Interest model training method and device, interest prediction method and device and electronic equipment

Abstract

The invention relates to the technical field of machine learning, in particular to an interest model training method and device, an interest prediction method and device and electronic equipment, and the method comprises the steps: constructing a fine tuning sample based on object association information of a sample object; the fine tuning sample comprises an input interest of the sample object and a label interest of the sample object; performing interest prediction on the sample object based on a pre-training language model and the input interest of the sample object to obtain a first prediction interest of the sample object; and performing model parameter fine adjustment on the pre-training language model based on the loss information of the first prediction interest of the sample object and the label interest of the sample object to obtain a target interest model for object interest prediction. According to the method, the information cocoon room can be broken, the object can contact more information, and the recommendation content is enriched; and in the cold start scene, the interested information can be recommended for the new object or the low-frequency object by predicting the potential interest of the new object or the low-frequency object.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a method for training an interest model, an interest prediction method, a device, and an electronic device. Background Art

[0002] Existing information recommendation algorithms, such as the classic collaborative filtering algorithm and the click sequence-based transformer algorithm, all rely on the object's historical operation information, which will cause the recommendation system to be overly dependent on the object's historical operation information. On the one hand, the content that the object is interested in will always be recommended, and the object will habitually be guided by his or her own interests, losing the opportunity to understand and come into contact with different things, thus creating an information cocoon. On the other hand, in the cold start scenario, for new or low-frequency business objects, their historical operation information is relatively sparse, and the recommendation system cannot accurately recommend information of interest to them. Summary of the Invention

[0003] The technical problem to be solved by this application is to provide a training method, an interest prediction method, a device and an electronic device for an interest model to solve the information cocoon problem caused by relying on the historical operation information of the object in the information recommendation process in the existing technology, as well as the problem of being unable to accurately recommend information in a cold start scenario.

[0004] In order to solve the above technical problems, on the one hand, this application provides a training method for an interest model, including:

[0005] Constructing a fine-tuning sample based on object association information of the sample object; the object association information includes the existing interests of the sample object, the fine-tuning sample includes the input interests of the sample object and the label interests of the sample object, the input interests are determined based on the existing interests of the sample object, and the label interests are associated with the input interests;

[0006] performing interest prediction on the sample object based on a pre-trained language model and the input interest of the sample object to obtain a first predicted interest of the sample object;

[0007] Fine-tune model parameters of the pre-trained language model based on the first predicted interest of the sample object and the loss information of the label interest of the sample object to obtain a target interest model for object interest prediction.

[0008] On the other hand, the present application provides an interest prediction method, comprising:

[0009] Get the target object's existing interest;

[0010] Based on the target interest model and the existing interest of the target object, the interest of the target object is predicted to obtain the target predicted interest of the target object; the target interest model is trained based on the training method of the above interest model.

[0011] On the other hand, the present application provides a training device for an interest model, comprising:

[0012] a fine-tuning sample construction module, configured to construct a fine-tuning sample based on object association information of a sample object; the object association information includes the existing interests of the sample object; the fine-tuning sample includes the input interests of the sample object and the label interests of the sample object; the input interests are determined based on the existing interests of the sample object; and the label interests are associated with the input interests;

[0013] a training prediction module, configured to predict the interest of the sample object based on the pre-trained language model and the input interest of the sample object, to obtain a first predicted interest of the sample object;

[0014] A parameter fine-tuning module is used to fine-tune the model parameters of the pre-trained language model based on the first predicted interest of the sample object and the loss information of the label interest of the sample object to obtain a target interest model for object interest prediction.

[0015] On the other hand, the present application provides an interest prediction device, comprising:

[0016] Interest acquisition module, used to obtain the existing interests of the target object;

[0017] An application prediction module is used to predict the interest of the target object based on the target interest model and the existing interest of the target object to obtain the target predicted interest of the target object; the target interest model is trained based on the training method of the above interest model.

[0018] On the other hand, the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the training method of the interest model or the interest prediction method as described above.

[0019] On the other hand, the present application provides a computer storage medium, which stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by a processor such as the above-mentioned training method of the interest model or the interest prediction method.

[0020] The implementation of the embodiments of the present application has the following beneficial effects:

[0021] In this application, by constructing fine-tuning samples to fine-tune the pre-trained language model, a target interest model is trained to predict the potential interests of objects. With the help of the general knowledge learned by the large language model, the problem of object interest exploration is effectively solved. In the constructed fine-tuning samples, the input interests of the sample objects correspond to the label interests of the sample objects one by one. The input interests of the sample objects are determined based on the sample objects' existing interests. The label interests of the sample objects are associated with the sample objects' existing interests. The constructed input interest-label interest sample pairs reflect the intrinsic correlation between interests. The fine-tuning samples constructed based on the sample pairs can fully cover the correlation between the different interests of each sample object. Then, based on the constructed fine-tuning samples, the parameters of the pre-trained language model are fine-tuned. The resulting target interest model can predict potential interests based on the correlation between different interests and existing interests, and then predict target interests associated with existing interests, thereby improving the accuracy of interest prediction. Furthermore, information recommendation for objects based on the predicted potential interests can break the information cocoon, allowing objects to be exposed to more information and enrich the recommended content. In addition, in cold start scenarios, by predicting the potential interests of new objects or low-frequency objects, interesting information can be recommended to them. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0023] Figure 1 This is a schematic diagram of the implementation environment provided by the embodiment of the present application;

[0024] Figure 2 This is a flow chart of a method for training an interest model provided in an embodiment of the present application;

[0025] Figure 3 This is a flow chart of a fine-tuning sample construction method provided in an embodiment of the present application;

