Recommendation processing method and device, storage medium and electronic equipment

Abstract

The invention discloses a recommendation processing method and device, a storage medium and electronic equipment, and the method comprises the steps: obtaining a user feature, an item feature and a scene context feature for a target recommendation task, inputting the user feature, the item feature and the scene context feature into a recommendation processing model, the recommendation processing model comprises scene expert networks corresponding to a plurality of recommendation project scenes under the target recommendation task, performing input feature screening on the scene context features to obtain target scene context features of each scene expert network, and performing recommendation processing on the target scene context features based on the target scene context features, the user features and the project features. And performing task recommendation by adopting the expert network of each scene to obtain a target task recommendation result, and outputting the target task recommendation result through the recommendation processing model.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a recommendation processing method, device, storage medium and electronic device. Background Art

[0002] With the development of computer technology, transaction recommendation systems based on machine learning have been widely used in various scenarios. Multi-scenario recommendation (or multi-task recommendation) is the core challenge of transaction recommendation systems. It requires the recommendation model to be able to process data features in multiple recommendation tasks at the same time and generate targeted and accurate recommendation results. Summary of the invention

[0003] This specification provides a recommendation processing method, device, storage medium and electronic device, and the technical solution is as follows:

[0004] In a first aspect, this specification provides a recommendation processing method, the method comprising:

[0005] Obtaining user features, project features, and scene context features for a target recommendation task, and inputting the user features, project features, and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task;

[0006] Performing input feature screening on the scene context features to obtain target scene context features of each scene expert network, and based on each of the target scene context features, the user features, and the project features, using each of the scene expert networks to perform task recommendation to obtain a target task recommendation result;

[0007] The target task recommendation result is outputted through the recommendation processing model.

[0008] In a feasible implementation manner, the step of filtering the scene context features as input features to obtain target scene context features of each scene expert network includes:

[0009] Determine a first random sampling mechanism corresponding to each of the scene expert networks;

[0010] Determine a first random screening probability distribution vector corresponding to the scene context feature based on the first random sampling mechanism;

[0011] The target scene context features of each of the scene expert networks are filtered from the scene context features based on the first random filtering probability distribution vector.

[0012] In a feasible implementation manner, the filtering of the target scene context features of each scene expert network from the scene context features based on the first random filtering probability distribution vector includes:

[0013] Based on the first random screening probability distribution vector and the scene context feature being input into the first model calculation formula, the target scene context feature of each scene expert network is screened by the first model calculation formula; the first model calculation formula satisfies the following formula:

[0014]

[0015] Among them, the is the target scene context feature of the scene expert network, the r 1 c is the first random screening probability distribution vector, the x c is the scene context feature, It indicates that a random sampling operation is performed on the scene context feature based on the first random screening probability distribution vector, and the squeeze() indicates a compression and cleaning operation after the random sampling operation.

[0016] In a feasible implementation, the step of performing task recommendation based on the target scene context features, the user features, and the project features using the scene expert network to obtain a target task recommendation result includes:

[0017] Generate a transaction feature combination of each scenario expert network based on the target scenario context features, the user features and the project features;

[0018] Performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result;

[0019] Based on the recommendation prediction results of each of the scenarios, task recommendation is performed to obtain a target task recommendation result, and the target task recommendation result is output through the recommendation processing model.

[0020] In a feasible implementation manner, performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result includes:

[0021] Transmitting each of the transaction feature combinations to each of the scenario expert networks respectively;

[0022] Determining the correlation weights of the user features and the project features with the target scene context features based on the scene expert network, and performing weighted processing on the target scene context features using a second model calculation formula based on the correlation weights to obtain weighted scene context features;

[0023] Combining the weighted scene context features, the user features and the project features using a third model calculation formula to obtain an expert transaction feature combination of each scene expert network;

[0024] Based on the expert transaction feature combination, a forward propagation operation is performed in the scenario expert network to perform scenario recommendation prediction to obtain an expert prediction result, and based on the expert prediction result, a scenario recommendation prediction result is obtained;

[0025] The second model calculation formula satisfies the following formula:

[0026]

[0027] Wherein, A is the weighted scene context feature, is the target scene context feature of the scene expert network, the x u is the user feature, the x m For the project features, the The weight representing the correlation between the user feature and the item feature and the target scene context feature.

[0028] The third model calculation formula satisfies the following formula:

[0029] x b ={x u ,x m ,A}

[0030] Among them, the x b is the expert transaction feature combination.

[0031] In a feasible implementation manner, performing task recommendation based on the recommendation prediction results of each of the scenarios to obtain a target task recommendation result, and outputting the target task recommendation result through the recommendation processing model, includes:

[0032] Based on the prediction results of the scenario recommendations, comprehensive processing of the prediction results is performed to obtain the target task recommendation results; or,

[0033] Determine the task scenario weight corresponding to each of the scenario recommendation prediction results, and perform weighted processing on the prediction results based on the task scenario weight and each of the scenario recommendation prediction results to obtain a target task recommendation result.

[0034] In a feasible implementation manner, the recommendation processing model further includes at least one gating network, and the determining of the task scenario weight corresponding to each of the scenario recommendation prediction results includes:

[0035] Determine a gated input feature combination for the gated network based on the user features, the project features, and the scene context features;

[0036] Based on the gated input feature combination, a forward propagation operation is performed in the gated network to predict the scene recommendation weight to obtain an expert prediction weight result, and the expert prediction weight result is used as the task scene weight corresponding to the scene recommendation prediction result.

[0037] In a feasible implementation manner, the determining of the gated input feature combination for the gated network based on the user feature, the project feature and the scene context feature includes:

[0038] Determine a second random sampling mechanism corresponding to the gating network, and determine a second random screening probability distribution vector corresponding to the scene context feature based on the second random sampling mechanism;

[0039] Filtering reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector;

[0040] A gated input feature combination of the gated network is generated based on the reference scene context feature, the user feature, and the item feature.

[0041] In a feasible implementation, the filtering of reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector includes:

[0042] Based on the second random screening probability distribution vector and the scene context feature being input into a fourth model calculation formula, the reference scene context feature of the gating network is screened by the fourth model calculation formula; the fourth model calculation formula satisfies the following formula:

[0043] Among them, the is the reference scene context feature of the scene expert network, is the second random screening probability distribution vector, the x c is the scene context feature, represents a random sampling operation performed on the scene context feature based on the second random screening probability distribution vector, and the squeeze() represents a compression and cleaning operation after the random sampling operation;

[0044] The generating of the gated input feature combination of the gated network based on the reference scene context feature, the user feature and the project feature comprises: combining the reference scene context feature, the user feature and the project feature using a fifth model calculation formula to obtain the gated input feature combination of the gated network;

[0045] The fifth model calculation formula satisfies the following formula:

[0046] Among them, the is the gated input feature combination, the x u is the user feature, the x m The project features.

[0047] In a feasible implementation manner, the step of performing weighted processing on the prediction results based on the task scenario weights and the recommended prediction results of each scenario to obtain the target task recommendation result includes:

[0048] Based on the task scenario weights and the recommended prediction results of each scenario, a sixth model calculation formula is used to perform weighted processing on the prediction results to obtain a target task recommendation result;

[0049] The sixth model calculation formula satisfies the following formula:

[0050]

[0051] Among them, the f k is the target task recommendation result, the f i (x b ) represents the scene recommendation prediction result of the i-th scene expert network, represents the task scenario weight for the i-th scenario expert network, and k represents the recommended task label.

[0052] In a second aspect, this specification provides a recommendation processing device, the device comprising:

[0053] A feature processing module, used to obtain user features, project features and scene context features for a target recommendation task, and input the user features, project features and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task;

[0054] A task recommendation module, used to perform input feature screening on the scene context features to obtain target scene context features of each scene expert network, and based on each target scene context feature, the user features and the project features, use each scene expert network to perform task recommendation to obtain a target task recommendation result;

[0055] The result output module is used to output the target task recommendation result through the recommendation processing model.

