Multi-scene feature learning method and device, computer program product and electronic equipment

Abstract

The embodiment of the invention relates to the technical field of computers, in particular to a multi-scene feature learning method and device, a computer program product and electronic equipment. The method comprises the following steps: acquiring a private feature of a scene, obtaining a weight feature associated with the scene according to the private feature, and zooming a shared feature of the scene according to the weight feature to obtain a target shared feature; performing domain division on the target shared features to obtain domain features, performing scenarized coding on the domain features to obtain scenarized domain features, and adjusting the scenarized domain features based on the weight features to obtain target scenarized domain features; and inputting the target scenarized domain features into a task network to predict the target scenarized domain features to obtain feature representations of tasks corresponding to the task network in different scenes. According to the method, the difference between different scenes can be captured, and the generalization ability and the prediction accuracy in different scenes are improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology, and more specifically, relate to a multi-scenario feature learning method and apparatus, a computer program product, and an electronic device. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims, and no description herein is admitted to be prior art by inclusion in this section.

[0003] Multi-scenario joint modeling aims to enable the model to understand the common characteristics and unique features of various scenarios at the same time, achieving the effect of both integration and differentiation. Common characteristics are usually easier to learn, while the core of multi-scenario modeling is to learn the special characteristics of each scene. To this end, relevant technologies start from feature extraction (allowing the model to perceive scene information), network structure design (enhancing the influence of scene characteristics through scene sub-networks), and optimization methods (such as contrastive learning) to enable the model to perceive and learn information from each scene.

[0004] In the prior art, one method is to optimize the model by adding a scene sub-feature network to strengthen the influence of scene features on the model output. The constructed scene sub-network can be: SubNet-logit, in which the output of the scene sub-network acts on the final output and is added to the regression result of the main network, such as the auxiliary network in the STAR (Star Topology AdaptiveRecommender, multi-scenario modeling) model; it can also be SubNet-NN: the output of the scene sub-network acts on the middle layer of the main network, and may use an attention mechanism or dynamic weights, such as the M2M (Multi-Scenario Multi-Task Meta Learning) and PPNet (Parameter Personalized Net) models; it can also be SubNet-embed: the scene sub-network acts on the input layer of the main network, for example, learning the importance score of the input feature, such as EPNet in PEPNet (Parameter and Embedding Personalized Network). Another is to use a multi-network structure for modeling, such as using MMoE or PLE structure, combined with SubNet-NN network, to fuse common information between scenes and specific scene information, or STAR multi-network structure, each scene corresponds to a unique network parameter, so that the model can share common information at the bottom layer, while learning specific scene information through independent or sub-network structure at the upper layer. This combination can improve the adaptability of the model to different scenes while maintaining the sharing and transfer of knowledge between tasks. Summary of the invention

[0005] In the first approach, the three existing technical paradigms focus on the optimization of the upper-layer modules of the network, but they usually assume that different scenarios are consistent in the underlying feature representation and adopt a simple shared bottom strategy for feature embedding. However, this assumption may not always hold true in practical applications, especially when dealing with complex cross-modal and cross-media business scenarios. This may cause the model to be unable to fully capture the differences between different scenarios, which in turn causes the problem of unbalanced effects in multiple scenarios.

[0006] In the second approach, scenario-specific adjustments in multi-network structure models may require additional hyperparameter adjustments. If the scenarios vary greatly, the shared architecture may not be sufficient to capture all features, and the commonalities and features between tasks may conflict, affecting the performance of the model on certain tasks.

[0007] To this end, there is a great need for an improved multi-scenario feature learning method and device, computer program product and electronic device to provide a multi-scenario feature learning method that improves the adaptability of the overall model network to training in different scenarios and improves the balance of the effect of multi-scenario modeling.

[0008] In this context, embodiments of the present disclosure are intended to provide a multi-scenario feature learning method and apparatus, a computer program product, and an electronic device.

[0009] According to one aspect of the present disclosure, a method for learning multi-scene features is provided, comprising:

[0010] Acquire a private feature of a scene, obtain a weight feature associated with the scene according to the private feature, and scale a shared feature of the scene according to the weight feature to obtain a target shared feature;

[0011] Performing domain division on the target shared features to obtain domain features, performing scenario encoding on the domain features to obtain scenario-based domain features, and adjusting the scenario-based domain features based on the weight features to obtain target scenario-based domain features;

[0012] The target scenario domain features are input into the task network to predict the target scenario domain features, and feature representations of the tasks corresponding to the task network in different scenarios are obtained.

