An object recommendation method, device, electronic device and storage medium

Abstract

This application relates to the field of computer technology, and in particular, to an object recommendation method, apparatus, electronic device, and storage medium. A positive sample set and a negative sample set of the object to be recommended are obtained, and at least one set of candidate feature combinations for screening alternative users of the object to be recommended is obtained based on the user portrait features of each positive sample and the user portrait features of each negative sample; from the at least one set of candidate feature combinations, at least one set of candidate feature combinations whose association degree with the user portrait features of each user in the positive sample set meets a set condition is selected as the target feature combination; according to the selected target feature combination, target users whose user portrait features match the target feature combination are determined from the alternative users; and the object to be recommended is recommended to the determined target users, improving the efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular, to an object recommendation method, apparatus, electronic device, and storage medium. Background Art

[0002] Population targeting is a very important link in the advertising system. The goal of population targeting is to identify the users most relevant to a specified advertisement as the potential user group. In related technologies, the target population can be determined based on the portrait index, mainly relying on the construction of user portraits. Then, the advertiser sets the population targeting conditions, extracts the corresponding target population, and then recommends advertisements to the target population.

[0003] However, in this method of related technologies, when determining the targeting conditions, it is usually necessary to pre-conduct population portrait insights and manually analyze the insight results, so as to select the target portrait features as the targeting conditions relying on the advertiser's experience, with low accuracy. And because it is not known which portrait features are different, insights and analyses will be carried out on as many portrait features as possible, and only single-dimensional portrait features can be analyzed and calculated separately, with a large amount of calculation and low efficiency. Summary of the Invention

[0004] Embodiments of this application provide an object recommendation method, apparatus, electronic device, and storage medium to improve the accuracy and efficiency of generating targeting conditions, thereby improving the accuracy of targeted recommendations.

[0005] The specific technical solutions provided by the embodiments of this application are as follows:

[0006] An embodiment of this application provides an object recommendation method, including:

[0007] Obtain a positive sample set and a negative sample set of the object to be recommended, where each positive sample included in the positive sample set represents a user with positive behavioral characteristics for the recommendation optimization goal, and each negative sample included in the negative sample set represents a user with negative behavioral characteristics for the recommendation optimization goal;

[0008] Based on the user portrait features of each positive sample and the user portrait features of each negative sample, obtain at least one set of candidate feature combinations for screening alternative users of the object to be recommended;

[0009] From the at least one set of candidate feature combinations, screen out at least one set of candidate feature combinations whose association degree with the user portrait features of each user in the positive sample set meets the set conditions as the target feature combination;

[0010] According to the screened target feature combination, determine target users from the alternative users whose user portrait features match the target feature combination;

[0011] Recommend the object to be recommended to the determined target users.

[0012] Another embodiment of the present application provides an object recommendation device, including:

[0013] A first acquisition module, configured to acquire a positive sample set and a negative sample set of the object to be recommended, where each positive sample included in the positive sample set represents a user with positive behavioral characteristics for the recommendation optimization target, and each negative sample included in the negative sample set represents a user with negative behavioral characteristics for the recommendation optimization target;

[0014] A second acquisition module, configured to obtain at least one set of candidate feature combinations for screening alternative users of the object to be recommended based on the user portrait features of each positive sample and the user portrait features of each negative sample;

[0015] A screening module, configured to screen out at least one set of candidate feature combinations whose association degree with the user portrait features of each user in the positive sample set meets the set conditions as the target feature combination;

[0016] A determination module, configured to determine, according to the screened target feature combination, target users whose user portrait features match the target feature combination from the alternative users;

[0017] A recommendation module, configured to recommend the object to be recommended to the determined target users.

[0018] Another embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the steps of any of the above object recommendation methods are implemented.

[0019] Another embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above object recommendation methods are implemented.

[0020] Another embodiment of the present application provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes any of the object recommendation methods provided in the above various alternative implementation manners.

[0021] In the embodiments of the present application, a positive sample set and a negative sample set of the object to be recommended are obtained. Based on the user portrait features of each positive sample and the user portrait features of each negative sample, at least one set of candidate feature combinations for screening alternative users of the object to be recommended is obtained, and the at least one set of candidate feature combinations can be processed to screen out at least one set of candidate feature combinations whose degree of association with the user portrait features of each user in the positive sample set meets the set conditions as the target feature combinations. According to the screened target feature combinations, target users whose user portrait features match the target feature combinations are determined from the alternative users, and then the object to be recommended can be recommended to the determined target users. In this way, feature analysis can be performed on the specified positive sample set and negative sample set, each candidate feature combination can be automatically generated, and directional analysis of feature combinations is achieved, rather than just single-dimensional directional analysis. Moreover, each candidate feature combination can be automatically analyzed and screened to determine the target feature combinations as the directional conditions for the recommended target users, realizing fast generation of directional condition recommendations according to different recommendation optimization goals, greatly improving the efficiency and also improving the accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is an application architecture diagram of the object recommendation method in the embodiments of the present application;

[0023] Figure 2 is a flowchart of an object recommendation method in the embodiments of the present application;

[0024] Figure 3 is a schematic diagram of the principle of generating candidate feature combinations in the embodiments of the present application;

[0025] Figure 4 is a flowchart of another object recommendation method in the embodiments of the present application;

[0026] Figure 5 is a schematic diagram of the interface for initiating population feature analysis in the embodiments of the present application;

[0027] Figure 6 is a schematic diagram of the interface for determining the positive and negative sample sets in the embodiments of the present application;

[0028] Figure 7 is a schematic diagram of the principle of industry popular orientation in the embodiments of the present application;

[0029] Figure 8 is a schematic diagram of the principle of industry high-quality orientation in the embodiments of the present application;

[0030] Figure 9 is a schematic diagram of the structure of the object recommendation device in the embodiments of the present application;

[0031] Figure 10Schematic structural diagram of the electronic device in the embodiment of the present application. Detailed implementation manners

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0033] For the convenience of understanding the embodiments of the present application, several concepts will be briefly introduced below:

[0034] Population targeting: By analyzing user behavior data, user attribute characteristics, etc., the common characteristics of potential target user groups are found, and then the object to be recommended is delivered to the target user groups with common characteristics. For example, in advertising population targeting, the actual users to whom the advertisement is finally delivered appear in the target user group of the advertisement.

[0035] Advertising insight analysis: That is, user portrait analysis. Insight analysis can help customers more comprehensively and meticulously understand the characteristic distribution of the attributes, interest classifications, focus points, geographical characteristics, etc. of the population. These characteristics can be used to optimize advertising creativity, guide marketing strategies, and provide a reference basis for further formulating delivery, etc.

[0036] Click Through Rate (CTR): The ratio of clicks generated after the exposure of population-targeted users, which is the ratio of the number of click users to the number of exposed users, and can be used to measure the click effect of a population-targeted user of an advertisement after exposure.

[0037] Click Value Rate (CVR): The ratio of conversion behaviors that occur after population-targeted users click on the advertisement, which is the ratio of the number of converted users to the number of click users, and can be used to measure the conversion effect of a population-targeted user of an advertisement after clicking on the advertisement. Among them, the conversion behavior can be behaviors such as downloading, registering, purchasing, etc., without limitation, and can be set by the advertiser.

[0038] Target Group Index (TGI): It is an index that reflects the strength or weakness of the target group within a specific research scope (such as geographical area, demographic field, media audience, product consumers, etc.). Its calculation formula is: [the proportion of the group with a certain characteristic in the target group / the proportion of the group with the same characteristic in the overall population] * standard number 100. TGI can characterize the significance of a user portrait feature in the target population. The larger the TGI, the more significant the user portrait feature.

