Information recommendation method and device, equipment, storage medium and program product

Abstract

The invention discloses an information recommendation method and device, equipment, a storage medium and a program product, and relates to the technical field of information processing, and the method comprises the steps: determining at least one dimension of reflection data matched with a target user in a reflection data set based on the user information and historical interaction information of the target user; the reflection data set comprises reflection data of multiple dimensions, and each dimension comprises at least one piece of reflection data; each piece of reflection data represents the preference condition of the target user in the dimension to which the reflection data belongs; and processing first prompt information constructed based on the historical interaction information and the reflection data of the at least one dimension through a generative model, so as to determine a target object matched with the target user in the at least one dimension in the recommended candidate object set for recommendation.

Description

Technical Field

[0001] The present application relates to the technical field of information processing, and in particular, to an information recommendation method, apparatus, device, storage medium, and program product. Background Art

[0002] With the rapid development of the Internet industry, online shopping and the way of using the Internet for information dissemination have become very common. Many online platforms (such as shopping platforms, food platforms, reading platforms, etc.) have information recommendation functions to recommend information that users may be interested in to them. However, the accuracy of the current recommendation functions is relatively low and cannot meet the personalized needs of users. Summary of the Invention

[0003] In view of the above problems, the present application provides an information recommendation method, apparatus, device, storage medium, and program product, and the specific solutions are as follows:

[0004] The first aspect of the present application provides an information recommendation method, including:

[0005] Based on the user information and historical interaction information of the target user, determine at least one dimension of reflection data that matches the target user in the reflection data set; the reflection data set includes reflection data of multiple dimensions, each dimension contains at least one piece of reflection data; each piece of reflection data represents the preference situation of the target user in the dimension to which the reflection data belongs;

[0006] Process the first prompt information constructed based on the historical interaction information and the at least one dimension of reflection data through a generative model to determine a target object that matches the target user in the at least one dimension in the recommendation candidate object set for recommendation.

[0007] In a possible implementation, the step of determining at least one dimension of reflection data that matches the target user in the reflection data set based on the user information and historical interaction information of the target user includes:

[0008] Encode the user information and historical interaction information to obtain encoded features;

[0009] Perform dimension prediction on the encoded features through a dimension selection model obtained based on reinforcement learning to obtain at least one dimension that matches the target user;

[0010] For each dimension in the at least one dimension, determine at least one piece of reflection data from at least one piece of reflection data in the dimension in the reflection data set.

[0011] In a possible implementation, the reflection data set includes at least one of the following reflection data sets:

[0012] A user-level reflection dataset, including: reflection data of multiple dimensions corresponding to the target user determined based on the historical interaction information;

[0013] A group-level reflection dataset, wherein the reflection data of each dimension is obtained by screening the reflection data of this dimension corresponding to each user in the user group containing the target user; the users in the user group are similar users;

[0014] A global-level reflection dataset, wherein the reflection data of each dimension is obtained by screening the reflection data of this dimension corresponding to all users.

[0015] In a possible implementation, when the reflection dataset includes three reflection datasets, for each dimension in the at least one dimension, determining at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection dataset includes:

[0016] If there is reflection data of this dimension corresponding to the target user, obtaining at least one piece of reflection data from at least one piece of reflection data of this dimension corresponding to the target user;

[0017] If there is no reflection data of this dimension corresponding to the target user, obtaining at least one piece of reflection data from at least one piece of reflection data of this dimension at the group level;

[0018] If there is no reflection data of this dimension at the group level, obtaining at least one piece of reflection data from at least one piece of reflection data of this dimension at the global level.

[0019] In a possible implementation, determining reflection data of at least one dimension matching the target user in the reflection dataset based on the user information and historical interaction information of the target user includes:

[0020] Constructing a second prompt message based on the user information, historical interaction information, and the reflection dataset; the second prompt message is used to indicate obtaining reflection data of at least one dimension matching the target user in the reflection dataset according to the user information and historical interaction information of the target user;

[0021] Inputting the second prompt message into the generative model to obtain reflection data of at least one dimension matching the target user output by the generative model.

[0022] In a possible implementation, processing a first prompt message constructed based on the historical interaction information and the reflection data of the at least one dimension through a generative model to determine a target object matching the target user in the at least one dimension in the recommended candidate object set, including:

[0023] Add the historical interaction information and the reflection data of the at least one dimension to the first prompt information template to obtain the first prompt information; the first prompt information is used to indicate to determine, with reference to the reflection data of the at least one dimension, a target object in the recommended candidate object set that matches the target user in the at least one dimension;

[0024] Input the first prompt information into the generative model to obtain the target object output by the generative model that matches the target user in the at least one dimension.

[0025] In a possible implementation, the reflection data of multiple dimensions corresponding to the target user is obtained in the following manner:

[0026] Determine the latest reflection data of each dimension based on the historical interaction information;

[0027] For each dimension, calculate the degree of improvement of the recommendation effect of the information recommendation method by using the latest reflection data of this dimension with the validation data set; the validation data set includes the historical interaction information of different users;

[0028] If the degree of improvement is greater than the target value, update the reflection data of the target user corresponding to this dimension based on the latest reflection data of this dimension.

[0029] In a possible implementation, the updating the reflection data of the target user corresponding to this dimension based on the latest reflection data of this dimension includes:

[0030] If the number of the reflection data of the target user corresponding to this dimension is less than the target number, add the latest reflection data of this dimension as the reflection data of the target user corresponding to this dimension;

[0031] If the number of the reflection data of the target user corresponding to this dimension is equal to the target number, screen out the target number of reflection data with the best improvement in the recommendation effect from the latest reflection data of this dimension and the target number of reflection data of the target user corresponding to this dimension; replace the reflection data of the target user corresponding to this dimension with the screened-out target number of reflection data.

[0032] In a possible implementation, each user corresponds to reflection data of multiple dimensions, including:

[0033] Each user corresponds to reflection data of at least two dimensions among the following dimensions;

[0034] Among them, the reflection data of the first dimension corresponding to each user represents an explicit preference determined at least based on the historical interaction information of this user;

[0035] The reflection data representation of the second dimension corresponding to each user represents the implicit preference determined based on the explicit preference of the user;

[0036] The reflection data representation of the third dimension corresponding to each user represents the collaborative preference of the user determined based at least on the historical interaction information of the user and the historical interaction information of each object in the recommended candidate object set.

[0037] A second aspect of the present application provides an information recommendation device, including:

[0038] A reflection matching module, configured to determine, based on the user information and historical interaction information of a target user, at least one dimension of reflection data that matches the target user in a reflection data set; the reflection data set includes reflection data of multiple dimensions, each dimension includes at least one piece of reflection data; each piece of reflection data represents the preference situation of the target user in the dimension to which the reflection data belongs;

[0039] A recommendation module, configured to process a first prompt message constructed based on the historical interaction information and the at least one dimension of reflection data through a generative model, so as to determine a target object that matches the target user in the at least one dimension in a recommended candidate object set for recommendation.

