A recommendation model training method and device, and an electronic device

Abstract

The application provides a recommendation model training method and device and electronic equipment. The recommendation model training method comprises: selecting sample data conforming to target field characteristics as source field selected sample data in a source field; using the source field selected sample data to form training data, training a recommendation model, and obtaining a first recommendation model for the target field. The recommendation model training method provided by the application migrates a large amount of complete sample data in the source field to the target field according to the characteristics of the target field, trains a recommendation model required by the target field according to the sample data in the source field, realizes sample data cross-scene migration of the source field to obtain a target field recommendation model, and effectively avoids the problem that the target field data is insufficient and a recommendation model cannot be constructed based on sample data in the target field.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine learning, in particular to a recommendation model training method and device and electronic equipment, and further relates to an information recommendation method. BACKGROUND

[0002] With the continuous progress of computer technology, deep learning has achieved great success and is widely used in various application scenarios, such as related information recommendation, data mining, etc.

[0003] Deep learning relies on unsupervised or semi-supervised feature learning algorithms and hierarchical feature extraction to achieve automatic extraction of data features, but in specific application scenarios, relying on deep learning technology to accurately extract data features requires a large amount of sample data to support, and the cost of obtaining sample data and the completeness of sample data itself have become decisive factors restricting the application effect of deep learning technology.

[0004] Although the development of transfer learning technology has made up for the problem of insufficient training data to some extent, the amount of sample data in the target scene also limits the application of transfer learning technology, and existing transfer learning technology cannot well adapt to tasks in these scenes, and related recommendation is a typical representative scene. Therefore, how to apply transfer learning technology to related recommendation scenarios has become a problem to be solved. SUMMARY

[0005] The present application provides a recommendation model training method to solve the problem of insufficient sample data in the target application scenario, which makes it difficult to apply transfer learning technology to achieve related recommendation in the target application scenario. The present application also relates to a recommendation model training device and electronic equipment. In addition, the present application also provides two different information recommendation methods.

[0006] The recommendation model training method provided by the present application comprises:

[0007] Selecting sample data consistent with the target field features in the source field as source field selected sample data;

[0008] Using the source field selected sample data to form training data, training a recommendation model to obtain a first recommendation model for the target field.

[0009] Optionally, the method further comprises:

[0010] Obtaining target field sample data in the target field data;

[0011] Using the target field sample data to train the first recommendation model to obtain a second recommendation model.

[0012] Optionally, the source domain selected sample data is obtained by filtering sample data from at least one of the following aspects: the use scenario of the sample data is consistent with the target domain, the classification of the object involved in the sample data is consistent with the target domain, and the classification of the subject involved in the sample data is consistent with the target domain.

[0013] Optionally, the source domain selected sample data is obtained by filtering sample data from at least one of the following aspects: the use scenario of the sample data is consistent with the target domain, the classification of the object involved in the sample data is consistent with the target domain, and the classification of the subject involved in the sample data is consistent with the target domain.

[0014] Optionally, the source domain selected sample data and the target domain sample data adopt data elements of the following data types: text data, structured data, or a combination of text data and structured data.

[0015] Optionally, the source domain selected sample data and the target domain sample data adopt data elements of the following data types: text data, structured data, or a combination of text data and structured data.

[0016] Optionally, the source domain selected sample data and the target domain sample data adopt data elements of the following data types: text data, structured data, or a combination of text data and structured data.

[0017] Optionally, the source domain selected sample data and the target domain sample data adopt data elements of the following data types: text data, structured data, or a combination of text data and structured data.

[0018] Optionally, the source domain selected sample data and the target domain sample data include the following data elements: trigger data, target data, and positive and negative sample identification.

[0019] The application also provides a related recommendation model obtaining device, which comprises:

[0020] A first sample data obtaining unit is configured to select sample data consistent with the target domain features in the source domain as source domain selected sample data.

[0021] A first recommendation model generating unit is configured to use the source domain selected sample data to form training data, train the recommendation model, and obtain a first recommendation model for the target domain.

[0022] The application also provides an electronic device, which comprises:

[0023] A processor;

[0024] A memory is configured to store a program of a method, and the program is configured to perform the following steps when read and executed by the processor: selecting sample data consistent with the target domain features in the source domain as source domain selected sample data; and using the source domain selected sample data to form training data, train the recommendation model, and obtain a first recommendation model for the target domain.

