[0032] S130. After the first stage of training is completed, the second stage of training is performed using the target scene samples.

[0033] In this embodiment of the disclosure, the second stage of training can be referred to as the Join stage (access, adaptation stage, or network parameter update stage). In this stage, the model no longer receives samples from the source scene, but focuses on adapting to the data distribution of the target scene.

[0034] For example, after training the model on historical data for several days in the first stage, the model enters the second stage, where the system only loads real-time sample data from immersive live streaming scenarios within the last 1-2 hours for fine-tuning.

[0035] S140. In the second stage of training, the parameters of the feature embedding layer are frozen, and only the parameters of the multi-task network layer are updated to obtain a prediction model adapted to the target scene.

[0036] In this embodiment of the disclosure, freezing parameters means stopping the calculation of the gradient of this part of the parameters during the backpropagation process and keeping its value unchanged.

[0037] For example, in frameworks like PyTorch or TensorFlow, the embedding layer can be set to be non-trainable or have fixed parameters. In this case, the parameters of the top fully connected layer (DNN) are updated only based on the loss value of the immersive scene samples, while the feature representations of the lower layers remain in the state trained in the first stage.

[0038] According to the scheme of this disclosure embodiment, by adopting a two-stage decoupled training strategy, the feature representation is fully trained using big data from the source scene in the first stage, and the overfitting or feature shift caused by small samples is prevented by freezing the embedding layer in the second stage. This solves the problem of insufficient model training due to the scarcity of samples in immersive scenes, thereby significantly improving the prediction accuracy and convergence speed of the target scene model without increasing the annotation cost.

[0039] In one possible implementation, S110 constructs a prediction model that includes a feature embedding layer and a multi-task network layer, specifically including the following steps: S111, the feature embedding layer is split into target scene-specific embedding, cross-scene shared embedding, and source scene-specific embedding.

[0040] In this embodiment of the disclosure, the feature embedding split design refers to logically dividing the vector space corresponding to the same feature ID into three independent parts. The target scene-specific embedding is used to represent feature attributes that are only effective in the target scene. The cross-scene shared embedding is used to represent common attributes shared in both scenes. The source scene-specific embedding is used to represent feature attributes that are only effective in the source scene.

[0041] For example, for the same live room ID, its embedding vector is split into three segments, each storing different semantic information.

[0042] S112. Set the target scene-exclusive embedding, cross-scene shared embedding, and source scene-exclusive embedding to have the same feature dimensions.

[0043] In this embodiment of the disclosure, the feature dimension refers to the length of the embedding vector. Setting the same dimension is to ensure the balance of parameters and the standardization of calculation, that is, to adopt an equivalent substitution strategy.

[0044] For example, if the original model's embedding has a total dimension of 8, this solution can expand the embedding to 12 dimensions, where the target scene-specific embedding is 4-dimensional, the cross-scene shared embedding is 4-dimensional, and the source scene-specific embedding is 4-dimensional.

[0045] According to the scheme of this disclosure embodiment, by adopting a three-stage embedding structure design of "exclusive-shared-exclusive", the problem of loss of scene characteristics caused by traditional coarse-grained sharing is improved, thereby improving the accuracy of cross-scene common features while retaining the specificity of each scene.

[0046] In one possible implementation, S120 performs the first stage of training using source scene samples and target scene samples, updating the parameters of the feature embedding layer and the multi-task network layer, specifically including the following steps: S121. When the current training sample is the source scene sample, the source scene-specific embedding and the cross-scene shared embedding are spliced ​​together as the network input.

[0047] In this embodiment of the disclosure, concatenation refers to joining two vectors along a dimensional direction. When the sample comes from the source scene, the model only activates the shared portion and the source scene-specific portion, while masking the target scene-specific portion.

[0048] For example, when a sample of a live streaming scene is input, the 4D live-specific embedding and the 4D shared embedding are extracted and concatenated into an 8D vector, which is then input into the multi-task network layer.

[0049] S122. When the current training sample is the target scene sample, the target scene-specific embedding and the cross-scene shared embedding are concatenated as network input.

[0050] For example, when inputting a sample of an immersive live streaming scene, the 4D immersive exclusive embedding and the 4D shared embedding are extracted and concatenated into an 8D vector as input.

[0051] S123. Based on network input, update the parameters of cross-scene shared embedding, corresponding exclusive embedding, and multi-task network layer simultaneously through backpropagation.

[0052] For example, based on the prediction error obtained by the network calculation of the concatenated vector, the embedding vector values ​​and network layer weights involved in this calculation are corrected simultaneously using the gradient descent algorithm.