[0026] Figure 4 This is a flow chart of another fine-tuning sample construction method provided in an embodiment of the present application;

[0027] Figure 5 This is a flow chart of a method for generating interest prediction guidance information provided by an embodiment of the present application;

[0028] Figure 6 This is a flow chart of another method for training an interest model provided in an embodiment of the present application;

[0029] Figure 7 This is a flow chart of an interest prediction method provided by an embodiment of the present application;

[0030] Figure 8 Schematic diagram of a training device for an interest model provided in an embodiment of the present application;

[0031] Figure 9 This is a schematic diagram of an interest prediction device provided in an embodiment of the present application;

[0032] Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions, and advantages of this application more clear, this application will be further described in detail below with reference to the accompanying drawings. It is clear that the embodiments described are only some of the embodiments of this application, and not all of them. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0034] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0035] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0036] It is understandable that in the specific implementation of this application, related data such as user information is involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0037] See also Figure 1 , which shows a schematic diagram of the implementation environment provided by an embodiment of the present application, the implementation environment may include: at least one object terminal 110, an interest processing server 120 and a recommendation server 130, the object terminal 110 and the interest processing server 120 can communicate data through the network, the interest processing server 120 and the recommendation server 130 can communicate data through the network, and the object terminal 110 and the recommendation server 130 can also communicate data through the network.

[0038] Specifically, the interest processing server 120 can obtain object-related information of the object from the object server 110. The object-related information may include object attribute information and information such as the object's existing interests. The interest processing server 120 can construct fine-tuning samples based on the object-related information and fine-tune the parameters of the pre-trained language model based on the constructed fine-tuning samples to obtain a target interest model for object interest prediction. Furthermore, the interest processing server 120 can predict the potential interests of each object based on the target interest model to obtain the target predicted interests corresponding to each object.

[0039] The recommendation server 130 may obtain the existing interests and target predicted interests of each object from the interest processing server 120 , and may then recommend information for each object based on the existing interests and target predicted interests.

[0040] The target terminal 110 can communicate with the interest processing server 120 based on a browser / server (B / S) model or a client / server (C / S) model, and can also communicate with the recommendation server 130 based on a browser / server (B / S) model or a client / server (C / S) model. The target terminal 110 may include a physical device such as a smartphone, tablet computer, laptop computer, digital assistant, smart wearable device, or vehicle-mounted terminal, and may also include software running on the physical device, such as an application. The operating system running on the target terminal 110 in the embodiments of the present application may include, but is not limited to, Android, iOS, Linux, Windows, etc.

[0041] Communication connections can be established between the object terminal 110 and the interest processing server 120, the interest processing server 120 and the recommendation server 130, and the interest processing server 120 and the recommendation server 130 via wired or wireless communication. The interest processing server 120 or the recommendation server 130 may include an independently running server, or a distributed server, or a server cluster consisting of multiple servers, wherein the server may be a cloud server.

[0042] It should be noted that, in one example, the interest processing server 120 and the recommendation server 130 are two separate servers, and are deployed on different platforms respectively; in another example, the interest processing server 120 and the recommendation server 130 can be integrated into one server, which is not specifically limited in this embodiment.

[0043] In order to solve the problem of information cocoon caused by the reliance on historical operation information of objects in the information recommendation process in the prior art, as well as the problem of inability to accurately recommend information in cold start scenarios, this application provides a training method for an interest model, the execution subject of which can be the above-mentioned interest processing server; please refer to Figure 2 , the method may include:

[0044] S210. Construct a fine-tuning sample based on the object association information of the sample object; the object association information includes the existing interest of the sample object, the fine-tuning sample includes the input interest of the sample object and the label interest of the sample object, the input interest is determined based on the existing interest of the sample object, and the label interest is associated with the input interest.

[0045] The object association information in this embodiment may include object attribute information and information such as the object's existing interests. When constructing a fine-tuning sample, the fine-tuning sample may be constructed based on the sample object's existing interests, or based on the sample object's existing interests and object attribute information.

[0046] Fine-tuning samples can be used in the supervised model fine-tuning process, so that when constructing fine-tuning samples based on the existing interests of sample objects, the input interests can be determined from the existing interests of sample objects, and the label interests associated with the input interests can be further determined, so that for each sample object, a sample pair of {input interest-label interest} can be constructed, that is, in the model fine-tuning training process, the input interest can be used as input information for model prediction; when constructing fine-tuning samples based on the existing interests of sample objects and object attribute information, the input interests can be determined from the existing interests of sample objects, and the label interests associated with the input interests can be further determined, so that for each sample object, a sample pair of {input interest, object attribute information-label interest} can be constructed, that is, in the model fine-tuning training process, the input interest and object attribute information can be used together as input information for model prediction.

[0047] In this embodiment, the input interest is determined based on the existing interests of the sample object. The input interest may include all interests in the existing interests, or the input interest may include some interests in the existing interests. When the input interest includes all interests in the existing interests, the corresponding label interest may be an interest predicted based on the existing interests, or an interest predicted based on the existing interests and object attribute information, that is, the label interest is predicted based on the input interest, and thus the label interest is associated with the input interest; when the input interest includes some interests in the existing interests, the corresponding label interest may be the remaining interests in the existing interests, that is, the collection of the input interest set and the label interest set is the existing interest. In this case, the input interest and the label interest are both the existing interests of the sample object. For the sample object, there is similarity or intrinsic connection between the existing interests, so there is an association between the existing interests, and thus the input interest is associated with the label interest; when the input interest includes some interests in the existing interests, the corresponding label interest may include the remaining interests in the existing interests and the interests predicted based on the input interests, or may include the remaining interests in the existing interests and the interests predicted based on the input interests and object attribute information.