[0056] In a third aspect, the present specification provides a computer storage medium, wherein the computer storage medium stores at least one instruction, wherein the instruction is suitable for being loaded by a processor and executing the method steps of one or more embodiments of the present specification.

[0057] In a fourth aspect, the present specification provides a computer program product, wherein the computer program product stores at least one instruction, wherein the instruction is suitable for being loaded by a processor and executing the method steps of one or more embodiments of the present specification.

[0058] In a fifth aspect, the present specification provides an electronic device, which may include: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps of one or more embodiments of the present specification.

[0059] The beneficial effects brought by the technical solutions provided by some embodiments of this specification include at least:

[0060] In one or more embodiments of the present specification, the electronic device obtains user features, project features and scene context features for the target recommendation task, inputs the user features, project features and scene context features into the recommendation processing model, screens the scene context features to obtain the target scene context features of each scene expert network, and based on the target scene context features, user features and project features, uses each scene expert network to perform task recommendation to obtain the target task recommendation result, and outputs the target task recommendation result through the recommendation processing model; by screening multiple scene expert networks in combination with scene context features, input differences are generated between the scene expert networks and they focus on different scene fields based on different scene features, which effectively enhances the task adaptability and feature utilization efficiency of the recommendation model, and fundamentally alleviates the expert polarization phenomenon. Specifically, the feature screening mechanism ensures that each scene expert network focuses on the specific context information of its corresponding scene, avoiding uneven load and performance differences caused by excessive participation of some experts in training or reasoning; at the same time, targeted input and independent task processing further strengthen the differences between expert networks, so that the scene expert networks can perform their respective duties and give full play to their capabilities in different scenarios, thereby improving the stability of the overall model, resource utilization efficiency and recommendation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in this specification or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 It is a scenario diagram of a recommended processing system provided in this specification;

[0063] Figure 2 It is a flowchart of a recommended treatment method provided in this manual;

[0064] Figure 3 It is a flowchart of an input feature screening provided in this manual;

[0065] Figure 4 It is a flowchart of a task recommendation provided in this manual;

[0066] Figure 5 It is a flowchart of a recommended treatment method provided in this manual;

[0067] Figure 6 It is a scenario diagram of a recommended processing model provided in this manual;

[0068] Figure 7 This is a schematic diagram for verifying the effect of a recommended treatment method provided in this manual;

[0069] Figure 8 It is a schematic diagram of the structure of a recommended processing device provided in this manual;

[0070] Fig. 9 It is a structural schematic diagram of an electronic device provided in this manual. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in this specification to clearly and completely describe the technical solutions in this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0072] In the description of this specification, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise clearly specified and limited, "including" and "having" and any of their variations are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices. For those of ordinary skill in the art, the specific meanings of the above terms in this specification can be understood in specific circumstances. In addition, in the description of this specification, unless otherwise specified, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B, which can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are an "or" relationship.

[0073] In related technologies, in transaction recommendation scenarios such as travel industry recommendations and application promotion marketing activities, multi-scenario / multi-task modeling is usually performed. For example, under the car life recommendation transaction, there are three recommendation scenarios: car maintenance, car rental, and car series, and these three scenarios are interrelated, so a recommendation model is needed for modeling, and this recommendation model must simultaneously perform recommendation prediction processing on the three scenarios (or can be understood as three tasks) to obtain recommendation prediction results (such as scenario scoring);

[0074] When performing multi-scenario / multi-task recommendations, a very popular recommendation model architecture is to use at least a scenario expert network for each scenario, integrate the output structures of each scenario expert network, and use the integrated results for task recommendations.

[0075] For example, the recommendation model can be decomposed into multiple scene expert networks, and then a gating network decides how to assign tasks to these expert models and integrate their outputs.

[0076] However, when using a network of experts based on different scenarios for multi-task recommendations, there is a certain probability of polarization, that is, some experts gradually become more or less popular than other experts. Specifically:

[0077] Unbalanced use of scene expert networks: Over time, some scene expert networks may be selected more often because they perform better than other experts on specific tasks. This may cause these scene expert networks to be overloaded, while other scene expert networks are rarely used or even completely idle.

[0078] Widening performance gap: Due to different usage frequencies, scene expert networks with high usage rates may be further optimized to improve their performance, while scene expert networks with low usage rates may experience performance degradation due to lack of sufficient training. This performance gap will gradually widen as training progresses, leading to a decline in the overall performance of the model.

[0079] Inefficient resource utilization: When some scene expert networks are overused and others are idle, the resource utilization of the entire model will be reduced, affecting the overall computational efficiency and cost-effectiveness.

[0080] Polarization refers to the trend that some scene expert networks gradually become more or less popular than other scene expert networks during the recommendation process based on the recommendation model (especially in the model training process of the recommendation model). Specifically, over time, some experts may be selected more because they perform better than other experts on specific tasks. This may cause these experts to be overloaded, while other experts are rarely used or even completely idle.

[0081] The present specification is described in detail below with reference to specific embodiments.

[0082] See also Figure 1 , is a schematic diagram of a scenario of a recommended processing system provided in this specification. Figure 1 As shown, the recommendation processing system may include at least a client cluster and a service platform 100 .

[0083] The client cluster may include at least one client, such as Figure 1 As shown, it specifically includes client 1 corresponding to user 1, client 2 corresponding to user 2, ..., client n corresponding to user n, where n is an integer greater than 0.

[0084] Each client in the client cluster may be an electronic device with communication function, including but not limited to: wearable device, handheld device, personal computer, tablet computer, vehicle-mounted device, smart phone, computing device or other processing device connected to wireless modem, etc. Electronic devices may be called different names in different networks, such as: user equipment, access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, electronic device, wireless communication device, user agent or user device, cellular phone, cordless phone, personal digital assistant (PDA), electronic device in 5G network or future evolution network, etc.

[0085] The service platform 100 can be a separate server device, such as a rack-mounted, blade, tower, or cabinet-mounted server device, or a workstation, mainframe computer, or other hardware device with strong computing power; it can also be a server cluster composed of multiple servers, and the servers in the service cluster can be composed in a symmetrical manner, wherein each server has equivalent functions and status in the transaction link, and each server can provide services to the outside independently, and the independent service can be understood as not requiring the assistance of other servers.

[0086] In one or more embodiments of the present specification, the service platform 100 may establish a communication connection with at least one client in the client cluster, and complete the data interaction in the recommendation processing process based on the communication connection, such as online transaction data interaction, such as the service platform 100 may make recommendations to the client based on the recommendation processing method of the present specification;

[0087] It should be noted that the service platform 100 and at least one client in the client cluster establish a communication connection through a network for interactive communication, wherein the network can be a wireless network or a wired network, the wireless network includes but is not limited to a cellular network, a wireless local area network, an infrared network or a Bluetooth network, and the wired network includes but is not limited to Ethernet, a universal serial bus (USB) or a controller local area network. In one or more embodiments of the specification, technologies and / or formats including Hyper Text Mark-up Language (HTML), Extensible Markup Language (XML), etc. are used to represent data (such as a target compressed package) exchanged through a network. In addition, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), Internet Protocol Security (IPsec), etc. can also be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can also be used to replace or supplement the above data communication technologies.

[0088] The recommendation processing system embodiment provided in this specification and the recommendation processing method in one or more embodiments belong to the same concept. The execution subject corresponding to the recommendation processing method involved in one or more embodiments of the specification can be the above-mentioned service platform 100; the execution subject corresponding to the recommendation processing method involved in one or more embodiments of the specification can also be the electronic device corresponding to the client, which is determined based on the actual application environment. The implementation process of the recommendation processing system embodiment can be detailed in the following method embodiment, which will not be repeated here.

[0089] based on Figure 1 The scenario diagram is shown, and the recommended processing method provided by one or more embodiments of this specification is introduced in detail below.

[0090] See also Figure 2 , is a flowchart of a recommendation processing method for one or more embodiments of this specification, which can be implemented by a computer program and can be run on a recommendation processing device based on the von Neumann system. The computer program can be integrated into an application or run as an independent tool application. The recommendation processing device can be an electronic device.