[0053] According to the solution of this disclosure embodiment, by adopting the method of dynamically assembling input vectors according to sample sources, the model input dimension is consistent in different scenarios (e.g., all are 8-dimensional), which solves the problem that heterogeneous scenario data is difficult to directly mix for training, thereby improving the efficiency of learning dual-scenario knowledge at the same time using the same network structure.

[0054] In one possible implementation, S123 updates the parameters of the cross-scene shared embedding, the corresponding unique embedding, and the multi-task network layer simultaneously through backpropagation, specifically including the following steps: S1231. Update the parameters of the cross-scene shared embedding based on the joint gradient generated by the source scene samples and the target scene samples.

[0055] In this embodiment of the disclosure, the joint gradient refers to the superposition of gradient information from different data sources. Since the shared embedding portion is used in both scenarios, it is simultaneously affected by the gradients from both scenarios.

[0056] For example, the shared 4D embedding vectors can learn common features between live and immersive scenes.

[0057] S1232. Update the parameters of the target scene-specific embedding based only on the gradient generated by the target scene samples.

[0058] For example, the immersive-exclusive 4D vector is updated solely by the loss of the immersive samples, unaffected by real-world data.

[0059] S1233, Update the parameters of the source scene-specific embedding based only on the gradient generated from the source scene samples.

[0060] For example, the 4D vector unique to the live sample is updated only by the loss of the live sample.

[0061] According to the scheme of this disclosure embodiment, by adopting a gradient-based differentiated update strategy, the problems of information interference and negative transfer in cross-scene learning are solved, thereby promoting the full transfer of shared features and the pure expression of unique features, and improving the robustness of feature expression.

[0062] In one possible implementation, the vector dimensions of cross-scene shared embedding, target scene-specific embedding, and source scene-specific embedding are all 4-dimensional, and all 12-dimensional embedding vectors of the feature embedding layer are updated simultaneously in the first stage of training.

[0063] In this embodiment of the disclosure, this means that in the model parameter space of the first stage, the Embedding layer has a total of 12 trainable parameters.

[0064] For example, compared to the traditional 8-dimensional model, this scheme expands the parameter space by 4 dimensions during the training phase, but can revert to 8 dimensions during inference applications through combination. It should be noted that the vector dimensions of cross-scene shared embeddings, target scene-specific embeddings, and source scene-specific embeddings can also be arbitrary.

[0065] According to the scheme of the present disclosure, by adopting a 4+4+4 dimensional configuration, the parameter scale and feature representation capability are balanced, which improves the problem that the excessive number of parameters leads to difficulty in convergence or the excessive number of parameters leads to insufficient expression, thereby helping to obtain better model performance under limited computing resources.

[0066] In one possible implementation, S140 freezes the parameters of the feature embedding layer during the second stage of training and updates only the parameters of the multi-task network layer, specifically including the following steps: S141. Obtain the target scene-specific embedding and cross-scene shared embedding corresponding to the target scene sample.

[0067] For example, in the Join phase of an immersive scene, the system only reads the 4D immersive exclusive embedding and the 4D shared embedding.

[0068] S142. Keep the parameter values ​​of target scene exclusive embedding and cross-scene shared embedding unchanged.

[0069] In this embodiment of the disclosure, keeping the parameter values ​​unchanged means that the gradients of these parameters are not calculated during backpropagation, or the gradients are set to zero.

[0070] For example, set the learning rate of the Embedding layer to 0.

[0071] S143. Embed the target scene exclusively and the cross-scene shared embed into the input multi-task network layer.

[0072] For example, these two 4-dimensional vectors can be concatenated into an 8-dimensional vector and input into an MLP network.

[0073] S144. Update the connection weights of the multi-task network layer based on the output of the multi-task network layer and the labels of the target scene samples.

[0074] For example, real conversion data generated by immersive scenarios can be used to fine-tune the weight matrix and bias terms in the MLP network.

[0075] According to the scheme of this disclosure embodiment, by adopting the strategy of freezing the bottom layer features and fine-tuning the upper layer network, the problem that small sample data can easily destroy the learned general feature distribution is avoided, thereby improving the model's adaptation speed and convergence speed in the target scene.

[0076] In one possible implementation, S120 performs the first stage of training using source scene samples and target scene samples, updating the parameters of the feature embedding layer and the multi-task network layer, specifically including the following steps: S124. From the original data stream of the source scene, select the samples in which the target behavior occurs as source scene samples.

[0077] In this embodiment of the disclosure, the target behavior refers to a salient behavior that is strongly correlated with the prediction task. For example, liking, commenting, or sharing, or clicking on a shopping cart or product card.

[0078] S125. Mix the source scene samples and the target scene samples according to the preset ratio.

[0079] In this embodiment of the disclosure, the preset ratio can be determined based on experience or hyperparameter search.

[0080] For example, the ratio of live click samples to immersive samples is approximately 5:1.