[0048] S220. Predict the interest of the sample object based on the pre-trained language model and the input interest of the sample object to obtain a first predicted interest of the sample object.

[0049] The pre-trained language model can be a BERT-based language model, a Transformer-based language model, etc., which is not specifically limited in this embodiment; the pre-trained language model has learned prior knowledge, so that the input interest in the fine-tuning sample corresponding to the specific task scenario is input into the pre-trained language model, and the pre-trained language model can predict the potential interest of the sample object based on the learned prior knowledge to obtain the corresponding first predicted interest; or the input interest and object attributes in the fine-tuning sample corresponding to the specific task scenario are input into the pre-trained language model, and the pre-trained language model can predict the potential interest of the sample object based on the learned prior knowledge to obtain the corresponding first predicted interest.

[0050] S230. Fine-tune the model parameters of the pre-trained language model based on the first predicted interest of the sample object and the loss information of the label interest of the sample object to obtain a target interest model for object interest prediction.

[0051] The label interest in the fine-tuning sample can be used as a prediction target, so that the first predicted interest actually predicted by the pre-trained language model can be compared with the label interest in the fine-tuning sample to determine the corresponding loss information, so that the model parameters can be fine-tuned based on the loss information of the label interest and the first predicted interest to obtain a target interest model for object interest prediction. Specifically, the fine-tuning method in this embodiment can adopt LoRA fine-tuning (Low-Rank Adaptation of Large Language Models, low-rank adaptation of large language models), SFT fine-tuning (Supervised Fine-Tuning) and other fine-tuning methods. It should be noted that the model fine-tuning methods that can be applied to this application in the prior art can be applied to this embodiment and are not limited here.

[0052] In this application, by constructing fine-tuning samples to fine-tune the pre-trained language model, a target interest model is trained to predict the potential interests of objects. With the help of the general knowledge learned by the large language model, the problem of object interest exploration is effectively solved. In the constructed fine-tuning samples, the input interests of the sample objects correspond to the label interests of the sample objects one by one. The input interests of the sample objects are determined based on the sample objects' existing interests. The label interests of the sample objects are associated with the sample objects' existing interests. The constructed input interest-label interest sample pairs reflect the intrinsic correlation between interests. The fine-tuning samples constructed based on the sample pairs can fully cover the correlation between the different interests of each sample object. Then, based on the constructed fine-tuning samples, the parameters of the pre-trained language model are fine-tuned. The resulting target interest model can predict potential interests based on the correlation between different interests and existing interests, and then predict target interests associated with existing interests, thereby improving the accuracy of interest prediction. Furthermore, information recommendation for objects based on the predicted potential interests can break the information cocoon, allowing objects to be exposed to more information and enrich the recommended content. In addition, in cold start scenarios, by predicting the potential interests of new objects or low-frequency objects, interesting information can be recommended to them.

[0053] In this embodiment, when the input interest includes part of the existing interests, the corresponding label interest can be the remaining interests in the existing interests, that is, the combination of the input interest set and the label interest set is the existing interest. In this case, the input interest and the label interest are both the existing interests of the sample object. For the sample object, there are similarities or internal connections between the existing interests, so that there is a correlation between the existing interests, and thus the input interest is associated with the label interest; please refer to Figure 3 , which shows a fine-tuning sample construction method, which may include:

[0054] S310. Perform random masking on the existing interests of the sample object to obtain at least one masked interest.

[0055] In this embodiment, random masking of interests can be specifically understood as randomly masking one or more interests in the existing interests, thereby obtaining at least one masked interest; each interest in the existing interests can be used as an object to be masked, and then the masked object is randomly masked. It should be noted that the number of masked interests in the existing interests and the number of unmasked interests in the existing interests are not specifically limited. The number of masked interests can be greater than the number of unmasked interests, the number of masked interests can be equal to the number of unmasked interests, and the number of masked interests can be less than the number of unmasked interests.

[0056] S320. Determine the at least one mask interest as the tag interest.

[0057] The masked interest can be determined as the label interest and can be used as the prediction target during the model fine-tuning process to calculate the loss with the actual prediction information of the pre-trained language model.

[0058] S330. Determine the interests of the sample object other than the tag interests as the input interests.

[0059] For example, the existing interests include {interest 1, interest 2, interest 3, interest 4}, and the masked interests obtained after random masking, that is, the label interests include {interest 2, interest 4}, and the corresponding input interests may include {interest 1, interest 3}.

[0060] S340. Determine the fine-tuning sample based on the tag interest and the input interest.

[0061] In this embodiment, both label interests and input interests can be determined based on the existing interests of the sample object, and label interests can be determined by random masking, and then the input interests can be determined, thereby improving the convenience and efficiency of determining label interests and input interests, and thus improving the convenience and efficiency of fine-tuning sample construction.

[0062] In this embodiment, when the input interest includes all interests in the existing interests, the corresponding tag interest can be an interest predicted based on the existing interests, or an interest predicted based on the existing interests and object attribute information, that is, the tag interest is predicted based on the input interest, so that the tag interest is associated with the input interest; please refer to Figure 4 , which shows another fine-tuning sample construction method, which may include:

[0063] S410. Generate interest prediction guidance information based on the interest prediction intention; the interest prediction guidance information is used to guide the generation method of the interest prediction result.