[0013] In an exemplary embodiment of the present disclosure, the acquiring of the private features of the scene and obtaining the weight features associated with the scene according to the private features include:

[0014] Acquire private statistical features, targets and identifiers of a scene, input the private statistical features, the targets and the identifiers into a gating network, and obtain weight features corresponding to the scene; wherein the gating network is a shallow fully connected neural network.

[0015] In an exemplary embodiment of the present disclosure, scaling the shared features of the scene according to the weight features to obtain target shared features includes:

[0016] Determine the common features of each scene, embed the common features, and obtain the shared features;

[0017] The target shared feature is obtained by multiplying the shared feature with the weight feature and concatenating the features obtained by the multiplication.

[0018] In an exemplary embodiment of the present disclosure, the domain division of the target shared features to obtain domain features, and the scenario encoding of the domain features to obtain scenario-based domain features include:

[0019] According to the feature types included in the shared features, domain division is performed on the target shared features to obtain a plurality of domain features;

[0020] The multiple domain features are respectively input into multiple expert sub-networks to obtain scenario-based domain features corresponding to the domain features, wherein the expert sub-network includes a multi-layer perceptron and an activation function.

[0021] In an exemplary embodiment of the present disclosure, the scenario domain feature can be adjusted based on the weight feature to obtain a target scenario domain feature, including:

[0022] Using the scenario-based domain feature, the domain feature, and the weight feature as inputs of a selection network, and obtaining a plurality of first scenario-based domain features through a plurality of the selection networks, wherein the selection network is a shallow perceptron neural network;

[0023] The multiple first scenario-based domain features are added together to obtain the target scenario-based domain feature corresponding to the scene.

[0024] In an exemplary embodiment of the present disclosure, inputting the target scenario domain features into the task network to predict the target scenario domain features, and obtaining feature representations of tasks corresponding to the task network in different scenarios, includes:

[0025] Determining tasks included in the scenario, and determining the number of task networks according to the tasks;

[0026] The target scenario domain features are respectively input into the plurality of task networks, the target scenario domain features are predicted, and feature representations of the tasks in different scenarios are obtained.

[0027] In an exemplary embodiment of the present disclosure, the target scenario domain features are respectively input into the plurality of task networks, the target scenario domain features are predicted, and feature representations of the tasks in different scenarios are obtained, including:

[0028] Inputting the target scenario domain feature into the task network to obtain an output feature of the target scenario domain feature in the first layer network of the task network;

[0029] Using the output features in the first layer of network and the weight features associated with the scene as inputs to the second layer of network of the task network, and using the output of the second layer of network and the weight features associated with the scene as inputs to the next layer of network;

[0030] Through each task network, feature representations of the tasks corresponding to the task network in multiple scenarios are obtained.

[0031] According to one aspect of the present disclosure, a multi-scenario feature learning device is provided, comprising:

[0032] A shared feature adjustment module, used to obtain private features of a scene, obtain weight features associated with the scene according to the private features, and scale the shared features of the scene according to the weight features to obtain target shared features;

[0033] A scenario-based feature correction module, configured to perform domain division on the target shared feature to obtain a domain feature, perform scenario-based encoding on the domain feature to obtain a scenario-based domain feature, and adjust the scenario-based domain feature based on the weight feature to obtain a target scenario-based domain feature;

[0034] The feature prediction module is used to input the target scenario domain features into the task network to predict the target scenario domain features, and obtain the feature representation of the task corresponding to the task network in different scenarios.

[0035] According to one aspect of the present disclosure, a computer storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the multi-scenario feature learning method described above is implemented.

[0036] According to one aspect of the present disclosure, there is provided an electronic device, including:

[0037] Processor; and

[0038] A memory, configured to store executable instructions of the processor;

[0039] Wherein, the processor is configured to execute any one of the multi-scene feature learning methods described above by executing the executable instructions.

[0040] According to the multi-scenario feature learning method implemented in the present disclosure, on the one hand, the shared features are scaled according to the weights of the private features to obtain the target shared features, and the private features of the underlying injection scenes of the model network are realized. The target shared features are domain-divided in the upper network to obtain domain features, and the domain features are encoded to obtain scenario-based domain features. The scenario-based domain features are adjusted based on the weight features to obtain target scenario-based domain features, and the target scenario-based domain features are predicted to obtain feature representations of the tasks in different scenarios, thereby fully utilizing the common information between different scenarios, while retaining the unique information of the scenarios, and improving the generalization ability of the model in various scenarios. On the other hand, On the one hand, after obtaining the target shared features, the target shared features are domain-divided to obtain domain features, the domain features are scenario-encoded to obtain scenario-based domain features, and the scenario-based domain features are adjusted based on the weight features to obtain target scenario-based domain features. The weight features are used to achieve fine-grained adjustment and optimization of the underlying differences of the scenes, and the adaptive fusion expression of the scenes by each network of the model is achieved, the scene adaptability is enhanced, and the balance of effects of different scenes is achieved. On the other hand, the private feature injection of the underlying scenes and the adaptive adjustment of the upper-level networks are used to achieve a more flexible and effective multi-scenario model, which reduces the redundancy of the model while maintaining the generalization ability and efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0042] Figure 1 A flowchart of a multi-scenario feature learning method according to an embodiment of the present disclosure is schematically shown;

