CN113901328A

CN113901328A - Information recommendation method and device, electronic equipment and storage medium

Info

Publication number: CN113901328A
Application number: CN202111382306.4A
Authority: CN
Inventors: 张学涛
Original assignee: Beijing Fangjianghu Technology Co Ltd
Current assignee: Beijing Fangjianghu Technology Co Ltd
Priority date: 2021-11-19
Filing date: 2021-11-19
Publication date: 2022-01-07

Abstract

The embodiment of the disclosure discloses an information recommendation method and device, an electronic device and a storage medium, wherein the method comprises the following steps: acquiring first user data corresponding to a first user, candidate object data corresponding to a candidate object and a multi-task scoring model obtained by pre-training, wherein the input of the multi-task scoring model comprises multi-task sharing input and independent input of each task; determining first characteristic data corresponding to shared input and second characteristic data of each task exclusive shared input based on the first user data and the candidate object data; obtaining a scoring result based on the first characteristic data, the second characteristic data and the multi-task scoring model; determining a target object to be recommended from all candidate objects based on the scoring result; and recommending the related information of the target object to the first user. The method and the device effectively improve the accuracy of the recommendation information.

Description

Information recommendation method and device, electronic equipment and storage medium

Technical Field

The present disclosure relates to information recommendation technologies, and in particular, to an information recommendation method and apparatus, an electronic device, and a storage medium.

Background

In the field of business services, such as the field of real estate, information recommendation services are generally required to be provided for users, in order to recommend object information, such as house source information, which better meets the requirements of the users to the users, a certain number of objects are generally required to be recalled from massive object information according to a certain rule as target objects recommended to the users, but in the prior art, the click probability of each object by the users is determined based on a single-task scoring model of the click probability, and then sequencing is performed based on the click probability, so that the target objects to be recommended to the users are determined.

Disclosure of Invention

The embodiment of the disclosure provides an information recommendation method and device, electronic equipment and a storage medium, and aims to solve the problems that information recommendation in the prior art is not accurate enough and the like.

In one aspect of the disclosed embodiments, an information recommendation method is provided, including: acquiring first user data corresponding to a first user, candidate object data corresponding to a candidate object and a multi-task scoring model obtained by pre-training, wherein the input of the multi-task scoring model comprises multi-task sharing input and independent input of each task; determining first characteristic data corresponding to the shared input and second characteristic data of each task exclusive shared input based on the first user data and the candidate object data; obtaining a scoring result based on the first feature data, each second feature data and the multi-task scoring model, wherein the scoring result comprises the score of each candidate object in each task; determining a target object to be recommended from all candidate objects based on the scoring result; and recommending the relevant information of the target object to the first user.

In an embodiment of the present disclosure, the multitask scoring model includes at least two tasks, a network structure of the multitask scoring model includes a shared network, a weight learning network corresponding to each task, and an output network of each task, where the shared network includes a plurality of parallel sub-networks; obtaining a scoring result based on the first feature data, each second feature data, and the multi-task scoring model, including: inputting the first characteristic data into each sub-network in the shared network to obtain a first output result corresponding to each sub-network;

for each second feature data, inputting the second feature data and the first feature data to a weight learning network of a corresponding task together to obtain a second output result of the weight learning network, wherein the second output result comprises weights of sub-networks in the shared network; for each second output result, inputting the second output result and each first output result into an output network of a corresponding task to obtain the output of the task; and taking the output of each task as the scoring result.

In an embodiment of the present disclosure, the determining a target object to be recommended from candidate objects based on the scoring result includes: and sequencing the candidate objects according to the scores of the candidate objects in at least one task, and taking the preset number of candidate objects with higher scores as the target objects.

In an embodiment of the present disclosure, before obtaining first user data corresponding to a first user, candidate object data corresponding to a candidate object, and a multi-task scoring model obtained by pre-training, the method further includes:

acquiring training characteristic data and corresponding label data, wherein the training characteristic data comprises first training characteristic data used for inputting a shared network and second training characteristic data corresponding to each task; inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into a weight learning network of a corresponding task to obtain a training output result of each task; determining the loss of each task based on the training output result of each task and the corresponding label data; determining a comprehensive loss based on the loss of each task; if the comprehensive loss meets a preset condition, finishing training to obtain the multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

In an embodiment of the present disclosure, the candidate object is a candidate house source, and the multi-task scoring model includes two tasks, namely a click task and a business opportunity task; the step of determining a target object to be recommended from the candidate objects based on the scoring result comprises the following steps: sorting the candidate house sources according to the scores of the candidate house sources in the click tasks or the business opportunity tasks, and taking the candidate house sources with higher scores in a first preset number as target house sources; or, sorting the candidate house sources according to the comprehensive scores of the candidate house sources in the click task and the business opportunity task, and taking the candidate house sources with higher scores in the first preset number as target house sources.

In an embodiment of the present disclosure, the determining, based on the first user data and the candidate object data, first feature data corresponding to the shared input and second feature data of a task-specific shared input includes: splitting the first user data and the candidate object data into first data and second data of each task according to a preset characteristic rule; and performing feature extraction on the first data to obtain first feature data, and performing feature extraction on each second data to obtain each second feature data.

In an embodiment of the present disclosure, acquiring candidate object data corresponding to a candidate object includes: acquiring original object data; recalling at least one group of object data from the original object data according to at least one recalling rule; and taking the recalled groups of object data as the candidate object data.

In another aspect of the embodiments of the present disclosure, a method for training a multi-task scoring model is provided, where inputs of a network structure of the multi-task scoring model include shared inputs of multiple tasks and individual inputs of each task; the method comprises the following steps: acquiring training user data, training object data and corresponding label data; determining first training characteristic data corresponding to shared input and second training characteristic data of task-independent shared input based on the training user data and the training object data; and training a pre-established multi-task scoring network based on the first training characteristic data and the second training characteristic data, and finishing training when the loss of the network meets a preset condition to obtain the multi-task scoring model.

In an embodiment of the present disclosure, the multitask scoring model includes at least two tasks, a network structure of the multitask scoring model includes a shared network, a weight learning network corresponding to each task, and an output network of each task, where the shared network includes a plurality of parallel sub-networks; the training of the pre-established multi-task scoring network based on the first training characteristic data and the second training characteristic data, and when the loss of the network meets a preset condition, ending the training to obtain the multi-task scoring model comprises: inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into a weight learning network of a corresponding task to obtain a training output result of each task; determining the loss of each task based on the training output result of each task and the corresponding label data; determining a comprehensive loss based on the loss of each task; if the comprehensive loss meets a preset condition, finishing training to obtain the multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and continuing training until the comprehensive loss meets the preset condition.

