WO2021135562A1 - 特征有效性评估方法、装置、电子设备及存储介质 - Google Patents

特征有效性评估方法、装置、电子设备及存储介质 Download PDF

Info

Publication number
WO2021135562A1
WO2021135562A1 PCT/CN2020/123721 CN2020123721W WO2021135562A1 WO 2021135562 A1 WO2021135562 A1 WO 2021135562A1 CN 2020123721 W CN2020123721 W CN 2020123721W WO 2021135562 A1 WO2021135562 A1 WO 2021135562A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
value
sample
combination
combined
Prior art date
Application number
PCT/CN2020/123721
Other languages
English (en)
French (fr)
Inventor
陈晓爽
于春功
Original Assignee
腾讯科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 腾讯科技(深圳)有限公司 filed Critical 腾讯科技(深圳)有限公司
Publication of WO2021135562A1 publication Critical patent/WO2021135562A1/zh
Priority to US17/703,822 priority Critical patent/US20220215296A1/en

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00Computing arrangements using knowledge-based models
    • G06N5/01Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound

Definitions

  • This application relates to machine learning technology in the field of Artificial Intelligence (AI), and in particular relates to a feature validity evaluation method, device, electronic device, and storage medium.
  • AI Artificial Intelligence
  • Artificial intelligence is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results.
  • Content recommendation is an application direction of artificial intelligence. It is specifically researched to recommend content of interest to the client, so as to achieve the purpose of content promotion.
  • the client's click-through rate of content is an important indicator for deciding to recommend content to the client.
  • the combination of different features may provide effective information for click-through rate estimation. Therefore, choosing the appropriate feature combination method to combine features is of great significance in content recommendation, and evaluating the effectiveness of the feature combination method is a key link.
  • the embodiments of the present application provide a feature validity evaluation method, device, electronic device, and storage medium, which can improve the efficiency of evaluating the validity of a large number of feature combinations.
  • the embodiment of the present application provides a feature validity evaluation method, and the method includes:
  • a feature combination mode set including multiple feature combination modes, the feature combination mode being a combination mode for the original feature of the content to be recommended;
  • the effectiveness of each feature combination manner is determined respectively, and the effectiveness is used to characterize the accuracy of content recommendation based on the features obtained by the corresponding feature combination manner.
  • the embodiment of the present application of the present invention also provides a feature validity evaluation device, the device includes:
  • the first construction unit is configured to construct a feature combination mode set including a plurality of feature combination modes, the feature combination mode being a combination mode for the original feature of the content to be recommended;
  • the acquiring unit is configured to acquire a feature value set corresponding to each of the feature combination modes, the feature value set includes the feature value of each combined feature, and the combined feature performs feature combination on the original feature based on the corresponding feature combination mode get;
  • the first determining unit is configured to determine the weight value of the corresponding combined feature based on the feature value of each of the combined features
  • the second construction unit is configured to construct a weight value set of a corresponding feature combination method based on the weight value of each combination feature;
  • the second determining unit is configured to determine the effectiveness of each feature combination manner based on the weight value set of each feature combination manner, and the effectiveness is used to characterize the feature combination based on the corresponding feature combination manner.
  • the accuracy of content recommendations is configured to determine the effectiveness of each feature combination manner based on the weight value set of each feature combination manner, and the effectiveness is used to characterize the feature combination based on the corresponding feature combination manner.
  • An embodiment of the present application also provides an electronic device, which includes:
  • Memory configured to store executable instructions
  • the processor is configured to implement the feature validity evaluation method provided in the embodiment of the present application when executing the executable instructions stored in the memory.
  • the embodiment of the present application also provides a storage medium storing executable instructions, which are used to implement the feature validity evaluation method provided by the embodiments of the present application when the executable instructions are executed by a processor.
  • the feature value set includes the feature value of each combination feature.
  • the weight value of the corresponding combination feature is determined, and the weight value set of the corresponding feature combination method is constructed based on the weight value of each combination feature.
  • the effectiveness of each feature combination method can be determined without time-consuming experiments and model training.
  • the effectiveness of a large number of feature combination methods can be evaluated in a short time, and the effectiveness of the combination method can be improved.
  • FIG. 1 is a schematic diagram of the architecture of a feature validity evaluation system 10 provided by an embodiment of this application;
  • FIG. 2 is a schematic diagram of the hardware structure of an electronic device 20 provided by an embodiment of the application.
  • FIG. 3 is a schematic diagram of the composition structure of a feature validity evaluation device 255 provided by an embodiment of this application;
  • FIG. 4 is a schematic structural diagram of a recommendation system provided by an embodiment of this application.
  • FIG. 5 is a schematic flowchart of a feature validity evaluation method provided by an embodiment of this application.
  • FIG. 6 is a schematic flowchart of a feature validity evaluation method provided by an embodiment of the application.
  • Content recommendation is an important application of the recommendation system. Among them, content recommendation is a method of selecting content that the user is interested in from a large amount of content to be recommended according to the needs of the user, and recommending the selected content to the user the process of. Among them, the content to be recommended may be a media file, or an item, etc., where the item may be a certain product or commodity.
  • Media files media in various forms (such as video, audio, graphics, and other media) available on the Internet, such as video files presented in the client, and articles including graphics (such as network articles or new Media articles), news, advertisements, etc.
  • Click-through rate refers to the ratio of the number of times that a certain content to be recommended in the client is clicked by the user to the number of times that it is displayed, that is, the click-through rate means that the content to be recommended is The probability that the user clicks.
  • the user's click-through rate is a very important indicator, especially in the online recommendation system, for the user to predict the click-through rate of a certain content to be recommended, so as to decide whether to recommend the content to be recommended to the user. An important way to improve user experience.
  • Click-through rate estimation is a method of predicting click-through rate. It can also be called a click-through rate estimation algorithm. This type of algorithm is based on the user and the relevant information of the content to be recommended, as well as other background data (such as the occurrence of click behaviors). The algorithm that predicts the probability that the user clicks on a content to be recommended in a certain context, such as the time of the user, the network environment the user is in, and the terminal device used (such as a mobile phone or a computer).
  • the click-through rate estimation algorithm can usually be implemented based on a certain mathematical model (such as the click-through rate estimation model).
  • the input features of the model can be original features or combined features to provide effective information for click-through rate estimation. .
  • the original feature refers to the feature directly obtained from the sample.
  • the original feature can include user features, content features, and background environment features.
  • user characteristics can be, for example, user identification (user ID), user's hobbies, user gender, user age, etc.
  • content characteristics can be, for example, content identification (content ID), content classification, content label, etc.
  • background environment characteristics can be, for example, It is the user's access time, the network environment (such as WiFi or cellular mobile network) where the user is visiting, and so on.
  • the combined feature refers to the feature obtained by combining one or more original features, and its value should traverse all possible value combinations of the original feature.
  • the value range of feature 1 is A and B
  • the value range of feature 2 is C, D, and E.
  • by comparing features 1 and Feature 2 performs feature combination to obtain multiple combined features.
  • the value range of the combined feature is AC, AD, AE, BC, BD, and BE; if the two original features are "feature 1: A, feature 2: C", Then the corresponding combination feature is AC.
  • feature 1 and feature 2 Based on the above-mentioned similar method, it is possible to realize the feature of multiple original features (such as two or more original features) Combine to get the corresponding multiple combination features.
  • the feature combination method refers to the method of combining the original features to obtain the combined feature
  • the combined feature refers to the specific feature value corresponding to the feature combination method.
  • “User ID-Content Tag” is a feature combination method
  • "Zhang San-Basketball” is the feature value of a combination feature corresponding to this feature combination method.
  • Feature engineering refers to the process of selecting appropriate original features and feature combination methods. In the recommendation system, because the efficiency of feature engineering directly affects the recommendation effect of the recommendation system, feature engineering is an extremely important processing link in the recommendation system.
  • the feature validity evaluation method provided by the embodiments of this application can be applied to the recommendation system, involving the field of artificial intelligence.
  • Artificial intelligence is a comprehensive technology of computer science. It attempts to understand the essence of intelligence and produce a new kind of human Intelligent machines that react in a similar way to intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.
  • Machine learning technology is an important application branch of artificial intelligence, involving many disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. Specializing in the study of how computers simulate or realize human learning behaviors in order to acquire new knowledge or skills, and reorganize the existing knowledge structure to continuously improve its own performance.
  • Machine learning technology is the core of artificial intelligence and the fundamental way to make computers intelligent.
  • the application of machine learning technology is widespread in all fields of artificial intelligence.
  • the embodiment of the application realizes the evaluation of the effectiveness of a large number of feature combination methods through machine learning technology, without the need for time-consuming experiments and model training, can evaluate the effectiveness of a large number of feature combination methods in a short time, and can improve the evaluation of the effectiveness of a large number of feature combination methods. The efficiency of the evaluation of the effectiveness of the feature combination method, thereby improving the recommendation effect of the recommendation system.
  • the usual problem is how to evaluate the effectiveness of the feature combination method, so as to determine whether the combined feature based on the feature combination method can be added to the click-through rate estimation model to determine whether the combination feature can be added to the click-through rate prediction model according to the evaluation result.
  • Rate estimates provide effective information.
  • related technologies provide a scheme for evaluating the effectiveness of feature combination methods.
  • the effectiveness of the feature combination method is mainly evaluated based on the experimental results of a large number of experiments. It is usually necessary for practitioners to screen out some candidate features based on the combination of the feature combination based on experience, which requires practitioners Having a high degree of familiarity with the product and a good sensitivity to the data can filter suitable features.
  • the method of manual feature selection requires a lot of trial and error. , Its quality and speed are difficult to be effectively guaranteed, and the experiment is very time-consuming.
  • the speed of evaluating the effectiveness of each feature combination method through experimental methods will be very slow and consume a lot of computing resources.
  • the number of feature combination methods that can be evaluated in a limited time is extremely limited, resulting in a greatly reduced efficiency in evaluating the effectiveness of a large number of feature combination methods.
  • the evaluation of the effectiveness of the feature combination method can be achieved through a full training method.
  • the original features are combined based on the feature combination method to be evaluated to obtain different features
  • the full training method obtains a model containing the feature through a full model training, and evaluates the prediction effect of the model as a basis for evaluating the effectiveness of the feature combination method.
  • the model training usually adopts the stochastic gradient descent (SGD, Stochastic Gradient Descent) or its deformation method.
  • SGD stochastic gradient descent
  • Stochastic Gradient Descent Stochastic Gradient Descent
  • the main problem of the full training method is that its calculation speed is slow, and it usually takes a lot of time to train. Therefore, it restricts the ability of practitioners to try more features, which in turn affects the efficiency of feature engineering.
  • the full training method is the Gradient Boosting Decision Tree method (referred to as the decision tree method) as an example for description.
  • the decision tree method maps the original feature to a new feature according to multiple judgment conditions. For example, the decision tree method divides all samples into 10 groups: samples that meet the gender of male and are between 20-25 years old, as the first group; The samples that meet the gender female and age between 20-25 years old are used as the second group; etc., when judging the effectiveness of a certain candidate feature combination method, the decision tree method uses information gain, information gain ratio, Gini coefficient, etc. Parameters are evaluated.
  • the decision tree method is mainly suitable for feature engineering of continuous features and discrete features with only a few values (such as gender). This is more effective in the early recommendation system, but the current recommendation system has a large number of discrete features called ID features, such as user ID, content ID, etc., these features have a large number of possible values, and the decision tree method The number of values of is very sensitive. Therefore, when judging the validity, the model complexity is so high that it cannot be applied to the online recommendation system. It also faces problems such as inaccurate judgments, making it difficult to apply the decision tree method to industrialization. Recommended in the system.
  • the embodiments of the present application provide a feature validity evaluation method by obtaining a feature value set corresponding to each feature combination mode, where the feature value set includes the feature value of each combined feature, based on The feature value of each combination feature is determined, and the weight value of the corresponding combination feature is determined, and the weight value set of the corresponding feature combination method is constructed based on the weight value of each combination feature.
  • the effectiveness of each feature combination method can be determined, and the efficiency of evaluating the effectiveness of a large number of feature combination methods can be improved; the appropriate target feature combination method can also be screened based on the effectiveness of the feature combination method, and obtained based on the target feature combination method
  • the target combination feature is used in the recommendation system for content recommendation, which significantly improves the efficiency of feature engineering, realizes more accurate click-through rate estimation, provides strong technical support for the recommendation system, and improves the recommendation effect of the recommendation system.
  • the electronic devices provided in the embodiments of the present application can be implemented as notebook computers, tablet computers, desktop computers, set-top boxes, and mobile devices (such as mobile phones, Various types of terminal devices such as portable music players, personal digital assistants, dedicated messaging devices, portable game devices, etc., can also be implemented as servers.
  • the servers can include but are not limited to any hardware devices that can perform calculations, such as independent
  • the physical server can also be a server cluster or distributed system composed of multiple physical servers, or it can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain names Services, security services, and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms; of course, the feature effectiveness evaluation method of the embodiment of the present application can also be implemented in collaboration with a terminal device and a server.
  • the server may be a cloud server, but the embodiment of the present application is not limited to a cloud server.
  • FIG. 1 is a schematic diagram of an optional architecture of the feature validity evaluation system 10 provided by an embodiment of the application.
  • the terminal 100 exemplarily shows the terminal 100-1 and the terminal 100-1).
  • the server 300 is connected through the network 200, where the network 200 may be a wide area network or a local area network, or a combination of the two, and uses a wireless link to realize data transmission.
  • the terminal 100 (such as the terminal 100-1) is configured to trigger a content acquisition request to the server 300 based on the content acquisition instruction, so as to request to acquire the content of its own interest.
  • the server 300 is configured to, in response to a content acquisition request sent by the terminal 100, construct a feature combination mode set including multiple feature combination modes; obtain a feature value set corresponding to each feature combination mode; determine the corresponding feature value based on the feature value of each combination feature Based on the weight value of each combination feature, construct the weight value set of the corresponding feature combination method; based on the weight value set of each feature combination method, determine the effectiveness of each feature combination method.