[0043] Figure 2 A schematic diagram of a multi-scenario feature learning system according to an embodiment of the present disclosure is schematically shown;

[0044] Figure 3 A flowchart of a method for scaling a shared feature of a scene according to a weight feature to obtain a target shared feature according to an embodiment of the present disclosure is schematically shown;

[0045] Figure 4 A schematic diagram of a feature scene-based coding correction module according to an embodiment of the present disclosure is schematically shown;

[0046] Figure 5 A flowchart of a method for performing domain division on target shared features to obtain domain features, performing scenario-based encoding on the domain features to obtain scenario-based domain features according to an embodiment of the present disclosure is schematically shown;

[0047] Figure 6 A flowchart of a method for adjusting a scenario domain feature based on a weight feature to obtain a target scenario domain feature according to an embodiment of the present disclosure is schematically shown;

[0048] Figure 7 A schematic diagram of a prediction layer task control module according to an embodiment of the present disclosure is schematically shown;

[0049] Figure 8 A flowchart of a method for inputting target scenario domain features into a task network to predict the target scenario domain features and obtaining feature representations of tasks corresponding to the task network in different scenarios according to an embodiment of the present disclosure is schematically shown;

[0050] Fig. 9 A flowchart of a method for respectively inputting target scenario domain features into multiple task networks, predicting the target scenario domain features, and obtaining feature representations of tasks in different scenarios according to an embodiment of the present disclosure is schematically shown;

[0051] Fig.10 A block diagram of a multi-scenario feature learning device according to an embodiment of the present disclosure is schematically shown;

[0052] Fig.11 A schematic diagram of a computer storage medium according to an embodiment of the present disclosure is shown;

[0053] Fig.12 A block diagram of an electronic device according to a disclosed embodiment is schematically shown.

[0054] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0055] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0056] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0057] According to an embodiment of the present disclosure, a multi-scene feature learning method, a multi-scene feature learning device, a computer program product, and an electronic device are provided.

[0058] In this article, when it is necessary to understand, the terms involved are explained as follows:

[0059] Adaptive feature learning: The weights and labels of features can be dynamically adjusted according to different user behaviors and contextual information, thereby improving the accuracy and personalization of recommendations. Feature representation can be dynamically adjusted according to the characteristics of different scenarios to improve the adaptability and performance of the model in specific scenarios.

[0060] Capturing multi-scenario differences: In multi-scenario modeling, scenarios can be defined based on different dimensions, such as the user's geographic location, time, device type used, or specific business needs. User behavior patterns and preferences may vary significantly in different scenarios. Capturing these differences is crucial to improving the model's prediction accuracy and user experience.

[0061] In this document, any number of elements in the drawings is used for illustration rather than limitation, and any naming is used only for distinction and does not have any limiting meaning.

[0062] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0064] The inventors discovered that the multi-scenario modeling in the related art assumes that different scenarios are consistent in the underlying feature representation and adopts a simple shared underlying strategy for feature embedding. When dealing with complex cross-modal and cross-media business scenarios, the model is unable to fully capture the differences between different scenarios.

[0065] In view of the above, the basic idea of ​​the present disclosure is to obtain the private features of the scene, obtain the weight features associated with the scene according to the private features of the scene, scale the common features according to the weight features to obtain the target shared features, distinguish the target shared features themselves into different domain features, perform scenario encoding on the domain features to obtain scenario domain features, and adjust the scenario domain features based on the weight features to obtain the target scenario domain features. The target scenario domain features are input into the task network, the target scenario domain features are predicted, and the feature representation of the task in different scenarios is obtained.

[0066] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0067] Exemplary Methods

[0068] Combine the following Figure 1 The multi-scenario feature learning method according to an exemplary embodiment of the present disclosure is described below. Figure 1 , the multi-scene feature learning method may include the following steps:

[0069] S110, acquiring private features of a scene, obtaining weight features associated with the scene according to the private features, scaling shared features of the scene according to the weight features, and obtaining target shared features;

[0070] S120, performing domain division on the target shared feature to obtain domain features, performing scenario encoding on the domain features to obtain scenario-based domain features, and adjusting the scenario-based domain features based on the weight features to obtain target scenario-based domain features;

[0071] S130: Input the target scenario domain features into the task network to predict the target scenario domain features, and obtain feature representations of the tasks corresponding to the task network in different scenarios.