In a further aspect of the disclosed embodiments, there is provided an information recommendation apparatus including: the system comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring first user data corresponding to a first user, candidate object data corresponding to candidate objects and a multi-task scoring model obtained by pre-training, and the input of the multi-task scoring model comprises multi-task sharing input and independent sharing input of each task; a first determining module, configured to determine, based on the first user data and the candidate object data, first feature data corresponding to the shared input and second feature data of a task-specific input; the first processing module is used for obtaining a scoring result based on the first characteristic data, the second characteristic data and the multi-task scoring model, wherein the scoring result comprises scores of the candidate objects in the tasks; the second determination module is used for determining a target object to be recommended from all candidate objects based on the scoring result; and the sending module is used for recommending the relevant information of the target object to the first user.

In an embodiment of the present disclosure, the multitask scoring model includes at least two tasks, a network structure of the multitask scoring model includes a shared network, a weight learning network corresponding to each task, and an output network of each task, where the shared network includes a plurality of parallel sub-networks; the first processing module comprises: a first unit, configured to input the first feature data into each sub-network in the shared network, and obtain a first output result corresponding to each sub-network; a second unit, configured to, for each second feature data, input the second feature data and the first feature data together into a weight learning network of a corresponding task, and obtain a second output result of the weight learning network, where the second output result includes weights of sub-networks in the shared network; a third unit, configured to input, for each second output result, the second output result and each first output result to an output network of a corresponding task, so as to obtain an output of the task; a fourth unit configured to take an output of each task as the scoring result.

In an embodiment of the present disclosure, the second determining module is specifically configured to: and sequencing the candidate objects according to the scores of the candidate objects in at least one task, and taking the preset number of candidate objects with higher scores as the target objects.

In an embodiment of the present disclosure, the apparatus further includes: the second acquisition module is used for acquiring training characteristic data and corresponding label data, wherein the training characteristic data comprises first training characteristic data used for inputting the shared network and second training characteristic data corresponding to each task; the second processing module is used for inputting the first training characteristic data into each sub-network of the shared network, inputting each second training characteristic data into the weight learning network of the corresponding task, and obtaining the training output result of each task; the third processing module is used for determining the loss of each task based on the training output result of each task and the corresponding label data; the fourth processing module is used for determining comprehensive loss based on the loss of each task; a fifth processing module to: if the comprehensive loss meets a preset condition, finishing training to obtain the multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

In an embodiment of the present disclosure, the candidate object is a candidate house source, and the multi-task scoring model includes two tasks, namely a click task and a business opportunity task; the second determining module is specifically configured to: sorting the candidate house sources according to the scores of the candidate house sources in the click tasks or the business opportunity tasks, and taking the candidate house sources with higher scores in a first preset number as target house sources; or, sorting the candidate house sources according to the comprehensive scores of the candidate house sources in the click task and the business opportunity task, and taking the candidate house sources with higher scores in the first preset number as target house sources.

In an embodiment of the present disclosure, the first determining module is specifically configured to: splitting the first user data and the candidate object data into first data and second data of each task according to a preset characteristic rule; and performing feature extraction on the first data to obtain first feature data, and performing feature extraction on each second data to obtain each second feature data.

In an embodiment of the present disclosure, the first obtaining module is specifically configured to: acquiring original object data;

recalling at least one group of object data from the original object data according to at least one recalling rule; and taking the recalled groups of object data as the candidate object data.

In yet another aspect of the disclosed embodiments, there is provided a training apparatus for a multi-task scoring model, wherein the inputs of a network structure of the multi-task scoring model include shared inputs of multiple tasks and individual inputs of each task; the device comprises: the third acquisition module is used for acquiring training user data, training object data and corresponding label data; a third determining module, configured to determine, based on the training user data and the training object data, first training feature data corresponding to a shared input and second training feature data of a shared input for each task; and the first training module is used for training a pre-established multi-task scoring network based on the first training characteristic data and the second training characteristic data, and when the loss of the network meets a preset condition, finishing the training to obtain the multi-task scoring model.

In an embodiment of the present disclosure, the multitask scoring model includes at least two tasks, a network structure of the multitask scoring model includes a shared network, a weight learning network corresponding to each task, and an output network of each task, where the shared network includes a plurality of parallel sub-networks; the first training module is specifically configured to: inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into a weight learning network of a corresponding task to obtain a training output result of each task; determining the loss of each task based on the training output result of each task and the corresponding label data; determining a comprehensive loss based on the loss of each task; if the comprehensive loss meets a preset condition, finishing training to obtain the multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

According to a further aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium storing a computer program for executing the method according to any one of the above-mentioned embodiments of the present disclosure.

According to still another aspect of the embodiments of the present disclosure, there is provided an electronic apparatus including: a processor; a memory for storing the processor-executable instructions; the processor is configured to read the executable instructions from the memory and execute the instructions to implement the method according to any of the above embodiments of the present disclosure.

According to yet another aspect of the embodiments of the present disclosure, there is provided a computer program product comprising a computer program/instructions which, when executed by a processor, implement the method according to any of the above-mentioned embodiments of the present disclosure.

According to the information recommendation method and device, the electronic device and the storage medium, scores of all tasks of all candidate house sources are obtained for different users based on the multi-task scoring model, the target object to be recommended is determined for the users based on the multi-task scoring result, more accurate object information is recommended for the users, the accuracy of the recommended information is effectively improved, the input features of the multi-task scoring model are divided according to the feature dependence of different tasks, optimization conflicts among the tasks in the training process are avoided, the multi-task scoring model obtained through training is better in performance, and the accuracy of the recommended information is further improved.

The technical solution of the present disclosure is further described in detail by the accompanying drawings and examples.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the disclosure and together with the description, serve to explain the principles of the disclosure.

The present disclosure may be more clearly understood from the following detailed description, taken with reference to the accompanying drawings, in which:

fig. 1 is an exemplary application scenario of an information recommendation method provided by the present disclosure;

FIG. 2 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the disclosure;

FIG. 3 is a flowchart of step 203 provided by an exemplary embodiment of the present disclosure;

FIG. 4 is a network architecture diagram of a multi-tasking scoring model provided by an exemplary embodiment of the present disclosure;

FIG. 5 is a schematic diagram of an exemplary network architecture for a multi-tasking scoring model provided by an exemplary embodiment of the present disclosure;

FIG. 6 is a flowchart illustrating a method for training a multi-tasking scoring model according to an exemplary embodiment of the present disclosure;

fig. 7 is a schematic structural diagram of an information recommendation device according to an exemplary embodiment of the present disclosure;

FIG. 8 is a block diagram of a first processing module provided in an exemplary embodiment of the present disclosure;

fig. 9 is a schematic structural diagram of an information recommendation device according to another exemplary embodiment of the present disclosure;

FIG. 10 is a schematic structural diagram of a training apparatus for a multi-task scoring model according to an exemplary embodiment of the present disclosure;

fig. 11 is a schematic structural diagram of an embodiment of an application of the electronic device of the present disclosure.