[0025] The application also provides an information recommendation method, comprising:

[0026] obtaining trigger data;

[0027] providing the trigger data to a first recommendation model of a target field, wherein the first recommendation model is used to provide target data as recommendation data according to the trigger data;

[0028] The recommendation model is obtained in the following manner:

[0029] selecting sample data conforming to the characteristics of the target field as selected sample data of a source field;

[0030] forming training data by using the selected sample data of the source field, training the recommendation model, and obtaining the first recommendation model for the target field.

[0031] The application also provides an information recommendation method, comprising:

[0032] obtaining trigger data;

[0033] providing the trigger data to a second recommendation model of a target field, wherein the second recommendation model is used to provide target data as recommendation data according to the trigger data;

[0034] The second recommendation model is obtained in the following manner:

[0035] selecting sample data conforming to the characteristics of the target field as selected sample data of a source field;

[0036] forming training data by using the selected sample data of the source field, training the recommendation model, and obtaining the first recommendation model for the target field;

[0037] collecting target field sample data from data of the target field;

[0038] training the first recommendation model by using the target field sample data, and obtaining the second recommendation model.

[0039] Compared with the prior art, the application has the following advantages:

[0040] The training method of the recommendation model provided in the application comprises the following steps: selecting sample data conforming to the characteristics of a target field as source field selected sample data in a source field; using the source field selected sample data to form training data, training a recommendation model, and obtaining a first recommendation model for the target field. The training method of the recommendation model provided in the application migrates a large amount of complete sample data in the source field to the target field according to the characteristics of the target field, trains a recommendation model required by the target field according to the sample data in the source field, realizes the cross-scene migration of the sample data in the source field to obtain the recommendation model of the target field, and effectively avoids the problem that the data quantity of the target field is insufficient, the data completeness is not high, and the sample data in the target field cannot be used to construct a recommendation model suitable for the target field. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The flow chart of the training method of the recommendation model provided in the first embodiment of the application;

[0042] Figure 2 The structural schematic diagram of the training device of the recommendation model provided in the second embodiment of the application;

[0043] Figure 3 The structural schematic diagram of the electronic device provided in the third embodiment of the application;

[0044] Figure 4 The flow chart of the information recommendation method provided in the fourth embodiment of the application;

[0045] Figure 5 The flow chart of the information recommendation method provided in the fifth embodiment of the application. DETAILED DESCRIPTION

[0046] In the following description, many specific details are set forth in order to provide a thorough understanding of the application. However, the application can be practiced without the specific details, and it is understood that the application can be implemented in many different ways, as would be understood by one skilled in the art, without departing from the spirit and scope of the application. Accordingly, the present application is not intended to be limited to the specific embodiments described below.

[0047] The application provides a training method and device of a recommendation model and an electronic device. The application also provides two information recommendation methods. The embodiments will be described in detail one by one in the following.

[0048] The training method of the recommendation model provided in the application is a cross-scene transfer learning method for training a model, and the core of the method is that: according to the characteristics of a target field, a large amount of complete sample data of a source field is transferred to the target field, and a recommendation model required by the target field is trained according to the sample data of the source field. The method effectively avoids the problem that the amount of data of the target field is insufficient, the data completeness is not high, and the target field recommendation model cannot be constructed based on the data information of the target field.

[0049] The training method of the recommendation model provided in the application aims to obtain a recommendation model required by a target field, and the recommendation model is used to recommend a recommendation object related to trigger data to a user according to a user trigger object.

[0050] Please refer to Figure 1 , which is a flowchart of the training method of the recommendation model provided in the first embodiment of the application. The method comprises the following steps:

[0051] Step S101: selecting sample data consistent with the characteristics of the target field as source field selected sample data in the source field.

[0052] In the first embodiment of the application, the sample data of the source field is the data that needs to be transferred and learned. Transfer learning is a type of learning method in machine learning. Specifically, transfer learning is to transfer the sample data of a trained model to a new model to assist the training of the new model and improve the use effect of the new model. In the process of model training, using the transfer learning can accelerate the training efficiency and use effect of the new model, and avoid obtaining a new model from zero learning.