[0040] A third aspect of the present application provides a computer program product, including computer-readable instructions, which when running on an electronic device, enable the electronic device to implement the information recommendation method of the first aspect or any implementation manner of the first aspect.

[0041] A fourth aspect of the present application provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0042] The memory is used to store a computer program;

[0043] The processor is configured to execute the computer program so that the electronic device can implement the information recommendation method of the first aspect or any implementation manner of the first aspect.

[0044] A fifth aspect of the present application provides a computer storage medium, which carries one or more computer programs, and when the one or more computer programs are executed by an electronic device, the electronic device can implement the information recommendation method of the first aspect or any implementation manner of the first aspect. Description of the Drawings

[0045] In combination with the accompanying drawings and with reference to the following specific embodiments, the above and other features, advantages and aspects of the various embodiments of the present disclosure will become more apparent. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic and that the original elements and elements are not necessarily drawn to scale.

[0046] Figure 1 It is a flowchart of an implementation of the information recommendation method provided by this application;

[0047] Figure 2 It is a flowchart of an implementation of determining at least one dimension of reflection data matching the target user in the reflection dataset based on the user information and historical interaction information of the target user provided by this application;

[0048] Figure 3 It is another flowchart of an implementation of determining at least one dimension of reflection data matching the target user in the reflection dataset based on the user information and historical interaction information of the target user provided by this application;

[0049] Figure 4 It is a flowchart of an implementation of obtaining reflection data of multiple dimensions corresponding to the target user provided by this application;

[0050] Figure 5 It is a comparison chart of the recommendation effects between the recommendation method of this application and the existing recommendation methods provided by this application;

[0051] Figure 6 It is a comparison chart of the recommendation effects when the recommendation method of this application uses a dimension selection model and when it does not use a dimension selection model provided by this application;

[0052] Figure 7 It is a comparison chart of the resource occupancy between the recommendation method of this application and the existing recommendation methods based on fine-tuning training provided by this application;

[0053] Figure 8 It is a schematic structural diagram of an information recommendation device provided by this application;

[0054] Figure 9 It is a schematic structural diagram of an electronic device provided by this application. Specific Embodiments

[0055] The following describes the embodiments of the present application in combination with the drawings in the embodiments of the present application. The terms used in the implementation part of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0056] The embodiments of the present application will be described below with reference to the accompanying drawings. As can be known to those of ordinary skill in the art, with the development of technology and the emergence of new scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0057] Terms such as "first" and "second" in the specification, claims and the above-mentioned accompanying drawings of the present application are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such terms can be interchanged under appropriate circumstances, which is only a way of distinguishing when describing objects with the same attributes in the embodiments of the present application. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, so that a process, method, system, product or device including a series of units does not have to be limited to those units, but may include other units not clearly listed or inherent to these processes, methods, products or devices.

[0058] The information recommendation method provided by the embodiments of the present application can be applied to information recommendation in a network platform, and the network platform can include, but is not limited to, any one of the following platforms: e-commerce platform, music platform, video platform, food platform, reading platform, etc.

[0059] Optionally, when the user opens the network platform, the information recommendation method of the present application can be started to recommend information to the user.

[0060] Alternatively, when the user performs a preset interaction on the network platform, the information recommendation method of the present application can be started to recommend information to the user. The preset interaction can include, but is not limited to, any one of the following interactions: performing a purchase interaction, performing a product usage interaction (such as listening to a song, or listening to a playlist, reading a book, reading a recipe, etc.).

[0061] As Figure 1 shown, it is a flowchart of an implementation of the information recommendation method provided by the embodiments of the present application, which can include:

[0062] Step S101: Based on the user information and historical interaction information of the target user, determine at least one (one or more) dimensions of reflection data that match the target user in the reflection data set. Among them, the reflection data set includes reflection data of multiple dimensions, each dimension contains at least one (one or more) reflection data; each reflection data characterizes the preference of the target user in the dimension to which the reflection data belongs.

[0063] The target user can be any user in the application scenario. For example, it can be any user in the e-commerce platform; for example, it can be any user in the music platform; for example, it can be any user in the video platform; for example, it can be any user in the food platform; for example, it can be any user in the reading platform, etc.

[0064] User information may include user profile information, and the profile information may include, but is not limited to, at least some of the following information: the user's gender, age, occupation, region, hobbies or shopping preferences set by the user, etc.

[0065] Historical interaction information may include, but is not limited to, at least one of the following information: the user's browsing records, purchase records, usage records (such as music listening records in a music platform, etc.), search records, etc. In different application scenarios, the historical interaction information may be the same or different.

[0066] The reflection dataset may be determined based on at least one of the following historical interaction information: the historical interaction information of the target user, and the historical interaction information of non-target users having an associated relationship with the target user.

[0067] The essence of matching with the target user is to match with the target user's user information and historical interaction information. The user information and historical interaction information of different users are usually different. Therefore, for different users, the dimensions of the reflection data determined in the reflection dataset may be different. That is to say, when judging the preferences of a user in this application, it is not judged from a fixed dimension, but from multiple dimensions, and the reflection data of the dimension matching the user is selected as the user's preference.

[0068] The dimension matching the target user may be only one dimension or multiple (i.e., at least two) dimensions. For each dimension matching the target user, only one piece of reflection data under that dimension may be obtained, or multiple pieces of reflection data under that dimension may be obtained.

[0069] Step S102: Process the first prompt information constructed based on the historical interaction information and the reflection data of the above at least one dimension through a generative model to determine a target object matching the target user in the above at least one dimension from the set of recommended candidate objects for recommendation.

[0070] The first prompt information constructed based on the historical interaction information of the target user and the reflection data of the above at least one dimension matching the target user may be input into the generative model to obtain a recommendation result generated by the generative model based on the first prompt information. This recommendation result is a target object determined by the generative model from the set of recommended candidate objects and matching the target user in the above at least one dimension.

[0071] The objects in the set of candidate recommended objects are the objects to be recommended to the user on the network platform.

[0072] The generative model can be a large model, such as a large language model (LLM), a multimodal large model (MLLMs), etc.

[0073] The generative model can adopt a model with a Transformer architecture or a model with other architectures, which is not specifically limited in this application.

[0074] The information recommendation method provided in the embodiments of this application combines user information and historical interaction information to screen out at least one dimension of reflection data representing the preferences of the target user that matches the target user from the pre-established reflection data in multiple dimensions, and then through the generative model, combines the historical interaction information of the target user and the at least one dimension of reflection data that matches the target user to screen out objects that match the target user in the above at least one dimension from the set of recommended candidate objects and recommend them to the target user, realizing personalized recommendation and improving the accuracy of the recommendation. Moreover, because at least one dimension of reflection data representing the preferences of the target user is considered during the recommendation, the recommendation result is interpretable.