[0039] Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or a local area network to achieve data computing, storage, processing, and sharing.

[0040] Cloud technology is the general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the cloud computing business model. It can form a resource pool, be used as needed, and is flexible and convenient. Cloud computing technology will become an important support. The back-end services of technical network systems require a large amount of computing and storage resources, such as video websites, picture websites, and more portal websites. With the high development and application of the Internet industry, in the future, each item may have its own identification mark and needs to be transmitted to the back-end system for logical processing. Data at different levels will be processed separately, and various industry data requires the support of a powerful system back-end, which can only be achieved through cloud computing.

[0041] Artificial Intelligence (AI) is to use a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results in theory, method, technology, and application systems. In other words, artificial intelligence is a comprehensive technology in computer science. It attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable the machine to have the functions of perception, reasoning, and decision-making.

[0042] Artificial intelligence technology is an interdisciplinary subject with a wide range of fields, including both hardware-level technology and software-level technology. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0043] Machine Learning (ML) is an interdisciplinary field that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning. For example, in the embodiments of this application, an ensemble tree model can be trained based on machine learning methods, and then, based on the features represented by the nodes of each tree in the trained ensemble tree model, traverse from the root node to the leaf node to obtain each candidate feature combination.

[0044] It should be noted that in the embodiments of this application, operations involving obtaining data such as user portrait features, positive behavior features, and negative behavior features require obtaining user permission or consent when these embodiments of this application are applied to specific products or technologies, and the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards of relevant countries and regions. For example, when relevant data needs to be obtained, relevant volunteers can be recruited and relevant agreements on authorizing data for volunteers can be signed, and then the data of these volunteers can be used for implementation; or, implementation can be carried out within the scope of an organization that has been authorized to allow, and the following implementation methods can be implemented using the data of internal members of the organization to conduct relevant identification for internal members; or, the relevant data used in specific implementations are all simulated data, such as simulated data generated in a virtual scenario.

[0045] With the research and progress of artificial intelligence technology, artificial intelligence technology has been studied and applied in multiple fields. For example, common ones include smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, driverless, autonomous driving, drones, robots, smart healthcare, smart customer service, etc. With the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0046] The solution provided in the embodiments of this application mainly relates to the machine learning technology of artificial intelligence, which is specifically described through the following embodiments:

[0047] Taking object advertising as an example, when making advertising recommendations, audience targeting is a very important link in the advertising system. Whether from the perspective of advertisers or from the perspective of traffic measurement, the process of advertising audience targeting is required. In related technologies, the audience targeting method usually relies on portrait indexing. For the experimental group and the control group, it determines from which portrait dimensions to conduct insights and submits insight tasks. Then, advertisers manually analyze the insight results and rely on the experience of advertisers to determine the most significant portrait features for the experimental group, that is, the target portrait features, as the targeting conditions. Then, based on the targeting conditions, the corresponding target audience is extracted. However, in this way, since it is not known which portrait features are different, insights and analyses are carried out on as many portrait features as possible, and only single-dimensional portrait features can be analyzed and calculated separately. The computational workload is large, the efficiency is low, the overall time consumption is long, and manual analysis is required, resulting in a large labor cost and low accuracy.

[0048] Therefore, to solve the above problems, an object recommendation method is provided in the embodiments of the present application. A positive sample set and a negative sample set of the object to be recommended are obtained. Based on the user portrait features of each positive sample and the user portrait features of each negative sample, at least one set of candidate feature combinations for screening alternative users of the object to be recommended is obtained. From at least one set of candidate feature combinations, a target feature combination is screened out. And according to the screened target feature combinations, target users whose user portrait features match the target feature combinations are determined from the alternative users, and the object to be recommended is recommended to the determined target users. In this way, the problem of determining the target features is transformed into the problem of determining the features that can best distinguish between positive and negative samples. The user portrait features of the positive sample set and the negative sample set are automatically analyzed to determine the target feature combinations. According to different recommendation optimization goals, targeting conditions can be quickly generated for recommendation, which can greatly reduce the manual operations of advertisers, improve efficiency, and can also perform combined targeting analysis on features, rather than just analyzing from a single dimension, improving the accuracy of determining targeting conditions, thereby improving the accuracy of object recommendation.

[0049] Refer to Figure 1 As shown in the figure, it is an application architecture schematic diagram of the object recommendation method in the embodiments of the present application, including a user terminal 100, a placement platform 200, and a server 300.

[0050] The user terminal 100 can be any intelligent device such as a smart phone, a tablet computer, a portable personal computer, a desktop computer, a smart TV, a smart robot, a vehicle-mounted electronic device, etc. Various application programs (Application, APP) can be installed on the user terminal 100. The server 300 can collect and obtain the behavior data of each user on each user terminal 100, the attribute information of the user, etc., so as to construct a user portrait.

[0051] The placement platform 200 can be installed in a terminal device, which can also be any intelligent device such as a smart phone, a tablet computer, a portable personal computer, a desktop computer, a smart TV, a smart robot, an in-vehicle electronic device, etc. It can provide services related to advertisement placement. For example, it can access behavioral data, population management, insight analysis, multi-terminal placement docking and other functions. It can select the population to be analyzed and user portrait features in the placement platform 200, and then generate an insight task.

[0052] The server 300 can provide various network services for the user terminal 100 and the placement platform 200. For different application programs, the server 300 can be regarded as the corresponding background server. For example, an advertiser selects the feature range to be insighted and the population to be insighted in the placement platform 200. Then, the server 300 can determine a positive sample set and a negative sample set for the population to be insighted, analyze the user portrait features of each user in the positive sample set and the negative sample set respectively, train an ensemble tree model, thereby obtaining at least one set of candidate feature combinations, screening out the target feature combination from them as the targeting condition, determining the target users who meet the target feature combination, and then recommending advertisements to the determined target users, that is, placing advertisements on the user terminal 100 where the target users are located. For example, when the target user opens a certain APP on the user terminal 100, the placed advertisement can be displayed in the APP.

[0053] Among them, the server 300 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, as well as big data and artificial intelligence platforms.

[0054] For example, when the server 300 is a cloud server, cloud computing in cloud technology can be used to automatically analyze and calculate user portrait features to quickly generate targeting condition recommendations, which can greatly improve the computing efficiency.

[0055] Among them, cloud computing refers to the delivery and usage model of IT infrastructure, which means obtaining the required resources in a on-demand and easily scalable manner through the network; generalized cloud computing refers to the delivery and usage model of services, which means obtaining the required services in a on-demand and easily scalable manner through the network. Such services can be related to IT and software, the Internet, or other services. Cloud computing is the product of the development and integration of traditional computer and network technologies such as grid computing, distributed computing, parallel computing, utility computing, network storage technologies, virtualization, and load balance.

[0056] With the development of the Internet, real-time data streams, and the diversification of connected devices, as well as the promotion of demands such as search services, social networks, mobile commerce, and open collaboration, cloud computing has developed rapidly. Different from previous parallel distributed computing, the emergence of cloud computing will drive a revolutionary change in the entire Internet model and enterprise management model conceptually.

[0057] Between the user terminal 100 and the server 300, and between the placement platform 200 and the server 300, they can be directly or indirectly connected through wired or wireless communication, which is not restricted in this application. For example, Figure 1As shown, take communication between each other realized based on the Internet connection as an example. Optionally, the above-mentioned Internet uses standard communication technologies and / or protocols. The Internet is usually the Internet, but it can also be any network, including but not limited to any combination of a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a mobile, wired or wireless network, a private network or a virtual private network. In some embodiments, technologies and / or formats including Hyper Text Mark-up Language (HTML), Extensible Markup Language (XML), etc. are used to represent the data exchanged through the network. In addition, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), Internet Protocol Security (IPsec), etc. can be used to encrypt all or some of the links. In other embodiments, custom and / or dedicated data communication technologies can also be used to replace or supplement the above data communication technologies.