  • the server 300 is also configured to determine the target feature combination method based on the effectiveness of each feature combination method, and determine the target recommendation content based on the target feature combination method.
  • the effectiveness of each feature combination method can be ranked. Obtain the corresponding ranking result, and then filter the feature combination mode of the target number from the feature combination mode set based on the ranking result, use the filtered feature combination mode as the target feature combination mode, and feature the original features of the recommended content based on the target feature combination mode Combine to obtain the target combination feature, to determine the click rate of the target user to recommend the content based on the target combination feature, and then determine the target recommended content based on the click rate, and push the target recommended content to the terminal 100.
  • the server 300 may be a separately configured server supporting various services, or may be configured as a server cluster.
  • the terminal 100 may present the target recommended content in a graphical interface 110 (for example, the graphical interface 110-1 of the terminal 100-1 and the graphical interface 110-2 of the terminal 100-2).
  • a graphical interface 110 for example, the graphical interface 110-1 of the terminal 100-1 and the graphical interface 110-2 of the terminal 100-2).
  • the electronic device can be implemented as a terminal device, can also be implemented as a server, and can also be implemented as a coordinated implementation of the terminal device and the server shown in FIG. 1 above.
  • FIG. 2 is a schematic diagram of an optional hardware structure of the electronic device 20 provided by an embodiment of the application. It can be understood that FIG. 2 only shows an exemplary structure of the electronic device instead of the entire structure. 2 shows part or all of the structure.
  • the electronic device 20 provided in the embodiment of the present application may include: at least one processor 210, a memory 250, at least one network interface 220, and a user interface 230.
  • the various components in the electronic device 20 are coupled together through the bus system 240.
  • the bus system 240 is configured to implement connection and communication between these components.
  • the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for clear description, various buses are marked as the bus system 240 in FIG. 2.
  • the processor 210 may be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP, Digital Signal Processor), or other programmable logic devices, discrete gates or transistor logic devices, or discrete hardware Components, etc., where the general-purpose processor may be a microprocessor or any conventional processor.
  • DSP Digital Signal Processor
  • the user interface 230 includes one or more output devices 231 that enable the presentation of media content, including one or more speakers and/or one or more visual display screens.
  • the user interface 230 also includes one or more input devices 232, including user interface components that facilitate user input, such as a keyboard, a mouse, a microphone, a touch screen display, a camera, and other input buttons and controls.
  • the memory 250 may be removable, non-removable, or a combination thereof.
  • Exemplary hardware devices include solid-state memory, hard disk drives, optical disk drives, and so on.
  • the memory 250 optionally includes one or more storage devices that are physically remote from the processor 210.
  • the memory 250 includes volatile memory or non-volatile memory, and may also include both volatile and non-volatile memory.
  • the non-volatile memory may be a read only memory (ROM, Read Only Memory), and the volatile memory may be a random access memory (RAM, Random Access Memory).
  • ROM read only memory
  • RAM Random Access Memory
  • the memory 250 described in the embodiment of the present application is intended to include any suitable type of memory.
  • the memory 250 can store data to support various operations. Examples of these data include programs, modules, and data structures, or a subset or superset thereof, as illustrated below.
  • the operating system 251 includes system programs for processing various basic system services and performing hardware-related tasks, such as a framework layer, a core library layer, a driver layer, etc., for implementing various basic services and processing hardware-based tasks;
  • the network communication module 252 is configured to reach other computing devices via one or more (wired or wireless) network interfaces 220.
  • Exemplary network interfaces 220 include: Bluetooth, Wireless Compatibility Authentication (WiFi), and Universal Serial Bus ( USB, Universal Serial Bus), etc.;
  • the presentation module 253 is configured to enable the presentation of information via one or more output devices 231 (for example, a display screen, a speaker, etc.) associated with the user interface 230 (for example, a user interface for operating peripheral devices and displaying content and information) );
  • output devices 231 for example, a display screen, a speaker, etc.
  • the user interface 230 for example, a user interface for operating peripheral devices and displaying content and information
  • the input processing module 254 is configured to detect one or more user inputs or interactions from one of the one or more input devices 232 and translate the detected inputs or interactions.
  • the feature validity evaluation device provided by the embodiments of the present application can be implemented in software.
  • FIG. 2 shows the feature validity evaluation device 255 stored in the memory 250, which can be in the form of programs, plug-ins, etc.
  • Software including a series of software modules.
  • FIG. 3 is a schematic diagram of an optional structure of the feature validity evaluation device 255 provided by an embodiment of the application.
  • the feature validity evaluation device 255 may include a first construction unit 2551, an acquisition unit 2552, and a first construction unit 2551.
  • the determination unit 2553, the second construction unit 2554, and the second determination unit 2555 have logical functions. Therefore, any combination or split can be performed according to the functions implemented by each software module.
  • the functions of each unit in the feature validity evaluation device 255 provided in the embodiment of the present application shown in FIG. 3 will be described below.
  • the feature validity evaluation device 255 provided in the embodiment of the present application may be implemented in hardware.
  • the feature validity evaluation device 255 provided in the embodiment of the present application may be in the form of a hardware decoding processor.
  • a processor which is programmed to execute the feature validity evaluation method provided in the embodiments of the present application.
  • a processor in the form of a hardware decoding processor may adopt one or more application specific integrated circuits (ASIC, Application Specific Integrated Circuit), DSP, Programmable Logic Device (PLD, Programmable Logic Device), Complex Programmable Logic Device (CPLD, Complex Programmable Logic Device), Field Programmable Gate Array (FPGA, Field-Programmable Gate Array) or other electronic components.
  • ASIC Application Specific Integrated Circuit
  • DSP Digital Signal Processing
  • PLD Programmable Logic Device
  • CPLD Complex Programmable Logic Device
  • FPGA Field-Programmable Gate Array
  • the recommendation system can be implemented as a server, and the server receives user requests sent by the terminal, according to user requests (user requests can include user ID, current time and other information ), extract the content to be recommended from the content database, and pass the user request and the content to be recommended to the feature center; the feature center organizes the user request and related information of the content to be recommended into a structured form (such as a string list, or In the form of key-value pairs), the original features of the content to be recommended (such as user ID, user age, content ID, content tag, etc.) are obtained, and the original features are passed to the feature combination module.
  • user requests can include user ID, current time and other information
  • extract the content to be recommended from the content database
  • pass the user request and the content to be recommended to the feature center the feature center organizes the user request and related information of the content to be recommended into a structured form (such as a string list, or In the form of key-value pairs)
  • the original features of the content to be recommended such as user ID, user age, content ID
  • a feature selection module is added to the recommendation system.
  • the feature selection module realizes the screening of feature combination methods to find effective feature combination methods.
  • the effective feature combination method is "user ID-content tag”.
  • the feature combination module can combine one or more original features based on the effective feature combination provided by the feature selection module to obtain multiple combined features (for example, based on the effective feature combination mode "User ID-Content Tag"
  • the combined feature obtained by the combination is "Zhang San-Basketball”
  • the combined feature is input as the input feature to the click-through rate estimation module to estimate the user's click-through rate for the recommended content; next, the re-ranking module According to the obtained click-through rate, the content to be recommended is sorted, and a target quantity of content is selected from the content to be recommended as the target recommended content; then, the recommendation system returns the target recommended content to the user.
  • the feature selection module selects an effective feature combination method (such as user ID-content tag) according to the offline log data recorded by the feature center, and then updates the feature combination method in the feature combination module to make the feature combination
  • the module outputs effective combined features to perform content recommendation model training or online click rate estimation based on the combined features.
  • the recommendation system will store the original features, recommendation results, and user clicks used in each recommendation process as logs. These logs can be used to obtain each combination of features.
  • Value for example, the feature combination mode is "user ID-content label", the feature value of a certain combination feature corresponding to the effective feature combination mode is "Zhang San-Basketball" weight value, in order to quickly estimate how much
  • the effectiveness of each feature combination method is used to determine the target feature combination method based on the effectiveness of multiple feature combination methods to guide subsequent recommendation operations.
  • the log system extracts user click data (ie samples) within a certain time range, and organizes each sample into the following format: Among them, i represents the i-th log; y (i) represents the click result of whether the user clicked; Is the estimated click-through rate of the sample, which is a value between 0 and 1; x (i) is the original feature sorted out by the feature center, which can be recorded in the form of key-value pairs, such as "User ID: Zhang 3. Content ID: Content A; User's concerned media: Media A, Media B; Content tag: Basketball" etc. It should be noted that the original feature can be a single value (such as a user ID) or multiple values (such as a media that the user pays attention to, usually more than one).
  • the click-through rate prediction model usually predicts the probability of a user clicking on content in the current environment through a certain algorithm based on input characteristics (such as user characteristics, content characteristics, background environment characteristics, etc.), that is, the click-through rate, which is between 0 and 1. Real number.
  • the click-through rate prediction model is usually modeled as the following form: There are many types of common click-through rate estimation models, for example, Logistic Regression (LR), Deep Neural Network (DNN) and so on. It should be noted that the embodiment of the application does not have any restriction on the type of the click-through rate estimation model.
  • the feature validity evaluation method may be implemented by a terminal or a server, and of course, the server and the terminal may cooperate with each other.
  • the server implementation is taken as an example below, such as implementation by the server 300 in FIG. 1, and description will be made in combination with the steps shown in FIG.
  • step 501 the server constructs a feature combination mode set including multiple feature combination modes.
  • the feature combination mode is a combination mode for the original features of the content to be recommended.
  • the server uses the above recommendation system for content recommendation, it needs to extract the relevant features of the content to be recommended, such as user features and content features, to combine user features and content features to obtain related combined features to improve content recommendation Accuracy.
  • the server may construct a feature combination mode set in the following manner: obtain multiple original features; determine multiple feature combination modes obtained by combining at least two original features from the multiple original features; based on multiple original features; The original feature and multiple feature combinations are used to construct a set of feature combination methods.
  • multiple original features are original features including user features and content features.
  • the multiple original features can be user features "user identification”, “media that the user pays attention to”, content features “posting media”, “content” Label” etc.
  • multiple original features include “user identification, media that the user pays attention to, posting media, content tag", and arbitrarily select two original features and combine them to obtain “user identification-media that the user pays attention to” and “user identification” -Multiple feature combination methods such as "posting media” and "media that users pay attention to-content tags”.
  • three original features can also be arbitrarily selected for combination, which is not limited here.
  • step 502 the feature value set corresponding to each feature combination mode is obtained, and the feature value set includes the feature value of each combined feature.
  • the combined feature is obtained by combining the original features based on the corresponding feature combination mode.
  • the feature value of the combined feature may be extracted based on historical log data, or may be part of historical log data obtained by sampling historical log data.
  • each original feature will not be combined with itself, but if an original feature includes multiple values, the original feature can be combined with itself. Therefore, a single original feature (that is, it is not combined with other The combination of the original features) can also be regarded as a special combination feature (that is, the feature obtained by "combining" only one original feature). For example, if the original feature is “media that the user pays attention to”, then the combined feature “media that the user pays attention to-the media that the user pays attention to” can be constructed. If in a sample, the media that the user pays attention to are media A and media B, then follow the above The feature value of the combined feature of the sample constructed by the method can be "media A-media A", “media A-media B” and "media B-media B".
  • step 503 based on the feature value of each combined feature, the weight value of the corresponding combined feature is determined.
  • the server may determine the weight value of each combination feature in the following manner: input the feature value of each combination feature into the weight calculation model to obtain the weight value corresponding to each combination feature output by the weight calculation model. That is to say, in the embodiment of the present application, the feature value of each combined feature can be input into the weight calculation model through a pre-trained weight calculation model, so as to obtain the weight value corresponding to each combined feature.
  • the server may train to obtain the weight calculation model in the following manner: input the feature value sample marked with the target weight value into the weight calculation model to obtain the weight value of the corresponding feature value sample output by the weight calculation model; Based on the output weight value and the target weight value, determine the value of the loss function of the weight calculation model; based on the value of the loss function, update the model parameters of the weight calculation model.
  • a weight calculation model can be built in advance based on deep learning methods, such as a convolutional neural network model used to calculate weight values, including an input layer, a hidden layer, and an output layer, which are used to calculate the weight value of each combination feature.
  • a weight calculation model is obtained.
  • the weight calculation model is trained based on the collected feature value samples to obtain optimized weight calculation model parameters.
  • the input feature value sample can be a sample only for a certain feature combination, or it can be a sample of all feature combinations, usually to speed up the model training, you can Only train for a certain feature combination method.
  • the server can obtain the weight value of the corresponding feature value sample in the following way: first obtain a large number of feature value samples, for example, can be obtained by sampling the relevant historical log data of certain content to be recommended, these features The value samples are marked with corresponding target weight values; before training, a large number of samples collected can be divided into training set and test set according to a certain ratio, and the feature value samples marked with target weight values in the training set are input into In the weight calculation model, the weight value of the corresponding feature value sample output by the weight calculation model is obtained. Further, the process of model training is the process of updating and adjusting each parameter in the model. The training sample data is input to the input layer of the weight calculation model, and after the hidden layer, it finally reaches the output layer and outputs the result.
  • a loss function is introduced, based on the weight value and target value of the feature value sample output by the weight calculation model
  • the weight value determines the value of the loss function; based on the value of the loss function, the backpropagation algorithm is used to update the parameters of the weight calculation model layer by layer until the loss function converges to achieve the constraint and adjustment of the parameters of the weight calculation model, thereby obtaining the calculation
  • a high-precision weight calculation model to determine the weight value of each combination feature based on the weight calculation model.