[0075] In an optional embodiment, a flowchart of an implementation for determining at least one dimension of reflection data that matches the target user in the reflection dataset based on the user information and historical interaction information of the target user is as Figure 2 shown, and it may include:

[0076] Step S201: Process the third prompt information constructed based on the user information and historical interaction information through a generative model to obtain at least one dimension that matches the target user.

[0077] The third prompt information is used to indicate which dimension or dimensions among multiple dimensions the user's preferences tend to be more in line with based on the user information and historical interaction information.

[0078] The user information and historical interaction information can be filled into the third prompt information template to obtain the third prompt information. The third prompt information template includes the identification information of each dimension among multiple dimensions.

[0079] Step S202: For each dimension among the at least one dimension, obtain at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection dataset.

[0080] At least one piece of reflection data can be randomly sampled from at least one piece of reflection data of this dimension in the reflection dataset, or at least one piece of reflection data can be randomly extracted according to a preset rule from at least one piece of reflection data of this dimension in the reflection dataset.

[0081] In an optional embodiment, an implementation manner of determining reflection data of at least one dimension that matches the target user in the reflection data set based on the user information and historical interaction information of the target user may be as follows:

[0082] Construct a second prompt message based on the user information, historical interaction information, and the reflection data set; the second prompt message is used to indicate obtaining reflection data of at least one dimension that matches the target user in the reflection data set according to the user information and historical interaction information of the target user.

[0083] The user information, historical interaction information of the target user, and the reflection data set may be filled into a second prompt message template to obtain the second prompt message.

[0084] Input the second prompt message into a generative model to obtain reflection data of at least one dimension obtained from the reflection data set and matching the target user output by the generative model.

[0085] When the generative model is a large model, due to the large number of parameters and complex calculations of the large model, the processing speed is relatively slow. To improve the information recommendation speed, a small model may be used to determine at least one dimension that matches the target user. Based on this, in an optional embodiment, another implementation flowchart of determining reflection data of at least one dimension that matches the target user in the reflection data set based on the user information and historical interaction information of the target user is as Figure 3 shown and may include:

[0086] Step S301: Encode the user information and historical interaction information of the target user to obtain encoded features.

[0087] Optionally, the user information of the target user may be encoded to obtain a first encoded feature; the historical interaction information of the target user may be encoded to obtain a second encoded feature; the first encoded feature and the second encoded feature are fused to obtain the final encoded feature.

[0088] Optionally, the first encoded feature and the second encoded feature may be concatenated to obtain the final encoded feature. In this case, the lengths of the first encoded feature and the second encoded feature may be the same or different.

[0089] Alternatively, the first encoded feature and the second encoded feature may be averaged to obtain the final encoded feature. In this case, the lengths of the first encoded feature and the second encoded feature are the same.

[0090] A pre-trained text encoder may be used to encode the user information and historical interaction information of the target user respectively.

[0091] Step S302: Perform dimension prediction on the encoded features through a dimension selection model obtained based on reinforcement learning to obtain at least one dimension that matches the target user.

[0092] In the embodiments of the present application, the dimension selection model is a small model with a small number of parameters. This can achieve more efficient recommendation within limited resources and time, reduce the dependence on large-scale data and training resources, improve the adaptability to different types of users, and enhance the recommendation effect. Optionally, the dimension selection model can be a model with the number of parameters less than the target number of parameters.

[0093] The dimension selection model performs prediction processing on the encoded features obtained in step S301, and selects at least one dimension that matches the target user from multiple dimensions. That is, it determines which dimension the user's preference is more inclined to.

[0094] The dimension selection model of the present application can be obtained through reinforcement learning training based on a preset training set and validation set. Among them, the training set is composed of user information and historical interaction information of multiple users, and the validation set is composed of historical interaction information of multiple users; the users in the training set and the validation set are the same, but the corresponding historical interaction data is different; the historical interaction record of each user is composed of interaction information at different consecutive historical moments of the user. Specifically, it can be trained through the following method:

[0095] For the user information and historical interaction information of each user in the training set, encode the user information and historical interaction information of the user to obtain the encoded features corresponding to the user; input the encoded features corresponding to the user into the dimension selection model to obtain the probabilities corresponding to each dimension output by the dimension selection model; the probability corresponding to each dimension represents the matching degree between the dimension and the user, and the greater the probability corresponding to the dimension, the higher the matching degree between the dimension and the user. Sort all dimensions in descending order of probability, and select the first N (N is an integer greater than or equal to 1) dimensions as at least one optimal dimension; score the at least one optimal dimension on the validation set to obtain the score of the at least one optimal dimension; determine the corresponding reward according to the score; feedback the reward to the dimension selection model for caching. After all the data of all users in the training set are calculated once, the dimension selection model updates the parameters of the dimension selection model according to all the rewards in the cache, and then uses the training set for the next round of training until the training end condition is met (for example, the number of iterations reaches the preset number, etc.). The reward is positively correlated with the score.

[0096] One implementation method for scoring at least one optimal dimension corresponding to each user on the validation set can be as follows: For each user, based on the difference between the performance metrics when making information recommendations using the reflection data of at least one optimal dimension and the performance metrics when making information recommendations without using the reflection data of at least one optimal dimension (this difference characterizes the degree of improvement in the recommendation effect of the information recommendation method when using the reflection data of the above at least one optimal dimension compared to not using the reflection data of the above at least one optimal dimension), score at least one optimal dimension. Among them, the performance metrics can include but are not limited to at least one of the following performance metrics: Normalized Discounted Cumulative Gain (NDCG), the hit rate of the recommendation list (i.e., the proportion of the objects selected by the user in the recommendation list among all the objects in the recommendation list). Among them, the larger the NDCG, the better the information recommendation effect. Similarly, the higher the hit rate, the better the information recommendation effect. The score is positively correlated with the difference in the Normalized Discounted Cumulative Gain and positively correlated with the difference in the hit rate.