[0058] It should be noted that the application architecture diagram in the embodiments of the present application is for more clearly illustrating the technical solutions in the embodiments of the present application, and does not constitute a limitation on the technical solutions provided in the embodiments of the present application. It can be applied to the advertising targeted placement scenario, but there is no limitation on other application architectures and business applications, and it can also be applied to other object targeted recommendation scenarios. The technical solutions provided in the embodiments of the present application are equally applicable to similar problems.

[0059] In each embodiment of the present application, the object recommendation method is applied to Figure 1 the application architecture shown as an example for illustrative purposes, and for ease of explanation, the object is mainly taken as an advertisement in the embodiments of the present application for explanation.

[0060] Based on the above embodiments, refer to Figure 2 As shown, it is a flowchart of an object recommendation method in an embodiment of the present application, which is described by taking an application to a server as an example. Specifically, the method includes:

[0061] Step 200: Obtain a positive sample set and a negative sample set of the object to be recommended.

[0062] Among them, each positive sample included in the positive sample set represents a user with positive behavioral characteristics for the recommendation optimization target, and each negative sample included in the negative sample set represents a user with negative behavioral characteristics for the recommendation optimization target.

[0063] In the embodiments of the present application, through the analysis of the demands for population targeting. For example, when targeting the advertising population, the needs of advertisers are usually to analyze the differences between the experimental group and the control group, and find out the typical user portrait characteristics, so as to use them as the population targeting conditions. To meet this need and save a large amount of manual analysis costs at the same time, a method that can automatically analyze user portrait characteristics in batches needs to be selected. Therefore, in the embodiments of the present application, considering that the model analysis method can be adopted. For example, if the experimental group is used as the positive sample, the control group is used as the negative sample, and the user portrait characteristics, etc. are used as the model characteristics, then finding the differential typical user portrait characteristics can be transformed into the problem of finding the characteristics that can best distinguish the positive and negative samples. First, the positive sample set and the negative sample set need to be obtained. When specifically performing step 200, there are the following several situations:

[0064] The first situation: The advertiser designates the populations of the control group and the experimental group.

[0065] Then determine the positive sample set and the negative sample set, specifically including: taking the uploaded users in the experimental group as the positive sample set, and taking the uploaded users in the control group as the negative sample set.

[0066] In the first situation, it is suitable for the advertiser to have already determined what the positive population and the negative population are respectively. For example, member users can be uploaded as the experimental group, and users with negative comments can be uploaded as the control group, so as to take the users in the experimental group as the positive sample set and the users in the control group as the negative sample set.

[0067] The second situation: The advertiser does not designate the populations of the control group and the experimental group.

[0068] Then determine the positive sample set and the negative sample set, specifically including: obtaining the recommendation optimization target and the historical reference object recommendation information, and according to the recommendation optimization target, screening out the users with positive behavioral characteristics for the recommendation optimization target from the historical reference object recommendation information to obtain the positive sample set, and screening out the users with negative behavioral characteristics for the recommendation optimization target to obtain the negative sample set.

[0069] In the second situation, it is not necessary for the advertiser to directly display and provide the populations of the control group and the experimental group. It is only necessary to set the recommendation optimization target and provide the relevant historical reference object recommendation information, and then the positive sample set and the negative sample set can be automatically obtained by extraction according to the recommendation optimization target and the historical reference object information.

[0070] Among them, taking an advertisement for an object as an example, the acquisition of historical reference object recommendation information can be, for example, that the advertiser provides an advertiser identifier. According to the advertiser identifier, the historical reference advertisement previously placed corresponding to the advertiser identifier is determined, and the historical reference advertisement recommendation information of the historical reference advertisement is obtained; for another example, it can be that the advertiser directly provides the identifier of the historical reference advertisement, and according to the identifier of the historical reference advertisement, the historical reference advertisement recommendation information is obtained; in addition, the historical reference object recommendation information can also be the recommendation information of the historical reference advertisements of other advertisers, such as other advertisements of the same type as the advertisement to be recommended, etc. The specific manner of obtaining the historical reference object recommendation information is not limited in the embodiments of the present application.

[0071] The recommended optimization target can be the click-through rate or the conversion rate, etc., and is not limited. For example, if the recommended optimization target is the click-through rate, the positive sample is the user who clicks on the object, and the negative sample is the user who is exposed to the object but does not click. For another example, if the recommended optimization target is the conversion rate, the positive sample is the user who converts the object, and the negative sample is the user who clicks but does not convert the object.

[0072] Step 210: Based on the user portrait features of each positive sample and the user portrait features of each negative sample, obtain at least one set of candidate feature combinations for screening alternative users of the object to be recommended.

[0073] When specifically performing step 210, it includes:

[0074] S1. Respectively determine the user portrait features of each positive sample in the positive sample set and the user portrait features of each negative sample in the negative sample set.

[0075] For example, the placement platform can provide multiple positioning dimensions, that is, multi-dimensional user portrait features. The advertiser can select the feature range from them. For example, the user portrait features in the system platform can be divided into six categories, namely basic attributes, interests and hobbies, user behavior, user status, user environment, and custom. Among them, the basic attributes can include gender, age, region, education level, etc., the interests and hobbies can include business interests, keywords, etc., the user behavior is the APP behavior, which can include active, paid users, etc., the user status can include marital status, living status, residential community price, etc., the user environment can include Internet access scenarios, device prices, operating systems, Internet connection methods, mobile operators, weather conditions, etc., and the custom can include web pages, QQ numbers, International Mobile Equipment Identity (IMEI), etc. Then, some of the user portrait features can be selected, or all of them can be selected to participate in the subsequent integrated tree model training, which is not limited in the embodiments of the present application.

[0076] In this way, after extracting the user portrait features of each positive sample and each negative sample, encoding processing is performed to generate a vector representation of the user portrait features that can be input into the integrated tree model for training. For example, the one-hot encoding method can be used for encoding processing to obtain a sparse discrete feature vector representation of the user. The embodiments of the present application do not limit this.

[0077] Furthermore, in the embodiments of the present application, the extracted user portrait features can also be filtered. Specifically, a possible implementation manner is provided in the embodiments of the present application to filter out the user portrait features with a coverage rate less than a preset threshold in the positive sample set or the negative sample set. Here, the coverage rate is the proportion of the user portrait feature in the positive sample set or the negative sample set. The preset threshold is, for example, one-thousandth, and can be set according to actual needs. In this way, the user portrait features with a low coverage rate, that is, a low usage frequency, can be filtered out, which can reduce the calculation amount and improve the efficiency and accuracy of subsequent model training.

[0078] S2. Based on the user portrait features of each positive sample and the user portrait features of each negative sample, obtain at least one group of candidate feature combinations for screening alternative users of the object to be recommended.

[0079] Among them, the alternative users can be all users, or the user group finally determined through screening for extracting target users. The embodiments of the present application do not specifically limit this.

[0080] In the embodiments of the present application, through research and analysis, in the supervised machine learning method, the decision tree model can automatically generate feature combinations and can give the information gain feature importance, which can be used to select typical features. And the integrated tree model among them has a better effect. For example, the eXtreme Gradient Boosting (XGBoost) model can better fit the positive and negative samples through multiple decision trees. Therefore, in the embodiments of the present application, the integrated tree model is used to analyze the user portrait features of the positive and negative samples and automatically generate feature combinations.

[0081] That is, the user portrait features of each positive sample in the positive sample set, each negative sample in the negative sample set, and the corresponding positive and negative sample labels are input into the integrated tree model for training. Through training, each positive sample or each negative sample can be assigned to the root nodes, nodes, and leaf nodes of each tree according to the user portrait features. Each root node, node, and leaf node corresponds to a user portrait feature.