[0064] The interest prediction intention in this embodiment can characterize the data format or expression form of the predicted interest, etc., and accordingly, corresponding prediction guidance information can be generated based on the interest prediction intention, and the prediction guidance information can be used to guide the preset general language model to output the predicted interest that meets the interest prediction intention. The interest prediction guidance information can specifically be a paragraph of text, a sentence or a question, which is used to guide the preset general language model to generate text or perform a specific task. It is the starting point for the interaction between the user and the model, and is used to clarify the user's needs and intentions; the role of the interest prediction guidance information is to guide the preset general language model to perform text generation or other natural language processing tasks. The preset general language model will understand the user's requirements based on the interest prediction guidance information and generate corresponding text as a response. The preset general language model can adopt models such as GPT-3 and GPT-4. Since calling this type of model consumes a lot of resources, it can be applied to small-batch prediction scenarios, such as the fine-tuning sample construction scenario in this embodiment.

[0065] For the method of generating interest prediction guidance information, please refer to Figure 5 , the method may include:

[0066] S510. Determine a candidate interest set based on the interest prediction intention; the predicted second predicted interests are all included in the candidate interest set.

[0067] In this embodiment, the predicted interest range may be predetermined to form a candidate interest set, so that all predicted interests fall within the candidate interests.

[0068] S520. Generate at least one predicted example; each predicted example includes input example interest and predicted example interest.

[0069] The prediction sample can indicate information such as the content and data format of the predicted output information; for example, the prediction sample a can be {wrong question-education}, where the input sample interest is a wrong question, and the corresponding prediction sample interest is education, that is, it is hoped that the preset general language model can predict the output interest of "education" when it receives the interest of "wrong question"; for another example, the prediction sample b can be {skin care products-beauty}, where the input sample interest is skin care products, and the corresponding prediction sample interest is education, that is, it is hoped that the preset general language model can predict the output interest of "beauty" when it receives the interest of "skin care products".

[0070] S530. Generate the interest prediction guidance information based on the candidate interest set and the at least one prediction example.

[0071] Once the candidate interest set and predicted samples are determined, corresponding interest prediction guidance information can be generated. In the actual implementation process, other information can be added to the interest prediction guidance information based on the implementation situation. For example, in the above-mentioned prediction samples, in addition to the input sample interest and the predicted sample interest, the interest level of the predicted sample interest can also be included. The interest level of any interest can represent the degree to which the corresponding object is interested in any interest.

[0072] In this embodiment, corresponding interest prediction guidance information is generated by generating a candidate interest set and prediction samples, so that the preset general language model can generate predicted interests that meet the prediction intention based on the interest prediction guidance information. When the prediction intention changes, the interest prediction guidance information can be adjusted accordingly, thereby improving the flexibility and convenience of interest prediction.

[0073] S420. Call a preset general language model based on the interest prediction guidance information, and perform interest prediction on the sample object based on the preset general language model and the existing interest of the sample object to obtain a second predicted interest of the sample object.

[0074] The preset general language model can use models such as GPT-3 and GPT-4. Since calling such models consumes a lot of resources, they can be applied to small-batch prediction scenarios, such as the fine-tuning sample construction scenario in this embodiment. The preset general language model can predict interests based on the existing interests of the input sample object to obtain a second predicted interest that the sample object may be interested in.

[0075] S430. Determine the label interest of the sample object based on the second predicted interest.

[0076] In one example, the second predicted interest can be directly determined as the tag interest. In another example, the second predicted interest may include multiple predicted interests, and the tag interest can be determined from the second predicted interests based on the interest levels of the multiple predicted interests. The interest level of any interest can represent the degree of interest of the corresponding subject in that interest. Specifically, the multiple predicted interests can be sorted in descending order of interest level, and a preset number of interests ranked at the top can be used as tag interests. Alternatively, interests with interest levels greater than a preset value can be used as tag interests.

[0077] S440. Determine the existing interest of the sample object as the input interest.

[0078] The existing interests of the sample objects can be directly used as input interests and used as input information of the model during the model fine-tuning process.

[0079] S450. Determine the fine-tuning sample based on the tag interest and the input interest.

[0080] In this embodiment, after determining the existing interests of the sample object, interest prediction can be further performed based on the existing interests of the sample object to obtain corresponding predicted interests, thereby expanding the interest range of the sample object. The predicted interests are then used as label interests, and fine-tuning samples are constructed based on the label interests. This can expand the interest range of the fine-tuning samples, as well as the richness of interests, making it easier for the fine-tuned model to explore more new interests.

[0081] In this embodiment, for objects with rich interests, the input interest and label interest can be determined by using a random interest mask method, or the label interest can be determined by calling a preset general language model based on interest prediction guidance information, with the existing interest used as the input interest; for objects with sparse interests, the label interest can be determined by calling a preset general language model based on interest prediction guidance information, with the existing interest used as the input interest.

[0082] The fine-tuning sample in this embodiment may include the input interest of the sample object, the label interest of the sample object, and the label profile information of the sample object. The label profile information is generated based on text processing of the object association information. Specifically, the label profile information of the sample object can be generated based on a preset general language model. Specifically, the existing interest of the sample object can be input into the preset general language model for profile prediction to generate corresponding label profile information, or the object association information of the sample object can be input into the preset general language model for profile prediction to generate corresponding label profile information. The object association information of the sample object includes the object attribute information of the sample object and the existing interest of the sample object. The preset general language model can adopt models such as GPT-3 and GPT-4. Since calling this type of model consumes a lot of resources, it can be applied to small batch prediction scenarios, such as the fine-tuning sample construction scenario in this embodiment. Accordingly, please refer to Figure 6 , which shows another method for training an interest model, which may include:

[0083] S610. Predict the interest of the sample object based on the pre-trained language model and the input interest of the sample object to obtain a first predicted interest of the sample object; predict the profile of the sample object based on the pre-trained language model and the object association information of the sample object to obtain predicted profile information of the sample object.