[0097] As an example, for each user, assuming that the historical interaction information corresponding to this user includes the historical interaction information of this user at t moments, a first prompt message can be constructed based on the historical interaction information of the first t - 1 moments of this user in the validation set and the reflection data of at least one optimal dimension of this user, and the first prompt message is input into the generative model to obtain a recommendation list (for the convenience of description and distinction, denoted as the first recommendation list). The first recommendation list is compared and analyzed with the historical interaction information of the t-th moment of this user, and the performance metrics are calculated (including at least one of the Normalized Discounted Cumulative Gain (NDCG) and the hit rate). For the convenience of description and distinction, the calculated performance metrics can be denoted as the performance metrics with reflection. The performance metrics with reflection can include at least one of the NDCG with reflection and the hit rate with reflection; A fourth prompt message is constructed based on the historical interaction information of the first t - 1 moments of this user in the validation set (obviously, the fourth prompt message does not include the reflection data of the above at least one optimal dimension), and the fourth prompt message is input into the generative model to obtain a recommendation list (for the convenience of description and distinction, denoted as the second recommendation list). The second recommendation list is compared and analyzed with the historical interaction information of the t-th moment of this user, and the performance metrics are calculated (including at least one of the Normalized Discounted Cumulative Gain (NDCG) and the hit rate). For the convenience of description and distinction, the calculated performance metrics can be denoted as the performance metrics without reflection. The performance metrics without reflection can include at least one of the NDCG without reflection and the hit rate without reflection;

[0098] Calculate the difference between the performance metrics with reflection and those without reflection. If the difference indicates that the performance metrics with reflection have improved compared to those without reflection (for example, the difference is greater than zero), it means that the information recommendation performance based on the reflection data has improved; the greater the difference, the more it indicates an improvement (that is, the better the improvement effect or the higher the improvement degree). Correspondingly, the scores corresponding to at least one optimal dimension are higher.

[0099] Step S303: For each of the at least one dimension mentioned above, obtain at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection dataset.

[0100] The reflection data under each dimension represents the specific preference situation of the target user under this dimension.

[0101] It is possible to randomly sample at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection dataset, or it is possible to randomly extract at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection dataset according to a preset rule.

[0102] In an optional embodiment, the reflection dataset includes at least one of the following reflection datasets:

[0103] The reflection dataset at the user level, including: reflection data of multiple dimensions corresponding to the target user determined based on the historical interaction information of the target user. That is to say, a reflection dataset can be established for each user in the network platform respectively, and the reflection dataset corresponding to each user is only determined based on the historical interaction information of this user.

[0104] The reflection dataset at the group level, where the reflection data of each dimension is obtained by screening the reflection data of this dimension corresponding to each user in the user group containing the target user; the users in the user group are similar users. The non-target users having an association relationship with the target user refer to the non-target users belonging to the same user group as the target user.

[0105] Optionally, all users in the network platform can be clustered to group similar users into one group, and a reflection dataset is established for each group of users respectively. As an example, users can be clustered based on the user information of the users.

[0106] The reflection dataset corresponding to each group of users also includes reflection data of multiple dimensions, and each dimension contains at least one piece of reflection data.

[0107] For each group of users, the reflection data of the same dimension corresponding to the group of users can be filtered out from the reflection data of the same dimension corresponding to each user in the group. For each dimension, the reflection data can be filtered based on the improvement effect of the reflection data on the information recommendation performance. The improvement effect of the reflection data on the information recommendation performance can be calculated through the validation dataset. For details, please refer to the foregoing embodiments and will not be elaborated here.

[0108] The reflection dataset at the global level, where the reflection data of each dimension is obtained by filtering the reflection data of this dimension corresponding to all users. That is, for the network platform, a reflection dataset is established. This reflection dataset also includes reflection data of multiple dimensions, and each dimension contains at least one piece of reflection data. The non-target users associated with the target user refer to the non-target users belonging to the same network platform as the target user.

[0109] Optionally, the reflection data of the same dimension corresponding to the network platform can be filtered out from the reflection data of the same dimension corresponding to each user in the network platform. For each dimension, the reflection data can be filtered based on the improvement effect of the reflection data on the information recommendation performance. The improvement effect of the reflection data on the information recommendation performance can be calculated through the validation dataset. For details, please refer to the foregoing embodiments and will not be elaborated here.

[0110] By setting reflection datasets at different levels, a complete rule for determining reflection data can be provided for different degrees of personalized needs. Optionally, in the case where the reflection dataset includes three reflection datasets, for each dimension in the above-mentioned at least one dimension, one implementation manner of determining at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection dataset can be:

[0111] If there is reflection data of this dimension corresponding to the target user, at least one piece of reflection data is obtained from at least one piece of reflection data of this dimension corresponding to the target user. When the target user has a lot of interactions on the network platform and the historical interaction information of the target user is sufficient, the reflection dataset corresponding to the target user can be determined based on the user's historical interaction information. At this time, for each dimension in the above-mentioned at least one dimension, at least one piece of reflection data is preferentially obtained from at least one piece of reflection data of this dimension corresponding to the target user.

[0112] If there is no reflection data of this dimension corresponding to the target user, at least one piece of reflection data is obtained from at least one piece of reflection data of this dimension at the group level.

[0113] In some cases, the interaction of the target user on the network platform is relatively less. For example, the target user has just registered on the network platform, or the interaction of the target user on the network platform is relatively less and the historical interaction information of the target user is relatively less, which is not sufficient to establish a reflection dataset for the target user. In this case, there may be no reflection dataset corresponding to the target user, or there is no reflection data for this dimension corresponding to the target user. At this time, at least one piece of reflection data can be obtained from at least one piece of reflection data for this dimension corresponding to the user group to which the target user belongs.

[0114] If there is no reflection data for this dimension at the group level, at least one piece of reflection data is obtained from at least one piece of reflection data for this dimension at the global level.

[0115] In some cases, there may be no user similar to the target user, that is, the target user is grouped by himself. In this case, there is no reflection dataset corresponding to the user group to which the target user belongs, or there is no reflection data for this dimension corresponding to the user group to which the target user belongs. At this time, at least one piece of reflection data can be obtained from at least one piece of reflection data for this dimension at the global level.

[0116] In an optional embodiment, before performing step S101, reflection data for multiple dimensions corresponding to the target user can be constructed or the reflection data for multiple dimensions corresponding to the target user can be updated.

[0117] Optionally, a flowchart of an implementation for obtaining reflection data for multiple dimensions corresponding to the target user provided in this application is as Figure 4 shown, and may include:

[0118] Step S401: Determine the latest reflection data for each dimension based on the historical interaction information of the target user.

[0119] Optionally, for each dimension, the fifth prompt information corresponding to this dimension constructed based on the historical interaction information of the target user can be processed by a generative model to obtain the latest reflection data for this dimension corresponding to the target user.

[0120] The historical interaction information of the target user can be filled into the fifth prompt information template corresponding to this dimension to obtain the fifth prompt information corresponding to this dimension; the fifth prompt information is input into the generative model to obtain the latest reflection data for this dimension output by the generative model.

[0121] Step S402: For each dimension, use the validation dataset to calculate the degree of improvement of the latest reflection data for this dimension on the recommendation effect of the information recommendation method. The validation dataset includes the historical interaction information of different users.