[0081] S126. Use the mixed sample data to iteratively train the prediction model to update the parameters of the feature embedding layer and the multi-task network layer.

[0082] For example, the mixed data can be shuffled and then fed into the model for training.

[0083] According to the solution of this disclosure embodiment, by adopting sample screening and ratio control, the problem that the massive noise data of the source scene may overwhelm the features of the target scene is alleviated, thereby enhancing the auxiliary role of the high-value information of the source scene on the target scene.

[0084] In one possible implementation, S120 performs the first stage of training using source scene samples and target scene samples, updating the parameters of the feature embedding layer and the multi-task network layer, specifically including the data acquisition, parsing, feature extraction, and fusion process: S1201, Data Acquisition and Interaction Behavior Characteristic Recording In response to interactive behaviors (such as clicks, views, and dwell times) generated by a target object (e.g., a user) in a live streaming room (source scene) or an immersive live streaming room (target scene), the system records the business attribute data associated with these interactive behaviors. Specifically, it records the target object's behavioral characteristics (such as user ID, historical click sequence, and active time period), live streaming room characteristics (such as live streaming room ID, streamer ID, and store level), and product characteristics (such as product ID, category ID, and price range). These features are typically sparse and discrete, used for subsequent ID mapping in the feature embedding layer.

[0085] S1202, Multimodal Analysis and Feature Encoding of Live Stream Data To capture the differences in scene content at a physical level, the system performs real-time analysis of live stream data from live or immersive live stream rooms, extracting multimodal features that characterize the scene content. This step involves the physical processing of video images, audio signals, and text data: Video image feature extraction: The system performs keyframe sampling on the live video stream (e.g., capturing one frame per second). A pre-trained convolutional neural network (CNN, such as ResNet or EfficientNet) is used to convolve the keyframe images to extract visual feature vectors representing the image's tone, texture, product outlines, and the anchor's body movements.

[0086] Audio and speech feature extraction: The system separates the audio track from the live stream and analyzes the broadcaster's speech data using automatic speech recognition (ASR) technology and acoustic models (such as VGGish). Acoustic feature vectors representing speech rate (fast / slow), pitch (high / low), volume energy, and emotional state (excited / calm) are extracted.

[0087] Text Feature Extraction: The system collects real-time bullet screen text, product descriptions, and transcripts of the host's spoken text in the live stream. Natural language processing models are used to semantically encode this text data, generating text feature vectors that represent semantic information. Text features may also include attribute features such as the live stream name, description, title, type, and category tags.

[0088] S1203, Feature Mapping and Fusion The system fuses the two types of features mentioned above to construct the network input vector: Feature mapping: The interaction behavior features (sparse ID features) obtained in S1201 are input into the feature embedding layer of the prediction model. According to the "exclusive-shared-exclusive" strategy of this embodiment, the ID features are mapped to corresponding exclusive embedding vectors and shared embedding vectors.

[0089] Feature fusion: The mapping result output from the feature embedding layer is fused with the multimodal features acquired by S1202. Fusion methods may include concatenation, element-wise addition, or weighted interaction through an attention mechanism.

[0090] Input network: The fused vector is used as the network input vector and fed into the subsequent multi-task network layers.

[0091] S1204, Model Calculation and Parameter Update Based on the network input vector, the predicted values ​​(such as click-through rate and conversion rate) for each task are calculated using a multi-task network layer. Based on the loss function value between the predicted value and the true label, the parameters of the feature embedding layer (for ID features) and the multi-task network layer (for the weights of the fused feature processing) are updated simultaneously using the backpropagation algorithm.

[0092] According to the scheme of this disclosure, by introducing multimodal feature extraction technology of video, audio and text, this scheme not only utilizes business logic features such as user ID, but also deeply processes the physical signal features of the live stream. This enables the model to perceive the essential differences in audiovisual expression between "live scenes" and "immersive scenes" (e.g., live streams have fast speech speed and the visuals are dominated by products; immersive streams have slow speech speed and the visuals are dominated by interaction), thereby significantly improving the robustness and accuracy of cross-scene transfer learning.

[0093] In one possible implementation, the multi-task network layer comprises multiple task prediction sub-networks. Updating the parameters of the feature embedding layer and the multi-task network layer in S120 or S140 specifically includes the following steps: S151. Calculate the loss function value between the predicted value output by each task prediction sub-network in the multi-task network layer and the true label.

[0094] In this embodiment of the disclosure, the task prediction subnetwork may be referred to as the Task Tower. The loss function value (Loss) is used to measure the prediction bias.

[0095] For example, calculate the MSE Loss for dwell time tasks, and the Cross-Entropy Loss for click and conversion tasks respectively.

