A multimedia information recommendation method and device

Abstract

The application relates to the computer technical field, in particular to a multimedia information recommendation method and device, wherein the method comprises the following steps: in response to a play request carrying object basic attributes and play conditions of a play object, obtaining a plurality of candidate multimedia information and extracting respective recommendation features; for one of the candidate multimedia information, according to a set information identifier, selecting at least one key information from the object basic attributes, the play conditions and the play content of the one candidate multimedia information and extracting key features; based on the key features, adjusting the weights of elements contained in the recommendation features of the one candidate multimedia information, and based on the weight adjustment result, obtaining an estimated conversion rate of the one candidate multimedia information; until at least one estimated conversion rate is obtained, selecting a recommended multimedia information from the at least one candidate multimedia information. The above method is used to improve the accuracy and efficiency of the play recommendation system in recommending multimedia resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a multimedia information recommendation method and device. BACKGROUND

[0002] At present, a playing application can obtain multimedia resources (such as promotional short videos, music, advertisements, etc.) to be played through a playing recommendation system, and the playing recommendation system can perform resource recommendation according to a playing reference condition carried in a multimedia resource recommendation request triggered by the playing application, so as to ensure the playing effect of the multimedia resources; wherein the playing reference condition includes but is not limited to: request triggering place, multimedia resource display position ID, basic information of a user of the playing application, etc.

[0003] In the related art, in the playing recommendation system, the playing application performs a resource recommendation request by accessing a software development kit (sdk) provided by the playing recommendation system. After the sdk collects the playing condition carried in the resource recommendation request, the sdk sends a multimedia resource request to a background server. The background server allocates each multimedia resource required by the playing application according to the received multimedia resource request and the basic information of the user of the playing application, and further obtains an estimated conversion rate generated by the playing application playing each multimedia resource; wherein the estimated conversion rate represents the conversion rate from the user of the playing application clicking the multimedia resource to becoming an effective active, registered, and paid user, and the calculation method is: conversion rate = (number of converted users / click amount of users clicking multimedia resources) * 100%. The background server generates a return result according to the multimedia resource whose estimated conversion rate reaches a set threshold, and returns the return result to the sdk of the playing application. The playing application can display the corresponding multimedia resource to the user through the return result.

[0004] However, the above-mentioned method of recommending multimedia resources, although it realizes the selective recommendation of multimedia resources for the playing application through the playing recommendation system, ignores the difference in the influence degree of different data on the conversion rate in the processing process of the playing reference condition of the playing application, the basic information of the user, and each multimedia resource allocated to the playing application, such as the influence degree of the user behavior (such as clicking, registering, paying, etc.) of the user on the conversion of the multimedia resource is often greater than that of other data, so that the accuracy of the estimated conversion rate obtained is poor, and many multimedia resource withdrawal operations are generated, so that the withdrawal of the multimedia resource seriously consumes the computing resources of the playing recommendation system and the media resources of the playing application, and also reduces the recommendation efficiency of the playing recommendation system.

[0005] Therefore, there is an urgent need for a multimedia information recommendation method and device to improve the accuracy and efficiency of the multimedia resource recommendation of the play recommendation system, save the computing resources of the play recommendation system and the media resources of the play application. SUMMARY

[0006] Embodiments of the present application provide a multimedia information recommendation method and device to improve the accuracy and efficiency of the multimedia resource recommendation of the play recommendation system, save the computing resources of the play recommendation system and the media resources of the play application.

[0007] In a first aspect, embodiments of the present application provide a multimedia information recommendation method, which comprises:

[0008] In response to a play request carrying object basic attributes and play conditions of a play object, obtaining a plurality of candidate multimedia information to be recommended, and extracting respective recommendation features of the plurality of candidate multimedia information;

[0009] For the plurality of candidate multimedia information, the following operations are respectively performed:

[0010] According to at least one information identifier, at least one key information is selected from the object basic attributes, the play conditions and the play content of one candidate multimedia information, and the key features of the at least one key information are extracted; each information identifier represents one kind of information that causes the historical estimated conversion rate to deviate;

[0011] Based on the key features, the weights of elements contained in the recommendation features of the one candidate multimedia information are adjusted, and based on the weight adjustment result, the estimated conversion rate of the one candidate multimedia information is obtained;

[0012] Based on the obtained at least one estimated conversion rate, a recommended multimedia information is selected from the at least one candidate multimedia information.

[0013] Optionally, in response to the play request carrying the object basic attributes and the play conditions of the play object, obtaining the plurality of candidate multimedia information to be recommended, and extracting the respective recommendation features of the plurality of candidate multimedia information, comprises:

[0014] In response to the play request, a recall algorithm is used to obtain a plurality of candidate multimedia information meeting the object basic attributes and the play conditions;

[0015] For the plurality of candidate multimedia information, the following operations are respectively performed:

[0016] Respectively extracting features from the object basic attributes, the play conditions and the play content of the one candidate multimedia information to obtain corresponding object features, condition features and content features;

[0017] crossing the object feature, the condition feature and the content feature to obtain the recommendation feature of the one candidate multimedia information.

[0018] Optionally, the extracting the key feature of the at least one key information comprises:

[0019] respectively extracting a feature of each of the at least one key information to obtain at least one preliminary key feature;

[0020] crossing the at least one preliminary key feature to obtain the key feature.

[0021] Optionally, the adjusting the weight of each element contained in the recommendation feature of the one candidate multimedia information based on the key feature, and obtaining the estimated conversion rate of the one candidate multimedia information based on the weight adjustment result comprises:

[0022] respectively performing twice mapping processing on the key feature to obtain a first weight feature and a second weight feature, the first weight feature and the second weight feature being respectively used to represent the weight of the associated element contained in the recommendation feature of the one candidate multimedia information and the key feature;

[0023] adjusting the weight of each element contained in the recommendation feature of the one candidate multimedia information according to the first weight feature, and performing classification mapping processing on the recommendation feature after the weight adjustment to obtain a preliminary estimated feature;

[0024] adjusting the weight of each element contained in the preliminary estimated feature according to the second weight feature, and performing classification mapping processing on the preliminary estimated feature after the weight adjustment to obtain the estimated conversion rate of the one candidate multimedia information.

[0025] Optionally, the method further comprises: determining a sparse recommendation feature according to the object basic attribute, the playing condition and the playing content of the one candidate multimedia information, the sparse recommendation feature representing the words of the object basic attribute, the playing condition and the playing content of the one candidate multimedia information respectively.

[0026] the performing classification mapping processing on the preliminary estimated feature after the weight adjustment to obtain the estimated conversion rate of the one candidate multimedia information comprises:

[0027] performing classification mapping processing on the preliminary estimated feature after the weight adjustment to obtain an intermediate estimated feature;

[0028] weighting the sparse recommendation feature and the intermediate estimation feature to obtain a final estimation feature, and obtaining the estimated conversion rate of the one candidate multimedia information based on the final estimation feature.

[0029] Optionally, the classifying and mapping the weight-adjusted recommendation feature to obtain the preliminary estimation feature comprises:

[0030] normalizing the weight-adjusted recommendation feature to obtain a standard recommendation feature, and classifying and mapping the standard recommendation feature to obtain the preliminary estimation feature;

[0031] The classifying and mapping the weight-adjusted preliminary estimation feature to obtain the estimated conversion rate of the one candidate multimedia information comprises:

[0032] normalizing the weight-adjusted preliminary estimation feature to obtain a standard preliminary estimation feature, and classifying and mapping the standard preliminary estimation feature to obtain the estimated conversion rate of the one candidate multimedia information.

[0033] Optionally, the classifying and mapping the key feature twice to obtain the first weight feature and the second weight feature comprises:

[0034] normalizing the key feature twice to obtain a first standard key feature and a second standard key feature, respectively, and classifying and mapping the first standard key feature and the second standard key feature to obtain a corresponding first preliminary weight feature and a second preliminary weight feature, respectively;

[0035] multiplying the first preliminary weight feature and the second preliminary weight feature by a set multiple to obtain the first weight feature and the second weight feature, respectively.

[0036] In the above method, the key feature is normalized to obtain the first standard key feature and the second standard key feature, respectively, which can prevent model overfitting and quickly obtain the first weight feature and the second weight feature of the one candidate multimedia information in a short time, improving the estimation efficiency. In the multiplication of the first preliminary weight feature and the second preliminary weight feature by the set multiple to obtain the first weight feature and the second weight feature, the set multiple can be set according to the influence degree of the estimated conversion rate of the corresponding element in the preliminary weight feature, so that each element in the first weight feature and the second weight feature obtained after multiplication by the set multiple can accurately represent the influence degree of the estimated conversion rate of the one candidate multimedia information, further improving the accuracy of obtaining the estimated conversion rate of the one candidate multimedia information.

[0037] Optionally, the at least one key information comprises one or more of a conversion indicator, a multimedia object category, and a play scene type.

[0038] The conversion indicator represents an object behavior in the object basic attribute that causes the multimedia information to be converted to generate the asset corresponding to the multimedia information.

[0039] The multimedia object category represents an object category of a multimedia object in the play content that is presented to the play object by the multimedia information.

[0040] The play scene type represents a page scene type in which the play object accesses a page to display the multimedia information in the play condition.

[0041] In a second aspect, an embodiment of the present application provides a multimedia information recommendation device, comprising:

[0042] A first feature extraction unit is configured to, in response to a play request carrying an object basic attribute of a play object and a play condition, acquire a plurality of candidate multimedia information to be recommended, and extract a recommendation feature of each of the plurality of candidate multimedia information.

[0043] For each of the plurality of candidate multimedia information, the following operations are performed:

[0044] A second feature extraction unit is configured to, according to at least one information identifier, select at least one key information from the object basic attribute, the play condition, and play content of one candidate multimedia information, and extract a key feature of the at least one key information. Each information identifier represents one type of information that causes a historical estimated conversion rate to deviate.

[0045] An estimation processing unit is configured to, based on the key feature, adjust a weight of each element contained in the recommendation feature of the one candidate multimedia information, and obtain an estimated conversion rate of the one candidate multimedia information based on a weight adjustment result.

[0046] An information selection unit is configured to, based on the obtained at least one estimated conversion rate, select a recommended multimedia information from the at least one candidate multimedia information.

[0047] Optionally, the first feature extraction unit is specifically configured to, in response to the play request, acquire the plurality of candidate multimedia information that meets the object basic attribute and the play condition by using a recall algorithm.

[0048] For each of the plurality of candidate multimedia information, the following operations are performed:

[0049] respectively, the object basic attribute, the playing condition and the playing content of the one candidate multimedia information are subjected to feature extraction to obtain corresponding object feature, condition feature and content feature;

[0050] The object feature, the condition feature and the content feature are subjected to feature cross processing to obtain the recommendation feature of the one candidate multimedia information.

[0051] Optionally, the second feature extraction unit is specifically configured to respectively subject each of the at least one key information to feature extraction to obtain at least one preliminary key feature;

[0052] The at least one preliminary key feature is subjected to feature cross processing to obtain the key feature.

[0053] Optionally, the estimation processing unit is specifically configured to subject the key feature to twice mapping processing respectively to obtain a first weight feature and a second weight feature, the first weight feature and the second weight feature being respectively used to represent the weight of the associated elements contained in the recommendation feature of the one candidate multimedia information and the key feature;

[0054] According to the first weight feature, the elements contained in the recommendation feature of the one candidate multimedia information are subjected to weight adjustment, and the recommendation feature after weight adjustment is subjected to classification mapping processing to obtain a preliminary estimation feature;

[0055] According to the second weight feature, the elements contained in the preliminary estimation feature are subjected to weight adjustment, and the preliminary estimation feature after weight adjustment is subjected to classification mapping processing to obtain the estimated conversion rate of the one candidate multimedia information.

[0056] Optionally, the estimation processing unit is further configured to determine a sparse recommendation feature according to the object basic attribute, the playing condition and the playing content of the one candidate multimedia information, the sparse recommendation feature representing the respective words of the object basic attribute, the playing condition and the playing content of the one candidate multimedia information.

[0057] Then the estimation processing unit is specifically configured to subject the preliminary estimation feature after weight adjustment to classification mapping processing to obtain an intermediate estimation feature;

[0058] The sparse recommendation feature and the intermediate estimation feature are weighted to obtain a final estimation feature, and based on the final estimation feature, the estimated conversion rate of the one candidate multimedia information is obtained.

[0059] Optionally, the estimation processing unit is specifically configured to normalize the recommended features after the weight adjustment to obtain standard recommended features, and perform classification mapping processing on the standard recommended features to obtain the preliminary estimation features.