[0091] It should be noted that the recommendation processing method involved in this application specification can be applied to the model training stage of the recommendation processing model, and can also be applied to the model application stage of the recommendation processing model. It is specifically set based on the actual application situation and is not specifically limited.

[0092] In actual recommendation application scenarios, the input features of the recommendation processing model involved in the recommendation system are usually divided into user features, project features (also called item features) and (scenario) context features.

[0093] User and item features are used as basic feature representations, while context features capture context information in a specific sub-scenario (i.e., recommended item scenario), including user and item behavior information in a specific sub-scenario. For example, in the multi-scenario recommendation involved in car life services, there may be three recommended item scenarios: car maintenance, car rental, and car series. Then the scenario context features are the user / item behavior information in these three scenarios.

[0094] User characteristics such as gender and age have the same meaning in different scenarios, but scenario context information has different meanings, indicating that the user's behavior characteristics in different scenarios are quite different. For example, in a payment platform's promotional benefits recommendation, coupons may include travel, livelihood, red envelopes, travel, movies, takeout and other coupons. These coupons have different issuance, click, and redemption characteristics in different scenarios.

[0095] Therefore, the technical concept of this solution is that if the scene expert networks corresponding to different recommendation project scenarios rely on different feature sets, this feature heterogeneity will bring unique feature information to each scene expert network, thereby making different scene expert networks focus on different domain knowledge and reduce polarization. Here, "input feature screening" (such as using the bagging network layer mechanism) is applied to randomly select a part of the contextual scene features, that is, all basic features of users and items will be sent to all expert networks, because these features are necessary for all experts. The contextual features related to the scene are randomly selected and sent to different scene expert networks after "input feature screening", so that differences are generated between the scene expert networks, and they focus on different scene fields based on different scene features.

[0096] The following is a detailed explanation of the recommended treatment methods, which include:

[0097] S102: Obtain user features, project features, and scene context features for a target recommendation task, and input the user features, project features, and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task;

[0098] User characteristics: refers to information related to the user, including static characteristics (such as age, gender, region) and dynamic behaviors (such as browsing history and purchasing preferences).

[0099] Item features: also known as item features or item features, refer to attributes related to the recommended content, such as product category, price, rating, etc.

[0100] Scenario context features: Features related to specific recommendation scenarios, capturing user behavior patterns in specific scenarios. For example, in the "travel" scenario, they may include travel time, transportation preferences, etc.

[0101] Recommendation processing model: used for recommendation processing, including at least a "scenario expert network" corresponding to multiple recommendation item scenarios under the target recommendation task, each scenario expert network focuses on the recommendation task under a specific scenario. Optionally, the recommendation processing model may also include at least one gating network;

[0102] In some embodiments, the scenario expert network and the recommendation processing model are created based on a machine learning model and trained using sample data.

[0103] Schematically, multiple groups of user features, item features, and scenario context features are extracted from data sources (such as user behavior databases, product information libraries, and context acquisition modules), and the three types of features extracted are input into the recommendation processing model. Each scenario expert network performs optimization processing for a specific recommendation item scenario, such as travel recommendations, shopping recommendations, or entertainment recommendations.

[0104] For example, assume that the model input data x is a d-dimensional vector extracted from the training data sample:

[0105] x={x u ,x m ,x c},

[0106] Among them, x u is a user feature, x m is the item feature, x c Scene context features.

[0107] S104: Screening the scene context features as input to obtain target scene context features of each scene expert network, and based on the target scene context features, the user features, and the project features, using each scene expert network to perform task recommendation to obtain a target task recommendation result;

[0108] Illustratively, the recommendation processing model applies "input feature screening" (such as using bagging random screening network layer mechanism) to randomly select a part of contextual scene features as the target scene context features for input to each expert network. That is to say, user features and item features will be sent to all scene expert networks, and the corresponding "target scene context features, the user features and the item features" will be input into the scene expert network for task recommendation to generate recommendation prediction input and obtain the target task recommendation result. In this way, differences are generated between the scene expert networks, and they focus on different scene areas based on different scene features.

[0109] S106: Outputting the target task recommendation result through the recommendation processing model.

[0110] Recommendation processing model output: Integrate the results of all scenario expert networks, or directly output the recommendation results of a certain scenario expert network.

[0111] Target task recommendation result: The final result generated by the recommendation processing model, which may include a recommendation list, rating prediction, or behavior probability. For example, recommending a set of optimal products or services, or predicting the probability of a user clicking.

[0112] For example, for a travel recommendation task, the output might be:

[0113] Target task recommendation results:

[0114] Recommended coupon: "20% bus discount coupon".

[0115] Recommended reason: "On rainy days, public transportation is more convenient during the morning rush hour."

[0116] Presentation method: The results are presented to users through the user interface or push notifications.

[0117] In one or more embodiments of the present specification, the electronic device obtains user features, project features and scene context features for the target recommendation task, inputs the user features, project features and scene context features into the recommendation processing model, screens the scene context features to obtain the target scene context features of each scene expert network, and based on the target scene context features, user features and project features, uses each scene expert network to perform task recommendation to obtain the target task recommendation result, and outputs the target task recommendation result through the recommendation processing model; by screening multiple scene expert networks in combination with scene context features, input differences are generated between the scene expert networks and they focus on different scene fields based on different scene features, which effectively enhances the task adaptability and feature utilization efficiency of the recommendation model, and fundamentally alleviates the expert polarization phenomenon. Specifically, the feature screening mechanism ensures that each scene expert network focuses on the specific context information of its corresponding scene, avoiding uneven load and performance differences caused by excessive participation of some experts in training or reasoning; at the same time, targeted input and independent task processing further strengthen the differences between expert networks, so that the scene expert networks can perform their respective duties and give full play to their capabilities in different scenarios, thereby improving the stability of the overall model, resource utilization efficiency and recommendation effect.

[0118] See also Figure 3 , Figure 3 This is a flow chart of an input feature screening process proposed in one or more embodiments of this specification. To specifically perform the input feature screening of the scene context features to obtain the target scene context features of each scene expert network, reference may be made to the following implementations:

[0119] S2002: Determine a first random sampling mechanism corresponding to each of the scene expert networks;

[0120] Scenario Expert Network: A network designed specifically for a recommendation scenario, such as a travel scenario recommendation network or a shopping scenario recommendation network.

[0121] Random sampling mechanism: A feature screening method to reduce or even avoid the polarization phenomenon of the recommendation processing model. The first random sampling mechanism determines which scene context features are retained as the expert network input for each scene expert network, thereby introducing feature diversity and enhancing the generalization ability of the model.

[0122] In an illustrative manner, an independent first random sampling mechanism is assigned to one or more scene expert networks to ensure that the focus of different scene expert networks on context features is random and differentiated. Optionally, a random sampling network layer (such as a bagging network layer) is set in the recommendation processing model to randomly select a part of the context scene features for the expert network, bringing unique feature information to each expert, thereby enabling different experts to focus on different domain knowledge and reducing polarization;

[0123] The implementation of the random sampling mechanism can be, but is not limited to, one or more of the following fittings:

[0124] Random sampling is performed using a Bernoulli distribution: each feature is sampled independently, with the probability of keeping or dropping controlled by a predefined value.

[0125] Random sampling using Gaussian distribution: Continuous value screening based on feature importance weight distribution.

[0126] Random sampling using fixed rules: Set a deterministic sampling scheme based on scenario requirements, such as retaining certain specific types of features.

[0127] S2004: Determine a first random screening probability distribution vector corresponding to the scene context feature based on the first random sampling mechanism;

[0128] (First) Random screening probability distribution vector: represents the probability of each scene context feature being selected, generated by a random sampling mechanism.

[0129] Vector form: [p 1 ,p 2 ,...,p n ], where p i ∈[0,1] represents the selection probability of the i-th feature.

[0130] Scenario context features: Input features related to the recommendation scenario, such as time, weather, location, etc. in the travel scenario.