[0072] In the multi-scenario feature learning method of the embodiment of the present disclosure, on the one hand, the shared features are scaled according to the weights of the private features to obtain the target shared features, the private features of the bottom layer of the model network are injected into the scene, the target shared features are domain-divided in the upper network to obtain domain features, the domain features are encoded to obtain scenario-based domain features, and the scenario-based domain features are adjusted based on the weight features to obtain target scenario-based domain features, the target scenario-based domain features are predicted to obtain feature representations of the tasks in different scenarios, and the common information between different scenarios is fully utilized while retaining the unique information of the scene, thereby improving the generalization ability of the model in various scenarios; on the other hand, On the one hand, after obtaining the target shared features, the target shared features are domain-divided to obtain domain features, the domain features are scenario-encoded to obtain scenario-based domain features, and the scenario-based domain features are adjusted based on the weight features to obtain the target scenario-based domain features. The weight features are used to achieve fine-grained adjustment and optimization of the underlying differences of the scenes, and the adaptive fusion expression of the scenes by each network of the model is achieved, the scene adaptability is enhanced, and the balance of effects of different scenes is achieved. On the other hand, the private feature injection of the underlying scenes and the adaptive adjustment of the upper-level networks are used to achieve a more flexible and effective multi-scenario model, which reduces the redundancy of the model while maintaining the generalization ability and efficiency of the model.

[0073] refer to Figure 2 As shown, the multi-scenario learning system of the present application may include a bottom-level feature adaptive weight adjustment module 210, a feature scenario coding correction module 220, and a prediction layer task control module 230. Among them, the bottom-level feature adaptive weight adjustment module 210 adaptively adjusts the importance of the feature according to the scene indication information to adapt to the heterogeneity of different scenes, and scales the shared features according to the weight features of the scene to obtain the target shared features. The feature scenario coding correction module 220 is used to perform domain division on the target shared features to obtain domain features, input the domain features into the expert subnetwork to obtain scenario-based domain features, and correct the output scenario-based domain features to obtain target scenario-based domain features. The prediction layer task control module 230 is used to input the target scenario-based domain features into different task networks to obtain feature representations of tasks in different scenes.

[0074] The following will be combined Figure 2 The multi-scene feature learning method is further explained and illustrated.

[0075] In step S110, private features of the scene are acquired, weight features associated with the scene are obtained according to the private features, and shared features of the scene are scaled according to the weight features to obtain target shared features.

[0076] Among them, scene features include public domain features and private domain features. Public domain features are shared features of multiple scenes, and private domain features, i.e. private features, are unique features of a certain scene. For each private feature, the weight of the private feature can be calculated through the gating network. After the weight of the private feature is obtained, the shared features of the scene can be scaled according to the weight to obtain the target shared features. Among them, the gating network is a shallow fully connected neural network. The scaling ratio of each feature is calculated according to the scene indicator, so that the same feature has different expressions in each scene. This method is similar to signal debugging, ensuring that the scaling ratio is limited to [0,2] to achieve adaptive weight adjustment of features in different scenes. At the same time, the weights in the gating network allow gradient backpropagation, which can enable the scene features to be dynamically adjusted as the model training progresses to more accurately reflect their importance in different scenes.

[0077] In this example embodiment, the step of acquiring the private features of the scene and obtaining the weight features associated with the scene according to the private features includes:

[0078] Obtain private statistical features, targets, and identifiers of the scene, input the private statistical features, the targets, and the identifiers into a gated network to obtain weight features corresponding to the scene; wherein the gated network is a shallow fully connected neural network. The activation function of the gated network can also be multiplied by the output of the LeakyReLU function, weighted addition, etc., which is not specifically limited in this example embodiment.

[0079] Specifically, the private statistical features, scene targets and scene identifiers of the scene are obtained, and the private statistical features, scene targets and scene identifiers of the scene are input into the gated network to obtain the weight features corresponding to the scene. Among them, the gated network is a shallow fully connected neural network, and the expression formula of the weight feature is: α = λ*sigmoid(relu(e sf ||e t ||e si )), where e sf , e t , e si They are scene-private statistical features, scene targets, and scene identifiers.

[0080] After obtaining the weighted features, the shared features of the scene can be scaled based on the weighted features. Figure 3 As shown, scaling the shared features of the scene according to the weight features to obtain target shared features includes:

[0081] Step S310: determining the common features of each scene, embedding the common features, and obtaining the shared features;

[0082] Step S320: Multiply the shared feature by the weight feature and concatenate the multiplied features to obtain the target shared feature.