[0060] The estimation processing unit is specifically configured to normalize the preliminary estimation features after the weight adjustment to obtain standard preliminary estimation features, and perform classification mapping processing on the standard preliminary estimation features to obtain the estimation conversion rate of the one candidate multimedia information.

[0061] Optionally, the estimation processing unit is specifically configured to perform twice normalization processing on the key features respectively to obtain first standard key features and second standard key features, and perform classification mapping processing on the first standard key features and the second standard key features respectively to obtain corresponding first preliminary weight features and second preliminary weight features.

[0062] The first preliminary weight features and the second preliminary weight features are multiplied by a set multiple respectively to obtain the first weight features and the second weight features.

[0063] Optionally, the at least one key information includes one or more of a conversion indicator, a multimedia object category, and a play scene type.

[0064] The conversion indicator represents an object behavior in the object basic attribute that causes the multimedia information to be converted to generate the multimedia information corresponding asset.

[0065] The multimedia object category represents an object category of a multimedia object displayed by the multimedia information to the play object in the play content.

[0066] The play scene type represents a page scene type in which the play object accesses a page to display the multimedia information in the play condition.

[0067] In a third aspect, an embodiment of the present application provides a computer device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes any one of the multimedia information recommendation methods in the first aspect.

[0068] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, including a computer program, and when the computer program runs on a computer device, the computer program is used to make the computer device execute any one of the multimedia information recommendation methods in the first aspect.

[0069] In a fifth aspect, an embodiment of the present application provides a computer program product. The computer program product includes a computer program stored in a computer readable storage medium. When a processor of a computer device reads the computer program from the computer readable storage medium, the processor executes the computer program, so that the computer device executes any one of the multimedia information recommendation methods in the first aspect.

[0070] The present application has the following beneficial effects:

[0071] The multimedia information recommendation method, device, computer device and storage medium provided by the embodiments of the present application, after the playing recommendation system receives a playing request carrying object basic attributes of a playing object and playing conditions, the playing recommendation system acquires a plurality of candidate multimedia information to be recommended in response to the playing request. In this way, the multimedia information to be recommended is preliminarily screened according to the playing request, and a plurality of candidate multimedia information to be recommended is obtained. Subsequently, each candidate multimedia information is estimated, and the recommended candidate multimedia information is determined according to the estimated conversion rate, so as to prevent unnecessary waste of computing resources caused by estimating a large number of multimedia information respectively without preliminary screening of the multimedia information. The recommendation features of each candidate multimedia information in the plurality of candidate multimedia information are acquired respectively, so that the recommendation features of one candidate multimedia information can represent the features of the one candidate multimedia information. According to the set information identifier, the key information corresponding to the information identifier is selected from the object basic attributes, the playing conditions and the playing content of the one candidate multimedia information. The information identifier represents one kind of information that causes the historical estimated conversion rate to deviate. In this way, at least one key information is selected according to at least one information identifier, and the key features of the at least one key information are extracted. The key features correspond to the features of at least one kind of information that causes the historical estimated conversion rate to deviate. Based on the key features, the weights of the elements in the recommendation features of the candidate multimedia information of the key features are adjusted, so that the corresponding key features in the recommendation features perform more accurately in the estimated conversion rate of the candidate multimedia information, and the accuracy of the estimated conversion rate of the multimedia information is improved.

[0072] Other features and advantages of the present application will be further described in the following description, and will become apparent from the description, or be learned through practice of the present application. The objects and other advantages of the present application can be achieved and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0073] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description only relate to the embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of the provided drawings.

[0074] Figure 1 An optional schematic diagram of an application scenario provided by the embodiments of the present application;

[0075] Figure 2 A flowchart of a multimedia information recommendation method provided by the embodiments of the present application;

[0076] Figure 3 A flowchart of a multimedia information recommendation method provided by the embodiments of the present application;

[0077] Figure 4A A schematic diagram of a feature extraction method of an object basic attribute provided by the embodiments of the present application;

[0078] Figure 4B A schematic diagram of a feature extraction method of a playing condition provided by the embodiments of the present application;

[0079] Figure 4C A schematic diagram of a feature extraction method of a playing content provided by the embodiments of the present application;

[0080] Figure 5A A schematic diagram of a feature extraction method of an object basic attribute provided by the embodiments of the present application;

[0081] Figure 5B A schematic diagram of a feature extraction method of a playing condition provided by the embodiments of the present application;

[0082] Figure 5C A schematic diagram of a feature extraction method of a playing content provided by the embodiments of the present application;

[0083] Figure 6 A schematic diagram of a feature cross processing method provided by the embodiments of the present application;

[0084] Figure 7 A schematic diagram of a feature extraction method of a key feature provided by the embodiments of the present application;

[0085] Figure 8 A schematic diagram of a feature extraction method of a key feature provided by the embodiments of the present application;

[0086] Figure 9 A flowchart of a multimedia information recommendation method provided by the embodiments of the present application;

[0087] Figure 10 A method diagram for mapping processing of key features based on a logistic regression algorithm is provided for the embodiments of the present application;

[0088] Figure 11 A method diagram for mapping processing of key features based on a deep neural network method is provided for the embodiments of the present application;

[0089] Figure 12 A method diagram for classification mapping processing of recommended features and preliminary estimated features of a candidate multimedia information based on a logistic regression algorithm is provided for the embodiments of the present application;

[0090] Figure 13 A method diagram for classification mapping processing of recommended features and preliminary estimated features of a candidate multimedia information based on a deep neural network method is provided for the embodiments of the present application;

[0091] Figure 14 A method diagram for obtaining first weight features and second weight features is provided for the embodiments of the present application;

[0092] Figure 15 A feature extraction method diagram for sparse recommended features is provided for the embodiments of the present application;

[0093] Figure 16 A multimedia information recommendation method diagram is provided for the embodiments of the present application;

[0094] Figure 17 A multimedia information recommendation method diagram is provided for the embodiments of the present application;

[0095] Figure 18 A multimedia information recommendation method diagram is provided for the embodiments of the present application;

[0096] Figure 19 A device diagram of a multimedia information recommendation device is provided for the embodiments of the present application;

[0097] Figure 20 A hardware component structure diagram of a computer device applying the embodiments of the present application is provided;

[0098] Figure 21 A hardware component structure diagram of another computer device applying the embodiments of the present application is provided. DETAILED DESCRIPTION

[0099] In order to make the purposes, technical solutions and advantages of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application. The embodiments in the present application and the features in the embodiments can be combined with each other in a non-conflicting manner. Moreover, although a logical order is shown in the flowchart, in some cases, the steps shown or described can be performed in an order different from that shown.

[0100] It can be understood that, in the following specific embodiments of the present application, data related to object basic attributes and the like are involved, and when the embodiments of the present application are applied to specific products or technologies, relevant permissions or consents need to be obtained, and the collection, use and processing of the relevant data need to comply with relevant laws, regulations and standards of countries and regions. For example, when relevant data needs to be obtained, relevant volunteers can be recruited and sign a volunteer authorization data agreement, and then the data of the volunteers can be used for implementation; or implementation is performed within the internal scope of an organization that has been authorized to allow, and the implementation of the following embodiments is performed by using the data of internal members to make relevant recommendations to the internal members; or the relevant data used in the specific implementation is all simulation data, for example, simulation data generated in a virtual scene.

[0101] In order to facilitate understanding of the technical solutions provided by the embodiments of the present application, some key terms used by the embodiments of the present application are explained first:

[0102] Recommendation algorithm: that is, Wide&Deep model. The core idea of the Wide&Deep model is to combine the memorization of a linear model (Wide) and the generalization of a deep neural network model (Deep), and to optimize the parameters of the two models in the training process, so as to achieve the optimal prediction ability of the overall model. The memorization is to find the correlation between items (a piece of data) or features from historical data. The generalization is the transmission of correlation, and new feature combinations that rarely or never appear in historical data are found.

[0103] Deep neural network (Deep Neural Networks, DNN): a neural network with a multi-layer network architecture.

[0104] Fully connected layer: is every node connected with all nodes of the previous layer, used to synthesize the features extracted in the front. Due to its full connection characteristics, the parameters of the fully connected layer are the most, which can reduce the influence of feature position on the classification result and improve the robustness of the entire deep neural network.

[0105] Regularization layer: including LN (layer normalization), which is proposed for natural language processing, is a method to convert input data into mean 0 and variance 1. Normalization (or normalization) is generally performed before the data is sent to the activation function, and the purpose is to hope that the input data will not fall into the saturation area of the activation function. Alleviate the gradient vanishing / gradient explosion phenomenon in DNN training, and accelerate the training speed of the model.

[0106] Activation function (Activation Function), a function running on the neuron of artificial neural network, responsible for mapping the input of the neuron to the output. Sigmoid function is often used as the activation function of neural network, which maps the variable to 0,1, which is a bilateral saturated activation function. ReLU function is used as the activation function of neural network, the gradient is constant when greater than 0, and the derivative of relu function is 0 when less than 0, so once the neuron activation value enters the negative half zone, the gradient will be 0, and the neuron will not undergo training. Only when the neuron activation value enters the positive half zone, there will be a gradient value, and the neuron will be trained at this time.

[0107] One-hot encoding, also known as one-bit effective encoding, is to use N-bit state registers to encode N states, each state has its own register bit, and only one bit is valid at any time. It can be used to convert classification data that computers cannot recognize into vectors containing only "0" and / or "1" that computers can recognize. For example, the classification data

male (gender), Z country (nationality), 20 (age)

[0108] Sparse feature: after one-hot encoding of information data of different categories, the obtained feature data will become sparse. For example, the feature vector [011000001] obtained by one-hot encoding of the above three classification data [male (gender), Z country (nationality), 20 (age)], the data sparsity of the gender dimension is one-third. If there are 100,000 items (one classification data), if the one-hot encoding is performed on the dimension of the item, the data sparsity of this dimension is one in 1,000,000. Therefore, the feature vector obtained by one-hot encoding is a sparse feature with sparsity.

[0109] Dense feature: compared with the sparse feature, the dense feature does not need to be normalized, and directly collects the size of the corresponding classification data in the corresponding dimension position, so as to more comprehensively represent the semantic information of all classification data. For example, the dense feature [1.7] of the classification data height can be obtained when the height of the classification data is 1.70.

[0110] FM (Factorization Machine, perceptual factorization machine): each feature is expressed by an embedding vector. FM is mainly used to solve the feature combination problem in the case of data sparsity.

[0111] FFM (Field-aware Factorization Machine, field-aware factorization machine): high-dimensional sparse features are grouped according to the source-field, and then all features in the group are used to jointly construct an embedding, so as to achieve the effect of feature dense dimension reduction, thereby reducing the input size of the model and effectively reducing the model parameters.

[0112] Play object: the play object involved in the embodiments of the present application can be a subject that performs interactive behavior in the client, for example, the play object can be an account in the client, or the play object can be a page accessed in the client.

[0113] Object basic attribute: the object basic attribute in the embodiments of the present application can be basic information of the play object provided by the client, for example, the object basic attribute is information such as gender, age, occupation, address, and object behavior (such as clicking, collecting, and paying) of the corresponding account.

[0114] Object feature: the object feature in the embodiments of the present application can be obtained by refining the object basic attribute of the corresponding account, for example, the object feature is obtained by refining the object basic attribute of the corresponding account, such as gender, age, occupation, address, and object behavior (such as clicking, collecting, and paying).

[0115] Page scene type: The type of the page scene corresponding to the page accessed by the playing object in the embodiments of the present application, for example, the page scene of the page accessed is a game, the corresponding page scene type is a game scene, the page scene of the page accessed is a video, the corresponding page scene type is a video scene, the page scene of the page accessed is shopping, the corresponding page scene type is a shopping scene, and the like.

[0116] Playing condition: The playing condition of the recommended multimedia information played by the corresponding account in the playing request can be set by the client in the embodiments of the present application, for example, the request occurrence place of the playing request, the playing address when the multimedia information is played in the client, and the like.

[0117] Condition feature: The condition feature in the embodiments of the present application can be obtained by extracting the playing condition of the recommended multimedia information played by the corresponding account in the playing request, for example, the condition feature is obtained by extracting the playing condition: the request occurrence place of the playing request, the playing address when the multimedia information is played in the client, and the like.

[0118] Playing content: The playing content in the embodiments of the present application is the content displayed to the playing object in the client by the candidate multimedia information, for example, the multimedia object in the candidate multimedia information, the category of the candidate multimedia information is a picture, a video, text, and the like, the name of the delivery party of the candidate multimedia information, and the like. The multimedia object can be an object that the playing object prefers to understand, for example, an article, a landscape, a tool, and the like.

