Object recommendation method and device, electronic equipment, storage medium and program product

Abstract

The embodiment of the invention discloses an object recommendation method and device, electronic equipment, a storage medium and a program product, and is used for obtaining a recommended object according to historical interaction data and at least two performance indexes, so that time and resources of a recommendation system are saved. The method comprises the steps of obtaining target information, wherein the target information comprises historical interaction data of using a target application program and a performance index weight sequence; the historical interaction data is obtained according to an interaction operation between the user and the historical recommendation object in at least one historical recommendation period; the performance index weight sequence comprises a weight value corresponding to each performance index in the at least two performance indexes; inputting the target information into a target recommendation model to obtain a recommendation object of the target application program in the current recommendation period; the target recommendation model is obtained by performing model training according to a sample training set; the sample performance index weight sequence comprises a sample weight value corresponding to each sample performance index in the at least two sample performance indexes.

Description

Technical Field

[0001] The present application relates to a recommendation system, and in particular to an object recommendation method, device, electronic device, storage medium and program product. Background Art

[0002] In recent years, a large number of studies have formulated the challenges of offline reinforcement learning (RL) as the problem of autoregressive sequence modeling through decision transformers (DT) or trajectory transformers. Recommender systems have re-examined the challenges of reinforcement learning from the perspective of supervised learning, leveraging the powerful modeling and generalization capabilities of transformer models. They directly map historical interaction data (which can be called states) and recommended objects (which can be called item sequences) to the best results through predictive capabilities, thereby bypassing the traditional complexity of solving RL problems. Some of the latest studies have demonstrated the application of DT in building recommendation systems. However, these studies mainly focus on the scenario of a single performance indicator, so how recommendation systems recommend recommended objects based on multiple performance indicators is an urgent problem to be solved. Summary of the invention

[0003] The embodiments of the present application provide an object recommendation method, device, electronic device, storage medium and program product for obtaining recommended objects based on historical interaction data and at least two performance indicators, thereby saving time and resources of the recommendation system.

[0004] The first aspect of the present application provides an object recommendation method, which may include: obtaining target information, wherein the target information includes historical interaction data of using a target application, and a performance indicator weight sequence; the historical interaction data is obtained according to the interaction operation between the user and the historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes the weight value corresponding to each of at least two performance indicators; the historical recommendation object includes the multimedia resources recommended in the historical recommendation cycle; inputting the target information into a target recommendation model to obtain the recommended object of the target application in the current recommendation cycle, wherein the recommended object of the current recommendation cycle is the recommended object that satisfies the performance indicator weight sequence and is recommended based on the historical interaction data, and the recommended object of the current recommendation cycle includes the multimedia resources recommended in the current recommendation cycle; wherein the target recommendation model is obtained by model training based on a sample training set, wherein the sample training set includes multiple groups of sample data, each group of the sample data includes sample recommendation cycle interaction data and a sample performance indicator weight sequence, and a reference recommendation object corresponding to the sample recommendation cycle; the sample performance indicator weight sequence includes the sample weight value corresponding to each of at least two sample performance indicators.

[0005] In the present technical solution, because the performance indicator weight sequence includes the weight value corresponding to each performance indicator of at least two performance indicators, compared with the solution in the prior art that only focuses on a single performance indicator, that is, "when optimizing one performance indicator and then optimizing another performance indicator, it is often found that the originally optimized performance indicator is deviated; as the number and complexity of performance indicators increase, this cycle becomes more and more cumbersome and requires more time and resources", the embodiment of the present application proposes a method for simultaneously recommending multiple performance indicators, taking into account at least two performance indicators, and the recommended objects obtained in the current recommendation cycle are recommended objects that meet the performance indicator weight sequence based on the historical interaction data, so that recommended objects that meet at least two performance indicators can be recommended, saving time and resources of the recommendation system.

[0006] Optionally, in some embodiments of the present application, the performance indicator weight sequence is pre-set, or is acquired according to a weight sequence setting instruction.

[0007] This technical solution provides an implementation method for obtaining a performance indicator weight sequence, thereby improving the feasibility of the solution.

[0008] Optionally, in some embodiments of the present application, the target information also includes user information using the target application, and each set of sample data also includes sample user information; wherein the recommended object is a recommended object that meets the performance indicator weight sequence based on the historical interaction data and the user information using the target application.

[0009] In this technical solution, the target information and each group of sample data are further explained, and user information is also included, which can improve the reliability of the target recommendation model obtained by model training.

[0010] Optionally, in some embodiments of the present application, the method may further include:

[0011] Obtaining the sample training set;

[0012] Input the sample recommendation period interaction data and the sample performance index weight sequence in each group of sample data into a step transformation module to obtain a transformation vector corresponding to each group of sample data; the transformation vector includes an initial state vector and an initial control vector, the initial state vector is obtained according to the sample recommendation period interaction data, and the initial control vector is obtained according to the sample recommendation period interaction data and the sample performance index weight sequence;

[0013] Inputting the transformation vector corresponding to each group of sample data into a sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data;

[0014] Obtaining the target recommendation model according to the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data;

[0015] Wherein, the initial recommendation model includes the step transformation module and the sequence transformation module.

[0016] In this technical solution, an implementation method is provided for obtaining a target recommendation model, thereby improving the feasibility of generating a target recommendation model.

[0017] Optionally, in some embodiments of the present application, the step transformation module includes a first encoder, a step transformer, a control signal encoder, and a state initializer; the step of inputting the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data into the step transformation module to obtain the transformation vector corresponding to each group of sample data may include:

[0018] Inputting the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data into the first encoder to obtain an interaction data embedding vector corresponding to the sample recommendation period interaction data and a performance indicator embedding vector corresponding to the sample performance indicator weight sequence;

[0019] Inputting the interaction data embedding vector into the state initializer to obtain the initial state vector;

[0020] Input the interaction data embedding vector and the performance indicator embedding vector into the step transformer to obtain the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation;

[0021] The interaction data embedding vector after the transformation in the step and the performance indicator embedding vector after the transformation in the step are input into the control signal encoder to obtain the initial control vector.

[0022] In the present technical solution, based on the sample recommendation cycle interaction data and the sample performance indicator weight sequence in each group of sample data, the initial state vector and initial control vector of each group of sample data are obtained through a series of processing by a step transformation module including a first encoder, a step transformer, a control signal encoder and a state initializer, thereby providing a data basis for subsequently obtaining the calculation recommendation object corresponding to each group of sample data through a sequence transformation module based on the initial state vector and the initial control vector, thereby improving the reliability of generating a target recommendation model.

[0023] Optionally, in some embodiments of the present application, inputting the interactive data embedding vector after the transformation in the step and the performance indicator embedding vector after the transformation in the step into the control signal encoder to obtain the initial control vector may include: splicing the performance indicator embedding vector after the transformation in the step and the interactive data embedding vector after the transformation in the step to obtain a spliced ​​vector; inputting the spliced ​​vector into the control signal encoder to obtain the initial control vector.

[0024] In this technical solution, the performance indicator embedding vector after the transformation in the step and the interactive data embedding vector after the transformation in the step are spliced ​​to obtain a spliced ​​vector; the spliced ​​vector is input into the control signal encoder to obtain the initial control vector. How to obtain the initial control vector is specifically described, which improves the feasibility of the solution.

[0025] Optionally, in some embodiments of the present application, the sequence transformation module includes a sequence transformer and a second encoder; inputting the transformation vector corresponding to each group of sample data into the sequence transformation module to obtain the calculation recommended object corresponding to each group of sample data may include: inputting the initial state vector and the initial control vector into the sequence transformer to obtain the vector after sequence transformation; inputting the vector after sequence transformation into the second encoder to obtain the calculation recommended object corresponding to each group of sample data.

[0026] In this technical solution, the transformation vector corresponding to each group of sample data is input into the sequence transformation module, and the calculated recommended objects corresponding to each group of sample data are specifically detailed. The calculated recommended objects corresponding to each group of sample data can subsequently be judged with the reference recommended objects corresponding to each group of sample data to determine the accuracy of the model training.

[0027] Optionally, in some embodiments of the present application, obtaining the target recommendation model based on the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data may include: using a loss function to obtain the target recommendation model based on the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data.

[0028] In this technical solution, a loss function is used to obtain the target recommendation model based on the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data. In theory, the smaller the value of the loss function, the higher the reliability of the target recommendation model.

[0029] Optionally, in some embodiments of the present application, the sample training set is a data-enhanced sample training set, and the data-enhanced sample training set includes a sample training set without data enhancement and a target sample training set, and the reference recommended objects in the target sample training set are obtained according to the historical recommended objects with which the user has interacted according to preset performance indicators.

[0030] In this technical solution, the sample training set after data enhancement can help the model generalization ability, improve the accuracy and stability of the recommendation model, and deal with the problem of insufficient or unbalanced data.

[0031] Optionally, in some embodiments of the present application, when the preset performance indicator is diversity, the reference recommended object in the target sample training set is the recommended object among the historical recommended objects with which the user has interacted, and has the smallest overlap with the reference recommended object in the sample training set without data enhancement; or, when the preset performance indicator is the cumulative score, the reference recommended object in the target sample training set is the recommended object among the historical recommended objects with which the user has interacted, excluding the reference recommended objects in the sample training set without data enhancement, and has the highest cumulative score.

[0032] In this technical solution, by explaining the data enhancement for different preset performance indicators, the DT model vision can be broadened, allowing the learned recommendation strategy to select more appropriate recommendation objects.

[0033] Optionally, in some embodiments of the present application, the method may further include: in response to a view trigger instruction in the current recommendation cycle, displaying the recommended objects of the target application in the current recommendation cycle through a display interface.

[0034] In the present technical solution, after obtaining the recommended objects of the target application in the current recommendation cycle, in response to a view trigger instruction in the current recommendation cycle, the recommended objects of the target application in the current recommendation cycle are displayed through a display interface, so that the current user can view, collect, forward, like, share, download, etc. according to his own needs, thereby meeting the different needs of users.

[0035] In a second aspect of the present application, an object recommendation device is provided, which may include:

[0036] An acquisition module is used to acquire target information, wherein the target information includes historical interaction data of using a target application and a performance indicator weight sequence; the historical interaction data is obtained based on interaction operations between a user and a historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes a weight value corresponding to each of at least two performance indicators; the historical recommendation object includes a multimedia resource recommended in the historical recommendation cycle;

[0037] A module is obtained, which is used to input the target information into a preset recommendation model to obtain the recommended objects of the target application in the current recommendation cycle, wherein the recommended objects of the current recommendation cycle are recommended objects that meet the performance indicator weight sequence based on the historical interaction data, and the recommended objects of the current recommendation cycle include multimedia resources recommended in the current recommendation cycle; wherein the target recommendation model is obtained by model training based on a sample training set, and the sample training set includes multiple groups of sample data, each group of the sample data includes sample recommendation cycle interaction data and a sample performance indicator weight sequence, and a reference recommended object corresponding to the sample recommendation cycle; the sample performance indicator weight sequence includes a sample weight value corresponding to each sample performance indicator of at least two sample performance indicators.