[0096] S152. Utilize an uncertainty weighting mechanism to dynamically adjust the weights of each loss function value.

[0097] In this embodiment of the disclosure, the Uncertainty Weighted Loss (UWL) mechanism is an algorithm that adaptively learns weights based on the Bayesian uncertainty principle.

[0098] For example, by introducing learnable noise parameters, the model can automatically reduce the weights of tasks with high uncertainty (difficult to predict) during training.

[0099] S153. Based on the weighted total loss value, perform gradient descent update operation.

[0100] For example, backpropagate the weighted total loss, calculate the gradient, and update the parameters.

[0101] According to the scheme of this disclosure embodiment, by adopting an uncertainty weighting mechanism, the dominance or conflict problem caused by the inconsistency of gradient magnitudes of different tasks in multi-task learning is alleviated, thereby helping to balance the optimization of multiple prediction tasks and improve overall performance.

[0102] In one possible implementation, the multi-task network layer comprises four parallel task prediction subnetworks, each configured to perform the following prediction tasks: Predicting the duration of the target audience's stay in the live stream.

[0103] For example, predicting how many seconds a user might stay in a live stream room.

[0104] Predicting the probability of a target audience clicking on a product card.

[0105] For example, predicting whether a user will click on the product list pop-up at the bottom of the screen.

[0106] Predicting the conversion probability after the target audience clicks on the product card.

[0107] For example, predicting whether a user will complete the order and payment after clicking on a product card.

[0108] Predicting the conversion probability of target audience after watching the live stream.

[0109] For example, predicting the probability that a user will eventually make a purchase while watching a live stream (regardless of whether they click on the product card).

[0110] According to the solution of this disclosure embodiment, by constructing a multi-task network that includes full-link behavior prediction, the problem that a single task cannot fully characterize the user's deep conversion intention is improved, thereby enhancing the accuracy of conversion intention prediction in the target scenario.

[0111] This disclosure provides an implementation process for a cross-scenario few-shot learning method, which combines feature embedding splitting, multi-task network construction, and a two-stage "Update-Join" training strategy.

[0112] like Figure 2 As shown, the implementation process mainly includes the following steps: S1. Constructing the Model Architecture First, the system constructs a prediction model that includes a feature embedding layer and a multi-task network layer.

[0113] Feature embedding layer: Designed as a three-stage structure of "exclusive-shared-exclusive". Specifically, for each feature ID (such as product ID or user ID), it is mapped to three independent 4-dimensional vectors: immersive exclusive embedding, cross-scene shared embedding, and real-time exclusive embedding.

[0114] Multi-task network layer: The multi-task network layer can include an immersive task tower and a live task tower. Both task towers contain four parallel task prediction subnetworks, which are used to predict: dwell time, probability of clicking on a product card, probability of conversion after clicking on a product card, and probability of conversion after watching the live stream.

[0115] S2. Data Preparation and Preprocessing The system obtains raw log data from the source scenario (live broadcast information stream) and the target scenario (immersive live broadcast stream), respectively.

[0116] Source scene processing: For live streaming data, the system filters and retains only samples where users have made "click" actions, while removing samples that are only exposed but not clicked, in order to align the behavioral spaces of the two scenes.

[0117] Target scenario processing: For immersive live streaming data, retain all interaction samples.

[0118] Sample mixing: The processed live click samples are mixed with immersive samples according to a preset ratio (e.g., 5:1) to form the training dataset for the first stage.

[0119] S3, Execute the first phase of training (Update Stage) This stage uses mixed historical data over a long period (such as the past 7 days) to train the model, with the core objective of learning general feature representations.

[0120] Dynamic input assembly: When a live sample is read, the model concatenates "cross-scene shared embedding + live exclusive embedding" as input, and hides the immersive exclusive part.

[0121] When an immersive sample is read, the model concatenates "immersive exclusive embedding + cross-scene shared embedding" as input, and hides the live exclusive part.

[0122] Joint gradient update: The backpropagation algorithm is used to update the parameters of all 12-dimensional embedding vectors (the shared part is affected by bidirectional gradients, and the unique part is affected by unidirectional gradients) and the parameters of the multi-task network layer simultaneously using the gradients generated by the mixed samples.

[0123] In this process, the Uncertainty Weighted Lever (UWL) mechanism is used to dynamically balance the Loss weights of the four tasks.

[0124] S4. Perform the second stage of training (Join Stage) This stage uses real-time data from short periods (such as the most recent hour) to fine-tune the model. The core purpose is to adapt it to the current distribution of the target scenario and prevent overfitting.

[0125] During this phase, live scene data is no longer input; only immersive samples of the immersive scene are input. Features are extracted from the immersive samples, and then feature embedding is used to split them into 4D immersive-specific embeddings and 4D immersive live scene shared embeddings (cross-scene shared embeddings).