[0053] The source field refers to a known scene with a large amount of complete sample data. For example: a certain e-commerce platform with a large amount of data in the e-commerce industry, a news platform with a certain influence in the news industry, etc.

[0054] The source field selected sample data is obtained by screening the sample data from at least one of the following aspects:

[0055] The use scene of the sample data is consistent with the target field, for example: the objects contained in the target field are used in the fitness scene (such as a gym), and then the source field selected sample data is composed of the sample data of the fitness-related objects in the source field.

[0056] The classification of the objects involved in the sample data is consistent with the target field, for example: the sample data required by the target field only involves data of the clothing category, and then the source field selected sample data is composed of the sample data of the clothing category in the source field.

[0057] The classification of the subject involved in the sample data is consistent with the target field, for example: assuming that the target field is aimed at children, then the source field is selected from the scene involving children in the source field.

[0058] It can be understood that in the first embodiment of the present application, although the sample data selected in the source field is consistent with the characteristics of the target field, the objects included in the target field and the objects in the source field are still different. Therefore, when the source field is constructed and the selected sample data is selected, the name of the recommended object cannot be used as the basis for constructing the alternative sample data.

[0059] In this embodiment, the sample data in the source field refers to the information describing the recommended event obtained in the scene of the source field according to the recommended event. Generally, corpus information describing the recommended event is selected. For example: according to the user's past purchase of item A, the user is recommended item B in a certain e-commerce platform, then item A is the object triggering the recommended event, and item B is the recommended object formed according to the triggering event. The information describing the recommended event is the information describing item A and the information describing item B.

[0060] When selecting sample data consistent with the characteristics of the target field in the source field, the source field recommendation log recorded by the system can be selected.

[0061] Specifically, it includes: obtaining a source field recommendation behavior log; the source field recommendation behavior log can record the recommended events occurring in the source field, for example: the process of the user purchasing item A and the recommended event occurring will be recorded in the system log. That is, the recommended events recorded in the system log include not only the information of item A but also the information of item B.

[0062] Therefore, at least one sample data can be formed according to the source field recommendation behavior log and the project data associated with the data recorded in the behavior log according to the predetermined sample data structure template.

[0063] In specific applications, the sample data is usually expressed in the form of binary classification, that is, according to the use of the sample data, the sample data is divided into positive sample data and negative sample data. Specifically, the number of successful recommendations can be preset, the sample data is selected from the source field recommendation log, and the sample data is classified. The data recorded in the recommended event that meets the number of successful recommendations and the associated project data are positive samples; the data recorded in the recommended event that does not meet the number of successful recommendations and the associated project data are negative samples.

[0064] The recommended success number is the number of times that a user responds after an object is recommended to the user. For example, a user purchases a product on an e-commerce platform, and the platform pushes a series of recommended products related to the product purchased by the user to the customer. If the user needs or is interested in the recommended product, the user will selectively purchase or browse the recommended product. If the user purchases or browses the recommended product, it means that the recommendation of the recommended product is successful. It can be understood that there are a large number of users on the e-commerce platform. If the number of times that the recommended product is successfully recommended is greater than the preset recommended success number, it means that the product previously purchased by the user is related to the recommended product. If the number of times that the recommended product is successfully recommended does not satisfy the preset recommended success number, it means that the product previously purchased by the user is not related to the recommended product.

[0065] In a specific application process, the source field selected sample data includes the following data elements: trigger data, target data, positive sample identifier, and negative sample identifier.

[0066] The trigger data is description information of an object triggering a recommendation event, and the target data is description information of a recommended object in the recommendation event.

[0067] The structure template of the source field selected sample data can be represented in the form of a triple, for example, the positive sample data can be <trigger data, target data, positive sample identifier>, and the negative sample data can be <trigger data, target data, negative sample identifier>. Optionally, to ensure the accuracy of the recommendation model obtained by sample training, the sampling ratio of the positive sample and the negative sample is generally 1:1.

[0068] Considering that if the similarity between the source field and the target field is insufficient, the transfer result obtained only according to the positive sample data and the negative sample data is not ideal, therefore, in this embodiment, structured information of the source field selected sample data also needs to be introduced. The structured information refers to information that is common to the trigger data and the target data in the sample data and can be obtained. It should be noted that the "common" here does not mean the same. For example, assuming that the trigger data corresponds to a shoe, and the target data also corresponds to a shoe, the structured information refers to information such as the name, color, price, and brand of the shoe.

