Methods, devices, equipment, and media for recommending items to multiple groups, in multiple scenarios, and for multiple tasks.

Abstract

This application discloses a method, apparatus, device, and medium for item recommendation across multiple user groups, scenarios, and tasks, relating to the field of artificial intelligence technology. The method involves acquiring user-side information, item-side information, and corresponding scene environment information in different scenarios; calculating weight dimensions, shared feature vectors, and environment dimensions to obtain a first dimension, a dense vector, and a second dimension; performing attention mechanism feature cross-calculation to obtain a third dimension; calculating item recommendation scores using the first dimension, dense vector, second dimension, and third dimension to obtain an item recommendation value; determining the target item recommendation value; identifying the target item information corresponding to the item recommendation value; and recommending the target item information to the user. This application can construct a three-dimensional perspective model encompassing multiple scenarios, user groups, and tasks, thereby improving the ability to represent differentiated user groups, meeting diverse user needs, and achieving comprehensive item recommendation across multiple user groups, scenarios, and tasks.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for recommending items to multiple groups of people, in multiple scenarios, and for multiple tasks. Background Technology

[0002] Current industry-standard recommendation systems with large multi-scenario, multi-task models still focus on modeling multiple optimization objectives (likes, shares, forwards) across various sub-scenarios in mobile applications. They fail to explicitly incorporate the dimension of multiple user groups into the model design, resulting in a poor ability to learn generalized needs from the same user group across scenarios and the migration and evolution of interests among different user groups across scenarios. Existing multi-scenario, multi-task models in recommendation systems often distribute single-modality, similar-type items (e.g., products, videos, news, merchants) across different sub-scenarios. They cannot leverage user behavior data across multiple mobile scenarios to deliver multi-modal, multi-type items according to the user's current real-time needs, thus hindering the model's ability to learn generalized needs from the same user group across scenarios and the migration and evolution of interests among different user groups across scenarios.

[0003] As can be seen from the above, how to construct a three-dimensional perspective model that encompasses multiple scenarios, multiple user groups, and multiple tasks, thereby improving the ability to represent the differences among different user groups, meeting the diversity of user needs, and realizing comprehensive multi-user, multi-scenario, and multi-task item recommendations, is a problem that needs to be solved in this field. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for item recommendation across multiple user groups, scenarios, and tasks. This method can construct a three-dimensional perspective model encompassing multiple scenarios, user groups, and tasks, thereby improving the ability to represent the differences among different user groups, meeting the diversity of user needs, and achieving comprehensive item recommendation across multiple user groups, scenarios, and tasks. The specific solution is as follows:

[0005] Firstly, this application discloses a method for recommending items across multiple groups, scenarios, and tasks, applied to a pre-defined model of multiple groups, scenarios, and tasks, including:

[0006] Acquire user-side information, item-side information, and corresponding scene environment information in different scenarios;

[0007] The user-side information, the item-side information, and the corresponding scene environment information are respectively subjected to weight dimension calculation, shared feature vector calculation, and environment dimension calculation to obtain the first dimension, dense vector, and second dimension;

[0008] The third dimension is obtained by performing attention mechanism feature cross-calculation based on the first dimension and the dense vector;

[0009] The item recommendation score is calculated using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value;

[0010] The item recommendation value that is greater than a preset threshold is taken as the target item recommendation value. The target item information corresponding to the item recommendation value is determined from the item-side information, and the target item information is recommended to the user.

[0011] Optionally, the multi-person, multi-scenario, multi-task model is constructed based on a weight dimension calculation layer, a shared feature vector calculation layer, an environment dimension calculation layer, an attention mechanism feature cross calculation layer, and an item recommendation score calculation layer.

[0012] Optionally, the step of calculating the weight dimension, the shared feature vector, and the environment dimension of the user-side information, the item-side information, and the corresponding scene environment information to obtain the first dimension, the dense vector, and the second dimension includes:

[0013] The weight dimension calculation layer is used to calculate the weight dimension of user-side information and item-side information in different scenarios to obtain the first dimension;

[0014] The shared feature vector calculation layer is used to calculate the shared feature vectors of the user-side information, the item-side information, and the corresponding scene environment information in different scenarios to obtain dense vectors;

[0015] The environmental dimension calculation layer is used to calculate the environmental dimension of the scene environment information under different scenarios to obtain the second dimension.

[0016] Optionally, the step of using the weighted dimension calculation layer to calculate the weighted dimension of user-side information and item-side information in different scenarios to obtain the first dimension includes:

[0017] Item information is determined from the item-side information in different scenarios, and the user's full-scenario reading sequence is determined from the user-side information in different scenarios. The weight dimension calculation layer is used to calculate the weight dimension of the item information and the user's full-scenario reading sequence to obtain the first dimension.