[0084] In this embodiment, the pre-trained language model has the functions of interest prediction and profile information prediction. Specifically, the existing interests of the sample object and the object association information of the sample object are used as input of the pre-trained language model in a preset format. When the pre-trained language model receives the corresponding input information, it can predict the interest of the sample object based on the input interest of the sample object to obtain the first predicted interest of the sample object, and predict the profile of the sample object based on the object association information of the sample object to obtain the predicted profile information of the sample object. Interest prediction and profile information prediction can be performed simultaneously, that is, the model receives the input information of the existing interests of the sample object and the object association information of the sample object, and the model can output the first predicted interest and predicted profile information.

[0085] S620. Based on the loss information of the first predicted interest of the sample object and the label interest of the sample object, as well as the loss information of the label profile information of the sample object and the predicted profile information of the sample object, fine-tune the model parameters of the pre-trained language model to obtain the target interest model.

[0086] Loss information can be calculated for predicted interest and predicted profile information respectively. Based on the loss information corresponding to the predicted interest and the predicted profile information, the model parameters can be fine-tuned to obtain the corresponding target interest model. The target interest model has the functions of both interest prediction and indirect information prediction.

[0087] In this application, by constructing fine-tuning samples to fine-tune the pre-trained language model, a target interest model for predicting the potential interest of the object is obtained through training. With the help of the general knowledge learned by the large language model, the problem of object interest exploration is effectively solved. In the constructed fine-tuning samples, the input interest of the sample object corresponds to the label interest of the sample object one by one, the input interest of the sample object is determined based on the existing interest of the sample object, and the label interest of the sample object is associated with the existing interest of the sample object. The constructed input interest-label interest sample pair reflects the intrinsic correlation between interests. The fine-tuning samples constructed based on the sample pair can fully cover the correlation between the different interests of each sample object, and then the parameters of the pre-trained language model are fine-tuned based on the constructed fine-tuning samples. The obtained target interest model can predict the potential interest based on the correlation between different interests and the existing interest, and then predict the target interest associated with the existing interest, thereby improving the accuracy of interest prediction. In addition, while predicting the interest, it is also possible to generate personalized profile information for each object so that the objects can better understand each other, facilitate subsequent information interaction, and improve the efficiency of interaction between objects.

[0088] Furthermore, this embodiment may also include a method for generating existing interests, which may specifically include:

[0089] Acquiring original interests of the sample object; the original interests include interests of multiple information types;

[0090] Information integration processing is performed on the interests of the multiple information types to obtain existing interests of the target information type.

[0091] Interests of various information types may include interests of text type, interests of graphic type, interests of video type, etc. In the process of integrating interests of various information types, interests of graphic type and interests of video type may be unified into interests of target information type. In this embodiment, interests of target information type may refer to interests of text type, that is, text recognition and image recognition may be performed on interests of graphic type to determine the interests of text type corresponding to the interests of graphic type, and image recognition may be performed on interests of video type to determine the interests of text type corresponding to the interests of video type.

[0092] There may be certain coupling and correlation between interests of various information types, and there may be common points of interest. Through information integration processing, not only can the interests be deduplicated, but also the complementarity between interests of different information types can be achieved, such as the complementarity of text interests, graphic interests and video interests, effectively reducing redundancy and improving complementarity.

[0093] After the target interest model is trained, you can use the target interest model to predict the interest of the target object. For details, see Figure 7 , which shows an interest prediction method, which may include:

[0094] S710. Obtain the existing interest of the target object.

[0095] In the case where the target object's existing interests include interests of multiple information types, information integration processing can be performed on the interests of the multiple information types to obtain the target information type interests of the target object, and the target information type interests can be determined as the target object's existing interests. The interests of the multiple information types may include text type interests, graphic type interests, video type interests, etc. In the process of information integration of the interests of the multiple information types, graphic type interests and video type interests can be unified into target information type interests. In this embodiment, the target information type interests can refer to text type interests, that is, text recognition and image recognition can be performed on the graphic type interests to determine the text type interests corresponding to the graphic type interests, and image recognition can be performed on the video type interests to determine the text type interests corresponding to the video type interests.

[0096] S720. Predict the interest of the target object based on the target interest model and the existing interest of the target object to obtain the target predicted interest of the target object; the target interest model is trained based on the above-mentioned interest model training method.

[0097] The training method of the target interest model can refer to the above content of this embodiment and will not be repeated here; the target interest model can predict the interest of the target object based on the input target object's existing interest to obtain the target predicted interest of the target object.

[0098] In this embodiment, the constructed input interest-label interest sample pairs reflect the inherent correlation between interests. The fine-tuning samples constructed based on these sample pairs can fully cover the correlation between the different interests of each sample object. The parameters of the pre-trained language model are then fine-tuned based on the constructed fine-tuning samples. The resulting target interest model can predict potential interests based on the correlation between different interests and existing interests, and then predict target interests related to existing interests, thereby improving the accuracy of interest prediction. Furthermore, recommending information to objects based on predicted potential interests can break the information cocoon, allowing objects to be exposed to more information and enriching the recommended content. In cold start scenarios, by predicting the potential interests of new or low-frequency objects, interesting information can be recommended to them.

[0099] Furthermore, the target interest model may have the functions of interest prediction and profile information generation at the same time. Specifically, interest prediction is performed on the target object based on the target interest model and the existing interest of the target object, and profile prediction is performed on the target object based on the target interest model and the existing interest of the target object to obtain the target predicted interest of the target object and the predicted profile information of the target object.

[0100] In a specific content business scenario, the target object's prediction profile information may be displayed on the content page created by the target object, or on the target object's personal page.