Detailed Description

Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that: the relative arrangement of the components and steps, the numerical expressions, and numerical values set forth in these embodiments do not limit the scope of the present disclosure unless specifically stated otherwise.

It will be understood by those of skill in the art that the terms "first," "second," and the like in the embodiments of the present disclosure are used merely to distinguish one element from another, and are not intended to imply any particular technical meaning, nor is the necessary logical order between them.

It is also understood that in embodiments of the present disclosure, "a plurality" may refer to two or more and "at least one" may refer to one, two or more.

It is also to be understood that any reference to any component, data, or structure in the embodiments of the disclosure, may be generally understood as one or more, unless explicitly defined otherwise or stated otherwise.

In addition, the term "and/or" in the present disclosure is only one kind of association relationship describing an associated object, and means that three kinds of relationships may exist, for example, a and/or B may mean: a exists alone, A and B exist simultaneously, and B exists alone. In addition, the character "/" in the present disclosure generally indicates that the former and latter associated objects are in an "or" relationship.

It should also be understood that the description of the various embodiments of the present disclosure emphasizes the differences between the various embodiments, and the same or similar parts may be referred to each other, so that the descriptions thereof are omitted for brevity.

Meanwhile, it should be understood that the sizes of the respective portions shown in the drawings are not drawn in an actual proportional relationship for the convenience of description.

The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the disclosure, its application, or uses.

Techniques, methods, and apparatus known to those of ordinary skill in the relevant art may not be discussed in detail but are intended to be part of the specification where appropriate.

It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, further discussion thereof is not required in subsequent figures.

The disclosed embodiments may be applied to electronic devices such as terminal devices, computer systems, servers, etc., which are operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well known terminal devices, computing systems, environments, and/or configurations that may be suitable for use with electronic devices, such as terminal devices, computer systems, servers, and the like, include, but are not limited to: personal computer systems, server computer systems, thin clients, thick clients, hand-held or laptop devices, microprocessor-based systems, set-top boxes, programmable consumer electronics, networked personal computers, minicomputer systems, mainframe computer systems, distributed cloud computing environments that include any of the above, and the like.

Electronic devices such as terminal devices, computer systems, servers, etc. may be described in the general context of computer system-executable instructions, such as program modules, being executed by a computer system. Generally, program modules may include routines, programs, objects, components, logic, data structures, etc. that perform particular tasks or implement particular abstract data types. The computer system/server may be practiced in distributed cloud computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules may be located in both local and remote computer system storage media including memory storage devices.

Summary of the disclosure

In the process of implementing the present disclosure, the inventor finds that, in the business service field such as the real estate field, when recommending object information such as house source information for a user, a service platform exposes a plurality of house sources ordered according to a certain factor for the user, the user clicks the interested house source, and checks the detailed house source information, and then, the service platform further generates a business behavior, which is an exposure-click-business machine path.

Brief description of the drawings

Fig. 1 is an exemplary application scenario of the information recommendation method provided in the present disclosure.

In the field of real estate, in order to generate more business opportunities, related real estate information is actively recommended to a user for the user to select, by means of the technical scheme provided by the disclosure, the possibility that the user generates at least two behaviors such as clicking behaviors and business opportunity behaviors on each real estate can be comprehensively considered for different users, and more accurate real estate information is provided for the user. The recommendation of the house source information can be realized based on a service platform and a user application program APP, the service platform can be deployed on any implementable electronic device, such as a server or other computer systems, the server can be a single server, a server cluster or a cloud server, the user application program APP can be deployed on user terminal equipment (such as a smart phone, a tablet computer, a desktop computer and the like), and a user can download and install the house source information according to the requirements of the user. The service platform can extract historical behaviors of historical users on each house source according to a large amount of historical user data to form training data for training the multitask scoring model, the multitask scoring model is used for scoring each candidate house source on a plurality of targets for the user when the house source is recommended for the user, the candidate house sources are ranked based on the multitask scoring result, and the candidate house source ranked on TOP-N is selected as the target house source to be recommended and recommended to the user for the user to check and select. As the ranking result can be used for integrating at least two tasks such as click behaviors, business opportunities and the like, the recommendation accuracy can be effectively improved, the business probability is improved, the business income is improved, and the user experience can be effectively improved.

Exemplary method

Fig. 2 is a flowchart illustrating an information recommendation method according to an exemplary embodiment of the present disclosure. The method comprises the following steps:

step 201, acquiring first user data corresponding to a first user, candidate object data corresponding to a candidate object, and a multi-task scoring model obtained by pre-training, wherein the input of the multi-task scoring model includes multi-task sharing input and independent input of each task.

The first user may be any user to recommend information, for example, the first user may be one of all currently maintained users, and the first user data may include basic information (for example, browsing behaviors, active times, and the like) of the first user, preference information (for example, living room preference, address preference, and the like obtained by aggregating behavior information of the user on the APP), query conditions (query conditions input by historical search of the user), and the like. The candidate object may be all or part of the currently maintained massive objects, or may be a batch of objects recalled from the massive objects according to one or more factors, for example, a batch of objects recalled as a candidate object according to one or more of click rate, popularity, and business probability. The candidate object data includes basic attribute information of the candidate object, for example, the candidate object is a candidate house source, and the candidate house source data includes basic attribute information of each candidate house source, for example, a position, a price, a number of rooms, and the like. The multi-task scoring model is used for scoring candidate objects on a plurality of tasks aiming at a user when the object is recommended for the user, and the score of each candidate object on each task represents the probability of the first user executing the task on the candidate object.

The multi-task scoring model needs to be obtained through pre-training, for example, the service platform can extract historical behaviors of the historical users on various house sources according to a large amount of historical user data to form training data for training to obtain the multi-task scoring model. Aiming at the problem that certain dependency characteristics of one task may negatively influence the optimization of another task in the training process, the network architecture of the multi-task scoring model disclosed by the invention is provided with shared input and independent input of each task aiming at the multiple tasks, the shared input is used for inputting multi-task shared characteristic data, and the independent input of each task is used for inputting the specific characteristic of each task, so that the independent characteristic of each task only influences the task without influencing other tasks, the characteristic dependency conflict among the tasks is effectively solved, and the performance of the multi-task scoring model is improved.

Step 202, determining first feature data corresponding to the shared input and second feature data of the task-independent shared input based on the first user data and the candidate object data.