[0119] Content feature: The content feature in the embodiments of the present application can be obtained by extracting the playing content displayed to the playing object in the client by the candidate multimedia information, for example, the multimedia object in the candidate multimedia information, the category of the candidate multimedia information is a picture, a video, text, and the like, the name of the delivery party of the candidate multimedia information, and the like.

[0120] Recommendation feature: The recommendation feature of the candidate multimedia information in the embodiments of the present application is obtained by combining the object feature, the condition feature, and the content feature of the candidate multimedia information. The object feature represents the semantic information of the basic attribute of the object, the condition feature represents the semantic information of the playing condition, and the content feature represents the playing content of the candidate multimedia information.

[0121] Weight feature: Or the preference weight feature, each element in the weight feature represents the weight of the corresponding element in the candidate multimedia information preferred in the pre-estimation processing, in other words, each element in the weight feature represents the overall preference of the corresponding element in the candidate multimedia information played by the playing object, the higher the weight value, the greater the possibility that the candidate multimedia information is preferred by the playing object, that is, the higher the possibility that the candidate multimedia information meets the preference demand of the playing object.

[0122] The design idea of the embodiments of the present application is briefly introduced as follows:

[0123] In the playing service of the client, the playing object can obtain playing resources by accessing the background server, such as obtaining page video resources, page text resources, page picture resources and the like of the corresponding access page by accessing the background server. The corresponding multimedia information can also be recommended to the playing object, and the recommended multimedia information is played to the playing object in the page scene of the access page, so as to improve the use rate of the recommended multimedia information by the playing object.

[0124] In the related art, in order to improve the use rate of the recommended multimedia information by the playing object, the background server receives the playing request sent by the client through the playing recommendation system, performs recall algorithm according to the related information carried in the playing request, filters the multimedia information to be recommended for the playing request, and then uses the recommendation algorithm to estimate the estimated conversion rate of each multimedia information to be recommended. The multimedia information with the highest estimated conversion rate or the multimedia information to be recommended with the estimated conversion rate meeting the set conversion rate threshold is returned to the client, and the client plays the recommended multimedia information to the playing object in the page scene of the access page. However, due to the generalization ability of the recommendation algorithm, the preferred elements of the playing object in the recommended features are generalized in the overall recommended features, which leads to deviation of the estimated conversion rate of the recommended multimedia information. Finally, the conversion rate of the multimedia information recommended to the client may be lower than the actual conversion rate of the multimedia information, the target asset generated by the conversion of the recommended multimedia information cannot be obtained due to the estimated conversion rate, and the recommended multimedia information also leads to waste of recommended multimedia information processing computing resources, invalid occupation of playing resources of the client by the recommended multimedia information, and waste of computing resources for withdrawing the recommended multimedia information.

[0125] In view of this, the embodiment of the present application provides a multimedia information recommendation method and device, computer equipment and a storage medium. After the playing recommendation system receives a playing request carrying object basic attributes and playing conditions of a playing object, the playing recommendation system acquires a plurality of candidate multimedia information to be recommended in response to the playing request. In this way, the multimedia information to be recommended is preliminarily screened according to the playing request, and a plurality of candidate multimedia information to be recommended is obtained. Subsequently, each candidate multimedia information is estimated, and the recommended candidate multimedia information is determined according to the estimated conversion rate, thereby preventing unnecessary waste of computing resources caused by estimating a large number of multimedia information without preliminary screening. The recommendation features of each candidate multimedia information in the plurality of candidate multimedia information are acquired respectively, so that the recommendation features of one candidate multimedia information can represent the features of the one candidate multimedia information. According to the set information identifier, the key information corresponding to the information identifier is selected from the object basic attributes, the playing conditions and the playing content of the one candidate multimedia information. The information identifier represents one kind of information that causes the historical estimated conversion rate to deviate. In this way, at least one key information is selected according to at least one information identifier, and the key features of the at least one key information are extracted. The key features correspond to the features of at least one kind of information that causes the historical estimated conversion rate to deviate. Based on the key features, the weights of the elements in the recommendation features of the candidate multimedia information of the key features are adjusted, so that the corresponding key features in the recommendation features are more accurate in the estimated conversion rate of the candidate multimedia information, and the accuracy of the estimated conversion rate of the multimedia information is improved.

[0126] The preferred embodiments of the present application are described below in conjunction with the accompanying drawings of the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application, and the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0127] As shown in Figure 1 , it is an application scenario diagram of the embodiment of the present application. The application scenario diagram includes any one of a plurality of terminal devices 110 and any one of a plurality of servers 120.

[0128] In the embodiments of the present application, the terminal device 110 includes but is not limited to a mobile phone, a tablet computer, a notebook computer, a desktop computer, an electronic book reader, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, and the like. The terminal device can be installed with a client related to the recommended multimedia information service. The client can be a software (such as a browser, a communication software, and the like), a webpage, an applet, or the like. The server 120 is a background server corresponding to the software or the webpage, the applet, or the like, or a background server specially used for recommending multimedia information to the client, which is not limited in the present application. The server 120 can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content distribution networks (CDN), and basic cloud computing services such as big data and artificial intelligence platforms.

[0129] It should be noted that the multimedia information recommendation method in the embodiments of the present application can be executed by a computer device, which can be the server 120 or the terminal device 110, that is, the method can be executed by the server 120 or the terminal device 110 alone, or by the server 120 and the terminal device 110 together. For example, when executed by the terminal device 110 and the server 120 together, the client in the terminal device 110 generates a play request according to the object basic attribute of the play object and the play condition of the client to the multimedia information when the multimedia information is needed, and sends the play request to the server 120. The server 120 receives the play request sent by the client, determines a plurality of candidate multimedia information to be recommended for the play request according to the object basic attribute and the play condition of the play object carried in the play request, extracts the recommendation features of the plurality of candidate multimedia information to be recommended, for any one of the plurality of candidate multimedia information to be recommended, selects at least one key information from the object basic attribute, the play condition, and the play content of the one candidate multimedia information according to at least one information identifier set, extracts the key features of the at least one key information, adjusts the weights of the elements contained in the recommendation features of the one candidate multimedia information based on the key features, and obtains the estimated conversion rate of the one candidate multimedia information based on the weight adjustment result, and then obtains the estimated conversion rate of each of the plurality of candidate multimedia information to be recommended, selects the recommended multimedia information from the at least one candidate multimedia information, generates a return result according to the recommended multimedia information, and returns to the client of the terminal device 110 to play to the play object.

[0130] It should be noted that, Figure 1The number of terminal devices and servers and the communication manner are not limited in fact, and are not specifically limited in the embodiments of the present application.

[0131] In addition, the multimedia information recommendation provided in the embodiments of the present application can be applied to various scenes, such as short video recommendation, news recommendation, novel recommendation, advertisement recommendation, and coupon recommendation.

[0132] The multimedia information recommendation method provided in the exemplary embodiments of the present application will be described below in combination with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above-mentioned application scenarios are only shown for the purpose of facilitating understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this respect.

[0133] Referring to Figure 2 FIG. 1 shows a flowchart of the multimedia information recommendation method provided in the embodiments of the present application, which is exemplarily described taking a server as an execution subject. The specific implementation process of the method is as follows.

[0134] Step 201: In response to a play request carrying object basic attributes of a play object and play conditions, a plurality of candidate multimedia information to be recommended is obtained, and recommendation features of the plurality of candidate multimedia information are extracted respectively;

[0135] In the embodiments of the present application, the play object can access a page when using a client, triggering the client to collect basic information of an account of the client corresponding to the play object to obtain the object basic attributes, which can include: an account identifier, age, gender, city, member level, access time, access frequency, conversion indicators of the play object to recommended multimedia information (such as registration, collection, payment, etc.), and the like. The information contained in the object basic attributes is not specifically limited here. The client is also triggered to collect play conditions corresponding to the page accessed by the play object, which can include: a request occurrence place of the play request, a play address of the recommended multimedia information provided by the client, a play time of the recommended multimedia information provided by the client, a time for waiting for the play object to perform a triggering behavior (object behavior corresponding to the conversion indicator, such as clicking, collecting, paying, registering, etc.) on the recommended multimedia information provided by the client, and the like. The specific setting of the play conditions is not limited here.

[0136] Step 202: For the plurality of candidate multimedia information, the following operations are performed respectively:

[0137] Step 2021: According to at least one information identifier, at least one key information is selected from the object basic attributes, the play conditions, and the play content of one candidate multimedia information, and key features of the at least one key information are extracted; each information identifier represents one kind of information that causes the historical estimated conversion rate to deviate;

[0138] In the embodiments of the present application, the at least one information identifier set can be an information identifier of information causing the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information, which is obtained according to the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information. For example, in a historical recommended multimedia information set in which the object basis information is basically the same and the playing conditions are basically the same, if the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information with different multimedia object types in the playing content is obvious, the information identifier of the multimedia object type can be set as one of the at least one information identifier set.

[0139] For the convenience of understanding, in an example, in a historical recommended multimedia information set, if the account identifier of the playing object, the age, the gender, the city, the access time, the access frequency, and the conversion indicator of the recommended multimedia information in the object basis attribute are basically the same, the playing request request occurrence place of the client, the playing address of the recommended multimedia information provided by the client, and the information of the game page scene type of the client are basically the same. In the historical recommended multimedia information set, the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information with the vehicle category as the multimedia object category in the playing content is small, and the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information with the mobile terminal as the multimedia object category in the playing content is large, that is, different multimedia object categories have a greater impact on the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information, the information identifier of the multimedia object category can be set as one of the at least one information identifier set.

[0140] In the embodiments of the present application, the at least one information identifier set can be an information identifier of information causing the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information, which is obtained according to the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information. For example, in a historical recommended multimedia information set in which the object basis information is basically the same, the playing conditions are basically the same, and the playing content is basically the same, if any one of the conversion indicator in the object basis attribute, the playing scene type in the playing condition, and the multimedia object category of each historical recommended multimedia information is different, it will cause the obvious change of the deviation between the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information, and the information identifier of the information of the conversion indicator, the playing scene type, and the multimedia object type can be set as the at least one information identifier set.

[0141] For ease of understanding, in an example, in a set of historical recommended multimedia information, if the object basic attribute of each historical recommended multimedia information includes the account identifier of the playing object: 001, age: 17, gender: male, city: BJ, and access frequency: more than 3 times are basically the same, the conversion index of the recommended multimedia information in the object basic attribute - object behavior includes: registration, click, collection, and payment. The playing condition of the client of each historical recommended multimedia information is basically the same, the request occurrence place of the playing condition of the client of the playing request is XX city-XX street, and the playing address *** of the recommended multimedia information provided by the client is basically the same. The playing scene type of the client in the playing condition includes: game page scene type, shopping page scene type, and video page scene type. The playing content of each historical recommended multimedia information is basically the same, the historical recommended multimedia information category in the playing content is video, and the historical recommended multimedia information material is a physical display. The multimedia object category in the playing content includes: vehicle category, terminal category, and food category. In the set of historical recommended multimedia information, any one of the information contents in the conversion index, the playing scene type, and the multimedia information object category in each historical recommended multimedia information is different, which will cause the deviation of the estimated conversion rate and the actual conversion rate of the historical recommended multimedia information to change obviously, and then the information identifier of the information of the conversion index, the playing scene type, and the multimedia object type can be set as the at least one information identifier.

[0142] Correspondingly, in the above example, the information corresponding to the set information identifier in the historical recommended multimedia information is key information, for example, the key information corresponding to the information identifier of the conversion index - object behavior is registration, click, payment, etc., the key information corresponding to the information identifier of the playing scene type is game page scene type, shopping page scene type, video page scene type, etc., and the key information corresponding to the information identifier of the multimedia object category is vehicle category, terminal category, food category, etc.

[0143] Step 2022: based on the key feature, adjusting the weight of each element contained in the recommendation feature of the one candidate multimedia information, and obtaining the estimated conversion rate of the one candidate multimedia information based on the weight adjustment result;

[0144] In the embodiment of the application, the key feature can adjust the weight of the corresponding element in each element contained in the recommendation feature of one candidate multimedia information, or the key feature can adjust the weight of all or part of the elements contained in the recommendation feature of one candidate multimedia information. Here, the specific adjustment mode of the weight of each element contained in the recommendation feature of one candidate multimedia information is not limited.