[0082] Among them, the number of trees in the integrated tree model is the first number, and the depth of each tree is the second number. For example, preferably, the first number can be set to 150-200. Of course, the optimal number of trees can also be determined by the effect of the validation set without limitation. At the same time, the depth of each tree cannot be set too large because the candidate feature combinations are finally obtained according to the features corresponding to the nodes passed from the root node to the leaf node of each tree. The depth of the tree determines the order of the final feature combination. Generally, a higher-order feature combination can describe the target user more accurately than a lower-order or single-dimensional feature. However, if the order of the feature combination is too large, the user scale corresponding to the feature combination will be too small. Therefore, the depth of the tree can be set by comprehensively considering the user scale corresponding to the combined features and the accuracy required for the targeted population. For example, the second number of the depth of the tree can be set to 3. Of course, there is no limitation in the embodiments of the present application.

[0083] When step S2 is executed, it specifically includes:

[0084] S2.1. Train an integrated tree model according to the user portrait features of each positive sample in the positive sample set and the user portrait features of each negative sample in the negative sample set to obtain a trained integrated tree model.

[0085] S2.2. For each tree in the trained integrated tree model, traverse from the root node of the tree to each leaf node, combine the user portrait features corresponding to the root node to each leaf node to obtain each highest-order feature combination, and obtain at least one group of candidate feature combinations according to the feature combinations in each non-empty subset of each highest-order feature combination. Among them, the order of the highest-order feature combination is the second number, and the order of at least one group of candidate feature combinations is from 1 to the second number.

[0086] In the embodiments of the present application, after training the integrated tree model, the final training result of the integrated tree model is not used, that is, it is not necessary to use the trained integrated tree model for classification. Instead, other by-products generated during training are used, that is, the features and association relationships represented by each node. For each tree in the integrated tree model, it is possible to traverse from the root node to the leaf node. Traversing from the root node to the leaf node can obtain a set of highest-order feature combinations. For example, if the second number is N, that is, the order of the highest-order feature combination is N, then the non-empty subsets of these highest-order feature combinations can be used as candidate feature combinations of orders 1 to N.

[0087] For example, refer to Figure 3 As shown, it is a schematic diagram of the principle of generating candidate feature combinations in the embodiments of the present application. Taking a tree in the integrated tree model and N being 3 as an example, as Figure 3As shown, the tree includes a root node A, with the corresponding user portrait feature being f1, two nodes B1 and B2, with the corresponding user portrait features being f2 and f3 respectively, and four leaf nodes C1, C2, C3, and C4, with the corresponding user portrait features being f4, f5, f6, and f7 respectively. By traversing from the root node A to each leaf node, four highest-order feature combinations can be obtained, namely f1 and f2 and f4, f1 and f2 and f5, f1 and f3 and f6, and f1 and f3 and f7. Then, non-empty subsets of each highest-order feature combination are determined respectively. Finally, candidate feature combinations of orders 1 to 3 can be obtained. Taking one of the highest-order feature combinations, f1 and f2 and f4, as an example, the candidate feature combinations determined by it are f1, f2, f4, f1 and f2, f1 and f4, f2 and f4, and f1 and f2 and f4.

[0088] Step 220: From at least one set of candidate feature combinations, screen out at least one set of candidate feature combinations whose association degrees with the user portrait features of each user in the positive sample set meet the set conditions as the target feature combinations.

[0089] In the embodiments of the present application, analysis and sorting can be respectively performed on the candidate feature combinations of each order. For example, sort the candidate feature combinations of the first order, sort the candidate feature combinations of the second order, etc. In this way, it is convenient for users to subsequently determine the target feature combinations from the candidate feature combinations of different orders according to the recommended accuracy requirements.

[0090] Among them, the association degrees between the user portrait features of each user in the positive sample set can be comprehensively characterized by using a significance index and a difference index, with the aim of finding more significant feature combinations in the positive sample set.

[0091] Step 230: According to the screened target feature combinations, determine target users from the alternative users whose user portrait features match the target feature combinations.

[0092] For example, if the alternative users are all users, and a target feature combination is f1 and f2 and f4, then target users can be extracted from the set of all users. The users whose user portrait features simultaneously have the features f1, f2, and f3 are the target users.

[0093] Step 240: Recommend the object to be recommended to the determined target users.

[0094] That is, the object to be recommended can be recommended to the user terminals where the target users are located to improve the recommendation accuracy and effect.

[0095] Next, in step 220 above, from at least one set of candidate feature combinations, at least one set of candidate feature combinations whose association degree with the user portrait features of each user in the positive sample set meets the set conditions is screened out as the target feature combination. A possible implementation manner is provided in the embodiments of the present application, which specifically includes:

[0096] S1.1. Calculate the proportions of at least one set of candidate feature combinations in the positive sample set, the negative sample set, and the preset full-scale user set respectively.

[0097] Among them, the proportion of a candidate feature combination in the positive sample set indicates the proportion of users with this candidate feature combination in the positive sample set among all users in the positive sample set. Similarly, the proportion in the negative sample set indicates the proportion of users with this candidate feature combination in the negative sample set among all users in the negative sample set, and the proportion in the full-scale user set indicates the proportion of users with this candidate feature combination in the full-scale user set among all users in the full-scale user set.

[0098] S1.2. Determine the positive sample set target group index corresponding to at least one set of candidate feature combinations respectively according to the proportions of at least one set of candidate feature combinations in the positive sample set and in the preset full-scale user set, and determine the negative sample set target group index corresponding to at least one set of candidate feature combinations respectively according to the proportions of at least one set of candidate feature combinations in the negative sample set and in the preset full-scale user set.

[0099] For example, taking one candidate feature combination as an example, the positive sample set target group index (TGI) and the negative sample set target group index (TGI) corresponding to this candidate feature combination are respectively:

[0100]

[0101] S1.3. Obtain the significance index and the difference index of at least one set of candidate feature combinations respectively according to the positive sample set target group index and the negative sample set target group index corresponding to at least one set of candidate feature combinations.

[0102] In the embodiments of the present application, through the analysis of the required target feature combinations, it can be seen that from the perspective of finally extracting the targeted target population, it is hoped to find feature combinations that are more significant in the positive population. Here, a problem of comparing positive and negative populations (i.e., the experimental group and the control group) is involved. Therefore, the target feature combinations that meet the requirements need to have two conditions: 1) Significance, that is, the TGI of the candidate feature combination, that is, the proportion of the candidate feature combination in the positive sample set or the negative sample set is relatively large compared to the proportion in the full user set. Otherwise, the number of samples containing the candidate feature combination in the training set may be small and the confidence level is low; 2) Difference, that is, the candidate feature combination is more significant in the positive population. Otherwise, using the candidate feature combination cannot distinguish between positive and negative populations. And if only the target group index of the positive sample set and the target group index of the negative sample set are used, two situations may occur: a) For example, the TGI of the positive sample set is 100 and the TGI of the negative sample set is 10. In this case, the ratio of the TGI of the positive sample set to the TGI of the negative sample set is large, but the size of the TGI of the positive sample set itself is not large enough and not significant enough to meet the significance requirement; b) For example, the TGI of the positive sample set is 10000 and the TGI of the negative sample set is 9000. In this case, the difference between the TGI of the positive sample set and the TGI of the negative sample set is large, but the ratio of the TGI of the positive sample set to the TGI of the negative sample set is small, that is, it is significant in both positive and negative sample sets and does not meet the difference requirement. Therefore, in the embodiments of the present application, based on the target group index of the positive sample set and the target group index of the negative sample set, a significance index and a difference index are defined.