  • the weight value of each combination feature can be obtained by optimizing the following function:
  • the weight values corresponding to the feature values of all the combined features of each sample in the feature combination mode such as F are added to obtain the predicted sample score, that is, the score of the predicted feature combination mode, exemplary, can be
  • the score of the corresponding feature combination method is predicted by the following method: Among them, w F, j represents the weight value of the combined feature; F represents the feature combination mode; j represents the feature value of each combined feature corresponding to the feature combination mode F; Indicates that the value of the i-th sample under the feature combination method F includes the combined feature j;
  • logarithmic error function which has the following form:
  • the optimal solution of the objective function of the above formula (3) generally cannot be expressed by an analytical expression. Therefore, its solution usually adopts an iterative method, such as stochastic gradient descent method, to obtain the weight value of each combination feature, which causes a large amount of model training to consume In feature engineering, it is necessary to evaluate the effectiveness of a large number of feature combination methods.
  • the method of using model training has high computational cost and slow speed.
  • the embodiment of the present application also provides a statistical method to determine the weight value of each combination feature, thereby speeding up the calculation of the weight value, and ensuring the efficiency of evaluating the effectiveness of a large number of feature combination methods. Promote.
  • the following describes the use of statistics to determine the weight value of each combination feature.
  • the server may determine the weight value of each combined feature in the following manner: determine the positive sample statistic corresponding to each feature value of the combined feature, and the negative sample statistic corresponding to each feature value of the combined feature; based on The positive sample statistic and the negative sample statistic corresponding to the value of each feature obtain the weight value of the corresponding combined feature.
  • the positive sample represents the sample data corresponding to the clicked content among the multiple content to be recommended displayed during the display process of the content to be recommended;
  • the negative sample represents the sample data corresponding to the content to be recommended during the display process of the content to be recommended , The sample data corresponding to the unclicked content among the displayed multiple content to be recommended.
  • the embodiment of the present application divides all samples corresponding to the collected feature value sets into positive samples and negative samples according to the click status of the content to be recommended.
  • the feature combination mode is "user identification-product type”
  • the feature value of each combination feature corresponding to the feature combination mode "user identification-product type” may include “Zhang San-cosmetics, Zhang San -Snacks, Zhang San-clothing, Zhang San-sports shoes", when content recommendation is based on the value of each feature, it will recommend “cosmetics, snacks, clothing and sports shoes” to "Zhang San” respectively, if "Zhang San” If you clicked on "cosmetics, snacks", and did not click on the others, then "Zhang San-cosmetics, Zhang San-snacks” are positive samples, and “Zhang San-clothing, Zhang San-sports shoes” are negative samples.
  • the analytical solution of the objective function is the weight value of the combined feature.
  • the server can determine the positive sample statistics including the feature value, which is the positive sample statistics corresponding to the feature value.
  • the negative sample statistic including the characteristic value is determined as the negative sample statistic corresponding to the characteristic value.
  • the weight value of the corresponding combined characteristic is determined based on the positive sample statistic and the negative sample statistic.
  • the following formula can be used to calculate the weight value of the corresponding combined feature:
  • w F, j is the weight value of the combined feature
  • F is the feature combination mode
  • j is the feature value of each combined feature corresponding to the feature combination mode F
  • weight value of the corresponding combination feature can be calculated according to formula (5), which is relatively simple and convenient. This is also the advantage of the statistical method over the weight calculation model training method.
  • the server can determine the positive sample statistics corresponding to each feature value of the combined feature in the following manner, and the combination The negative sample statistics corresponding to each feature value of the feature: determine the feature value statistics of the combined feature in the positive sample including the feature value, and the feature value statistics of the combined feature in the negative sample including the feature value; based on The feature value statistic of the combined feature in the positive sample including the feature value is determined, and the positive sample statistic corresponding to the feature value is determined; the feature value is determined based on the feature value statistic of the combined feature in the negative sample including the feature value The corresponding negative sample statistics.
  • multiple feature values of combined features may appear in a sample. For example, if Zhang San watched a certain variety show in which a certain star participated, the content tags of the sample include “a certain star” and "a certain variety show”. At the same time, the feature value of the combined feature corresponding to the sample "user ID-content tag” can include "Zhang San-a certain star” and "Zhang San-a variety show", and its click behavior may also be derived from these two characteristics One of the values. Therefore, the contribution of this sample to "Zhang San-a star” should be less than a sample that only contains the characteristic value of "Zhang San-a star” (for example, in another sample, Zhang San watched the star once personally Interview).
  • the weight value w F, j of the corresponding combined feature is calculated by the above formula (5).
  • the more the feature value of a sample the more it takes the weight of the value j for one of the features
  • the weight value w F,j of the corresponding combination feature calculated according to formula (5) is positive infinity (this means that the recommendation system determines that the click rate of the content is 100%), which is unreasonable result.
  • ⁇ 1 and ⁇ 2 are non-negative real numbers.
  • this formula (10) adds two new items, namely ⁇ 1
  • these two new items will play a leading role, making the weight value of the combined feature tend to 0; and when the number of samples is large, the original objective function term will play a role.
  • the weight value of the combined feature tends to the weight value calculated by formula (5).
  • the server can obtain the corresponding statistics based on the positive sample statistics and the negative sample statistics corresponding to each feature value in the following manner
  • the weight value of the combined feature when When it is less than ⁇ 1 , the weight value of the combined feature is zero; When greater than ⁇ 1 , the weight value of the combined feature is the ratio of the difference to the first sum, and the difference is And ⁇ 1 , the first sum value is the sum of N′ and ⁇ 2 ; when When it is less than - ⁇ 1 , the weight value of the combined feature is the ratio of the second sum value to the first sum value, and the second sum value is Sum with ⁇ 1 .
  • ⁇ 1 and ⁇ 2 are both non-negative real numbers; N′ is obtained according to the ratio of the product value to the third sum value; the product value is the positive sample statistic corresponding to the characteristic value, and the negative sample statistic corresponding to the characteristic value.
  • the product value of; the third sum value is the sum of the positive sample statistic corresponding to the characteristic value and the negative sample statistic corresponding to the characteristic value; According to the positive sample statistic corresponding to the characteristic value, the ratio of the negative sample statistic corresponding to the characteristic value is obtained.
  • the embodiment of the present application may use the following method to replace formula (5) to approximate the weight value w F,j of the corresponding combined feature:
  • F is the feature combination mode
  • j is the feature value of each combination feature corresponding to the feature combination mode F
  • the server can determine the positive sample statistics corresponding to each feature value of the combined feature in the following manner, and the feature selection of the combined feature Negative sample statistics corresponding to the value: determine the corresponding first prediction accuracy when the positive sample including the feature value is the training sample, and the second prediction accuracy when the negative sample including the feature value is the training sample; based on the first prediction Accuracy, the positive sample statistic corresponding to the feature value is determined; based on the second prediction accuracy, the negative sample statistic corresponding to the feature value is determined.
  • the first prediction accuracy represents the accuracy of the recommendation based on the positive sample of the feature value
  • the second prediction accuracy represents the accuracy of the recommendation based on the negative sample of the feature value
  • the evaluation of the importance of a certain combination feature based on the combination of feature combinations should be based on the existing click-through rate estimation model, that is, the added combination feature can play a role in the click-through rate estimation model.
  • the objective function shown in formula (4) can be adjusted to the following form to consider the impact of the click-through rate estimation model:
  • the above objective function (12) also does not have an analytical solution.
  • the embodiment of the present application uses the following formula to modify the positive sample statistics and the negative sample statistics corresponding to the characteristic values:
  • the weight value w F, j of the corresponding combination feature can be calculated by formula (5).
  • the weight value w F, j of the corresponding combination feature can be calculated by formula (5).
  • the server can determine the positive sample statistic corresponding to each feature value of the combined feature in the following manner, and the corresponding value of each feature of the combined feature Negative sample statistics: determine the weight value of the first sample sample and the weight value of the second sample sample; based on the weight value of the first sample sample, determine the positive sample statistic corresponding to the feature value; based on the weight value of the second sample sample , Determine the negative sample statistic corresponding to the characteristic value.
  • the first sampling sample represents a sample of a first proportion drawn from a positive sample including a characteristic value
  • the second sampling sample represents a sample of a second proportion drawn from a negative sample including the characteristic value
  • the embodiment of the present application proposes to determine the positive sample statistics and the negative sample statistics corresponding to the feature value through a sampling method. At this time, with The calculation formula is:
  • the weight value w F, j of the corresponding combined feature can be calculated by formula (5).
  • ⁇ % samples can be randomly selected from the positive samples containing feature j, and ⁇ % of the samples can be randomly selected from the positive samples containing feature j.
  • a sample of ⁇ % is randomly selected from the negative sample of j, and then calculated according to the above formulas (15) and (16) with
  • ⁇ % and ⁇ % can be set according to actual needs, and are not limited here.
  • the sample weights are different in the above different solutions. In formulas (6)(7), the weight of the sample is 1; in formulas (8)(9)(13)(14), the weight of the sample is the value of the corresponding formula inside the summation sign.
  • the server can also determine the positive sample statistics corresponding to each feature value of the combined feature and the negative sample statistics corresponding to each feature value of the combined feature in the following way: determine that the first part of the sample includes the feature value The weight value of the positive sample of the value, and the weight value of the first sample in the second part of the sample; determine the weight value of the negative sample in the first part of the sample including the feature value, and the weight of the second sample in the second part of the sample Value; based on the weight value of the positive sample including the feature value in the first part of the sample, and the weight value of the first sample sample in the second part of the sample, determine the positive sample statistic corresponding to the feature value; based on the first part of the sample including the feature The weight value of the negative sample and the weight value of the second sample in the second part of the sample are used to determine the negative sample statistic corresponding to the characteristic value.
  • the first sampling sample represents the first proportion of samples drawn from the positive samples that belong to the second part of the sample and includes the feature value
  • the second sampling sample represents the sample that belongs to the second part of the sample and includes the feature value The second proportion of samples drawn from the negative samples.
  • Part A the overall sample into two parts, namely Part A and Part B. All samples in Part A participate in statistics, while samples in Part B are sampled.
  • samples with a higher degree of importance for example, the part with a very large prediction deviation of the click-through rate estimation model
  • Part B the samples with a lower degree of importance
  • the weight value w F,j of the corresponding combined feature can be calculated by formula (5).
  • the above-mentioned improvement plan for each combination feature in each sample corresponding to multiple feature values, and the improvement plan for the inaccurate prediction accuracy of the click-through rate estimation model for a certain feature are all adjustments with That is, adjust the weight of each sample in the summation number.
  • the corresponding sample weight can be multiplied, which can be calculated by the following formula:
  • step 504 based on the weight value of each combined feature, a set of weight values of the corresponding feature combination mode is constructed.
  • the weight value set of the feature combination mode includes the weight value corresponding to each combined feature, and the combined feature can be obtained by combining the original features based on the corresponding feature combination mode.
  • step 505 based on the set of weight values of each feature combination mode, the effectiveness of each feature combination mode is determined respectively.
  • the effectiveness of the feature combination mode is used to characterize the accuracy of content recommendation based on the features obtained by the combination of the corresponding feature combination mode.
  • the server may determine the effectiveness of each feature combination method by: weighting the weight values of all the combination features in the weight value set to obtain a score corresponding to the feature combination method; and a score based on each feature combination method , Respectively determine the effectiveness of each feature combination.
  • the server can respectively determine the effectiveness of each feature combination method based on the score of each feature combination method in the following way: compare the score of each feature combination method with the target score to obtain the comparison result corresponding to each feature combination method; based on the comparison As a result, the effectiveness of the corresponding combination of features is determined.
  • each weight value in the weight value set can be weighted to obtain the corresponding feature combination Method of scoring; compare the scores of each feature combination method with the corresponding target score to obtain a comparison result, so as to determine the effectiveness of the corresponding feature combination method based on the comparison result. It can be seen that the application of the foregoing embodiment realizes the calculation of the effectiveness of each feature combination mode, thereby realizing the screening of the target feature combination mode according to the effectiveness of each feature combination mode.
  • the area under the ROC curve and the coordinate axis can be used to calculate the comparison results of the scores of each feature combination method and the target score, so as to determine the feature combination method Effectiveness.
  • the score of the feature combination mode is used to characterize the likelihood that the content to be recommended will be clicked by the user when recommending based on the combined features obtained by the corresponding feature combination mode.
  • the feature effectiveness evaluation method further includes: after determining the effectiveness of each feature combination mode, based on the ranking of the effectiveness of each feature combination mode, filter the feature combination mode with the target number from the feature combination mode set.
  • a target feature combination method based on the target feature combination method, the original features are combined to obtain the target combination feature, and content recommendation is based on the target combination feature.
  • each feature combination method is sorted in descending order of effectiveness, and the features are The multiple feature combination methods in the combination method set are screened to obtain a feature combination method with a high effective target number as the target feature combination method.
  • the feature combination is performed based on the determined target feature combination method to obtain the content recommendation method. Target combination characteristics.
  • the combination method is determined as the target feature combination method.
  • the validity threshold can also be set in advance, the validity of each feature combination mode is compared with the validity threshold, and each feature combination mode whose validity reaches the validity threshold is determined as the target feature combination mode.
  • the feature effectiveness evaluation method further includes: after determining the effectiveness of each feature combination mode, based on the ranking of the effectiveness of each feature combination mode, select the target number of feature combinations from the feature combination mode set.
  • the method is used as the first candidate feature combination method; based on the first candidate feature combination method and the original feature, multiple second candidate feature combination methods are generated; from the multiple second candidate feature combination methods, the feature combination method that meets the screening conditions is selected as the combination method Target feature combination method: Based on the target feature combination method, the original features are combined to obtain the target combination feature, which can be used for content recommendation based on the target combination feature.
  • the embodiments of the present application amplify the first candidate feature combination modes to obtain more feature combination modes as the second Candidate feature combination method.
  • the server may generate multiple second candidate feature combination modes in the following manner: based on the first candidate feature combination mode, the original features are combined to obtain the combined feature; the combined feature is determined to be obtained by combining at least one original feature Multiple feature combination methods of; Based on multiple feature combination methods and the first candidate feature combination method, multiple second candidate feature combination methods are generated. It can be seen that the embodiment of the present application generates more second candidate feature combination methods based on the first candidate feature combination method and the original features, thereby increasing the diversity of feature combination methods, so as to obtain more effective feature combination methods, and provide The accuracy of content recommendations.
