Training method and device of recommendation model

Abstract

Embodiments of the present specification provide a recommendation model training method and device. The recommendation model training method comprises: obtaining a mixed scene sample set, the mixed scene sample set comprising feature data of a plurality of recommendation items, and the feature data of the plurality of recommendation items corresponding to a plurality of different scenes; performing model training on an initial recommendation model according to the mixed scene sample set to obtain an intermediate recommendation model; the initial recommendation model comprising a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit comprising a plurality of expert sub-units, a gating sub-unit corresponding to each recommendation target, and a normalization sub-unit corresponding to each recommendation target; obtaining a plurality of single-scene sample sets; performing model training on the intermediate recommendation model according to each single-scene sample set to obtain a recommendation model; and the recommendation model being used to determine a recommendation item that satisfies a multi-target recommendation condition determined by a preset scene and each recommendation target in the preset scene.

Description

Technical Field

[0001] This document relates to the field of machine learning, and in particular to a method and apparatus for training a recommendation model. Background Technology

[0002] With the development of internet technology, recommendation models are being applied more and more widely. On the one hand, in the same recommendation scenario, it is often necessary to consider multiple different recommendation goals. If recommendation predictions are made separately according to each recommendation goal, conflicting prediction results are likely to be obtained. On the other hand, in different recommendation scenarios, the recommendation goals to be considered may be different, and the number of users accessing the platform may also vary greatly. Therefore, multi-goal models or multi-scenario models have certain limitations and cannot meet the increasingly complex and diverse recommendation needs. Summary of the Invention

[0003] This specification provides one or more embodiments of a method for training a recommendation model. The method includes: acquiring a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items for multiple scenarios. Based on the mixed-scene sample set, an initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scene unit corresponding to each recommendation objective in a preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. Multiple single-scene sample sets are acquired; each single-scene sample set includes feature data of a recommendation item for one scenario. Based on each single-scene sample set, the intermediate recommendation model is trained to obtain a recommendation model; the recommendation model is used to determine recommendation items in a preset scenario that satisfy multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0004] This specification provides an information recommendation method through one or more embodiments. The information recommendation method includes: obtaining a set of candidate recommendation items for a target scene; the set of candidate recommendation items includes feature data of multiple candidate recommendation items. The candidate recommendation item set is input into the recommendation model for prediction processing to obtain recommendation item information that satisfies the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each recommendation target; the recommendation model is trained in the following manner: obtaining a mixed scenario sample set; the mixed scenario sample set includes feature data of recommendation items for multiple scenarios; training an initial recommendation model based on the mixed scenario sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation target in the preset target set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation target, and a normalization sub-unit corresponding to each recommendation target; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario; obtaining multiple single-scenario sample sets; each single-scenario sample set includes feature data of recommendation items for one scenario; training the intermediate recommendation model based on each single-scenario sample set to obtain the recommendation model.

[0005] This specification provides one or more embodiments of a training apparatus for a recommendation model, comprising: a first sample acquisition module configured to acquire a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items for multiple scenarios. A first model training module configured to train an initial recommendation model based on the mixed-scene sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scene unit corresponding to each recommendation objective in a preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. A second sample acquisition module configured to acquire multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for one scenario. A second model training module configured to train the intermediate recommendation model based on each single-scene sample set to obtain a recommendation model; the recommendation model is used to determine recommendation items in a preset scenario that satisfy multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0006] This specification provides an information recommendation device in one or more embodiments, including: a set acquisition module configured to acquire a set of candidate recommendation items for a target scene; the set of candidate recommendation items includes feature data of multiple candidate recommendation items. An ensemble prediction module is configured to input the candidate recommendation item set into a recommendation model for prediction processing to obtain recommendation item information that satisfies multi-objective recommendation conditions. The multi-objective recommendation conditions are determined by the target scenario and each recommendation target. The recommendation model is trained as follows: A mixed scenario sample set is acquired; the mixed scenario sample set includes feature data of recommendation items from multiple scenarios. Based on the mixed scenario sample set, an initial recommendation model is trained to obtain an intermediate recommendation model. The initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation target in a preset target set. The multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation target, and a normalization sub-unit corresponding to each recommendation target. The multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. Multiple single-scenario sample sets are acquired; each single-scenario sample set includes feature data of a recommendation item from one scenario. Based on each single-scenario sample set, the intermediate recommendation model is trained to obtain the recommendation model.

[0007] This specification provides one or more embodiments of a training device for a recommendation model, comprising: a processor; and a memory configured to store computer-executable instructions, which, when executed, cause the processor to: acquire a mixed-scene sample set; the mixed-scene sample set including feature data of recommendation items for multiple scenarios. Based on the mixed-scene sample set, train an initial recommendation model to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scene unit corresponding to each recommendation objective in a preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. Acquire multiple single-scene sample sets; each single-scene sample set includes feature data of recommendation items for one scenario. Based on each single-scene sample set, train the intermediate recommendation model to obtain a recommendation model; the recommendation model is used to determine recommendation items in a preset scenario that satisfy multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0008] This specification provides an information recommendation device according to one or more embodiments, including: a processor; and a memory configured to store computer-executable instructions, which, when executed, cause the processor to: acquire a set of candidate recommendations for a target scene; the set of candidate recommendations includes feature data of a plurality of candidate recommendations. The candidate recommendation item set is input into the recommendation model for prediction processing to obtain recommendation item information that satisfies the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each recommendation target; the recommendation model is trained in the following manner: obtaining a mixed scenario sample set; the mixed scenario sample set includes feature data of recommendation items for multiple scenarios; training an initial recommendation model based on the mixed scenario sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation target in the preset target set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation target, and a normalization sub-unit corresponding to each recommendation target; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario; obtaining multiple single-scenario sample sets; each single-scenario sample set includes feature data of recommendation items for one scenario; training the intermediate recommendation model based on each single-scenario sample set to obtain the recommendation model.

[0009] This specification provides one or more embodiments of a storage medium for storing computer-executable instructions, which, when executed by a processor, implement the following process: acquiring a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items for multiple scenarios. Based on the mixed-scene sample set, an initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scene unit corresponding to each recommendation objective in a preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. Acquiring multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for one scenario. Based on each single-scene sample set, the intermediate recommendation model is trained to obtain a recommendation model; the recommendation model is used to determine recommendation items in a preset scenario that satisfy multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0010] This specification provides one or more embodiments of a storage medium for storing computer-executable instructions, which, when executed by a processor, perform the following process: obtaining a set of candidate recommendations for a target scene; the set of candidate recommendations includes feature data of multiple candidate recommendations. The candidate recommendation item set is input into the recommendation model for prediction processing to obtain recommendation item information that satisfies the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each recommendation target; the recommendation model is trained in the following manner: obtaining a mixed scenario sample set; the mixed scenario sample set includes feature data of recommendation items for multiple scenarios; training an initial recommendation model based on the mixed scenario sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation target in the preset target set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation target, and a normalization sub-unit corresponding to each recommendation target; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario; obtaining multiple single-scenario sample sets; each single-scenario sample set includes feature data of recommendation items for one scenario; training the intermediate recommendation model based on each single-scenario sample set to obtain the recommendation model. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in one or more embodiments of this specification 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 some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating the training method of a recommendation model provided in one or more embodiments of this specification;

[0013] Figure 2 A schematic diagram illustrating the data flow for training an initial recommendation model, provided in one or more embodiments of this specification.

[0014] Figure 3 A schematic diagram illustrating the data flow for training an intermediate recommendation model, provided in one or more embodiments of this specification.

[0015] Figure 4 A gradient backhaul diagram provided for one or more embodiments of this specification;

[0016] Figure 5A flowchart illustrating an information recommendation method provided in one or more embodiments of this specification;

[0017] Figure 6 A schematic diagram of a training device for a recommendation model provided in one or more embodiments of this specification;

[0018] Figure 7 A schematic diagram of an information recommendation device provided for one or more embodiments of this specification;

[0019] Figure 8 This is a schematic diagram of the structure of a training device for a recommended model provided in one or more embodiments of this specification.