[0131] Schematically, a selection probability is generated for the scene context feature according to the first random sampling mechanism. For example, using the Bernoulli distribution, set [p 1 ,p 2 ,...,p n ], where p i=0.7 means that the i-th feature has a 70% probability of being retained.

[0132] Optionally, the context features of the target scene are randomly screened by the first random sampling mechanism. The probability of screening the context features is denoted as p, and the feature combination after screening is denoted as x b :

[0133]

[0134] in, is x c First, the random sampling distribution is determined according to the probability p to generate r c (First) Randomly select the probability distribution vector, and then according to r c Probability vector for randomly selecting x c Some of the features in the output

[0135] For example, the first random sampling mechanism uses Bernoulli distribution for random sampling, which can be referred to Determine the Bernoulli random sampling distribution according to the probability p, and then generate r c (First) Randomly filter the probability distribution vector.

[0136] S2006: Filter target scene context features of each scene expert network from the scene context features based on the first random screening probability distribution vector.

[0137] Target scene context features: A subset of features that are filtered and assigned to a scene expert network.

[0138] Screening: retain or discard features based on random screening probability distribution vectors;

[0139] Schematically, the scene expert network is based on the random selection probability distribution vector for each scene context feature x i ∈x c , generate a random number r i～ U(0,1). If r i ≤p i , then the feature is retained; otherwise it is discarded. Each scene expert network only receives its own target scene context features.

[0140] In a feasible implementation manner, the step of filtering the target scene context features of each scene expert network from the scene context features based on the first random screening probability distribution vector may be performed in the following manner:

[0141] Based on the first random screening probability distribution vector and the scene context feature being input into the first model calculation formula, the target scene context feature of each scene expert network is screened by the first model calculation formula to provide feature diversity for the expert network; the first model calculation formula satisfies the following formula:

[0142]

[0143] Among them, the is the target scene context feature of the scene expert network, the r 1 c is the first random screening probability distribution vector, the x c is the scene context feature;

[0144] Said It indicates that a random sampling operation is performed on the scene context feature based on the first random screening probability distribution vector; such as element-by-element multiplication (Hadamard product) of the scene context feature, and some features are filtered through the first random screening probability distribution vector.

[0145] The squeeze() method represents a compression and cleaning operation after the random sampling operation, removing possible redundant zero vectors and retaining only valid features.

[0146] The function of squeeze() is to clean and compress the features after random sampling, remove redundant dimensions or eliminate invalid features, and ensure that the dimensions and content of the features meet the input requirements of the subsequent model. The first random screening probability distribution vector is used to multiply the scene context features element by element, which may set some feature values ​​to zero. At this time, the compression and cleaning operation is used to filter the results to retain only non-zero features, concentrate the retained context features, remove the redundant positions occupied by zero values, and form a compact feature representation for subsequent processing.

[0147] In this specification, through the implementation of S2002-S2006, the recommendation processing model can dynamically filter scene context features, provide each scene expert network with an exclusive target feature subset, and effectively enhance the task adaptability and differentiation of the expert network. This random sampling mechanism introduces the diversity of feature selection of multiple expert networks, significantly alleviates the expert polarization phenomenon, and improves the generalization ability and resource utilization efficiency of the model.

[0148] See also Figure 4 , Figure 4This is a flowchart of a task recommendation proposed in one or more embodiments of this specification. To specifically perform the task recommendation based on the context features of each target scene, the user features and the project features, using each scene expert network to obtain the target task recommendation result, you can refer to the following implementation methods:

[0149] S3002: Generate a transaction feature combination of each scenario expert network based on the target scenario context features, the user features, and the project features;

[0150] The transaction feature combination of each scene expert network after screening is recorded as x b :

[0151]

[0152] S3004: Performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result;

[0153] Scenario recommendation prediction: Each scenario expert network generates a prediction result for a specific scenario, such as a recommendation list, click probability, or behavior score, based on a combination of input transaction features. The output form of scenario recommendation prediction depends on the goal of the scenario task.

[0154] In each of the scene expert networks, the "user features and project features" and the target scene context features selected by random sampling are used for attention operation. The purpose of attention is to find the features with greater relevance to the specific scene among the basic features of "user features and project features" based on the scene features unique to different experts to perform scene prediction and obtain scene prediction results. In this way, different experts can pay more attention to knowledge in specific fields and filter out noise through attention to retain important information.

[0155] Schematically, the scene expert network performs forward propagation on the transaction feature combination. During the forward propagation process, the "user features and item features" are used with the target scene context features after random sampling to perform attention operations, extract feature representations, perform scene recommendation prediction processing, and generate scene prediction results.

[0156] Example

[0157] Assume that the following transaction feature combination is input into the "taxi recommendation network":

[0158] User characteristics: age 30, gender female.

[0159] Project features: The coupon type is a taxi coupon, and the amount is 10 yuan.

[0160] Scene context features: The time is the morning rush hour and the weather is rainy. Output scene prediction result: The probability of the user clicking on the coupon is 0.85.

[0161] In a feasible implementation manner, the scenario recommendation prediction for the transaction feature combination based on each scenario expert network is performed to obtain a scenario recommendation prediction result, which can be performed in the following manner:

[0162] A2: transmitting each transaction feature combination to each scenario expert network respectively;

[0163] A4: determining the correlation weights of the user features and the project features with the target scene context features based on the scene expert network, and performing weighted processing on the target scene context features using a second model calculation formula based on the correlation weights to obtain weighted scene context features;

[0164] The second model calculation formula satisfies the following formula:

[0165]

[0166] Wherein, A is the weighted scene context feature, is the target scene context feature of the scene expert network, the x u is the user feature, the x m For the project features, the The weight representing the correlation between the user feature and the item feature and the target scene context feature.

[0167] Furthermore, through the attention mechanism, user features and project features are weighted to improve the relevance of contextual features.

[0168] Said It means calculating the correlation between user features, item features and target scene context features, and outputting the attention weight.

[0169] Said and It means that the target scene context features are weighted by the attention weight to extract the important features of the specific scene. The obtained weighted scene context features incorporate the domain relevance of user features and project features.

[0170] A6: Combining the weighted scene context features, the user features and the project features using a third model calculation formula to obtain an expert transaction feature combination of each scene expert network;

[0171] The third model calculation formula satisfies the following formula:

[0172] x b ={x u ,x m ,A}

[0173] The weighted scene context feature, the user feature and the project feature are combined by a third model calculation formula, and the x b is the expert transaction feature combination.

[0174] A8: Based on the expert transaction feature combination, a forward propagation operation is performed in the scenario expert network to perform scenario recommendation prediction to obtain an expert prediction result, and a scenario recommendation prediction result is obtained based on the expert prediction result;

[0175] The expert prediction result can be expressed as f i (x b );

[0176] S3006: Perform task recommendation based on the recommendation prediction results of each of the scenarios to obtain a target task recommendation result, and output the target task recommendation result through the recommendation processing model.

[0177] Task recommendation: Integrate the prediction results of different scenarios to generate the final task-level recommendation output, usually a recommendation list or ranking result.

[0178] Target task recommendation result: The final output of the model, which is the recommendation content generated after combining the prediction results of all scenario recommendations.

[0179] Schematically, scenario prediction results are collected from expert networks of each scenario, and the scenario prediction results of each scenario are weighted or fused according to task requirements to obtain target task recommendation results, which are then passed through a recommendation processing model.

[0180] In a feasible implementation, based on the prediction results of the scenario recommendations, the prediction results are comprehensively processed to obtain the target task recommendation results;

[0181] In a feasible implementation, the task scenario weight corresponding to each of the scenario recommendation prediction results is determined, and the prediction results are weighted based on the task scenario weight and each of the scenario recommendation prediction results to obtain a target task recommendation result.

[0182] Example - In the travel recommendation task:

[0183] Scenario prediction results:

[0184] The predicted click probability of the “Public Transportation Recommendation Network” is 0.6;

[0185] The predicted click probability of the taxi recommendation network is 0.85;

[0186] The "Bicycle Recommendation Network" predicts a click probability of 0.4.