[0122] The degree of improvement of the recommendation effect of the information recommendation method by the latest reflection data of this dimension can be characterized by the difference between the performance metrics (such as at least one of NDCG and hit rate) when using the latest reflection data of this dimension for information recommendation and the performance metrics of information recommendation without using the latest reflection data of this dimension. For the specific obtaining process, reference can be made to the implementation manner of obtaining the degree of improvement of the recommendation effect of the information recommendation method by the reflection data when the dimension selection model is trained by reinforcement learning as described above, which will not be elaborated here.

[0123] Step S403: If the degree of improvement is greater than the target value, update the reflection data of this dimension corresponding to the target user based on the latest reflection data of this dimension.

[0124] Updating the reflection data of this dimension corresponding to the target user may include, but is not limited to: directly adding the latest reflection data as the reflection data of this dimension corresponding to the target user, or replacing a certain reflection data of this dimension corresponding to the target user with the latest reflection data, or keeping the reflection data of this dimension corresponding to the target user unchanged (that is, discarding the latest reflection data).

[0125] If the degree of improvement is less than or equal to the target value, keep the reflection data of this dimension corresponding to the target user unchanged, that is, do not update the reflection data of this dimension corresponding to the target user.

[0126] In an optional embodiment, when updating the reflection data of this dimension corresponding to the target user based on the latest reflection data of this dimension, the way of updating the reflection data of this dimension corresponding to the target user based on the latest reflection data of this dimension can be determined according to the number of existing reflection data of this dimension corresponding to the target user. Specifically:

[0127] If the number of existing reflection data of this dimension corresponding to the target user is less than the target number, add the latest reflection data of this dimension as the reflection data of this dimension corresponding to the target user.

[0128] If the number of existing reflection data of this dimension corresponding to the target user is equal to the target number, select the target number of reflection data with the best improvement in recommendation effect from the latest reflection data of this dimension and the target number of existing reflection data of this dimension corresponding to the target user; replace the reflection data of this dimension corresponding to the target user with the selected target number of reflection data.

[0129] Among them, the target number of reflection data with the best improvement in recommendation effect may or may not include the latest reflection data.

[0130] By continuously iterating and optimizing the reflection dataset corresponding to the target user, the adaptability and accuracy of information recommendation are enhanced. This optimizable mechanism innovatively realizes a generative model recommendation system that can be optimized without fine-tuning, without fine-tuning costs and can quickly respond to new user needs, avoiding personalized recommendation losses caused by single and fixed reflection dimensions, and further improving the personalization and accuracy of recommendations by maintaining the reflection dataset in multiple dimensions.

[0131] In an optional embodiment, the reflection data corresponding to each user in multiple dimensions may include: the reflection data corresponding to each user in at least two of the following three dimensions.

[0132] Among them, the reflection data of the first dimension corresponding to each user represents the explicit preference determined at least based on the historical interaction information of the user. The explicit preference may be the content of the information input or selected by the user in the historical interaction information, including but not limited to: the name of the object input or selected by the user, attributes (including but not limited to at least one of the following: brand, style, style, function, etc.). For example, if a user searches for "light-colored T-shirts" on an e-commerce platform, the user's explicit preferences are T-shirts and light colors. If the user also searches for brand A, the user's explicit preferences also include brand A.

[0133] As an example, corresponding to the first dimension, a way to obtain the latest reflection data of the first dimension corresponding to the target user by processing the fifth prompt information corresponding to the first dimension constructed based on the historical interaction information of the target user through a generative model can be:

[0134] Fill the user information of the target user, the sequence of objects with which the user has interacted, the set of recommended candidate objects, the sequence of recommended objects predicted at the historical moment, the object selected by the user in the sequence of recommended objects predicted at the historical moment (which may be empty, that is, no object is selected in the sequence of recommended objects predicted at the historical moment), and the search record and selection record of the user into the fifth prompt information template corresponding to the first dimension to obtain the fifth prompt information corresponding to the first dimension. The fifth prompt information corresponding to the first dimension instructs the generative model to generate the explicit preference of the target user.

[0135] Input the fifth prompt information corresponding to the first dimension into the generative model to obtain the latest explicit preference of the target user in the first dimension generated by the generative model.

[0136] The reflection data representation of the second dimension corresponding to each user represents the implicit preferences determined based on the explicit preferences of that user. The implicit preferences can be the user's preference situation summarized according to the user's display preferences, which can be information other than the user's explicit preferences in the historical interaction information, or information not recorded in the historical information (for example, the causal relationship between the objects interacted by the user, that is, why the user interacts with the next object). For example, if a user has searched for "light-colored T-shirts" on an e-commerce platform, and in the user's purchase records, there are 10 records of purchasing T-shirts, among which 8 times white T-shirts are purchased and 2 times light blue T-shirts are purchased, then it can be determined that the user's implicit preference is: white. Another example is that in the user's 10 purchase records, 6 are of brand A, 2 are of brand B, and 2 are of brand C. Since the user has not searched for brand A, it can be determined that the user's implicit preferences also include brand A.

[0137] As an example, for the second dimension, an implementation method of processing the fifth prompt information corresponding to the second dimension constructed based on the historical interaction information of the target user through a generative model to obtain the latest reflection data of the target user corresponding to the second dimension can be:

[0138] Fill the user information of the target user, the sequence of objects interacted by the user, the set of recommended candidate objects, the sequence of recommended objects predicted at historical moments, the object selected by the user from the sequence of recommended objects predicted at historical moments (which may be empty, that is, no object is selected from the sequence of recommended objects predicted at historical moments), and the user's search record and selection record into the fifth prompt information template corresponding to the second dimension to obtain the fifth prompt information corresponding to the second dimension. The fifth prompt information corresponding to the second dimension instructs the generative model to generate the implicit preferences of the target user.

[0139] Input the fifth prompt information corresponding to the second dimension into the generative model to obtain the latest implicit preferences of the target user in the second dimension generated by the generative model.

[0140] The reflection data representation of the third dimension corresponding to each user represents the collaborative preferences of the user determined at least based on the historical interaction information of the user and the historical interaction information of each object in the set of recommended candidate objects.

[0141] Optionally, the collaborative filtering information of the user can be determined based on the historical interaction information of the user and the historical interaction information of each object in the set of recommended candidate objects. The collaborative filtering information includes: the ratings of each object interacted by the user, and the predicted ratings of the user for each object in the set of recommended candidate objects.

[0142] Optionally, a collaborative filtering model can be trained in advance based on the true ratings of each user in the network platform for the objects they have interacted with (objects that the user has purchased or used, which must be objects that the user has searched for and / or browsed). The input of the collaborative filtering model is the user information of each user and the objects that the user has interacted with, and the output of the collaborative filtering model is the predicted scores of the user for each object they have interacted with. With the goal that the predicted scores of each object output by the collaborative filtering model approach the true ratings of each object, the parameters of the collaborative filtering model are updated. That is to say, compared with the collaborative filtering model before parameter update, the scores of each object predicted by the collaborative filtering model after parameter update are closer to the true ratings of the objects.