[0145] In an example, assuming that the at least one key information includes a conversion indicator: registration, a play scene type: game page scene type, and a multimedia object type: mobile terminal, the at least one key feature can correspondingly represent key features of the conversion indicator: registration, the play scene type: game page scene type, and the multimedia object type: mobile terminal. Further, the key features of registration, the game page scene type, and the mobile terminal can adjust weights of elements corresponding to registration, the game page scene type, and the mobile terminal in the recommendation features of a candidate multimedia information.

[0146] In an example, assuming that the at least one key information includes a conversion indicator: registration, a play scene type: game page scene type, and a multimedia object type: mobile terminal, the at least one key feature can correspondingly represent key features of the conversion indicator: registration, the play scene type: game page scene type, and the multimedia object type: mobile terminal. Further, the key features of registration, the game page scene type, and the mobile terminal can adjust weights of elements corresponding to registration, the game page scene type, and the mobile terminal in the recommendation features of a candidate multimedia information.

[0147] In an example, assuming that the at least one key information includes a conversion indicator: registration, a play scene type: game page scene type, and a multimedia object type: mobile terminal, the at least one key feature can correspondingly represent key features of the conversion indicator: registration, the play scene type: game page scene type, and the multimedia object type: mobile terminal. Further, the key features of registration, the game page scene type, and the mobile terminal can adjust weights of elements corresponding to registration, the game page scene type, and the mobile terminal in the recommendation features of a candidate multimedia information. Specifically, the weights of all or part of elements corresponding to registration, play object age, play object gender, play object education, play object occupation, the game page scene type, a play request occurrence location, and the mobile terminal can be adjusted according to an association relationship between the elements. For example, a key feature of a conversion indicator of payment can be associated with an age, an occupation, and the like of a play object.

[0148] In an example, assuming that the at least one key information includes the conversion indicator: collection, and the play scene type: novel page scene type, and the multimedia object type: novel, and the related material information: documentary in the candidate multimedia information, then correspondingly, the at least one key feature can represent the key features of the conversion indicator: collection, and the play scene type: novel page scene type, and the multimedia object type: novel, and the related material information: documentary in the candidate multimedia information. Further, the key features of collection, novel page scene type, novel, and documentary can adjust the weights of all elements or part of elements of the recommendation features of a candidate multimedia information corresponding to collection, play object age, play object gender, play object education, play object major, play object occupation, novel page scene type, play request location, novel, history, and documentary, etc. Specifically, the association between elements can also be set, for example, the key feature of the related material information: documentary in the candidate multimedia information can be associated with the major and occupation of the play object.

[0149] The weight adjustment result can be obtained after the recommendation features of a candidate multimedia information are processed by weight adjustment and related estimation. Wherein, the weight adjustment of the recommendation features based on the key features can be set according to the specific estimation processing mode. For example, the weight adjustment result can be obtained by adjusting the weights of the recommendation features based on the key features before the estimation processing of the recommendation features, or can be obtained by adjusting the intermediate estimation features obtained based on the recommendation features during the estimation process of the recommendation features, or can be obtained by adjusting the final estimation features obtained after the estimation processing of the recommendation features based on the key features. Here, the specific process of weight adjustment result and weight adjustment is not limited, and can be set as needed.

[0150] In an example, the estimation processing can be a logistic regression algorithm, and before the recommendation features are processed by the logistic regression algorithm, the weights of the recommendation features can be adjusted based on the key features. Alternatively, the weights of the recommendation features after the logistic regression algorithm can also be adjusted based on the key features. Assuming that the estimation processing includes three times of logistic regression algorithm processing, then the weights of the recommendation features can be adjusted based on the key features before or after any one of the first, second, and third logistic regression algorithm, or twice of the logistic regression algorithm, or three times of the logistic regression algorithm.

[0151] In an example, the estimation processing can be a deep neural network, and before the recommendation features are input into a regularization layer in the deep neural network, the recommendation features can be adjusted in weight based on the key features. Alternatively, the recommendation features can be adjusted in weight based on the key features after the recommendation features are input into the regularization layer in the deep neural network. Assuming that the deep neural network includes processing of three regularization layers, the recommendation features can be adjusted in weight based on the key features before or after any one of the first regularization layer, the second regularization layer, and the third regularization layer, or before or after any two of the first regularization layer, the second regularization layer, and the third regularization layer, or before or after all of the first regularization layer, the second regularization layer, and the third regularization layer.

[0152] Step 203: selecting a recommended multimedia information from the at least one candidate multimedia information based on the obtained at least one estimated conversion rate.

[0153] In the embodiments of the present application, the selection manner of selecting the recommended multimedia information from the at least one candidate multimedia information can be: selecting a candidate multimedia information with the highest estimated conversion rate from the at least one candidate multimedia information as the recommended multimedia information, or selecting a candidate multimedia information with an estimated conversion rate meeting a set threshold from the at least one candidate multimedia information as the recommended multimedia information, or selecting a candidate multimedia information with an estimated conversion rate meeting a set threshold and a multimedia object type meeting a set multimedia object type from the at least one candidate multimedia information as the recommended multimedia information, and the like. Here, the specific selection manner of the recommended multimedia information is not limited.

[0154] In an example, the estimated conversion rate of the candidate multimedia information A is 65%, the estimated conversion rate of the candidate multimedia information B is 39%, and the estimated conversion rate of the candidate multimedia information C is 70%, and the candidate multimedia information C with the highest estimated conversion rate can be selected as the recommended multimedia information.

[0155] In an example, the estimated conversion rate of the candidate multimedia information A is 75%, the estimated conversion rate of the candidate multimedia information B is 39%, and the estimated conversion rate of the candidate multimedia information C is 80%, and the candidate multimedia information C with an estimated conversion rate meeting a set threshold of 80% can be selected as the recommended multimedia information.

[0156] In an example, the estimated conversion rate of the candidate multimedia information A is 75%, the multimedia object category is a vehicle, the estimated conversion rate of the candidate multimedia information B is 39%, the multimedia object category is a mobile terminal, and the estimated conversion rate of the candidate multimedia information C is 80%, the multimedia object category is a novel, and the candidate multimedia information A with an estimated conversion rate meeting a set threshold of 60% and a set multimedia object type of a vehicle can be selected as the recommended multimedia information.

[0157] In the method, after receiving a play request carrying object basic attributes and a play condition of a play object, a plurality of candidate multimedia information to be recommended is obtained in response to the play request. In this way, preliminary screening of the recommended multimedia information is implemented for the play request, and a plurality of candidate multimedia information to be recommended is obtained. Subsequently, each candidate multimedia information is estimated, and the recommended candidate multimedia information is determined according to the estimated conversion rate, thereby preventing unnecessary waste of computing resources caused by estimating a large number of multimedia information respectively without preliminary screening of the multimedia information. The recommendation features of each of the plurality of candidate multimedia information are obtained respectively, so that the recommendation features of one candidate multimedia information can represent the characteristics of the one candidate multimedia information. According to the set information identifier, the key information corresponding to the information identifier is selected from the object basic attributes, the play condition, and the play content of the one candidate multimedia information. The information identifier represents one kind of information that causes the historical estimated conversion rate to deviate. In this way, at least one key information is selected according to at least one information identifier, and the key features of the at least one key information are extracted. The key features correspond to the characteristics of at least one kind of information that causes the historical estimated conversion rate to deviate. Based on the key features, the weights of the elements in the recommendation features of the candidate multimedia information of the key features are adjusted, so that the corresponding key features in the recommendation features are more accurate in the estimated conversion rate of the candidate multimedia information, and the accuracy of the estimated conversion rate of the multimedia information is improved.

[0158] The embodiment of the present application also provides a multimedia information recommendation method based on the above-mentioned Figure 2 process steps, as shown in Figure 3 step 201, the response to the play request carrying the object basic attributes and the play condition of the play object, obtaining a plurality of candidate multimedia information to be recommended, and extracting the respective recommendation features of the plurality of candidate multimedia information, comprising:

[0159] Step 301, in response to the play request, a plurality of candidate multimedia information meeting the object basic attributes and the play condition is obtained by using a recall algorithm;

[0160] For the plurality of candidate multimedia information, the following operations are performed respectively:

[0161] Step 3021, respectively extracting features from the object basic attributes, the play condition, and the play content of the one candidate multimedia information to obtain corresponding object features, condition features, and content features;

[0162] In the embodiment of the present application, the object basic attributes carried in the play request can be collected by the client when the client accesses the page of the play object, or can be collected by the client when the client is installed in the terminal device and stored locally. Here, the specific acquisition method of the object basic attributes is not limited.

[0163] In an example, the object basic attribute can include: a play object age, a play object gender, a city where the play object is located, a historical browsing number of times of the play object to the access page, a historical browsing average duration of the play object to the access page, a conversion index of the play object, and the like, where the specific basic information contained in the object basic attribute is not limited. The play condition can include: a play request occurrence place, a play recommended multimedia information address, a maximum duration of playing the recommended multimedia information, a play scene type, and the like, where the specific condition information contained in the play condition is not limited. The play content can include: a multimedia information object category, a multimedia information category (such as a picture category, a video category, a text or voice category, and the like), multimedia information material related information (such as if the multimedia information object category is a novel, the corresponding multimedia information material related information can be a documentary novel or a romance novel or a suspense novel, and the like; if the multimedia information object category is a picture, the corresponding multimedia information material related information can be an animation picture or a real object novel, and the like), and the like, where the specific content information contained in the play content is not limited.

[0164] In the embodiments of the present application, the object basic attribute, the play condition, and the play content of a candidate multimedia information can be respectively subjected to feature extraction to obtain corresponding object features, condition features, and content features, or each attribute information in the object basic attribute can be respectively subjected to feature extraction to obtain attribute features, and the attribute features can be subjected to feature cross to obtain object features; each specific play condition in the play condition can be respectively subjected to feature extraction to obtain specific play condition features, and the specific play condition features can be subjected to feature cross to obtain condition features; each item of content in the play content can be respectively subjected to feature extraction to obtain specific content features, and the specific content features can be subjected to feature cross to obtain content features. The specific feature extraction manner of the present application for the object basic attribute, the play condition, and the play content of a candidate multimedia information is not limited. Through the first feature extraction method, the generalization association relationship between each item of information (attribute information of the object basic attribute, each specific play condition of the play condition, and each item of content of the play content) can be increased. Through the second feature extraction method, the features of each field (such as the object basic attribute is one field, the play condition is one field, and the play content is one field) can be reserved, which can be set according to specific needs.

[0165] Based on the above example, the object basic attribute can include: the play object age: 20, the play object gender: female, the play object city: BJ, the play object historical browsing times of the access page: 30, the play object historical browsing average duration of the access page: 30min, the play object conversion index: collection, etc. Here, the specific basic information contained in the object basic attribute is not limited. The play condition can include: the play request occurrence place: BJ, the play recommended multimedia information address: XX street in BJ city, the maximum duration of the play recommended multimedia information: 2min, the play scene type: game page scene type, etc. Here, the specific condition information contained in the play condition is not limited. The play content can include: the multimedia information object category: mobile terminal, the multimedia information category: video category, the multimedia information material related information: physical display, etc. Here, the specific content information contained in the play content is not limited. As shown in Figure 4A , in this example, the object basic attribute is subjected to feature extraction to obtain the corresponding object feature, as shown in Figure 4B , in this example, the play condition is subjected to feature extraction to obtain the corresponding condition feature, as shown in Figure 4C , in this example, the play content of a candidate multimedia information is subjected to feature extraction to obtain the corresponding content feature. As shown in Figure 5A , in this example, each attribute information in the object basic attribute is subjected to feature extraction to obtain each attribute feature, and the attribute features are subjected to feature cross to obtain the object feature, as shown in Figure 5B , in this example, each specific play condition in the play condition is subjected to feature extraction to obtain each specific play condition feature, and the specific play condition features are subjected to feature cross to obtain the condition feature, as shown in Figure 5C , in this example, each item of content in the play content is subjected to feature extraction to obtain each specific content feature, and the specific contents are subjected to feature cross to obtain the content feature.

[0166] Step 3022, the object feature, the condition feature and the content feature are subjected to feature cross processing to obtain the recommendation feature of the one candidate multimedia information.

[0167] Based on the object feature, the condition feature and the content feature obtained by the feature selection method corresponding to the above example, such as Figure 4A , 4B , 4C or Figure 5A , 5B , 5C, the object feature, the condition feature and the content feature are subjected to feature cross processing by the feature cross processing method provided in the embodiments of the present application, as shown in Figure 6 , the object feature, the condition feature and the content feature corresponding to one candidate multimedia information are subjected to feature cross to obtain the recommendation feature of the one candidate multimedia information.