Because the network architecture of the multi-task scoring model comprises the shared input and the independent input of each task, after the first user data and the candidate object data are obtained, the first feature data corresponding to the shared input and the second feature data of the independent input of each task can be determined based on the first user data and the candidate object data according to the preset feature rule. The preset feature rules can be set according to actual requirements, and the division of the features needs to be consistent with the division of the features during training.

And step 203, obtaining a scoring result based on the first characteristic data, the second characteristic data and the multi-task scoring model, wherein the scoring result comprises the score of each candidate object in each task.

Wherein the score of each candidate object at each task represents the probability that the first user performed the task on the candidate object. After the first feature data and the second feature data of each task are obtained, the first feature data can be used as shared input of the multi-task scoring model, the second feature data can be used as independent input of each task and input into the multi-task scoring model, and scores of each task are obtained through the multi-task scoring model and serve as scoring results.

And step 204, determining a target object to be recommended from the candidate objects based on the scoring result.

After the scoring result is obtained, the candidate objects can be ranked based on the scoring result, and the target object to be recommended is determined from the candidate objects according to the ranking result.

In an optional example, the sorting may be performed according to a score of any task, or may be performed according to a comprehensive score of at least two tasks in the multiple tasks, which may be specifically set according to an actual requirement, and this embodiment is not limited.

For the house source recommendation service, the multi-task scoring model may include two tasks, namely a click task and a business task, and may be sorted according to scores of the business tasks, or according to scores of the click task, or according to a comprehensive score of the click task and the business task, where the comprehensive score may be a sum of the scores of the two tasks, or may be weighted average according to different weights, and may be specifically set according to actual requirements.

The obtained target object may include one or more objects, and the specific recommended number may be set according to actual requirements.

Step 205, recommending the relevant information of the target object to the first user.

After the target object to be recommended is obtained, the related information of the target object can be recommended to the first user.

In an optional example, in order to further improve the user experience, the recommended target object may be a list of objects sorted according to the scores, so that the terminal device may display according to the sorting result, so that the user may see the most relevant recommendation information first.

In an optional example, specific content of the related information of the recommended target object may be set according to actual requirements, for example, some important information of the target house source may be recommended first, a display list is formed, and when the user clicks each house source in the list, the detailed information page may be entered to view detailed information of the house source. The specific display mode is not limited.

According to the information recommendation method provided by the embodiment of the disclosure, scores of tasks of candidate house sources are obtained for different users based on the multi-task scoring model, the target object to be recommended is determined for the users based on the multi-task scoring result, more accurate object information is recommended for the users, the accuracy of recommended information is effectively improved, the input features of the multi-task scoring model are divided according to the feature dependence of different tasks, optimization conflicts among the tasks in the training process are avoided, and therefore the multi-task scoring model obtained through training is better in performance, and the accuracy of recommended information is further improved.

In an alternative example, fig. 3 is a flowchart of step 203 provided by an exemplary embodiment of the present disclosure. In this example, the multitask scoring model includes at least two tasks, a network structure of the multitask scoring model includes a shared network, a weight learning network corresponding to each task, and an output network of each task, and the shared network includes a plurality of parallel sub-networks; the corresponding step 203 may specifically include the following steps:

step 2031, inputting the first feature data into each sub-network in the shared network, and obtaining a first output result corresponding to each sub-network.

Wherein the sub-networks of the shared network are used for learning different combinations of the first characteristic data.

Step 2032, for each second feature data, inputting the second feature data and the first feature data together to the weight learning network of the corresponding task, and obtaining a second output result of the weight learning network, where the second output result includes the weight of each sub-network in the shared network.

The weight learning network of each task is used for obtaining the weight of the first output result of each sub-network to the task, and the input of the weight learning network is the exclusive characteristic (namely, the second characteristic data of the task) and the shared characteristic (namely, the first characteristic data) of the task.

Step 2033, for each second output result, inputting the second output result and each first output result to the output network of the corresponding task, and obtaining the output of the task.

Step 2034, the output of each task is taken as the scoring result.

And weighting the first output result of each sub-network to obtain the input of the output network of the corresponding task through the obtained different weights of each sub-network in different tasks.

Each task is provided with an output network, a second output result output by the weight learning network of the task and a first output result output by each sub-network of the sharing network are input to the output network corresponding to the task, the output network can weight the first output result of each sub-network based on the weight of the second output result and obtain the output of the task after other related processing, and the results of the first output results of each sub-network after different weighting represent different influences of sharing characteristics on different tasks, so that the obtained scoring result has higher accuracy on each task.

Illustratively, fig. 4 is a schematic network structure diagram of a multitask scoring model provided by an exemplary embodiment of the present disclosure. In this example, two tasks are taken as an example, where the number of the subnetworks included in the shared network is 4, the number of the subnetworks in practical application may be set according to actual requirements, and the number of network layers included in each subnetwork may also be set according to actual requirements, for example, the subnetwork may be a deep neural network including an input layer, a hidden layer, and an output layer, where the hidden layer may include one or more neural network layers, and a circle represents a neuron node (node for short), where the number of network layers and the number of neuron nodes in each layer are exemplary, and may be set according to actual requirements in practical application. Different sub-networks in the shared network are used for learning different combination situations of the first feature data, so that the specific structures of the different sub-networks are different, or the network parameters with the same structure are different, and the specific configuration can be set according to actual requirements. The weight learning network of each task is used for learning the weight of each sub-network in the task, the specific structure of the weight learning network can be a deep neural network comprising an input layer, a hidden layer and an output layer, wherein the hidden layer can comprise one or more neural network layers, the number of the hidden layers and the number of neuron nodes of each layer can be set according to actual requirements, and the weight represents the importance degree of the learning output result of the sub-network on the task, so that different feature combinations learned by each sub-network of the shared network have different influences on different tasks, in the training optimization process, the optimization conflict among the tasks with larger differences can be effectively reduced, and the effectiveness and the accuracy of the multi-task scoring model on each task are further improved; the output network of each task may be any practicable classification network, and specifically may include an input layer, a hidden layer, and an output layer, where the output result is a probability that a user performs the task on an object, or the output result may include a probability that the user performs the task and a probability that the user does not perform the task, and may be specifically set according to actual requirements. It should be noted that, based on the connection relationship between the partial networks, the input layer of one partial network may be the output layer of another partial network.