[0126] Freeze parameters: Set the feature embedding layer (all 12-dimensional vectors) to be untrainable, keeping the general feature representation learned in the first stage unchanged.

[0127] Network fine-tuning: The model only reads "immersive exclusive embedding + cross-scene shared embedding" as input.

[0128] Calculate the loss between the predicted results and the true labels of the immersive samples.

[0129] The weight parameters of multi-task network layers (such as fully connected layers) are updated only based on this loss.

[0130] S5, Online Reasoning and Applications The trained model is deployed into the immersive live streaming recommendation system. When a user enters the immersive live streaming room, the system extracts features in real time. The model reads the corresponding "immersive-specific embeddings" and "cross-scene shared embeddings" and concatenates them. The multi-task network outputs prediction results (such as estimated conversion rate, CVR).

[0131] The system sorts live streams based on their estimated scores and recommends high-potential live streams to users.

[0132] It should be noted that there are entity associations or behavioral link associations between the source scene and the target scene.

[0133] Entity association refers to two scenarios sharing some or all of the user ID or item ID. Behavioral link association refers to two scenarios that may have a pre-guided (e.g., clicking from the source scenario to enter the target scenario) or parallel complementary (e.g., different content sections under the same APP) relationship in terms of business.

[0134] The combination of source scene and target scene can include, but is not limited to, the following examples: Example 1: Comprehensive e-commerce recommendation scenario Source scenario: The "You May Also Like" information feed recommendation on the app's homepage. This scenario has the highest traffic, with user-generated exposure and click samples reaching hundreds of millions, primarily reflecting users' generalized interests.

[0135] Target scenarios: "New Product Launch" or "Flash Sale" channels. Due to the short launch time of new products or the narrow audience of luxury goods, the interaction samples are extremely sparse (small sample size).

[0136] Implementation logic: Use the massive click data from the homepage to train the shared embedding (learn users' basic preferences for categories), then freeze the embedding, and use a small amount of new product click data to fine-tune the network in order to predict users' purchase intentions for new products.

[0137] Example 2: Video content distribution scenario Source scenario: Short video browsing scenario. Users scroll quickly and interact frequently, allowing for rapid capture of users' instantaneous interests and activity levels.

[0138] Target scenario: Long-form videos or paid movie / TV sections. High user decision-making costs and sparse viewing behavior.

[0139] Implementation logic: Even if there is little activity in the long video area, the system can still accurately recommend movies by sharing and embedding the user's behavioral characteristics in the short video area (such as what types of clips they like to watch) and combining them with a small amount of playback history in the long video area.

[0140] Example 3: Local life service (O2O) scenario Source scenario: Food delivery recommendations. This is a high-frequency, essential scenario, accumulating a massive amount of user taste preferences and geographical location characteristics.

[0141] Target scenarios: Wedding photography or home renovation recommendations. These are extremely low-frequency scenarios, resulting in a very limited sample size.

[0142] Implementation logic: Utilize user spending power and preferences embedded in the food delivery scenario to assist in estimating the conversion rate of high-priced services in low-frequency scenarios.

[0143] According to the scheme of this disclosure embodiment, through the above-described complete implementation process, the system can fully utilize the massive click data accumulated in real-world scenarios to "warm up" feature representations, and then accurately adapt to immersive scenarios by freezing features and fine-tuning the network. This process design effectively avoids the problems of difficult cold starts and insufficient training with small samples in immersive scenarios, and improves the efficiency of cross-scenario knowledge transfer and application effects.

[0144] This disclosure also provides a live streaming room recommendation method, which is applied to the online service phase after the model training is completed. Figure 3 This is a flowchart illustrating a live-stream recommendation method according to another embodiment of this disclosure, such as... Figure 3 As shown, the method includes: S310. In response to a user's browsing request in the target scenario, acquire the user's real-time interactive behavior characteristics and the multimodal characteristics of the candidate live streaming rooms. The multimodal characteristics include at least one of video image features, audio speech features, and text features obtained by parsing the live stream data of the candidate live streaming rooms.

[0145] In this embodiment of the disclosure, when a user issues a browsing request (e.g., swiping up to switch videos) in a target scene (e.g., an immersive live streaming channel), the system acquires two types of key features in real time for model inference: Real-time interactive behavior characteristics: Obtain the user's real-time behavior sequence up to the current moment (e.g., which cards were just clicked in the live list, how long the user stayed in the current session).

[0146] Multimodal features of candidate live stream rooms: Obtain the current live stream status of candidate live stream rooms. The system parses the live stream data and extracts current video image features (such as the current frame), audio speech features (such as the current speaking speed of the anchor), and text features (such as current hot words in the bullet comments).