[0101] The constructed input interest-label interest sample pairs reflect the inherent correlation between interests. The fine-tuning samples constructed based on these sample pairs can fully cover the correlations between the different interests of each sample object. The parameters of the pre-trained language model are then fine-tuned based on the constructed fine-tuning samples. The resulting target interest model can predict potential interests based on the correlations between different interests and existing interests, and then predict target interests related to existing interests, thereby improving the accuracy of interest prediction. Furthermore, while predicting interests, it can also generate personalized profile information for each object, allowing them to better understand each other and facilitate subsequent information exchange, thereby improving the efficiency of interaction between objects.

[0102] When the target predicted interest of the target object is predicted, the target predicted interest can be output as an interest in a specified format for easy online loading and use; for example, it can be applied to recommendation systems, social scenarios, etc.

[0103] When the target predicted interests of the target object are predicted, content push based on the target predicted interests can be achieved. Specifically, the following steps may be performed:

[0104] Determining target recommended content corresponding to the target object based on the existing interests of the target object and the target predicted interests of the target object;

[0105] Push the target recommended content to the target object.

[0106] It should be noted that during the actual push process, content can be pushed based on both the target subject's existing interests and the target predicted interests. Specifically, content can be pushed based on both the target subject's existing interests and the target predicted interests. In the initial stages of recommendation, the number of recommended content based on existing interests can be determined to be greater than the number of recommended content based on the target predicted interests to avoid the target subject's difficulty adapting to the large-scale recommended content based on the predicted interests. The effectiveness of the target predicted interests can be evaluated based on the subject's click-through rate on the recommended content. The target predicted interests may include multiple interests, each with a corresponding interest level. The interest level of any interest can represent the subject's level of interest in that interest. The number of recommended content can be determined based on the interest levels of each of the multiple interests, for example, the number of recommended content for interests with high interest levels is greater than the number of recommended content for interests with low interest levels. To ensure that the recommended content based on the target predicted interests is exposed to the target subject, interventions can be made in the recommendation system's recommendation strategy, such as allocating exposure quotas for recommended content based on the target predicted interests or determining exposure quotas based on the interest levels of multiple interests within the target predicted interests.

[0107] In this embodiment, when the target predicted interest of the target object is predicted, recommended content corresponding to the target predicted interest can be pushed to the target object. Information recommendation for the object based on the predicted potential interest can break the information cocoon, allow the object to be exposed to more information, and enrich the recommended content; and in the cold start scenario, by predicting the potential interests of new objects or low-frequency objects, information of interest can be recommended to them.

[0108] In one example, each object can have corresponding existing interests, exploration interests, and object profile:

[0109] Subject 1:

[0110] Interested: Holiday: 0.26861683, Wrong Question Book: 0.26861683, Software: 0.26861683, Luffa: 0.15735158, Cracked Version: 0.15699892, Tomato: 0.15108718, Portfolio: 0.15108718, Vest: 0.13128781

[0111] Exploration interests: Online education: 0.289616, Digital: 0.278616, Recipes: 0.268616, Content industry: 0.258616, Food culture: 0.248616, Creative photography: 0.238616

[0112] Profile of the candidate: A holiday academic master, a software fanatic, and an explorer of food and photography!

[0113] Subject 2:

[0114] Already interested in: Skin care products: 0.13320568

[0115] Exploration interests: Beauty and makeup: 0.245391, body plastic surgery: 0.198375, fashion events: 0.215678, food culture: 0.172849

[0116] Profile of the subject: Passionate about shopping and skin care, the pursuit of beauty never stops.

[0117] The decimal after each interest represents the interest level of that interest. The interest level of any interest can represent the degree of interest of the corresponding subject in that interest. For example, for subject 1, their interest level for vacation is 0.26861683, while their interest level for tomatoes is 0.15108718. This indicates that subject 1 is more interested in vacation.

[0118] The following takes Object 3 as an example to illustrate the specific implementation process of this application:

[0119] Subject 3:

[0120] Interested in: lottery, Super Lotto, finance

[0121] Exploration interests: Funds: 0.321346, Foreign Exchange: 0.289676, Bonds: 0.269676, Stocks: 0.24698, Macroeconomics: 0.233976, Industrial Economics: 0.228164, Financial Management and Investment Skills: 0.22073, Company News: 0.203425

[0122] Profile of the person: He is passionate about lottery, sensitive to numbers, and financial information is his new pursuit!

[0123] Recommended titles based on existing interests and exploration interests include:

[0124] 1. Company A's market value evaporated by X overnight, and its third-quarter results fell short of expectations.

[0125] 2. Prices may rise again

[0126] 3. [Financial Management Classroom] Investment Tips

[0127] 4. Three-Minute Trader: 10 Investment Tips

[0128] 5. Comprehensive Stock Investment Skills: Technical Indicator Analysis and Practice

[0129] …

[0130] By recommending relevant content, the activity of the target and the click-through rate of the content can be increased.

[0131] This application can be applied to the construction and recommendation distribution of interests of information flow subjects, recommending relevant content to subjects, improving user experience, breaking the information cocoon of active subjects, and exploring more potential interests for active subjects, allowing active subjects to access more information, allowing active subjects to acquire information and knowledge more comprehensively and effectively. At the same time, for subjects with sparse interests, it can also use a small number of interests to explore the subject's potential interests. At the same time, based on the subject's interests, a brief introduction to the subject can be designed and displayed on the subject's personal page or next to the author's name to increase the fun of interaction.

[0132] It should be noted that any of the methods in this embodiment can be combined based on actual implementation conditions and have corresponding beneficial effects, which will not be described in detail here.