[0020] Figure 9 This is a schematic diagram of the structure of an information recommendation device provided for one or more embodiments of this specification. Detailed Implementation

[0021] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0022] This specification provides an example of a training method for a recommendation model:

[0023] The training method for the recommendation model provided in this embodiment involves acquiring a mixed-scene sample set and using this mixed-scene sample set as common model input data for different recommendation objectives and different scenarios; training an initial recommendation model based on the mixed-scene sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scene unit corresponding to each recommendation objective in a preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. Through the above model structure, the underlying multi-objective structure receives full-scene batch samples for training, establishing a heterogeneous multi-objective model; multiple single-scene sample sets are acquired; each single-scene sample set... This includes feature data for recommendations within a single scenario; based on each single-scenario sample set, intermediate recommendation models are trained separately to obtain recommendation models. Training data is aggregated according to scenarios, meaning each batch represents the same scenario. Training is conducted intensively on the mapping relationships of single scenarios with a small number of training parameters, leveraging the large information volume of the entire batch, to obtain a refined objective function mapping relationship for each scenario. This allows for the introduction of a large amount of scenario data through a unified standard data stream, with common parts of the model used for full-scenario description, truly enriching the matching between users and different recommendation scenarios, better meeting user needs. It simplifies the model architecture to achieve information gain sharing while maintaining the uniqueness of scenario and objective information; it resolves gradient conflicts through partial gradient backpropagation; and it enables the training of a unified multi-objective, multi-scenario model to serve a large number of online scenarios, significantly reducing computational resources and enhancing model real-time performance.

[0024] Reference Figure 1 The training method for the recommendation model provided in this embodiment specifically includes steps S102 to S108.

[0025] Step S102: Obtain a mixed scene sample set; the mixed scene sample set includes feature data of recommendation items from multiple scenes.

[0026] The mixed scenario sample set can include multiple mixed scenario samples, each of which can include feature data of a recommendation item for a scenario, and the mixed scenario sample set corresponds to at least two different scenarios.

[0027] In multiple scenarios, each scenario can be a recommendation scenario that recommends people, things, or other items to users.

[0028] In addition, based on the number of users accessing the recommended scenario, the recommended scenario can be divided into large scenarios and small scenarios.

[0029] Specifically, scenarios where the number of users accessing the app exceeds a preset threshold can be defined as large scenarios. For example, the number of users accessing the homepage of a shopping app is in the tens of millions, and the product recommendations on the homepage correspond to a large scenario. Alternatively, recommendation scenarios where the number of users accessing the app is less than the preset threshold can be defined as small scenarios. For example, if a user enters the "My XX" page from the shopping app homepage and then enters the first subpage from the "My XX" page, and the number of users accessing the first subpage is in the tens of thousands, the product recommendations on the first subpage correspond to a small scenario.

[0030] Recommendations can be people, for example, a social media platform recommends other users that the user might want to follow; these other users are the social media platform's recommendations. Recommendations can be things, for example, a shopping platform recommends products that the user might be interested in; these products are the shopping platform's recommendations. Recommendations can be news, for example, a news app recommends articles that the user might be interested in; these articles are the news app's recommendations, and so on.

[0031] Each scenario can have one or more recommendations.

[0032] For example, social media platforms can recommend other users that users are interested in, as well as advertisements, and so on.

[0033] The feature data for each recommendation item can be data that reflects the characteristics that distinguish it from other recommendations of the same type. For example, in the product recommendation scenario on the homepage of a shopping app, the feature data for product 1 includes, but is not limited to, its size, weight, color, style, price, brand, store name, etc.

[0034] Furthermore, when recommending people, objects, or other items to users in recommendation scenarios, not only is the feature data of the recommended items needed, but also the user's preferences. Therefore, each hybrid scenario sample can also include the user's feature data. For example, the user's feature data can include the user's personal attribute information, the user's likes history, the user's feedback on historical recommendations, and so on.

[0035] It should be noted that, considering that the aforementioned user characteristic data may involve user privacy, in order to protect user privacy and security, this characteristic data can only be obtained with the corresponding user's authorization.

[0036] Each mixed-scenario sample can also include a recommendation tag, which can reflect the user's level of approval for the recommended items.

[0037] Multiple mixed-scene samples in a mixed-scene sample can correspond to the same recommendation item.

[0038] Step S104: Train the initial recommendation model based on the mixed scenario sample set to obtain an intermediate recommendation model. The initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in the preset objective set. The multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalized sub-unit corresponding to each recommendation objective. The multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario.

[0039] Based on the mixed scenario sample set, the initial recommendation model is trained to obtain an intermediate recommendation model. This can be done by iteratively training the initial recommendation model or by using other predetermined model training methods.

[0040] The initial recommendation model includes at least: multi-objective units and multi-scene units corresponding to each recommendation objective in the preset objective set.

[0041] The preset target set can include multiple recommended targets. Each of these recommended targets can correspond to a recommendation parameter, which can be used to sort multiple identical recommended items. Furthermore, each recommended target can correspond to a multi-scene unit.

[0042] In recommendation scenarios, it is very common to have more than one recommendation goal. When making recommendations to users, it is often necessary to consider multiple recommendation goals, and there may be contradictions between the various recommendation goals. For example, a user may like to see a certain type of advertisement but may not be willing to buy the corresponding product.

[0043] In one specific implementation, the recommendation item is multimedia data; the recommendation parameters include at least one of the following: click pass rate, conversion rate, and playback duration.

[0044] In practice, the recommended items can be multimedia data, and the recommended parameters can be CTR (Click-Through-Rate), CVR (Conversion Rate), GMV (Gross Merchandise Volume), playback duration, etc.

[0045] CTR can be considered the click-through rate of an online advertisement, determined by dividing the actual number of clicks on the advertisement by the number of times the advertisement was displayed. The actual number of clicks on an online advertisement can be the number of times the advertisement reached the corresponding target page. Online advertisements can be at least one of several types, including image ads, text ads, keyword ads, ranking ads, and video ads.

[0046] For example, after entering keywords into a search engine, the relevant web pages are sorted in order based on factors such as bidding, and users will choose the websites they are interested in and click on them; the total number of searches for a website is called the total number of searches, and the percentage of times users click on and enter the website is called the click-through rate (CTR).

[0047] CVR can be the conversion rate from a user clicking on an ad to becoming a valid activated, registered, or even paying user.

[0048] GMV can be the total amount of transactions within a specified time period, including the amount of paid orders and the amount of unpaid orders. Specifically, GMV can include one or more of the following: sales revenue, cancelled order amount, rejected order amount, and returned order amount.

[0049] When the recommendation target corresponds to CTR, the recommendation target can be used to indicate the recommendation order of the corresponding recommendation items based on the numerical value of CTR. Other recommendation parameters can refer to CTR.

[0050] A multi-objective unit may include multiple expert sub-units, a gating sub-unit corresponding to each recommended objective, and a normalization sub-unit corresponding to each recommended objective.

[0051] The expert subunit can be an Expert model, which is a neural network. It's important to note that this Expert model is part of the initial recommendation model, not a separate model from it.

[0052] It is important to note that there is a pre-configured correspondence between expert sub-units and recommendation targets.

[0053] For example, each recommendation target corresponds to one of the multiple expert sub-units, or each recommendation target corresponds to two of the multiple expert sub-units, and so on.

[0054] The correspondence between expert subunits and recommendation targets is illustrated in the following example: a preset target set includes recommendation target 1 and recommendation target 2, and there are three expert subunits: Expert A, Expert B, and Expert C. Recommendation target 1 corresponds to Expert A and Expert B, and recommendation target 2 corresponds to Expert B and Expert C.

[0055] The gated subunit can be a Gate network, which is another type of neural network.

[0056] The normalization subunit can be a Softmax layer, which is used to "compress" a K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ(z), such that each element is in the range (0,1) and the sum of all elements is 1.

[0057] A multi-scenario unit may include a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario.