[0038] In a third aspect of the present application, an electronic device is provided, comprising a memory, one or more processors, and a display, wherein the memory is coupled to the processor, the memory is used to store computer program code, the computer program code comprises computer instructions, and when the processor executes the computer instructions, the electronic device executes any one of the methods described in the first aspect.

[0039] In a fourth aspect of the present application, a chip is provided, which includes a processor, a memory, and a display. The processor may be a logic circuit, an integrated circuit, or a general-purpose processor, etc. The memory stores instructions. The processor may implement the method in the above-mentioned first aspect by reading the software code stored in the memory. The memory may be integrated in the processor, or may be located outside the processor and exist independently.

[0040] In a fifth aspect of the present application, a chip system is provided, which is applied to an electronic device, the chip system comprising one or more interface circuits and one or more processors, and a display, the interface circuit and the processor are interconnected via a line, the interface circuit is used to receive a signal from a memory of the terminal device and send the signal to the processor, the signal comprising a computer instruction stored in the memory, and when the processor executes the computer instruction, the electronic device executes any one of the methods described in the first aspect.

[0041] In a sixth aspect of the present application, a computer program product is provided, the computer program product comprising: a computer program (also referred to as code, or instruction), which enables a computer to execute the method in the first aspect when the computer program is executed.

[0042] In a seventh aspect of the present application, a computer-readable storage medium is provided, which stores a computer program (also referred to as code or instruction) which, when executed on a computer, enables the computer to execute the method in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments and the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained based on these drawings.

[0044] Figure 1 A schematic diagram of an embodiment of the object recommendation method in the embodiment of the present application;

[0045] Figure 2 This is another schematic diagram of an embodiment of the object recommendation method in the embodiment of the present application;

[0046] Figure 3A This is a schematic diagram of an embodiment of the initial recommendation model in the embodiments of the present application;

[0047] Figure 3B This is another schematic diagram of an initial recommendation model in an embodiment of the present application;

[0048] Figure 3C This is another schematic diagram of an initial recommendation model in an embodiment of the present application;

[0049] Figure 3D A schematic diagram of the initial recommendation model in the embodiment of the present application;

[0050] Figure 4A A schematic diagram of the influence of the weight values ​​of multiple performance indicators predefined in the MocDT model framework on the evaluation performance indicators in an embodiment of the present application;

[0051] Figure 4B A schematic diagram of the impact of different enhancement strategies on recommendation performance in an embodiment of the present application;

[0052] Figure 4C This is a schematic diagram of the impact of different enhancement rates on recommended performance in an embodiment of the present application;

[0053] Figure 5 This is a schematic diagram of an embodiment of an object recommendation device in an embodiment of the present application;

[0054] Figure 6 This is a schematic diagram of an embodiment of an electronic device in an embodiment of the present application;

[0055] Figure 7 This is a schematic diagram of another embodiment of the electronic device in the embodiment of the present application. DETAILED DESCRIPTION

[0056] The embodiments of the present application provide an object recommendation method, device, electronic device, storage medium and program product for obtaining recommended objects based on historical interaction data and at least two recommendation strategies, thereby saving time and resources of the recommendation system.

[0057] In order to make the technical personnel in the technical field better understand the scheme of the present application, the technical scheme in the embodiment of the present application will be described below in conjunction with the drawings in the embodiment of the present application. Obviously, the described embodiment is only a part of the embodiment of the present application, not all of the embodiments. Based on the embodiments in the present application, they should all fall within the scope of protection of the present application.

[0058] Recommender systems facilitate smooth interactions between users and digital platforms by continuously adjusting recommendation strategies to reflect the evolution of user preferences. The core goal is to build a series of recommendations that meet users' immediate needs while promoting long-term and sustained engagement, thereby improving user retention. However, optimizing long-term user engagement faces significant challenges, primarily because it requires predicting and measuring the impact of recommendation strategies on future outcomes. As an alternative, these recommendation systems use immediate feedback signals (also known as immediate performance indicators, such as clicks, purchases, and ratings) as proxy indicators for optimizing future engagement.

[0059] Although this approach is feasible in practice, it also faces severe challenges. The most critical challenge is to quantify the impact of current behavior on the future, which increases the complexity of optimizing the immediate performance indicator. In addition, the need to coordinate between multiple often conflicting performance indicators (such as accuracy and diversity) further increases the complexity. For example, a recommendation strategy that focuses too much on recommending objects that users have previously preferred to improve immediate performance indicators may cause users to feel bored, which in turn leads to a gradual decline in user engagement, highlighting the complex balance required by recommendation strategies. As a result, existing recommendation methods often find themselves caught in a continuous cycle of training and evaluation: after optimizing one performance indicator, when optimizing another performance indicator, they often find that they have deviated from the originally optimized performance indicator. As the number and complexity of performance indicators increase, this cycle becomes more and more cumbersome and requires more time and resources.

[0060] In recent years, a large number of studies have formulated the challenges of offline reinforcement learning (RL) as an autoregressive sequence modeling problem through decision transformers (DT) or trajectory transformers. With the powerful modeling and generalization capabilities of transformer models, these studies have revisited the challenges of reinforcement learning from the perspective of supervised learning, directly mapping state and action sequences to optimal results through prediction capabilities, thereby bypassing the traditional complexity of solving RL problems. Some recent studies have demonstrated the application of DT in building recommendation systems. However, these studies mainly focus on single performance indicator scenarios, ignoring the complex multi-performance indicator dynamics inherent in the recommendation environment, and over-rely on immediate performance indicators (such as Normalized Discounted Cumulative Gain (NDCG) and Recall) when evaluating these models, and fail to fully explore the recommendation strategy planning capabilities.

[0061] In an embodiment of the present application, an object recommendation method is proposed, by obtaining target information, the target information includes historical interaction data using a target application, and a performance indicator weight sequence; the historical interaction data is obtained by interactive operations between a user and a historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes a weight value corresponding to each of at least two performance indicators; the target information is input into a preset recommendation model to obtain a recommended object for the current recommendation cycle of the target application. Because the performance indicator weight sequence includes a weight value corresponding to each of at least two performance indicators, compared to the prior art solution that only focuses on a single performance indicator, the embodiment of the present application takes into account at least two performance indicators, so that recommended objects that meet at least two performance indicators can be recommended, saving time and resources of the recommendation system.

[0062] The technical solution of the present application is further described below by way of embodiments. Figure 1 As shown, it is a schematic diagram of an embodiment of the object recommendation method in the embodiment of the present application, which may include:

[0063] 101. Obtain target information, where the target information includes historical interaction data using a target application and a performance indicator weight sequence.

[0064] The historical interaction data is obtained based on the interaction operations between the user and the historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes the weight value corresponding to each performance indicator of at least two performance indicators.

[0065] The historical recommendation object includes multimedia resources recommended in the historical period, which may be multimedia resources about items (also called commodities) (for example, pictures about an item, and / or multimedia resources such as videos), or multimedia works published by other users, which are not limited in the specific embodiments of the present application. For example, multimedia resources may include at least one of images, videos, audio and text.

[0066] Optionally, in some possible implementations, the target information also includes information about a user using the target application.

[0067] User information may include user identification (ID, also known as user account), mobile phone number, login password, login device, gender, age, education, occupation, region, etc.

[0068] The historical interaction data may include: historical recommended objects of at least one historical recommendation cycle, and interaction information between users and historical recommended objects, such as satisfaction, click-through rate, accuracy, etc.

[0069] The performance indicator weight sequence of each recommendation cycle can be considered as a recommendation strategy, that is, the recommendation strategy includes weight values ​​corresponding to multiple performance indicators. Different recommendation strategies may focus on different performance indicators, which can be represented by corresponding weight values. For example, if the weight value of performance indicator A is the largest, it can be said that the recommendation strategy focuses on performance indicator A. The weight value corresponding to each performance indicator can represent the priority of the corresponding performance indicator.

[0070] It should be noted that the weight value corresponding to each recommended performance indicator can be a specific value or a range of values, which is not specifically limited in the embodiments of the present application.

[0071] The weight value corresponding to each performance indicator in at least two performance indicators may be the same or different. For example: Under normal circumstances, the larger the weight value, the higher the priority of the corresponding performance indicator, and the smaller the weight value, the lower the priority of the corresponding performance indicator; or, in some cases, the smaller the weight value, the higher the priority of the corresponding performance indicator, and the larger the weight value, the lower the priority of the corresponding performance indicator. The embodiment of the present application is explained by taking the example that the larger the weight value, the higher the priority of the corresponding performance indicator. In the present technical solution, it is more intuitive and accurate to characterize the priority of the performance indicator by the weight value.

[0072] Optionally, in some possible implementations, the performance indicator weight sequence is pre-set, or is acquired according to a weight sequence setting instruction.

[0073] In the present technical solution, the performance indicator weight sequence may be pre-set, or may be obtained according to a weight sequence setting instruction, or may be obtained by modifying the obtained performance indicator weight sequence, and the embodiments of the present application do not make specific limitations. Among them, the obtained performance indicator weight sequence is modified, and specifically, the target weight value corresponding to the target performance indicator in the performance indicator weight sequence is modified, without the need to retrain the target recommendation model. The present technical solution provides an implementation method for obtaining a performance indicator weight sequence, which improves the feasibility of the solution.

[0074] Optionally, in some possible implementations, the performance indicator weight sequence of the current recommendation cycle is the same as or different from the performance indicator weight sequence of the previous recommendation cycle.

[0075] The performance indicator weight sequence of the current recommendation cycle may be the same as or different from the performance indicator weight sequence of the previous recommendation cycle; if the performance indicator weight sequence of the current recommendation cycle is different from the performance indicator weight sequence of the previous recommendation cycle, it means that there is a different weight value corresponding to at least one performance indicator in the performance indicator weight sequence of the current recommendation cycle and the performance indicator weight sequence of the previous recommendation cycle.

[0076] In this technical solution, the performance indicator weight sequence of the current recommendation cycle is the same as or different from the performance indicator weight sequence of the previous recommendation cycle. If they are different, it means that the performance indicator weight sequence can be adjusted in time according to historical interaction data, so as to present recommended objects that meet different performance indicators to users.