[0187] Fusion rules: According to user preferences and task scenarios, give the "taxi recommendation network" a higher weight, such as w 打车 =0.7, the final fusion score R final =0.7*0.85+0.2*0.6+0.1*0.4=0.745

[0188] Output recommendation result: recommend the user to receive a "taxi voucher" and display the reason for the recommendation: "It is more convenient to take a taxi during the morning rush hour on rainy days."

[0189] In this specification, each scenario expert network independently predicts the transaction feature combination, and the solution ensures that scenario-specific information is fully mined; in S3006, the recommendation prediction results of each scenario are integrated and comprehensively processed, effectively balancing the complexity of multi-scenario tasks and generating the global optimal task recommendation results. This design not only alleviates the expert polarization phenomenon, but also improves the personalization, accuracy and task adaptability of the recommendation system, meeting the recommendation needs of multi-scenario tasks.

[0190] Optionally, the recommendation processing model further includes at least one gating network, which is a neural network used to dynamically assign weights to scene expert networks, determine the weight distribution of different scene expert networks based on input features, and guide the fusion of recommendation task results. That is, the recommendation processing model is composed of multiple scene expert networks and at least one gating network, and the task scene weights corresponding to the prediction results of each scene recommendation can be determined by:

[0191] S4010: Determine a gated input feature combination for the gated network based on the user feature, the project feature, and the scene context feature;

[0192] S4012: Based on the gated input feature combination, a forward propagation operation is performed in the gated network to predict the scene recommendation weight to obtain an expert prediction weight result, and the expert prediction weight result is used as the task scene weight corresponding to the scene recommendation prediction result.

[0193] Further, the explanations of S4010-S4012 can refer to the explanations of subsequent steps, such as Figure 5 As shown, Figure 5 This is a flowchart of a recommended processing method, specifically:

[0194] S4002: Obtain user features, project features, and scene context features for a target recommendation task, and input the user features, project features, and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task and at least one gating network;

[0195] like Figure 6 As shown, Figure 6 is a scenario diagram of a recommendation processing model. Figure 6 In the recommendation processing model, there are three scene expert networks (Expert1, Expert2, Expert3) and two gated networks (gate1, gate2). The user feature is represented as: user feature, the item feature is represented as: item feature, the scene context feature is represented as: contextual feature; the attention mechanism is represented as attention;

[0196] S4004: Screening the scene context features as input to obtain target scene context features of each scene expert network;

[0197] Indicatively, in Figure 6 The three scene expert networks are applied with bagging network layers to randomly select a part of context scene features, that is, all basic features of users and items will be sent to all expert networks, because these features are necessary for all experts. They are E1-bagging-layer network layer corresponding to the scene expert network of Expert1, E2-bagging-layer network layer corresponding to the scene expert network of Expert2, and E3-bagging-layer network layer corresponding to the scene expert network of Expert3; and, for the gated network, bagging network layers are also used to randomly select a part of context scene features, which are E1-bagging-layer network layer corresponding to the gated network gate1, and E2-bagging-layer network layer corresponding to the gated network gate2; the bagging network layer uses a random sampling mechanism (such as Bernoulli distribution) to generate feature subsets. Different bagging layers ensure that each expert network receives different context feature subsets, increasing the diversity of features and the differences in task division.

[0198] Figure 6 The Output Layer combines the weights of the gating network output with the scene recommendation results of the expert network to generate the final recommendation results. Each target recommendation task corresponds to an output (such as output 1 and output 2).

[0199] Schematically, the scene context features are sampled through bagging and then input into three scene-specific expert networks and two gating networks. Then, in the scene expert network, the basic features of the user and item and the scene context features after bagging are used for attention operation. The purpose of attention is to find the features of the basic features of the user and item that are more relevant to the specific scene based on the scene features unique to different scene expert networks. In this way, different scene expert networks can pay more attention to the knowledge in specific fields and filter out the noise through attention to retain important information.

[0200] Figure 6 The recommendation model described achieves accurate processing of multi-scenario recommendation tasks through the collaboration of scene expert network and gating network, combined with feature screening (Bagging layer), feature correlation calculation (Attention mechanism) and dynamic weight allocation (gating network). The model design fully considers the issues of feature diversity, task adaptability and balanced resource allocation, and provides an efficient and flexible solution for multi-task recommendation scenarios.

[0201] S4006: Generate a transaction feature combination of each scenario expert network based on the target scenario context features, the user features, and the project features;

[0202] For details, please refer to the explanations of other method steps in this manual, which will not be repeated here.

[0203] S4008: Performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result;

[0204] For details, please refer to the explanations of other method steps in this manual, which will not be repeated here.

[0205] S4010: Determine a gated input feature combination for the gated network based on the user feature, the project feature, and the scene context feature;

[0206] Gated input feature combination: A feature set generated specifically for the gating network, including user features, item features, and context features, which are used to calculate the weights of each scene expert network.

[0207] In a feasible implementation manner, the determining of the gated input feature combination for the gated network based on the user feature, the project feature, and the scene context feature may refer to the following method:

[0208] B2: Determine a second random sampling mechanism corresponding to the gating network, and determine a second random screening probability distribution vector corresponding to the scene context feature based on the second random sampling mechanism;

[0209] Second random sampling mechanism: A feature screening method to reduce or even avoid the polarization phenomenon of the recommendation processing model. The second random sampling mechanism determines which scene context features are retained as the input of the gating network for each gating network, thereby introducing feature diversity and enhancing the generalization ability of the model.

[0210] In an illustrative manner, an independent second random sampling mechanism is assigned to one or more gated networks to ensure that different gated networks have random and differentiated focus on contextual features. Optionally, a random sampling network layer (such as a bagging network layer) is set in the recommendation processing model to randomly select a portion of contextual scene features for the gated network, bringing unique feature information to each gated network, thereby enabling different experts and gated networks to focus on different domain knowledge and greatly reducing polarization;

[0211] The implementation of the random sampling mechanism can be, but is not limited to, one or more of the following fittings:

[0212] Random sampling is performed using a Bernoulli distribution: each feature is sampled independently, with the probability of keeping or dropping controlled by a predefined value.

[0213] Random sampling using Gaussian distribution: Continuous value screening based on feature importance weight distribution.

[0214] Random sampling using fixed rules: Set a deterministic sampling scheme based on scenario requirements, such as retaining certain specific types of features.

[0215] B4: filtering reference scene context features of the gating network from the scene context features based on the second random screening probability distribution vector;

[0216] Reference scene context features: A subset of features that are filtered and assigned to a gating network.

[0217] Screening: retain or discard features based on random screening probability distribution vectors;

[0218] (Second) Random screening probability distribution vector: represents the probability of each scene context feature being selected, generated by a random sampling mechanism.

[0219] Vector form: [p 1 ,p 2 ,...,p n ], where p i ∈[0,1] represents the selection probability of the i-th feature.

[0220] Scenario context features: Input features related to the recommendation scenario, such as time, weather, location, etc. in the travel scenario.

[0221] In a feasible implementation, the filtering of reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector includes:

[0222] Based on the second random screening probability distribution vector and the scene context feature being input into a fourth model calculation formula, the reference scene context feature of the gating network is screened by the fourth model calculation formula; the fourth model calculation formula satisfies the following formula:

[0223] Among them, the is the reference scene context feature of the scene expert network, is the second random screening probability distribution vector, the x c is the scene context feature, It indicates that a random sampling operation is performed on the scene context features based on the second random screening probability distribution vector, such as element-by-element multiplication (Hadamard product) of the scene context features, and some features are filtered through the second random screening probability distribution vector.

[0224] squeeze() represents the compression and cleaning operation after the random sampling operation; it removes the redundant zero vectors that may be generated and only retains the valid features. The function of squeeze() is to clean and compress the features after random sampling, remove redundant dimensions or eliminate invalid features, and ensure that the dimensions and content of the features meet the input requirements of the subsequent model. The second random screening probability distribution vector is used to multiply the scene context features element by element, which may set some feature values ​​to zero. At this time, the compression and cleaning operation is used to filter the results to retain only non-zero features, concentrate the retained context features, remove the redundant positions occupied by zero values, and form a compact feature representation for subsequent processing.