[0058] Shared network subunits can be common NNs (Neural Networks), i.e., shared neural networks.

[0059] In multiple scenarios, the private network subunit corresponding to the i-th scenario can be a scene. i NN stands for the private neural network of the i-th scene. i can be a natural number greater than 0, and the maximum value of i is determined by the number of scenes corresponding to these multiple scenes.

[0060] For example, the mixed scene sample set corresponds to 10 scenes, meaning the maximum value of i is 10. Within these 10 scenes, the private network subunit corresponding to the first scene could be scene1 NN, the private network subunit corresponding to the second scene could be scene2 NN, and so on, up to scene1 NN. 10 The 10 scenes share a single common neural network (NN). That is, within each scene unit corresponding to a recommendation goal, there are common NNs, scene1NNs, scene2NNs, and so on. 10 NN.

[0061] The fully connected subunit can be a Scenario-Specific FCN (Scenario-Specific FullyConvolutional Networks).

[0062] In one specific implementation, the model parameters of the fully connected subunit in the initial recommendation model are determined based on the shared parameters of the corresponding shared network subunit and the private parameters of the corresponding private network subunit; the shared network subunit corresponding to the fully connected subunit corresponds to the same recommendation goal as the fully connected subunit; and the private network subunit corresponding to the fully connected subunit corresponds to the same scenario and the same recommendation goal as the fully connected subunit.

[0063] The shared parameters of a shared network subunit can be network parameters determined based on that shared network subunit; the private parameters of a private network subunit can be network parameters determined based on that private network subunit.

[0064] For example, the preset target set includes recommended target 1 and recommended target 2.

[0065] The mixed scene sample set corresponds to 5 scenes: Scene 1, Scene 2... Scene 5.

[0066] The multi-scenario units corresponding to the recommended objective 1 include: shared network subunit 1, private network subunit 1 corresponding to the first scenario, private network subunit 2 corresponding to the second scenario, ..., private network subunit 5 corresponding to the fifth scenario, and fully connected subunit 1 corresponding to the first scenario, fully connected subunit 2 corresponding to the second scenario, ..., fully connected subunit 5 corresponding to the fifth scenario.

[0067] The multi-scenario units corresponding to the recommended objective 2 include: shared network subunit 2, private network subunit 6 corresponding to the first scenario, private network subunit 7 corresponding to the second scenario, ..., private network subunit 10 corresponding to the fifth scenario, and fully connected subunit 6 corresponding to the first scenario, fully connected subunit 7 corresponding to the second scenario, ..., fully connected subunit 10 corresponding to the fifth scenario.

[0068] The model parameters of the fully connected subunit 1 are determined based on the shared parameters of the shared network subunit 1 and the private parameters of the private network subunit 1;

[0069] The model parameters of the fully connected subunit 2 are determined based on the shared parameters of the shared network subunit 1 and the private parameters of the private network subunit 2;

[0070] ...

[0071] The model parameters of the fully connected subunit 5 are determined based on the shared parameters of the shared network subunit 1 and the private parameters of the private network subunit 5;

[0072] The model parameters of the fully connected subunit 6 are determined based on the shared parameters of the shared network subunit 2 and the private parameters of the private network subunit 6;

[0073] The model parameters of the fully connected subunit 7 are determined based on the shared parameters of the shared network subunit 2 and the private parameters of the private network subunit 7;

[0074] ...

[0075] The model parameters of the fully connected subunit 10 are determined based on the shared parameters of the shared network subunit 2 and the private parameters of the private network subunit 10.

[0076] It is important to note that during the process of training the initial recommendation model based on the mixed scenario sample set to obtain the intermediate recommendation model, the model parameters of the fully connected sub-units in the multi-scenario units corresponding to each recommendation target are updated based on shared parameters, private parameters, and the output data of the multi-target units.

[0077] In one specific implementation, each recommendation target corresponds to a preset number of expert sub-units; the multi-target unit also includes a computation unit corresponding to each recommendation target; the gating sub-unit corresponding to the same recommendation target is connected to the normalization sub-unit, and the output of the gating sub-unit is the input of the normalization sub-unit; the normalization sub-unit corresponding to the same recommendation target is connected to the computation unit, and the output of the normalization sub-unit is the input of the computation unit; the preset number of expert sub-units corresponding to the same recommendation target are connected to the computation unit, and the output of the expert sub-units is the input of the computation unit.

[0078] The preset quantity can be 2, or it can be a natural number greater than 2.

[0079] For example, the preset target set includes recommended target 1 and recommended target 2, and the number of multiple expert sub-units is 3, namely Expert A, Expert B and Expert C. Among them, recommended target 1 corresponds to gating sub-unit 1, normalization sub-unit 1, Expert A, Expert B and calculation unit 1; recommended target 2 corresponds to gating sub-unit 2, normalization sub-unit 2, Expert B, Expert C and calculation unit 2.

[0080] The gating subunit 1 corresponding to the recommended target 1 is connected to the normalization subunit 1, and the output of the gating subunit 1 is the input of the normalization subunit 1.

[0081] The gating subunit 2 corresponding to the recommended target 2 is connected to the normalization subunit 2, and the output of the gating subunit 2 is the input of the normalization subunit 2.

[0082] Normalization subunit 1, corresponding to recommendation target 1, is connected to computation unit 1, and the output of normalization subunit 1 is the input of computation unit 1. Furthermore, Expert A and Expert B, corresponding to recommendation target 1, are also connected to computation unit 1, with the output of Expert A and the output of Expert B serving as inputs to computation unit 1. Therefore, the input of computation unit 1 includes the output of normalization subunit 1, the output of Expert A, and the output of Expert B.

[0083] Normalization subunit 2, corresponding to recommendation target 2, is connected to computation unit 2, and its output serves as the input to computation unit 2. Furthermore, Expert B and Expert C, also corresponding to recommendation target 2, are connected to computation unit 2, with Expert B's output and Expert C's output serving as inputs to computation unit 2. Therefore, the input to computation unit 2 includes the outputs of normalization subunit 2, Expert B, and Expert C.

[0084] In one specific implementation, the expert subunit is used to generate corresponding first intermediate data based on the feature data; the gating subunit is used to generate multiple initial weight values ​​based on the feature data; each initial weight value corresponds to an expert subunit; the normalization subunit is used to normalize the initial weight values ​​to obtain the corresponding target weight values; and the calculation unit is used to generate second intermediate data corresponding to the recommendation target based on the first intermediate data and the target weight values.

[0085] The number of initial weight values ​​generated by each gating subunit is determined by the number of expert subunits corresponding to the same recommendation goal as that gating subunit. For example, if gating subunit 1, Expert A, and Expert B correspond to recommendation goal 1, then gating subunit 1 is used to generate initial weight value 1 corresponding to Expert A and initial weight value 2 corresponding to Expert B based on the feature data.

[0086] Normalizing the initial weight values ​​can be done by using the softmax function to convert the initial weight values ​​into a value within a preset range, which can be [0,1].

[0087] The calculation unit is used to generate second intermediate data corresponding to the recommendation target based on the first intermediate data and the target weight values. This can be achieved by assigning multiple target weight values ​​to the first intermediate data generated by the corresponding expert subunits, and then performing a weighted summation to obtain the second intermediate data. The second intermediate data corresponding to each recommendation target can be the input values ​​of the multi-scenario units corresponding to each recommendation target.

[0088] For example, Feature_all_scene is used to represent feature data for recommendations across multiple scenes.

[0089] The preset target set includes recommended target 1 and recommended target 2, and there are three expert sub-units: Expert A, Expert B, and Expert C. Recommended target 1 corresponds to gating sub-unit 1, normalization sub-unit 1, Expert A, Expert B, and calculation unit 1; recommended target 2 corresponds to gating sub-unit 2, normalization sub-unit 2, Expert B, Expert C, and calculation unit 2.