[0077] Optionally, in some possible implementations, the method may further include: in a case where a modification instruction for the performance indicator weight sequence is detected, modifying the performance indicator weight sequence to obtain a modified performance indicator weight sequence;

[0078] Inputting the target information into the target recommendation model to obtain the recommended objects of the current recommendation cycle of the target application may include: inputting the historical interaction data of the target application and the modified performance indicator weight sequence into the target recommendation model to obtain the recommended objects of the current recommendation cycle of the target application; or,

[0079] The user information, historical interaction data, and modified performance indicator weight sequence of the target application are input into the target recommendation model to obtain the recommended objects of the current recommendation cycle of the target application.

[0080] Exemplarily, the performance indicator weight sequence includes weight values ​​corresponding to at least two performance indicators, for example, in the previous recommendation cycle of the current recommendation cycle, the at least two performance indicators include performance indicator A, performance indicator B and performance indicator C, the weight value corresponding to performance indicator A is A1, the weight value corresponding to performance indicator B is B1, and the weight value corresponding to performance indicator C is C1. If the weight value A1 is greater than the weight value B1, and the weight value B1 is greater than the weight value C1, it indicates that the priority of performance indicator A is greater than the priority of performance indicator B, and the priority of performance indicator B is greater than the priority of performance indicator C.

[0081] For example, in the current recommendation cycle, at least two performance indicators include performance indicator A, whose corresponding weight value is A1, performance indicator B, whose corresponding weight value is B1, and performance indicator C, whose corresponding weight value is C1. At this time, the requirement for performance indicator B is relatively high, and the weight value corresponding to performance indicator B can be updated from B1 to B2. At this time, if the weight value B2 is greater than the weight value A1, and the weight value A1 is greater than the weight value C1, it indicates that the priority of performance indicator B is greater than that of performance indicator A, and the priority of performance indicator A is greater than that of performance indicator C.

[0082] In the present technical solution, the performance indicator weight sequence can be modified; wherein, the performance indicator weight sequence can be modified, specifically, the target weight value corresponding to the target performance indicator in the performance indicator weight sequence can be modified. The priority of different performance indicators can be adjusted by adjusting the weight values ​​corresponding to different performance indicators. Thus, the historical interaction data of the target application and the modified performance indicator weight sequence are input into the target recommendation model to obtain the recommended objects of the current recommendation cycle of the target application; or, the user information, historical interaction data, and the modified performance indicator weight sequence of the target application are input into the target recommendation model to obtain the recommended objects of the current recommendation cycle of the target application. Therefore, the corresponding weight values ​​can be updated according to different performance indicators or the requirements of different performance indicators, without the need to retrain the recommendation model, thereby greatly improving the efficiency of obtaining recommended objects.

[0083] Optionally, in some possible implementations, performance indicators may include but are not limited to diversity, accuracy, precision, recall, harmonic mean of precision and recall, average accuracy, normalized discounted cumulative gain, click-through rate, conversion rate, coverage, novelty, surprise, user satisfaction, retention rate, prediction error, time delay, stability, robustness, and interpretability.

[0084] In the recommendation system, common indicators for evaluating the performance of the recommendation model may include but are not limited to the following:

[0085] Accuracy: measures the proportion of user behaviors correctly predicted by the recommendation system.

[0086] Precision: Measures the proportion of items recommended by the recommendation system that are actually accepted by users.

[0087] Recall: It measures how much of all relevant items the recommendation system can retrieve.

[0088] F1 Score: The harmonic mean of precision and recall, used to measure the overall performance of the recommendation system.

[0089] Mean Average Precision (MAP): measures the average accuracy of each recommended item in the recommendation list.

[0090] Normalized Discounted Cumulative Gain (NDCG): This measures the quality of recommendations by taking into account the relevance and position of items in the recommendation list.

[0091] Click-Through Rate (CTR): measures the user's click behavior on the recommended results.

[0092] Conversion Rate: Measures the proportion of users who make a purchase or other targeted actions after clicking on a recommendation.

[0093] Coverage: Measures the diversity of the types or quantities of items that a recommendation system can recommend.

[0094] Diversity: Measures the diversity of items in the recommendation list.

[0095] Novelty: measures the ability of a recommender system to recommend items that are previously unknown or uncommon to the user.

[0096] Serendipity: measures the ability of a recommendation system to recommend items that users do not expect but ultimately find interesting.

[0097] User Satisfaction: Measure the user's satisfaction with the recommendation results through surveys or feedback.

[0098] Retention Rate: Measures the user's continued use of the recommendation system over a period of time.

[0099] Prediction Error: measures the accuracy of the recommendation system in predicting user behavior.

[0100] Latency: measures the speed at which a recommendation system generates recommendation results.

[0101] Stability: Measures the consistency of the recommendation system over time or across different user groups.

[0102] Robustness: Measures the ability of a recommender system to resist manipulated or noisy data.

[0103] Explainability: measures the ability of a recommendation system to provide reasons for its recommendations.

[0104] In this technical solution, a specific description is given for different performance indicators, which improves the feasibility of the solution.

[0105] 102. Input the target information into the target recommendation model to obtain the recommended objects of the target application in the current recommendation cycle.

[0106] The recommended objects of the current recommendation cycle are recommended objects that meet the performance indicator weight sequence based on the historical interaction data, and the recommended objects of the current recommendation cycle include multimedia resources recommended in the current recommendation cycle.

[0107] The target recommendation model is obtained by model training based on the sample training set. The sample training set includes multiple groups of sample data. Each group of sample data includes sample recommendation period interaction data and sample performance index weight sequence, as well as reference recommendation objects corresponding to the sample recommendation period. The sample performance index weight sequence includes the sample weight value corresponding to each sample performance index in at least two sample performance indicators. The sample recommendation period interaction data is obtained based on the interaction operation between the user and the reference recommendation object in at least one sample recommendation period.

[0108] Exemplarily, the recommended objects of the current recommendation cycle include multimedia resources recommended in the current recommendation cycle, which may be multimedia resources about items (also referred to as commodities) (e.g., pictures about an item, and / or multimedia resources such as videos), or multimedia works published by other users, which are not limited in the specific embodiments of the present application. For example, multimedia resources may include at least one of images, videos, audios, and texts.

[0109] Optionally, in some possible implementations, each group of sample data also includes sample user information; that is, the target recommendation model is obtained by performing model training based on the sample training set, the sample training set includes multiple groups of sample data, each group of sample data includes sample user information, sample recommendation cycle interaction data and sample performance indicator weight sequence, and reference recommendation objects corresponding to the sample recommendation cycle; and the sample user information is user information using the target application.

[0110] That is, the target information also includes user information using the target application, and each set of sample data also includes sample user information; wherein the recommended object of the current recommendation cycle is the recommended object that meets the performance indicator weight sequence based on the historical interaction data and the user information using the target application. For example, if the user information using the target application is a lawyer, then the recommended object of the current recommendation cycle is likely to be a book or multimedia work related to law.

[0111] In this technical solution, the target recommendation model is obtained by model training, and the characteristics of the sample user information are also considered. Then, the target information input into the target recommendation model may also include the user information of the target application. Thus, the recommended object of the current recommendation cycle is the recommended object that meets the performance indicator weight sequence based on the historical interaction data and the user information of the target application, which can better meet the needs of users.

[0112] Optionally, in some possible implementations, the method may further include: in response to a viewing trigger instruction in the current recommendation cycle, displaying the recommended objects of the target application in the current recommendation cycle through a display interface.

[0113] Exemplarily, after obtaining the recommended objects of the target application in the current recommendation cycle, if the current user happens to need to view them, then the electronic device can display the recommended objects of the target application in the current recommendation cycle through the display interface in response to a view trigger instruction in the current recommendation cycle, thereby displaying the recommended objects of the current recommendation cycle to the current user for the current user to view, collect, forward, like, share, download, and other operations.

[0114] In the present technical solution, after obtaining the recommended objects of the target application in the current recommendation cycle, in response to a view trigger instruction in the current recommendation cycle, the recommended objects of the target application in the current recommendation cycle are displayed through a display interface, so that the current user can view, collect, forward, like, share, download, etc. according to his or her own needs, thereby meeting the different needs of users.

[0115] In the embodiment of the present application, because the performance indicator weight sequence includes the weight value corresponding to each performance indicator of at least two performance indicators, compared with the solution in the prior art that only focuses on a single performance indicator, that is, "when optimizing one performance indicator and then optimizing another performance indicator, it is often found that the originally optimized performance indicator is deviated; as the number and complexity of performance indicators increase, this cycle becomes more and more cumbersome and requires more time and resources." In the embodiment of the present application, a method for making recommendations for multiple performance indicators at the same time is proposed, taking into account at least two performance indicators, and the recommended objects obtained in the current recommendation cycle are recommended objects that meet the performance indicator weight sequence based on the historical interaction data, so that recommended objects that meet at least two performance indicators can be recommended, saving time and resources of the recommendation system.

[0116] like Figure 2 As shown, it is a schematic diagram of another embodiment of the object recommendation method in the embodiment of the present application, which may include:

[0117] 201. Obtain a sample training set.

[0118] The sample training set includes: multiple groups of sample data, each group of sample data includes sample recommendation period interaction data and sample performance index weight sequence, and reference recommendation object corresponding to the sample recommendation period; the sample performance index weight sequence includes sample weight values ​​corresponding to each sample performance index in at least two sample performance indicators. The sample recommendation period interaction data is obtained according to the interaction operation between the user and the reference recommendation object in at least one sample recommendation period.

[0119] Optionally, in some possible implementations, each set of sample data also includes sample user information, and the sample user information is information about users who use the target application.

[0120] Exemplarily, the sample user information may include the sample user identification (Identity, ID, also referred to as user account), mobile phone number, login password, login device, gender, age, education level, occupation, region and other information.

[0121] The sample recommendation cycle interaction data may include: recommended objects of at least one sample recommendation cycle, and interaction information between users and reference recommended objects, such as satisfaction, click-through rate, accuracy, etc.

[0122] The sample performance indicator weight sequence of each sample recommendation cycle can be considered as a sample recommendation strategy, that is, the sample recommendation strategy includes weight values ​​corresponding to multiple sample performance indicators. Different sample recommendation strategies may focus on different sample performance indicators, which can be represented by corresponding weight values. For example, the weight value of sample performance indicator A is the largest, which can indicate that the sample recommendation strategy focuses on sample performance indicator A. The weight value corresponding to each sample performance indicator can represent the priority of the corresponding sample performance indicator.

[0123] The weight value corresponding to each sample performance indicator in at least two sample performance indicators may be the same or different. For example: Under normal circumstances, the larger the weight value, the higher the priority of the corresponding sample performance indicator, and the smaller the weight value, the lower the priority of the corresponding sample performance indicator; or, in some cases, the smaller the weight value, the higher the priority of the corresponding sample performance indicator, and the larger the weight value, the lower the priority of the corresponding sample performance indicator. The embodiment of the present application is explained by taking the example that the larger the weight value, the higher the priority of the corresponding sample performance indicator. In the present technical solution, it is more intuitive and accurate to characterize the priority of the sample performance indicator by the weight value.