  • the target feature combination mode can be multiple, and the original feature can be combined based on each target feature combination mode, or the original feature can be combined based on the most effective target feature combination mode to obtain the target combination.
  • the server can perform content recommendation based on the target combination feature in the following manner: the target combination feature is used as the input feature and input into the click-through rate estimation model to obtain the click-through rate of the content to be recommended by the target user; based on the click-through rate, Select a target amount of content from the content to be recommended as the target recommended content; return the target recommended content to the target user.
  • the feature value set corresponding to each feature combination mode is obtained, where the feature value set includes the feature value of each combined feature , Determine the weight value of the corresponding combination feature based on the feature value of each combination feature, thereby constructing the weight value set of the corresponding feature combination method based on the weight value of each combination feature, so, based on the weight value set of each feature combination method ,
  • the effectiveness of each feature combination method can be determined without time-consuming experiments and model training. It can evaluate the effectiveness of a large number of feature combination methods in a short time and improve the effectiveness of a large number of feature combination methods. The efficiency of the recommendation system, thereby improving the recommendation effect of the recommendation system.
  • FIG. 6 is a schematic flowchart of the feature validity evaluation method provided by the embodiment of the present application.
  • the feature The effectiveness evaluation method can be implemented by the terminal, can also be implemented by the server, and can also be implemented by the server and the terminal in cooperation.
  • the cooperative implementation of the server and the terminal is taken as an example.
  • the cooperative implementation of the terminal 100-1 and the server 300 in FIG. 1 is combined with the steps shown in FIG. 6 to describe the implementation of the feature validity evaluation method provided by the embodiment of the present application.
  • the feature validity evaluation method provided by the embodiment of the present application may include the following steps:
  • step 601 the terminal initiates a content acquisition request to the server.
  • the user corresponding to the terminal triggers the content acquisition instruction through the interface of the terminal, and the terminal generates a content acquisition request in response to the content acquisition instruction, and sends the content acquisition request to the server.
  • step 602 after receiving the content acquisition request, the server constructs a feature combination mode set including multiple feature combination modes.
  • the feature combination mode is a combination mode for the original features of the content to be recommended.
  • multiple original features can be obtained, that is, the related original features of the user or the content to be recommended, such as "user identification, content tag", etc., and the obtained multiple original features can be combined in pairs to Form multiple feature combination methods, so as to construct a feature combination method set based on the acquired multiple original features and multiple feature combination methods.
  • more than two original features can also be selected and combined to form multiple feature combinations.
  • step 603 the server obtains the feature value set corresponding to each feature combination mode.
  • the feature value set includes the feature value of each combined feature, and the combined feature can be obtained by feature combination of the original feature based on the corresponding feature combination mode.
  • the feature value of the combined feature may be extracted based on historical log data, or may be part of historical log data obtained by sampling historical log data.
  • step 604 the server determines a positive sample statistic corresponding to each feature value of the combined feature, and a negative sample statistic corresponding to each feature value of the combined feature.
  • the positive sample represents the sample data corresponding to the clicked content among the multiple content to be recommended displayed during the display process of the content to be recommended;
  • the negative sample represents the amount of content displayed during the display process of the content to be recommended Sample data corresponding to unclicked content among the content to be recommended.
  • step 605 the server obtains the weight value of the corresponding combined feature based on the positive sample statistic and the negative sample statistic corresponding to each feature value.
  • the server may also use a weight calculation model to determine the weight value of each combination feature.
  • the feature value of each combination feature is input into the weight calculation model to obtain the corresponding output of the weight calculation model. The weight value of each combination feature.
  • step 606 the server constructs a weight value set of the corresponding feature combination mode based on the weight value of each combination feature.
  • the weight value set of the feature combination mode includes the weight value corresponding to each combined feature, and the combined feature can be obtained by combining the original features based on the corresponding feature combination mode.
  • step 607 the server weights the weight values of all the combined features in the weight value set to obtain a score corresponding to the feature combination mode.
  • the score of the feature combination mode is used to characterize the likelihood that the content to be recommended will be clicked by the user when content is recommended based on the combined features obtained by the corresponding feature combination mode.
  • step 608 the server compares the score of each feature combination mode with the target score to obtain a comparison result corresponding to each feature combination mode.
  • the comparison result between the score of each feature combination mode and the target score can be calculated through accuracy indicators such as AUC and Logloss.
  • step 609 the server determines the validity of the corresponding feature combination based on the comparison result.
  • the effectiveness of the feature combination mode is used to characterize the accuracy of content recommendation based on the features obtained by the corresponding feature combination mode combination.
  • step 610 based on the ranking of the effectiveness of each feature combination mode, the server selects the feature combination mode of the target number from the feature combination mode set as the target feature combination mode.
  • the server determines the effectiveness of each feature combination mode, it can also filter the feature combination mode set to obtain a target number of feature combination modes from the feature combination mode set as the first candidate feature combination mode based on the ranking of the effectiveness of each feature combination mode; Based on the first candidate feature combination mode and the original feature, multiple second candidate feature combination modes are generated; from the multiple second candidate feature combination modes, the feature combination mode that meets the screening conditions is selected as the target feature combination mode.
  • step 611 the server performs feature combination on the original features based on the target feature combination mode to obtain the target combination feature.
  • step 612 the server determines the target recommended content based on the target combination feature, and returns the target recommended content to the terminal.
  • the server can determine the target recommendation content based on the target combination feature in the following way: use the target combination feature as the input feature and input it into the click-through rate estimation model to obtain the click-through rate of the content to be recommended by the target user; The content of the target quantity is selected as the target recommendation content.
  • step 613 the terminal presents the target recommended content.
  • the feature value set corresponding to each feature combination mode is obtained, where the feature value set includes the feature value of each combined feature.
  • the weight value of the corresponding combination feature is determined, and the weight value set of the corresponding feature combination method is constructed based on the weight value of each combination feature.
  • One use scenario is as follows: an APP builds a recommendation system from scratch, and uses the input features (original features of the content to be recommended) as the user ID, content ID, content tag, and city where the user is located to predict the click-through rate of the target user. Generally, it is difficult for each original feature to predict the click-through rate separately, but the combined feature obtained by combining the original features may provide effective information for the click-through rate.
  • Possible feature combinations include “posting media-user ID”, “posting media-content ID”, “posting media-content tag”, “posting media-user city”, “posting media-user ID-content tag”, “Posting media-user's city-content label” etc., can determine the scores of the above-mentioned feature combination methods through the feature effectiveness evaluation method proposed in the embodiments of this application, and compare the scores based on the determined feature combination methods with the target scores As a result, the most effective combination of several target features (for example, "posting media-user's city”) is selected to estimate the click-through rate of the news APP.
  • the recommendation system of an APP has a feature automatic screening tool, which can be selected from some original features of the content to be recommended, such as user ID, content ID, content tag, user city, etc.
  • Feature combination methods For these feature combination methods, the feature effectiveness evaluation method proposed in the embodiments of this application can further evaluate the effectiveness of each feature combination method, and screen out the most effective target feature combination method for the number of targets to use Estimated click-through rate for the APP.
  • the feature validity evaluation device 255 provided in the embodiment of the present application.
  • the feature validity evaluation device 255 may include:
  • the first construction unit 2551 is configured to construct a feature combination mode set including multiple feature combination modes, where the feature combination mode is a combination mode for the original feature of the content to be recommended; the obtaining unit 2552 is configured to obtain each of the feature combinations The feature value set corresponding to the mode, the feature value set includes the feature value of each combined feature, and the combined feature is obtained by combining the original features based on the corresponding feature combination mode; the first determining unit 2553 is configured to be based on The feature value of each of the combined features is determined, and the weight value of the corresponding combined feature is determined; the second construction unit 2554 is configured to construct the weight value set of the corresponding feature combination method based on the weight value of each of the combined features; second The determining unit 2555 is configured to determine the effectiveness of each feature combination method based on the weight value set of each feature combination method, and the effectiveness is used to characterize the content of the feature combination obtained based on the corresponding feature combination method. Recommended accuracy.
  • the first determining unit includes:
  • the first determining subunit is configured to determine a positive sample statistic corresponding to each feature value of the combined feature, and a negative sample statistic corresponding to each feature value of the combined feature;
  • the second determining subunit is configured to obtain the corresponding weight value of the combined feature based on the positive sample statistic and the negative sample statistic corresponding to the value of each feature.
  • the first determining subunit determines the positive sample statistic corresponding to each feature value of the combined feature, and the negative sample statistic corresponding to each feature value of the combined feature, you can use Realize in the following way:
  • determining the positive sample statistic that includes the feature value is the positive sample statistic corresponding to the feature value
  • the negative sample statistic that includes the characteristic value is determined to be the negative sample statistic corresponding to the characteristic value.
  • the first determining subunit determines the positive sample statistic corresponding to each feature value of the combined feature, and the negative sample statistic corresponding to each feature value of the combined feature, it may be Implemented in the following way:
  • the negative sample statistic corresponding to the feature value is determined.
  • the first determining subunit determines the positive sample statistic corresponding to each feature value of the combined feature, and the negative sample statistic corresponding to each feature value of the combined feature, it may be Implemented in the following way:
  • the first prediction accuracy represents the accuracy of recommendation based on the positive sample of the feature value
  • the second prediction accuracy represents the accuracy of recommendation based on the negative sample of the feature value
  • the first determining subunit determines the positive sample statistic corresponding to each feature value of the combined feature, and the negative sample statistic corresponding to each feature value of the combined feature, it may be Implemented in the following way:
  • the first sampling sample represents a first proportion of samples drawn from a positive sample including the characteristic value
  • the second sampling sample represents a sample drawn from a negative sample including the characteristic value Sample of the second proportion
  • a negative sample statistic corresponding to the characteristic value is determined.
  • the first determining subunit determines the positive sample statistic corresponding to each feature value of the combined feature, and the negative sample statistic corresponding to each feature value of the combined feature, it may be Implemented in the following way:
  • the second determining subunit obtains the corresponding weight value of the combined feature based on the positive sample statistic and the negative sample statistic corresponding to the value of each feature, it can be implemented in the following manner:
  • w F, j is the weight value of the combined feature
  • F is the feature combination mode
  • j is the feature value of each combined feature corresponding to the feature combination mode F
  • the second determining subunit obtains the corresponding weight value of the combined feature based on the positive sample statistic and the negative sample statistic corresponding to the value of each feature, it can be implemented in the following manner :
  • the weight value of the combined feature is the ratio of the difference to the first sum, and the difference is The difference with ⁇ 1 , the first sum value is the sum of N′ and ⁇ 2 ;
  • the weight value of the combined feature is the ratio of the second sum value to the first sum value, and the second sum value is And the sum of ⁇ 1;
  • the positive sample represents the sample data corresponding to the clicked content among the multiple content to be recommended displayed during the display process of the content to be recommended;
  • the negative sample represents the sample data displayed during the display process of the content to be recommended Sample data corresponding to the unclicked content among the multiple to-be-recommended content.
  • the first determining unit determines the weight value of the corresponding combined feature based on the feature value of each of the combined features
  • the feature value of each of the combined features is input into the weight calculation model, and the weight value of each of the combined features output by the weight calculation model is obtained.
  • the second determining unit separately determines the effectiveness of each feature combination manner based on the weight value set of each feature combination manner, it can be implemented in the following manner:
  • the second determining unit separately determines the effectiveness of each feature combination manner based on the score of each feature combination manner, it can be implemented in the following manner:
  • the feature validity evaluation device further includes:
  • the first screening unit is configured to sort based on the effectiveness of each of the feature combination modes, and obtain a target number of feature combination modes from the feature combination mode set as the target feature combination modes;
  • the feature combination unit is configured to perform feature combination on the original feature based on the target feature combination mode to obtain a target combination feature
  • the content recommendation unit is configured to perform content recommendation based on the target combination feature.
  • the feature validity evaluation device further includes:
  • the second screening unit is configured to sort based on the effectiveness of each of the feature combination modes, and obtain a target number of feature combination modes from the feature combination mode set as the first candidate feature combination modes;
  • a generating unit configured to generate a plurality of second candidate feature combination modes based on the first candidate feature combination mode and the original feature
  • the third screening unit is configured to select a feature combination mode that meets the screening conditions from the plurality of second candidate feature combination modes as the target feature combination mode;
  • the feature combination unit is configured to perform feature combination on the original feature based on the target feature combination mode to obtain a target combination feature
  • the content recommendation unit is configured to perform content recommendation based on the target combination feature.
  • the content recommendation unit performs content recommendation based on the target combination feature, it can be implemented in the following manner:
  • An embodiment of the present application also provides an electronic device, which includes:
  • Memory configured to store executable instructions
  • the processor is configured to execute the executable instructions stored in the memory to implement the above-mentioned feature validity evaluation method provided in the embodiment of the present application.
  • the embodiment of the present application also provides a storage medium that stores executable instructions.
  • executable instructions When executed by a processor, they are used to implement the above-mentioned feature validity evaluation method provided by the embodiments of the present application.
  • the storage medium may specifically be a computer-readable storage medium, such as Ferromagnetic Random Access Memory (FRAM), ROM, PROM, erasable programmable read-only memory (EPROM, Erasable Programmable Read-Only Memory, Electrically Erasable Programmable Read-Only Memory (EEPROM, Electrically Erasable Programmable Read-Only Memory), Flash Memory, Magnetic Surface Memory, Optical Disk or CD-ROM , Compact Disc Read-Only Memory), etc.; it can also be a variety of devices including one or any combination of the above-mentioned memories.
  • FRAM Ferromagnetic Random Access Memory
  • ROM read-only memory
  • PROM erasable programmable read-only memory
  • EPROM Erasable Programmable Read-Only Memory
  • EEPROM Electrically Erasable Programmable Read-Only Memory
  • Flash Memory Magnetic Surface Memory, Optical Disk or CD-ROM , Compact Disc Read-Only Memory
  • the executable instructions may be in the form of programs, software, software modules, scripts or codes, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and their It can be deployed in any form, including being deployed as an independent program or as a module, component, subroutine or other unit suitable for use in a computing environment.