[0168] In the method, a recall algorithm is used to preliminarily screen the multimedia information to be recommended to the client according to the object basic attribute and the playing condition, and a plurality of candidate multimedia information meeting the object basic attribute and the playing condition is obtained. In this way, the calculation resources consumed for the preliminary selection of the multimedia information without using the recall algorithm can be reduced. The corresponding object feature, condition feature and content feature are extracted from the object basic attribute, the playing condition and the playing content of the candidate multimedia information. In this way, the recommendation feature of the candidate multimedia information includes the object feature of the object basic attribute, the condition feature of the playing condition and the content feature of the playing content of the candidate multimedia information. The extracted object feature, condition feature and content feature are cross-processed to obtain the recommendation feature of the candidate multimedia information. In this way, the object feature, condition feature and content feature are mapped into the recommendation feature with the feature dimension number required by the model, the nonlinear capability of the model is increased, and the prediction effect of the model is improved.

[0169] The embodiment of the application further provides a key feature extraction method based on the above Figure 2 and Figure 3 Corresponding process steps, the key feature of the at least one key information is extracted, including:

[0170] The feature extraction is performed on each key information in the at least one key information respectively to obtain at least one preliminary key feature.

[0171] The at least one preliminary key feature is cross-processed to obtain the key feature. That is, the way of obtaining the key feature can be that the feature extraction is performed on the at least one key information to obtain the key feature, or the feature extraction is performed on each key information in the at least one key information respectively to obtain at least one preliminary key feature, and the at least one preliminary key feature is cross-processed to obtain the key feature. The specific feature extraction manner of the feature extraction performed on the at least one key information is not limited in the application, and can be set according to specific needs.

[0172] In an example, if the at least one key information includes the conversion index of the playing object: registration, the playing scene type: game page scene type and the multimedia object category: mobile game, as shown in Figure 7 the feature extraction is performed on the conversion index of the playing object: registration, the playing scene type: game page scene type and the multimedia object category: mobile game to obtain the key feature. As shown in Figure 8As shown, the conversion indicators of the playing object are subjected to feature extraction to obtain preliminary key features, the playing scene types are subjected to feature extraction to obtain preliminary key features, and the multimedia object categories are subjected to feature extraction to obtain preliminary key features. The three preliminary key features are subjected to feature cross processing to obtain key features.

[0173] In the above method, feature extraction is performed on each of the at least one key information to obtain at least one preliminary key feature. In this way, the at least one preliminary key feature represents the characteristics of the at least one key information. The at least one preliminary key feature is subjected to feature cross processing to obtain key features. In this way, the feature cross processing maps the at least one preliminary key feature into key features with a feature dimension number required by the model, increases the nonlinear capability of the model, and improves the accuracy of the subsequent generation of weight features by the model.

[0174] Based on the above key feature extraction method, the embodiments of the present application further provide a multimedia information recommendation method, as shown in Figure 9 For ease of understanding, based on the flow steps described in Figure 2 and Figure 3 , step 2022, the key features are subjected to weight adjustment based on the elements contained in the recommendation features of the one candidate multimedia information, and based on the weight adjustment result, an estimated conversion rate of the one candidate multimedia information is obtained, including:

[0175] Step 901, the key features are subjected to two times of mapping processing respectively to obtain first weight features and second weight features, and the first weight features and the second weight features are respectively used to represent the weights of the associated elements contained in the recommendation features of the one candidate multimedia information and the key features;

[0176] In the embodiments of the present application, the specific method of mapping processing the key features can be a logistic regression algorithm, a deep neural network method, etc., and the number of times of mapping processing can be one, two, three or more times, etc. Here, the specific method and the number of times of mapping processing the key features are not limited.

[0177] For ease of understanding, the embodiments of the present application provide a method of mapping processing the key features based on a logistic regression algorithm, as shown in Figure 10As shown in the figure, in this embodiment, it is assumed that the key feature is 64-dimensional, and the 64-dimensional key feature is subjected to two times of logistic regression algorithm to obtain 64-dimensional first weight feature and 64-dimensional second weight feature. Different parameters can be used in the two times of logistic regression algorithm to obtain the corresponding weight feature. In addition, the key feature dimension herein does not limit the key feature dimension in actual implementation, for example, the key feature dimension can be 36-dimensional, 48-dimensional, 72-dimensional, etc. The dimensions of the first weight feature and the second weight feature do not limit the key feature dimension in actual implementation, for example, the dimensions of the first weight feature and the second weight feature can be 36-dimensional, 48-dimensional, 72-dimensional, etc. The dimensions of the first weight feature and the second weight feature can be the same as or different from the key feature dimension, and can be set as needed, for example, 64-dimensional key feature can be subjected to logistic regression algorithm to obtain 36-dimensional / 48-dimensional / 72-dimensional first weight feature and second weight feature, and 36-dimensional key feature can be subjected to logistic regression algorithm to obtain 24-dimensional / 48-dimensional / 72-dimensional first weight feature and second weight feature. The dimensions of the key feature, the first weight feature and the second weight feature are not limited herein.

[0178] The embodiment of the present application provides a method for mapping and processing a key feature based on a deep neural network method, as shown in the figure. Figure 11 As shown in the figure, in this embodiment, it is assumed that the key feature is 64-dimensional, and the 64-dimensional key feature is subjected to two times of deep neural network method to obtain 64-dimensional first weight feature and 64-dimensional second weight feature. In this embodiment, the deep neural network can include: 64-dimensional key feature→fully connected layer→regularization layer→activation function→fully connected layer→activation function→64-dimensional first weight feature, and 64-dimensional key feature→fully connected layer→regularization layer→activation function→fully connected layer→activation function→64-dimensional second weight feature. The dimensions of the key feature, the first weight feature and the second weight feature are not limited herein, and the dimensions of the key feature, the first weight feature and the second weight feature can be the same or different.

[0179] In step 902, the elements included in the recommended feature of the one candidate multimedia information are subjected to weight adjustment according to the first weight feature, and the recommended feature subjected to weight adjustment is subjected to classification mapping processing to obtain preliminary estimated feature.

[0180] In an example, it is assumed that the key feature is 【0 0 4 0 8 0 9 0 2 16 1 6】, the first weight feature is 【1 1 0.2 1 0.4 1 0.45 1 0.1 1.8 0.1 0.5】, the second weight feature is 【1 1 0.2 1 0.6 1 0.6 1 0.1 1.4 0.1 0.4】, and the recommended feature is 【6 9 4 10 8 22 9 3 2 16 1 6】; the first weight feature is multiplied by the recommended feature, i.e., 【1 1 0.2 1 0.4 1 0.45 1 0.1 1.8 0.1 0.5】*【6 9 4 10 8 22 9 3 2 16 1 6】=【6*1 9*1 4*0.2 10*1 8*0.4 22*1 9*0.45 3*1 2*0.1 16*1.8 1*0.1 6*0.5】=【6 9 0.810 3.2 22 4.05 3 0.2 28.8 0.1 3】, to obtain the recommended feature after weight adjustment. It should be noted that, in order to facilitate understanding, the multiplication of the first weight feature and the recommended feature of one candidate multimedia information is taken as the weight adjustment method in the example, and the weight adjustment method of the first weight feature and the recommended feature of one candidate multimedia information can also be weighting, multiplication after weighting of the first weight feature and the recommended feature of one candidate multimedia information, etc. Here, the specific method of weight adjustment of the first weight feature and the recommended feature of one candidate multimedia information is not limited, and can be set as required. In addition, the dimensions and specific values of the features involved in the example are only used to facilitate understanding of the multimedia information recommendation method of the present application, and do not limit the multimedia information recommendation method of the present application.

[0181] In the embodiments of the present application, the specific method of classification mapping processing of the recommended feature after weight adjustment can be a logistic regression algorithm, a deep neural network method, etc., and the number of times of classification mapping processing can be one, two, three or more times, etc. Here, the specific method and number of times of classification mapping processing of the recommended feature after weight adjustment are not limited.

[0182] In step 903, the elements included in the preliminary estimation feature are subjected to weight adjustment according to the second weight feature, and the preliminary estimation feature after weight adjustment is subjected to classification mapping processing, to obtain the estimated conversion rate of the one candidate multimedia information.

[0183] In the above example, it is assumed that the preliminary estimated feature is 【0.1 0.1 0.4 0.3 0.8 0.2 0.9 0.1 0.2 0.6 1 0.6】, the second weight feature is *the preliminary estimated feature is = 【1 1 0.2 1 0.6 1 0.6 1 0.1 1.4 0.1 0.4】 * 【0.1 0.1 0.4 0.3 0.8 0.2 0.9 0.1 0.2 0.6 1 0.6】 = 【1 1 0.2 1 0.6 1 0.6 1 0.1 1.4 0.1 0.4】 = 【0.1 0.1 0.08 0.3 0.48 0.2 0.54 0.1 0.02 0.84 0.1 0.24】 to obtain the weight-adjusted preliminary estimated feature. It should be noted that, in order to facilitate understanding, the example takes the multiplication of the second weight feature and the preliminary estimated feature of one candidate multimedia information as the weight adjustment method, and the weight adjustment method of the second weight feature and the preliminary estimated feature of one candidate multimedia information can also be weighting, multiplication after weighting of the second weight feature and the preliminary estimated feature of one candidate multimedia information, or the like. Here, the specific method of weight adjustment of the second weight feature and the preliminary estimated feature of one candidate multimedia information is not limited, and can be set as needed. The specific method of weight adjustment of the first weight feature and the recommended feature of one candidate multimedia information can be the same as or different from the specific method of weight adjustment of the second weight feature and the preliminary estimated feature of one candidate multimedia information. In addition, the dimensions and specific values of the features involved in the example are only used to facilitate understanding of the multimedia information recommendation method of the present application, and do not limit the multimedia information recommendation method of the present application.

[0184] In the embodiments of the present application, the specific method of classification mapping processing of the weight-adjusted preliminary estimated feature can be a logistic regression algorithm, a deep neural network method, or the like, and the number of times of classification mapping processing can be one, two, three, or more times, or the like. Here, the specific method and number of times of classification mapping processing of the weight-adjusted preliminary estimated feature are not limited.

[0185] In the embodiments of the present application, the specific method of classification mapping processing of the recommended feature and the preliminary estimated feature of one candidate multimedia information can be a logistic regression algorithm, a deep neural network method, or the like, and the number of times of classification mapping processing can be one, two, three, or more times, or the like. Here, the specific method and number of times of classification mapping processing of the recommended feature and the preliminary estimated feature of one candidate multimedia information are not limited.

[0186] For the convenience of understanding, the embodiments of the present application give a method of classification mapping processing of the recommended feature and the preliminary estimated feature of one candidate multimedia information based on a logistic regression algorithm, as follows: Figure 12As shown in the figure, in this embodiment, it is assumed that the recommendation feature of a candidate multimedia information is 64-dimensional, each element contained in the recommendation feature of a candidate multimedia information is adjusted in weight according to the 64-dimensional first weight feature, the 64-dimensional weight-adjusted recommendation feature is obtained, and the weight-adjusted recommendation feature is processed by a logistic regression algorithm to obtain a 64-dimensional preliminary estimation feature. Each element contained in the 64-dimensional preliminary estimation feature is adjusted in weight according to the second weight feature, the 64-dimensional weight-adjusted preliminary estimation feature is obtained, and the weight-adjusted preliminary estimation feature is further processed by a logistic regression algorithm to obtain the estimated conversion rate of a candidate multimedia information.

[0187] The embodiment of the present application gives a method of classifying and mapping the recommendation feature and the preliminary estimation feature of a candidate multimedia information based on a deep neural network method, as shown in the figure. Figure 13 As shown in the figure, in this embodiment, the recommendation feature of a candidate multimedia information is 64-dimensional, each element contained in the recommendation feature of a candidate multimedia information is adjusted in weight according to the 64-dimensional first weight feature, the 64-dimensional weight-adjusted recommendation feature is obtained, and the weight-adjusted recommendation feature is processed by a logistic regression algorithm to obtain a 64-dimensional preliminary estimation feature. Each element contained in the 64-dimensional preliminary estimation feature is adjusted in weight according to the second weight feature, the 64-dimensional weight-adjusted preliminary estimation feature is obtained, and the weight-adjusted preliminary estimation feature is further processed by a logistic regression algorithm to obtain the estimated conversion rate of a candidate multimedia information. Here, the dimensions of the recommendation feature, the weight-adjusted recommendation feature, the preliminary estimation feature, the weight-adjusted preliminary estimation feature, and the first weight feature and the second weight feature are only an example, and the dimensions of the recommendation feature, the weight-adjusted recommendation feature, the preliminary estimation feature, the weight-adjusted preliminary estimation feature, and the first weight feature and the second weight feature can be any dimensions set as needed, such as 36 dimensions, 48 dimensions, 72 dimensions, etc., and the dimensions of each feature can be the same or different. If the dimensions of the features are different, they can be padded by zero or other methods, or they can be padded by feature combination in the classification and mapping process. Here, the dimensions of each feature are not limited.