[0090] The input value of Expert A can be Feature_all_scene, and the output value of Expert A can be represented by Exp A, as shown in (1):

[0091] Exp A=Expert A(Feature_all_scene) (1)

[0092] Expert B and Expert C are similar to Expert A, as shown in (2) and (3):

[0093] Exp B=Expert B(Feature_all_scene) (2)

[0094] Exp C=Expert C(Feature_all_scene) (3)

[0095] The first intermediate data includes Exp A, Exp B, and Exp C.

[0096] The gated subunit 1 can be represented by Domain_gate1. The input value of Domain_gate1 can be Feature_all_scene, and the output value can be two initial weight values, which are uniformly represented by Dg1, as shown in (4):

[0097] Dg1=Domain_gate1(Feature_all_scene) (4)

[0098] The gating subunit 2 is similar to the gating subunit 1, as shown in (5):

[0099] Dg2=Domain_gate2(Feature_all_scene) (5)

[0100] Normalization subunit 1 is used to normalize the initial weight value Dg1 to obtain the corresponding target weight value softmax(Dg1), and normalization subunit 2 is used to normalize the initial weight value Dg2 to obtain the corresponding target weight value softmax(Dg2).

[0101] The calculation unit 1 is used to generate the second intermediate data corresponding to the recommendation target 1 based on the first intermediate data and the target weight value. The second intermediate data can be represented by Out1, and the initial weight value Dg1 includes D1 and D2, specifically as shown in (6):

[0102] Out1 = D1 * Exp A + D2 * Exp B (6)

[0103] The calculation unit 2 is used to generate the second intermediate data corresponding to the recommendation target 2 based on the first intermediate data and the target weight value. The second intermediate data can be represented by Out2, and the initial weight value Dg2 includes D3 and D4, as shown in (7):

[0104] Out2 = D3 * Exp B + D4 * Exp C (7)

[0105] In one specific implementation, the preset target set includes a first recommended target and a second recommended target; the multi-target unit includes a first expert subunit, a second expert subunit, and a third expert subunit; the first recommended target corresponds to the first expert subunit and the second expert subunit; the second recommended target corresponds to the second expert subunit and the third expert subunit.

[0106] In specific implementation, the preset target set includes a first recommended target and a second recommended target. The first recommended target corresponds to the first expert subunit and the second expert subunit, and the second recommended target corresponds to the second expert subunit and the third expert subunit. The first recommended target and the second recommended target share the second expert subunit.

[0107] In one specific implementation, the multi-target unit further includes a first embedded subunit and multiple second embedded subunits; the first embedded subunit is connected to each expert subunit respectively, and the output of the first embedded subunit is the input of the expert subunit; each second embedded subunit is connected to a gating subunit, and the output of the second embedded subunit is the input of the gating subunit.

[0108] The input to the first embedding subunit is the feature data of the recommendation items for multiple scenarios, and the input to each of the multiple second embedding subunits is the feature data of the recommendation items for multiple scenarios.

[0109] For example, the preset target set includes recommended target 1 and recommended target 2, and the number of multiple expert sub-units is 3, namely Expert A, Expert B, and Expert C. Among them, recommended target 1 corresponds to gating sub-unit 1, Expert A, and Expert B; recommended target 2 corresponds to gating sub-unit 2, Expert B, and Expert C.

[0110] The multi-target unit further includes a first embedded subunit, a second embedded subunit 1, and a second embedded subunit 2. The first embedded subunit is connected to Expert A, Expert B, and Expert C respectively, and its output serves as the input to Expert A, Expert B, and Expert C. The second embedded subunit 1 is connected to gated subunit 1, and its output serves as the input to gated subunit 1; the second embedded subunit 2 is connected to gated subunit 2, and its output serves as the input to gated subunit 2.

[0111] In one specific implementation, the first embedding subunit is used to transform the feature data to obtain the corresponding first feature vector; the expert subunit is used to generate the corresponding first intermediate data based on the first feature vector; the second embedding subunit is used to transform the feature data to obtain the corresponding second feature vector; and the gating subunit is used to generate multiple initial weight values ​​based on the second feature vector.

[0112] The first embedding subunit is used to perform embedding processing on the feature data to obtain the corresponding first feature vector. The expert subunit is used to generate the corresponding first intermediate data based on the first feature vector. This is similar to the aforementioned technical concept of "the expert subunit is used to generate the corresponding first intermediate data based on the feature data", and can be referred to the corresponding description section.

[0113] The second embedding subunit is used to perform embedding processing on the feature data to obtain the corresponding second feature vector. The gating subunit is used to generate multiple initial weight values ​​based on the second feature vector. This is similar to the aforementioned technical concept of "the gating subunit is used to generate multiple initial weight values ​​based on the feature data", and can be referred to the corresponding description section.

[0114] In one specific implementation, each recommendation target corresponds to a recommendation parameter; the fully connected sub-unit is used to determine the prediction result of the target recommendation parameter in the corresponding scenario of the fully connected sub-unit based on the second intermediate data; the target recommendation parameter is determined by the recommendation target corresponding to the multi-scenario unit including the fully connected sub-unit.

[0115] Each multi-scene unit corresponds to a recommendation target. Each multi-scene unit includes a fully connected sub-unit.

[0116] The target recommendation parameters can be the recommendation parameters that correspond to the same recommendation target as the fully connected subunit, that is, the recommendation parameters that correspond to the same recommendation target as the multi-scene units that include the fully connected subunit.

[0117] The fully connected subunit is used to determine the prediction results of the target recommendation parameters in the corresponding scenario based on private parameters, shared parameters, and second intermediate data.

[0118] The prediction result of the target recommendation parameter can be the ranking prediction result of the numerical values ​​of the target recommendation parameters of multiple similar recommendation items, the prediction result of recommendation items whose target recommendation parameter values ​​are greater than or equal to or less than a preset parameter threshold, or other types of prediction results determined by the target recommendation parameter.

[0119] Figure 2 This is a schematic diagram illustrating the data flow for training an initial recommendation model, provided in one or more embodiments of this specification.

[0120] like Figure 2 As shown, the multi-target unit includes a first embedded subunit 204 and a second embedded subunit 206 (i.e., Figure 2 The first embedded subunit 1) and the second embedded subunit 208 (i.e. Figure 2 The first embedded subunit 2), gate subunit 210 (i.e., gate subunit 1), gate subunit 212 (i.e., gate subunit 2), expert subunit A214, expert subunit B216, expert subunit C218, normalization subunit 220 and normalization subunit 222.

[0121] The mixed scene sample set 202 is the input data for the first embedded subunit 204, the second embedded subunit 206, and the second embedded subunit 208.

[0122] Recommended target 1 corresponds to the following multi-scenario units: private network subunit 224 (i.e. Figure 2 Private network subunit 1 corresponding to recommended target 1), ..., private network subunit 228 (i.e. Figure 2 The recommended target 1 corresponds to the private network subunit i) and the shared network subunit 226 (i.e. Figure 2 The network consists of the public network sub-unit corresponding to the recommended target 1, the scene-specific fully connected neural network 230 corresponding to scenario 1, ..., the scene-specific fully connected neural network 232 corresponding to scenario i. Here, i can be a natural number greater than 1. Figure 2 Only a few scenes are shown schematically.

[0123] The scenario-specific fully connected neural network 230 corresponding to scenario 1 outputs the prediction results of recommended target 1 and scenario 1 234, ..., and the scenario-specific fully connected neural network 232 corresponding to scenario i outputs the prediction results of recommended target 1 and scenario i 236.

[0124] Recommended target 2 corresponds to the following multi-scenario units: private network subunit 238 (i.e. Figure 2Private network subunit 1 corresponding to recommended target 2), ..., private network subunit 242 (i.e. Figure 2 The recommended target 2 corresponds to the private network subunit k), and the shared network subunit 240 (i.e. Figure 2 The network consists of the public network subunit corresponding to recommendation target 2, the scenario-specific fully connected neural network 244 corresponding to scenario 1, ..., the scenario-specific fully connected neural network 246 corresponding to scenario k. Here, k can be a natural number greater than 1. k can be the same as or different from i. The scenarios corresponding to each recommendation target can be the same or different. Figure 2 Only a few scenes are shown schematically.