[0225] For example, the second random sampling mechanism uses Bernoulli distribution for random sampling, which can be referred to According to the probability p, the Bernoulli random sampling distribution is determined, and then the (Second) Randomly screen the probability distribution vector.

[0226] B6: Generate a gated input feature combination of the gated network based on the reference scene context features, the user features, and the project features.

[0227] In a feasible implementation manner, generating the gated input feature combination of the gated network based on the reference scene context feature, the user feature and the project feature includes: combining the reference scene context feature, the user feature and the project feature using a fifth model calculation formula to obtain the gated input feature combination of the gated network;

[0228] The fifth model calculation formula satisfies the following formula:

[0229] Among them, the is the gated input feature combination, the x u is the user feature, the x m The project features.

[0230] S4012: Based on the gated input feature combination, a forward propagation operation is performed in the gated network to predict the scene recommendation weight to obtain an expert prediction weight result, and the expert prediction weight result is used as the task scene weight corresponding to the scene recommendation prediction result.

[0231] Forward propagation operation: The gated input feature combination passes through each layer of the gated network, and the weight distribution of the expert network for each scene is calculated and output.

[0232] Scenario recommendation weight prediction: According to the output of the gating network, weights are dynamically assigned to each scenario expert network to fuse their prediction results.

[0233] Expert prediction weight result: The weight value output by the gating network indicates the contribution ratio of each scene expert network in the current recommendation task.

[0234] Said represents the task scenario weight for the i-th scenario expert network;

[0235] S4014: performing weighted processing on the prediction results based on the task scenario weights and the recommended prediction results of each scenario to obtain a target task recommendation result;

[0236] Task scene weight: The weight generated by the gating network for each scene expert network is used to weight the scene prediction results.

[0237] Target task recommendation result: The final recommended content or result is generated by weighted fusion of the prediction results of all scenario expert networks.

[0238] The weight distribution of the scene expert network is dynamically adjusted through the gating network to ensure that the expertise of experts in different scenes can be fully utilized according to task requirements during the recommendation result fusion process, improving the accuracy and personalization of recommendations. At the same time, this dynamic weight distribution mechanism effectively avoids the phenomenon of expert polarization, makes resource utilization more balanced, and significantly improves the overall performance of the model.

[0239] In a feasible implementation manner, the step of performing weighted processing on the prediction results based on the task scenario weights and the recommended prediction results of each scenario to obtain the target task recommendation result includes:

[0240] Based on the task scenario weights and the recommended prediction results of each scenario, a sixth model calculation formula is used to perform weighted processing on the prediction results to obtain a target task recommendation result;

[0241] The sixth model calculation formula satisfies the following formula:

[0242]

[0243] Among them, the f k is the target task recommendation result, the f i (x b ) represents the scene recommendation prediction result of the i-th scene expert network, represents the task scenario weight for the i-th scenario expert network, and k represents the recommended task label.

[0244] Example

[0245] The following are the recommended prediction results for the hypothetical scenario:

[0246] 「Bus Recommendation Network」outputs click probability R 公交 =0.4;

[0247] The taxi recommendation network outputs the click probability R 打车 =0.85;

[0248] 「Bicycle Recommendation Network」outputs click probability R 自行车 =0.3.

[0249] The corresponding scene weight is: w = [0.65, 0.25, 0.10]

[0250] The final result is calculated by weighted fusion:

[0251] R final =0.65*0.4+0.25*0.85+0.10*0.3=0.515

[0252] The final recommendation may be: "Recommend users to collect taxi coupons because it is more convenient to take a taxi during the morning rush hour on rainy days."

[0253] In one or more embodiments of this specification, the recommendation processing model combines the dynamic collaboration of the scene expert network and the gating network to effectively improve the accuracy and adaptability of the recommendation task. The scene expert network focuses on extracting feature information in a specific recommendation scenario, while the gating network achieves intelligent fusion of the expert network output by dynamically calculating the scene weights. The combination of feature screening and weight allocation not only alleviates the expert polarization phenomenon, but also optimizes the resource utilization between scenarios, ensuring that the model can efficiently and balancedly generate personalized recommendation results in multi-scenario and multi-task recommendations, bringing significant improvements to user experience and business goals.

[0254] In order to verify the beneficial effects of implementing one or more of the recommended processing methods in this manual, experimental verification was carried out, such as Figure 7 As shown, Figure 7 This is a schematic diagram of the effect verification of a recommendation processing method. For example, if the target recommendation task is to estimate the coupon recommendation task in the travel industry, then the gating network of the travel task may give a higher weight to the "travel expert network with travel scenario characteristics". The network weights of the recommendation processing model used for coupon recommendation are visualized as follows Figure 7 As shown, the horizontal axis represents that expert 1 (E1) was randomly input with the features of Ele.me, Feizhu, and Taopiaopiao. Expert 2 (E2) was randomly input with the features of Ele.me, Taobao, and Taopiaopiao. The vertical axis represents the weights of the different feature groups of the two experts in four different prediction tasks: takeaway, online shopping, transportation, and movie. It can be seen that since expert 1 has Feizhu features that expert 2 does not have, the transportation task has a higher weight on expert 1, especially, expert 1's Feizhu feature group has the highest weight. Expert 2 has Taobao features that expert 1 does not have, so it contributes more to the e-commerce task.

[0255] It can be seen from the above that by executing the recommendation processing method, after passing through the bagging layer, the features input by different experts are different, which will also cause the expert network to have weight differences when estimating different tasks. Each expert network focuses on knowledge in a specific field, each performs its own duties, and each expert network is an indispensable part. Therefore, the problem of a certain expert network stopping updating and several expert networks occupying a dominant position is avoided. Each expert is very important, thus avoiding the occurrence of polarization. In the related art, in order to avoid polarization, either dropout is added, such as stopping training of certain experts with a certain probability, which will damage the training effect of the expert network. Either additional loss function terms are added to avoid polarization, and adding additional loss functions will also damage the learning effect of the original task loss function. The recommendation processing method involved in this specification neither adds dropout nor loss function, but makes the expert network different through clever feature design, thereby avoiding polarization and not reducing the original task learning effect.

[0256] In one or more embodiments of this specification, a novel expert network model structure is proposed, which can effectively alleviate the polarization problem of the multi-expert network architecture and improve the performance of the recommendation processing model. Through the bagging layer, the recommendation processing model generates different feature subgroups for experts, and uses the attention mechanism to capture features to reduce irrelevant noise data. Currently, expert networks are very popular in the field of large models (LLM), and this specification applies the bagging method to MoE, which is innovative and solves the polarization pain point problem in the recommendation processing model of the multi-expert network architecture and improves the effect.

[0257] The following will be combined Figure 8 , the recommended processing device provided in this manual is introduced in detail. It should be noted that Figure 8 The recommended processing device shown is used to perform this instruction Figure 1 to Figure 7 For the convenience of explanation, only the parts related to this specification are shown. For the specific technical details not disclosed, please refer to this specification. Figure 1 to Figure 7 The embodiment shown.

[0258] See also Figure 8 , which shows a schematic diagram of the structure of the recommendation processing device of this specification. The recommendation processing device 1 can be implemented as all or part of the electronic device through software, hardware or a combination of both. According to some embodiments, the recommendation processing device 1 includes a feature processing module 11, a task recommendation module 12 and a result output module 13, which are specifically used to:

[0259] A feature processing module 11 is used to obtain user features, project features and scene context features for a target recommendation task, and input the user features, project features and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task;

[0260] A task recommendation module 12 is used to perform input feature screening on the scene context features to obtain target scene context features of each scene expert network, and based on each target scene context feature, the user feature and the project feature, use each scene expert network to perform task recommendation to obtain a target task recommendation result;

[0261] The result output module 13 is used to output the target task recommendation result through the recommendation processing model.