[0124] 202. According to the sample training set, the initial recommendation model is trained to obtain a target recommendation model.

[0125] In this technical solution, a sample training set is obtained; based on the sample training set, the initial recommendation model is trained to obtain a target recommendation model; an implementation method for performing model training to obtain a target recommendation model is provided, which improves the reliability of generating the recommendation model, so that when the target recommendation model is subsequently used to output the corresponding recommendation object, the accuracy is higher.

[0126] The following first introduces the adaptation of the multi-performance indicator decision transformer (DT) in this application for this specific task, and then discusses in detail the method of the multi-performance indicator controllable decision transformer (MocDT) proposed in this application.

[0127] Model the historical interaction data as state s t , the output recommendation object is regarded as action a t , some instantaneous performance indicators (such as click-through rate or rating, etc.) are regarded as cumulative ratings r(s t , a t ), through this modeling approach, the recommendation process is transformed into a decision problem. Considering the overall level performance indicators related to the future trajectory, a sliding window can be used to capture the future trajectory, and this sliding window changes continuously during the recommendation process. Formally, the H-step future sequence after timestamp t is expressed as Here, each future sequence has multiple performance indicators corresponding to it, for example: Future sequence, multiple performance indicators are

[0128] It should be noted that in the training process of the target recommendation model, the calculation of the weight values ​​corresponding to these performance indicators depends on different functions. Here we will give examples of two common performance indicators in the recommendation setting: cumulative score and diversity. Among them, the accuracy, precision, recall rate, average accuracy, click-through rate, conversion rate, coverage, user satisfaction, etc. mentioned above can be collectively referred to as scores.

[0129] (1) Cumulative score

[0130] r(s t , a t ) indicates the user's preference for the recommended object a t The cumulative score is calculated as follows:

[0131]

[0132] In formula 1, t represents the current timestamp, which can be understood as the current recommendation period, and t′=t+1 represents a time step after the current timestamp t, which can be understood as the next recommendation period. Represents the cumulative score for one time step after the current timestamp t.

[0133] (2) Diversity

[0134] Jaccard index, also known as Jaccard similarity coefficient, is used to compare the similarities and differences between finite sample sets. The larger the Jaccard coefficient value, the higher the sample similarity.

[0135] The ratio of the number of elements in the intersection of two sets A and B to the number of elements in the union of A and B is called the Jaccard similarity coefficient of the two sets, represented by the user symbol J(A, B). The Jaccard similarity coefficient is an indicator to measure the similarity between two sets.

[0136]

[0137] If J(A, B) is used to represent the Jaccard similarity of the categories of two item sets A and B, then the average diversity can be quantified as follows:

[0138]

[0139] In formula 3, t represents the current timestamp, t′=t+1 represents a time step after the current timestamp t, Indicates the similarity between the current timestamp t and the recommended object corresponding to a time step after the current timestamp. Indicates the diversity of recommended objects corresponding to the current timestamp t and a time step after the current timestamp t.

[0140] The following is a detailed description of how to train the initial recommendation model according to the sample training set to obtain the target recommendation model, as shown below:

[0141] 1. Each set of sample data in the sample training set includes: sample recommendation period interaction data and sample performance indicator weight sequence.

[0142] like Figure 3A As shown, it is a schematic diagram of an embodiment of the initial recommendation model in the embodiment of the present application. The initial recommendation model includes a step transformation module and a sequence transformation model module. Optionally, in some embodiments of the present application, the initial recommendation model is trained according to the sample training set to obtain a target recommendation model, which may include: inputting the sample recommendation cycle interaction data and the sample performance index weight sequence in each group of sample data into the step transformation module to obtain a transformation vector corresponding to each group of sample data; the transformation vector includes an initial state vector and an initial control vector, the initial state vector is obtained according to the sample recommendation cycle interaction data, and the initial control vector is obtained according to the sample recommendation cycle interaction data and the sample performance index weight sequence; the transformation vector corresponding to each group of sample data is input into the sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data; according to the calculation recommendation object corresponding to each group of sample data and the reference recommendation object corresponding to each group of sample data, the target recommendation model is obtained. The steps in this implementation method can directly replace step 202.

[0143] In this technical solution, the sample recommendation period interaction data (Historicalitem features) and the sample performance index weight sequence (O1, O2...) in each group of sample data can be input into the step transformation module for transformation, and the transformation vector (including the initial state vector and the initial control vector ); then the transformation vector corresponding to each group of sample data is input into the sequence transformation module to obtain the calculated recommendation object corresponding to each group of sample data, and then the target recommendation model is obtained according to the calculated recommendation object corresponding to each group of sample data and the reference recommendation object corresponding to each group of sample data. A specific implementation method is provided for how to obtain the target recommendation model, which improves the feasibility of generating the target recommendation model.

[0144] Further, such as Figure 3BAs shown, it is a schematic diagram of another embodiment of the initial recommendation model in the embodiment of the present application. The step transformation module includes a first encoder, a step transformer, a control signal encoder and a state initializer. Optionally, in some embodiments of the present application, the sample recommendation cycle interaction data and the sample performance index weight sequence in each group of sample data are input into the step transformation module to obtain the transformation vector corresponding to each group of sample data, which may include:

[0145] The sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data are input into the first encoder to obtain the interaction data embedding vector corresponding to the sample recommendation period interaction data and the performance indicator embedding vector corresponding to the sample performance indicator weight sequence; the interaction data embedding vector is input into the state initializer to obtain the initial state vector; the interaction data embedding vector and the performance indicator embedding vector are input into the step transformer to obtain the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation; the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation are input into the control signal encoder to obtain the initial control vector.

[0146] Exemplarily, the step conversion module includes a first encoder, a step converter, a control signal encoder, and a state initializer. The sample recommendation period interaction data and the sample performance index weight sequence in each group of sample data can be input into the first encoder to obtain the interaction data embedding vector corresponding to the sample recommendation period interaction data. And the performance indicator embedding vector corresponding to the sample performance indicator weight sequence

[0147] Optionally, the first encoder includes a second target encoder and a third target encoder, and the sample recommendation period interaction data in each group of sample data is input into the second target encoder to obtain the interaction data embedding vector The sample performance index weight sequence in each group of sample data is input into the third target encoder to obtain the performance index embedding vector

[0148] Then, the interaction data is embedded into the vector Input to the state initializer to get the initial state vector Embed interaction data into vectors And the performance indicator embedding vector Input into the step transformer to obtain the interaction data embedding vector after each layer of step transformation and the performance indicator embedding vector after each layer of step transformation; input the interaction data embedding vector after each layer of step transformation and the performance indicator embedding vector after each layer of step transformation into the control signal encoder to obtain the initial control vector of each layer.

[0149] Assuming that the step transformer includes 3 layers, then after the first layer of step transformer processing, the interaction data embedding vector after the first layer of step transformer is obtained And the performance indicator embedding vector after the first layer step transformation After being processed by the second-layer step transformer, the interaction data embedding vector after the second-layer step transformation is obtained. And the performance indicator embedding vector after the second layer step transformation After being processed by the third-layer step transformer, the interaction data embedding vector after the third-layer step transformation is obtained. And the performance indicator embedding vector after the third step transformation

[0150] In the present technical scheme, based on the sample recommendation cycle interaction data and the sample performance indicator weight sequence in each group of sample data, the initial state vector and initial control vector of each group of sample data are obtained through a series of processing by the step transformation module including the first encoder, the step transformer, the control signal encoder and the state initializer, thereby providing a data basis for the subsequent calculation recommendation object corresponding to each group of sample data obtained through the sequence transformation module based on the initial state vector and the initial control vector, thereby improving the reliability of generating the target recommendation model.

[0151] Optionally, in some embodiments of the present application, the interaction data embedding vector after the step transformation and the performance indicator embedding vector after the step transformation are input into the control signal encoder to obtain an initial control vector, which may include: splicing the performance indicator embedding vector after the step transformation and the interaction data embedding vector after the step transformation to obtain a spliced ​​vector; inputting the spliced ​​vector into the control signal encoder to obtain an initial control vector.

[0152] Considering the inherent conflicts among multiple performance indicators, it is impractical to achieve the optimal state of all performance indicators. Therefore, we can set weights for these performance indicators and define a control signal c t Then use it to guide the recommendation model to generate recommended objects that meet the corresponding weight values ​​of these performance indicators. t is a representation of the expected future trajectory, given by the historical interaction data s t The weight values ​​corresponding to a set of performance indicators The parameters of this model are g φ .

[0153] So,

[0154] Among them, c t is the output of the control signal encoder, is the input of the control signal encoder.

[0155] Exemplarily, the performance indicator embedding vector after each layer of step transformation and the interaction data embedding vector after each layer of step transformation can be spliced ​​to obtain a spliced ​​vector for each layer; the spliced ​​vector for each layer is input into the control signal encoder to obtain an initial control vector for each layer.

[0156] For example: embed the performance indicators after the first layer of steps into the vector And the interaction data embedding vector after the first layer step transformation Perform splicing to obtain the first layer of splicing vector The first layer of concatenated vectors is input to the control signal encoder, and the first layer of initial control vectors is output.

[0157] The performance indicators after the second step transformation are embedded into the vector And the interaction data embedding vector after the second layer step transformation Splice to get the second layer of splicing vector The second layer concatenation vector is input to the control signal encoder, and the output second layer initial control vector

[0158] The performance index after the third step transformation is embedded into the vector And the interaction data embedding vector after the third step transformation Splice to get the third layer of splicing vector The third layer concatenation vector is input to the control signal encoder, and the output third layer initial control vector

[0159] In this technical solution, the performance indicator embedding vector after the step transformation and the interactive data embedding vector after the step transformation are spliced ​​to obtain a spliced ​​vector; the spliced ​​vector is input into the control signal encoder to obtain an initial control vector. How to obtain the initial control vector is specifically explained, which improves the feasibility of the solution.

[0160] Further, such as Figure 3C As shown, it is a schematic diagram of another embodiment of the initial recommendation model in the embodiment of the present application. The sequence transformation module includes a sequence transformer and a second encoder. Optionally, in some embodiments of the present application, the transformation vector corresponding to each group of sample data is input into the sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data, which may include: inputting the initial state vector and the initial control vector into the sequence transformer to obtain the vector after sequence transformation; inputting the vector after sequence transformation into the second encoder to obtain the calculation recommendation object corresponding to each group of sample data.