[0125] The scenario-specific fully connected neural network 244 corresponding to scenario 1 outputs the recommended target 2 and the prediction result of scenario 1 248, ..., the scenario-specific fully connected neural network 246 corresponding to scenario k outputs the recommended target 2 and the prediction result of scenario k 250.

[0126] Based on the mixed-scene sample set, the training objective of the initial recommendation model is to enable the intermediate recommendation model to possess underlying general multi-objective task prediction capabilities, supporting multi-task objective information sharing and private domain information differentiation. The upper-layer model initially fits the general shared information of homogeneous tasks with the specific scene network.

[0127] Step S106: Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of recommendation items for one scene.

[0128] Each single-scene sample set can include multiple single-scene samples, and each single-scene sample can include feature data of one recommendation item for a specific scene. Within each single-scene sample set, the scenes corresponding to all single-scene sample sets are the same; that is, each single-scene sample set corresponds to a single scene.

[0129] Step S108: Train the intermediate recommendation model according to each single-scene sample set to obtain the recommendation model; the recommendation model is used to determine the recommendation items that meet the multi-objective recommendation conditions in the preset scene; the multi-objective recommendation conditions are determined by the preset scene and each recommendation objective.

[0130] The intermediate recommendation model is trained separately for each single-scene sample set to obtain the recommendation model. Alternatively, the intermediate recommendation model can be iteratively trained separately for each single-scene sample set, thereby continuously adjusting the model parameters corresponding to that scene during the training process.

[0131] It should be emphasized that the differences between this step and the aforementioned step S104 include:

[0132] (a1) Different training data

[0133] The training data used in step S104 is a mixed scene sample set, in which the feature data of different scenes are mixed together in a shuffled order. However, the training data used in this step is a single scene sample set, that is, each batch of training data corresponds to only one scene, and the feature data of different scenes will not be used in the same batch of model training.

[0134] (a2) Different gradient backpropagation methods

[0135] Step S104 uses a full gradient backpropagation method, meaning the training loss is used to drive the parameter updates of the entire model. This step uses a partial gradient backpropagation method, meaning the training loss is used to drive the parameter updates of some parameters in the corresponding scenario, while the other part of the parameters can be frozen. After freezing, the model training will not adjust the other part of the parameters. For example, the training loss is used to drive the private parameters of private network sub-units to update the parameters, while the shared parameters of public network sub-units and the model parameters corresponding to multi-objective units are all frozen.

[0136] The recommendation model is used to determine the recommended items that meet the multi-objective recommendation conditions in a preset scenario; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0137] In practice, the mapping relationship between scenarios and multi-objective recommendation conditions can be configured, and the weight values ​​of each recommendation objective can be configured in each multi-objective recommendation condition.

[0138] Based on the preset scenario and the mapping relationship between the scenario and the multi-objective recommendation conditions, the multi-objective recommendation conditions to be satisfied in the preset scenario can be determined. Based on the weight values ​​of each recommendation objective in the multi-objective recommendation conditions and the prediction results of each recommendation parameter in the preset scenario, it can be determined whether there are recommendation items in the preset scenario that satisfy the multi-objective recommendation conditions, and if so, the recommendation items that satisfy the multi-objective recommendation conditions can be identified.

[0139] In one specific implementation, the training loss of the intermediate recommendation model includes multiple fully connected losses; each fully connected loss corresponds to a fully connected sub-unit; the fully connected losses are used to update the private parameters of the private network sub-units corresponding to the fully connected sub-units.

[0140] Each fully connected subunit can be used to generate a corresponding fully connected loss based on private parameters, shared parameters, and second intermediate data corresponding to the same recommendation goal. This fully connected loss is used to update the private parameters of the private network subunit corresponding to the fully connected subunit.

[0141] It is important to note that during the training of the intermediate recommendation model based on the single-scene sample set for each scenario, the fully connected sub-units corresponding to other scenarios outside of that scenario do not participate in the model training. That is, only the multiple fully connected sub-units corresponding to that scenario can generate the fully connected loss, and each fully connected sub-unit corresponds to a recommendation target in the preset target set.

[0142] Because the network structure of private network sub-units is relatively simple, they can fit well even in small scenarios, avoiding poor fitting results caused by an excessively large number of samples in small scenarios.

[0143] Figure 3 This is a schematic diagram illustrating the data flow for training an intermediate recommendation model, provided in one or more embodiments of this specification.

[0144] like Figure 3 As shown, the multi-target unit includes a first embedded subunit 304 and a second embedded subunit 306 (i.e., Figure 3 The second embedded subunit 1) and the second embedded subunit 308 (i.e. Figure 3 The second embedded subunit 2), gate subunit 310 (i.e., gate subunit 1), gate subunit 312 (i.e., gate subunit 2), expert subunit A314, expert subunit B316, expert subunit C318, normalization subunit 320 and normalization subunit 322.

[0145] The single-scene sample set 302 is the input data for the first embedded subunit 304, the second embedded subunit 306, and the second embedded subunit 308.

[0146] Recommended target 1 corresponds to the following multi-scenario units: private network subunit 324 (i.e. Figure 3 Private network subunit 1 corresponding to recommended target 1), ..., private network subunit 328 (i.e. Figure 3 The recommended target 1 corresponds to the private network subunit i) and the shared network subunit 326 (i.e. Figure 3 The network consists of the public network sub-unit corresponding to the recommended target 1, the scene-specific fully connected neural network 330 corresponding to scenario 1, ..., the scene-specific fully connected neural network 332 corresponding to scenario i. Here, i can be a natural number greater than 1. Figure 3 Only a few scenes are shown schematically.

[0147] The scenario-specific fully connected neural network 330 corresponding to scenario 1 outputs the prediction results 334 for the recommended target 1 and scenario 1, ..., and the scenario-specific fully connected neural network 332 corresponding to scenario i outputs the prediction results 336 for the recommended target 1 and scenario i.

[0148] Recommended target 2 corresponds to the following multi-scenario units: private network subunit 338 (i.e. Figure 3 Private network subunit 1 corresponding to recommended target 2), ..., private network subunit 342 (i.e. Figure 3 The recommended target 2 corresponds to the private network subunit k), and the shared network subunit 340 (i.e. Figure 3 The network consists of the public network sub-unit corresponding to recommendation target 2, the scene-specific fully connected neural network 344 corresponding to scenario 1, ..., the scene-specific fully connected neural network 346 corresponding to scenario k. Here, k can be a natural number greater than 1. k can be the same as or different from i. The scenarios corresponding to each recommendation target can be the same or different. Figure 3 Only a few scenes are shown schematically.

[0149] The scenario-specific fully connected neural network 344 for scenario 1 outputs the recommended target 2 and the prediction result of scenario 1 348, ..., the scenario-specific fully connected neural network 346 for scenario k outputs the recommended target 2 and the prediction result of scenario k 350.

[0150] like Figure 3 As shown, the multi-objective unit and each multi-scene unit are blocked. It's important to note that this "block" means that the multi-objective unit can still send data unidirectionally to each multi-scene unit, but the training loss generated by each multi-scene unit cannot be transmitted back due to the block. Therefore, this training loss will not change the model parameters in the multi-objective unit. The model parameters in the multi-objective unit can be considered frozen parameters.

[0151] Furthermore, the communication between the public network subunit and the scene-specific fully connected neural network is blocked. It's important to note that this "blocking" means that the public network subunit can still send data unidirectionally to each scene-specific fully connected neural network, but the training loss generated by each scene-specific fully connected neural network cannot be transmitted back due to the block. Therefore, this training loss will not change the model parameters in the public network subunit. The model parameters in the public network subunit can be considered frozen parameters.

[0152] Figure 4 This is a schematic diagram of gradient backhaul provided for one or more embodiments of this specification.

[0153] Figure 4 It includes the "Initial" diagram, the "Phase 1" diagram, and the "Phase 2" diagram.

[0154] The “Initial” graph is used to represent the gradient backpropagation information of scenario 1, scenario 2... scenario n in the initial recommendation model.