[0018] Optionally, the step of using the shared feature vector calculation layer to calculate shared feature vectors for the user-side information, the item-side information, and the corresponding scene environment information in different scenarios to obtain dense vectors includes:

[0019] The user-side information and the item-side information in different scenarios are combined and cross-referenced to obtain cross-features;

[0020] The basic features of the item, the basic features of the user, the basic features of the item, and the environmental features of the corresponding scene environment are determined from the user-side information, the item-side information, and the corresponding scene environment information in different scenarios. The shared feature vector calculation layer is used to calculate the shared feature vector of the basic features of the item, the user features, the environmental features, and the cross features to obtain the dense vector.

[0021] Optionally, the step of using the environment dimension calculation layer to perform environment dimension calculation on the scene environment information under different scenarios to obtain the second dimension includes:

[0022] Scene statistical features are determined from the scene environment information under different scenarios, and the environment dimension calculation layer is used to calculate the environment dimension of the scene statistical features to obtain the second dimension.

[0023] Optionally, the step of calculating the item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value includes:

[0024] The item recommendation score calculation layer in the multi-person, multi-scenario, and multi-task model is used to concatenate features of the first dimension, the dense vector, the second dimension, and the third dimension to obtain a feature concatenation vector. Then, the item recommendation score is calculated on the feature concatenation vector to obtain the item recommendation value.

[0025] Secondly, this application discloses a multi-person, multi-scenario, multi-task item recommendation device, applied to a preset multi-person, multi-scenario, multi-task model, including:

[0026] The information acquisition module is used to acquire user-side information, item-side information, and corresponding scene environment information in different scenarios.

[0027] The calculation module is used to perform weight dimension calculation, shared feature vector calculation and environment dimension calculation on the user-side information, the item-side information and the corresponding scene environment information, respectively, to obtain the first dimension, the dense vector and the second dimension;

[0028] The feature cross calculation module is used to perform attention mechanism feature cross calculation based on the first dimension and the dense vector to obtain the third dimension;

[0029] The recommendation value calculation module is used to calculate the item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value.

[0030] The item information recommendation module is used to take the item recommendation value that is greater than a preset threshold as the target item recommendation value, determine the target item information corresponding to the item recommendation value from the item-side information, and recommend the target item information to the user.

[0031] Thirdly, this application discloses an electronic device, comprising:

[0032] Memory, used to store computer programs;

[0033] A processor is used to execute the computer program to implement the aforementioned method for recommending items to multiple groups of people, in multiple scenarios, and on multiple tasks.

[0034] Fourthly, this application discloses a computer storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed method for recommending items to multiple groups of people in multiple scenarios and with multiple tasks.

[0035] As can be seen, this application provides a method for recommending items to multiple users, in multiple scenarios, and across multiple tasks. This method includes acquiring user-side information, item-side information, and corresponding scenario environment information in different scenarios; calculating weight dimensions, shared feature vectors, and environment dimensions for the user-side information, item-side information, and corresponding scenario environment information, respectively, to obtain a first dimension, a dense vector, and a second dimension; performing attention mechanism feature cross-calculation based on the first dimension and the dense vector to obtain a third dimension; calculating an item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain an item recommendation value; using the item recommendation value greater than a preset threshold as the target item recommendation value; determining the target item information corresponding to the item recommendation value from the item-side information; and recommending the target item information to the user side. This application is applied to the constructed three-dimensional perspective model that integrates multiple scenarios, multiple user groups, and multiple tasks. It calculates user-side information, item-side information, and corresponding scene environment information in different scenarios, and finally obtains the first dimension, dense vector, second dimension, and third dimension to obtain the item recommendation value. Combined with diversified information services for multiple scenarios, multiple user groups, and multiple tasks, it improves the ability to differentiate the representation of different user groups, meets the diversity of user needs, and realizes comprehensive item recommendation for multiple user groups, multiple scenarios, and multiple tasks. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0037] Figure 1 This application discloses a flowchart of a method for recommending items to multiple groups of people in multiple scenarios and with multiple tasks.

[0038] Figure 2 This is a diagram illustrating the overall structure of a multi-person, multi-scenario, multi-task model disclosed in this application.

[0039] Figure 3 This application discloses an overall flowchart of a system engineering approach for implementing item recommendation.