[0161] Exemplarily, the sequence conversion module includes a sequence converter and a second encoder. The initial state vector corresponding to each set of sample data can be and the initial control vector, input to the sequence transformer to obtain the vector after sequence transformation Then the vector after sequence transformation Input to the second encoder to obtain the calculation recommendation object a corresponding to each set of sample data t It should be noted that the number of layers l of the step converter and the sequence converter is the same. Therefore, the sequence converter is also described here by taking 3 layers as an example.

[0162] The first layer initial control vector and the initial state vector After being processed by the first layer sequence transformer, the second layer state vector is obtained Then the second layer initial control vector and the second layer state vector After being processed by the second-layer sequence transformer, the third-layer state vector is obtained Then the third layer initial control vector and the third-layer state vector After being processed by the third-layer sequence transformer, the fourth-layer state vector is obtained For the fourth layer state vector After encoding by the second encoder, the calculation recommendation object a corresponding to each set of sample data is obtained t ′.

[0163] In this technical solution, the transformation vector corresponding to each set of sample data is input into the sequence transformation module to obtain each set of sample data. t 'Then we can refer to the recommended object a corresponding to each set of sample data t Make a judgment to determine the accuracy of model training.

[0164] Exemplarily, the first encoder in the initial recommendation model includes two submodules: an object encoder (ObjEncoder) and a history encoder (Hist Encoder). The object encoder generates a weight sequence of sample performance indicators for sample recommendation period t. Transformed into the corresponding performance indicator embedding vector The weight values ​​of all input performance indicators are normalized. The history encoder uses the gated recurrent unit (GRU) model to transcode the sample recommendation cycle interaction data into the initial history embedding vector Here, the superscript 0 of each vector indicates that it is the initial embedding vector for the transformer in the subsequent step.

[0165] The step transformer processes the encoded embedding vectors and feeds them into L step transformer layers to enrich their representation. The output is then fed into the control signal encoder to construct the control signal c t .

[0166] Control signal encoder pair c t g in φ Here, we only need to transform the representation vector of the first layer of the step transformer Then, the control vector corresponding to the L layer of the sequence transformer is obtained by concatenating the multi-layer perceptron in the control signal encoder.

[0167] The state initializer further embeds the history into a vector Extracted into the initial state vector of the sequence transformer

[0168] The sequence transformer integrates the initial inputs from the control signal encoder and the state initializer. It systematically generates a sequence of recommended objects through its L sequence transformer layers, and finally reaches the time range of t+H steps through H steps. Each iterative step is decoded by a decoder to produce a specific action (i.e., recommend one or more recommended objects) to align the recommended objects with multiple performance indicators.

[0169] Together, these components form the MocDT model. This complex architecture is designed to tailor the recommendation path, ensuring that each sequence of recommended objects is precisely tailored to the multiple performance metrics required.

[0170] 2. Each set of sample data in the sample training set includes: sample user information, sample recommendation cycle interaction data, and sample performance indicator weight sequence.

[0171] refer to Figure 3AAs shown, the initial recommendation model includes a step transformation module and a sequence transformation model module. Optionally, in some embodiments of the present application, the initial recommendation model is trained according to the sample training set to obtain the target recommendation model, which may include: inputting the sample user information, sample recommendation cycle interaction data and sample performance index weight sequence in each group of sample data into the step transformation module to obtain the transformation vector corresponding to each group of sample data; the transformation vector includes an initial state vector and an initial control vector, the initial state vector is obtained according to the sample recommendation cycle interaction data, and the initial control vector is obtained according to the sample user information, the sample recommendation cycle interaction data and the sample performance index weight sequence; the transformation vector corresponding to each group of sample data is input into the sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data; according to the calculation recommendation object corresponding to each group of sample data and the reference recommendation object corresponding to each group of sample data, the target recommendation model is obtained. The steps in this implementation method can directly replace step 202.

[0172] In this technical solution, the sample user information (User features), sample recommendation period interaction data (Historical item features) and sample performance index weight sequence (O1, O2...) in each set of sample data can be input into the step transformation module for transformation, and the transformation vector (including the initial state vector and the initial control vector ); then the transformation vector corresponding to each group of sample data is input into the sequence transformation module to obtain the calculated recommendation object corresponding to each group of sample data, and then the target recommendation model is obtained according to the calculated recommendation object corresponding to each group of sample data and the reference recommendation object corresponding to each group of sample data. An implementation method is provided for how to obtain the target recommendation model, which improves the feasibility of generating the target recommendation model.

[0173] refer to Figure 3B As shown, the step transformation module includes a first encoder, a step transformer, a control signal encoder and a state initializer. Optionally, in some embodiments of the present application, the sample user information, the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data are input into the step transformation module to obtain the transformation vector corresponding to each group of sample data, which may include:

[0174] The sample user information, sample recommendation period interaction data and sample performance indicator weight sequence in each group of sample data are input into the first encoder to obtain a user information embedding vector corresponding to the sample user information, an interaction data embedding vector corresponding to the sample recommendation period interaction data, and a performance indicator embedding vector corresponding to the sample performance indicator weight sequence; the interaction data embedding vector is input into the state initializer to obtain an initial state vector; the user information embedding vector, the interaction data embedding vector and the performance indicator embedding vector are input into the step transformer to obtain a user information embedding vector after step transformation, an interaction data embedding vector after step transformation and a performance indicator embedding vector after step transformation; the user information embedding vector after step transformation, the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation are input into the control signal encoder to obtain an initial control vector.

[0175] Exemplarily, the step conversion module includes a first encoder, a step converter, a control signal encoder, and a state initializer. The sample user information, sample recommendation cycle interaction data, and sample performance indicator weight sequence in each group of sample data can be input into the first encoder to obtain a user information embedding vector corresponding to the sample user information. The interaction data embedding vector corresponding to the sample recommendation period interaction data And the performance indicator embedding vector corresponding to the sample performance indicator weight sequence

[0176] Optionally, the first encoder includes a first target encoder, a second target encoder and a third target encoder, and the sample user information in each group of sample data is input into the first target encoder to obtain the user information embedding vector The sample recommendation cycle interaction data in each group of sample data is input into the second target encoder to obtain the interaction data embedding vector The sample performance index weight sequence in each group of sample data is input into the third target encoder to obtain the performance index embedding vector

[0177] Then, the interaction data is embedded into the vector Input to the state initializer to get the initial state vector Embed user information into vector and interaction data embedding vector And the performance indicator embedding vector Input into the step transformer to obtain the user information embedding vector after each layer of step transformation, the interaction data embedding vector after each layer of step transformation, and the performance indicator embedding vector after each layer of step transformation; input the user information embedding vector after each layer of step transformation, the interaction data embedding vector after each layer of step transformation, and the performance indicator embedding vector after each layer of step transformation into the control signal encoder to obtain the initial control vector of each layer.

[0178] Assuming that the step transformer includes 3 layers, then after the first layer of step transformer processing, the user information embedding vector after the first layer of step transformer is obtained The first layer of transformed interaction data embedding vector And the performance indicator embedding vector after the first layer step transformation After being processed by the second-layer step transformer, the user information embedding vector after the second-layer step transformation is obtained The second-layer transformed interaction data embedding vector And the performance indicator embedding vector after the second layer step transformation After being processed by the third-layer step transformer, the user information embedding vector after the third-layer step transformation is obtained. The third step transforms the interaction data embedding vector And the performance indicator embedding vector after the third step transformation

[0179] In the present technical scheme, based on the sample user information, sample recommendation cycle interaction data and sample performance indicator weight sequence in each group of sample data, the initial state vector and initial control vector of each group of sample data are obtained through a series of processing by the step transformation module including the first encoder, the step transformer, the control signal encoder and the state initializer, thereby providing a data basis for the subsequent calculation of the recommendation object corresponding to each group of sample data through the sequence transformation module based on the initial state vector and the initial control vector, thereby improving the reliability of generating the target recommendation model.

[0180] Optionally, in some embodiments of the present application, the user information embedding vector after the step transformation, the interaction data embedding vector after the step transformation, and the performance indicator embedding vector after the step transformation are input into the control signal encoder to obtain an initial control vector, which may include: splicing the user information embedding vector after the step transformation, the performance indicator embedding vector after the step transformation, and the interaction data embedding vector after the step transformation to obtain a spliced ​​vector; inputting the spliced ​​vector into the control signal encoder to obtain an initial control vector.

[0181] Considering the inherent conflicts among multiple performance indicators, it is impractical to achieve the optimal state of all performance indicators. Therefore, we can set weights for these performance indicators and define a control signal c t Then use it to guide the recommendation model to generate recommended objects that meet the corresponding weight values ​​of these performance indicators. t is a representation of the expected future trajectory, given by the historical interaction data s t The weight values ​​corresponding to a set of performance indicators The parameters of this model are g φ .

[0182] So,

[0183] Among them, c t is the output of the control signal encoder, It is the input of the control signal encoder.

[0184] Exemplarily, the user information embedding vector after each layer of step transformation, the performance indicator embedding vector after each layer of step transformation, and the interaction data embedding vector after each layer of step transformation can be spliced ​​to obtain a spliced ​​vector for each layer; the spliced ​​vector for each layer is input into the control signal encoder to obtain an initial control vector for each layer.

[0185] For example: embed the user information after the first layer of transformation into a vector The performance indicator embedding vector after the first layer transformation And the interaction data embedding vector after the first layer step transformation Perform splicing to obtain the first layer of splicing vector The first layer of concatenated vectors is input to the control signal encoder, and the first layer of initial control vectors is output.

[0186] The user information after the second step is embedded into the vector The performance index embedding vector after the second-layer transformation And the interaction data embedding vector after the second layer step transformation Splice to get the second layer of splicing vector The second layer concatenation vector is input to the control signal encoder, and the output second layer initial control vector

[0187] The user information after the third step is embedded into the vector The performance index after the third step is embedded in the vector And the interaction data embedding vector after the third step transformation Splice to get the third layer of splicing vector The third layer concatenation vector is input to the control signal encoder, and the output third layer initial control vector

[0188] In this technical solution, the user information embedding vector after the step transformation, the performance indicator embedding vector after the step transformation, and the interactive data embedding vector after the step transformation are spliced ​​to obtain a spliced ​​vector; the spliced ​​vector is input into the control signal encoder to obtain an initial control vector. How to obtain the initial control vector is specifically explained, which improves the feasibility of the solution.

[0189] For further reference, Figure 3C As shown, the sequence transformation module includes a sequence transformer and a second encoder. Optionally, in some embodiments of the present application, inputting the transformation vector corresponding to each group of sample data into the sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data may include: inputting the initial state vector and the initial control vector into the sequence transformer to obtain the vector after sequence transformation; inputting the vector after sequence transformation into the second encoder to obtain the calculation recommendation object corresponding to each group of sample data.