[0143] Optionally, after training, for any user, if the user has rated the objects they have interacted with, the ratings of the objects the user has interacted with can be directly obtained. If the user has not rated the objects they have interacted with, the user information of the user and the objects the user has interacted with can be input into the collaborative filtering model to obtain the ratings of the user for each object they have interacted with. The user and each object in the recommended candidate object set are input into the collaborative filtering model to obtain the ratings of the user for each object in the recommended candidate object set.

[0144] As an example, for the third dimension, an implementation manner of processing the fifth hint information corresponding to the third dimension constructed based on the historical interaction information of the target user through a generative model to obtain the latest reflection data of the target user corresponding to the third dimension can be:

[0145] The user information of the target user, the sequence of objects they have interacted with, the recommended candidate object set, the sequence of recommended objects predicted at the historical moment, the object selected by the user from the sequence of recommended objects predicted at the historical moment (which may be empty, that is, no object is selected from the sequence of recommended objects predicted at the historical moment), the collaborative filtering information, and the search record and selection record of the user are filled into the fifth hint information template corresponding to the third dimension to obtain the fifth hint information corresponding to the third dimension. The fifth hint information corresponding to the third dimension instructs the generative model to generate the collaborative preference of the target user.

[0146] The fifth hint information corresponding to the third dimension is input into the generative model to obtain the latest explicit preference of the target user in the third dimension generated by the generative model.

[0147] This application provides personalized reflection for users through three dimensions: explicit preference, implicit preference, and user-item collaborative preference, comprehensively capturing the dynamic preferences of users (that is, understanding user needs more comprehensively), and improving the personalization, accuracy, and interpretability of recommendations.

[0148] In an optional embodiment, an implementation manner of processing the first prompt information constructed based on historical interaction information and reflection data in at least one dimension by a generative model to determine a target object that matches the target user in at least one dimension from a set of recommended candidate objects may be as follows:

[0149] Add the historical interaction information of the target user and the reflection data in at least one dimension that matches the target user to the first prompt information template to obtain the first prompt information; the first prompt information is used to indicate referring to the reflection data in at least one dimension to determine a target object that matches the target user of the historical interaction information in at least one dimension from the set of recommended candidate objects.

[0150] Input the first prompt information into the generative model to obtain the target object output by the generative model that matches the target user in the above at least one dimension.

[0151] Next, a comparison and explanation of the recommendation effect of this application will be given.

[0152] As Figure 5 shown, it is a comparison chart of the recommendation effects of the information recommendation method ( Figure 5 denoted as MoRE in

[0153] this application) and existing recommendation methods.

[0154] Among them, there are three categories of existing recommendation methods, namely traditional deep recommendation methods (such as Caser, FDSA, BERT4Rec, GRU4Rec, SASRec, etc.), untrained large language model (LLM)-based recommendation methods (such as LLM4RS, LLMRank(CoT), etc.), and trained large language model (LLM)-based recommendation methods (such as LC-Rec, BinLLM, Re2LLM, etc.).

[0154] The data used for performance comparison are three subsets within the Amazon Review Data (2018) dataset: Amazon Arts ( Figure 5 denoted as Arts in Figure 5 ), Amazon Video Games ( Figure 5 denoted as Games in

[0155] The metrics used for recommendation performance are: HR@5, HR@10, N@5, N@10. Among them, HR@k (k = 5 or 10) is the hit rate of the top-k recommendation list, and N@k (k = 5 or 10) is the normalized discounted cumulative gain (NDCG) of the top-k recommendation list. The higher the value of HR@k, the better the recommendation performance. Similarly, the higher the value of N@k, the better the recommendation performance.

[0156] from Figure 5 It can be seen that the HR@k and N@k of this application are higher than those of the existing recommended solutions, indicating that the recommended solution of this application has improved recommendation performance compared with the existing recommended solutions. In order to more clearly understand the degree of performance improvement of this application, for each indicator in each data set, the optimal value of the indicator of the recommended solution of this application relative to the indicator of each existing recommended solution is calculated here (such as Figure 5 The improvement compared to the underlined value in Figure 5 As shown in the last row of , it can be seen that the improvement of this application compared with the best solution of the existing recommended solutions can reach up to 14.19% and the lowest is 2.13%.

[0157] like Figure 6 As shown, the recommended method provided for this application uses a dimension selection model ( Figure 6 A comparison of the recommendation effects when the dimension selection model is used (denoted as MoRE in the figure) and when the dimension selection model is not used.

[0158] Among them, when the dimension selection model is not used, a random selection method can be used ( Figure 6 Random) or the greedy selection method ( Figure 6 Greedy), or use three dimensions fixedly ( Figure 6 Ref EP Ref IP Ref CF ) (the reflection data of each dimension is fixed).

[0159] The data used for performance comparison is Amazon Arts, Crafts and Sewing.

[0160] The recommended performance indicators are N@5 and N@10.

[0161] Depend on Figure 6 It can be seen that the recommendation performance of the present application is better when the reinforcement learning-based dimension selection model is used than when the reinforcement learning-based dimension selection model is not used.

[0162] like Figure 7 As shown, this is the recommended method for this application ( Figure 7 MoRE (RL)) and existing fine-tuning training-based recommendation methods ( Figure 7 The selected one is the comparison chart of resource usage of LC-Rec (Fine-Tune), BinLLM (Fine-Tune), and Re2LLM (RL).

[0163] The data used for performance comparison are three subsets of the Amazon Review Data (2018) dataset: AmazonArts ( Figure 7 Arts), Amazon Video Games ( Figure 7 Games), Amazon MusicalInstruments ( Figure 7 Instruments).

[0164] The indicators used for resource usage are: Time and GRAM. Time represents the time required for information recommendation, in hours. Figure 7 It is recorded as Time / h; GRAM represents the display memory required for information recommendation, in GB. Figure 7 It is expressed as GRAM / GB. The smaller the value of Time, the shorter the recommendation time, the less resources are occupied, and the lower the training / inference cost. Similarly, the smaller the value of GRAM, the less resources are occupied and the lower the training / inference cost.

[0165] from Figure 7 It can be seen that the resource occupation of this application is less than that of the existing recommended solutions.

[0166] Corresponding to the method embodiment, the present application also provides an information recommendation device. A structural diagram of the information recommendation device provided in the embodiment of the present application is as follows: Figure 8 As shown, it may include:

[0167] Reflection matching module 801 and recommendation module 802;

[0168] The reflection matching module 801 is used to determine, based on the user information and historical interaction information of the target user, at least one dimension of reflection data matching the target user in the reflection data set; the reflection data set includes reflection data of multiple dimensions, each dimension includes at least one reflection data; each reflection data represents the preferences of the target user in the dimension to which the reflection data belongs;

[0169] The recommendation module 802 is used to process the first prompt information constructed based on the historical interaction information and the reflective data of the at least one dimension through a generative model to determine the target object that matches the target user in the at least one dimension in the recommendation candidate object set for recommendation.