[0083] In the following, step S310 and step S320 are further explained and illustrated. Specifically, the common features of the scene are determined, all the common features are embedded to obtain the shared features of the scene, and the shared features are multiplied by the weighted features and the multiplied features are concatenated to obtain the target shared features.

[0084] The target shared feature can be expressed as Q s =[α1Q1||…||α N Q N ], where N is the number of features, α i is the weight feature.

[0085] In this example embodiment, weight features of different scenarios are determined through a gating network, thereby highlighting influential features and weakening unimportant features.

[0086] For example, in music software, scenes may include daily recommendations, private FM, search and listening to songs, etc. Different scenes include different features. Among them, in these multiple scenes, features such as song playback time and collection can be used as shared features, and other features in each scene can be used as private features.

[0087] In this example embodiment, weight features associated with the scene are obtained through private features, and shared features of the scene are scaled through the weight features, thereby achieving adaptive adjustment of the shared features and capturing the underlying differences of different scenes.

[0088] In step S120, domain division is performed on the target shared features to obtain domain features, scenario encoding is performed on the domain features to obtain scenario-based domain features, and the scenario-based domain features are adjusted based on the weight features to obtain target scenario-based domain features.

[0089] The target shared features themselves can be divided into different domain features, which can be divided into user configuration features, user behavior features, audio features, etc. In this example embodiment, the domain features are not specifically limited. The domain features are input into the expert sub-network for encoding to obtain scenario-based domain features. In the expert sub-network, the model is allowed to flexibly adapt to the complexity and differences of different domain features. Figure 4 As shown, the feature scenario encoding correction module includes an expert sub-network, a selection network and a gating network.

[0090] In this example embodiment, reference Figure 5 As shown, the target shared features are domain-divided to obtain domain features, and the domain features are scene-encoded to obtain scene-based domain features, including:

[0091] Step S510. According to the feature types included in the shared features, the target shared features are domain-divided to obtain a plurality of domain features;

[0092] Step S520. Input the multiple domain features into multiple expert sub-networks respectively to obtain scenario-based domain features corresponding to the domain features, wherein the expert sub-network includes a multi-layer perceptron and an activation function.

[0093] Below, step S510 and step S520 will be further explained and illustrated. Specifically, first, according to the feature types included in the shared features, the target shared features are domain-divided to obtain multiple domain features. In this example embodiment, the number of domain features is not specifically limited. After obtaining multiple domain features, the domain features can be respectively input into the corresponding expert sub-networks, and the scenario-based domain features corresponding to the domain features are obtained through the expert sub-networks. Among them, the expert sub-network is composed of a multi-layer perceptron and an activation function. The encoding method of the expert sub-network can also be encoded according to a shallow fully connected network, softmax function encoding, etc., which is not specifically limited in this example embodiment.

[0094] When the scenario-based domain features are obtained, the scenario-based domain features can be input into the selection network to modify the scenario-based domain features. The selection network is used to determine the weight distribution of each scenario-based domain feature between different expert networks to optimize the fusion of features. Figure 6 As shown, the scenario domain feature can be adjusted based on the weight feature to obtain the target scenario domain feature, including:

[0095] Step S610. Using the scenario-based domain feature, the domain feature and the weight feature as inputs of a selection network, and obtaining a plurality of first scenario-based domain features through a plurality of the selection networks, wherein the selection network is a shallow perceptron neural network;

[0096] Step S620: Add the multiple first scenario-based domain features to obtain the target scenario-based domain feature corresponding to the scene.

[0097] In the following, step S610 and step S620 will be further explained and illustrated. Specifically, the scenario-based domain features, the domain features obtained by the expert sub-network, and the weighted features of the scene are used as inputs of the selection network, and multiple first scenario-based domain features are obtained through multiple selection networks. After the first scenario-based domain features are obtained, the multiple first scenario-based domain features can be added to obtain the target scenario-based domain features corresponding to the scene. The target scenario-based domain features can be expressed as:

[0098]

[0099] in, represents the kth expert sub-network, K s is the sub-expert network The number of represents the output vector generated by the concatenation of a single feature domain and scene gating, which is obtained by a simple linear transformation with a softmax activation function:

[0100]

[0101] Among them, w sh x is the domain feature.

[0102] In this example embodiment, the scenario-based domain features output by the expert sub-network are corrected through domain features and weight features, which helps to transfer knowledge between different scenarios and improves the learning efficiency and effect of the model.

[0103] In step S130, the target scenario domain features are input into the task network to predict the target scenario domain features, and feature representations of the tasks corresponding to the task network in different scenarios are obtained.