[0190] Exemplarily, the sequence conversion module includes a sequence converter and a second encoder. The initial state vector corresponding to each set of sample data can be and the initial control vector, input to the sequence transformer to obtain the vector after sequence transformation Then the vector after sequence transformation Input to the second encoder to obtain the calculation recommendation object a corresponding to each set of sample data t It should be noted that the number of layers l of the step converter and the sequence converter is the same. Therefore, the sequence converter is also described here by taking 3 layers as an example.

[0191] The first layer initial control vector and the initial state vector After being processed by the first layer sequence transformer, the second layer state vector is obtained Then the second layer initial control vector and the second layer state vector After being processed by the second-layer sequence transformer, the third-layer state vector is obtained Then the third layer initial control vector and the third-layer state vector After being processed by the third-layer sequence transformer, the fourth-layer state vector is obtained For the fourth layer state vector After encoding by the second encoder, the calculation recommendation object a corresponding to each set of sample data is obtained t ′.

[0192] In this technical solution, the transformation vector corresponding to each set of sample data is input into the sequence transformation module, and the calculation recommendation object corresponding to each set of sample data is specifically detailed, and the calculation recommendation object a corresponding to each set of sample data is t 'Then we can refer to the recommended object a corresponding to each set of sample data t Make a judgment to determine the accuracy of model training.

[0193] like Figure 3D As shown in FIG. 1 , a schematic diagram of the initial recommendation model in an embodiment of the present application is shown. The first encoder includes three submodules: a user encoder (User Encoder), an object encoder (Obj Encoder), and a history encoder (HistEncoder). The user encoder converts the sample user information (including user ID and various available user features, etc.) of the sample recommendation period t into a user embedding vector The target encoder recommends sample performance index weight sequence for sample period t Transformed into the corresponding performance indicator embedding vector The weight values ​​of all input performance indicators are normalized. The history encoder uses the gated recurrent unit (GRU) model to transcode the sample recommendation cycle interaction data into the initial history embedding vector The superscript 0 in each vector indicates that it is the initial embedding vector for the subsequent Step Transformer.

[0194] The step transformer processes the encoded embedding vectors and feeds them into L step transformer layers to enrich their representation. The output is then fed into the control signal encoder to construct the control signal c t .

[0195] Control signal encoder pair c t g in φ Here, we only need to transform the representation vector of the first layer of the step transformer Then, the control vector corresponding to the L layer of the sequence transformer is obtained by concatenating the multi-layer perceptron in the control signal encoder.

[0196] The state initializer further embeds the history into a vector Extracted into the initial state vector of the sequence transformer

[0197] The sequence transformer integrates the initial inputs from the control signal encoder and the state initializer. It systematically generates a sequence of recommended objects through its L sequence transformer layers, and finally reaches the time range of t+H steps through H steps. Each iterative step is decoded by a decoder to produce a specific action (i.e., recommend one or more recommended objects) to align the recommended objects with multiple performance indicators.

[0198] Together, these components form the MocDT model. This complex architecture is designed to tailor the recommendation path, ensuring that each sequence of recommended objects is precisely tailored to the multiple performance metrics required.

[0199] Optionally, in some embodiments of the present application, obtaining a target recommendation model based on the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data may include: using a loss function to obtain the target recommendation model based on the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data.

[0200] For example, a loss function, such as a negative log-likelihood loss function, can be used to train a policy network π based on a sequence transformer in a sequence transformer module. θ The expression of the loss function is as follows:

[0201]

[0202] Among them, the reference recommended object for the correct answer is a t , because p(a t )=1, so p(a) in the formula can be omitted t ) term, so the expression of the loss function can be simplified as follows:

[0203]

[0204] E represents the expectation, and D refers to the input target information, including user information (User features) and historical interaction data (Historical item features). The purpose of the above formula is to improve all reference recommendation objects a t The probability predicted by the target recommendation model.

[0205] In this technical solution, a loss function is used to obtain a target recommendation model based on the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data. In theory, the smaller the value of the loss function, the higher the reliability of the target recommendation model, thereby improving the accuracy of the target recommendation model.

[0206] Optionally, in some possible implementations, performing model training on the initial recommendation model to obtain a target recommendation model based on the sample training set may include: performing data augmentation on the sample training set to obtain a data-enhanced sample training set; and performing model training on the initial recommendation model to obtain a target recommendation model based on the data-enhanced sample training set.

[0207] In recommendation systems, data augmentation is an important technique that expands and improves the sample training set of the recommendation system by generating new training data, thereby improving the accuracy and stability of the model. Data augmentation methods usually include random perturbations, data mixing, and data generation. These methods can help the model generalize, improve the accuracy and stability of the recommendation model, and deal with the problem of insufficient or unbalanced data.

[0208] Optionally, in some possible implementations, data enhancement is performed on the sample training set to obtain a sample training set after data enhancement, which may include: obtaining a target sample training set corresponding to a preset recommendation strategy based on the sample training set, wherein the reference recommendation objects in the target sample training set are obtained according to the preset recommendation strategy based on historical recommendation objects with which the user has interacted; and using the sample training set and the target sample training set as the sample training set after data enhancement.

[0209] In this technical solution, an implementation method of data enhancement is provided, which expands and improves the sample training set of the recommendation system by generating a new training data set, thereby improving the accuracy and stability of the model.

[0210] Optionally, in some embodiments of the present application, the sample training set is a data-enhanced sample training set, and the data-enhanced sample training set includes a data-unenhanced sample training set and a target sample training set, and the reference recommended objects in the target sample training set are obtained according to a preset recommendation strategy based on historical recommended objects with which the user has interacted.

[0211] In this technical solution, an implementation method of data enhancement is provided, which expands and improves the sample training set of the recommendation system by generating a new training data set, thereby improving the accuracy and stability of the model.

[0212] Optionally, in some embodiments of the present application, when the preset performance indicator is diversity, the reference recommended object in the target sample training set is the recommended object with the smallest overlap with the reference recommended object in the sample training set without data enhancement among the historical recommended objects with which the user has interacted; or,

[0213] When the preset performance indicator is the cumulative score, the reference recommended object in the target sample training set is the recommended object with the highest cumulative score among the historical recommended objects with which the user has interacted, excluding the reference recommended object in the sample training set without data enhancement.

[0214] The effectiveness of the decision transformer (DT) model depends on the quality of the offline sample training set (also called sample dataset). The DT model is biased towards the more likely to satisfy the specified future trajectory c t Reference recommendation object a t Therefore, without comprehensive offline trajectories, the learned recommendation strategy may not work well. Introducing a target trajectory sequence that can reflect the desired future trajectory synthesis can broaden the DT model vision and allow the learned recommendation strategy to select more appropriate recommendation objects.

[0215] For example, consider a recommendation trajectory τ. At timestamp t, the user has been recommended a list of historical recommended objects h t ={α t′ |t′≤t-1}. We construct a synthetic target trajectory sequence Then concatenate it with the original trajectory sequence: As the enhanced trajectory. All recommended objects in the target trajectory sequence are selected from the recommended objects that the user has interacted with in the offline sample training set. For example, here are three recommendation strategies as examples, as shown below:

[0216] If the preset performance indicator is the cumulative score, it means that the recommendation strategy focuses on the cumulative score. In each recommendation cycle t′≥t, the recommendation object list h t′ In the example, the recommended object with the highest cumulative score that has not yet appeared is selected as the next recommended object.

[0217] If the preset performance indicator is diversity, it means that the recommendation strategy focuses on diversity. Then a greedy strategy can be adopted. In each recommendation cycle t′>t, a target object in the recommended object list h is selected. t′ The recommended object with the smallest category overlap is selected as the next recommended object.

[0218] If the preset performance indicator is random, it means that the recommendation strategy is random, and the recommended objects are selected indiscriminately and included in the target trajectory sequence.

[0219] In this technical solution, by explaining the data enhancement for different preset performance indicators, the DT model vision can be broadened, allowing the learned recommendation strategy to select more appropriate recommendation objects.

[0220] 203. Obtain target information, where the target information includes historical interaction data using the target application and a recommendation strategy weight sequence.

[0221] 204. Input the target information into the target recommendation model to obtain the recommended objects of the target application in the current recommendation cycle.

[0222] It should be noted that steps 203-204 and Figure 1 Steps 101 - 102 in the illustrated embodiment are similar and will not be described in detail here.

[0223] This application can be widely applied to recommendation system scenarios. For example, this method can be used on electronic product platforms and short video platforms to simultaneously improve the recommendation performance of multiple performance indicators.

[0224] In an embodiment of the present application, a multi-performance indicator controllable decision transformer model is proposed. The model innovatively applies a control signal to represent multiple performance indicators. Through the weight values ​​corresponding to the multiple performance indicators, the recommendation model can be guided to generate a sequence of recommended objects consistent with the multiple performance indicators, and there is no need to retrain the recommendation model when the weight values ​​of the performance indicators change. In order to ensure that the recommendation model can effectively map multiple performance indicators to recommended objects, different data enhancement strategies can be used to simulate various interaction trajectories in existing offline interactions. The recommendation model is tested through a series of randomly selected performance indicators, and a sequence of recommended objects is generated in a sequential manner. Whether the trajectory generated by the recommendation model meets multiple performance indicators is used as the criterion for measuring success.

[0225] The embodiments of the present application can optimize the long-term performance indicators of the recommendation system, such as long-term user participation, satisfaction, etc. It can also optimize multiple performance indicators at the same time, and there is no need to retrain the recommendation model when the weight value of the performance indicator changes.

[0226] To evaluate the object recommendation method in the embodiment of the present application, three public recommendation datasets—dataset A, dataset B, and dataset C—can be used for testing, as shown in Table 1 below:

[0227] Dataset user Recommended Interaction data category Dataset A 6040 3706 1000209 19 Dataset B 7176 10728 12530806 31 Dataset C 27285 7551 1436609 46

[0228] Table 1

[0229] Exemplarily, this experiment selected a classic sequence recommender, five multi-performance index recommenders, one RL-based multi-performance index recommender, two multi-performance index controllable algorithms, and two DT-based recommenders as baseline models. The performance was measured by analyzing the entire test user group and calculating the mean and standard deviation of the generated results containing 10 recommended object sequences. This evaluation cumulatively evaluated the ability to adjust the results in terms of score (abbreviated as score) and diversity.

[0230] Here, we take the settings in an environment focusing on diversity as an example, because the same is true in an environment focusing on other performance indicators such as cumulative scores.

[0231] The recommendation performance of the embodiment of the present application is significantly improved. For the diverse performance indicators, the recommendation model proposed in the present application performs better than other models in each data set.