[0133] See also Figure 8 This embodiment further provides a training device for an interest model, which may include:

[0134] A fine-tuning sample construction module 810 is configured to construct a fine-tuning sample based on object association information of a sample object; the object association information includes the existing interests of the sample object; the fine-tuning sample includes the input interests of the sample object and the label interests of the sample object; the input interests are determined based on the existing interests of the sample object; and the label interests are associated with the input interests.

[0135] A training prediction module 820 is configured to predict the interest of the sample object based on the pre-trained language model and the input interest of the sample object to obtain a first predicted interest of the sample object;

[0136] The parameter fine-tuning module 830 is used to fine-tune the model parameters of the pre-trained language model based on the first predicted interest of the sample object and the loss information of the label interest of the sample object to obtain a target interest model for object interest prediction.

[0137] Furthermore, the fine-tuning sample construction module 810 includes:

[0138] a mask processing module, configured to perform random mask processing on the existing interests of the sample object to obtain at least one masked interest;

[0139] A first determining module, configured to determine the at least one masked interest as the tag interest;

[0140] a second determining module, configured to determine an interest other than the label interest among the existing interests of the sample object as the input interest;

[0141] A third determining module is configured to determine the fine-tuning sample based on the label interest and the input interest.

[0142] Furthermore, the fine-tuning sample construction module 810 includes:

[0143] A first generating module is configured to generate interest prediction guidance information based on the interest prediction intention; the interest prediction guidance information is used to guide the generation method of the interest prediction result;

[0144] a model calling module, configured to call a preset universal language model based on the interest prediction guidance information, and perform interest prediction on the sample object based on the preset universal language model and the existing interest of the sample object to obtain a second predicted interest of the sample object;

[0145] a fourth determining module, configured to determine a label interest of the sample object based on the second predicted interest;

[0146] a fifth determining module, configured to determine the existing interest of the sample object as the input interest;

[0147] A sixth determining module is configured to determine the fine-tuning sample based on the label interest and the input interest.

[0148] Furthermore, the first generating module includes:

[0149] a candidate interest set determining module, configured to determine a candidate interest set based on the interest prediction intention; wherein the predicted second predicted interests are all included in the candidate interest set;

[0150] A prediction sample generation module, configured to generate at least one prediction sample; each prediction sample includes an input sample interest and a predicted sample interest;

[0151] The second generating module is configured to generate the interest prediction guidance information based on the candidate interest set and the at least one prediction example.

[0152] Furthermore, the fine-tuning sample further includes label brief information of the sample object, and the label brief information is generated based on text processing of the object association information;

[0153] The device further comprises:

[0154] a profile prediction module, configured to predict a profile of the sample object based on the pre-trained language model and the object association information of the sample object, and obtain predicted profile information of the sample object;

[0155] The parameter fine-tuning module 830 includes:

[0156] A joint fine-tuning module is used to fine-tune the model parameters of the pre-trained language model based on the loss information of the first predicted interest of the sample object and the label interest of the sample object, as well as the label profile information of the sample object and the loss information of the predicted profile information of the sample object to obtain the target interest model.

[0157] Furthermore, the device further comprises:

[0158] An original interest acquisition module, configured to acquire the original interests of the sample object; the original interests include interests of various information types;

[0159] The information integration module is used to perform information integration processing on the interests of the multiple information types to obtain existing interests of the target information type.

[0160] See also Figure 9 This embodiment further provides an interest prediction device, including:

[0161] Interest acquisition module 910, used to acquire the existing interests of the target object;

[0162] The application prediction module 920 is used to predict the interest of the target object based on the target interest model and the existing interest of the target object to obtain the target predicted interest of the target object; the target interest model is trained based on the training method of the above interest model.

[0163] Furthermore, the application prediction module 920 includes:

[0164] A joint prediction module is used to predict the interest of the target object based on the target interest model and the existing interest of the target object, and to predict the profile of the target object based on the target interest model and the existing interest of the target object, so as to obtain the target predicted interest of the target object and the predicted profile information of the target object.

[0165] Furthermore, the device further comprises:

[0166] A recommended content determination module, configured to determine target recommended content corresponding to the target object based on the existing interests of the target object and the target predicted interests of the target object;

[0167] The content push module is used to push the target recommended content to the target object.

[0168] The apparatus provided in the above embodiments can execute the method provided in any embodiment of the present application, and has the functional modules and beneficial effects corresponding to the execution of the method. For technical details not fully described in the above embodiments, please refer to the method provided in any embodiment of the present application.

[0169] This embodiment further provides a computer-readable storage medium, in which at least one instruction or at least one program is stored. The at least one instruction or at least one program is loaded by a processor and executed by any of the above methods of this embodiment.

[0170] According to one aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the above methods.

[0171] Figure 10 This is a block diagram of an electronic device for training an interest model or predicting interest according to an exemplary embodiment. The electronic device may be a server, and its internal structure may be as shown in FIG. Figure 10As shown. The electronic device includes a processor, a memory and a network interface connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the electronic device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, it implements a training method for an interest model, or an interest prediction method.

[0172] Those skilled in the art will understand that Figure 10 The structure shown in the figure is only a block diagram of a part of the structure related to the scheme of the present disclosure, and does not constitute a limitation on the electronic device to which the scheme of the present disclosure is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0173] This specification provides method operation steps as described in the embodiments or flowcharts, but more or fewer operation steps may be included based on routine or non-creative work. The steps and order listed in the embodiments are only one way of executing the order of many steps and do not represent the only execution order. When an actual system or interrupt product is executed, it can be executed sequentially or in parallel according to the method shown in the embodiments or the drawings (for example, in a parallel processor or multi-threaded processing environment).