[0040] Figure 4 This is a schematic diagram of the structure of a multi-person, multi-scenario, multi-task item recommendation device disclosed in this application;

[0041] Figure 5 This application provides a structural diagram of an electronic device. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Current industry-standard recommendation systems with large-scale multi-scenario, multi-task models still focus on modeling multiple optimization objectives (likes, shares, forwards) across various sub-scenarios in mobile applications. They fail to explicitly incorporate the dimension of multiple user groups into the model design, resulting in a poor ability to learn generalized needs from the same user group across scenarios and the migration and evolution of interests among different user groups across scenarios. Existing multi-scenario, multi-task models in recommendation systems often distribute single-modality, similar-type items (e.g., products, videos, news, merchants) across different sub-scenarios. They cannot leverage user behavior data across multiple scenarios on mobile devices to deliver multi-modal, multi-type items according to the user's current real-time needs. Consequently, the models cannot effectively learn generalized needs from the same user group across scenarios and the migration and evolution of interests among different user groups across scenarios. Therefore, constructing a three-dimensional model encompassing multiple scenarios, user groups, and tasks to improve the ability to represent the differences among user groups, meet the diversity of user needs, and achieve comprehensive multi-user, multi-scenario, multi-task item recommendation remains a challenge in this field.

[0044] See Figure 1 As shown in the figure, this invention discloses a method for recommending items to multiple groups of people in multiple scenarios and with multiple tasks. This method is applied to a pre-defined model of multiple groups of people in multiple scenarios and with multiple tasks, and may specifically include:

[0045] Step S11: Obtain user-side information, item-side information, and corresponding scene environment information in different scenarios.

[0046] The multi-person, multi-scenario, multi-task model described in this application is constructed based on a weight dimension calculation layer, a shared feature vector calculation layer, an environment dimension calculation layer, an attention mechanism feature cross-calculation layer, and an item recommendation score calculation layer. The overall design diagram of the specific multi-person, multi-scenario, multi-task model is shown below. Figure 2 As shown, it specifically includes: a weight dimension calculation layer, a shared feature vector calculation layer, an environment dimension calculation layer, an attention mechanism feature cross calculation layer (feature cross layer based on attention mechanism), and an item recommendation score calculation layer (including feature splicing layer, multi-expert network, task weight, and fully connected layer, etc.).

[0047] In this embodiment, acquiring user-side information, item-side information, and corresponding scene environment information in different scenarios mainly involves data sampling. Data sampling focuses on the following aspects: analyzing the distribution of user groups, item consumption, and different consumption behaviors across multiple scenarios; merging sub-scenarios with fewer users; downsampling data from sub-scenarios with a large number of users and denser behaviors; sampling and differentiating data on different consumption behaviors (clicks, adding to favorites, liking, purchasing, sharing) in each scenario; and ensuring the coverage of the target user group by the data from each sub-scenarios.

[0048] User-side information encompasses user preference information, user segmentation information, basic user information, statistical features based on user behavior in various scenarios, and high-order implicit user representation vectors generated by the model, all generated based on the user's full-scenario interaction behavior sequence for all recommendation sub-scenarios. The specific user-side information is shown in Table 1:

[0049] Table 1

[0050]

[0051]

[0052] The item-side information covers all items that can be distributed and displayed to users in all recommendation sub-scenarios. It further abstracts the characteristics of different types of items to facilitate cross-calculation with user information in the model, improving generalization ability. Item types include news, hot post comments, intelligent assistant encyclopedia, intelligent assistant stock and fund selection strategies, sector analysis, instructional videos, real-time stock technical pattern events, and paid features. Specific item-side information is shown in Table 2.

[0053] Table 2

[0054]

[0055] The scene environment information includes the name of the sub-scene requested by the upstream and the sensitive environmental features of the current scene. The specific scene environment information is shown in Table 3:

[0056] Table 3

[0057] Scene environment information Coverage scenarios Sub-scene name all Mobile operating system platform all User active / passive refresh Information Flow Target of the scene Smart assistant, individual stock page Market fluctuations all Individual stock price change Smart Assistant Scene-based proactive search, input question Smart Assistant User Intent Smart Assistant

[0058] Step S12: Perform weight dimension calculation, shared feature vector calculation, and environment dimension calculation on the user-side information, the item-side information, and the corresponding scene environment information, respectively, to obtain the first dimension, the dense vector, and the second dimension.

[0059] In this embodiment, the weight dimension calculation layer is used to calculate the weight dimension of user-side information and item-side information under different scenarios to obtain the first dimension; the shared feature vector calculation layer is used to calculate the shared feature vector of user-side information, item-side information and corresponding scene environment information under different scenarios to obtain the dense vector; and the environment dimension calculation layer is used to calculate the environment dimension of scene environment information under different scenarios to obtain the second dimension.

[0060] The specific weight dimension calculation process is as follows: determine the item information from the item-side information in different scenarios, determine the user's full-scenario reading sequence from the user-side information in different scenarios, and use the weight dimension calculation layer to calculate the weight dimension of the item information and the user's full-scenario reading sequence to obtain the first dimension.