Illustratively, fig. 5 is a schematic diagram of an exemplary network structure of a multi-task scoring model provided by an exemplary embodiment of the present disclosure. Wherein, the numbers on the nodes are used for clearly expressing the number of the nodes on each layer, the shared network comprises 4 sub-networks, the input layer of each sub-network comprises 7 nodes, namely 7 nodes sharing the input, the first sub-network (from the left) comprises 2 layers in total, and comprises an output layer (comprising 4 nodes) besides the input layer, the second sub-network comprises 3 layers in total, and comprises an intermediate layer (comprising 16 nodes) and an output layer (comprising 4 nodes) besides the input layers of the 7 nodes; the third sub-network comprises a total of 4 layers, two intermediate layers (16 nodes each) and one output layer (4 nodes) in addition to the input layer; the 4 th sub-network comprises 5 layers in total, and also comprises 3 intermediate layers (the number of nodes is respectively 16, 32 and 16) and an output layer (4 nodes) besides the input layer; it can be seen that the first subnetwork is simpler in structure and is used for learning simple feature combinations; the weight learning network of the task 1 comprises 5 layers in total, an input layer (9 nodes) comprises 2 nodes of the task 1 which share the input independently and 7 nodes of the share input, the number of the 3 middle layer nodes is 16, and the number of the output layer nodes is 4; the weight learning network of the task 2 comprises 5 layers in total, an input layer (9 nodes) comprises 2 nodes of the task 2 for sharing input and 7 nodes of the sharing input, the number of the 3 intermediate layer nodes is 16, and the number of the output layer nodes is 4; the output network of the task 1 comprises 5 layers, the input layer (20 nodes) comprises 4 nodes of the output layer of the weight learning network of the task 1 and output layer nodes of each sub-network, the number of 3 middle layer nodes is respectively 16, 32 and 16, and the number of output layer nodes is 1; similarly, the output network of task 2 includes 5 layers, the input layer (20 nodes) includes 4 nodes of the output layer of the weight learning network of task 2 and output layer nodes of each sub-network, the number of 3 intermediate layer nodes is 16, 32 and 16 respectively, and the number of output layer nodes is 1. The exclusive input of the task 1 can be, for example, two characteristics of click times and click room source average price, the exclusive input of the task 2 can be, for example, two characteristics of business times and business room source average price, and the shared input can include, for example, a characteristic formed by a user related characteristic and a room source related characteristic.

In practical applications, the number of output layer nodes of the weight learning network may not be the same as the number of subnetworks, for example, the number of subnetworks is 4, and the number of output layer nodes of the weight learning network is 3, that is, it means that different tasks may select a part of subnetworks in a plurality of subnetworks for weighting. The structures of the weighted learning networks of different tasks may also be set to be different, and the structures of the output networks of different tasks may also be set to be different, and specifically may be set according to actual requirements, and are not limited to the structures in the above-mentioned figures.

In an alternative example, the step 204 of determining the target object to be recommended from the candidate objects based on the scoring result includes: and sequencing the candidate objects according to the scores of the candidate objects in at least one task, and taking the preset number of candidate objects with higher scores as target objects.

Specifically, after the scores of the candidate objects in the tasks are obtained, the ranking of the candidate objects may be performed according to the score of any one of the tasks, or may be performed according to the composite score of at least two of the tasks, where the composite score may be the sum of the scores of the tasks, or may be a weighted average according to different weights, and may be specifically set according to actual requirements. For example, the house source recommends services, and the scores of the candidate house sources on the click task and the business task are obtained, and during sorting, the sorting can be performed according to the score of the business task, the sorting can also be performed according to the score of the click task, and the sorting can also be performed according to the comprehensive score of the click task and the business task.

In an optional example, the multitask scoring model needs to be obtained through pre-training, that is, before the first user data corresponding to the first user, the candidate object data corresponding to the candidate object, and the multitask scoring model obtained through pre-training are obtained in step 201, the method of the present disclosure further includes: acquiring training characteristic data and corresponding label data, wherein the training characteristic data comprises first training characteristic data used for inputting a shared network and second training characteristic data corresponding to each task; inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into the weight learning network of the corresponding task to obtain a training output result of each task; determining the loss of each task based on the training output result of each task and the corresponding label data; determining a comprehensive loss based on the loss of each task; if the comprehensive loss meets the preset condition, finishing training to obtain a multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

The label data is whether each task is executed by the history user on each object, for example, whether the history user clicks a certain house source or not and whether business opportunity is generated or not. Each piece of training sample data includes user data of a historical user, object data of an object, and a behavior tag of the historical user on the object, where the behavior tag may be represented by 1 or 0, for example, 1 represents that the historical user has executed a task corresponding to the behavior on the object, for example, has executed a click task on the house source, that is, has generated a click behavior on the house source, and the training feature data may be obtained based on the training sample data and a preset feature rule, and the obtained training feature data includes first training feature data that is shared and input and second training feature data of each task. The first training feature data and each second training feature data are input to a corresponding network portion, the output result of each task can be obtained, which is called as the training output result, the loss of each task can be determined by adopting the preset loss function based on the training output result of each task and the corresponding label data, further, the comprehensive loss can be determined based on the loss of each task and a preset weighting rule, the preset weighting rule can be set according to the actual requirement, if the comprehensive loss meets the preset condition, the training tends to be convergent, the training can be finished to obtain a trained multi-task scoring model, if the comprehensive loss does not meet the preset condition, the parameters need to be adjusted to continue the training, the parameters may include network parameters, the number of layers of each network, the number of subnetworks, and the like, and may be specifically set according to actual requirements, which is not limited in this disclosure.

In an alternative example, the training process is described in detail by taking training optimization of a multi-task scoring model for house source matching as an example. Referring to the network structure of fig. 4, task 1 is a click task, task 2 is a business task, and the training of the model includes the following steps:

1. preparing a data set

For example, the service data of a recent period (e.g., 7 days, 30 days) may be applied, and the behavior of the user on each house source is extracted as a label, that is, the label includes two values, i.e., whether to click and whether to generate a business opportunity, which may be represented by different symbols. Based on the business data, the user and the house source data are associated through the user identification (such as the user ID) and the house source identification (such as the house source ID), and the business data comprise but are not limited to user basic information, user preference information, user behavior information, house source basic information and the like.

And performing association integration on the behaviors of each house source, the user data and the house source data by each user in a period of time to serve as a training data set, wherein each sample data is, for example, < user data, house source data, behavior/label >.

2. Building model networks

The model network can be constructed by using a TensorFlow or PyTorch framework, the network with a certain network layer structure is constructed according to business logic, and finally two target tasks are output and correspond to a click task and a business opportunity task.

3. Model training

After a model network is constructed, training optimization is performed based on the obtained training data set, optionally, the training optimization mode may adopt any implementable optimizer, such as an Adam optimizer, an SGD, and the like, and may be specifically set according to actual requirements.

4. Parameter adjustment

And adjusting model parameters according to the output result of the model training and the label data, wherein the parameters can comprise the number of the network layers, the number of the sub-networks and the like.