[0188] In the method, the key features are respectively mapped twice to obtain the first weight feature and the second weight feature. In this way, the part model corresponding to the mapping processing of the key features can be set as one part model or two part models. When the part model is set as one part model, the part model maps the key features twice to flexibly obtain the first weight feature and the second weight feature, so that the weights of the elements represented by the first weight feature and the second weight feature have high robustness. When the part model is set as two part models, the two part models respectively map the key features to obtain the first weight feature and the second weight feature, and the model parameters in the two part models are not completely the same, so that the weights of the elements represented by the first weight feature and the second weight feature have high robustness. Before the recommended features of a candidate multimedia information are classified and mapped, the elements contained in the recommended features of a candidate multimedia information are adjusted in weight according to the first weight feature to obtain weight-adjusted recommended features, and the weight-adjusted recommended features are classified and mapped. In this way, the weights of the elements corresponding to the key features in the weight-adjusted recommended features are accurate, so that the preliminary estimated features obtained through the classification and mapping processing can accurately represent the influence of the elements in the weight-adjusted recommended features on the estimated conversion rate of a candidate multimedia information. Further, the elements in the preliminary estimated features are adjusted in weight according to the second weight feature, and the weight-adjusted preliminary estimated features are classified and mapped, so that the elements in the weight-adjusted and classified and mapped preliminary estimated features accurately represent the influence on the estimated conversion rate of a candidate multimedia information, thereby improving the accuracy of the estimated conversion rate of a candidate multimedia information.

[0189] The embodiment of the present application further provides a weight feature acquisition method based on the above Figure 9 The multimedia information recommendation method comprises the following steps:

[0190] In step 9011, the key features are respectively normalized twice to obtain the first standard key feature and the second standard key feature, and the first standard key feature and the second standard key feature are respectively classified and mapped to obtain the corresponding first preliminary weight feature and the second preliminary weight feature.

[0191] In an example, the key features can be normalized by a normalization method of 0-mean standardization (Z-score method) or a normalization method of a regularization layer of a deep neural network method, to obtain first standard key features and second standard key features. Further, the first standard key features and the second standard key features can be respectively classified and mapped by a logistic regression algorithm to obtain corresponding first preliminary weight features and second preliminary weight features, or the first standard key features and the second standard key features can be respectively classified and mapped by a fully connected layer of a deep neural network method to obtain corresponding first preliminary weight features and second preliminary weight features. The normalization method of the key features and the classification and mapping method of the first standard key features and the second standard key features are not limited in the present application.

[0192] Step 9012, respectively multiplying the first preliminary weight features and the second preliminary weight features by a set multiple to obtain the first weight features and the second weight features.

[0193] In an example, the way of multiplying the first preliminary weight features and the second preliminary weight features by a set multiple can be directly multiplying the first preliminary weight features and the second preliminary weight features by a set multiple, respectively, or can be multiplying the first preliminary weight features and the second preliminary weight features by a set multiple through a set function, respectively. The specific scaling method of the first preliminary weight features and the second preliminary weight features is not limited here.

[0194] In an example, as shown in FIG. 9, the first preliminary weight features and the second preliminary weight features can be respectively multiplied by a set multiple to obtain the first weight features and the second weight features. Figure 14As shown, the key features can be normalized by the normalization method of the regularization layer of the deep neural network method to obtain the first standard key features and the second standard key features, and the first standard key features and the second standard key features are classified and mapped by the full connection layer of the deep neural network method to obtain the corresponding first preliminary weight features and the second preliminary weight features, and the first preliminary weight features and the second preliminary weight features are processed by the set activation function to realize multiplication with the set multiple to obtain the first weight features and the second weight features. The weight features are obtained. Specifically, at least one key information includes: multimedia object category: mobile terminal, conversion index: collection, playing scene type: game page scene type, etc. According to the at least one key information, 64-dimensional key features are obtained, the 64-dimensional key features are processed by Sparsefm (perception factor decomposition machine) for feature combination to obtain 64-dimensional key features after feature combination processing, the 64-dimensional key features after feature combination processing are classified and mapped by the full connection layer to obtain 64-dimensional key features after classification and mapping, the 64-dimensional key features after classification and mapping are normalized by the regularization layer to obtain the first standard key features, the first standard key features are processed by the relu activation function layer, at this time, the element value of each element in the first standard key features after the relu activation function layer is in the range of 【0-1】, and the first standard key features after the relu activation function layer are classified and mapped by the full connection layer to obtain the first preliminary weight features, the first preliminary weight features are processed by the 2*sigmoid activation function layer to obtain the first weight features, at this time, the element value of each element in the first weight features processed by the 2*sigmoid activation function layer is in the range of 【0-2】.

[0195] The 64-dimensional key features are obtained by encoding, the 64-dimensional key features are processed by Sparsefm (perception factor decomposition machine) for feature combination to obtain 64-dimensional key features after feature combination processing, the 64-dimensional key features after feature combination processing are classified and mapped by the full connection layer to obtain 64-dimensional key features after classification and mapping, the 64-dimensional key features after classification and mapping are normalized by the regularization layer to obtain the second standard key features, the second standard key features are processed by the relu activation function layer, at this time, the element value of each element in the second standard key features after the relu activation function layer is in the range of 【0-1】, and the second standard key features after the relu activation function layer are classified and mapped by the full connection layer to obtain the second preliminary weight features, the second preliminary weight features are processed by the 2*sigmoid activation function layer to obtain the second weight features, at this time, the element value of each element in the second weight features processed by the 2*sigmoid activation function layer is in the range of 【0-2】.

[0196] In the method, the first weight feature and the second weight feature are acquired according to the key features, and the weight adjustment is performed on the recommended features and the preliminary estimated features of the candidate multimedia information respectively before the normalization processing of the recommended features and the preliminary estimated features of the candidate multimedia information respectively, so that the weight of the corresponding element in the key features can be prevented from being weakened in the normalization processing, and the recommended features and the preliminary estimated features after the weight adjustment can accurately retain the weight of the corresponding element in the key features in the overall recommended features.

[0197] The embodiment of the application further provides a multimedia information recommendation method based on the above Figure 9 The corresponding multimedia information recommendation method further comprises: determining sparse recommended features according to the object basic attributes, the playing conditions and the playing content of the candidate multimedia information, the sparse recommended features representing the respective words of the object basic attributes, the playing conditions and the playing content of the candidate multimedia information.

[0198] In the embodiment of the application, as shown in the figure, Figure 15 The sparse recommended features are determined according to the object basic attributes, the playing conditions and the playing content of the candidate multimedia information, if the dimension number of the sparse recommended features is less than the set sparse feature dimension number, the sparse feature dimension number can be increased through the feature combination, for example, the sparse recommended features with 36 dimensions are increased to the sparse recommended features with 64 dimensions through the sparse select product layer in the neural network, if the dimension number of the sparse recommended features is greater than the set sparse feature dimension number, the sparse feature dimension number can be reduced through the feature combination, for example, the sparse recommended features with 72 dimensions are reduced to the sparse recommended features with 64 dimensions through the sparse select product layer in the neural network. The feature combination through the sparse select product layer in the neural network provided herein is only an example, and the feature combination can also be performed through the domain-aware factorization machine, and the specific manner of the feature combination is not limited herein.

[0199] The classification mapping processing of the preliminary estimated features after the weight adjustment in step 903 obtains the estimated conversion rate of the candidate multimedia information, comprising:

[0200] Step 9031, the classification mapping processing of the preliminary estimated features after the weight adjustment obtains the intermediate estimated features.

[0201] In the embodiments of the present application, the preliminary estimated features after weight adjustment can be classified and mapped by a full connection layer in a logistic regression algorithm or a deep neural network method to obtain intermediate estimated features. The specific method of classifying and mapping the preliminary estimated features after weight adjustment is not limited here.

[0202] At step 9032, the sparse recommendation features are weighted with the intermediate estimated features to obtain final estimated features, and an estimated conversion rate of the one candidate multimedia information is obtained based on the final estimated features.

[0203] Based on the above Figure 12 According to the examples, the present application provides another multimedia information recommendation method. In one example, as shown in Figure 16 After obtaining the 64-dimensional preliminary estimated features after weight adjustment, the preliminary estimated features after weight adjustment are further processed by a logistic regression algorithm to obtain intermediate estimated features. The intermediate estimated features are weighted with sparse recommendation features to obtain final estimated features, and an estimated conversion rate of the one candidate multimedia information can be obtained based on the final estimated features.

[0204] Based on the above Figure 13 According to the examples, the present application provides another multimedia information recommendation method. In one example, as shown in Figure 17 After obtaining the 64-dimensional preliminary estimated features after weight adjustment, the preliminary estimated features after weight adjustment are further processed by a logistic regression algorithm to obtain intermediate estimated features. The intermediate estimated features are weighted with sparse recommendation features to obtain final estimated features, and an estimated conversion rate of the one candidate multimedia information can be obtained based on the final estimated features.

[0205] In the above method, the sparse recommendation features are determined according to the object basic attributes, the playing conditions and the playing content of the one candidate multimedia information. The sparse recommendation features are weighted with the intermediate estimated features, so that the final estimated features not only contain the semantic features of each word of the one candidate multimedia information, but also contain the word features of each word of the one candidate multimedia information. The word itself characteristics represented by the intermediate estimated features are compensated by the sparse recommendation features, and the accuracy of the estimated conversion rate of the one candidate multimedia information is improved.

[0206] The embodiments of the present application also provide a multimedia information recommendation method based on the above Figure 9 According to the corresponding multimedia information recommendation method, the classified and mapped processing of the recommendation features after weight adjustment at step 902 to obtain preliminary estimated features includes:

[0207] The weight-adjusted recommended features are normalized to obtain standard recommended features, and the standard recommended features are classified and mapped to obtain the preliminary estimated features;

[0208] In the embodiments of the present application, the weight-adjusted recommended features can be normalized by a normalization method of 0-mean standardization (Z-score method) or a normalization method of a regularization layer of a deep neural network method to obtain standard recommended features. Further, the standard recommended features can be classified and mapped by a logistic regression algorithm to obtain preliminary estimated features, or the standard recommended features can be classified and mapped by a fully connected layer of a deep neural network method to obtain preliminary estimated features. The normalization method of the weight-adjusted recommended features and the classification and mapping method of the standard recommended features are not specifically limited in the present application.

[0209] The weight-adjusted preliminary estimated features are classified and mapped in step 903 to obtain the estimated conversion rate of the one candidate multimedia information, including:

[0210] The weight-adjusted preliminary estimated features are normalized to obtain standard preliminary estimated features, and the standard preliminary estimated features are classified and mapped to obtain the estimated conversion rate of the one candidate multimedia information.

[0211] In the embodiments of the present application, the weight-adjusted preliminary estimated features can be normalized by a normalization method of 0-mean standardization (Z-score method) or a normalization method of a regularization layer of a deep neural network method to obtain standard preliminary estimated features. Further, the standard preliminary estimated features can be classified and mapped by a logistic regression algorithm, or the standard preliminary estimated features can be classified and mapped by a fully connected layer of a deep neural network method to obtain the estimated conversion rate of the one candidate multimedia information. The normalization method of the weight-adjusted preliminary estimated features and the classification and mapping method of the standard preliminary estimated features are not specifically limited in the present application.

[0212] In the embodiments of the present application, the preliminary estimated features after weight adjustment can be normalized by a normalization method of 0-mean standardization (Z-score method) or a normalization method of a regularization layer of a deep neural network method, to obtain standard preliminary estimated features; further, the standard preliminary estimated features can be classified and mapped by a logistic regression algorithm, to obtain intermediate estimated features; the sparse recommended features and the intermediate estimated features are weighted to obtain final estimated features, and based on the final estimated features, an estimated conversion rate of a candidate multimedia information is obtained, or the standard preliminary estimated features can be classified and mapped by a fully connected layer of a deep neural network method, to obtain intermediate estimated features; the sparse recommended features and the intermediate estimated features are weighted to obtain final estimated features, and based on the final estimated features, an estimated conversion rate of a candidate multimedia information is obtained. The normalization method of the preliminary estimated features after weight adjustment and the classified and mapped method of the standard preliminary estimated features are not limited in the present application. In the above method, the recommended features after weight adjustment and the preliminary estimated features after weight adjustment are normalized respectively, to prevent model overfitting, and to quickly obtain an estimated conversion rate of a candidate multimedia information in a short time, to improve the estimation efficiency.