[0174] The structure shown in this embodiment is only a partial structure related to the scheme of the present application, and does not constitute a limitation on the device to which the scheme of the present application is applied. The specific device may include more or fewer components than shown, or combine certain components, or have different arrangements of components. It should be understood that the methods, devices, etc. disclosed in this embodiment can be implemented in other ways. For example, the device embodiment described above is only schematic. For example, the division of the modules is only a division of logical functions. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or unit modules.

[0175] Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of this application. The aforementioned storage medium includes: a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc., various media that can store program code.

[0176] Those skilled in the art may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this specification can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0177] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for training an interest model, characterized in that: include: Constructing a fine-tuning sample based on object association information of the sample object; the object association information includes the existing interests of the sample object, the fine-tuning sample includes the input interests of the sample object and the label interests of the sample object, the input interests are determined based on the existing interests of the sample object, and the label interests are associated with the input interests; performing interest prediction on the sample object based on a pre-trained language model and the input interest of the sample object to obtain a first predicted interest of the sample object; Fine-tune model parameters of the pre-trained language model based on the first predicted interest of the sample object and the loss information of the label interest of the sample object to obtain a target interest model for object interest prediction.

2. The method according to claim 1, characterized in that The constructing of a fine-tuning sample based on the object association information of the sample object includes: Performing random masking processing on the existing interests of the sample object to obtain at least one masked interest; determining the at least one mask interest as the tag interest; Determining the interests of the sample object other than the label interests as the input interests; The fine-tuning sample is determined based on the tag interest and the input interest.

3. The method according to claim 1, characterized in that The constructing of a fine-tuning sample based on the object association information of the sample object includes: Generate interest prediction guidance information based on the interest prediction intention; the interest prediction guidance information is used to guide the generation method of the interest prediction result; calling a preset general language model based on the interest prediction guidance information, and performing interest prediction on the sample object based on the preset general language model and the existing interest of the sample object to obtain a second predicted interest of the sample object; determining a label interest of the sample object based on the second predicted interest; determining the existing interests of the sample object as the input interests; The fine-tuning sample is determined based on the tag interest and the input interest.

4. The method according to claim 3, characterized in that The generating of interest prediction guidance information based on interest prediction intention includes: Determine a candidate interest set based on the interest prediction intention; the predicted second predicted interests are all included in the candidate interest set; Generate at least one prediction sample; each prediction sample includes an input sample interest and a predicted sample interest; The interest prediction guidance information is generated based on the candidate interest set and the at least one prediction example.

5. The method according to claim 1, characterized in that The fine-tuning sample further includes label brief information of the sample object, wherein the label brief information is generated based on text processing of the object association information; The method further comprises: Performing profile prediction on the sample object based on the pre-trained language model and the object association information of the sample object to obtain predicted profile information of the sample object; Fine-tuning the model parameters of the pre-trained language model based on the loss information of the first predicted interest of the sample object and the label interest of the sample object to obtain a target interest model for object interest prediction includes: Based on the loss information of the first predicted interest of the sample object and the label interest of the sample object, as well as the loss information of the label profile information of the sample object and the predicted profile information of the sample object, the model parameters of the pre-trained language model are fine-tuned to obtain the target interest model.

6. The method according to claim 1, characterized in that The method further comprises: Acquiring original interests of the sample object; the original interests include interests of multiple information types; Information integration processing is performed on the interests of the multiple information types to obtain existing interests of the target information type.

7. An interest prediction method, characterized in that: include: Get the target object's existing interest; Based on the target interest model and the existing interest of the target object, interest prediction is performed on the target object to obtain the target predicted interest of the target object; the target interest model is trained based on the method according to any one of claims 1-6.

8. The method according to claim 7, characterized in that The performing interest prediction on the target object based on the target interest model and the existing interest of the target object to obtain the target predicted interest of the target object includes: Based on the target interest model and the existing interest of the target object, interest prediction is performed on the target object, and based on the target interest model and the existing interest of the target object, a profile prediction is performed on the target object to obtain the target predicted interest of the target object and the predicted profile information of the target object.

9. The method according to claim 7, characterized in that After predicting the target object's interest based on the target interest model and the target object's existing interest to obtain the target predicted interest of the target object, the method further includes: Determining target recommended content corresponding to the target object based on the existing interests of the target object and the target predicted interests of the target object; Push the target recommended content to the target object.

10. A training device for an interest model, characterized in that: include: A fine-tuning sample construction module, used to construct a fine-tuning sample based on object association information of the sample object; The object association information includes the existing interests of the sample object, the fine-tuning sample includes the input interests of the sample object and the label interests of the sample object, the input interests are determined based on the existing interests of the sample object, and the label interests are associated with the input interests; a training prediction module, configured to predict the interest of the sample object based on the pre-trained language model and the input interest of the sample object, to obtain a first predicted interest of the sample object; A parameter fine-tuning module is used to fine-tune the model parameters of the pre-trained language model based on the first predicted interest of the sample object and the loss information of the label interest of the sample object to obtain a target interest model for object interest prediction.

11. An interest prediction device, characterized in that: include: Interest acquisition module, used to obtain the existing interests of the target object; An application prediction module is used to predict the interest of the target object based on a target interest model and the existing interest of the target object to obtain the target predicted interest of the target object; the target interest model is trained based on the method described in any one of claims 1-6.

12. An electronic device, characterized in that: The device includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the training method of the interest model as described in any one of claims 1 to 6, or the interest prediction method as described in any one of claims 7 to 9.

13. A computer storage medium, characterized in that The storage medium stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded by the processor and executed by the training method of the interest model as described in any one of claims 1 to 6, or the interest prediction method as described in any one of claims 7 to 9.

Images