[0170] The information recommendation device provided by the embodiment of the present application screens out at least one dimension of reflection data representing the preferences of the target user that matches the target user from the pre-established reflection data in multiple dimensions in combination with the user information and historical interaction information, and then, through a generative model, screens out an object that matches the target user in the above at least one dimension from the recommendation candidate object set in combination with the historical interaction information of the target user and the at least one dimension of reflection data that matches the target user, and recommends it to the target user, realizing personalized recommendation and improving the accuracy of the recommendation. Moreover, because the reflection data representing the preferences of the target user is considered during the recommendation, the recommendation result is interpretable.

[0171] In an optional embodiment, when the reflection matching module 801 determines at least one dimension of reflection data that matches the target user in the reflection data set based on the user information and historical interaction information of the target user, it is used for:

[0172] Encode the user information and historical interaction information to obtain encoded features;

[0173] Perform dimension prediction on the encoded features through a dimension selection model obtained based on reinforcement learning to obtain at least one dimension that matches the target user;

[0174] For each dimension in the at least one dimension, determine at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection data set.

[0175] In an optional embodiment, the reflection data set includes at least one of the following reflection data sets:

[0176] The reflection data set at the user level, including: reflection data of multiple dimensions corresponding to the target user determined based on the historical interaction information;

[0177] The reflection data set at the group level, where the reflection data of each dimension is obtained by screening the reflection data of this dimension corresponding to each user in the user group including the target user; the users in the user group are similar users;

[0178] The reflection data set at the global level, where the reflection data of each dimension is obtained by screening the reflection data of this dimension corresponding to all users.

[0179] In an optional embodiment, when the reflection data set includes three reflection data sets, for each dimension in the at least one dimension, when the reflection matching module 801 determines at least one piece of reflection data from at least one piece of reflection data of this dimension in the reflection data set, it is used for:

[0180] If there is reflection data for this dimension corresponding to the target user, obtain at least one piece of reflection data from at least one piece of reflection data for this dimension corresponding to the target user;

[0181] If there is no reflection data for this dimension corresponding to the target user, obtain at least one piece of reflection data from at least one piece of reflection data for this dimension at the group level;

[0182] If there is no reflection data for this dimension at the group level, obtain at least one piece of reflection data from at least one piece of reflection data for this dimension at the global level.

[0183] In an optional embodiment, when the reflection matching module 801 determines reflection data for at least one dimension that matches the target user in the reflection data set based on the user information and historical interaction information of the target user, it is used for

[0184] Construct a second prompt message based on the user information and historical interaction information, and the reflection data set; the second prompt message is used to indicate obtaining reflection data for at least one dimension that matches the target user in the reflection data set according to the user information and historical interaction information of the target user;

[0185] Input the second prompt message into the generative model to obtain reflection data for at least one dimension that matches the target user output by the generative model.

[0186] In an optional embodiment, when the recommendation module 802 processes the first prompt message constructed based on the historical interaction information and the reflection data for at least one dimension through a generative model to determine a target object that matches the target user in at least one dimension in the recommended candidate object set, it is used for:

[0187] Fill the historical interaction information and the reflection data for at least one dimension into the first prompt message template to obtain the first prompt message; the first prompt message is used to indicate determining a target object that matches the target user of the historical interaction information in at least one dimension in the recommended candidate object set with reference to the reflection data for at least one dimension;

[0188] Input the first prompt message into the generative model to obtain the target object that matches the target user in at least one dimension output by the generative model.

[0189] In an optional embodiment, the information recommendation device of the present application further includes a reflection data maintenance module, which is used to obtain reflection data for multiple dimensions corresponding to the target user:

[0190] Determine the latest reflection data for each dimension based on the historical interaction information;

[0191] For each dimension, use the validation data set to calculate the improvement degree of the recommendation effect of the information recommendation method by the latest reflection data of this dimension; the validation data set includes historical interaction information of different users;

[0192] If the improvement degree is greater than the target value, update the reflection data of the target user corresponding to this dimension based on the latest reflection data of this dimension.

[0193] In an optional embodiment, when the reflection data maintenance module updates the reflection data of the target user corresponding to this dimension based on the latest reflection data of this dimension, it is used for:

[0194] If the number of reflection data of the target user corresponding to this dimension is less than the target number, add the latest reflection data of this dimension as the reflection data of the target user corresponding to this dimension;

[0195] If the number of reflection data of the target user corresponding to this dimension is equal to the target number, among the latest reflection data of this dimension and the target number of reflection data of the target user corresponding to this dimension, screen out the target number of reflection data with the optimal improvement of the recommendation effect; replace the reflection data of the target user corresponding to this dimension with the screened target number of reflection data.

[0196] In an optional embodiment, each user corresponds to reflection data of multiple dimensions, including:

[0197] Each user corresponds to reflection data of at least two of the following dimensions;

[0198] Among them, the reflection data of the first dimension corresponding to each user represents the explicit preference determined at least based on the historical interaction information of this user;

[0199] The reflection data of the second dimension corresponding to each user represents the implicit preference determined based on the explicit preference of this user;

[0200] The reflection data of the third dimension corresponding to each user represents the collaborative preference of this user determined at least based on the historical interaction information of this user and the historical interaction information of each object in the recommended candidate object set.

[0201] An electronic device is also provided in the embodiments of the present application. Refer to Figure 9As shown, it shows a schematic structural diagram of an electronic device suitable for implementing the embodiments of the present application. The electronic device in the embodiments of the present application can be a terminal device (such as a car machine, a large-screen device, a smart home, a mobile phone, a tablet computer, a laptop computer, a desktop computer, etc.), or a server (which can be a single server, a server cluster, or a cloud server, etc.). Figure 9 The electronic device shown is only an example and should not impose any limitations on the functions and usage scope of the embodiments of the present application.

[0202] As Figure 9 shown, the electronic device may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 901, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 902 or the program loaded from the storage device 908 into the random access memory (RAM) 903. When the electronic device is powered on, various programs and data required for the operation of the electronic device are also stored in the RAM 903. The processing device 901, the ROM 902, and the RAM 903 are connected to each other through a bus 904. The input / output (I / O) interface 905 is also connected to the bus 904.

[0203] Generally, the following devices can be connected to the I / O interface 905: an input device 906 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 907 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 908 including, for example, a memory card, a hard disk, etc.; and a communication device 909. The communication device 909 can allow the electronic device to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 9 it shows an electronic device having various devices, it should be understood that it is not required to implement or have all the shown devices. Instead, more or fewer devices can be implemented or had.