[0262] In a feasible implementation manner, the step of filtering the scene context features as input features to obtain target scene context features of each scene expert network includes:

[0263] Determine a first random sampling mechanism corresponding to each of the scene expert networks;

[0264] Determine a first random screening probability distribution vector corresponding to the scene context feature based on the first random sampling mechanism;

[0265] The target scene context features of each of the scene expert networks are filtered from the scene context features based on the first random filtering probability distribution vector.

[0266] In a feasible implementation manner, the filtering of the target scene context features of each scene expert network from the scene context features based on the first random filtering probability distribution vector includes:

[0267] Based on the first random screening probability distribution vector and the scene context feature being input into the first model calculation formula, the target scene context feature of each scene expert network is screened by the first model calculation formula; the first model calculation formula satisfies the following formula:

[0268]

[0269] Among them, the is the target scene context feature of the scene expert network, the r 1 c is the first random screening probability distribution vector, the x c is the scene context feature, It indicates that a random sampling operation is performed on the scene context feature based on the first random screening probability distribution vector, and the squeeze() indicates a compression and cleaning operation after the random sampling operation.

[0270] In a feasible implementation, the step of performing task recommendation based on the target scene context features, the user features, and the project features using the scene expert network to obtain a target task recommendation result includes:

[0271] Generate a transaction feature combination of each scenario expert network based on the target scenario context features, the user features and the project features;

[0272] Performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result;

[0273] Based on the recommendation prediction results of each of the scenarios, task recommendation is performed to obtain a target task recommendation result, and the target task recommendation result is output through the recommendation processing model.

[0274] In a feasible implementation manner, performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result includes:

[0275] Transmitting each of the transaction feature combinations to each of the scenario expert networks respectively;

[0276] Determining the correlation weights of the user features and the project features with the target scene context features based on the scene expert network, and performing weighted processing on the target scene context features using a second model calculation formula based on the correlation weights to obtain weighted scene context features;

[0277] Combining the weighted scene context features, the user features and the project features using a third model calculation formula to obtain an expert transaction feature combination of each scene expert network;

[0278] Based on the expert transaction feature combination, a forward propagation operation is performed in the scenario expert network to perform scenario recommendation prediction to obtain an expert prediction result, and based on the expert prediction result, a scenario recommendation prediction result is obtained;

[0279] The second model calculation formula satisfies the following formula:

[0280]

[0281] Wherein, A is the weighted scene context feature, is the target scene context feature of the scene expert network, the x uis the user feature, the x m For the project features, the The weight representing the correlation between the user feature and the item feature and the target scene context feature.

[0282] The third model calculation formula satisfies the following formula:

[0283] x b ={x u ,x m ,A}

[0284] Among them, the x b is the expert transaction feature combination.

[0285] In a feasible implementation manner, performing task recommendation based on the recommendation prediction results of each of the scenarios to obtain a target task recommendation result, and outputting the target task recommendation result through the recommendation processing model, includes:

[0286] Based on the prediction results of the scenario recommendations, comprehensive processing of the prediction results is performed to obtain the target task recommendation results; or,

[0287] Determine the task scenario weight corresponding to each of the scenario recommendation prediction results, and perform weighted processing on the prediction results based on the task scenario weight and each of the scenario recommendation prediction results to obtain a target task recommendation result.

[0288] In a feasible implementation manner, the recommendation processing model further includes at least one gating network, and the determining of the task scenario weight corresponding to each of the scenario recommendation prediction results includes:

[0289] Determine a gated input feature combination for the gated network based on the user features, the project features, and the scene context features;

[0290] Based on the gated input feature combination, a forward propagation operation is performed in the gated network to predict the scene recommendation weight to obtain an expert prediction weight result, and the expert prediction weight result is used as the task scene weight corresponding to the scene recommendation prediction result.

[0291] In a feasible implementation manner, the determining of the gated input feature combination for the gated network based on the user feature, the project feature and the scene context feature includes:

[0292] Determine a second random sampling mechanism corresponding to the gating network, and determine a second random screening probability distribution vector corresponding to the scene context feature based on the second random sampling mechanism;

[0293] Filtering reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector;

[0294] A gated input feature combination of the gated network is generated based on the reference scene context feature, the user feature, and the item feature.

[0295] In a feasible implementation, the filtering of reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector includes:

[0296] Based on the second random screening probability distribution vector and the scene context feature being input into a fourth model calculation formula, the reference scene context feature of the gating network is screened by the fourth model calculation formula; the fourth model calculation formula satisfies the following formula:

[0297] Among them, the is the reference scene context feature of the scene expert network, is the second random screening probability distribution vector, the x c is the scene context feature, represents a random sampling operation performed on the scene context feature based on the second random screening probability distribution vector, and the squeeze() represents a compression and cleaning operation after the random sampling operation;

[0298] The generating of the gated input feature combination of the gated network based on the reference scene context feature, the user feature and the project feature comprises: combining the reference scene context feature, the user feature and the project feature using a fifth model calculation formula to obtain the gated input feature combination of the gated network;

[0299] The fifth model calculation formula satisfies the following formula:

[0300] Among them, the is the gated input feature combination, the x u is the user feature, the x m The project features.

[0301] In a feasible implementation manner, the step of performing weighted processing on the prediction results based on the task scenario weights and the recommended prediction results of each scenario to obtain the target task recommendation result includes:

[0302] Based on the task scenario weights and the recommended prediction results of each scenario, a sixth model calculation formula is used to perform weighted processing on the prediction results to obtain a target task recommendation result;

[0303] The sixth model calculation formula satisfies the following formula:

[0304]

[0305] Among them, the f k is the target task recommendation result, the f i (x b ) represents the scene recommendation prediction result of the i-th scene expert network, represents the task scenario weight for the i-th scenario expert network, and k represents the recommended task label.

[0306] It should be noted that the recommendation processing device provided in the above embodiment only uses the division of the above functional modules as an example when executing the recommendation processing method. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the recommendation processing device provided in the above embodiment and the recommendation processing method embodiment belong to the same concept, and the implementation process thereof is detailed in the method embodiment, which will not be repeated here.

[0307] The above serial numbers in this specification are for description only and do not represent the advantages or disadvantages of the embodiments.

[0308] The present specification also provides a computer storage medium, which can store multiple instructions, which are suitable for being loaded and executed by a processor as described above. Figure 1 to Figure 7 The recommended processing method of the embodiment shown in the figure can be found in the specific execution process. Figure 1 to Figure 7 The specific description of the illustrated embodiment will not be repeated here.

[0309] The present specification also provides a computer program product, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor as described above. Figure 1 to Figure 7 The recommended processing method of the embodiment shown in the figure can be found in the specific execution process. Figure 1 to Figure 7 The specific description of the illustrated embodiment will not be repeated here.

[0310] Please refer to Fig. 9 , is a block diagram of a structure of an electronic device provided in an embodiment of this specification. The electronic device in this specification may include one or more of the following components: a processor 1010, a memory 1020, an input device 1030, an output device 1040, and a bus 1050. The processor 1010, the memory 1020, the input device 1030, and the output device 1040 may be connected via a bus 1050.

[0311] The processor 1010 may include one or more processing cores. The processor 1010 uses various interfaces and lines to connect various parts of the entire electronic device, and executes various functions of the electronic device and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 1020, and calling data stored in the memory 1020. Optionally, the processor 1010 can be implemented in at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), and programmable logic array (PLA). The processor 1010 can integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. Among them, the CPU mainly processes the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing display content; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 1010, but may be implemented separately through a communication chip.

[0312] The memory 1020 may include a random access memory (RAM) or a read-only memory (ROM). Optionally, the memory 1020 includes a non-transitory computer-readable storage medium. The memory 1020 may be used to store instructions, programs, codes, code sets, or instruction sets.

[0313] The input device 1030 is used to receive input instructions or data, and the input device 1030 includes but is not limited to a keyboard, a mouse, a camera, a microphone or a touch device. The output device 1040 is used to output instructions or data, and the output device 1040 includes but is not limited to a display device and a speaker. In the embodiment of this specification, the input device 1030 can be a temperature sensor for obtaining the operating temperature of the electronic device. The output device 1040 can be a speaker for outputting audio signals.