[0103] Specifically, 1) The average value of the target group index of the positive sample set and the target group index of the negative sample set corresponding to at least one group of candidate feature combinations is respectively used as the significance index of at least one group of candidate feature combinations.

[0104] For example, the significance index is:

[0105]

[0106] 2) The ratio of the target group index of the positive sample set and the target group index of the negative sample set corresponding to at least one group of candidate feature combinations is respectively used as the difference index of at least one group of candidate feature combinations.

[0107] For example, the difference index is:

[0108]

[0109] S1.4. Respectively obtain the final sorting index of at least one group of candidate feature combinations according to the significance index and the difference index of at least one group of candidate feature combinations.

[0110] In the embodiments of the present application, when sorting at least one group of candidate feature combinations, it is necessary to fuse the significance index and the difference index. Specifically:

[0111] 1) Normalize the significance index and the difference index of at least one group of candidate feature combinations respectively.

[0112] For example, the normalization process can be defined as:

[0113]

[0114] where mean represents the mean value and std represents the variance.

[0115] In this way, after normalizing at least one group of candidate feature combinations respectively, it can be ensured that at least one group of candidate feature combinations are all within the same value range, which is convenient for comparison. The final sorting index of at least one group of candidate feature combinations is more comparable, thus improving the accuracy of sorting.

[0116] 2) Take the sum of the normalized significance index and the normalized difference index of at least one group of candidate feature combinations respectively as the final sorting index of at least one group of candidate feature combinations.

[0117] For example, the final sorting index = the normalized significance index + the normalized difference index.

[0118] In this way, the final sorting index of at least one group of candidate feature combinations can be used to sort at least one group of candidate feature combinations corresponding to at least one order respectively. The larger the final sorting index, the more accurate the target user characterized by the candidate feature combination.

[0119] S1.5. Sort the at least one group of candidate feature combinations of each order from high to low according to the final sorting index of the at least one group of candidate feature combinations.

[0120] S1.6. Screen out the target feature combinations from the at least one group of candidate feature combinations according to the sorting result and the preset number of target users.

[0121] In the embodiments of the present application, several possible implementation manners are provided for screening out the target feature combinations from the at least one group of candidate feature combinations according to the sorting result and the preset number of target users:

[0122] The first implementation manner:

[0123] Specifically, it includes: 1) According to the specified order number, sequentially screen out the candidate feature combinations that can extract the number of candidate feature combinations satisfying the preset number of target users from the at least one group of candidate feature combinations corresponding to the specified order number from high to low according to the sorting result, as the target feature combinations.

[0124] That is, the advertiser can select from the candidate feature combinations of which order to determine the target feature combination and extract the target users according to its own needs. For example, if an advertiser has a small budget and hopes that the advertisement placement is more accurate, since the higher the order of the candidate feature combination, the more accurate the determined target users are, the advertiser can select a higher order. For example, the highest order is 3, and the specified order is 3. Moreover, the advertiser can also set the number of target users to be recommended. Then, from each candidate feature combination of the 3rd order, according to the sorting result from high to low, the target feature combination can be determined to ensure that the target users extracted according to the determined target feature combination can meet the requirement of the number of target users.

[0125] 2) If the number of target users that can be extracted by at least one group of candidate feature combinations corresponding to the specified order does not meet the preset number of target users, then screen from at least one group of candidate feature combinations corresponding to other orders according to the sorting result from high to low until a candidate feature combination that can extract the number of target users meeting the preset number of target users is screened out to obtain the target feature combination.

[0126] For example, if the target users extracted based on each candidate feature combination of the 3rd order are still less than the preset number of target users, the target feature combination can be automatically determined from other orders. Preferably, it is determined from the order adjacent to the specified order, either one order higher or one order lower, to try to meet the accuracy requirements of the advertiser. Of course, the embodiments of the present application do not limit this.

[0127] The second implementation manner:

[0128] Specifically, it includes: according to the sorting result from high to low, screen out from at least one group of candidate feature combinations of the highest order to at least one group of candidate feature combinations of the lower order in turn the candidate feature combinations that can extract the number of target users meeting the preset number of target users as the target feature combination.

[0129] That is to say, in the embodiments of the present application, it is not necessary for the advertiser to specify the order in advance, but only need to set the required number of target users. At this time, according to the number of target users, first determine from each candidate feature combination of the highest order, and then determine the target feature combination from each candidate feature combination of the lower order in turn until the determined target feature combination can extract the number of target users meeting the preset number of target users.

[0130] Further, to further improve accuracy, after sorting at least one set of candidate feature combinations for each order from high to low, and before screening out the target feature combination from at least one set of candidate feature combinations according to the sorting result and the preset number of target users, the at least one set of sorted candidate feature combinations can also be filtered and screened first. Specifically, a possible implementation manner is provided in the embodiments of the present application. According to the sorting result, for at least one set of candidate feature combinations for each order, the candidate feature combinations with a preset number before sorting are screened out, that is, the candidate feature combinations that are not sorted among the top preset number can be filtered out. Then, when screening the target feature combination later, it can be screened from the candidate feature combinations with a preset number before sorting. In this way, the candidate feature combinations with relatively low sorting may not have obvious significance and difference, and may not be accurate for target user extraction, so they can be filtered out, which can further improve the accuracy of determining the targeting conditions and the accuracy of recommendation.

[0131] Further, in the embodiments of the present application, at least one set of candidate feature combinations is automatically generated based on the integrated tree model and sorted. Then, the at least one set of sorted candidate feature combinations can be displayed. When determining the target feature combination, the embodiments of the present application can also support manual selection of the target feature combination. For example, the advertiser can refer to the sorting results of each candidate feature combination and manually select the required candidate feature combination as the target feature combination. The embodiments of the present application do not limit this.

[0132] In the embodiments of the present application, a positive sample set and a negative sample set of the object to be recommended are obtained. Based on the user portrait features of each positive sample and the user portrait features of each negative sample, at least one set of candidate feature combinations for screening alternative users of the object to be recommended is obtained. From at least one set of candidate feature combinations, at least one set of candidate feature combinations whose association degree with the user portrait features of each user in the positive sample set meets the set conditions is screened out as the target feature combination. According to the screened target feature combination, target users whose user portrait features match the target feature combination are determined from the alternative users. Then, the object to be recommended can be recommended to the determined target users. In this way, by training an ensemble tree model, the user portrait features of each user in the specified positive sample set and negative sample set can be analyzed, and each candidate feature combination can be automatically generated, so as to screen out the target feature combination from them as the directional condition, without manual analysis and screening, independent of the advertiser's experience, greatly improving the efficiency, accuracy, and reducing the overall time consumption. For example, the analysis of tens of thousands or even more user portrait features can be completed within 30 minutes. Moreover, it can not only support comparison with the full user set, but also support direct comparison of two specific populations, and can automatically generate candidate feature combinations. It not only supports first-order user portrait feature analysis, but also supports high-order user portrait feature combination analysis, and also improves the accuracy of the determined target users and the accuracy of recommendation.

[0133] Based on the above embodiments, the following uses a specific application scenario to illustrate the overall solution of the object recommendation method in the embodiments of the present application. The object recommendation method in the embodiments of the present application can be divided into the following parts: 1) sample generation; 2) feature engineering; 3) model training; 4) candidate feature combination generation; 5) result ranking; 6) directional extraction, corresponding to Figure 4 each step in Figure 4 As shown, it is a schematic flowchart of another object recommendation method in the embodiments of the present application, including:

[0134] Step 400: Determine the positive sample set and the negative sample set.

[0135] In the sample generation part, the users in the experimental group are used as the positive sample set, and the users in the control group are used as the negative sample set.