[0061] Weight Dimension Calculation Layer: The user's reading item sequence is derived from the user's reading behavior sequence across the entire application scenario. Scene features are added to calculate the similarity between the user's historical behavior sequence and the target item. The network input consists of three vectors: the representation vectors of the user's historical sequence and the target item (batch size, number of behavior sequences + 1, vector dimension), the vectors of the user's historical clicked items and the scene to which the target item belongs (batch size, number of behavior sequences + 1, vector dimension), and a mask vector representing the number of the user's historical sequences (batch size, number of behavior sequences). The representation and scene vectors of the target item are extracted and copied several times from the behavior sequence, with dimensions (batch size, number of behavior sequences, vector dimension) to ensure consistency with the historical sequence vector dimension. An attention mechanism implemented through a feedforward neural network calculates the attention score weights of the historical sequence relative to the target item in terms of representation and scene. When calculating sequence weights, it is important to incorporate the mask features of the behavior sequence and calculate the weight score on the user's actual historical sequence. The formula for calculating attention weights is as follows:

[0062]

[0063]

[0064] in, For the i items in the user's sequence, For the target item, p(B) i (This refers to the user's historical sequence.) For the scene s to which the items in the user's sequence belong, For the scene s to which the target item belongs, p(B) s () represents the number of scenes.

[0065] After applying the softmax (normalized exponential) activation function, the weights for sequence representation and scene are obtained respectively. and The dimension is (number of batches, number of behavior sequences, 1). Finally, considering the weights of representation and scenario, the output vector of the user's historical sequence is obtained. The formula for calculating the weighted output vector of the user's historical sequence is as follows:

[0066]

[0067] Multiply the user's historical sequence vector by the weights of the scene and representation, and sum them over the second dimension of the output vector. The first dimension is (number of batches, vector dimension).

[0068] The specific shared feature vector calculation process is as follows: The user-side information and item-side information under different scenarios are combined and cross-referenced to obtain cross-features; Basic item features, user features, and environmental features are determined from the user-side information, item-side information, and corresponding scene environment information under different scenarios; The shared feature vector calculation layer is used to calculate the shared feature vector for the basic item features, user features, environmental features, and cross-features to obtain the dense vector. Specifically, the combination and cross-reference involves pairwise combination and cross-reference of scene features, user group features, and important item features (concatenating two features) to generate new features (i.e., cross-features). Taking Table 4 as an example, this allows for the acquisition of potential connections between different user groups, different types of items, and scenarios.

[0069] Table 4

[0070] Original feature A Original feature B Cross features Item type Scene Name Cross1 Is it a new user? Scene Name Cross2 Item type Is it a new user? Cross3

[0071] Shared Feature Vector Calculation Layer: The shared features in the scenario include user KYC (Know Your Customer) features, target item basic features, environmental features, and cross features. After these discrete features pass through the vector generation layer, the output is a dense vector with dimensions (batch size, vector dimension). The dimension of each discrete feature in the model remains consistent, all being 16-dimensional vectors.

[0072] The specific environmental dimension calculation process is as follows: determine the scene statistical features from the scene environment information under different scenarios, and use the environmental dimension calculation layer to calculate the environmental dimension of the scene statistical features to obtain the second dimension.

[0073] Environment-specific computation layer: Scene-specific features mainly consist of statistical characteristics of user behavior in various scenes. The data distribution of these features differs significantly across different scenes; therefore, we consider modeling these features separately. Scene-specific statistical features are concatenated together, with the dimension being (number of batches, number of features). The concatenated vector is then passed through a single-layer fully connected network to obtain the output vector, with the second dimension being (number of batches, output vector dimension).

[0074] Step S13: Perform attention mechanism feature cross-calculation based on the first dimension and the dense vector to obtain the third dimension.

[0075] In this embodiment, an attention-based feature cross-layer is used for attention mechanism feature cross-computation. The input to the feature cross-layer is the scene-shared feature vector after the shared feature vector calculation layer and the output vector of the weight dimension calculation layer. The input to the feature cross-layer is n features with dimensions (batch size, 1, vector dimension). The inner product of each feature vector is used to obtain a vector V with dimensions (batch size, (n*n+1) / 2, vector dimension). Vector V is then passed through a fully connected layer of ReLU (Rectified Linear Unit) and softmax to obtain the weight 'a' of each cross feature vector, with dimensions (batch size, (n*n+1) / 2). The following formula is used:

[0076]

[0077] Where f_Att=() represents the feature intersection of the attention mechanism, x i x j V represents the i-th and j-th eigenvalues. i V j It is a vectorized representation of features i and j. The final output F of the feature cross layer is obtained by multiplying the weight vector a and the vector V and summing them in the second dimension of the vectors. The third dimension is (batch size, vector dimension).