[0314] In addition, those skilled in the art will appreciate that the structure of the electronic device shown in the above drawings does not constitute a limitation on the electronic device, and the electronic device may include more or fewer components than shown, or combine certain components, or arrange the components differently. For example, the electronic device also includes a radio frequency circuit, an input unit, a sensor, an audio circuit, a wireless fidelity (WIFI) module, a power supply, a Bluetooth module and other components, which will not be described in detail here.

[0315] In the embodiments of this specification, the execution subject of each step may be the electronic device described above. Optionally, the execution subject of each step is the operating system of the electronic device. The operating system may be an Android system, an IOS system, or other operating systems, which is not limited in the embodiments of this specification.

[0316] exist Fig. 9 In the electronic device, the processor 1010 can be used to call the program stored in the memory 1020 and execute it to implement the recommended processing method described in the various method embodiments of this specification.

[0317] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0318] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions. For example, the user characteristics, project characteristics and scene context characteristics involved in this specification are all obtained with full authorization.

[0319] The above disclosure is only the preferred embodiment of this specification, which certainly cannot be used to limit the scope of rights of this specification. Therefore, equivalent changes made according to the claims of this specification are still within the scope covered by this specification.

Claims

1. A recommendation processing method, the method comprising: Obtaining user features, project features, and scene context features for a target recommendation task, and inputting the user features, project features, and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task; Performing input feature screening on the scene context features to obtain target scene context features of each scene expert network, and based on each of the target scene context features, the user features, and the project features, using each of the scene expert networks to perform task recommendation to obtain a target task recommendation result; The target task recommendation result is outputted through the recommendation processing model.

2. According to the method of claim 1, the step of filtering the scene context features as input features to obtain target scene context features of each scene expert network comprises: Determine a first random sampling mechanism corresponding to each of the scene expert networks; Determine a first random screening probability distribution vector corresponding to the scene context feature based on the first random sampling mechanism; The target scene context features of each of the scene expert networks are filtered from the scene context features based on the first random filtering probability distribution vector.

3. The method according to claim 2, wherein the step of filtering target scene context features of each scene expert network from the scene context features based on the first random screening probability distribution vector comprises: Based on the first random screening probability distribution vector and the scene context feature being input into a first model calculation formula, the target scene context feature of each scene expert network is screened by the first model calculation formula; The first model calculation formula satisfies the following formula: Among them, the is the target scene context feature of the scene expert network, the r1 c is the first random screening probability distribution vector, the x c is the scene context feature, It indicates that a random sampling operation is performed on the scene context feature based on the first random screening probability distribution vector, and the squeeze() indicates a compression and cleaning operation after the random sampling operation.

4. The method according to claim 1, wherein the step of performing task recommendation based on the target scene context features, the user features and the project features using each of the scene expert networks to obtain a target task recommendation result comprises: Generate a transaction feature combination of each scenario expert network based on the target scenario context features, the user features and the project features; Performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result; Based on the recommendation prediction results of each of the scenarios, task recommendation is performed to obtain a target task recommendation result, and the target task recommendation result is output through the recommendation processing model.

5. According to the method of claim 4, the step of performing scenario recommendation prediction on the transaction feature combination based on each scenario expert network to obtain a scenario recommendation prediction result comprises: Transmitting each of the transaction feature combinations to each of the scenario expert networks respectively; Determining the correlation weights of the user features and the project features with the target scene context features based on the scene expert network, and performing weighted processing on the target scene context features using a second model calculation formula based on the correlation weights to obtain weighted scene context features; Combining the weighted scene context features, the user features and the project features using a third model calculation formula to obtain an expert transaction feature combination of each scene expert network; Based on the expert transaction feature combination, a forward propagation operation is performed in the scenario expert network to perform scenario recommendation prediction to obtain an expert prediction result, and based on the expert prediction result, a scenario recommendation prediction result is obtained; The second model calculation formula satisfies the following formula: Wherein, A is the weighted scene context feature, is the target scene context feature of the scene expert network, the x u is the user feature, the x m For the project features, the A weight representing the correlation between the user feature and the item feature and the target scene context feature; The third model calculation formula satisfies the following formula: x b ={x u ,x m ,A} Among them, the x b is the expert transaction feature combination.

6. The method according to claim 4, wherein the step of performing task recommendation based on each of the scenario recommendation prediction results to obtain a target task recommendation result, and outputting the target task recommendation result through the recommendation processing model comprises: Based on the prediction results of the scenarios, the prediction results are comprehensively processed to obtain the target task recommendation results; or, Determine the task scenario weight corresponding to each of the scenario recommendation prediction results, and perform weighted processing on the prediction results based on the task scenario weight and each of the scenario recommendation prediction results to obtain a target task recommendation result.

7. According to the method of claim 6, the recommendation processing model further comprises at least one gating network, and the determining of the task scenario weight corresponding to each of the scenario recommendation prediction results comprises: Determine a gated input feature combination for the gated network based on the user features, the project features, and the scene context features; Based on the gated input feature combination, a forward propagation operation is performed in the gated network to predict the scene recommendation weight to obtain an expert prediction weight result, and the expert prediction weight result is used as the task scene weight corresponding to the scene recommendation prediction result.

8. The method according to claim 6, wherein determining a gated input feature combination for the gated network based on the user feature, the item feature, and the scene context feature comprises: Determine a second random sampling mechanism corresponding to the gating network, and determine a second random screening probability distribution vector corresponding to the scene context feature based on the second random sampling mechanism; Filtering reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector; A gated input feature combination of the gated network is generated based on the reference scene context feature, the user feature, and the item feature.

9. The method according to claim 8, wherein the step of filtering the reference scene context features of the gating network from the scene context features based on the second random filtering probability distribution vector comprises: Based on the second random screening probability distribution vector and the scene context feature being input into a fourth model calculation formula, the reference scene context feature of the gating network is screened by the fourth model calculation formula; The fourth model calculation formula satisfies the following formula: Among them, the is the reference scene context feature of the scene expert network, is the second random screening probability distribution vector, the x c is the scene context feature, represents a random sampling operation performed on the scene context feature based on the second random screening probability distribution vector, and the squeeze() represents a compression and cleaning operation after the random sampling operation; The generating of the gated input feature combination of the gated network based on the reference scene context feature, the user feature and the project feature comprises: combining the reference scene context feature, the user feature and the project feature using a fifth model calculation formula to obtain the gated input feature combination of the gated network; The fifth model calculation formula satisfies the following formula: Among them, the is the gated input feature combination, the x u is the user feature, the x m The project features.

10. The method according to claim 6, wherein the step of performing weighted processing on the prediction results based on the task scenario weights and the recommended prediction results of each scenario to obtain the target task recommendation result comprises: Based on the task scenario weights and the recommended prediction results of each scenario, a sixth model calculation formula is used to perform weighted processing on the prediction results to obtain a target task recommendation result; The sixth model calculation formula satisfies the following formula: Among them, the f k is the target task recommendation result, the f i (x b ) represents the scene recommendation prediction result of the i-th scene expert network, represents the task scenario weight for the i-th scenario expert network, and k represents the recommended task label.

11. A recommendation processing device, comprising: A feature processing module, used to obtain user features, project features and scene context features for a target recommendation task, and input the user features, project features and scene context features into a recommendation processing model, wherein the recommendation processing model includes a scene expert network corresponding to a plurality of recommendation project scenes under the target recommendation task; A task recommendation module, used to perform input feature screening on the scene context features to obtain target scene context features of each scene expert network, and based on each target scene context feature, the user features and the project features, use each scene expert network to perform task recommendation to obtain a target task recommendation result; The result output module is used to output the target task recommendation result through the recommendation processing model.

12. A computer storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the method steps according to any one of claims 1 to 10.

13. A computer program product, the computer program product storing at least one instruction, wherein the at least one instruction is loaded by a processor and executes the method steps according to any one of claims 1 to 10.

14. An electronic device comprising: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps as claimed in any one of claims 1 to 10.

Images