[0213] In an example, based on the method for obtaining the first weight features and the second weight features in the above Figure 14 and the multimedia information recommendation method in the above Figure 17 , the embodiments of the present application provide a multimedia information recommendation method, as shown in the above Figure 18 , 64-dimensional dense recommended features are extracted according to the object basic attributes, the playing conditions and the playing content of a candidate multimedia information, the 64-dimensional dense recommended features after feature crossing are obtained by processing with fwffm (a domain-aware factorization machine), the 64-dimensional dense recommended features after feature crossing are weighted adjusted based on the first weight features to obtain recommended features after weight adjustment, the recommended features after weight adjustment are normalized by a regularization layer to obtain standard recommended features, the standard recommended features are processed by a fully connected layer to obtain preliminary estimated features, the preliminary estimated features are weighted adjusted based on the second weight features to obtain preliminary estimated features after weight adjustment, the preliminary estimated features after weight adjustment are normalized by a regularization layer to obtain standard preliminary estimated features, the standard preliminary estimated features are processed by a relu activation function layer to obtain standard preliminary estimated features after activation function, the standard preliminary estimated features after activation function are processed by a fully connected layer to obtain intermediate estimated features, the intermediate estimated features and the sparse recommended features are weighted to obtain final estimated features, and based on the final estimated features, an estimated conversion rate of a candidate multimedia information is obtained. The above Figure 18The multimedia information recommendation method in the method can be a prediction model of a deep neural network method, which can be trained based on a historical recommended multimedia information set. The historical recommended multimedia information in the historical recommended multimedia information set includes an object basic attribute corresponding to the recommended multimedia information, a playing condition, and playing content of the recommended multimedia information, at least one key information of the recommended multimedia information, and a corresponding predicted conversion rate and an actual conversion rate. In the application process of the prediction model, information causing a deviation between the predicted conversion rate and the actual conversion rate of the recommended multimedia information in the generated recommended multimedia information set can be statistically analyzed, and at least one information identifier set is updated according to the information identifier of the information.

[0214] In the above method processes, the at least one key information includes one or more of a conversion indicator, a multimedia object category, and a playing scene type; the conversion indicator represents an object behavior in the object basic attribute that causes the multimedia information to be converted to generate an asset corresponding to the multimedia information; for example, an object behavior such as registration, payment, and collection of a playing object can cause the multimedia information to be converted to generate an asset corresponding to the multimedia information, and the object behaviors such as registration, payment, and collection can be the conversion indicator.

[0215] The multimedia object category represents an object category of a multimedia object displayed to the playing object in the playing content; for example, the multimedia object displayed to the playing object can be a novel such as "The Da Vinci Code" and "Red Dream", and the object category of the multimedia object is a novel, and the multimedia object displayed to the playing object can be an AA car and a BB car, and the object category of the multimedia object is a vehicle.

[0216] The playing scene type represents a page scene type of a page displayed to the playing object in the playing condition; for example, if the page accessed by the playing object is a shopping page, the page scene type of the page displayed to the playing object in the playing condition is a shopping page scene type, and if the page accessed by the playing object is a game page, the page scene type of the page displayed to the playing object in the playing condition is a game page scene type.

[0217] Based on the same concept, the embodiment of the present application provides a multimedia information recommendation device, as shown in the device. Figure 19 The device includes:

[0218] A first feature extraction unit 1901 is configured to, in response to a playing request carrying an object basic attribute of a playing object and a playing condition, acquire a plurality of candidate multimedia information to be recommended, and extract a respective recommendation feature of each of the plurality of candidate multimedia information.

[0219] For the plurality of candidate multimedia information, the following operations are respectively performed:

[0220] The second feature extraction unit 1902 is configured to select at least one key information from the object basic attribute, the play condition and the play content of one candidate multimedia information according to at least one information identifier set, and extract a key feature of the at least one key information; each information identifier represents one kind of information that causes the historical estimated conversion rate to deviate;

[0221] The estimation processing unit 1903 is configured to adjust the weights of elements contained in the recommendation feature of the one candidate multimedia information based on the key feature, and obtain an estimated conversion rate of the one candidate multimedia information based on the weight adjustment result.

[0222] The information selection unit 1904 is configured to select a recommended multimedia information from the at least one candidate multimedia information based on the obtained at least one estimated conversion rate.

[0223] Optionally, the first feature extraction unit 1901 is specifically configured to, in response to the play request, acquire a plurality of candidate multimedia information meeting the object basic attribute and the play condition by using a recall algorithm.

[0224] The following operations are respectively performed on the plurality of candidate multimedia information:

[0225] The object feature, the condition feature and the content feature are respectively obtained by performing feature extraction on the object basic attribute, the play condition and the play content of the one candidate multimedia information.

[0226] The recommendation feature of the one candidate multimedia information is obtained by performing feature cross processing on the object feature, the condition feature and the content feature.

[0227] Optionally, the second feature extraction unit 1902 is specifically configured to perform feature extraction on each key information in the at least one key information respectively, to obtain at least one preliminary key feature.

[0228] The key feature is obtained by performing feature cross processing on the at least one preliminary key feature.

[0229] Optionally, the estimation processing unit 1903 is specifically configured to perform twice mapping processing on the key feature respectively, to obtain a first weight feature and a second weight feature, and the first weight feature and the second weight feature are respectively used to represent the weights of associated elements contained in the recommendation feature of the one candidate multimedia information and the key feature.

[0230] The recommendation feature of the one candidate multimedia information is adjusted in weight according to the first weight feature, and the adjusted recommendation feature is classified and mapped to obtain a preliminary estimation feature;

[0231] The preliminary estimation feature is adjusted in weight according to the second weight feature, and the adjusted preliminary estimation feature is classified and mapped to obtain the estimated conversion rate of the one candidate multimedia information.

[0232] Optionally, the estimation processing unit 1903 is further configured to,

[0233] The sparse recommendation feature is determined according to the object basic attribute, the playing condition and the playing content of the one candidate multimedia information, and the sparse recommendation feature represents the respective words of the object basic attribute, the playing condition and the playing content of the one candidate multimedia information.

[0234] The estimation processing unit 1903 is specifically configured to perform classification and mapping processing on the adjusted preliminary estimation feature to obtain an intermediate estimation feature.

[0235] The sparse recommendation feature and the intermediate estimation feature are weighted to obtain a final estimation feature, and the estimated conversion rate of the one candidate multimedia information is obtained based on the final estimation feature.

[0236] Optionally, the estimation processing unit 1903 is specifically configured to,

[0237] The adjusted recommendation feature is normalized to obtain a standard recommendation feature, and the standard recommendation feature is classified and mapped to obtain the preliminary estimation feature.

[0238] The classification and mapping processing on the adjusted preliminary estimation feature to obtain the estimated conversion rate of the one candidate multimedia information includes:

[0239] The adjusted preliminary estimation feature is normalized to obtain a standard preliminary estimation feature, and the standard preliminary estimation feature is classified and mapped to obtain the estimated conversion rate of the one candidate multimedia information.

[0240] Optionally, the estimation processing unit 1903 is specifically configured to,

[0241] The key feature is normalized twice to obtain a first standard key feature and a second standard key feature, respectively, and the first standard key feature and the second standard key feature are classified and mapped to obtain a first preliminary weight feature and a second preliminary weight feature, respectively.

[0242] The first preliminary weight feature and the second preliminary weight feature are respectively multiplied by a set multiple to obtain the first weight feature and the second weight feature.

[0243] Optionally, the at least one key information comprises one or more of a conversion index, a multimedia object category, and a playing scene type.

[0244] The conversion index represents an object behavior in the object basic attribute that causes the multimedia information to be converted to generate the asset corresponding to the multimedia information.

[0245] The multimedia object category represents an object category of a multimedia object in the playing content that is presented to the playing object by the multimedia information.

[0246] The playing scene type represents a page scene type in which the playing object accesses a page to display the multimedia information in the playing condition.

[0247] Based on the same inventive concept as the method embodiments, the present embodiment also provides a computer device. In an embodiment, the computer device can be a server, such as the server 120 shown in Figure 1 In this embodiment, the structure of the computer device can be as shown in Figure 20 , and includes a memory 2001, a communication module 2003, and one or more processors 2002.

[0248] The memory 2001 is configured to store computer programs executed by the processor 2002. The memory 2001 can mainly include a program storage area and a data storage area, where the program storage area can store an operating system and programs required for running instant messaging functions, etc.; and the data storage area can store various instant messaging information and operation instruction sets, etc.

[0249] The memory 2001 can be a volatile memory, such as a random-access memory (RAM); the memory 2001 can also be a non-volatile memory, such as a read-only memory, a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD); or the memory 2001 can be any other medium capable of carrying or storing desired computer programs in the form of instructions or data structures and capable of being accessed by a computer, but is not limited to this. The memory 2001 can be a combination of the above memories.

[0250] Processor 2002 may include one or more central processing units (CPUs) or digital processing units, etc. Processor 2002 is used to implement the above-mentioned multimedia information recommendation method when calling computer programs stored in memory 2001.

[0251] The communication module 2003 is used to communicate with terminal devices and other servers.

[0252] This application embodiment does not limit the specific connection medium between the memory 2001, communication module 2003, and processor 2002. This application embodiment... Figure 20 The memory 2001 and the processor 2002 are connected via a bus 2004, which is in... Figure 20 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered as limiting information. The Bus 2004 can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 20 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0253] The memory 2001 stores a computer storage medium, which stores computer-executable instructions for implementing the multimedia information recommendation method of this application embodiment. The processor 2002 is used to execute the aforementioned multimedia information recommendation method, such as... Figure 2 or Figure 3 or Figure 9 As shown.

[0254] In another embodiment, the computer device can also be other computer devices, such as... Figure 1 The terminal device 110 is shown. In this embodiment, the structure of the computer device can be as follows: Figure 21 As shown, it includes components such as: communication component 2110, memory 2120, display unit 2130, camera 2140, sensor 2150, audio circuit 2160, Bluetooth module 2170, processor 2180, etc.

[0255] The communication component 2110 is used to communicate with the server. In some embodiments, it may include a Circuit-Based Wireless Fidelity (WiFi) module. WiFi is a short-range wireless transmission technology, and computer devices can use WiFi modules to help users send and receive information.

[0256] The memory 2120 can be used to store software programs and data. The processor 2180 performs various functions and data processing of the terminal device 110 by running the software programs or data stored in the memory 2120. The memory 2120 can include a high-speed random access memory, and can further include a non-volatile memory such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. The memory 2120 stores an operating system that enables the terminal device 110 to operate. In the present application, the memory 2120 can store an operating system and various application programs, and can further store a computer program for performing the multimedia information recommendation method in the embodiments of the present application.

[0257] The display unit 2130 can also be used to display information input by a user or information provided to a user, and a graphical user interface (GUI) of various menus of the terminal device 110. Specifically, the display unit 2130 can include a display screen 2132 arranged on the front of the terminal device 110. The display screen 2132 can be configured in the form of a liquid crystal display, a light-emitting diode, etc. The display unit 2130 can be used to display a multimedia information recommendation user interface and the like in the embodiments of the present application.

[0258] The display unit 2130 can also be used to receive input digital or character information, and generate signal inputs related to user settings and function controls of the terminal device 110. Specifically, the display unit 2130 can include a touch screen 2131 arranged on the front of the terminal device 110, which can collect touch operations of a user thereon or therearound, such as clicking a button, dragging a scroll box, etc.

[0259] The touch screen 2131 can be overlaid on the display screen 2132, or the touch screen 2131 can be integrated with the display screen 2132 to realize the input and output functions of the terminal device 110. After integration, the touch screen 2131 and the display screen 2132 can be simply referred to as a touch display screen. In the present application, the display unit 2130 can display application programs and corresponding operation steps.

[0260] The camera 2140 can be used to capture still images, and a user can publish comments on images captured by the camera 2140 through an application. The camera 2140 can be one or multiple. An object generates an optical image through a lens and projects the optical image onto a photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts an optical signal into an electrical signal, and then transmits the electrical signal to the processor 2180 to convert the electrical signal into a digital image signal.