[0155] The “Phase 1” diagram is used to represent the gradient backpropagation information of scenario 1, scenario 2... scenario n during the training process of the initial recommendation model to obtain the intermediate recommendation model.

[0156] The “Phase 2” diagram is used to represent the gradient backpropagation information of scenario 1, scenario 2... scenario n during the training process of the intermediate recommendation model to obtain the recommendation model.

[0157] In summary, the training method for the recommendation model provided in this embodiment obtains a mixed-scene sample set and uses this mixed-scene sample set as common model input data for different recommendation goals and different scenarios; based on the mixed-scene sample set, the initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scene unit corresponding to each recommendation goal in the preset target set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation goal, and a normalization sub-unit corresponding to each recommendation goal; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario. Through the above model structure, the underlying multi-objective structure receives full-scene batch samples for training, establishing a heterogeneous multi-objective model; multiple single-scene sample sets are obtained; each single-scene sample set includes a... The system gathers feature data for recommendation items in each scenario. Based on the sample set of each single scenario, intermediate recommendation models are trained separately to obtain the recommendation models. Training data is aggregated according to scenario, meaning each batch represents the same scenario. Training is then conducted intensively on the mapping relationships of single scenarios with fewer training parameters, leveraging the large information volume of the entire batch. This allows for the acquisition of refined objective function mapping relationships for each scenario. In this way, a large amount of scenario data can be introduced through a unified standard data stream. The common parts of the model provide a full-scenario description of users and items, truly achieving rich matching between users and different recommendation scenarios, better meeting user needs. The system can streamline the model architecture to achieve information gain sharing while maintaining the uniqueness of scenario and target information. Gradient conflicts can be resolved through partial gradient backpropagation. Training a unified multi-objective, multi-scenario model can serve a large number of online scenarios, significantly reducing computational resources and enhancing model real-time performance.

[0158] This specification provides an example of an information recommendation method:

[0159] Figure 5 This is a flowchart illustrating an information recommendation method provided in one or more embodiments of this specification.

[0160] like Figure 5 The information recommendation method shown includes steps S502-S504:

[0161] Step S502: Obtain the set of candidate recommendation items for the target scene; the set of candidate recommendation items includes feature data of multiple candidate recommendation items.

[0162] Step S504: Input the candidate recommendation item set into the recommendation model for prediction processing to obtain recommendation item information that meets the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each recommendation target; the recommendation model is trained as follows: obtain a mixed scenario sample set; the mixed scenario sample set includes feature data of recommendation items in multiple scenarios; train the initial recommendation model based on the mixed scenario sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation target in the preset target set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation target, and a normalization sub-unit corresponding to each recommendation target; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario; obtain multiple single-scenario sample sets; each single-scenario sample set includes feature data of recommendation items in one scenario; train the intermediate recommendation model based on each single-scenario sample set to obtain the recommendation model.

[0163] The information recommendation method provided in this embodiment is similar to the training method of the recommendation model provided in the above embodiments during the execution process. Please refer to the relevant content of the above embodiments when reading this embodiment.

[0164] The following is an example of a training device for a recommendation model provided in this specification:

[0165] In the above embodiments, a method for training a recommendation model is provided, and correspondingly, a device for training a recommendation model is also provided, which will be described below with reference to the accompanying drawings.

[0166] Reference Figure 6 The diagram shows a schematic of a training device for a recommendation model provided in this embodiment.

[0167] Since the apparatus embodiments correspond to the method embodiments, the descriptions are relatively simple. For relevant parts, please refer to the corresponding descriptions of the method embodiments provided above. The apparatus embodiments described below are merely illustrative.

[0168] This embodiment provides a training device for a recommendation model, including:

[0169] The first sample acquisition module 602 is configured to acquire a mixed scene sample set; the mixed scene sample set includes feature data of recommendation items for multiple scenes;

[0170] The first model training module 604 is configured to train an initial recommendation model based on the mixed scenario sample set to obtain an intermediate recommendation model. The initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in a preset objective set. The multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective. The multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario.

[0171] The second sample acquisition module 606 is configured to acquire multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for the scene.

[0172] The second model training module 608 is configured to train the intermediate recommendation model according to each single-scene sample set to obtain a recommendation model; the recommendation model is used to determine the recommendation items that meet the multi-objective recommendation conditions in the preset scene; the multi-objective recommendation conditions are determined by the preset scene and each recommendation objective.

[0173] The following is an embodiment of an information recommendation device provided in this specification:

[0174] In the above embodiments, an information recommendation method is provided, and correspondingly, an information recommendation device is also provided, which will be described below with reference to the accompanying drawings.

[0175] Reference Figure 7 The diagram illustrates an information recommendation device provided in this embodiment.

[0176] Since the apparatus embodiments correspond to the method embodiments, the descriptions are relatively simple. For relevant parts, please refer to the corresponding descriptions of the method embodiments provided above. The apparatus embodiments described below are merely illustrative.

[0177] This embodiment provides an information recommendation device, including:

[0178] The set acquisition module 702 is configured to acquire a set of candidate recommendation items for a target scene; the set of candidate recommendation items includes feature data of multiple candidate recommendation items;

[0179] The set prediction module 704 is configured to input the set of candidate recommendation items into the recommendation model for prediction processing to obtain recommendation item information that satisfies the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each of the recommendation objectives.

[0180] The recommendation model is trained in the following manner:

[0181] Obtain a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items from multiple scenarios;

[0182] Based on the mixed scenario sample set, the initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in the preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario;

[0183] Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for that scene;

[0184] Based on each single-scenario sample set, the intermediate recommendation model is trained to obtain the recommendation model.

[0185] The following is an example of a training device for a recommended model provided in this manual:

[0186] Corresponding to the training method for a recommendation model described above, and based on the same technical concept, one or more embodiments of this specification also provide a training device for a recommendation model, which is used to execute the training method for the recommendation model provided above. Figure 8 This is a schematic diagram of the structure of a training device for a recommended model provided in one or more embodiments of this specification.

[0187] This embodiment provides a training device for a recommendation model, comprising:

[0188] like Figure 8 As shown, the training device for the recommendation model can vary significantly due to differences in configuration or performance. It may include one or more processors 801 and memory 802, with memory 802 storing one or more application programs or data. Memory 802 can be temporary or persistent storage. The application programs stored in memory 802 may include one or more modules (not shown), each module including a series of computer-executable instructions from the training device for the recommendation model. Furthermore, processor 801 may be configured to communicate with memory 802, executing the series of computer-executable instructions in memory 802 on the training device for the recommendation model. The training device for the recommendation model may also include one or more power supplies 803, one or more wired or wireless network interfaces 804, one or more input / output interfaces 805, one or more keyboards 806, etc.

[0189] In one specific embodiment, the training device for the recommendation model includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for the training device of the recommendation model, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following:

[0190] Obtain a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items from multiple scenarios;

[0191] Based on the mixed scenario sample set, the initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in the preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario;

[0192] Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for that scene;

[0193] Based on each single-scenario sample set, the intermediate recommendation model is trained to obtain a recommendation model; the recommendation model is used to determine the recommendation items that meet the multi-objective recommendation conditions in the preset scenario; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0194] This specification provides an example of a storage medium as follows:

[0195] Corresponding to the training method of the recommendation model described above, based on the same technical concept, one or more embodiments of this specification also provide a storage medium.

[0196] The storage medium provided in this embodiment is used to store computer-executable instructions, which, when executed by a processor, implement the following process:

[0197] Obtain a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items from multiple scenarios;

[0198] Based on the mixed scenario sample set, the initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in the preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario;

[0199] Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for that scene;

[0200] Based on each single-scenario sample set, the intermediate recommendation model is trained to obtain a recommendation model; the recommendation model is used to determine the recommendation items that meet the multi-objective recommendation conditions in the preset scenario; the multi-objective recommendation conditions are determined by the preset scenario and each recommendation objective.