  • executable instructions may but do not necessarily correspond to files in the file system, and may be stored as part of files that store other programs or data, for example, in a HyperText Markup Language (HTML, HyperText Markup Language) document
  • HTML HyperText Markup Language
  • One or more scripts in are stored in a single file dedicated to the program in question, or in multiple coordinated files (for example, a file storing one or more modules, subroutines, or code parts).
  • executable instructions can be deployed to be executed on one computing device, or on multiple computing devices located in one location, or on multiple computing devices that are distributed in multiple locations and interconnected by a communication network Executed on.
  • This combination feature can be a target feature combination method determined based on the effectiveness of each feature combination method.
  • the target combination feature obtained by the combination is based on the target combination feature Content recommendation can improve the recommendation effect of the recommendation system.
  • the electronic device constructs a feature combination mode set including multiple feature combination modes, where the feature combination mode is a combination mode for the original feature of the content to be recommended; obtains the feature value set corresponding to each feature combination mode, The feature value set includes the feature value of each combined feature, and the combined feature is obtained by feature combination of the original feature based on the corresponding feature combination mode; based on the feature value of each combined feature, the corresponding combined feature is determined Based on the weight value of each combination feature, construct the weight value set of the corresponding feature combination mode; based on the weight value set of each feature combination mode, determine the effectiveness of each feature combination mode, so The effectiveness is used to characterize the accuracy of content recommendation based on the features obtained by combining the corresponding feature combination methods; in this way, based on the weight value set of each feature combination method, the effectiveness of each feature combination method can be determined without cost Time-based experiments and model training can evaluate the effectiveness of a large number of feature combination methods in a short time, improve the efficiency of evaluating the effectiveness of a large

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computational Linguistics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

一种特征有效性评估方法、装置、电子设备及存储介质,涉及人工智能技术;方法包括:构建包括多个特征组合方式的特征组合方式集合,特征组合方式为针对待推荐内容的原始特征的组合方式;获取各特征组合方式对应的特征值集合,特征值集合包括各组合特征的特征取值,组合特征基于相应的特征组合方式对原始特征进行特征组合得到;基于各组合特征的特征取值,确定相应的组合特征的权重值;基于各组合特征的权重值,构建相应的特征组合方式的权重值集合;基于各特征组合方式的权重值集合,分别确定各特征组合方式的有效性。

Description

特征有效性评估方法、装置、电子设备及存储介质
相关申请的交叉引用
本申请基于申请号为202010007053.1、申请日为2020年01月03日的中国专利申请提出,并要求该中国专利申请的优先权,该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本申请涉及人工智能(AI,Artificial Intelligence)领域中的机器学习技术,尤其涉及一种特征有效性评估方法、装置、电子设备及存储介质。
背景技术
人工智能是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能,感知环境、获取知识并使用知识获得最佳结果的理论、方法和技术及应用系统。
内容推荐是人工智能的一个应用方向,具体研究向客户端推荐感兴趣的内容,从而实现内容推广的目的。客户端对内容的点击率,是决定向客户端推荐内容的重要指标。不同特征的组合可能对点击率预估提供有效的信息,因此,选择合适的特征组合方式对特征进行组合在内容推荐中具有重要意义,而评估特征组合方式的有效性是关键环节。
在相关技术中,主要采用实验方法或模型训练方法评估特征组合方式的有效性,然而,相关技术的这些评估方法均无法对大量的特征组合方式的有效性进行快速评估。
发明内容
本申请实施例提供一种特征有效性评估方法、装置、电子设备及存储介质,能够提高对大量的特征组合方式的有效性进行评估的效率。
本申请实施例提供一种特征有效性评估方法,所述方法包括:
构建包括多个特征组合方式的特征组合方式集合,所述特征组合方式为针对待推荐内容的原始特征的组合方式;
获取各所述特征组合方式对应的特征值集合,所述特征值集合包括各组合特征的特征取值,所述组合特征基于相应的特征组合方式对所述原始特征进行特征组合得到;
基于各所述组合特征的特征取值,确定相应的组合特征的权重值;
基于各所述组合特征的权重值,构建相应的特征组合方式的权重值集合;
基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,所述有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
本发明本申请实施例还提供一种特征有效性评估装置,所述装置包括:
第一构建单元,配置为构建包括多个特征组合方式的特征组合方式集合,所述特征组合方式为针对待推荐内容的原始特征的组合方式;
获取单元,配置为获取各所述特征组合方式对应的特征值集合,所述特征值集合包括各组合特征的特征取值,所述组合特征基于相应的特征组合方式对所述原始特征进行特征组合得到;
第一确定单元,配置为基于各所述组合特征的特征取值,确定相应的组合特征的权重值;
第二构建单元,配置为基于各所述组合特征的权重值,构建相应的特征组合方式的权重值集合;
第二确定单元,配置为基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,所述有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
本申请实施例还提供一种电子设备,所述电子设备包括:
存储器,配置为存储可执行指令;
处理器,配置为执行所述存储器中存储的可执行指令时,实现本申请实施例提供的特征有效性评估方法。
本申请实施例还提供一种存储介质,存储有可执行指令,所述可执行指令被处理器执行时,用于实现本申请实施例提供的特征有效性评估方法。
应用本申请实施例提供的特征有效性评估方法、装置、电子设备及存储介质,对于给定的包括多个特征组合方式的特征组合方式集合,通过获取各特征组合方式对应的特征值集合,这里特征值集合包括各组合特征的特征取值,基于各组合特征的特征取值,确定相应的组合特征的权重值,从而基于各组合特征的权重值,构建相应的特征组合方式的权重值集合,如此,基于各特征组合方式的权重值集合,就可以确定各特征组合方式的有效性,无需进行耗时的实验和模型训练,能够实现在短时间内评估大量的特征组合方式的有效性,提高对大量的特征组合方式的有效性进行评估的效率,进而提升推荐系统的推荐效果。
附图说明
图1为本申请实施例提供的特征有效性评估系统10的架构示意图;
图2为本申请实施例提供的电子设备20的硬件结构示意图;
图3为本申请实施例提供的特征有效性评估装置255的组成结构示意图;
图4为本申请实施例提供的推荐系统的结构示意图;
图5为本申请实施例提供的特征有效性评估方法的流程示意图;
图6为本申请实施例提供的特征有效性评估方法的流程示意图。
具体实施方式
为了使本申请实施例的目的、技术方案和优点更加清楚,下面将结合附图对本申请作进一步地详细描述,所描述的实施例不应视为对本申请的限制,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例,都属于本申请保护的范围。
在以下的描述中,涉及到“一些实施例”,其描述了所有可能实施例的子集,但是可以理解,“一些实施例”可以是所有可能实施例的相同子集或不同子集,并且可以在不冲突的情况下相互结合。
在以下的描述中,所涉及的术语“第一”、“第二”等仅仅是用于区别类似的对象,不代表针对对象的特定的顺序或先后次序,可以理解地,“第一”、“第二”等在允许的 情况下可以互换特定的顺序或先后次序,以使这里描述的本申请实施例能够以除了在这里图示或描述的以外的顺序实施。
除非另有定义,本申请实施例所使用的所有的技术和科学术语与属于本申请实施例的技术领域的技术人员通常理解的含义相同。本申请中所使用的术语只是为了描述具体的实施例的目的,不是旨在限制本申请。
在对本申请实施例进行详细说明之前,先对本申请实施例中涉及的名词和术语进行说明,本申请实施例中涉及的名词和术语适用于如下的解释。
1)推荐系统,内容推荐是推荐系统的一个重要应用,其中,内容推荐是一种根据用户的需求,在大量的待推荐内容中选取用户感兴趣的内容,并将所选取的内容推荐给用户的过程。其中,待推荐内容可以是媒体文件,也可以是物品等,这里的物品可以是某个产品或者商品。
2)媒体文件,在互联网中可获取的各种形式(比如视频、音频、图文等媒体形式)的媒体,例如客户端中呈现的视频文件、包括图文形式的文章(比如网络文章或新媒体文章)、新闻、广告等。
3)点击率(CTR,Click-Through Rate),是指客户端中的某一待推荐内容被用户点击的次数与被显示的次数的比值,也就是说,点击率是指该待推荐内容被用户点击的概率。在推荐系统中,用户的点击率是一个非常重要的指标,尤其是在在线推荐系统中,对于用户对某个待推荐内容的点击率进行预测,从而决定是否向用户推荐该待推荐内容,是提高用户体验的重要方式。
4)点击率预估,是对点击率进行预测的方法,也可称为点击率预估算法,该类算法是根据用户和待推荐内容的相关信息,以及其它的背景数据(比如点击行为发生的时间、用户所处的网络环境、使用的终端设备如是使用手机还是电脑)等,预测用户在某一背景下点击某个待推荐内容的概率的算法。
其中,点击率预估算法通常可基于一定的数学模型(如点击率预估模型)实现,该模型的输入特征可以是原始特征,也可以是组合特征,以对点击率预估提供有效的信息。
5)原始特征,是指从样本中直接得到的特征,原始特征可以包括用户特征、内容特征、背景环境特征。其中,用户特征例如可以是用户标识(用户ID)、用户的兴趣爱好、用户性别、用户年龄等;内容特征例如可以是内容标识(内容ID)、内容分类、内容标签等;背景环境特征例如可以是用户访问时间、用户访问时所处的网络环境(比如WiFi或者蜂窝移动网络)等。
6)组合特征,是指通过对一个或多个原始特征进行特征组合所得到的特征,其取值应遍历原始特征的所有可能的取值组合。
举例来说,如果有两个原始特征,分别为特征1和特征2,特征1的取值范围为A和B,特征2的取值范围为C、D和E,那么,通过对特征1和特征2进行特征组合,可以得到多个组合特征,组合特征的取值范围为AC、AD、AE、BC、BD和BE;如果两个原始特征为“特征1:A,特征2:C”,那么对应得到的组合特征为AC。以上是通过对两个原始特征(特征1和特征2)进行特征组合得到组合特征的示例性说明,基于上述类似的方法,可以实现对多个原始特征(比如两个以上的原始特征)进行特征组合,以得到相应的多个组合特征。
7)特征组合方式,是指将原始特征进行特征组合得到组合特征的方式,而组合特征,是指在该特征组合方式下所对应的具体特征值。例如,“用户ID-内容标签”为一种特征组合方式,而“张三-篮球”则是该特征组合方式下所对应的一个组合特征的特征取值。
8)特征工程,是指一种选择合适的原始特征和特征组合方式的过程。在推荐系统 中,由于特征工程的效率直接影响到推荐系统的推荐效果,因此,特征工程是推荐系统中极其重要的处理环节。