[0104] refer to Figure 7 As shown, the prediction layer task control module includes a deep neural network task tower and a gating network. The deep neural network task tower is used to predict the target scenario domain features; the gating network is used to adjust the parameters of each layer of the deep neural network task tower. Among them, the number of deep neural network task towers is determined by the number of tasks, the number of gating networks and the number of network layers included in the deep neural network task tower are the same as the number of scenes. In this example embodiment, the number of scenes is not specifically limited.

[0105] In this example embodiment, reference Figure 8 As shown, the target scenario domain features are input into the task network to predict the target scenario domain features, and feature representations of the tasks corresponding to the task network in different scenarios are obtained, including:

[0106] Step S810. Determine the tasks included in the scenario, and determine the number of task networks according to the tasks;

[0107] Step S820. Input the target scenario domain features into the multiple task networks respectively, predict the target scenario domain features, and obtain feature representations of the tasks in different scenarios.

[0108] In the following, step S810 and step S820 will be further explained and illustrated. Specifically, the tasks included in the scene are determined, and the number of task networks is determined according to the number of tasks, wherein the task network is a deep neural network task tower. After determining the number of deep neural network task towers, the target scenario domain features are respectively input into a plurality of deep neural network task towers, and the target scenario domain features are predicted by the deep neural network task tower to obtain feature representations of the tasks in different scenes.

[0109] For further reference, Fig. 9As shown, the target scenario domain features are respectively input into the plurality of task networks, the target scenario domain features are predicted, and feature representations of the tasks in different scenarios are obtained, including:

[0110] Step S910. Input the target scenario domain feature into the task network to obtain the output feature of the target scenario domain feature in the first layer network of the task network;

[0111] Step S920. Use the output features in the first layer network and the weight features associated with the scene as inputs to the second layer network of the task network, and use the output of the second layer network and the weight features associated with the scene as inputs to the next layer network;

[0112] Step S930. Obtain feature representations of tasks corresponding to each task network in multiple scenarios through each task network.

[0113] Below, step S910-step S930 will be further explained and illustrated. Specifically, first, the target scenario domain features are input into the first layer of the deep neural network task tower to obtain the output features of the target scenario domain features in the first layer of the network; then, the output features of the first layer of the network and the first weight features associated with the first scene are used as the input of the second layer of the network, and the output of the second layer of the network is obtained through the second layer of the network. The output of the second layer of the network and the second weight features associated with the second scene are input into the third layer of the network to obtain the output features of the third layer of the network, and the output features of the third layer of the network and the third weight features associated with the third scene are multiplied to obtain the feature representation of the tasks corresponding to each task network in multiple scenes.

[0114] The output of each layer of the deep neural network task tower is dynamically adjusted by a gating network, which can be expressed as:

[0115]

[0116] U l+1 =f(O l W l +b l ),l∈{1,…,n}

[0117] Among them, U l (x) is the output vector of each layer of the deep neural network task tower; W l represents the weight matrix between layer l and layer l+1; b l represents the bias parameter of the lth layer; U l+1is the output of the l+1 layer network, f is the activation function, the first n-1 layers use relu as the activation function, and the last layer uses unscaled sigmoid as the activation function; the final estimation uses cross entropy as the loss function. The feature output part of the scene information does not do gradient backpropagation to prevent the underlying shared network from being affected by the upper network layer, and realizes personalized processing for specific tasks.

[0118] For example, in music software, the scenarios may include daily recommendations, private FM, search and listen to songs, etc. The tasks included in the scenarios may be completion rate, collection rate, completion rate after collection, etc. Then, the prediction layer task control module may include at least three deep neural network task towers, each of which includes at least three layers of networks, and correspondingly there are also three gated networks to calculate the weight features corresponding to each scenario.

[0119] In this example embodiment, the features of different domains are refined so that the model can combine scene information for high-order representation, implement upper-level task network hierarchical regulation, and provide finer-grained network parameters, thereby personalizing the parameters of the prediction layer task regulation module.

[0120] Exemplary Devices

[0121] After introducing the multi-scenario feature learning method of the exemplary embodiment of the present disclosure, next, refer to Fig.10 A multi-scenario feature learning device according to an exemplary embodiment of the present disclosure is described.

[0122] refer to Fig.10 As shown, the multi-scenario feature learning device 10 of the exemplary embodiment of the present disclosure may include: a shared feature adjustment module 1010, a scenario feature correction module 1020 and a feature prediction module 1030; wherein:

[0123] A shared feature adjustment module 1010 is used to obtain private features of a scene, obtain weight features associated with the scene according to the private features, and scale the shared features of the scene according to the weight features to obtain target shared features;

[0124] A scenario-based feature correction module 1020 is used to perform domain division on the target shared feature to obtain a domain feature, perform scenario-based encoding on the domain feature to obtain a scenario-based domain feature, and adjust the scenario-based domain feature based on the weight feature to obtain a target scenario-based domain feature;

[0125] The feature prediction module 1030 is used to input the target scenario domain features into the task network to predict the target scenario domain features, and obtain feature representations of the tasks corresponding to the task network in different scenarios.