[0147] S320. Using the training method provided in any embodiment of this disclosure, a prediction model adapted to the target scene is obtained, and based on real-time interactive behavior characteristics and multimodal characteristics, the predicted probability of user behavior towards candidate live streaming rooms is obtained.

[0148] The prediction model adapted to the target scene, obtained by the training method described in the foregoing embodiments, is invoked, and the following steps are performed: Mapping: Input the user's real-time interaction behavior features into the model's feature embedding layer. At this point, the model reads the frozen and optimized "target scene-specific embedding" and "cross-scene shared embedding" parameter tables and outputs the corresponding feature vectors.

[0149] Fusion: The vector output by the feature embedding layer is concatenated and fused with the real-time multimodal features (dense vectors) obtained by S320 to obtain the network input vector.

[0150] Calculation: Input the network input vector into the multi-task network layer of the prediction model to calculate the predicted probability of user behavior for each candidate live room (e.g., the probability of expected viewing time > 30 seconds, the probability of expected product click).

[0151] S330. Based on the predicted probability of behavior, select the target live streaming room from the candidate live streaming rooms for recommendation.

[0152] Candidate live streams are ranked based on calculated behavioral prediction probabilities, or on a weighted comprehensive score calculated using multi-task probabilities (such as dwell time prediction, product card click prediction, click conversion prediction, and view conversion prediction). The target live stream with the highest score is selected, and its video stream is pushed to the user's client interface for display.

[0153] According to the scheme of this disclosure embodiment, the recommendation method achieves accurate capture of users' "instant interests" by real-time parsing of the multimodal features of the live stream and combining it with users' real-time interactive behavior during the inference phase. In particular, by combining a prediction model that has undergone two-stage training and adaptation, it can effectively utilize users' historical behavioral inertia in the source scene in immersive scenarios (small sample scenarios) and accurately match the content attributes of the current live stream, thereby significantly improving the click-through rate and user retention time of the recommendation.

[0154] Figure 4 This is a schematic diagram of the structure of a model training device 400 provided according to an embodiment of the present disclosure. For example... Figure 4 As shown, the device includes: Module 401 is used to build a prediction model that includes a feature embedding layer and a multi-task network layer; The first update module 402 is used to perform the first stage of training using source scene samples and target scene samples, and update the parameters of the feature embedding layer and the multi-task network layer. Execution module 403 is used to perform the second phase of training using target scene samples after the first phase of training is completed; The second update module 404 is used to freeze the parameters of the feature embedding layer in the second stage of training and update only the parameters of the multi-task network layer to obtain a prediction model adapted to the target scene.

[0155] In one possible implementation, building module 401 is used for: The feature embedding layer is divided into target scene-specific embedding, cross-scene shared embedding, and source scene-specific embedding. The target scene-exclusive embedding, cross-scene shared embedding, and source scene-exclusive embedding are set to have the same feature dimensions.

[0156] In one possible implementation, the first update module 402 is used for: Given that the current training samples are source scene samples, the source scene-specific embedding and the cross-scene shared embedding are concatenated as network input; When the current training samples are target scene samples, the target scene-specific embedding and the cross-scene shared embedding are concatenated as network input; Based on network input, the parameters of cross-scene shared embedding, corresponding exclusive embedding, and multi-task network layer are updated simultaneously through backpropagation.

[0157] In one possible implementation, the first update module 402 is used for: Based on the joint gradient generated from the source scene samples and the target scene samples, update the parameters of the cross-scene shared embedding; Update the parameters of the target scene-specific embedding based solely on the gradient generated from the target scene samples; Update the parameters of the source scene-specific embedding based solely on the gradient generated from the source scene samples.

[0158] In one possible implementation, the vector dimensions of cross-scene shared embedding, target scene-specific embedding, and source scene-specific embedding are all 4-dimensional, and all 12-dimensional embedding vectors of the feature embedding layer are updated simultaneously in the first stage of training.

[0159] In one possible implementation, the second update module 404 is used for: Obtain the target scene-specific embedding and cross-scene shared embedding corresponding to the target scene sample; Keep the parameter values ​​for target-scene exclusive embedding and cross-scene shared embedding unchanged; The target scenario-specific embedding and cross-scenario shared embedding are input into the multi-task network layer; Update the connection weights of the multi-task network layer based on the output of the multi-task network layer and the labels of the target scene samples.

[0160] In one possible implementation, the first update module 402 is used for: From the raw data stream of the source scene, samples in which the target behavior occurs are selected as source scene samples; The source scene samples and the target scene samples are mixed according to a preset ratio; The prediction model is iteratively trained using the mixed sample data to update the parameters of the feature embedding layer and the multi-task network layer.