[0078] Step S14: Calculate the item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value.

[0079] In this embodiment, the item recommendation score calculation layer in the multi-person, multi-scenario, and multi-task model is used to perform feature concatenation on the first dimension, the dense vector, the second dimension, and the third dimension to obtain a feature concatenation vector. Then, the item recommendation score is calculated on the feature concatenation vector to obtain the item recommendation value.

[0080] The item recommendation score calculation layer (which includes a feature concatenation layer, a multi-expert network, a fully connected layer, and task weights) concatenates the outputs of layers 1-4 into a single vector. The input to the multi-scene network layer is the output vector of the feature vector concatenation layer, representing both scene features and user grouping features. The output vector of the concatenation layer passes through the expert network layer to obtain a vector e composed of N expert vectors, with dimensions (batch size, vector dimension, number of experts). Simultaneously, the output vector of the feature concatenation layer passes through a two-layer fully connected weight network consisting of ReLU and softmax activation functions to obtain a task-specific weight vector g for each of the n expert vectors. t (x), dimension (number of batches, number of experts). The vector representation table of scene features and the vector representation table of user group features will pass through two fully connected layers, with ReLU and sigmoid activation functions respectively, outputting a scene-based weight vector g. s (x) and the weight vector g based on user grouping u (x), with dimensions (number of batches, number of experts). As shown in the formula:

[0081]

[0082] For each task, there is a multi-expert vector e and a scenario-based weight vector g. s (x), the weight vector g based on user grouping u (x), the multi-objective weights g corresponding to the task t (x), the output vector of each task is obtained by weighted summation along the dimension of the number of experts i, with the dimension being (number of batches, vector dimension).

[0083] The multi-person, multi-scenario, multi-task model in this application is a multi-task model with CTR (Click-Through-Rate) and CVR (Conversion Rate) tasks. The vector representations of the obtained CTR and CVR tasks are then passed through three fully connected layers with ReLU activation function and then through the sigmoid activation function to obtain the CTR and CVR scores. Multiplying the CTR score and CVR score gives the exposure-click-conversion rate (CTCVR) score.

[0084] Step S15: Take the item recommendation value that is greater than the preset threshold as the target item recommendation value, determine the target item information corresponding to the item recommendation value from the item-side information, and recommend the target item information to the user side.

[0085] The entire process of recommending target item information using the multi-person, multi-scenario, and multi-task model in this application is as follows: Figure 3 As shown, (1) when a user enters any scenario, they first request a unified user understanding service; (2) the user understanding service generates user requirements based on the current user ID (Identity document) and the user's historical information across all scenarios using a generative model, and sends the user requirements to the item center so that the item center can determine the item-side information; (3) it generates user profile information, user group information, user representation vector and other user-side data, and sends the user-side information to the data engineering department; (4) the data engineering department performs data sampling, aggregation and preprocessing operations on the user-side information, item-side information and scenario information; (5) using the multi-person multi-scenario multi-task model of this application, it scores the items according to the above-mentioned item recommendation method of this application to determine the target item recommendation value; (6) the target item information is recommended to the user.

[0086] In addition, this application can also evaluate multi-person, multi-scenario, and multi-task models: the offline evaluation of the model mainly evaluates the quality of multi-scenario and multi-person modeling schemes from two dimensions: scenario and user group. The main process is as follows: (1) Training data preparation: prepare training datasets for each scenario and multi-scenario training datasets that integrate each scenario; (2) Validation data preparation: prepare validation datasets for each scenario and multi-scenario validation datasets that integrate each scenario; (3) Test data preparation: prepare test datasets for each scenario and test datasets that differentiate people in each scenario (such as new user datasets and old user datasets); (4) Model training: train independent models for each scenario using training datasets and validation datasets for each scenario (different from multi-scenario models); train multi-scenario and multi-person models using multi-scenario training datasets and multi-scenario validation datasets. If the AUC (Area Under the Curve) of each model does not improve for 5 consecutive epochs (hyperparameters), training is stopped and the offline model is exported; (5) Model evaluation: use the model for each scenario and the multi-scenario model to predict the test data for each scenario and use AUC to evaluate the quality of the model. After completing the offline model training and evaluation, and the online ranking service code development and testing, the model is exported as a .pb file and uploaded to HDFS (Hadoop Distributed File System). TFServing (a high-performance open-source library) is then started to load the model and enable real-time inference for ranking services in various online scenarios.