[0126] According to an exemplary embodiment of the present disclosure, the shared feature adjustment module includes:

[0127] Acquire private statistical features, targets and identifiers of a scene, input the private statistical features, the targets and the identifiers into a gating network, and obtain weight features corresponding to the scene; wherein the gating network is a shallow fully connected neural network.

[0128] According to an exemplary embodiment of the present disclosure, the shared feature adjustment module includes:

[0129] A common feature determination module, used to determine the common features of each scene, embed the common features, and obtain the shared features;

[0130] The shared feature generation module is used to obtain the target shared feature by multiplying the shared feature with the weight feature and concatenating the multiplied features.

[0131] According to an exemplary embodiment of the present disclosure, the scenario-based feature correction module includes:

[0132] A domain feature division module, used for performing domain division on the target shared feature according to the feature types included in the shared feature to obtain a plurality of domain features;

[0133] The scenario-based domain feature generation module is used to input the multiple domain features into multiple expert sub-networks respectively to obtain scenario-based domain features corresponding to the domain features, wherein the expert sub-network includes a multi-layer perceptron and an activation function.

[0134] According to an exemplary embodiment of the present disclosure, the scenario-based feature correction module includes:

[0135] a first scenario-based domain feature acquisition module, configured to use the scenario-based domain feature, the domain feature, and the weight feature as inputs of a selection network, and obtain multiple first scenario-based domain features through multiple selection networks, wherein the selection network is a shallow perceptron neural network;

[0136] The target scenario domain feature generation module is used to add the multiple first scenario domain features to obtain the target scenario domain feature corresponding to the scene.

[0137] According to an exemplary embodiment of the present disclosure, the feature prediction module comprises:

[0138] A task quantity determination module, used to determine the tasks included in the scenario, and determine the number of task networks according to the tasks;

[0139] The task feature prediction module is used to input the target scenario domain features into the multiple task networks respectively, predict the target scenario domain features, and obtain the feature representation of the task in different scenarios.

[0140] According to an exemplary embodiment of the present disclosure, the feature prediction module includes:

[0141] A first output module, used to input the target scenario domain feature into the task network to obtain an output feature of the target scenario domain feature in a first layer network of the task network;

[0142] A network input feature determination module, used to use the output features in the first layer of the network and the weight features associated with the scene as inputs to the second layer of the task network, and use the output of the second layer of the network and the weight features associated with the scene as inputs to the next layer of the network;

[0143] The task feature acquisition module is used to obtain feature representations of tasks corresponding to the task network in multiple scenarios through each task network.

[0144] Since the functional modules of the multi-scenario feature learning device of the disclosed embodiment are the same as those in the disclosed embodiment of the multi-scenario feature learning method, they will not be described in detail here.

[0145] Exemplary Storage Media

[0146] After introducing the multi-scenario feature learning method and device according to the exemplary embodiment of the present disclosure, Fig.11 A computer-readable storage medium according to an exemplary embodiment of the present disclosure is described.

[0147] refer to Fig.11 As shown, a program product 1100 for implementing the above method according to an embodiment of the present disclosure is described, which can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

[0148] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0149] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0150] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the foregoing.

[0151] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0152] Exemplary Electronic Devices

[0153] After introducing the storage medium of the exemplary embodiment of the present disclosure, next, reference is made to Fig.12 An electronic device according to an exemplary embodiment of the present disclosure is described.

[0154] Fig.12The electronic device 1200 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0155] like Fig.12 As shown, the electronic device 1200 is in the form of a general computing device. The components of the electronic device 1200 may include, but are not limited to: the at least one processing unit 1210, the at least one storage unit 1220, a bus 1230 connecting different system components (including the storage unit 1220 and the processing unit 1210), and a display unit 1240.

[0156] The storage unit stores a program code, which can be executed by the processing unit 1210, so that the processing unit 1210 performs the steps according to various exemplary embodiments of the present disclosure described in the above "Exemplary Method" section of this specification. For example, the processing unit 1210 can perform the following steps: Figure 1 Steps S110 to S130 shown in FIG.

[0157] The storage unit 1220 may include a volatile storage unit, such as a random access storage unit (RAM) 12201 and / or a cache storage unit 12202 , and may further include a read-only storage unit (ROM) 12203 .

[0158] The storage unit 1220 may also include a program / utility 12204 having a set (at least one) of program modules 12205, such program modules 12205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0159] The bus 1230 may include a data bus, an address bus, and a control bus.