[0161] In one possible implementation, the first update module 402 is used for: In response to the click behavior of the target object in the live broadcast room of the source scene, multi-dimensional data including the target object's behavioral characteristics, live broadcast room characteristics and product characteristics are recorded in real time; Use multidimensional data as source scene samples; The target object profile features, live broadcast room attribute features, and product information features are extracted from the source scene samples and the target scene samples, respectively, and used as the original features of the input feature embedding layer.

[0162] In one possible implementation, the multi-task network layer contains multiple task prediction subnetworks; In one possible implementation, the first update module 402 is used for: Calculate the loss function value between the predicted value output by each task prediction subnetwork in the multi-task network layer and the true label; By utilizing an uncertainty weighting mechanism, the weights of each loss function value are dynamically adjusted; Based on the weighted total loss value, perform a gradient descent update operation.

[0163] In one possible implementation, the multi-task network layer comprises four parallel task prediction subnetworks, each configured to perform the following prediction tasks: Predicting the duration of the target audience's stay in the live stream; Predicting the probability of a target audience clicking on a product card; Predicting the conversion probability after the target audience clicks on the product card; Predicting the conversion probability of target audience after watching the live stream.

[0164] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0165] Figure 5This is a structural schematic diagram of the live streaming room recommendation device 500 provided according to an embodiment of this disclosure. Figure 5 As shown, the device includes: The feature acquisition module 501 is used to respond to the user's browsing request in the target scenario and acquire the user's real-time interactive behavior features and the multimodal features of the candidate live streaming room; wherein, the multimodal features include at least one of video image features, audio speech features and text features obtained by parsing the live streaming data of the candidate live streaming room. The prediction module 502 is used to obtain the predicted probability of user behavior for candidate live streaming rooms based on the real-time interactive behavior characteristics and multimodal characteristics by using a prediction model adapted to the target scene obtained by the training method of any embodiment of the present disclosure. The recommendation module 503 is used to select target live streaming rooms from candidate live streaming rooms for recommendation based on the probability of behavior prediction.

[0166] In the technical solutions disclosed herein, the acquisition, storage, and application of any type of information, such as user personal information, comply with relevant laws and regulations and do not violate public order and good morals.

[0167] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0168] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0169] like Figure 6 As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded into random access memory (RAM) 603 from storage unit 608. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0170] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0171] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as model training methods or live streaming recommendation methods. For example, in some embodiments, the model training method or live streaming recommendation method can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the model training method or live streaming recommendation method described above can be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform a model training method or a live streaming recommendation method by any other suitable means (e.g., by means of firmware).

[0172] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0173] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0174] In the context of this disclosure, 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 fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0175] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0176] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0177] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0178] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0179] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A model training method, comprising: Construct a prediction model that includes a feature embedding layer and a multi-task network layer; The first stage of training is performed using source scene samples and target scene samples to update the parameters of the feature embedding layer and the multi-task network layer; wherein the source scene samples and the target scene samples contain interactive behavior features of the target object and multimodal features representing scene content, and the multimodal features include at least one of video image features, audio speech features and text features; After the first stage of training is completed, the second stage of training is performed using the target scene samples; In the second stage of training, the parameters of the feature embedding layer are frozen, and only the parameters of the multi-task network layer are updated to obtain a prediction model adapted to the target scene.

2. The method according to claim 1, wherein, The construction of the prediction model, which includes a feature embedding layer and a multi-task network layer, includes: The feature embedding layer is divided into target scene-specific embedding, cross-scene shared embedding, and source scene-specific embedding. The target scene-specific embedding, the cross-scene shared embedding, and the source scene-specific embedding are configured to have the same feature dimensions.

3. The method according to claim 2, wherein, The first stage of training, which utilizes source scene samples and target scene samples to update the parameters of the feature embedding layer and the multi-task network layer, includes: When the current training sample is the source scene sample, the cross-scene shared embedding and the source scene-specific embedding are concatenated as the network input; When the current training sample is the target scene sample, the target scene-specific embedding and the cross-scene shared embedding are concatenated as the network input; Based on the network input, the parameters of the cross-scene shared embedding, the corresponding exclusive embedding, and the multi-task network layer are updated simultaneously through backpropagation.

4. The method according to claim 3, wherein, The step of simultaneously updating the parameters of the cross-scenario shared embedding, the corresponding exclusive embedding, and the multi-task network layer through backpropagation includes: The parameters of the cross-scene shared embedding are updated based on the joint gradient generated by the source scene samples and the target scene samples. Update the parameters of the target scene-specific embedding based solely on the gradient generated from the target scene samples; Update the parameters of the source scene-specific embedding based solely on the gradient generated from the source scene samples.