[0087] In this embodiment, user-side information, item-side information, and corresponding scene environment information under different scenarios are obtained; weight dimension calculation, shared feature vector calculation, and environment dimension calculation are performed on the user-side information, the item-side information, and the corresponding scene environment information to obtain a first dimension, a dense vector, and a second dimension; attention mechanism feature cross calculation is performed based on the first dimension and the dense vector to obtain a third dimension; an item recommendation score is calculated using the first dimension, the dense vector, the second dimension, and the third dimension to obtain an item recommendation value; the item recommendation value greater than a preset threshold is taken as the target item recommendation value, the target item information corresponding to the item recommendation value is determined from the item-side information, and the target item information is recommended to the user side. This application is applied to the constructed three-dimensional perspective model that integrates multiple scenarios, multiple user groups, and multiple tasks. It calculates user-side information, item-side information, and corresponding scene environment information in different scenarios, and finally obtains the first dimension, dense vector, second dimension, and third dimension to obtain the item recommendation value. Combined with diversified information services for multiple scenarios, multiple user groups, and multiple tasks, it improves the ability to differentiate the representation of different user groups, meets the diversity of user needs, and realizes comprehensive item recommendation for multiple user groups, multiple scenarios, and multiple tasks.

[0088] This application constructs a model from a three-dimensional perspective of multiple scenarios, multiple groups, and multiple tasks, enabling the model to have the ability to represent different user groups in a differentiated manner. This is achieved in the following two aspects: (1) Feature construction: By explicitly splicing and crossing user group features and scenario-related features, the association between user groups and scenarios is explicitly constructed. For example, the values ​​of user group features (new and old users) and scenario name features are spliced ​​together to generate new features, which are then input into the model; (2) Model design: User group features are input into a weight generation network composed of fully connected networks to generate the weight coefficients of user group features in different tasks and scenarios, so that the model can better learn the preferences of different user groups in different scenarios. For heterogeneous items with various data structures such as long text, short text, video, chart, and image, which are distributed in multiple scenarios according to user needs, this invention innovates in the following two aspects: (1) Data feature construction: The features of different forms of items are further abstracted so that they can be cross-calculated with user information in the model to improve generalization ability. For example, the service links, service areas, preferred groups, and service styles of all items are defined, and these features correspond to the preferences of different user groups. For information, Q&A titles, videos, charts and comments on intelligent Q&A results pages, we extract semantic features such as entity words, financial indicators, and emotional intent through natural language, image, and speech technologies. These two parts of features define the explicit representation of diverse heterogeneous items, and then use the dense vector encoding layer of the deep model to perform unified vector representation of different types of heterogeneous items in different scenarios; (2) In terms of model design: by performing vector cross-operation on the item business class label representation vector and the user business preference class feature in the model layer, the recommendation ranking model can more efficiently distribute the cold start of newly deployed item types across scenarios to specific users. By constructing a feature cross-layer based on the attention mechanism of item feature vector and scenario feature vector, the model can better learn to generalize the preference distribution of the same item type being consumed in different scenarios.

[0089] See Figure 4 As shown, this embodiment of the invention discloses a multi-person, multi-scenario, multi-task item recommendation device, applied to a preset multi-person, multi-scenario, multi-task model, which may specifically include:

[0090] Information acquisition module 11 is used to acquire user-side information, item-side information and corresponding scene environment information in different scenarios;

[0091] The calculation module 12 is used to perform weight dimension calculation, shared feature vector calculation and environment dimension calculation on the user-side information, the item-side information and the corresponding scene environment information, respectively, to obtain the first dimension, the dense vector and the second dimension.

[0092] The feature cross calculation module 13 is used to perform attention mechanism feature cross calculation based on the first dimension and the dense vector to obtain the third dimension;

[0093] The recommendation value calculation module 14 is used to calculate the item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value.

[0094] The item information recommendation module 15 is used to take the item recommendation value that is greater than a preset threshold as the target item recommendation value, determine the target item information corresponding to the item recommendation value from the item-side information, and recommend the target item information to the user side.

[0095] In this embodiment, user-side information, item-side information, and corresponding scene environment information under different scenarios are obtained; weight dimension calculation, shared feature vector calculation, and environment dimension calculation are performed on the user-side information, the item-side information, and the corresponding scene environment information to obtain a first dimension, a dense vector, and a second dimension; attention mechanism feature cross calculation is performed based on the first dimension and the dense vector to obtain a third dimension; an item recommendation score is calculated using the first dimension, the dense vector, the second dimension, and the third dimension to obtain an item recommendation value; the item recommendation value greater than a preset threshold is taken as the target item recommendation value, the target item information corresponding to the item recommendation value is determined from the item-side information, and the target item information is recommended to the user side. This application is applied to the constructed three-dimensional perspective model that integrates multiple scenarios, multiple user groups, and multiple tasks. It calculates user-side information, item-side information, and corresponding scene environment information in different scenarios, and finally obtains the first dimension, dense vector, second dimension, and third dimension to obtain the item recommendation value. Combined with diversified information services for multiple scenarios, multiple user groups, and multiple tasks, it improves the ability to differentiate the representation of different user groups, meets the diversity of user needs, and realizes comprehensive item recommendation for multiple user groups, multiple scenarios, and multiple tasks.