The method completes training optimization of a plurality of tasks through once modeling, effectively improves optimization efficiency, realizes respective learning of association and conflict among different tasks through combination of a sharing network and a weight learning network, ensures that the dependency characteristics of one task, which have negative influence on optimization of other tasks, are only used for dependency of the task and do not participate in the sharing network, realizes that the condition of selecting the more important characteristic combination of the task in the sharing network has larger weight and the less important characteristic combination of the task has smaller weight through the weight learning network, realizes dependency selection of different tasks on different characteristics, and further improves optimization efficiency and model performance.

In an optional example, the candidate object is a candidate house source, and the multi-task scoring model comprises two tasks of a click task and a business opportunity task; determining a target object to be recommended from the candidate objects based on the scoring result, wherein the method comprises the following steps: sorting the candidate house sources according to the scores of the candidate house sources in the click tasks or the business opportunity tasks, and taking the candidate house sources with higher scores in a first preset number as target house sources; or, sorting the candidate house sources according to the comprehensive scores of the candidate house sources in the click task and the business opportunity task, and taking the candidate house sources with higher scores in the first preset number as target house sources.

In an optional example, determining, based on the first user data and the candidate object data, first feature data corresponding to the shared input and second feature data of the task-specific shared input includes: splitting the first user data and the candidate object data into first data and second data of each task according to a preset characteristic rule; and performing feature extraction on the first data to obtain first feature data, and performing feature extraction on each second data to obtain each second feature data.

The preset feature rules can be set according to whether features depended on by different tasks in an actual scene have negative influence on optimization of other tasks. The feature extraction may be implemented in any way, and the disclosure is not limited.

In one optional example, the candidate objects may be mass objects currently maintained. That is, the method of the present disclosure is applied to a recall stage in a matching service, where the recall stage refers to that a user finds a certain number of objects from a large number of objects, for example, finds a certain number of house resources from a large number of house resources. The target house source obtained for the user is taken as a recall result.

In an optional example, obtaining candidate object data corresponding to the candidate object includes: acquiring original object data; recalling at least one group of object data from the original object data according to at least one recalling rule; and taking the recalled groups of object data as candidate object data. That is, the method of the present disclosure may also be used in a sorting stage of matching services, where a batch of objects recalled in each path are obtained through multiple recalls (e.g., according to click volume, popularity, and quotient probability), the objects recalled in each path are merged and then used as candidate objects, scoring is performed on the basis of the multi-task scoring model of the present disclosure, the candidate objects are sorted, and the object most likely to be acted by the user is arranged at a front position.

Any of the information recommendation methods provided by the embodiments of the present disclosure may be executed by any suitable device having data processing capability, including but not limited to: terminal equipment, a server and the like. Alternatively, any information recommendation method provided by the embodiments of the present disclosure may be executed by a processor, for example, the processor may execute any information recommendation method mentioned by the embodiments of the present disclosure by calling a corresponding instruction stored in a memory. And will not be described in detail below.

Another embodiment of the present disclosure may further provide a training method of a multi-task scoring model, which is used for training the multi-task scoring model in the above embodiments. The method can be executed by any implementable electronic device, such as a server, and the electronic device may be the same device as the electronic device of the information recommendation method or a different device. Fig. 6 is a flowchart illustrating a training method of a multi-task scoring model according to an exemplary embodiment of the disclosure. In the embodiment, the input of the network structure of the multi-task scoring model comprises shared input of multiple tasks and independent input of each task; the method comprises the following steps:

step 301, training user data, training object data and corresponding label data are obtained.

Step 302, based on the training user data and the training object data, determining first training feature data corresponding to the shared input and second training feature data of each task-independent shared input.

And 303, training the pre-established multi-task scoring network based on the first training characteristic data and the second training characteristic data, and finishing the training when the loss of the network meets a preset condition to obtain a multi-task scoring model.

In an optional example, the multitask scoring model comprises at least two tasks, the network structure of the multitask scoring model comprises a shared network, a weight learning network corresponding to each task and an output network of each task, and the shared network comprises a plurality of parallel sub-networks; training a pre-established multi-task scoring network based on first training characteristic data and second training characteristic data, and when the loss of the network meets a preset condition, finishing the training to obtain a multi-task scoring model, wherein the training comprises the following steps: inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into the weight learning network of the corresponding task to obtain a training output result of each task; determining the loss of each task based on the training output result of each task and the corresponding label data; determining a comprehensive loss based on the loss of each task; if the comprehensive loss meets the preset condition, finishing training to obtain a multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

It should be noted that the specific operations of the model training process have been described in detail in the foregoing embodiments, and are not described herein again.

Exemplary devices

Fig. 7 is a schematic structural diagram of an information recommendation device according to an exemplary embodiment of the present disclosure. The apparatus of this embodiment may be used to implement the corresponding method embodiment of the present disclosure, and the apparatus shown in fig. 7 includes: a first obtaining module 501, a first determining module 502, a first processing module 503, a second determining module 504 and a sending module 505.

The first obtaining module 501 is configured to obtain first user data corresponding to a first user, candidate object data corresponding to a candidate object, and a multi-task scoring model obtained through pre-training, where an input of the multi-task scoring model includes a multi-task sharing input and an exclusive input of each task.

The first determining module 502 is configured to determine, based on the first user data and the candidate object data acquired by the first acquiring module 501, first feature data corresponding to the shared input and second feature data of the task-specific shared input.

The first processing module 503 is configured to obtain a scoring result based on the first feature data, the second feature data, and the multi-task scoring model determined by the first determining module 502, where the scoring result includes a score of each candidate object in each task.

A second determining module 504, configured to determine, based on the scoring result obtained by the first processing module 503, a target object to be recommended from the candidate objects.

A sending module 505, configured to recommend the relevant information of the target object obtained by the second determining module 504 to the first user.

In an alternative example, fig. 8 is a schematic structural diagram of a first processing module according to an exemplary embodiment of the disclosure. In this example, the network structure of the multitask scoring model includes a shared network, a weight learning network corresponding to each task, and an output network of each task, where the shared network includes a plurality of parallel sub-networks; accordingly, the first processing module 503 includes: a first unit 5031, a second unit 5032, a third unit 5033 and a fourth unit 5034. A first unit 5031, configured to input the first feature data determined by the first determining module 502 into each sub-network in the shared network, and obtain a first output result corresponding to each sub-network; a second unit 5032, configured to, for each second feature data determined by the first determining module 502, input the second feature data and the first feature data together into a weight learning network of the corresponding task, and obtain a second output result of the weight learning network, where the second output result includes weights of sub-networks in the sharing network; a third unit 5033, configured to input, for each second output result obtained by the second unit 5032, the second output result and each first output result into an output network of the corresponding task, so as to obtain an output of the task; a fourth unit 5034 for taking the output of each task obtained by the third unit 5033 as a scoring result.