[0201] It should be noted that the embodiments concerning the storage medium in this specification and the embodiments concerning the training method of the recommendation model in this specification are based on the same inventive concept. Therefore, the specific implementation of this embodiment can be referred to the implementation of the corresponding method described above, and the repeated parts will not be described again.

[0202] Corresponding to the information recommendation method described above, based on the same technical concept, one or more embodiments of this specification also provide a training device for a recommendation model, which is used to execute the information recommendation method provided above. Figure 9 This is a schematic diagram of the structure of a training device for a recommended model provided in one or more embodiments of this specification.

[0203] This embodiment provides a training device for a recommendation model, comprising:

[0204] like Figure 9As shown, the training device for the recommendation model can vary significantly due to differences in configuration or performance. It may include one or more processors 901 and memory 902, with memory 902 storing one or more application programs or data. Memory 902 can be temporary or persistent storage. The application programs stored in memory 902 may include one or more modules (not shown), each module including a series of computer-executable instructions from the training device for the recommendation model. Furthermore, processor 901 may be configured to communicate with memory 902, executing the series of computer-executable instructions in memory 902 on the training device for the recommendation model. The training device for the recommendation model may also include one or more power supplies 903, one or more wired or wireless network interfaces 904, one or more input / output interfaces 905, one or more keyboards 906, etc.

[0205] In one specific embodiment, the training device for the recommendation model includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer-executable instructions for the training device of the recommendation model, and is configured to be executed by one or more processors. The one or more programs include computer-executable instructions for performing the following:

[0206] Obtain a set of candidate recommendation items for the target scenario; the set of candidate recommendation items includes feature data of multiple candidate recommendation items;

[0207] The candidate recommendation set is input into the recommendation model for prediction processing to obtain recommendation information that meets the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each recommendation objective.

[0208] The recommendation model is trained in the following manner:

[0209] Obtain a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items from multiple scenarios;

[0210] Based on the mixed scenario sample set, the initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in the preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario;

[0211] Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for that scene;

[0212] Based on each single-scenario sample set, the intermediate recommendation model is trained to obtain the recommendation model.

[0213] This specification provides an example of a storage medium as follows:

[0214] Corresponding to the information recommendation method described above, and based on the same technical concept, one or more embodiments of this specification also provide a storage medium.

[0215] The storage medium provided in this embodiment is used to store computer-executable instructions, which, when executed by a processor, implement the following process:

[0216] Obtain a set of candidate recommendation items for the target scenario; the set of candidate recommendation items includes feature data of multiple candidate recommendation items;

[0217] The candidate recommendation set is input into the recommendation model for prediction processing to obtain recommendation information that meets the multi-objective recommendation conditions; the multi-objective recommendation conditions are determined by the target scenario and each recommendation objective.

[0218] The recommendation model is trained in the following manner:

[0219] Obtain a mixed-scene sample set; the mixed-scene sample set includes feature data of recommendation items from multiple scenarios;

[0220] Based on the mixed scenario sample set, the initial recommendation model is trained to obtain an intermediate recommendation model; the initial recommendation model includes a multi-objective unit and a multi-scenario unit corresponding to each recommendation objective in the preset objective set; the multi-objective unit includes multiple expert sub-units, a gating sub-unit corresponding to each recommendation objective, and a normalization sub-unit corresponding to each recommendation objective; the multi-scenario unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scenario, and a private network sub-unit corresponding to each scenario;

[0221] Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of a recommendation item for that scene;

[0222] Based on each single-scenario sample set, the intermediate recommendation model is trained to obtain the recommendation model.

[0223] It should be noted that the embodiments concerning storage media in this specification and the embodiments concerning information recommendation methods in this specification are based on the same inventive concept. Therefore, the specific implementation of this embodiment can be referred to the implementation of the corresponding method described above, and the repeated parts will not be described again.

[0224] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0225] In the 1930s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many improvements to the methodology today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that an improvement to the methodology cannot be implemented using a hardware physical module. For example, a Programmable Logic Device (PLD) (e.g., a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program a digital system themselves to "integrate" it onto a PLD, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0226] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0227] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0228] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, when implementing the embodiments of this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0229] Those skilled in the art will understand that one or more embodiments of this specification can be provided as a method, system, or computer program product. Therefore, one or more embodiments of this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0230] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0231] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0232] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0233] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0234] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0235] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0236] It should also be noted that 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 limitation, 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.

[0237] One or more embodiments of this specification can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a particular task or implement a particular abstract data type. One or more embodiments of this specification can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0238] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0239] The above description is merely an embodiment of this document and is not intended to limit the scope of this document. Various modifications and variations can be made to this document by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this document should be included within the scope of the claims of this document.

Claims

1. A method for training a recommendation model, comprising: obtaining a mixed scene sample set; the mixed scene sample set comprising feature data of recommendation items of multiple scenes; training an initial recommendation model based on the mixed scene sample set to obtain an intermediate recommendation model; the initial recommendation model comprising a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit comprising a plurality of expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit comprising a shared network sub-unit, a fully connected sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit being connected in sequence, the output of the gating sub-unit being the input of the normalization sub-unit, and the output of the normalization sub-unit being the input of the calculation unit; the expert sub-unit being connected to the calculation unit, and the output of the expert sub-unit being the input of the calculation unit; the model parameters of the fully connected sub-unit in the initial recommendation model being determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; obtaining a plurality of single scene sample sets; each single scene sample set comprising feature data of recommendation items of one scene; training the intermediate recommendation model based on each single scene sample set to obtain a recommendation model; the recommendation model being used to determine recommendation items meeting multi-target recommendation conditions in a preset scene. the multi-target recommendation conditions being determined by the preset scene and each recommendation target.

2. The method of claim 1, the shared network sub-unit corresponding to the fully connected sub-unit corresponding to the same recommendation target; the private network sub-unit corresponding to the fully connected sub-unit corresponding to the same scene and the same recommendation target.

3. The method of claim 2, the training loss of the intermediate recommendation model comprising a plurality of fully connected losses; each fully connected loss corresponding to one fully connected sub-unit; the fully connected loss being used to update the private parameters of the private network sub-unit corresponding to the fully connected sub-unit.

4. The method of claim 2, the expert sub-unit being used to generate corresponding first intermediate data based on the feature data; the gating sub-unit being used to generate a plurality of initial weight values based on the feature data; each initial weight value corresponding to one expert sub-unit; the normalization sub-unit being used to normalize the initial weight values to obtain corresponding target weight values; the calculation unit being used to generate second intermediate data corresponding to the recommendation target based on the first intermediate data and the target weight values.

5. The method of claim 4, each recommendation target corresponding to one recommendation parameter. The full connection sub-unit is configured to determine a prediction result of a target recommendation parameter of the full connection sub-unit under a corresponding scene according to the second intermediate data; the target recommendation parameter is determined by a recommendation target corresponding to the multi-scene unit including the full connection sub-unit.

6. The method of claim 2, The multi-target unit further comprises a first embedding sub-unit and a plurality of second embedding sub-units; The first embedding sub-unit is connected with each of the expert sub-units, and an output of the first embedding sub-unit is an input of the expert sub-unit; Each of the second embedding sub-units is connected with one of the gating sub-units, and an output of the second embedding sub-unit is an input of the gating sub-unit.

7. The method of claim 6, The first embedding sub-unit is configured to perform conversion processing on the feature data to obtain a corresponding first feature vector; The expert sub-unit is configured to generate corresponding first intermediate data according to the first feature vector; The second embedding sub-unit is configured to perform conversion processing on the feature data to obtain a corresponding second feature vector; The gating sub-unit is configured to generate a plurality of initial weight values according to the second feature vector.

8. The method of claim 2, The preset target set comprises a first recommendation target and a second recommendation target; The multi-target unit comprises a first expert sub-unit, a second expert sub-unit, and a third expert sub-unit; The first recommendation target corresponds to the first expert sub-unit and the second expert sub-unit; The second recommendation target corresponds to the second expert sub-unit and the third expert sub-unit.

9. The method of claim 5, wherein the recommendation item is multimedia data; The recommendation parameter comprises at least one of a click-through rate, a conversion rate, and a play duration.