[0069] Step S102, using the source field selected sample data to form training data, training a recommendation model to obtain a first recommendation model for the target field.

[0070] This step is a process of transferring the source field selected sample data to the target field for learning.

[0071] Since the description information of the selected sample data in the source field may not conform to the description habits of the target field, using sample data formed by text data, structured data, or a combination of text data and structured data for transfer learning may not achieve a more accurate recommendation effect.

[0072] Therefore, when using the selected sample data in the source field to form training data, the selected sample data in the source field needs to be aligned, that is, the selected sample data in the source field is aligned to form training data.

[0073] The alignment of the selected sample data in the source field refers to adjusting the description information of the trigger data and the description information of the target data in the sample data to conform to the description information of the target field, for example: assuming that the description information in the selected sample data in the source field is Chinese, and the description information required by the target field is English, then alignment refers to converting the Chinese description information in the selected sample data into English description information. For another example: assuming that the description information in the selected sample data in the source field includes the Internet, and the description information used by the target field to describe the Internet is the World Wide Web, then it refers to modifying the Internet in the selected sample data to the World Wide Web.

[0074] In the first embodiment of the present application, the untrained recommendation model loads the basic model parameters using a general pre-trained recommendation model. The pre-trained language model is a general language model. The pre-trained language model can learn the semantic expression of the training text by using a large amount of easily accessible training text and a deep structure model (for example: Bert / XLnet), and can transfer the learned content to other task scenarios (for example: semantic similarity, sentiment classification, entity recognition, etc.).

[0075] Since the selected sample data not only includes positive sample data and negative sample data composed of information describing the recommendation event, but also includes source field structured information generated according to the information describing the recommendation event, and the information describing the recommendation event is also aligned with the description habits of the target field. Therefore, the first recommendation model obtained by training the recommendation model based on the positive sample data, the negative sample data, and the structured information can not only be used as a recommendation model for the target field, but also the input object and the output object of the recommendation model in the target field are similar in structure (for example: similar in category, similar in semantics, similar in emotion, etc.).

[0076] Since the recommendation model training method provided in the embodiment not only considers the description information of the recommendation event in the source field, but also considers the similarity in structure of the data contained in the positive sample and the negative sample. Therefore, in the case where the source field and the target field have a certain correlation, the first recommendation model can be deployed as a recommendation model for the target field.

[0077] In the target field, in order to make the first recommendation model more accurate to complete the related recommendation task of the target field, the training method of the recommendation model in the first embodiment of the application also involves fine-tuning of the first model.

[0078] Fine-tuning refers to training the first recommendation model based on the existing first recommendation model for different related recommendation tasks using limited sample data in the target field, so that the first recommendation model adapts to the recommendation task of the target field to improve the recommendation effect.

[0079] The process of fine-tuning includes obtaining target field sample data in the data of the target field, and using the target field sample data to train the first recommendation model to obtain a second recommendation model.

[0080] The method of obtaining the target field sample data is basically the same as the method of obtaining the selected sample data of the source field. It is obtained based on the recommendation behavior log of the target field. For details, please refer to the process of obtaining the selected sample data of the source field, which will not be described further.

[0081] It can be understood that the second recommendation model obtained by training the first recommendation model based on the target field sample data can be better adapted to the target field to complete the related recommendation task.

[0082] After obtaining the target field recommendation model, the target field recommendation model can be deployed and served offline / online based on a specific use scenario and a server side (the server side includes a server or a cloud side) to realize the related recommendation function of the target field in the current use scenario. Of course, in other embodiments, the target field recommendation model can also be deployed on a terminal device, and this embodiment is not limited.

[0083] In order to verify the recommendation effect of the application and the target field recommendation model, the target field recommendation model is trained and evaluated according to the AIREC intelligent recommendation engine.

[0084] Since the input target object and the output to-be-recommended object of the target field recommendation model can only be related or unrelated, AUC (Area Under Curve) is used as an evaluation index for evaluating the target field recommendation model.

[0085] AUC is a model evaluation index in the field of machine learning, and its value is between [0, 1]. The larger the value of AUC, the higher the accuracy of the model.

[0086] Next, the scoring process of the recommendation model and the result of the scoring are explained.