9)响应于,用于表示所执行的操作所依赖的条件或者状态,当满足所依赖的条件或状态时,所执行的一个或多个操作可以是实时的,也可以具有设定的延迟;在没有特别说明的情况下,所执行的多个操作不存在执行先后顺序的限制。
本申请实施例提供的特征有效性评估方法可应用于推荐系统中,涉及人工智能领域,人工智能是计算机科学的一个综合技术,它企图了解智能的实质,并生产出一种新的能以人类智能相似的方式做出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法,使机器具有感知、推理与决策的功能。
机器学习技术是人工智能的重要应用分支,涉及概率论、统计学、逼近论、凸分析、算法复杂度理论等多门学科。专门研究计算机怎样模拟或实现人类的学习行为,以获取新的知识或技能,重新组织已有的知识结构使之不断改善自身的性能。机器学习技术是人工智能的核心,是使计算机具有智能的根本途径,机器学习技术的应用遍及人工智能的各个领域。本申请实施例通过机器学习技术实现对大量的特征组合方式的有效性进行评估,而无需进行耗时的实验和模型训练,可以在短时间内评估大量的特征组合方式的有效性,能够提高对特征组合方式的有效性进行评估的效率,进而提升推荐系统的推荐效果。
在特征工程中,通常面临的问题是如何评估特征组合方式的有效性,以便根据评估结果确定基于该特征组合方式进行组合的组合特征,是否可以新增至点击率预估模型中,以对点击率预估提供有效的信息。为解决该技术问题,相关技术提供了评估特征组合方式的有效性的方案。
在相关技术的一些方案中,主要依赖大量实验的实验结果评估特征组合方式的有效性,通常需要从业人员根据经验筛选出基于该特征组合方式组合得到的一些备选的特征,这就要求从业人员对产品具有较高的熟悉度,以及对数据具有较好的敏感度,才能筛选合适的特征,但由于不同的从业人员的经验并不相同,因此,人工特征选择的方法需要大量的试错环节,其质量和速度均难以得到有效的保证,且实验非常耗时,当存在大量的特征组合方式时,通过实验方法评估每个特征组合方式的有效性的速度会很慢,消耗计算资源多,在有限时间内能够评估的特征组合方式的数量是极其有限的,导致对大量的特征组合方式的有效性进行评估的效率大大降低。
在相关技术的另一些方案中,可通过全量训练方法实现特征组合方式的有效性的评估,在实际实施时,首先,基于待评估的特征组合方式对原始特征进行组合得到不同的特征,其次,全量训练方法在加入每个特征之后,通过一次全量的模型训练,得到含有该特征的模型,并评估该模型的预估效果,作为评估特征组合方式的有效性的依据。其中,模型训练通常采用随机梯度下降法(SGD,Stochastic Gradient Descent)或其变形方法。然而,全量训练方法的主要问题在于其计算速度较慢,通常需要大量的时间来训练,因此,制约了从业人员尝试更多特征的能力,进而影响了特征工程的效率。
下面以全量训练方法为梯度提升决策树(Gradient Boosting Decision Tree)方法(简称决策树方法)为例进行说明。决策树方法根据多个判断条件将原始特征映射到一个新的特征,例如,决策树方法将所有样本分为10组:满足性别男、年龄在20-25岁之间样本,作为第1组;满足性别女、年龄在20-25岁之间的样本,作为第2组;等等,在判断某个候选特征组合方式的有效性时,决策树方法采用信息增益、信息增益比、基尼系数等参数进行评估。
然而,决策树方法主要适用于连续特征和只有少量取值的离散特征(例如性别)的 特征工程。这在早期的推荐系统中较为有效,但目前的推荐系统具有大量被称为ID类特征的离散特征,例如用户ID、内容ID等,这些特征具有大量可能的取值,而决策树方法对特征的取值数量非常敏感,因此,在判断有效性时面临着模型复杂度极高,以至于无法应用于在线推荐系统,还会面临着判断不准确等问题,使得决策树方法难以应用于工业化的推荐系统中。
由此可见,相关技术的方案均无法对大量的特征组合方式的有效性进行快速评估。为至少解决相关技术的上述技术问题,本申请实施例提供了一种特征有效性评估方法,通过获取各特征组合方式对应的特征值集合,这里特征值集合包括各组合特征的特征取值,基于各组合特征的特征取值,确定相应的组合特征的权重值,从而基于各组合特征的权重值,构建相应的特征组合方式的权重值集合,如此,基于各特征组合方式的权重值集合,就可以确定各特征组合方式的有效性,提高了对大量的特征组合方式的有效性进行评估的效率;还可以基于特征组合方式的有效性筛选出合适的目标特征组合方式,基于目标特征组合方式得到目标组合特征,以用于推荐系统进行内容推荐,显著提高特征工程的效率,实现更准确的点击率预估,为推荐系统提供强有力的技术支持,提升推荐系统的推荐效果。
下面说明实施本申请实施例的特征有效性评估方法的电子设备的示例性应用,本申请实施例提供的电子设备可以实施为笔记本电脑,平板电脑,台式计算机,机顶盒,移动设备(例如移动电话、便携式音乐播放器、个人数字助理、专用消息设备、便携式游戏设备)等各种类型的终端设备,也可以实施为服务器,该服务器可以包括但不限于任何可以进行计算的硬件设备,如可以是独立的物理服务器,也可以是多个物理服务器构成的服务器集群或者分布式系统,还可以是提供云服务、云数据库、云计算、云函数、云存储、网络服务、云通信、中间件服务、域名服务、安全服务、以及大数据和人工智能平台等基础云计算服务的云服务器;当然,本申请实施例的特征有效性评估方法也可以为终端设备和服务器协同实施。这里,服务器可以为云端服务器,但本申请实施例中并不仅限于云端服务器。
下面以终端设备和服务器协同实施为例,参考附图对本申请实施例的特征有效性评估系统的示例性应用进行说明。参见图1,图1为本申请实施例提供的特征有效性评估系统10的一个可选的架构示意图,为实现支撑的一个示例性应用,终端100(示例性示出了终端100-1和终端100-2)通过网络200连接服务器300,其中,网络200可以是广域网或者局域网,又或者是二者的组合,使用无线链路实现数据传输。
在一些实施例中,终端100(如终端100-1),配置为基于内容获取指令触发向服务器300发送内容获取请求,以请求获取自身感兴趣的内容。
服务器300,配置为响应于终端100发送的内容获取请求,构建包括多个特征组合方式的特征组合方式集合;获取各特征组合方式对应的特征值集合;基于各组合特征的特征取值,确定相应的组合特征的权重值;基于各组合特征的权重值,构建相应的特征组合方式的权重值集合;基于各特征组合方式的权重值集合,分别确定各特征组合方式的有效性。
接下来,服务器300还配置为基于各特征组合方式的有效性确定目标特征组合方式,以基于目标特征组合方式确定目标推荐内容,在实际实施时,可以对各特征组合方式的有效性进行排序,得到相应的排序结果,进而基于排序结果从特征组合方式集合中筛选目标数量的特征组合方式,将筛选得到的特征组合方式作为目标特征组合方式,基于目标特征组合方式对待推荐内容的原始特征进行特征组合,得到目标组合特征,以基于目标组合特征确定目标用户对待推荐内容的点击率,进而基于点击率确定目标推荐内容, 并将目标推荐内容推送至终端100。
这里,在实际应用中,服务器300既可以为单独配置的支持各种业务的一个服务器,亦可以配置为一个服务器集群。
终端100接收到目标推荐内容后,可以在图形界面110(例如终端100-1的图形界面110-1和终端100-2的图形界面110-2)中对上述目标推荐内容进行呈现。
接下来继续对实施本申请实施例的特征有效性评估方法的电子设备的硬件结构进行说明。电子设备可以实施为终端设备,也可以实施为服务器,还可以为上述图1示出的终端设备和服务器的协同实施。
参见图2,图2为本申请实施例提供的电子设备20的一个可选的硬件结构示意图,可以理解,图2仅仅示出了电子设备的示例性结构而非全部结构,根据需要可以实施图2示出的部分结构或全部结构。本申请实施例提供的电子设备20可以包括:至少一个处理器210、存储器250、至少一个网络接口220和用户接口230。电子设备20中的各个组件通过总线系统240耦合在一起。可以理解,总线系统240配置为实现这些组件之间的连接通信。总线系统240除包括数据总线之外,还包括电源总线、控制总线和状态信号总线。但是为了清楚说明起见,在图2中将各种总线都标为总线系统240。
处理器210可以是一种集成电路芯片,具有信号的处理能力,例如通用处理器、数字信号处理器(DSP,Digital Signal Processor),或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等,其中,通用处理器可以是微处理器或者任何常规的处理器等。
用户接口230包括使得能够呈现媒体内容的一个或多个输出装置231,包括一个或多个扬声器和/或一个或多个视觉显示屏。用户接口230还包括一个或多个输入装置232,包括有助于用户输入的用户接口部件,比如键盘、鼠标、麦克风、触屏显示屏、摄像头、其他输入按钮和控件。
存储器250可以是可移除的,不可移除的或其组合。示例性的硬件设备包括固态存储器,硬盘驱动器,光盘驱动器等。存储器250可选地包括在物理位置上远离处理器210的一个或多个存储设备。
存储器250包括易失性存储器或非易失性存储器,也可包括易失性和非易失性存储器两者。非易失性存储器可以是只读存储器(ROM,Read Only Memory),易失性存储器可以是随机存取存储器(RAM,Random Access Memory)。本申请实施例描述的存储器250旨在包括任意适合类型的存储器。
在一些实施例中,存储器250能够存储数据以支持各种操作,这些数据的示例包括程序、模块和数据结构或者其子集或超集,下面示例性说明。
操作系统251,包括用于处理各种基本系统服务和执行硬件相关任务的系统程序,例如框架层、核心库层、驱动层等,用于实现各种基础业务以及处理基于硬件的任务;
网络通信模块252,配置为经由一个或多个(有线或无线)网络接口220到达其他计算设备,示例性的网络接口220包括:蓝牙、无线相容性认证(WiFi)、和通用串行总线(USB,Universal Serial Bus)等;
呈现模块253,配置为经由一个或多个与用户接口230相关联的输出装置231(例如,显示屏、扬声器等)使得能够呈现信息(例如,用于操作外围设备和显示内容和信息的用户接口);
输入处理模块254,配置为对一个或多个来自一个或多个输入装置232之一的一个或多个用户输入或互动进行检测以及翻译所检测的输入或互动。
在一些实施例中,本申请实施例提供的特征有效性评估装置可以采用软件方式实 现,图2示出了存储在存储器250中的特征有效性评估装置255,其可以是程序和插件等形式的软件,包括一系列的软件模块。参见图3,图3为本申请实施例提供的特征有效性评估装置255的一个可选的组成结构示意图,例如,特征有效性评估装置255可以包括第一构建单元2551、获取单元2552、第一确定单元2553、第二构建单元2554和第二确定单元2555,这些单元的功能是逻辑上的,因此,根据各软件模块所实现的功能可以进行任意的组合或拆分。这里,需要说明的是,对于图3所示的本申请实施例提供的特征有效性评估装置255中的各个单元的功能,将在下文进行说明。
在另一些实施例中,本申请实施例提供的特征有效性评估装置255可以采用硬件方式实现,作为示例,本申请实施例提供的特征有效性评估装置255可以是采用硬件译码处理器形式的处理器,其被编程以执行本申请实施例提供的特征有效性评估方法,例如,硬件译码处理器形式的处理器可以采用一个或多个应用专用集成电路(ASIC,Application Specific Integrated Circuit)、DSP、可编程逻辑器件(PLD,Programmable Logic Device)、复杂可编程逻辑器件(CPLD,Complex Programmable Logic Device)、现场可编程门阵列(FPGA,Field-Programmable Gate Array)或其他电子元件。
结合上述对本申请实施例的特征有效性评估系统及电子设备的说明,在对本申请实施例提供的特征有效性评估方法进行说明之前,先对特征有效性评估方法应用的推荐系统的结构进行说明。
参见图4,图4为本申请实施例提供的推荐系统的结构示意图,推荐系统可实施为服务器,服务器接收终端发送的用户请求,根据用户请求(用户请求中可包括用户ID、当前时刻等信息),从内容数据库中提取出待推荐内容,将用户请求和待推荐内容共同传递至特征中心;特征中心将用户请求及待推荐内容的相关信息整理成结构化的形式(比如字符串列表,或者键值对等形式),从而得到待推荐内容的原始特征(比如用户ID、用户年龄、内容ID、内容标签等),并将原始特征传递至特征组合模块。
在本申请实施例中,在推荐系统中增加了特征选择模块,通过特征选择模块实现特征组合方式的筛选,以寻找有效的特征组合方式,比如有效的特征组合方式为“用户ID-内容标签”,而特征组合模块则可基于特征选择模块提供的有效的特征组合方式,对一个或多个原始特征进行特征组合,得到多个组合特征(比如基于该有效的特征组合方式“用户ID-内容标签”组合得到的一个组合特征为“张三-篮球”);接下来,将组合特征作为输入特征输入至点击率预估模块,以预估用户对待推荐内容的点击率;接下来,重排序模块根据得到的点击率,对待推荐内容进行排序,从待推荐内容中选取目标数量的内容作为目标推荐内容;再接下来,推荐系统将目标推荐内容返回给用户。
这里,在实际实施时,特征选择模块根据特征中心记录的离线日志数据,选择有效的特征组合方式(如用户ID-内容标签),并由此更新特征组合模块中的特征组合方式,使特征组合模块输出有效的组合特征,以基于该组合特征进行内容推荐模型的训练或线上点击率预估。
需要说明的是,在实现线上内容推荐的同时,推荐系统会将每次推荐过程所用到的原始特征、推荐结果、用户的点击情况存储成日志,这些日志可用于得到各个组合特征的每种取值(比如特征组合方式为“用户ID-内容标签”,该有效的特征组合方式所对应的某个组合特征的特征取值为“张三-篮球”)的权重值,以便快速估计出多个特征组合方式的有效性,以便基于多个特征组合方式的有效性确定目标特征组合方式,指导后续推荐操作。
下面对日志及点击率预估模型进行说明。日志系统提取一定时间范围内的用户点击数据(即样本),并将每条样本整理为如下格式:
Figure PCTCN2020123721-appb-000001
其中,i表示第i条 日志;y (i)表示用户是否点击的点击结果;
Figure PCTCN2020123721-appb-000002
是预估得到的该样本的点击率,是一个0至1之间的数值;x (i)为特征中心整理出的原始特征,可以记录为键-值对的形式,如“用户ID:张三;内容ID:内容A;用户的关注媒体:媒体A、媒体B;内容标签:篮球”等。需要说明的是,原始特征可以是单个值(如用户ID),也可以是多个值(如用户关注的媒体,通常不止一个)。
点击率预估模型通常根据输入特征(如用户特征、内容特征、背景环境特征等),通过一定的算法预测出用户在当前环境下点击内容的概率,即点击率,为0至1之间的实数。点击率预估模型通常被建模为如下形式:
Figure PCTCN2020123721-appb-000003
常见的点击率预估模型的种类很多,例如,逻辑斯蒂回归模型(LR,Logistic Regression)、深度神经网络模型(DNN,Deep Neural Network)等。需要说明的是,本申请实施例对点击率预估模型的类型没有任何限制。
下面将结合本申请实施例提供的特征有效性评估系统、电子设备的示例性应用以及推荐系统的说明,对本申请实施例提供的特征有效性评估方法的实现进行说明。