[0261] The terminal device can further include at least one sensor 2150, such as an acceleration sensor 2151, a distance sensor 2152, a fingerprint sensor 2153, a temperature sensor 2154. The terminal device can also be configured with a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, a light sensor, a motion sensor, and other sensors.

[0262] The audio circuit 2160, the speaker 2161, and the microphone 2162 can provide an audio interface between the user and the terminal device 110. The audio circuit 2160 can convert the received audio data into an electrical signal and transmit the electrical signal to the speaker 2161, which converts the electrical signal into a sound signal and outputs the sound signal. The terminal device 110 can also be configured with a volume button for adjusting the volume of the sound signal. On the other hand, the microphone 2162 converts the collected sound signal into an electrical signal, which is received by the audio circuit 2160 and converted into audio data. The audio data is then output to the communication component 2110 for transmission to another terminal device 110, for example, or to the memory 2120 for further processing.

[0263] The Bluetooth module 2170 is used to interact with other Bluetooth devices having a Bluetooth module through a Bluetooth protocol. For example, the terminal device can establish a Bluetooth connection with a wearable computer device (e.g., a smart watch) having a Bluetooth module through the Bluetooth module 2170, and thus interact with the wearable computer device.

[0264] The processor 2180 is the control center of the terminal device, which connects all parts of the terminal device through various interfaces and lines, executes various functions of the terminal device and processes data by running or executing software programs stored in the memory 2120 and calling data stored in the memory 2120. In some embodiments, the processor 2180 can include one or more processing units; the processor 2180 can also integrate an application processor and a baseband processor, wherein the application processor mainly processes the operating system, the user interface, and the application program, and the baseband processor mainly processes wireless communication. It can be understood that the above-mentioned baseband processor can also not be integrated into the processor 2180. In the present application, the processor 2180 can run the operating system, the application program, the user interface display and touch response, and the multimedia information recommendation method of the embodiments of the present application. In addition, the processor 2180 is coupled with the display unit 2130.

[0265] In some possible implementation manners, each aspect of the multimedia information recommendation method provided in the present application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on a computer device, the computer program is used to make the computer device execute the steps in the multimedia information recommendation method according to various example embodiments of the present application described above in the specification, for example, the computer device can execute the steps in the multimedia information recommendation method according to various example embodiments of the present application described above in the specification. Figure 2or Figure 3 or Figure 9 the steps shown in FIG. 12.

[0266] The program product of the embodiments of the present application can employ any combination of one or more computer-readable media. The computer-readable media can be a computer- readable signal medium, or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include the following: an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0267] The program product of the embodiments of the present application can employ a compact disc read-only memory (CD-ROM) and include a computer program, and can be run on a computer device. However, the program product of the present application is not limited thereto, and in the present document, the computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with a command execution system, apparatus, or device.

[0268] The computer-readable signal medium can include a data signal propagated in baseband or propagated as a carrier wave in a propagated data signal, in which the computer-readable program is embodied. Such propagated signal can take a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium that is not a storage medium and that can communicate, propagate, or transport a program for use by or in connection with a command execution system, apparatus, or device.

[0269] The computer program contained in the computer-readable medium can be transmitted using any suitable medium, including, but not limited to, wireless, wired, optical fiber, RF, and the like, or any suitable combination thereof.

[0270] Computer programs for performing operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++, or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer programs can be executed entirely on a user's computer device, executed partially on a user device, executed as a stand-alone software package, executed partially on a user computer device and partially on a remote computer device, or entirely on a remote computer device or server. In situations in which remote computer device is used, the remote computer device can be connected to the user's computer device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the like, or the remote computer device can be connected to the Internet or an extranet through a provider service.

[0271] It should be noted that, although several units or sub-units of the apparatus are mentioned in the above detailed description, such division is merely exemplary and not mandatory. Indeed, according to an embodiment of the present application, features and functions of two or more units described above can be embodied in one unit. Conversely, features and functions of one unit described above can be further divided into units embodied by several units.

[0272] Moreover, while operations of the methods of the present application are described in a particular order in the figures, this is not required or implied, and the desired results can be achieved without performing all of the operations shown, or performing the operations in the particular order shown. Additionally or alternatively, certain steps can be omitted, combined into a single step, and / or split into multiple steps.

[0273] Those skilled in the art will appreciate that embodiments of the present application can be devised for a method, a system, or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) embodying computer programs of instructions executable on a computer.

[0274] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0275] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0276] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0277] While the preferred embodiments of the application have been described, additional variations and modifications can be employed, as will be appreciated by those of ordinary skill in the art, once armed with the foregoing disclosure. Accordingly, the appended claims as filed and as ultimately allowed in the patent granted are intended to encompass within their scope all such variations and modifications as are within the scope of the application.

[0278] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A multimedia information recommendation method characterized by comprising: The method comprises: in response to a play request carrying object basic attributes and play conditions of a play object, obtaining a plurality of candidate multimedia information to be recommended, and extracting respective recommendation features of the plurality of candidate multimedia information; for the plurality of candidate multimedia information, the following operations are respectively performed: selecting at least one key information from the object basic attributes, the play conditions and the play content of a candidate multimedia information according to at least one information identifier, and extracting key features of the at least one key information; each information identifier represents information that causes the deviation between the estimated conversion rate and the actual conversion rate of the recommended multimedia information according to the statistical history; the candidate multimedia information is any one of the plurality of candidate multimedia information; based on the key features, adjusting the weights of elements contained in the recommendation features of the candidate multimedia information, and obtaining the estimated conversion rate of the candidate multimedia information based on the weight adjustment result; based on the obtained at least one estimated conversion rate, selecting recommended multimedia information from the plurality of candidate multimedia information; the method comprises: performing twice mapping processing on the key features respectively to obtain first weight features and second weight features, the first weight features and the second weight features being respectively used to represent the weights of associated elements contained in the recommendation features of the candidate multimedia information and the key features; adjusting the weights of elements contained in the recommendation features of the candidate multimedia information according to the first weight features, and performing classification mapping processing on the weight-adjusted recommendation features to obtain preliminary estimated features; adjusting the weights of elements contained in the preliminary estimated features according to the second weight features, and performing classification mapping processing on the weight-adjusted preliminary estimated features to obtain the estimated conversion rate of the candidate multimedia information.

2. The method of claim 1, wherein, the method comprises: in response to the play request, a recall algorithm is used to obtain a plurality of candidate multimedia information meeting the object basic attributes and the play conditions; for the plurality of candidate multimedia information, the following operations are respectively performed: respectively extracting features of the object basic attributes, the play conditions and the play content of the candidate multimedia information to obtain corresponding object features, condition features and content features; performing feature cross processing on the object features, the condition features and the content features to obtain the recommendation features of the candidate multimedia information.

3. The method of claim 1 or 2, wherein, the method comprises: respectively extracting features of each key information in the at least one key information to obtain at least one preliminary key feature; perform feature cross processing on the at least one preliminary key feature to obtain the key feature.

4. The method of claim 1, wherein, Further comprising: determining sparse recommendation features according to the object basic attributes, the playing conditions, and the playing content of the one candidate multimedia information, the sparse recommendation features representing respective terms of the object basic attributes, the playing conditions, and the playing content of the one candidate multimedia information; then performing classification mapping processing on the preliminary estimated feature after weight adjustment to obtain the estimated conversion rate of the one candidate multimedia information, comprising: performing classification mapping processing on the preliminary estimated feature after weight adjustment to obtain an intermediate estimated feature; weighting the sparse recommendation features and the intermediate estimated feature to obtain a final estimated feature, and obtaining the estimated conversion rate of the one candidate multimedia information based on the final estimated feature.

5. The method of claim 1, wherein, the classification mapping processing on the recommendation feature after weight adjustment to obtain a preliminary estimated feature, comprising: performing normalization processing on the recommendation feature after weight adjustment to obtain a standard recommendation feature, and performing classification mapping processing on the standard recommendation feature to obtain the preliminary estimated feature; the classification mapping processing on the preliminary estimated feature after weight adjustment to obtain the estimated conversion rate of the one candidate multimedia information, comprising: performing normalization processing on the preliminary estimated feature after weight adjustment to obtain a standard preliminary estimated feature, and performing classification mapping processing on the standard preliminary estimated feature to obtain the estimated conversion rate of the one candidate multimedia information.

6. The method of claim 1, wherein, the two mapping processing on the key feature to obtain a first weight feature and a second weight feature, comprising: performing twice normalization processing on the key feature to obtain a first standard key feature and a second standard key feature, respectively, and performing classification mapping processing on the first standard key feature and the second standard key feature to obtain corresponding first preliminary weight features and second preliminary weight features, respectively; multiplying the first preliminary weight feature and the second preliminary weight feature by a set multiple, respectively, to obtain the first weight feature and the second weight feature.

7. The method of claim 1, wherein, the at least one key information comprises one or more of a conversion indicator, a multimedia object category, and a playing scene type; the conversion indicator represents an object behavior in the object basic attributes that causes conversion of the multimedia information to generate the asset corresponding to the multimedia information; the multimedia object category represents an object category of a multimedia object displayed by the multimedia information to the playing object in the playing content; the playing scene type represents a page scene type in which the playing object accesses a page to display the multimedia information in the playing condition.

8. A multimedia information recommendation apparatus characterized by comprising: the device comprises: a first feature extraction unit configured to, in response to a playing request carrying object basic attributes of a playing object and playing conditions, obtain a plurality of candidate multimedia information to be recommended, and extract recommendation features of the plurality of candidate multimedia information respectively; for the plurality of candidate multimedia information, the following operations are performed respectively: The second feature extraction unit is configured to select at least one key information from the object basic attribute, the playing condition, and the playing content of one candidate multimedia information according to at least one information identifier, and extract a key feature of the at least one key information; each information identifier represents information that causes a deviation between a predicted conversion rate and an actual conversion rate of the recommended multimedia information according to a statistical history; and the one candidate multimedia information is any one of the plurality of candidate multimedia information. The estimation processing unit is configured to adjust weights of elements included in a recommendation feature of the one candidate multimedia information based on the key feature, and obtain a predicted conversion rate of the one candidate multimedia information based on a weight adjustment result. The information selection unit is configured to select a recommended multimedia information from the plurality of candidate multimedia information based on the obtained at least one predicted conversion rate. The estimation processing unit is specifically configured to perform twice mapping processing on the key feature respectively, to obtain a first weight feature and a second weight feature, and the first weight feature and the second weight feature are respectively used to represent weights of associated elements included in the recommendation feature of the one candidate multimedia information and the key feature. The elements included in the recommendation feature of the one candidate multimedia information are adjusted in weights based on the first weight feature, and the recommendation feature after the weight adjustment is classified and mapped to obtain a preliminary estimated feature. The elements included in the preliminary estimated feature are adjusted in weights based on the second weight feature, and the preliminary estimated feature after the weight adjustment is classified and mapped to obtain the predicted conversion rate of the one candidate multimedia information.

9. The apparatus of claim 8, wherein, The first feature extraction unit is specifically configured to, in response to the playing request, acquire the plurality of candidate multimedia information that meets the object basic attribute and the playing condition by using a recall algorithm. The following operations are performed respectively for the plurality of candidate multimedia information: Features are extracted from the object basic attribute, the playing condition, and the playing content of the one candidate multimedia information respectively to obtain corresponding object features, condition features, and content features. The object features, the condition features, and the content features are cross-processed to obtain the recommendation feature of the one candidate multimedia information.

10. The apparatus of claim 8 or 9, wherein, The second feature extraction unit is specifically configured to extract features from each key information in the at least one key information respectively to obtain at least one preliminary key feature. The at least one preliminary key feature is cross-processed to obtain the key feature.

11. A computer-readable non-transitory storage medium, characterized in that, The computer-readable nonvolatile storage medium stores a program, and when the program runs on the computer, the computer is caused to implement the method in any one of claims 1 to 7.

12. A computer device, comprising: The computer-readable nonvolatile storage medium stores a program, and when the program runs on the computer, the computer is caused to implement the method in any one of claims 1 to 7. The computer-readable nonvolatile storage medium stores a program, and when the program runs on the computer, the computer is caused to implement the method in any one of claims 1 to 7. The computer-readable nonvolatile storage medium stores a program, and when the program runs on the computer, the computer is caused to implement the method in any one of claims 1 to 7.

13. A computer program product, characterised in that, The computer program is stored in a computer readable storage medium. When a processor of a computer device reads the computer program from the computer readable storage medium, the processor executes the computer program, so that the computer device executes the method according to any one of claims 1 to 7.

Images