5. The method according to claim 4, wherein, The cross-scene shared embedding, the target scene-specific embedding, and the source scene-specific embedding all have a 4-dimensional vector dimension, and all 12-dimensional embedding vectors of the feature embedding layer are updated simultaneously during the first stage of training.

6. The method according to claim 2, wherein, In the second stage of training, freezing the parameters of the feature embedding layer and updating only the parameters of the multi-task network layer includes: Obtain the target scene-specific embedding and cross-scene shared embedding corresponding to the target scene sample; Keep the parameter values ​​of the target scene-specific embedding and the cross-scene shared embedding unchanged; The target scene-specific embedding and the cross-scene shared embedding are input into the multi-task network layer; The connection weights of the multi-task network layer are updated based on the output of the multi-task network layer and the label of the target scene sample.

7. The method according to claim 1, wherein, The first stage of training, which utilizes source scene samples and target scene samples to update the parameters of the feature embedding layer and the multi-task network layer, includes: From the raw data stream of the source scene, samples in which the target behavior occurs are selected as the source scene samples; The source scene samples and the target scene samples are mixed according to a preset ratio; The prediction model is iteratively trained using the mixed sample data to update the parameters of the feature embedding layer and the multi-task network layer.

8. The method according to claim 1, wherein performing the first stage of training using source scene samples and target scene samples to update the parameters of the feature embedding layer and the multi-task network layer comprises: In response to the interactive behavior of the target object in the live broadcast room or immersive live broadcast room, the behavioral characteristics of the target object, the characteristics of the live broadcast room and the characteristics of the product are recorded as the interactive behavior characteristics; The live stream data of the live broadcast room or the immersive live broadcast room is analyzed; Keyframe images, anchor voice data, or text interaction data are extracted from the live stream data and feature-encoded respectively to obtain the video image features, the audio voice features, or the text interaction features, which are used as the multimodal features. The interactive behavior features are input into the feature embedding layer of the prediction model for feature mapping, and the mapping result is fused with the multimodal features to obtain the network input vector; Based on the network input vector, calculations are performed using the multi-task network layer of the prediction model; The parameters of the feature embedding layer and the multi-task network layer are updated based on the loss function value between the predicted value output by the prediction model and the true label.

9. The method according to claim 1, wherein, The multi-task network layer contains multiple task prediction sub-networks; Updating the parameters of the feature embedding layer and the multi-task network layer includes: Calculate the loss function value between the predicted value output by each task prediction sub-network in the multi-task network layer and the true label; The weights of each loss function value are dynamically adjusted using an uncertainty weighting mechanism. Based on the weighted total loss value, perform a gradient descent update operation.

10. The method according to claim 1, wherein, The multi-task network layer comprises four parallel task prediction sub-networks, each configured to perform the following prediction tasks: Predicting the duration of the target audience's stay in the live stream; The probability prediction of the target object clicking on the product card; The conversion probability prediction after the target object clicks on the product card; The conversion probability prediction of the target audience after watching the live stream.

11. A live streaming recommendation method, comprising: In response to a user's browsing request in a target scenario, the system acquires the user's real-time interactive behavior characteristics and the multimodal characteristics of the candidate live streaming rooms; wherein, the multimodal characteristics include at least one of video image features, audio speech features, and text features obtained by parsing the live streaming data of the candidate live streaming rooms; Using the prediction model adapted to the target scenario obtained by the training method as described in any one of claims 1-10, the predicted probability of the user's behavior towards the candidate live streaming room is obtained based on the real-time interactive behavior features and the multimodal features. Based on the predicted behavior probability, a target live stream is selected from the candidate live streams for recommendation.

12. A model training device, comprising: Build modules are used to construct prediction models that include feature embedding layers and multi-task network layers; The first update module is used to perform the first stage of training using source scene samples and target scene samples, and update the parameters of the feature embedding layer and the multi-task network layer. An execution module is used to perform a second phase of training using the target scene samples after the first phase of training is completed; The second update module is used to freeze the parameters of the feature embedding layer during the second stage of training and update only the parameters of the multi-task network layer to obtain a prediction model adapted to the target scene.

13. A live streaming recommendation device, comprising: The feature acquisition module is used to respond to a user's browsing request in a target scenario and acquire the user's real-time interactive behavior features and the multimodal features of the candidate live streaming rooms; wherein, the multimodal features include at least one of video image features, audio speech features and text features obtained by parsing the live streaming data of the candidate live streaming rooms; The prediction module is used to obtain the predicted probability of the user's behavior towards the candidate live room based on the real-time interactive behavior features and the multimodal features by using a prediction model adapted to the target scene obtained by the training method as described in any one of claims 1-10. The recommendation module is used to select a target live streaming room from the candidate live streaming rooms for recommendation based on the predicted probability of the behavior.

14. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-11.

15. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-11.

16. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-11.

Images