参见图5,图5为本申请实施例提供的特征有效性评估方法的流程示意图,在一些实施例中,该特征有效性评估方法可由终端实施,也可由服务器实施,当然也可由服务器及终端协同实施。下面以服务器实施为例,如通过图1中的服务器300实施,结合图5示出的步骤进行说明。
在步骤501中,服务器构建包括多个特征组合方式的特征组合方式集合。
在本申请实施例中,特征组合方式为针对待推荐内容的原始特征的组合方式。在实际应用中,当服务器采用上述推荐系统进行内容推荐时,需要提取待推荐内容的相关特征,比如用户特征和内容特征,以对用户特征和内容特征等进行组合得到相关组合特征以提高内容推荐的精确度。
在一些实施例中,服务器可通过如下方式构建特征组合方式集合:获取多个原始特征;确定在多个原始特征中将至少两个原始特征进行组合所得到的多个特征组合方式;基于多个原始特征及多个特征组合方式,构建特征组合方式集合。
这里,多个原始特征即为包括用户特征和内容特征的原始特征,比如,该多个原始特征可以是用户特征“用户标识”、“用户关注的媒体”,内容特征“发文媒体”、“内容标签”等。
具体来说,在构建特征组合方式集合时,首先要获取多个原始特征,然后多次从多个原始特征中选择至少两个特征并进行组合,以得到多个特征组合方式。示例性地,比如多个原始特征包括“用户标识、用户关注的媒体、发文媒体、内容标签”,任意选择两个原始特征进行组合,可以得到“用户标识-用户关注的媒体”、“用户标识-发文媒体”、“用户关注的媒体-内容标签”等多个特征组合方式。在实际实施时,除了可以任意选择两个原始特征之外,还可以任意选择三个原始特征进行组合,这里不作限定。在得到多个特征组合方式后,基于该多个特征组合方式与多个原始特征,构建特征组合方式集合。由于该多个原始特征可以被看作一种特殊的特征组合方式,那么可以将所得到多个特征组合方式、与多个原始特征均作为特征组合方式集合中的各特征组合方式,以此构建特征组合方式集合。
示例性的,可采用如下方式构建特征组合方式集合:首先,将获取的所有的原始特征进行组合得到D1,D 1={特征1,特征2,…,特征n};其次,将多个原始特征中的任意两个特征进行两两组合,构造多个特征组合方式,用D2表示:D 2={(特征1,特征1),…(特征1,特征n),…,(特征n,特征n)};最后,构造 集合D=D 1∪D 2,即为包括多个特征组合方式的特征组合方式集合。
需要说明的是,在构建对原始特征进行两两组合得到D 2时,可以根据先验知识,例如仅将用户特征与内容特征进行组合、将用户特征与背景环境特征进行组合等,从而遍历较少的特征,降低处理器的计算量。然而,在实际应用中,如果处理器的计算能力很强,则可以进一步构建两个以上的简单特征组合得到的特征作为候选特征,直至遍历所有的候选特征构造多个特征组合方式,进而形成特征组合方式集合。
在步骤502中,获取各特征组合方式对应的特征值集合,特征值集合包括各组合特征的特征取值。
在本申请实施例中,组合特征是基于相应的特征组合方式对原始特征进行特征组合得到的。这里,组合特征的特征取值可以是基于历史日志数据提取得到的,也可以是对历史日志数据进行采样得到的部分历史日志数据。
在实际应用中,组合特征有时会比单个原始特征更加重要,例如,内容标签作为一个独立的原始特征,并不是一个重要特征,因为每个标签对应的内容都可能有大量用户愿意观看,但是具体到某个用户而言,则可能出现该用户只对含有某几个关键词的内容感兴趣。因此,用户ID和内容标签组合得到的组合特征就是一个很重要的特征。例如,用户张三喜欢看篮球类的内容,那么,利用特征组合方式“用户ID-内容标签”得到的组合特征的值为“张三-篮球”的样本进行点击率预估的准确度,显然就比仅利用单个原始特征“篮球”的样本进行点击率预估的准确度要大大提高。
需要指出的是,一般来说,各个原始特征不会与自身进行组合,但如果一个原始特征包括多个值时,则该原始特征可以与自身进行组合,因此,单个原始特征(即未与其他的原始特征进行组合的特征)也可视为一种特殊的组合特征(即只有一个原始特征“组合”得到的特征)。例如,原始特征为“用户关注的媒体”,则可以构造组合特征“用户关注的媒体-用户关注的媒体”,如果在某个样本中,用户关注的媒体为媒体A、媒体B,则按前述方式构造该样本的组合特征的特征取值可以为“媒体A-媒体A”、“媒体A-媒体B”和“媒体B-媒体B”。
在步骤503中,基于各组合特征的特征取值,确定相应的组合特征的权重值。
在一些实施例中,服务器可通过如下方式确定各组合特征的权重值:将各组合特征的特征取值输入至权重计算模型中,得到权重计算模型输出的对应各组合特征的权重值。也就是说,本申请实施例可通过预先训练的权重计算模型,将各组合特征的特征取值分别输入到权重计算模型中,从而得到各组合特征对应的权重值。
在一些实施例中,服务器可通过如下方式训练得到权重计算模型:将标注有目标权重值的特征取值样本,输入至权重计算模型,得到权重计算模型输出的对应特征取值样本的权重值;基于输出的权重值及目标权重值,确定权重计算模型的损失函数的值;基于损失函数的值,更新权重计算模型的模型参数。
在实际应用中,可基于深度学习方法预先构建权重计算模型,比如用于计算权重值的卷积神经网络模型,包括输入层、隐藏层及输出层,用于计算各组合特征的权重值,以得到权重计算模型,在权重计算模型构建完成后,基于采集的特征取值样本对该权重计算模型进行训练,以得到优化的权重计算模型参数。在实际实施时,在进行权重计算模型训练时,输入的特征取值样本可以是仅针对某个特征组合方式的样本,还可以是所有特征组合方式的样本,通常为加快模型训练的速度,可以仅针对某个特征组合方式的样本进行训练。
在实际实施时,服务器可通过如下方式得到对应特征取值样本的权重值:首先获取大量的特征取值样本,比如可以是对某些待推荐内容的相关历史日志数据进行采样得到的,这些特征取值样本分别标注有对应的目标权重值;在进行训练之前,可以将采集的 大量样本按照一定的比例划分为训练集和测试集,将训练集中标注有目标权重值的特征取值样本输入到权重计算模型中,以得到权重计算模型输出对应特征取值样本的权重值。进一步地,模型训练的过程即是对模型中各参数的更新调整过程,将训练样本数据输入到权重计算模型的输入层,经过隐藏层,最后达到输出层并输出结果,由于权重计算模型的输出结果与实际结果可能有误差,则需要计算输出结果与实际值之间的误差,并将误差从输出层向隐藏层进行反向传播,直至传播到输入层,进而在反向传播的过程中,根据误差调整模型参数的值;整个训练过程不断迭代上述步骤,直至收敛,以减小模型输出的误差。
基于此,在本申请实施例中,为了降低权重计算模型输出的权重值与目标权重值之间可能存在的误差,引入了损失函数,基于权重计算模型输出的特征取值样本的权重值及目标权重值,确定损失函数的值;基于损失函数的值,运用反向传播算法逐层更新权重计算模型的参数,直至损失函数收敛,以实现对权重计算模型的参数的约束和调整,从而得到计算精度高的权重计算模型,以基于该权重计算模型确定各组合特征的权重值。
在实际应用中,在基于权重计算模型确定各组合特征的权重值时,可通过优化如下函数得到各组合特征的权重值:
Figure PCTCN2020123721-appb-000004
其中,
Figure PCTCN2020123721-appb-000005
表示基于各组合特征的权重值预测的点击率(0至1之间的实数),y (i)表示用户是否点击的点击结果(点击时为1,未点击时为0),loss为目标函数。在实际应用中,在将训练集中标注有目标权重值的特征取值样本输入到权重计算模型中,以得到权重计算模型输出对应特征取值样本的权重值后,在测试集数据的每个样本中,将每个样本在特征组合方式比如F下的所有组合特征的特征取值对应的权重值进行相加,得到预测的样本得分,即得到预测的特征组合方式的评分,示例性的,可通过如下方式预测对应特征组合方式的评分:
Figure PCTCN2020123721-appb-000006
其中,w F,j表示组合特征的权重值;F表示特征组合方式;j表示该特征组合方式F下所对应的各组合特征的特征取值;
Figure PCTCN2020123721-appb-000007
表示第i个样本在该特征组合方式F下的取值包含组合特征j;
Figure PCTCN2020123721-appb-000008
表示预测的特征组合方式的评分。接下来,可通过如下方式
Figure PCTCN2020123721-appb-000009
计算预测的点击率。
这里,常见的目标函数包括平方误差函数(square loss)、对数误差函数(Logloss)等。在点击率预估过程中,最常用的目标函数为对数误差函数,其形式如下:
Figure PCTCN2020123721-appb-000010
实际应用中,由于对于所有的正样本而言,均有y (i)=1;对于所有的负样本而言,均有y (i)=0,因此,上述公式(2)可变换成如下形式:
Figure PCTCN2020123721-appb-000011
上述公式(3)的目标函数的最优解一般无法用解析表达式表示,因此,其求解通常采用迭代的方法,如随机梯度下降法来获取各组合特征的权重值,导致模型训练消耗大量的时间,而在特征工程中,需要对大量的特征组合方式的有效性进行评估,采用模型训练的方法的计算成本高,速度慢。
为避免采用耗时的模型训练,本申请实施例还提供了采用统计的方式确定各组合特征的权重值,从而加快权重值的计算,保证针对大量的特征组合方式的有效性进行评估的效率得以提升。下面对采用统计的方式确定各组合特征的权重值进行说明。
在一些实施例中,服务器可通过如下方式确定各组合特征的权重值:确定组合特征 的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量;基于各特征取值对应的正样本统计量与负样本统计量,得到相应的组合特征的权重值。
在本申请实施例中,正样本,表征在待推荐内容的展示过程中,所展示的多个待推荐内容中被点击的内容对应的样本数据;负样本,表征在待推荐内容的展示过程中,所展示的多个待推荐内容中未被点击的内容对应的样本数据。
也就是说,本申请实施例根据待推荐内容的点击状态,将采集到的特征值集合对应的所有样本划分为正样本和负样本。示例性的,当特征组合方式为“用户标识-商品类型”时,该特征组合方式“用户标识-商品类型”下所对应的各组合特征的特征取值可包括“张三-化妆品、张三-零食、张三-服饰、张三-运动鞋”,基于每个特征取值进行内容推荐时,即向“张三”分别推荐“化妆品、零食、服饰和运动鞋”,如果“张三”分别点击了“化妆品、零食”,其他并未点击,那么,“张三-化妆品、张三-零食”即为正样本,“张三-服饰、张三-运动鞋”即为负样本。
在实际应用中,确定组合特征的权重值时,只需要考虑组合特征的特征取值样本的一些统计量即可。因此,公式(3)的目标函数可变换成如下形式:
Figure PCTCN2020123721-appb-000012
该目标函数的解析解即为组合特征的权重值。
在一些实施例中,考虑一种最简单的情况,即组合特征的特征取值仅有一个,此时,服务器可确定包括特征取值的正样本统计量,为特征取值对应的正样本统计量,以及确定包括特征取值的负样本统计量,为特征取值对应的负样本统计量。
在确定特征取值对应的正样本统计量,以及特征取值对应的负样本统计量之后,基于该正样本统计量及负样本统计量,确定相应的组合特征的权重值。在实际实施时,基于各特征取值对应的正样本统计量与负样本统计量,可采用如下公式计算相应的组合特征的权重值:
Figure PCTCN2020123721-appb-000013
其中,w F,j为组合特征的权重值;F为特征组合方式;j为该特征组合方式F下所对应的各组合特征的特征取值;
Figure PCTCN2020123721-appb-000014
为特征取值j对应的正样本统计量;
Figure PCTCN2020123721-appb-000015
为特征取值j对应的负样本统计量。
在实际应用中,针对这种组合特征的特征取值仅有一个的情况,可采用如下公式计算特征取值对应的正样本统计量,以及特征取值对应的负样本统计量:
Figure PCTCN2020123721-appb-000016
Figure PCTCN2020123721-appb-000017
结合公式(5)至(7)可知,如果含有特征取值j的正样本的数量越多,那么组合特征的权重值越大;反之,如果含有该特征取值j的负样本的数量越多,那么组合特征的权重值越小。在一些实施例中,如果含有特征取值j的正样本统计量大于含有特征取值j的负样本统计量,则w F,j>0,表示该组合特征对点击率预估的贡献是正向的,反之则是负向的。
根据上述方法,只需要统计出每个特征取值对应的
Figure PCTCN2020123721-appb-000018
即可根据公式(5)计算出相应的组合特征的权重值,较为简单便捷,这也是统计的方式相对于权重计算模型训练方法的优势。
上述实施例考虑的仅是最简单的情况,然而,在工程应用时,还可能面对其他不同的情况,因此,需要针对实际情况对上述实施例进行改进。下面针对不同的场景对统计量的改进计算方法进行说明。
在一些实施例中,针对某个特征组合方式,在一个样本中出现多个组合特征的特征取值时,服务器可通过如下方式确定组合特征的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量:确定包括特征取值的正样本中组合特征的特征取值统计量,以及包括特征取值的负样本中组合特征的特征取值统计量;基于包括特征取值的正样本中组合特征的特征取值统计量,确定特征取值对应的正样本统计量;基于包括特征取值的负样本中组合特征的特征取值统计量,确定特征取值对应的负样本统计量。
在实际应用中,针对某种特征组合方式(如“用户ID-内容标签”),在一个样本中可能出现多个组合特征的特征取值。例如,张三观看了一部有某球星参与的某综艺节目,那么该样本的内容标签包括“某球星”和“某综艺节目”。同时,该样本“用户ID-内容标签”所对应的组合特征的特征取值可包括“张三-某球星”和“张三-某综艺节目”,其点击行为也可能来源于这两个特征取值中的某一个。因此,该样本对“张三-某球星”的贡献,应该小于一个只包含“张三-某球星”特征取值的样本(例如,在另一个样本中,张三观看了该球星的一次个人采访)。
为了解决该问题,在本申请实施例中,可通过如下公式计算
Figure PCTCN2020123721-appb-000019
Figure PCTCN2020123721-appb-000020
Figure PCTCN2020123721-appb-000021
Figure PCTCN2020123721-appb-000022
然后,通过上述公式(5)计算相应的组合特征的权重值w F,j,在该实施例中,一个样本的特征取值越多,那么它对其中的某个特征取值j的权重的贡献就越小,因为该样本为正样本或负样本,有可能是因为其他特征取值决定的,而并非是由特征取值j决定的。在实际应用中,由于推荐系统中的特征大多数是稀疏的,因此可能出现某个特征在训练样本中只出现了极少次数的情况。例如,“内容ID”在训练样本的某个内容中只出现了一次,并且被点击,可视为正样本。在该情况下,按照公式(5)计算得到的相应的组合特征的权重值w F,j为正无穷(这意味着推荐系统确定该内容的点击率为100%),这是一个不合理的结果。
为了解决该问题,可以将公式(4)变换为如下形式:
Figure PCTCN2020123721-appb-000023