[0136] Step 401: Respectively determine the user portrait features of each user in the positive sample set and the negative sample set.

[0137] Step 402: Train an ensemble tree model according to the user portrait features of each user in the positive sample set and the negative sample set.

[0138] Among them, the ensemble tree model is, for example, an XGBoost model, which is not limited in the embodiments of the present application.

[0139] Step 403: Based on the trained ensemble tree model, obtain each candidate feature combination.

[0140] Specifically, for each tree in the trained ensemble tree model, traverse from the root node of the tree to each leaf node, combine the user portrait features corresponding to the root node to each leaf node to obtain each highest-order feature combination, and use the feature combinations in each non-empty subset of each highest-order feature combination as each candidate feature combination.

[0141] Since the ensemble tree model is used for classification, through continuous learning based on the user portrait features of positive samples and negative samples, the positive samples and negative samples can be classified. The root node and the nodes from the node to the leaf node in the ensemble tree model all correspond to user portrait features. The candidate feature combinations obtained through such traversal and combination can be used to screen the alternative users of the object to be recommended.

[0142] Step 404: Screen out the target feature combination from each candidate feature combination.

[0143] Specifically, obtain the final sorting index of each candidate feature combination, and sort each candidate feature combination of each order from high to low according to the final sorting index of each candidate feature combination. Thus, according to the sorting result and the preset number of target users, the target feature combination can be screened out from each candidate feature combination, that is, the determined target feature combination is usually the first N candidate feature combinations in the sorting result.

[0144] Step 405: Determine the target users that match the target feature combination according to the screened target feature combination.

[0145] It should be noted that the above steps 400 - 404 are the automatic analysis part of the user portrait features, and step 405 is the target user extraction part. In this way, in the embodiment of the present application, based on the ensemble model, the automatic analysis of each user portrait feature is realized, and the candidate feature combination is generated. Furthermore, the target feature combination is analyzed and determined, and the target users are determined, and the object is recommended to the target users, realizing the targeted recommendation function of the object, improving the accuracy and efficiency. For example, it is applied to the advertising positioning and placement service scenario to improve the advertising placement effect. Through analysis and verification, compared with the method of manually selecting the targeted population by advertisers in the related technology, the method in the embodiment of the present application has an average increase of 15% in the click-through rate and an average increase of 20% in the conversion rate in terms of the advertising placement effect. While improving the advertising placement effect, it also reduces the manual analysis time consumption and cost of advertisers or operators.

[0146] Based on the above embodiments, the object recommendation method in the embodiments of the present application will be described from the product side below. Taking the application in the advertising targeted placement scenario as an example, in the population insight analysis provided by the placement platform in the embodiments of the present application, a population feature analysis function can be provided to perform user portrait feature analysis and quickly generate targeted condition recommendations. For example, refer to Figure 5 As shown in

[0147] After the user selects the "Population Feature Analysis" function, the population to be analyzed can be selected. For example, refer to Figure 6 As shown in Figure 6 As shown in

[0148] In the placement platform, "My Population" and "Advertising Population" can be provided. "My Population" represents the population owned by the placement platform itself, and "Advertising Population" represents the population that has interacted with the advertisement. The advertiser can select the population to be analyzed according to the needs, and then the positive sample set and negative sample set can be determined from it. In addition, the advertiser can also select the feature range in the placement platform, that is, the range of user portrait features to be analyzed, so as to generate an insight task, which can automatically analyze the user portrait features of the positive sample set and negative sample set, generate each candidate feature combination, and perform sorting and screening to determine the target feature combination.

[0149] In addition, for the problems existing in the related technology, in the embodiments of the present application, in addition to the automatic targeted recommendation based on the ensemble tree model, several possible implementation manners are also provided. The targeted recommendation results at the industry granularity can be calculated based on the industry granularity, as follows:

[0150] The first implementation manner: Industry popular targeting, that is, targeting with relatively high user portrait features in object recommendation.

[0151] Specifically: 1) For each feature combination, obtain the object information of the historical targeted recommendation using the feature combination and the object information of the historical non - targeted recommendation using the feature combination.

[0152] Among them, the object information for historical feature combination targeted recommendation at least includes the number of objects for which feature combination targeted recommendation is used, and the object information for historical non - use of feature combination targeted recommendation at least includes the number of objects for which feature combination targeted recommendation is not used.

[0153] Moreover, in the embodiments of the present application, the object information for historical use of feature combination targeted recommendation and the object information for historical non - use of feature combination targeted recommendation can be obtained from a single industry or from the entire industry, and the embodiments of the present application do not limit this.

[0154] 2) Determine the sum of the number of objects for which feature combination targeted recommendation is used and the number of objects for which feature combination targeted recommendation is not used, and use the ratio of the number of objects for which feature combination targeted recommendation is used to this sum as the usage frequency index of this feature combination.

[0155] For example, taking the object as an advertisement, refer to Figure 7 As shown, it is a schematic diagram of the industry - popular targeting principle in the embodiments of the present application. As shown in Figure 7 Shown, Figure 7 In (1) of Figure 7 is the advertisement for the entire industry, and in (2) of or

[0156] 3) Based on the usage frequency index corresponding to each feature combination, sort each feature combination respectively.

[0157] 4) According to the sorting result and the preset target user quantity, screen out the target feature combination from each feature combination, and determine the target users who meet the target feature combination.

[0158] 5) Recommend objects to the determined target users.

[0159] The second implementation method: industry - high - quality targeting, that is, the targeting with better conversion effect during the object recommendation process.

[0160] Specifically: 1) For each feature combination, obtain the number of first objects that have had an object click behavior but have not been targeted recommended based on the feature combination;

[0161] Obtain the number of second objects that have had an object click behavior and have been targeted recommended based on the feature combination;

[0162] Obtain the number of third objects that have had an object click and conversion behavior but have not been targeted recommended based on the feature combination;

[0163] Obtain the number of fourth objects that have undergone object clicks and conversions and are targeted and recommended based on feature combinations.

[0164] 2) Determine a conversion index of the feature combination according to the first object quantity, the second object quantity, the third object quantity, and the fourth object quantity.

[0165] For example, see Figure 8 As shown, it is a schematic diagram of the industry high-quality orientation principle in the embodiment of this application, such as Figure 8 As shown, A represents the number of first objects, B represents the number of second objects, C represents the number of third objects, and B represents the number of fourth objects. The conversion index used for sorting is:

[0166]

[0167] 3) Based on the conversion index corresponding to each feature combination, each feature combination is sorted.

[0168] 4) According to the sorting results and the preset number of target users, a target feature combination is screened out from each feature combination, and target users who meet the target feature combination are determined.

[0169] 5) Recommend objects to the identified target users.

[0170] The third implementation method: industry potential targeting, that is, targeting with good conversion effect and low usage frequency in the object recommendation process.

[0171] The third implementation is based on the first and second implementations, specifically: 1) removing the target feature combination determined based on the sorting based on the conversion index from the target feature combination determined based on the sorting based on the usage frequency, to determine the final target feature combination.

[0172] For example, industry potential orientation = industry quality orientation - industry popularity orientation.

[0173] 2) Identify target users who meet the final target characteristics and recommend objects to the identified target users.

[0174] It should be noted that for the above three implementation manners in the embodiments of the present application, automatic analysis can also be performed on each feature combination without manual analysis, which can also solve the problems of large computing performance and labor cost in the related art, improve efficiency and accuracy. However, compared with the targeted recommendation based on the integrated tree model in the above embodiments, the analysis ranges of the experimental group and the control group cannot be specified, and only based on the industry granularity, recommendations cannot be customized according to advertisers, and the analysis dimension is also relatively single. That is to say, compared with the targeted recommendation based on the integrated tree model, the above three implementation manners may have lower accuracy and efficiency. However, compared with the related art, they can still solve the problems existing in the related art. These several targeted recommendation manners should all fall within the scope protected by the present application.