[0087] The evaluation successively increases the training sample data to train the recommendation model to construct the target field recommendation model:

[0088] 1. Without training sample, the untrained recommendation model is used as the third recommendation model;

[0089] 2. On the basis of the third recommendation model, a small amount of sample data of the source field is used to train the recommendation model to construct the recommendation model, and the fourth recommendation model is obtained;

[0090] 3. On the basis of the fourth recommendation model, a small amount of sample data of the target field is used to train the recommendation model to construct the fifth recommendation model;

[0091] 4. On the basis of the fifth recommendation model, a large amount of sample data of the source field and a small amount of sample data of the target field are used to train the recommendation model to construct the sixth recommendation model. Based on the AIREC intelligent recommendation engine, the following evaluation data is obtained:

[0092]

[0093]

[0094] The experimental results show that the recommendation effect obtained by directly applying the recommendation model to the target field is not ideal, but the recommendation effect of the recommendation model trained by the sample data of the source field and fine-tuned by a small amount of sample data of the target field is obviously improved. At the same time, on the basis of a large amount of training sample data of the source field, the recommendation effect of the recommendation model fine-tuned by a small amount of sample data of the target field is the best.

[0095] In addition, on the basis of the experiment, the application also verifies the influence of the structured information of the sample data on the recommendation effect of the recommendation model. The experiment proves that, in the target field, the recommendation model trained by the sample data constructed by the structured information has a stable improvement of 0.2% to 0.3% in the recommendation effect compared with the sample data not constructed by the structured information.

[0096] In summary, the training method of the recommendation model provided by the application migrates a large amount of complete sample data of the source field to the target field according to the characteristics of the target field, obtains the recommendation model required by the target field according to the sample data of the source field, realizes the purpose of cross-scene migration of the sample data of the source field to obtain the target field recommendation model, and effectively avoids the problem that the data quantity of the target field is insufficient and the data completeness is not high, and the sample data of the target field cannot be used to construct the recommendation model suitable for the target field.

[0097] The above embodiment introduces a training method of a recommendation model, and specifically introduces the training process and advantages of the recommendation model. Correspondingly, the second embodiment of the present application provides a training device of a recommendation model, please refer to Figure 2 , which is a structural schematic diagram of the training device of the recommendation model according to the second embodiment of the present application. Since the device embodiment is basically similar to the above-mentioned method embodiment, the description is relatively simple, and the related parts refer to the part of the above-mentioned method embodiment. The device embodiment described below is only illustrative.

[0098] The device comprises:

[0099] The first sample data acquisition unit 201 selects sample data conforming to the target field feature as source field selected sample data in the source field.

[0100] The first recommendation model generation unit 202 is used to form training data using the source field selected sample data, train the recommendation model, and obtain the first recommendation model for the target field.

[0101] Optionally, the device further comprises:

[0102] The second sample data acquisition unit is used to obtain target field sample data in the data of the target field.

[0103] The second recommendation model generation unit is used to train the first recommendation model using the target field sample data, and obtain the second recommendation model.

[0104] Optionally, the source field selected sample data is obtained by screening sample data from at least one of the following aspects: the use scene of the sample data is consistent with the target field, the classification of the items involved in the sample data is consistent with the target field, and the classification of the subjects involved in the sample data is consistent with the target field.

[0105] Optionally, the first recommendation model generation unit 202 comprises:

[0106] The alignment sub-unit is used to form training data after aligning the source field selected sample data.

[0107] Optionally, the source field selected sample data and the target field sample data adopt data elements of the following data types: text data, structured data, or a combination of text data and structured data.

[0108] Optionally, the first sample data acquisition unit 201 comprises:

[0109] The log acquisition unit is used to acquire source field recommendation behavior logs.

[0110] The first sample data obtaining subunit is configured to form at least one sample data according to the source field recommendation behavior log and item data associated with the data recorded in the behavior log, and according to a predetermined sample data structure template.

[0111] Optionally, the source field selected sample data and the target field sample data include the following data elements: trigger data, target data, positive and negative sample identification.

[0112] In the above embodiments, a recommendation method and device of a recommendation model are respectively provided. Correspondingly, the third embodiment of the present application further provides an electronic device. Since the electronic device is basically similar to the above-mentioned method embodiments and device embodiments, the description is relatively simple, and the relevant parts refer to the above-mentioned method embodiments and device embodiments. The electronic device described below is only illustrative.