In an optional example, the second determining module 504 is specifically configured to: and sequencing the candidate objects according to the scores of the candidate objects in at least one task, and taking the preset number of candidate objects with higher scores as target objects.

Fig. 9 is a schematic structural diagram of an information recommendation device according to another exemplary embodiment of the present disclosure.

In one optional example, the apparatus of the present disclosure further comprises: a second obtaining module 506, a second processing module 507, a third processing module 508, a fourth processing module 509, and a fifth processing module 510. A second obtaining module 506, configured to obtain training feature data and corresponding label data, where the training feature data includes first training feature data used for inputting to the shared network and second training feature data corresponding to each task; the second processing module 507 is configured to input the first training feature data acquired by the second acquisition module to each sub-network of the shared network, and input each second training feature data acquired by the second acquisition module to the weight learning network of the corresponding task, so as to obtain a training output result of each task; a third processing module 508, configured to determine a loss of each task based on the training output result of each task and the corresponding label data obtained by the second processing module; a fourth processing module 509, configured to determine a comprehensive loss based on the loss of each task determined by the third processing module; a fifth processing module 510 for: if the comprehensive loss obtained by the fourth processing module meets the preset condition, finishing the training and obtaining a multi-task scoring model; and if the comprehensive loss obtained by the fourth processing module does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

In an optional example, the candidate object is a candidate house source, and the multi-task scoring model comprises two tasks of a click task and a business opportunity task; the second determining module 504 is specifically configured to: sorting the candidate house sources according to the scores of the candidate house sources in the click tasks or the business opportunity tasks, and taking the candidate house sources with higher scores in a first preset number as target house sources; or, sorting the candidate house sources according to the comprehensive scores of the candidate house sources in the click task and the business opportunity task, and taking the candidate house sources with higher scores in the first preset number as target house sources.

In an optional example, the first determining module 502 is specifically configured to: splitting the first user data and the candidate object data into first data and second data of each task according to a preset characteristic rule; and performing feature extraction on the first data to obtain first feature data, and performing feature extraction on each second data to obtain each second feature data.

In an optional example, the first obtaining module 501 is specifically configured to: acquiring original object data; recalling at least one group of object data from the original object data according to at least one recalling rule; and taking the recalled groups of object data as candidate object data.

Yet another embodiment of the present disclosure further provides a training apparatus for a multi-task scoring model, where the apparatus of this embodiment may be used to implement an embodiment of a training method for a multi-task scoring model corresponding to the present disclosure, and fig. 10 is a schematic structural diagram of a training apparatus for a multi-task scoring model provided in an exemplary embodiment of the present disclosure, where an input of a network structure of the multi-task scoring model in this example includes a shared input of multiple tasks and an exclusive input of each task; the device includes: a third acquisition module 401, a third determination module 402 and a first training module 403.

The third obtaining module 401 is configured to obtain training user data, training object data, and corresponding label data. A third determining module 402, configured to determine, based on the training user data and the training object data, first training feature data corresponding to the shared input and second training feature data of the task-specific input. The first training module 403 is configured to train a pre-established multi-task scoring network based on the first training feature data and the second training feature data, and when the loss of the network meets a preset condition, end the training to obtain a multi-task scoring model.

In an optional example, the multitask scoring model comprises at least two tasks, the network structure of the multitask scoring model comprises a shared network, a weight learning network corresponding to each task and an output network of each task, and the shared network comprises a plurality of parallel sub-networks; correspondingly, the first training module 403 is specifically configured to: inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into the weight learning network of the corresponding task to obtain a training output result of each task; determining the loss of each task based on the training output result of each task and the corresponding label data; determining a comprehensive loss based on the loss of each task; if the comprehensive loss meets the preset condition, finishing training to obtain a multi-task scoring model; and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

In addition, an embodiment of the present disclosure also provides an electronic device, including:

a memory for storing a computer program;

a processor, configured to execute the computer program stored in the memory, and when the computer program is executed, implement the information recommendation method according to any of the above embodiments of the present disclosure.

Fig. 11 is a schematic structural diagram of an embodiment of an application of the electronic device of the present disclosure. As shown in fig. 11, the electronic device includes one or more processors and memory.

The processor may be a Central Processing Unit (CPU) or other form of processing unit having data processing capabilities and/or instruction execution capabilities, and may control other components in the electronic device to perform desired functions.

The memory may include one or more computer program products that may include various forms of computer-readable storage media, such as volatile memory and/or non-volatile memory. The volatile memory may include, for example, Random Access Memory (RAM), cache memory (cache), and/or the like. The non-volatile memory may include, for example, Read Only Memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium and executed by a processor to implement the information recommendation methods of the various embodiments of the present disclosure described above and/or other desired functions.

In one example, the electronic device may further include: an input device and an output device, which are interconnected by a bus system and/or other form of connection mechanism (not shown).

The input device may also include, for example, a keyboard, a mouse, and the like.

The output device may output various information including the determined distance information, direction information, and the like to the outside. The output devices may include, for example, a display, speakers, a printer, and a communication network and remote output devices connected thereto, among others.

Of course, for simplicity, only some of the components of the electronic device relevant to the present disclosure are shown in fig. 11, omitting components such as buses, input/output interfaces, and the like. In addition, the electronic device may include any other suitable components, depending on the particular application.

In addition to the above methods and apparatus, embodiments of the present disclosure may also be a computer program product comprising computer program instructions that, when executed by a processor, cause the processor to perform the steps in the information recommendation methods according to the various embodiments of the present disclosure described in the above sections of this specification.

The computer program product may write program code for carrying out operations for embodiments of the present disclosure in any combination of one or more programming languages, including an object oriented programming language such as Java, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's computing device and partly on a remote computing device, or entirely on the remote computing device or server.

Furthermore, embodiments of the present disclosure may also be a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, cause the processor to perform steps in an information recommendation method according to various embodiments of the present disclosure described in the above section of the present specification.

The computer-readable storage medium may take any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may include, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

Those of ordinary skill in the art will understand that: all or part of the steps for implementing the method embodiments may be implemented by hardware related to program instructions, and the program may be stored in a computer readable storage medium, and when executed, the program performs the steps including the method embodiments; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.

The foregoing describes the general principles of the present disclosure in conjunction with specific embodiments, however, it is noted that the advantages, effects, etc. mentioned in the present disclosure are merely examples and are not limiting, and they should not be considered essential to the various embodiments of the present disclosure. Furthermore, the foregoing disclosure of specific details is for the purpose of illustration and description and is not intended to be limiting, since the disclosure is not intended to be limited to the specific details so described.