[0096] In some specific embodiments, the multi-person, multi-scenario, multi-task model is constructed based on a weight dimension calculation layer, a shared feature vector calculation layer, an environment dimension calculation layer, an attention mechanism feature cross calculation layer, and an item recommendation score calculation layer.

[0097] In some specific embodiments, the computing module 12 may specifically include:

[0098] The first dimension determination module is used to calculate the weight dimension of user-side information and item-side information in different scenarios using the weight dimension calculation layer to obtain the first dimension;

[0099] The dense vector determination module is used to perform shared feature vector calculation on the user-side information, the item-side information and the corresponding scene environment information under different scenarios using the shared feature vector calculation layer, so as to obtain a dense vector;

[0100] The second dimension determination module is used to perform environmental dimension calculation on the scene environment information under different scenarios using the environmental dimension calculation layer to obtain the second dimension.

[0101] In some specific embodiments, the computing module 12 may specifically include:

[0102] The weight dimension calculation module is used to determine item information from the item-side information in different scenarios and to determine the user's full-scenario reading sequence from the user-side information in different scenarios. The weight dimension calculation layer is used to calculate the weight dimension of the item information and the user's full-scenario reading sequence to obtain the first dimension.

[0103] In some specific embodiments, the computing module 12 may specifically include:

[0104] The cross-feature determination module is used to combine and cross-reference the user-side information and the item-side information in different scenarios to obtain cross-features;

[0105] The shared feature vector calculation module is used to determine the basic features of the item, the user features, and the environmental features from the user-side information, the item-side information, and the corresponding scene environment information in different scenarios, respectively. The shared feature vector calculation layer is used to calculate the shared feature vector of the basic features of the item, the user features, the environmental features, and the cross features to obtain the dense vector.

[0106] In some specific embodiments, the computing module 12 may specifically include:

[0107] The environment dimension calculation module is used to determine scene statistical features from the scene environment information under different scenarios, and to perform environment dimension calculation on the scene statistical features using the environment dimension calculation layer to obtain the second dimension.

[0108] In some specific embodiments, the item information recommendation module 13 may specifically include:

[0109] The item recommendation score calculation module is used to perform feature concatenation on the first dimension, the dense vector, the second dimension, and the third dimension using the item recommendation score calculation layer in the multi-person, multi-scenario, and multi-task model to obtain a feature concatenation vector. Then, the item recommendation score is calculated on the feature concatenation vector to obtain the item recommendation value.

[0110] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the multi-person, multi-scenario, multi-task item recommendation method performed by the electronic device as disclosed in any of the foregoing embodiments.

[0111] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0112] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.

[0113] The operating system 221 manages and controls the various hardware devices on the electronic device 20 and the computer program 222 to enable the processor 21 to perform calculations and processing on the data 223 in the memory 22. It can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the multi-person, multi-scenario, multi-task item recommendation method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the multi-person, multi-scenario, multi-task item recommendation device from external devices, and data collected by its own input / output interface 25, etc.

[0114] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0115] Furthermore, this application also discloses a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the multi-person, multi-scenario, and multi-task item recommendation method disclosed in any of the foregoing embodiments.

[0116] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0117] The above provides a detailed description of a method, apparatus, device, and storage medium for recommending items to multiple groups of people, in multiple scenarios, and for multiple tasks. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for recommending items to multiple groups of people, in multiple scenarios, and across multiple tasks, characterized in that, Applied to pre-defined multi-person, multi-scenario, multi-task models, including: Acquire user-side information, item-side information, and corresponding scene environment information in different scenarios; The user-side information, the item-side information, and the corresponding scene environment information are respectively subjected to weight dimension calculation, shared feature vector calculation, and environment dimension calculation to obtain the first dimension, dense vector, and second dimension; The third dimension is obtained by performing attention mechanism feature cross-calculation based on the first dimension and the dense vector; The item recommendation score is calculated using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value; The item recommendation value that is greater than a preset threshold is taken as the target item recommendation value. The target item information corresponding to the item recommendation value is determined from the item-side information, and the target item information is recommended to the user. The multi-person, multi-scenario, multi-task model is constructed based on a weight dimension calculation layer, a shared feature vector calculation layer, an environment dimension calculation layer, an attention mechanism feature cross calculation layer, and an item recommendation score calculation layer. The step of performing weight dimension calculation, shared feature vector calculation, and environmental dimension calculation on the user-side information, the item-side information, and the corresponding scene environment information to obtain a first dimension, a dense vector, and a second dimension includes: using the weight dimension calculation layer to perform weight dimension calculation on the user-side information and item-side information under different scenarios to obtain a first dimension; using the shared feature vector calculation layer to perform shared feature vector calculation on the user-side information, the item-side information, and the corresponding scene environment information under different scenarios to obtain a dense vector; and using the environmental dimension calculation layer to perform environmental dimension calculation on the scene environment information under different scenarios to obtain a second dimension.