[0113] Please refer to Figure 3 The electronic device provided by the third embodiment of the present application includes:

[0114] The processor 301;

[0115] The memory 302 is configured to store the program of the method. When the program is read and executed by the processor, the following steps are performed: selecting sample data conforming to the target field features as source field selected sample data in the source field; using the source field selected sample data to form training data, training the recommendation model, and obtaining a first recommendation model for the target field.

[0116] The fourth embodiment of the present application provides an information recommendation method. Please refer to Figure 4 The information recommendation method flowchart provided by the fourth embodiment of the present application is compared with the first embodiment. The present embodiment is the use process of the first recommendation model provided by the first embodiment. The same parts of the present embodiment and the first embodiment are not described again.

[0117] The information recommendation method provided by the fourth embodiment of the present application includes the following steps:

[0118] Step S401: obtaining trigger data.

[0119] Step S402: providing the trigger data to the first recommendation model of the target field. The first recommendation model is configured to provide target data as recommendation data according to the trigger data.

[0120] The first recommendation model is obtained in the following way:

[0121] Selecting sample data conforming to the features of the target field as source field selected sample data in the source field;

[0122] Using the source field selected sample data to form training data, training the recommendation model, and obtaining a first recommendation model for the target field.

[0123] In a specific use scenario, the information recommendation method described in this embodiment is configured with a server (the server includes a server and / or a cloud). Of course, in other embodiments, the method can also be configured with a terminal device, and this embodiment does not limit this.

[0124] In this embodiment, the trigger data refers to data triggered by a user, for example, information corresponding to a product triggered by a user for browsing / purchasing in an application platform is trigger data.

[0125] After the trigger data is provided to the recommendation model of the target field, the recommendation model will provide recommendation data similar to the trigger data to the terminal service interface corresponding to the user according to the input trigger data.

[0126] The fifth embodiment of the present application also provides an information recommendation method, please refer to Figure 5 The information recommendation method flowchart provided by the fifth embodiment of the present application is compared with the first embodiment described above. The use process of the second recommendation model provided by the first embodiment is basically similar to the first embodiment and the fourth embodiment described above. The same parts of this embodiment and the first embodiment and the fourth embodiment are not described again.

[0127] The information recommendation method provided by the fifth embodiment of the present application includes the following steps:

[0128] Step S501, obtaining trigger data.

[0129] Step S502, providing the trigger data to the second recommendation model of the target field, and the second recommendation model is used to provide target data as recommendation data according to the trigger data.

[0130] The recommendation model is obtained in the following way:

[0131] Selecting sample data conforming to the features of the target field as source field selected sample data in the source field;

[0132] Using the source field selected sample data to form training data, training the recommendation model, and obtaining a first recommendation model for the target field.

[0133] In the data of the target field, target field sample data is collected and obtained;

[0134] The first recommendation model is trained by using the target field sample data to obtain a second recommendation model.

[0135] The application discloses the above-mentioned preferred embodiments, but is not intended to limit the application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the application. Therefore, the protection scope of the application should be subject to the scope defined by the claims of the application.

[0136] In one typical arrangement, the computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0137] The memory can include non-persistent memory in the computer readable medium, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM), or flash memory. The memory is an example of computer readable media.

[0138] 1. Computer readable media includes both permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology for storing information. Information can be computer readable instructions, data structures, program modules 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 technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. According to the definition in this paper, computer readable media does not include non-transitory computer readable media, such as modulated data signals and carriers.

[0139] 2. Those skilled in the art should understand that the embodiments of the application can be provided as a system or an electronic device. Therefore, the application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the application can be in the form of a computer program product implemented 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.

Claims

1. A method for training a recommendation model, characterized in that, include: Select sample data from the source domain that match the characteristics of the target domain as the selected sample data from the source domain; The selected sample data from the source domain is used to form training data, and the recommendation model is trained to obtain a first recommendation model for the target domain. The step of selecting sample data in the source domain that matches the characteristics of the target domain includes: Obtain source domain recommendation behavior logs and recommendation success counts, wherein the recommendation success counts represent the number of responses to the recommended object; based on the recommendation success counts, the source domain recommendation behavior logs, and the item data associated with the data recorded in the behavior logs, form at least one sample data according to a predetermined sample data structure template.