In the present specification, the embodiments are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same or similar parts in the embodiments are referred to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

The block diagrams of devices, apparatuses, systems referred to in this disclosure are only given as illustrative examples and are not intended to require or imply that the connections, arrangements, configurations, etc. must be made in the manner shown in the block diagrams. These devices, apparatuses, devices, systems may be connected, arranged, configured in any manner, as will be appreciated by those skilled in the art. Words such as "including," "comprising," "having," and the like are open-ended words that mean "including, but not limited to," and are used interchangeably therewith. The words "or" and "as used herein mean, and are used interchangeably with, the word" and/or, "unless the context clearly dictates otherwise. The word "such as" is used herein to mean, and is used interchangeably with, the phrase "such as but not limited to".

The methods and apparatus of the present disclosure may be implemented in a number of ways. For example, the methods and apparatus of the present disclosure may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order for the steps of the method is for illustration only, and the steps of the method of the present disclosure are not limited to the order specifically described above unless specifically stated otherwise. Further, in some embodiments, the present disclosure may also be embodied as programs recorded in a recording medium, the programs including machine-readable instructions for implementing the methods according to the present disclosure. Thus, the present disclosure also covers a recording medium storing a program for executing the method according to the present disclosure.

It is also noted that in the devices, apparatuses, and methods of the present disclosure, each component or step can be decomposed and/or recombined. These decompositions and/or recombinations are to be considered equivalents of the present disclosure.

The previous description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other aspects without departing from the scope of the disclosure. Thus, the present disclosure is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

The foregoing description has been presented for purposes of illustration and description. Furthermore, this description is not intended to limit embodiments of the disclosure to the form disclosed herein. While a number of example aspects and embodiments have been discussed above, those of skill in the art will recognize certain variations, modifications, alterations, additions and sub-combinations thereof.

Claims

1. An information recommendation method, comprising:

acquiring first user data corresponding to a first user, candidate object data corresponding to a candidate object and a multi-task scoring model obtained by pre-training, wherein the input of the multi-task scoring model comprises multi-task sharing input and independent input of each task;

determining first characteristic data corresponding to the shared input and second characteristic data of each task exclusive shared input based on the first user data and the candidate object data;

obtaining a scoring result based on the first feature data, each second feature data and the multi-task scoring model, wherein the scoring result comprises the score of each candidate object in each task;

determining a target object to be recommended from all candidate objects based on the scoring result;

and recommending the relevant information of the target object to the first user.

2. The method of claim 1, wherein the multitask scoring model comprises at least two tasks, the network structure of the multitask scoring model comprises a shared network, a weight learning network corresponding to each task and an output network of each task, and the shared network comprises a plurality of parallel sub-networks;

obtaining a scoring result based on the first feature data, each second feature data, and the multi-task scoring model, including:

inputting the first characteristic data into each sub-network in the shared network to obtain a first output result corresponding to each sub-network;

for each second feature data, inputting the second feature data and the first feature data to a weight learning network of a corresponding task together to obtain a second output result of the weight learning network, wherein the second output result comprises weights of sub-networks in the shared network;

for each second output result, inputting the second output result and each first output result into an output network of a corresponding task to obtain the output of the task;

and taking the output of each task as the scoring result.

3. The method according to claim 2, wherein the determining a target object to be recommended from the candidate objects based on the scoring result comprises:

and sequencing the candidate objects according to the scores of the candidate objects in at least one task, and taking the preset number of candidate objects with higher scores as the target objects.

4. The method of claim 2, wherein before obtaining the first user data corresponding to the first user, the candidate object data corresponding to the candidate object, and the multi-task scoring model obtained by pre-training, the method further comprises:

acquiring training characteristic data and corresponding label data, wherein the training characteristic data comprises first training characteristic data used for inputting a shared network and second training characteristic data corresponding to each task;

inputting the first training characteristic data into each sub-network of the shared network, and inputting each second training characteristic data into a weight learning network of a corresponding task to obtain a training output result of each task;

determining the loss of each task based on the training output result of each task and the corresponding label data;

determining a comprehensive loss based on the loss of each task;

if the comprehensive loss meets a preset condition, finishing training to obtain the multi-task scoring model;

and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and entering the next round of training until the comprehensive loss meets the preset condition.

5. The method of claim 2, wherein the candidate object is a candidate house source, and the multi-task scoring model comprises two tasks, a click task and a business task;

the step of determining a target object to be recommended from the candidate objects based on the scoring result comprises the following steps:

sorting the candidate house sources according to the scores of the candidate house sources in the click tasks or the business opportunity tasks, and taking the candidate house sources with higher scores in a first preset number as target house sources; alternatively, the first and second electrodes may be,

and sequencing the candidate house sources according to the comprehensive scores of the candidate house sources in the click task and the business opportunity task, and taking the candidate house sources with higher scores in a first preset number as target house sources.

6. The method of claim 1, wherein determining first feature data corresponding to the shared input and second feature data of task-specific shared inputs based on the first user data and the candidate object data comprises:

splitting the first user data and the candidate object data into first data and second data of each task according to a preset characteristic rule;

and performing feature extraction on the first data to obtain first feature data, and performing feature extraction on each second data to obtain each second feature data.

7. The method according to any one of claims 1-6, wherein obtaining candidate object data corresponding to the candidate objects comprises:

acquiring original object data;

recalling at least one group of object data from the original object data according to at least one recalling rule;

and taking the recalled groups of object data as the candidate object data.

8. The training method of the multi-task scoring model is characterized in that the input of the network structure of the multi-task scoring model comprises shared input of multiple tasks and independent input of each task;

the method comprises the following steps:

acquiring training user data, training object data and corresponding label data;

determining first training characteristic data corresponding to shared input and second training characteristic data of task-independent shared input based on the training user data and the training object data;

and training a pre-established multi-task scoring network based on the first training characteristic data and the second training characteristic data, and finishing training when the loss of the network meets a preset condition to obtain the multi-task scoring model.

9. The method of claim 8, wherein the multitask scoring model comprises at least two tasks, the network structure of the multitask scoring model comprises a shared network, a weight learning network corresponding to each task and an output network of each task, and the shared network comprises a plurality of parallel sub-networks;

the training of the pre-established multi-task scoring network based on the first training characteristic data and the second training characteristic data, and when the loss of the network meets a preset condition, ending the training to obtain the multi-task scoring model comprises:

determining a comprehensive loss based on the loss of each task;

and if the comprehensive loss does not meet the preset condition, adjusting parameters according to preset rules, and continuing training until the comprehensive loss meets the preset condition.

10. A computer program product comprising computer programs/instructions, characterized in that the computer programs/instructions, when executed by a processor, implement the method of any of the preceding claims 1-9.