2. The item recommendation method for multiple groups, multiple scenarios, and multiple tasks according to claim 1, characterized in that, The step of using the weighted dimension calculation layer to calculate the weighted dimension of user-side information and item-side information in different scenarios to obtain the first dimension includes: Item information is determined from the item-side information in different scenarios, and the user's full-scenario reading sequence is determined from the user-side information in different scenarios. The weight dimension calculation layer is used to calculate the weight dimension of the item information and the user's full-scenario reading sequence to obtain the first dimension.

3. The item recommendation method for multiple groups, multiple scenarios, and multiple tasks according to claim 1, characterized in that, The step of using the shared feature vector calculation layer to calculate shared feature vectors for the user-side information, the item-side information, and the corresponding scene environment information under different scenarios to obtain dense vectors includes: The user-side information and the item-side information in different scenarios are combined and cross-referenced to obtain cross-features; The basic features of the item, the basic features of the user, the basic features of the item, and the environmental features of the corresponding scene environment are determined from the user-side information, the item-side information, and the corresponding scene environment information in different scenarios. The shared feature vector calculation layer is used to calculate the shared feature vector of the basic features of the item, the user features, the environmental features, and the cross features to obtain the dense vector.

4. The item recommendation method for multiple groups, multiple scenarios, and multiple tasks according to claim 1, characterized in that, The step of using the environment dimension calculation layer to perform environment dimension calculations on the scene environment information under different scenarios to obtain the second dimension includes: Scene statistical features are determined from the scene environment information under different scenarios, and the environment dimension calculation layer is used to calculate the environment dimension of the scene statistical features to obtain the second dimension.

5. The method for recommending items to multiple groups, multiple scenarios, and multiple tasks according to any one of claims 1 to 4, characterized in that, The step of calculating the item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value includes: The item recommendation score calculation layer in the multi-person, multi-scenario, and multi-task model is used to concatenate features of the first dimension, the dense vector, the second dimension, and the third dimension to obtain a feature concatenation vector. Then, the item recommendation score is calculated on the feature concatenation vector to obtain the item recommendation value.

6. A multi-person, multi-scenario, multi-task item recommendation device, characterized in that, Applied to pre-defined multi-person, multi-scenario, multi-task models, including: The information acquisition module is used to acquire user-side information, item-side information, and corresponding scene environment information in different scenarios. The calculation module is used to perform weight dimension calculation, shared feature vector calculation and environment dimension calculation on the user-side information, the item-side information and the corresponding scene environment information, respectively, to obtain the first dimension, the dense vector and the second dimension; The feature cross calculation module is used to perform attention mechanism feature cross calculation based on the first dimension and the dense vector to obtain the third dimension; The recommendation value calculation module is used to calculate the item recommendation score using the first dimension, the dense vector, the second dimension, and the third dimension to obtain the item recommendation value. The item information recommendation module is used to take the item recommendation value that is greater than a preset threshold as the target item recommendation value, determine the target item information corresponding to the item recommendation value from the item-side information, and recommend the target item information to the user side. The multi-person, multi-scenario, multi-task model is constructed based on a weight dimension calculation layer, a shared feature vector calculation layer, an environment dimension calculation layer, an attention mechanism feature cross calculation layer, and an item recommendation score calculation layer. The step of performing weight dimension calculation, shared feature vector calculation, and environmental dimension calculation on the user-side information, the item-side information, and the corresponding scene environment information to obtain a first dimension, a dense vector, and a second dimension includes: using the weight dimension calculation layer to perform weight dimension calculation on the user-side information and item-side information under different scenarios to obtain a first dimension; using the shared feature vector calculation layer to perform shared feature vector calculation on the user-side information, the item-side information, and the corresponding scene environment information under different scenarios to obtain a dense vector; and using the environmental dimension calculation layer to perform environmental dimension calculation on the scene environment information under different scenarios to obtain a second dimension.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor is configured to execute the computer program to implement the multi-person, multi-scenario, multi-task item recommendation method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the multi-person, multi-scenario, multi-task item recommendation method as described in any one of claims 1 to 5.

Images