2. The method according to claim 1, characterized in that, The method further includes: Obtain target domain sample data from the data in the target domain; Using the target domain sample data, the first recommendation model is trained to obtain the second recommendation model.

3. The method according to claim 1, characterized in that, The selected sample data in the source domain is obtained by filtering the sample data from at least one of the following perspectives: the usage scenario of the sample data is consistent with the target domain, the classification of the items involved in the sample data is consistent with the target domain, and the classification of the subjects involved in the sample data is consistent with the target domain.

4. The method according to claim 1, characterized in that, The step of forming training data using the selected sample data from the source neighborhood includes: aligning the selected sample data from the source neighborhood to form training data.

5. The method according to claim 2, characterized in that, The selected sample data from the source domain and the sample data from the target domain use data elements of the following data types: text data, structured data, or a combination of text data and structured data.

6. The method according to claim 2, characterized in that, The source domain selected sample data and the target domain sample data include the following data elements: trigger data, target data, and positive and negative sample identifiers.

7. A device for obtaining a relevant recommendation model, characterized in that, include: The first sample data acquisition unit is used to select sample data in the source domain that matches the characteristics of the target domain as the selected sample data in the source domain; The first recommendation model generation unit is used to form training data using the selected sample data from the source domain, train the recommendation model, and obtain a first recommendation model for the target domain. The first sample data acquisition unit is further configured to: Obtain source domain recommendation behavior logs and recommendation success counts, wherein the recommendation success counts represent the number of responses to the recommended object; based on the recommendation success counts, the source domain recommendation behavior logs, and the item data associated with the data recorded in the behavior logs, form at least one sample data according to a predetermined sample data structure template.

8. An electronic device, characterized in that, include: processor; A memory is used to store a program for a method, which, when read and executed by the processor, performs the following steps: selecting sample data in the source domain that matches the features of the target domain as selected sample data in the source domain; using the selected sample data in the source domain to form training data, training a recommendation model, and obtaining a first recommendation model for the target domain; wherein, selecting sample data in the source domain that matches the features of the target domain includes: acquiring a recommendation behavior log and the number of successful recommendations in the source domain, wherein the number of successful recommendations is used to represent the number of responses to the recommended object; forming at least one sample data according to a predetermined sample data structure template based on the number of successful recommendations, the recommendation behavior log in the source domain, and the item data associated with the data recorded in the behavior log.

9. An information recommendation method, characterized in that, include: Get the trigger data; The trigger data is provided to a first recommendation model in the target domain, and the first recommendation model is used to provide target data as recommendation data based on the trigger data. The first recommendation model is obtained in the following way: Select sample data from the source domain that match the characteristics of the target domain as the selected sample data from the source domain; The selected sample data from the source domain is used to form training data, and the recommendation model is trained to obtain a first recommendation model for the target domain. The step of selecting sample data in the source domain that matches the characteristics of the target domain includes: Obtain source domain recommendation behavior logs and recommendation success counts, wherein the recommendation success counts represent the number of responses to the recommended object; based on the recommendation success counts, the source domain recommendation behavior logs, and the item data associated with the data recorded in the behavior logs, form at least one sample data according to a predetermined sample data structure template.

10. An information recommendation method, characterized in that, include: Get the trigger data; The trigger data is provided to a second recommendation model in the target domain, and the second recommendation model is used to provide target data as recommendation data based on the trigger data. The second recommendation model is obtained in the following way: Select sample data from the source domain that match the characteristics of the target domain as the selected sample data from the source domain; The selected sample data from the source domain is used to form training data, and the recommendation model is trained to obtain a first recommendation model for the target domain. Sample data of the target domain are collected from the data in the target domain. Using the target domain sample data, the first recommendation model is trained to obtain the second recommendation model; The step of selecting sample data in the source domain that matches the characteristics of the target domain includes: Obtain source domain recommendation behavior logs and recommendation success counts, wherein the recommendation success counts represent the number of responses to the recommended object; based on the recommendation success counts, the source domain recommendation behavior logs, and the item data associated with the data recorded in the behavior logs, form at least one sample data according to a predetermined sample data structure template.

Images