[0232] For classic sequential methods, their architectures use the management cumulative score as the only performance indicator, so they are less effective in scenarios with multiple performance indicators. For multiple performance indicator models without control functions, the lack of control capabilities weakens their flexibility and limits their ability to adjust recommendation results based on specific desired states.

[0233] For the model equipped with control function, its design of optimizing multiple performance indicators within a specified window size enables the model to directly pursue long-term multiple performance indicators instead of greedily pursuing immediate performance indicators.

[0234] Compared with the classic DT model, MocDT can simultaneously focus on and control the optimization of multiple performance indicators. This controllability of multiple performance indicators significantly improves its applicability and effectiveness in complex recommendation scenarios.

[0235] The control capability of the embodiment of the present application is significantly improved. Figure 4A As shown, it is a schematic diagram of the influence of the weight values ​​of multiple performance indicators predefined in the MocDT model framework on the evaluation performance indicators in the embodiment of the present application. The weight values ​​of these performance indicators are expressed in the form of a tuple, and the first and second elements correspond to the cumulative score and diversity, respectively. During the test, 10 recommended objects can be generated according to the weight values ​​of the given performance indicators, and the performance indicators of the cumulative score and diversity can be evaluated after each recommendation cycle.

[0236] The results of 10 recommendation cycles are shown in the figure with smaller patterns (such as squares, triangles (including triangles in different orientations), circles, pentagons, diamonds, etc.), depicting the instantaneous state, while the larger pattern expresses the average results, providing a strong measure of performance consistency. The empirical results of Datasets A, B, and C emphasize the key role of the weight values ​​of performance indicators in guiding the performance of the recommendation system. It is worth noting that the relative positions of the recommendation results corresponding to the weight values ​​of the nine performance indicators remain unchanged in each recommendation cycle, which means that the MocDT model has reliable multi-performance indicator control capabilities.

[0237] Dataset A shows obvious clustering effects on the weight values ​​of the nine performance indicators. The evaluation results of the weight values ​​of each performance indicator are closely grouped, which indicates that there is a strong collaborative filtering signal in dataset A, and the recommendation model is good at learning the mapping from multiple performance indicators to recommended objects.

[0238] On the contrary, although the recommendation models trained on datasets B and C fluctuate in different time periods, the relative position of the weight value results of each performance indicator remains unchanged, which shows that the control of multiple performance indicators of the recommendation model has a certain inherent stability. In dataset B, there is an inverse correlation between diversity and cumulative score, indicating that in scenarios such as short videos in dataset B, user preferences tend to be homogenized, and too diverse recommendations may reduce user engagement.

[0239] In sharp contrast, dataset C shows a positive correlation between diversity and cumulative score. As the weight value of the performance indicator increases, the diversity also increases accordingly. This reflects the content community background of dataset C, that is, users have an affinity for a variety of content. Improving content diversity is in line with user preferences and is likely to increase user satisfaction with recommendations.

[0240] In order to explore the specific impact of data augmentation on the performance of the recommendation model, different augmentation strategies and the effects of adjusting the augmentation rate are used here. To clearly present the results, we focus on reporting one core indicator: the cumulative score. Two different performance indicator weight values ​​are used here, namely (0.5, 1.0) and (1.0, 0.5), to explore the impact of different priority performance indicators.

[0241] Enhance the impact of the strategy. Figure 4B As shown in FIG. 1 , it is a schematic diagram showing the influence of different enhancement strategies on the recommendation performance in the embodiment of the present application. Figure 4B The analysis found that the three datasets responded differently to the enhancement strategies. In dataset A, the performance of the enhanced data is significantly improved over that of the unenhanced dataset. Data augmentation introduces a wider range of data changes, enabling the model to more effectively learn the mapping from multiple performance indicators to recommended objects. The four different enhancement strategies all produced similar diversification effects on the original unified dataset. In contrast, the cumulative scores of datasets B and C did not change much after data augmentation, which may be because these datasets themselves already contain rich scene changes. In particular, the rating-centric enhancement strategy performs better than random enhancement and diversity enhancement, a result consistent with its aim to promote high-rated user-recommendation object interaction, thereby improving the predictive ability of the recommendation model in terms of ratings.

[0242] The effect of enhancement rate. Figure 4CAs shown, it is a schematic diagram of the effect of different enhancement rates on recommendation performance in the embodiment of the present application. The increase in the enhancement rate has almost no effect on the cumulative score of data set A. This may be because the data set itself has sufficient diversity and collaborative filtering signals, and excessive data does not directly translate into performance improvement. On the contrary, the performance of data set B grows steadily with the increase in the enhancement rate, which is consistent with the positive orientation of data enhancement. Data enhancement helps the model to manage the complexity of negatively correlated indicators, thereby optimizing the accuracy of reasoning. However, as the enhancement rate of data set C increases, the indicators are decreasing. However, data enhancement introduces a large number of unconventional instances as negative samples, which distorts the distribution of training data and ultimately hinders the ability of the model to learn the real user preference patterns reflected in the test set.

[0243] like Figure 5 As shown, it is a schematic diagram of an embodiment of an object recommendation device in an embodiment of the present application, which may include:

[0244] The acquisition module 501 is used to acquire target information, wherein the target information includes historical interaction data of using the target application and a performance indicator weight sequence; the historical interaction data is obtained according to the interaction operation between the user and the historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes the weight value corresponding to each of the at least two performance indicators; the historical recommendation object includes the multimedia resources recommended in the historical recommendation cycle;

[0245] The obtaining module 502 is used to input the target information into a preset recommendation model to obtain the recommended object of the target application in the current recommendation cycle, wherein the recommended object of the current recommendation cycle is a recommended object that satisfies the performance indicator weight sequence and is recommended based on the historical interaction data, and the recommended object of the current recommendation cycle includes the multimedia resources recommended in the current recommendation cycle;

[0246] Among them, the target recommendation model is obtained by model training based on a sample training set, and the sample training set includes multiple groups of sample data, each group of sample data includes sample recommendation cycle interaction data and a sample performance indicator weight sequence, and a reference recommendation object corresponding to the sample recommendation cycle; the sample performance indicator weight sequence includes a sample weight value corresponding to each sample performance indicator of at least two sample performance indicators.

[0247] Optionally, in some possible implementations, the performance indicator weight sequence is pre-set, or is acquired according to a weight sequence setting instruction.

[0248] Optionally, in some possible implementations, the target information also includes user information using the target application, and each set of sample data also includes sample user information; wherein the recommended object of the current recommendation cycle is a recommended object that meets the performance indicator weight sequence based on the historical interaction data and the user information using the target application.

[0249] Optionally, in some possible implementations, the acquisition module 501 is further used to acquire the sample training set;

[0250] Obtaining module 502 is specifically used to input the sample recommendation cycle interaction data and the sample performance indicator weight sequence in each group of sample data into a step transformation module to obtain a transformation vector corresponding to each group of sample data; the transformation vector includes an initial state vector and an initial control vector, the initial state vector is obtained based on the sample recommendation cycle interaction data, and the initial control vector is obtained based on the sample recommendation cycle interaction data and the sample performance indicator weight sequence; inputting the transformation vector corresponding to each group of sample data into a sequence transformation module to obtain a calculated recommendation object corresponding to each group of sample data; obtaining the target recommendation model based on the calculated recommendation object corresponding to each group of sample data and the reference recommendation object corresponding to each group of sample data; wherein the initial recommendation model includes the step transformation module and the sequence transformation module.

[0251] Optionally, in some possible implementations, the step conversion module includes a first encoder, a step converter, a control signal encoder, and a state initializer;

[0252] Obtaining module 502 is specifically used to input the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data into the first encoder to obtain the interaction data embedding vector corresponding to the sample recommendation period interaction data and the performance indicator embedding vector corresponding to the sample performance indicator weight sequence; input the interaction data embedding vector into the state initializer to obtain the initial state vector; input the interaction data embedding vector and the performance indicator embedding vector into the step converter to obtain the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation; input the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation into the control signal encoder to obtain the initial control vector.

[0253] Optionally, in some possible implementations, a module 502 is obtained, which is specifically used to splice the performance indicator embedding vector after the transformation in the step and the interaction data embedding vector after the transformation in the step to obtain a spliced ​​vector; and input the spliced ​​vector into the control signal encoder to obtain the initial control vector.

[0254] Optionally, in some possible implementations, the sequence conversion module includes a sequence converter and a second encoder;

[0255] Obtaining module 502 is specifically used to input the initial state vector and the initial control vector into the sequence transformer to obtain the vector after sequence transformation; input the vector after sequence transformation into the second encoder to obtain the calculation recommendation object corresponding to each group of sample data.

[0256] Optionally, in some possible implementations, a module 502 is obtained, which is specifically used to use a loss function to obtain the target recommendation model according to the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data.

[0257] Optionally, in some possible implementations, the sample training set is a data-enhanced sample training set, and the data-enhanced sample training set includes a data-unenhanced sample training set and a target sample training set, and the reference recommended objects in the target sample training set are obtained according to preset performance indicators based on historical recommended objects with which the user has interacted.

[0258] Optionally, in some possible implementations, when the preset performance indicator is diversity, the reference recommended object in the target sample training set is a recommended object among historical recommended objects with which the user has interacted, and which has the smallest overlap with the reference recommended object in the sample training set without data enhancement; or,

[0259] When the preset performance indicator is a cumulative score, the reference recommended object in the target sample training set is the recommended object with the highest cumulative score among the historical recommended objects with which the user has interacted, excluding the reference recommended object in the sample training set without data enhancement.

[0260] Optionally, in some possible implementations, the object recommendation device further includes a display module 503;

[0261] The display module 503 is used to display the recommended objects of the target application in the current recommendation period through a display interface in response to a viewing trigger instruction in the current recommendation period.

[0262] like Figure 6As shown, it is a schematic diagram of an embodiment of an electronic device in the present application, which may include Figure 5 The object recommendation device shown.

[0263] like Figure 7 As shown, it is a schematic diagram of another embodiment of the electronic device in the embodiment of the present application, which may include: a memory 701, a processor 702, and a display 703. The memory 701 stores a computer program that can be run on the processor 702. When the electronic device executes the program, the above Figure 1 and Figure 2 The method of the illustrated embodiment.

[0264] In an embodiment of the present application, the processor 702 is used to obtain target information, wherein the target information includes historical interaction data of using a target application and a performance indicator weight sequence; the historical interaction data is obtained according to the interaction operation between the user and the historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes the weight value corresponding to each of the at least two performance indicators; the historical recommendation object includes the multimedia resources recommended in the historical recommendation cycle; the target information is input into a preset recommendation model to obtain the recommended object of the target application in the current recommendation cycle, the recommended object of the current recommendation cycle is the recommended object that satisfies the performance indicator weight sequence and is recommended according to the historical interaction data, and the recommended object of the current recommendation cycle includes the multimedia resources recommended in the current recommendation cycle;

[0265] Among them, the target recommendation model is obtained by model training based on a sample training set, and the sample training set includes multiple groups of sample data, each group of sample data includes sample recommendation cycle interaction data and a sample performance indicator weight sequence, and a reference recommendation object corresponding to the sample recommendation cycle; the sample performance indicator weight sequence includes a sample weight value corresponding to each sample performance indicator of at least two sample performance indicators.