[0175] Based on the same inventive concept, an object recommendation device is further provided in the embodiments of the present application. The object recommendation device may be a hardware structure, a software module, or a combination of a hardware structure and a software module. Based on the above embodiments, refer to Figure 9 As shown, the object recommendation device in the embodiments of the present application specifically includes:

[0176] A first obtaining module 90, configured to obtain a positive sample set and a negative sample set of the object to be recommended, where each positive sample included in the positive sample set represents a user with positive behavior characteristics for the recommendation optimization target, and each negative sample included in the negative sample set represents a user with negative behavior characteristics for the recommendation optimization target;

[0177] A second obtaining module 91, configured to obtain at least one group of candidate feature combinations for screening alternative users of the object to be recommended based on the user portrait features of each positive sample and the user portrait features of each negative sample;

[0178] A screening module 92, configured to screen out at least one group of candidate feature combinations whose association degrees with the user portrait features of each user in the positive sample set meet the set conditions from the at least one group of candidate feature combinations as the target feature combinations;

[0179] A determining module 93, configured to determine, according to the screened target feature combinations, target users whose user portrait features match the target feature combinations from the alternative users;

[0180] A recommendation module 94, configured to recommend the object to be recommended to the determined target users.

[0181] Optionally, when obtaining the positive sample set and the negative sample set of the object to be recommended, the first obtaining module 90 is specifically configured to: use the uploaded experimental group users as the positive sample set, and use the uploaded control group users as the negative sample set.

[0182] Optionally, the first obtaining module 90 is specifically configured to: obtain the recommended optimization target and historical reference object recommendation information, and according to the recommended optimization target, screen out users with positive behavior characteristics for the recommended optimization target from the historical reference object recommendation information to obtain a positive sample set, and screen out users with negative behavior characteristics for the recommended optimization target to obtain a negative sample set.

[0183] Optionally, the second obtaining module 91 is further configured to:

[0184] Filter out user portrait features with a coverage less than a preset threshold in the positive sample set or the negative sample set.

[0185] Optionally, when obtaining at least one set of candidate feature combinations for screening alternative users of the to-be-recommended object based on the user portrait features of each positive sample and the user portrait features of each negative sample, the second obtaining module 91 is specifically configured to:

[0186] Train an ensemble tree model according to the user portrait features of each positive sample in the positive sample set and the user portrait features of each negative sample in the negative sample set to obtain a trained ensemble tree model, where the number of trees in the ensemble tree model is the first number, and the depth of each tree is the second number;

[0187] For each tree in the trained ensemble tree model, traverse from the root node to each leaf node, combine the user portrait features corresponding to from the root node to each leaf node to obtain each highest-order feature combination, and according to the feature combinations in each non-empty subset of each highest-order feature combination, obtain at least one set of candidate feature combinations, where the order of the highest-order feature combination is the second number, and the order of the at least one set of candidate feature combinations is from 1 to the second number.

[0188] Optionally, when screening out at least one set of candidate feature combinations that meet the set conditions for the correlation degree between the at least one set of candidate feature combinations and the user portrait features of each user in the positive sample set as the target feature combination, the screening module 92 is specifically configured to:

[0189] Calculate the proportions of the at least one set of candidate feature combinations in the positive sample set, the negative sample set, and the preset full-scale user set respectively;

[0190] Respectively determine the positive sample set target population index corresponding to the at least one set of candidate feature combinations according to the proportions of the at least one set of candidate feature combinations in the positive sample set and in the preset full-scale user set, and respectively determine the negative sample set target population index corresponding to the at least one set of candidate feature combinations according to the proportions of the at least one set of candidate feature combinations in the negative sample set and in the preset full-scale user set;

[0191] Obtain the significance index and difference index of the at least one set of candidate feature combinations respectively according to the positive sample set target population index and the negative sample set target population index corresponding to the at least one set of candidate feature combinations;

[0192] Obtain the final ranking index of the at least one set of candidate feature combinations respectively according to the significance index and difference index of the at least one set of candidate feature combinations;

[0193] Rank the at least one set of candidate feature combinations of each order from high to low respectively according to the final ranking index of the at least one set of candidate feature combinations;

[0194] Screen out the target feature combinations from the at least one set of candidate feature combinations according to the ranking result and the preset number of target users.

[0195] Optionally, when obtaining the significance index and difference index of the at least one set of candidate feature combinations respectively according to the positive sample set target population index and the negative sample set target population index corresponding to the at least one set of candidate feature combinations, the screening module 92 is specifically configured to:

[0196] Respectively take the average value of the positive sample set target population index and the negative sample set target population index corresponding to the at least one set of candidate feature combinations as the significance index of the at least one set of candidate feature combinations;

[0197] Respectively take the ratio of the positive sample set target population index to the negative sample set target population index corresponding to the at least one set of candidate feature combinations as the difference index of the at least one set of candidate feature combinations.

[0198] Optionally, when obtaining the final ranking index of the at least one set of candidate feature combinations respectively according to the significance index and difference index of the at least one set of candidate feature combinations, the screening module 92 is specifically configured to:

[0199] Normalize the significance index and difference index of the at least one set of candidate feature combinations respectively;

[0200] Respectively take the sum of the normalized significance index and the normalized difference index of the at least one set of candidate feature combinations as the final ranking index of the at least one set of candidate feature combinations.

[0201] Optionally, when screening out the target feature combinations from the at least one set of candidate feature combinations according to the ranking result and the preset number of target users, the screening module 92 is specifically configured to:

[0202] According to the specified order, starting from the at least one group of candidate feature combinations corresponding to the specified order in descending order of the sorting result, filter out the candidate feature combinations that can extract the target user quantity meeting the preset target user quantity as the target feature combination;

[0203] If the target user quantity that can be extracted by the at least one group of candidate feature combinations corresponding to the specified order does not meet the preset target user quantity, then filter from the at least one group of candidate feature combinations corresponding to other orders in descending order of the sorting result until the candidate feature combinations that can extract the target user quantity meeting the preset target user quantity are filtered out to obtain the target feature combination.

[0204] Optionally, when filtering out the target feature combination from the at least one group of candidate feature combinations according to the sorting result and the preset target user quantity, the filtering module 92 is specifically configured to:

[0205] In descending order of the sorting result, starting from the at least one group of candidate feature combinations of the highest order to the at least one group of candidate feature combinations of the lower order, filter out the candidate feature combinations that can extract the target user quantity meeting the preset target user quantity as the target feature combination.

[0206] Based on the above embodiments, refer to Figure 10 The following is a schematic structural diagram of an electronic device in an embodiment of the present application.

[0207] An embodiment of the present application provides an electronic device, which may be a terminal or a server. In the embodiment of the present application, the electronic device is taken as an example of a server for illustration. The electronic device may include a processor 1010 (Center Processing Unit, CPU), a memory 1020, an input device 1030, an output device 1040, etc.

[0208] The memory 1020 may include a read-only memory (ROM) and a random access memory (RAM), and provide program instructions and data stored in the memory 1020 to the processor 1010. In the embodiment of the present application, the memory 1020 may be used to store the program of any object recommendation method in the embodiment of the present application.

[0209] The processor 1010 is configured to execute any object recommendation method in the embodiment of the present application according to the program instructions obtained by calling the program instructions stored in the memory 1020.

[0210] Based on the above embodiments, in the embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the object recommendation method in any method embodiment described above is implemented.

[0211] Based on the above embodiments, in the embodiments of the present application, a computer program product or a computer program is further provided. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the object recommendation method in any of the above method embodiments.