10. An information recommendation method, comprising: obtaining a candidate recommendation item set of a target scene; The candidate recommendation item set comprises feature data of a plurality of candidate recommendation items; inputting the candidate recommendation item set into a recommendation model for prediction processing to obtain recommendation item information satisfying a multi-target recommendation condition; the multi-target recommendation condition is determined by the target scene and each recommendation target; The recommendation model is obtained by training in the following manner: obtaining a mixed scene sample set; the mixed scene sample set comprises feature data of recommendation items of a plurality of scenes; According to the mixed scene sample set, model training is performed on an initial recommendation model to obtain an intermediate recommendation model; the initial recommendation model includes a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit includes a plurality of expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit are sequentially connected, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected to the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; model parameters of the fully connected sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; a plurality of single-scene sample sets are obtained; each single-scene sample set includes feature data of a recommendation item of one scene; model training is respectively performed on the intermediate recommendation model according to each single-scene sample set to obtain the recommendation model.

11. A recommendation model training device, comprising: a first sample obtaining module configured to obtain a mixed scene sample set; the mixed scene sample set includes feature data of recommendation items of a plurality of scenes; a first model training module configured to perform model training on an initial recommendation model according to the mixed scene sample set to obtain an intermediate recommendation model; the initial recommendation model includes a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit includes a plurality of expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit includes a shared network sub-unit, a fully connected sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit are sequentially connected, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected to the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; model parameters of the fully connected sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; a second sample obtaining module configured to obtain a plurality of single-scene sample sets; each single-scene sample set includes feature data of a recommendation item of one scene; a second model training module configured to perform model training on the intermediate recommendation model respectively according to each of the single-scene sample sets, to obtain a recommendation model; the recommendation model is used to determine a recommended item that meets a multi-target recommendation condition in a preset scene; the multi-target recommendation condition is determined by the preset scene and each of the recommended targets.

12. An information recommendation apparatus, comprising: a set obtaining module configured to obtain a candidate recommended item set of a target scene; the candidate recommended item set comprises feature data of a plurality of candidate recommended items; a set prediction module configured to input the candidate recommended item set into a recommendation model for prediction processing, to obtain recommended item information that meets a multi-target recommendation condition; the multi-target recommendation condition is determined by the target scene and each of the recommended targets; the recommendation model is obtained through training in the following manner: obtaining a mixed-scene sample set; the mixed-scene sample set comprises feature data of recommended items of a plurality of scenes; performing model training on an initial recommendation model according to the mixed-scene sample set, to obtain an intermediate recommendation model; the initial recommendation model comprises a multi-target unit and a multi-scene unit corresponding to each of the recommended targets in a preset target set; the multi-target unit comprises a plurality of expert sub-units corresponding to each of the recommended targets, a gating sub-unit corresponding to each of the recommended targets, a normalization sub-unit corresponding to each of the recommended targets, and a calculation unit corresponding to each of the recommended targets; the multi-scene unit comprises a shared network sub-unit, a fully connected sub-unit corresponding to each of the scenes, and a private network sub-unit corresponding to each of the scenes; the gating sub-unit, the normalization sub-unit, and the calculation unit are connected in sequence, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected to the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; model parameters of the fully connected sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; obtaining a plurality of single-scene sample sets; each of the single-scene sample sets comprises feature data of recommended items of one of the scenes; performing model training on the intermediate recommendation model respectively according to each of the single-scene sample sets, to obtain the recommendation model.

13. A training device of a recommendation model, comprising: a processor; and a memory configured to store computer-executable instructions that, when executed, cause the processor to: obtain a mixed-scene sample set; the mixed-scene sample set comprises feature data of recommended items of a plurality of scenes; According to the mixed scene sample set, model training is performed on an initial recommendation model to obtain an intermediate recommendation model; the initial recommendation model includes a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit includes a plurality of expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit includes a shared network sub-unit, a full connection sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit are connected in sequence, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected to the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; model parameters of the full connection sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; A plurality of single-scene sample sets are obtained; each single-scene sample set includes feature data of a recommendation item of one scene; According to each single-scene sample set, model training is performed on the intermediate recommendation model to obtain a recommendation model; the recommendation model is used to determine a recommendation item that meets a multi-target recommendation condition in a preset scene; The multi-target recommendation condition is determined by the preset scene and each recommendation target.

14. An information recommendation device, comprising: a processor; and a memory configured to store computer executable instructions that, when executed, cause the processor to: obtain a candidate recommendation item set of a target scene; the candidate recommendation item set includes feature data of a plurality of candidate recommendation items; input the candidate recommendation item set into a recommendation model for prediction processing to obtain recommendation item information that meets a multi-target recommendation condition; the multi-target recommendation condition is determined by the target scene and each recommendation target; the recommendation model is trained in the following way: obtain a mixed scene sample set; the mixed scene sample set includes feature data of recommendation items of a plurality of scenes; According to the mixed scene sample set, an initial recommendation model is subjected to model training to obtain an intermediate recommendation model; the initial recommendation model comprises a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit comprises a plurality of expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit comprises a shared network sub-unit, a full connection sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit are sequentially connected, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected with the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; the model parameters of the full connection sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; a plurality of single-scene sample sets are obtained; each single-scene sample set comprises feature data of a recommendation item of one scene; according to each single-scene sample set, the intermediate recommendation model is subjected to model training to obtain the recommendation model.

15. A storage medium for storing computer executable instructions, which, when executed by a processor, implement the following processes: a mixed scene sample set is obtained; the mixed scene sample set comprises feature data of recommendation items of a plurality of scenes; according to the mixed scene sample set, an initial recommendation model is subjected to model training to obtain an intermediate recommendation model; the initial recommendation model comprises a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit comprises a plurality of expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit comprises a shared network sub-unit, a full connection sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit are sequentially connected, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected with the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; the model parameters of the full connection sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; a plurality of single-scene sample sets are obtained; each single-scene sample set comprises feature data of a recommendation item of one scene; According to each single-scene sample set, model training is respectively performed on the intermediate recommendation model, and a recommendation model is obtained; the recommendation model is used to determine a recommendation item meeting a multi-target recommendation condition in a preset scene; The multi-target recommendation condition is determined by the preset scene and each recommendation target.

16. A storage medium for storing computer executable instructions, which, when executed by a processor, implement the following processes: Obtain a candidate recommendation item set of a target scene; the candidate recommendation item set includes feature data of multiple candidate recommendation items; Input the candidate recommendation item set into a recommendation model for prediction processing, and obtain recommendation item information meeting a multi-target recommendation condition; the multi-target recommendation condition is determined by the target scene and each recommendation target; The recommendation model is obtained through the following training: Obtain a mixed scene sample set; the mixed scene sample set includes feature data of recommendation items of multiple scenes; According to the mixed scene sample set, model training is performed on an initial recommendation model, and an intermediate recommendation model is obtained; the initial recommendation model includes a multi-target unit and a multi-scene unit corresponding to each recommendation target in a preset target set; the multi-target unit includes multiple expert sub-units corresponding to each recommendation target, a gating sub-unit corresponding to each recommendation target, a normalization sub-unit corresponding to each recommendation target, and a calculation unit corresponding to each recommendation target; the multi-scene unit includes a shared network sub-unit, a full connection sub-unit corresponding to each scene, and a private network sub-unit corresponding to each scene; the gating sub-unit, the normalization sub-unit, and the calculation unit are connected in sequence, the output of the gating sub-unit is the input of the normalization sub-unit, and the output of the normalization sub-unit is the input of the calculation unit; the expert sub-unit is connected with the calculation unit, and the output of the expert sub-unit is the input of the calculation unit; the model parameters of the full connection sub-unit in the initial recommendation model are determined based on shared parameters of the corresponding shared network sub-unit and private parameters of the corresponding private network sub-unit; Obtain multiple single-scene sample sets; each single-scene sample set includes feature data of recommendation items of one scene; According to each single-scene sample set, model training is respectively performed on the intermediate recommendation model, and the recommendation model is obtained.

Images