[0266] Optionally, in some possible implementations, the performance indicator weight sequence is pre-set, or is acquired according to a weight sequence setting instruction.

[0267] Optionally, in some possible implementations, the target information also includes user information using the target application, and each set of sample data also includes sample user information; wherein the recommended object of the current recommendation cycle is a recommended object that meets the performance indicator weight sequence based on the historical interaction data and the user information using the target application.

[0268] Optionally, in some possible implementations, the processor 702 is further configured to obtain the sample training set;

[0269] Processor 702 is specifically used to input the sample recommendation cycle interaction data and the sample performance indicator weight sequence in each group of sample data into a step transformation module to obtain a transformation vector corresponding to each group of sample data; the transformation vector includes an initial state vector and an initial control vector, the initial state vector is obtained based on the sample recommendation cycle interaction data, and the initial control vector is obtained based on the sample recommendation cycle interaction data and the sample performance indicator weight sequence; input the transformation vector corresponding to each group of sample data into a sequence transformation module to obtain a calculated recommendation object corresponding to each group of sample data; obtain the target recommendation model based on the calculated recommendation object corresponding to each group of sample data and the reference recommendation object corresponding to each group of sample data; wherein the initial recommendation model includes the step transformation module and the sequence transformation module.

[0270] Optionally, in some possible implementations, the step conversion module includes a first encoder, a step converter, a control signal encoder, and a state initializer;

[0271] Processor 702 is specifically used to input the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data into the first encoder to obtain the interaction data embedding vector corresponding to the sample recommendation period interaction data and the performance indicator embedding vector corresponding to the sample performance indicator weight sequence; input the interaction data embedding vector into the state initializer to obtain the initial state vector; input the interaction data embedding vector and the performance indicator embedding vector into the step converter to obtain the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation; input the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation into the control signal encoder to obtain the initial control vector.

[0272] Optionally, in some possible implementations, the processor 702 is specifically used to splice the performance indicator embedding vector after the transformation in the step and the interaction data embedding vector after the transformation in the step to obtain a spliced ​​vector; and input the spliced ​​vector into the control signal encoder to obtain the initial control vector.

[0273] Optionally, in some possible implementations, the sequence conversion module includes a sequence converter and a second encoder;

[0274] Processor 702 is specifically used to input the initial state vector and the initial control vector into the sequence transformer to obtain the vector after sequence transformation; input the vector after sequence transformation into the second encoder to obtain the calculation recommendation object corresponding to each group of sample data.

[0275] Optionally, in some possible implementations, the processor 702 is specifically configured to use a loss function to obtain the target recommendation model according to the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data.

[0276] Optionally, in some possible implementations, the sample training set is a data-enhanced sample training set, and the data-enhanced sample training set includes a data-unenhanced sample training set and a target sample training set, and the reference recommended objects in the target sample training set are obtained according to preset performance indicators based on historical recommended objects with which the user has interacted.

[0277] Optionally, in some possible implementations, when the preset performance indicator is diversity, the reference recommended object in the target sample training set is a recommended object among historical recommended objects with which the user has interacted, and which has the smallest overlap with the reference recommended object in the sample training set without data enhancement; or,

[0278] When the preset performance indicator is a cumulative score, the reference recommended object in the target sample training set is the recommended object with the highest cumulative score among the historical recommended objects with which the user has interacted, excluding the reference recommended object in the sample training set without data enhancement.

[0279] Optionally, in some possible implementations, the display 703 is used to display the recommended objects of the target application in the current recommendation cycle through a display interface in response to a view trigger instruction in the current recommendation cycle.

[0280] 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. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from a website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (Digital Subscriber Line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, server or data center. The computer-readable storage medium can be any available medium that a computer can store or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a DVD), or a semiconductor medium (eg, a solid state disk (SSD)).

[0281] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0282] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. 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 units, which can be electrical, mechanical or other forms.

[0283] 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 on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0284] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0285] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk.

[0286] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned 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. An object recommendation method, characterized in that: include: Acquire target information, the target information including historical interaction data using the target application and a performance indicator weight sequence; The historical interaction data is obtained according to the interaction operation between the user and the historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes the weight value corresponding to each of the at least two performance indicators; the historical recommendation object includes the multimedia resources recommended in the historical recommendation cycle; Input the target information into a target recommendation model to obtain a recommended object of the target application in a current recommendation cycle, wherein the recommended object of the current recommendation cycle is a recommended object that satisfies the performance indicator weight sequence and is recommended based on the historical interaction data, and the recommended object of the current recommendation cycle includes a multimedia resource recommended in the current recommendation cycle; Among them, the target recommendation model is obtained by model training based on a sample training set, and the sample training set includes multiple groups of sample data, each group of sample data includes sample recommendation cycle interaction data and a sample performance indicator weight sequence, and a reference recommendation object corresponding to the sample recommendation cycle; the sample performance indicator weight sequence includes a sample weight value corresponding to each sample performance indicator of at least two sample performance indicators.

2. The method according to claim 1, characterized in that The performance indicator weight sequence is preset, or is acquired according to a weight sequence setting instruction.

3. The method according to claim 1 or 2, characterized in that: The target information also includes user information using the target application, and each set of sample data also includes sample user information; wherein, the recommended object of the current recommendation cycle is a recommended object that meets the performance indicator weight sequence based on the historical interaction data and the user information using the target application.

4. The method according to claim 1 or 2, characterized in that: The method further comprises: Obtaining the sample training set; Input the sample recommendation period interaction data and the sample performance index weight sequence in each group of sample data into a step transformation module to obtain a transformation vector corresponding to each group of sample data; the transformation vector includes an initial state vector and an initial control vector, the initial state vector is obtained according to the sample recommendation period interaction data, and the initial control vector is obtained according to the sample recommendation period interaction data and the sample performance index weight sequence; Inputting the transformation vector corresponding to each group of sample data into a sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data; Obtaining the target recommendation model according to the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data; The initial recommendation model includes the step transformation module and the sequence transformation module.

5. The method according to claim 4, characterized in that The step transformation module includes a first encoder, a step transformer, a control signal encoder and a state initializer; the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data are input into the step transformation module to obtain a transformation vector corresponding to each group of sample data, including: Inputting the sample recommendation period interaction data and the sample performance indicator weight sequence in each group of sample data into the first encoder to obtain an interaction data embedding vector corresponding to the sample recommendation period interaction data and a performance indicator embedding vector corresponding to the sample performance indicator weight sequence; Inputting the interaction data embedding vector into the state initializer to obtain the initial state vector; Input the interaction data embedding vector and the performance indicator embedding vector into the step transformer to obtain the interaction data embedding vector after step transformation and the performance indicator embedding vector after step transformation; The interaction data embedding vector after the transformation in the step and the performance indicator embedding vector after the transformation in the step are input into the control signal encoder to obtain the initial control vector.

6. The method according to claim 5, characterized in that The step of inputting the interactive data embedding vector transformed in the step and the performance indicator embedding vector transformed in the step into the control signal encoder to obtain the initial control vector comprises: splicing the performance indicator embedding vector transformed in the step and the interaction data embedding vector transformed in the step to obtain a spliced ​​vector; The concatenated vector is input into the control signal encoder to obtain the initial control vector.

7. The method according to claim 5, characterized in that The sequence transformation module includes a sequence transformer and a second encoder; the transformation vector corresponding to each group of sample data is input into the sequence transformation module to obtain the calculation recommendation object corresponding to each group of sample data, including: Inputting the initial state vector and the initial control vector into the sequence converter to obtain a vector after sequence conversion; The vector after the sequence transformation is input into the second encoder to obtain the calculation recommendation object corresponding to each group of sample data.

8. The method according to claim 4, characterized in that The obtaining the target recommendation model according to the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data includes: The target recommendation model is obtained by using a loss function according to the calculated recommended objects corresponding to each group of sample data and the reference recommended objects corresponding to each group of sample data.

9. The method according to claim 1 or 2, characterized in that: The sample training set is a data-enhanced sample training set, which includes a data-unenhanced sample training set and a target sample training set. The reference recommended objects in the target sample training set are obtained according to preset performance indicators based on historical recommended objects with which users have interacted.

10. The method according to claim 9, characterized in that When the preset performance indicator is diversity, the reference recommended object in the target sample training set is a recommended object with the smallest overlap with the reference recommended object in the sample training set without data enhancement among the historical recommended objects with which the user has interacted; or, When the preset performance indicator is a cumulative score, the reference recommended object in the target sample training set is the recommended object with the highest cumulative score among the historical recommended objects with which the user has interacted, excluding the reference recommended object in the sample training set without data enhancement.

11. The method according to claim 1 or 2, characterized in that: The method further comprises: In response to the viewing trigger instruction in the current recommendation cycle, the recommended objects of the target application in the current recommendation cycle are displayed through a display interface.

12. An object recommendation device, characterized in that: include: An acquisition module, used to acquire target information, wherein the target information includes historical interaction data using the target application and a performance indicator weight sequence; The historical interaction data is obtained according to the interaction operation between the user and the historical recommendation object in at least one historical recommendation cycle; the performance indicator weight sequence includes the weight value corresponding to each of the at least two performance indicators; The historical recommendation objects include multimedia resources recommended in the historical recommendation period; A obtaining module, used for inputting the target information into a preset recommendation model to obtain a recommended object of the target application in a current recommendation cycle, wherein the recommended object of the current recommendation cycle is a recommended object satisfying the performance indicator weight sequence recommended based on the historical interaction data, and the recommended object of the current recommendation cycle includes a multimedia resource recommended in the current recommendation cycle; Among them, the target recommendation model is obtained by model training based on a sample training set, and the sample training set includes multiple groups of sample data, each group of sample data includes sample recommendation cycle interaction data and a sample performance indicator weight sequence, and a reference recommendation object corresponding to the sample recommendation cycle; the sample performance indicator weight sequence includes a sample weight value corresponding to each sample performance indicator of at least two sample performance indicators.

13. An electronic device, characterized in that: The electronic device comprises a memory, a processor, and a display, wherein the memory stores a computer program that can be run on the processor, and when the electronic device executes the program, the method according to any one of claims 1 to 11 is implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

15. A computer program product having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 11 is implemented.

Images