[0212] Those of ordinary skill in the art can understand that all or part of the steps to implement the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other various media that can store program codes.

[0213] Alternatively, if the above-integrated unit is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks or optical disks and other various media that can store program codes.

Claims

1. An object recommendation method, characterized in that, Including: Obtaining a positive sample set and a negative sample set of the object to be recommended, where each positive sample included in the positive sample set represents a user with positive behavioral characteristics for the recommendation optimization target, and each negative sample included in the negative sample set represents a user with negative behavioral characteristics for the recommendation optimization target; Based on the user portrait features of each positive sample and the user portrait features of each negative sample, obtaining at least one set of candidate feature combinations for screening alternative users of the object to be recommended; Based on the non-empty subsets of each feature included in each candidate feature combination, screening out at least one set of candidate feature combinations that meet the set conditions for the degree of association with the user portrait features of each user in the positive sample set as the target feature combination; According to the screened target feature combination, determining target users from the alternative users whose user portrait features match the target feature combination; Recommending the object to be recommended to the determined target users; Among them, the obtaining at least one set of candidate feature combinations for screening alternative users of the object to be recommended based on the user portrait features of each positive sample and the user portrait features of each negative sample includes: based on the trained ensemble tree model, allocating each positive sample and each negative sample to different leaf nodes, and obtaining the at least one set of candidate feature combinations based on the path from the root node to the leaf node.

2. The method according to claim 1, characterized in that Obtaining the positive sample set and the negative sample set of the object to be recommended specifically includes: Taking the uploaded experimental group users as the positive sample set and the uploaded control group users as the negative sample set; or, Obtaining the recommendation optimization target and the historical reference object recommendation information, and according to the recommendation optimization target, screening out users with positive behavioral characteristics for the recommendation optimization target from the historical reference object recommendation information to obtain the positive sample set, and screening out users with negative behavioral characteristics for the recommendation optimization target to obtain the negative sample set.

3. The method according to claim 1, wherein Further including: Filtering out user portrait features with a coverage rate less than a preset threshold in the positive sample set or the negative sample set.

4. The method according to claim 1, wherein The obtaining at least one set of candidate feature combinations for screening alternative users of the object to be recommended based on the user portrait features of each positive sample and the user portrait features of each negative sample specifically includes: Training an ensemble tree model according to the user portrait features of each positive sample in the positive sample set and the user portrait features of each negative sample in the negative sample set to obtain a trained ensemble tree model, where the number of trees in the ensemble tree model is the first number, and the depth of each tree is the second number; For each tree in the trained ensemble tree model respectively, traversing from the root node of the tree to each leaf node, combining the user portrait features corresponding to from the root node to each leaf node to obtain each highest-order feature combination, and obtaining at least one set of candidate feature combinations according to the feature combinations in the non-empty subsets of each highest-order feature combination, where the order of the highest-order feature combination is the second number, and the order of the at least one set of candidate feature combinations is from 1 to the second number.

5. The method according to any one of claims 1 to 4, characterized in that From the at least one set of candidate feature combinations, screen out at least one set of candidate feature combinations whose association degrees with the user portrait features of each user in the positive sample set meet the set conditions as the target feature combinations, specifically including: Calculate the proportions of the at least one set of candidate feature combinations in the positive sample set, the negative sample set, and the preset full-scale user set respectively; Determine the positive sample set target population indexes corresponding to the at least one set of candidate feature combinations respectively according to the proportions of the at least one set of candidate feature combinations in the positive sample set and in the preset full-scale user set, and determine the negative sample set target population indexes corresponding to the at least one set of candidate feature combinations respectively according to the proportions of the at least one set of candidate feature combinations in the negative sample set and in the preset full-scale user set; Obtain the significance indexes and difference indexes of the at least one set of candidate feature combinations respectively according to the positive sample set target population indexes and negative sample set target population indexes corresponding to the at least one set of candidate feature combinations, and obtain the final sorting indexes of the at least one set of candidate feature combinations respectively according to the significance indexes and difference indexes of the at least one set of candidate feature combinations; Sort the at least one set of candidate feature combinations of each order from high to low according to the final sorting indexes of the at least one set of candidate feature combinations; Screen out the target feature combinations from the at least one set of candidate feature combinations according to the sorting result and the preset target user quantity; 6. The method according to claim 5, wherein Obtain the significance indexes and difference indexes of the at least one set of candidate feature combinations respectively according to the positive sample set target population indexes and negative sample set target population indexes corresponding to the at least one set of candidate feature combinations, specifically including: Take the averages of the positive sample set target population indexes and negative sample set target population indexes corresponding to the at least one set of candidate feature combinations respectively as the significance indexes of the at least one set of candidate feature combinations; Take the ratios of the positive sample set target population indexes and negative sample set target population indexes corresponding to the at least one set of candidate feature combinations respectively as the difference indexes of the at least one set of candidate feature combinations; 7. The method according to claim 5, wherein Obtain the final sorting indexes of the at least one set of candidate feature combinations respectively according to the significance indexes and difference indexes of the at least one set of candidate feature combinations, specifically including: Normalize the significance indexes and difference indexes of the at least one set of candidate feature combinations respectively; Take the sums of the normalized significance indexes and normalized difference indexes of the at least one set of candidate feature combinations respectively as the final sorting indexes of the at least one set of candidate feature combinations; 8. The method according to claim 5, characterized in that, Screen out the target feature combinations from the at least one set of candidate feature combinations according to the sorting result and the preset target user quantity, specifically including: According to the specified order, sequentially screen out the candidate feature combinations that can extract the preset target user quantity from the at least one set of candidate feature combinations corresponding to the specified order from high to low according to the sorting result as the target feature combinations; If the number of target users that can be extracted by at least one set of candidate feature combinations corresponding to the specified order does not meet the preset number of target users, then start screening from at least one set of candidate feature combinations corresponding to other orders in descending order according to the sorting result until a candidate feature combination that can extract a number of target users meeting the preset number of target users is screened out to obtain the target feature combination.

9. The method according to claim 5, wherein Based on the sorting result and the preset number of target users, screening out the target feature combination from the at least one set of candidate feature combinations specifically includes: In descending order according to the sorting result, successively screening out from at least one set of candidate feature combinations of the highest order to at least one set of candidate feature combinations of the lower order a candidate feature combination that can extract a number of target users meeting the preset number of target users as the target feature combination.

10. An object recommendation device, characterized in that, It includes: A first obtaining module, configured to obtain a positive sample set and a negative sample set of the object to be recommended, where each positive sample included in the positive sample set represents a user with positive behavior characteristics for the recommendation optimization goal, and each negative sample included in the negative sample set represents a user with negative behavior characteristics for the recommendation optimization goal; A second obtaining module, configured to obtain at least one set of candidate feature combinations for screening alternative users of the object to be recommended based on the user portrait features of each positive sample and the user portrait features of each negative sample; A screening module, configured to, based on non-empty subsets of each feature included in each candidate feature combination, screen out at least one set of candidate feature combinations whose association degree with the user portrait features of each user in the positive sample set meets the set condition from the at least one set of candidate feature combinations as the target feature combination; A determination module, configured to determine, according to the screened target feature combination, target users whose user portrait features match the target feature combination from the alternative users; A recommendation module, configured to recommend the object to be recommended to the determined target users; When the second obtaining module is configured to obtain at least one set of candidate feature combinations for screening alternative users of the object to be recommended based on the user portrait features of each positive sample and the user portrait features of each negative sample, it is specifically configured to: based on the trained ensemble tree model, allocate each positive sample and each negative sample to different leaf nodes, and based on the path from the root node to the leaf node, obtain the at least one set of candidate feature combinations.

11. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1-9.

Images