[0160] The electronic device 1200 may also communicate with one or more external devices 1300 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), and such communication may be performed via an input / output (I / O) interface 1250. The electronic device 1200 also includes a display unit 1240, which is connected to the input / output (I / O) interface 1250 for display. In addition, the electronic device 1200 may also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) via a network adapter 1260. As shown, the network adapter 1260 communicates with other modules of the electronic device 1200 via a bus 1230. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 1200, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0161] It should be noted that although the song relationship identification device and several modules or sub-modules of the song relationship identification device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be concretized in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided to be concretized by multiple units / modules.

[0162] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0163] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A multi-scene feature learning method, characterized in that: include: Acquire a private feature of a scene, obtain a weight feature associated with the scene according to the private feature, and scale a shared feature of the scene according to the weight feature to obtain a target shared feature; Performing domain division on the target shared features to obtain domain features, performing scenario encoding on the domain features to obtain scenario-based domain features, and adjusting the scenario-based domain features based on the weight features to obtain target scenario-based domain features; The target scenario domain features are input into the task network to predict the target scenario domain features, and feature representations of the tasks corresponding to the task network in different scenarios are obtained.

2. The multi-scene feature learning method according to claim 1, characterized in that: The acquiring of the private features of the scene, and obtaining the weight features associated with the scene according to the private features, includes: Acquire private statistical features, targets and identifiers of a scene, input the private statistical features, the targets and the identifiers into a gating network, and obtain weight features corresponding to the scene; wherein the gating network is a shallow fully connected neural network.

3. The multi-scene feature learning method according to claim 1, characterized in that: Scaling the shared features of the scene according to the weight features to obtain target shared features includes: Determine the common features of each scene, embed the common features, and obtain the shared features; The target shared feature is obtained by multiplying the shared feature with the weight feature and concatenating the features obtained by the multiplication.

4. The multi-scene feature learning method according to claim 1, characterized in that: The performing domain division on the target shared features to obtain domain features, and performing scenario-based encoding on the domain features to obtain scenario-based domain features, includes: According to the feature types included in the shared features, domain division is performed on the target shared features to obtain a plurality of domain features; The multiple domain features are respectively input into multiple expert sub-networks to obtain scenario-based domain features corresponding to the domain features, wherein the expert sub-network includes a multi-layer perceptron and an activation function.

5. The multi-scene feature learning method according to claim 4, characterized in that: The step of adjusting the scenario domain feature based on the weight feature to obtain a target scenario domain feature includes: Using the scenario-based domain feature, the domain feature, and the weight feature as inputs of a selection network, and obtaining a plurality of first scenario-based domain features through a plurality of the selection networks, wherein the selection network is a shallow perceptron neural network; The multiple first scenario-based domain features are added together to obtain the target scenario-based domain feature corresponding to the scene.

6. The multi-scene feature learning method according to claim 1, characterized in that: The step of inputting the target scenario domain features into the task network to predict the target scenario domain features, and obtaining feature representations of tasks corresponding to the task network in different scenarios, includes: Determining tasks included in the scenario, and determining the number of task networks according to the tasks; The target scenario domain features are respectively input into the plurality of task networks, the target scenario domain features are predicted, and feature representations of the tasks in different scenarios are obtained.

7. The multi-scene feature learning method according to claim 6, characterized in that: The step of inputting the target scenario domain features into the plurality of task networks respectively, predicting the target scenario domain features, and obtaining feature representations of the tasks in different scenarios includes: Inputting the target scenario domain feature into the task network to obtain an output feature of the target scenario domain feature in the first layer network of the task network; Using the output features in the first layer of network and the weight features associated with the scene as inputs to the second layer of network of the task network, and using the output of the second layer of network and the weight features associated with the scene as inputs to the next layer of network; Through each task network, feature representations of the tasks corresponding to the task network in multiple scenarios are obtained.

8. A multi-scene feature learning device, characterized in that: include: A shared feature adjustment module, used to obtain private features of a scene, obtain weight features associated with the scene according to the private features, and scale the shared features of the scene according to the weight features to obtain target shared features; A scenario-based feature correction module, configured to perform domain division on the target shared feature to obtain a domain feature, perform scenario-based encoding on the domain feature to obtain a scenario-based domain feature, and adjust the scenario-based domain feature based on the weight feature to obtain a target scenario-based domain feature; The feature prediction module is used to input the target scenario domain features into the task network to predict the target scenario domain features, and obtain the feature representation of the task corresponding to the task network in different scenarios.

9. A computer program product, comprising a computer program product, characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: processor; as well as A memory, configured to store executable instructions of the processor; The processor is configured to perform the method of any one of claims 1 to 7 by executing the executable instructions.

Images