[0204] The embodiments of the present application also provide a computer program product including computer-readable instructions. When the computer-readable instructions run on an electronic device, the electronic device is enabled to implement any one of the information recommendation methods provided by the embodiments of the present application.

[0205] The embodiments of the present application also provide a computer-readable storage medium. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can be enabled to implement any one of the information recommendation methods provided by the embodiments of the present application.

[0206] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines.

[0207] It should be understood that in the embodiments of this application, the dependent claims, each embodiment, and features can be combined with each other or coexist, and all can achieve the solution of the foregoing technical problems.

[0208] Through the description of the above embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware. Of course, it can also be implemented by means of dedicated hardware, including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. Generally, functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures used to implement the same function can also be various, such as analog circuits, digital circuits, or dedicated circuits. However, for this application, in more cases, software program implementation is a better implementation method. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disc of a computer, and includes several instructions to enable a computer device (which can be a personal computer, training device, or network device, etc.) to execute the methods described in each embodiment of this application.

[0209] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. Professionals can use different methods to implement the described functions for each specific solution, but such implementation should not be considered to exceed the scope of this application.

[0210] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from a website, a computer, a training device, or a data center to another website, a computer, a training device, or a data center by wire (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be stored by a computer or a data storage device such as a training device or a data center that includes one or more integrated available media. The available medium may be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0211] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0212] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An information recommendation method, comprising: Based on the user information and historical interaction information of a target user, determining at least one dimension of reflection data in a reflection data set that matches the target user; the reflection data set includes reflection data of multiple dimensions, each dimension contains at least one piece of reflection data; each piece of reflection data characterizes the preference situation of the target user in the dimension to which the reflection data belongs; Processing a first prompt message constructed based on the historical interaction information and the at least one dimension of reflection data through a generative model to determine a target object that matches the target user in the at least one dimension in a set of recommended candidate objects for recommendation.

2. The method according to claim 1, wherein the determining at least one dimension of reflection data in the reflection data set that matches the target user based on the user information and historical interaction information of the target user includes: Encoding the user information and historical interaction information to obtain encoded features; Performing dimension prediction on the encoded features through a dimension selection model obtained based on reinforcement learning to obtain at least one dimension that matches the target user; For each dimension in the at least one dimension, determining at least one piece of reflection data from at least one piece of reflection data in the dimension in the reflection data set.

3. The method according to claim 2, wherein the reflection data set includes at least one of the following reflection data sets: User-level reflection dataset, including: Reflection data of multiple dimensions corresponding to the target user determined based on the historical interaction information; A reflection data set at the group level, wherein the reflection data of each dimension is obtained by screening the reflection data of each dimension corresponding to each user in a user group containing the target user; the users in the user group are similar users; A reflection data set at the global level, wherein the reflection data of each dimension is obtained by screening the reflection data of each dimension corresponding to all users.

4. The method according to claim 3, in the case where the reflection data set includes three reflection data sets, for each dimension in the at least one dimension, determining at least one piece of reflection data from at least one piece of reflection data in the dimension in the reflection data set includes: If there is reflection data of the dimension corresponding to the target user, obtaining at least one piece of reflection data from at least one piece of reflection data of the dimension corresponding to the target user; If there is no reflection data of the dimension corresponding to the target user, obtaining at least one piece of reflection data from at least one piece of reflection data of the dimension at the group level; If there is no reflection data of the dimension at the group level, obtaining at least one piece of reflection data from at least one piece of reflection data of the dimension at the global level.

5. The method according to claim 1, wherein the determining at least one dimension of reflection data in the reflection data set that matches the target user based on the user information and historical interaction information of the target user includes: Constructing a second prompt message based on the user information, historical interaction information, and the reflection data set; The second prompt information is used to indicate that at least one dimension of reflection data matching the target user is obtained from the reflection dataset according to the user information and historical interaction information of the target user; The second prompt information is input into the generative model to obtain at least one dimension of reflection data output by the generative model and matching the target user.

6. The method according to claim 1, wherein the generative model processes the first prompt information constructed based on the historical interaction information and the at least one dimension of reflection data to determine a target object in the recommendation candidate object set that matches the target user in the at least one dimension, including: Adding the historical interaction information and the at least one dimension of reflection data to a first prompt information template to obtain the first prompt information; The first prompt information is used to indicate that a target object that matches the target user of the historical interaction information in the at least one dimension is determined in the recommendation candidate object set with reference to the at least one dimension of reflection data; The first prompt information is input into the generative model to obtain a target object output by the generative model and matching the target user in the at least one dimension.

7. The method according to claim 3, wherein the reflection data of multiple dimensions corresponding to the target user is obtained by the following method: Determining the latest reflection data of each dimension based on the historical interaction information; For each dimension, calculating the degree of improvement of the recommendation effect of the information recommendation method by the latest reflection data of this dimension using a validation dataset; the validation dataset includes the historical interaction information of different users; If the degree of improvement is greater than the target value, updating the reflection data of the corresponding dimension of the target user based on the latest reflection data of this dimension.

8. The method according to claim 7, wherein updating the reflection data of the corresponding dimension of the target user based on the latest reflection data of this dimension includes: If the number of reflection data of the corresponding dimension of the target user is less than the target number, adding the latest reflection data of this dimension as the reflection data of the corresponding dimension of the target user; If the number of reflection data of the corresponding dimension of the target user is equal to the target number, screening out the target number of reflection data with the optimal improvement in the recommendation effect from the latest reflection data of this dimension and the target number of reflection data of the corresponding dimension of the target user; replacing the reflection data of the corresponding dimension of the target user with the screened target number of reflection data.

9. The method according to claim 3, wherein each user corresponds to reflection data of multiple dimensions, including: Each user corresponds to reflection data of at least two of the following dimensions; Among them, the reflection data of the first dimension corresponding to each user represents explicit preferences determined at least based on the historical interaction information of this user; The reflection data of the second dimension corresponding to each user represents implicit preferences determined based on the explicit preferences of this user; The reflection data representation in the third dimension corresponding to each user is determined at least based on the historical interaction information of the user and the collaborative preferences of the user determined from the historical interaction information of each object in the set of recommended candidate objects.

10. An information recommendation device, comprising: A reflection matching module, configured to determine at least one dimension of reflection data that matches the target user in a reflection data set based on the user information and historical interaction information of the target user; the reflection data set includes reflection data in multiple dimensions, and each dimension contains at least one piece of reflection data; each piece of reflection data represents the preference of the target user in the dimension to which the reflection data belongs; A recommendation module, configured to process the first prompt information constructed based on the historical interaction information and the at least one dimension of reflection data through a generative model, so as to determine a target object that matches the target user in the at least one dimension in a set of recommended candidate objects for recommendation.

Images