其中,λ 1和λ 2是非负实数,相较于公式(4),该公式(10)新增了两项,即λ 1|w F,j|和
Figure PCTCN2020123721-appb-000024
被称为正则项,当样本数量较少时,新增的这两项会起到主导作用,使得组合特征的权重值趋向于0;而当样本数量较多时,原有的目标函数项会起到主导作用,使得组合特征的权重值趋于公式(5)计算得到的权重值。
由于上述公式(10)没有解析解,因此,在一些实施例中,针对取值过少的样本,服务器可通过如下方式基于各特征取值对应的正样本统计量与负样本统计量,得到相应的组合特征的权重值:当
Figure PCTCN2020123721-appb-000025
小于λ 1时,组合特征的权重值为零;当
Figure PCTCN2020123721-appb-000026
大于λ 1时,组合特征的权重值为差值与第一和值的比值,差值为
Figure PCTCN2020123721-appb-000027
与λ 1的差值,第一和值 为N′与λ 2的和值;当
Figure PCTCN2020123721-appb-000028
小于-λ 1时,组合特征的权重值为第二和值与第一和值的比值,第二和值为
Figure PCTCN2020123721-appb-000029
与λ 1的和值。
其中,λ 1和λ 2均为非负实数;N′根据乘积值与第三和值的比值得到;乘积值为特征取值对应的正样本统计量,与特征取值对应的负样本统计量的乘积值;第三和值为特征取值对应的正样本统计量,与特征取值对应的负样本统计量的和值;
Figure PCTCN2020123721-appb-000030
根据特征取值对应的正样本统计量,与特征取值对应的负样本统计量的比值得到。
示例性的,本申请实施例可采用如下方法替代公式(5),近似计算相应的组合特征的权重值w F,j
Figure PCTCN2020123721-appb-000031
其中,
Figure PCTCN2020123721-appb-000032
F为特征组合方式;j为该特征组合方式F下所对应的各组合特征的特征取值;
Figure PCTCN2020123721-appb-000033
为特征取值j对应的正样本统计量;
Figure PCTCN2020123721-appb-000034
为特征取值j对应的负样本统计量。
这里,当正样本的数量或者负样本的数量较少时,使得N′也较小,那么
Figure PCTCN2020123721-appb-000035
则w F,j被强制限制为0;而随着样本数量不断增多,w F,j会变成非零值,但λ 1和λ 2的存在会降低w F,j的绝对值,也即降低w F,j对特征组合方式的有效性评估的影响,只有当样本数量充分多时,w F,j才会不断接近
Figure PCTCN2020123721-appb-000036
表明此时可以充分相信样本所得到的统计数据。
在一些实施例中,针对某特征对点击率预估模型的预测精度不准确的情形,服务器可通过如下方式确定组合特征的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量:确定包括特征取值的正样本为训练样本时对应的第一预测精度,以及包括特征取值的负样本为训练样本时对应的第二预测精度;基于第一预测精度,确定特征取值对应的正样本统计量;基于第二预测精度,确定特征取值对应的负样本统计量。
这里,第一预测精度表征基于特征取值的正样本进行推荐的准确度;第二预测精度表征基于特征取值的负样本进行推荐的准确度。
在实际应用中,评估基于特征组合方式组合得到的某个组合特征的重要性,要建立在已有点击率预估模型的基础上,即所增加的组合特征对点击率预估模型能起到多大的补充作用。如果按照前述实施例确定组合特征的特征取值对应的正样本统计量和负样本统计量,可能存在如下情况:计算得到的特征有效性较高,但点击率预估模型中已有其他特征能起到类似的效果,因此将这个特征输入至点击率预估模型中,并不会显著提升点击率预估模型的预测精度。
例如,如果已知点击率预估模型针对每个样本所预估出的点击率,可将公式(4)所示的目标函数调整为如下形式,以考虑点击率预估模型的影响:
Figure PCTCN2020123721-appb-000037
(12),其中,
Figure PCTCN2020123721-appb-000038
满足
Figure PCTCN2020123721-appb-000039
用于表示基于点击率预估模型得到的该样本的评分,如果
Figure PCTCN2020123721-appb-000040
表明点击率预估模型输出的该用户的点击概率大于50%,反之,则表明点击率预估模型输出的该用户的点击概率小于50%。
上述目标函数(12)也不存在解析解,本申请实施例采用如下公式修正特征取值对应的正样本统计量和负样本统计量:
Figure PCTCN2020123721-appb-000041
Figure PCTCN2020123721-appb-000042
然后,可通过公式(5)计算相应的组合特征的权重值w F,j,在该实施例中,对于一个正样本来说,点击率预估模型预测得越不准确,那么该样本就更应该被予以考虑;对于一个负样本来说,点击率预估模型预测得越不准确,那么该样本也更应该被予以考虑。
在一些实施例中,在工程应用中,针对训练集合过大的样本情形,服务器可通过如下方式确定组合特征的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量:确定第一采样样本的权重值及第二采样样本的权重值;基于第一采样样本的权重值,确定特征取值对应的正样本统计量;基于第二采样样本的权重值,确定特征取值对应的负样本统计量。
这里,第一采样样本,表征从包括特征取值的正样本中抽取的第一比例的样本;第二采样样本,表征从包括特征取值的负样本中抽取的第二比例的样本。
在实际应用中,工程应用中的数据通常可达到TB甚至PB量级,尽管通过上述实施例进行简单的统计量即可进行特征组合方式的有效性评估,但如果特征组合方式的数据量过大,这种评估仍然是不现实的。因此,在训练集样本较大时,本申请实施例提出了通过采样方法确定特征取值对应的正样本统计量和负样本统计量,此时,
Figure PCTCN2020123721-appb-000043
Figure PCTCN2020123721-appb-000044
的计算公式为:
Figure PCTCN2020123721-appb-000045
Figure PCTCN2020123721-appb-000046
然后,可通过公式(5)计算相应的组合特征的权重值w F,j,这里,在训练集样本较大时,可以从含有特征j的正样本中随机抽取α%的样本,从含有特征j的负样本中随机抽取β%的样本,进而按照上述公式(15)和(16)计算
Figure PCTCN2020123721-appb-000047
Figure PCTCN2020123721-appb-000048
这里,α%和β%可根据实际需要进行设定,在此不做限定。样本权重在上述不同的解决方案中是不相同的。在公式(6)(7)中,样本的权重为1;在公式(8)(9)(13)(14)中,样本的权重则是求和号内部的对应式子的值。
在另一些情形下,服务器还可通过以下方式确定组合特征的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量:确定第一部分样本中包括特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值;确定第一部分样本中包括特征取值的负样本的权重值,以及第二部分样本中第二采样样本的权重值;基于第一部分样本中包括特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,确定特征取值对应的正样本统计量;基于第一部分样本中包括特征取值的负样本的权重值,以及第二部分样本中第二采样样本的权重值,确定特征取值对应的负样本统计量。
这里,第一采样样本,表征从归属于第二部分样本且包括特征取值的正样本中抽取 的第一比例的样本;第二采样样本,表征从归属于第二部分样本且包括特征取值的负样本中抽取的第二比例的样本。
具体来说,在实际应用中,还可以通过部分采样的方法进行估计。所谓部分采样,是将整体样本分为两部分,即为A部分和B部分,A部分的所有样本均参与统计,而对B部分的样本进行采样处理。这样就可以将重要程度较高的样本(例如点击率预估模型预测偏差非常大的部分)放入A部分,保证这部分样本被充分考虑,而将重要程度较低的样本放入B部分进行采样处理,以减小计算量。此时,
Figure PCTCN2020123721-appb-000049
Figure PCTCN2020123721-appb-000050
可通过如下方式计算得到:
Figure PCTCN2020123721-appb-000051
Figure PCTCN2020123721-appb-000052
然后,可通过公式(5)计算相应的组合特征的权重值w F,j
需要说明的是,上述实施例给出了四个不同方向上的优化方案,来确定组合特征的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量,各个方案互相之间没有矛盾,可以彼此进行组合。
示例性的,以针对每个样本中各组合特征对应多个特征取值,以及针对某特征对点击率预估模型的预测精度不准确的改进方案进行组合为例进行说明。
上述针对每个样本中各组合特征对应多个特征取值的改进方案,以及针对某特征对点击率预估模型的预测精度不准确的改进方案均为调整
Figure PCTCN2020123721-appb-000053
Figure PCTCN2020123721-appb-000054
即调整求和号中每个样本的权重,在两者进行组合时,相应的样本权重相乘即可,即可通过如下公式进行计算:
Figure PCTCN2020123721-appb-000055
Figure PCTCN2020123721-appb-000056
在步骤504中,基于各组合特征的权重值,构建相应的特征组合方式的权重值集合。
这里,特征组合方式的权重值集合中包括对应各组合特征的权重值,该组合特征可基于相应的特征组合方式对原始特征进行组合得到。
在步骤505中,基于各特征组合方式的权重值集合,分别确定各特征组合方式的有效性。
在本申请实施例中,特征组合方式的有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
在一些实施例中,服务器可通过如下方式确定各特征组合方式的有效性:将权重值集合中的所有组合特征的权重值进行加权,得到对应特征组合方式的评分;基于各特征组合方式的评分,分别确定各特征组合方式的有效性。
这里,服务器可通过如下方式基于各特征组合方式的评分,分别确定各特征组合方式的有效性:将各特征组合方式的评分与目标评分进行比较,得到对应各特征组合方式的比较结果;基于比较结果,确定对应各特征组合方式的有效性。
在确定了每个特征组合方式对应的权重值集合后,计算相应的特征组合方式的有效 性,在实际应用中,可将权重值集合中的每个权重值进行加权,从而得到相应的特征组合方式的评分;将各特征组合方式的评分与对应的目标评分进行比较,得到比较结果,从而基于该比较结果,确定相应的特征组合方式的有效性。可见,应用上述实施例,实现了对各特征组合方式的有效性的计算,从而实现根据各特征组合方式的有效性对目标特征组合方式的筛选。
在实际实施时,可通过ROC曲线下与坐标轴围成的面积(AUC,Area Under Curve)、Logloss等精度指标,计算各特征组合方式的评分与目标评分的比较结果,从而确定各特征组合方式的有效性。这里,特征组合方式的评分,用于表征基于相应的特征组合方式得到的组合特征进行推荐时,待推荐内容被用户点击的可能性大小。
在一些实施例中,特征有效性评估方法还包括:在确定各特征组合方式的有效性后,基于各特征组合方式的有效性的排序,从特征组合方式集合中筛选得到目标数量的特征组合方式作为目标特征组合方式;基于目标特征组合方式,对原始特征进行特征组合,得到目标组合特征,以基于目标组合特征进行内容推荐。
也就是说,本申请实施例在确定了每个特征组合方式的有效性后,根据每个特征组合方式的有效性的大小,将各特征组合方式按照有效性从大到小进行排序,对特征组合方式集合中的多个特征组合方式进行筛选,以得到有效性高的目标数量的特征组合方式,作为目标特征组合方式,以基于确定的目标特征组合方式进行特征组合,得到用于内容推荐的目标组合特征。
在实际实施时,可以预先设置目标数量,将排序靠前的目标数量的特征组合方式,作为目标特征组合方式,示例性的,目标数量设定为30,则将排序靠前的前30个特征组合方式确定为目标特征组合方式。当然,还可以预先设置有效性阈值,将每个特征组合方式的有效性与有效性阈值进行比较,确定有效性达到该有效性阈值的每个特征组合方式作为目标特征组合方式。
在另一些实施例中,特征有效性评估方法还包括:在确定各特征组合方式的有效性后,基于各特征组合方式的有效性的排序,从特征组合方式集合中筛选得到目标数量的特征组合方式作为第一候选特征组合方式;基于第一候选特征组合方式及原始特征,生成多个第二候选特征组合方式;从多个第二候选特征组合方式中,选取符合筛选条件的特征组合方式作为目标特征组合方式;基于目标特征组合方式,对原始特征进行特征组合,得到目标组合特征,以基于目标组合特征进行内容推荐。
这里,为了进一步筛选出更加有效的特征组合方式,本申请实施例在筛选得到第一候选特征组合方式后,对第一候选特征组合方式进行扩增,以得到更多的特征组合方式作为第二候选特征组合方式。
在一些实施例中,服务器可通过如下方式生成多个第二候选特征组合方式:基于第一候选特征组合方式,对原始特征进行组合得到组合特征;确定组合特征与至少一个原始特征进行组合所得到的多个特征组合方式;基于多个特征组合方式及第一候选特征组合方式,生成多个第二候选特征组合方式。可见,本申请实施例基于第一候选特征组合方式及原始特征,生成了更多的第二候选特征组合方式,从而增加了特征组合方式的多样性,以得到更多有效的特征组合方式,提供内容推荐的准确性。
这里,在得到目标特征组合方式后,即可基于该目标特征组合方式,对获取的多个原始特征进行特征组合,得到目标组合特征。其中,目标特征组合方式可以是多个,可以基于每个目标特征组合方式对上述原始特征进行特征组合,也可以基于有效性最高的目标特征组合方式对上述原始特征进行特征组合,从而得到目标组合特征,以基于该目标组合特征对上述待推荐内容进行推荐。
在一些实施例中,服务器可通过如下方式基于目标组合特征进行内容推荐:将目标 组合特征作为输入特征,输入至点击率预估模型中,得到目标用户对待推荐内容的点击率;基于点击率,从待推荐内容中选取目标数量的内容作为目标推荐内容;将目标推荐内容返回至目标用户。
采用本申请实施例提供的技术方案,对于给定的包括多个特征组合方式的特征组合方式集合,通过获取各特征组合方式对应的特征值集合,这里特征值集合包括各组合特征的特征取值,基于各组合特征的特征取值,确定相应的组合特征的权重值,从而基于各组合特征的权重值,构建相应的特征组合方式的权重值集合,如此,基于各特征组合方式的权重值集合,就可以确定各特征组合方式的有效性,无需进行耗时的实验和模型训练,能够实现在短时间内评估大量的特征组合方式的有效性,提高对大量的特征组合方式的有效性进行评估的效率,进而提升推荐系统的推荐效果。
接下来将说明本申请实施例在一个实际的应用场景中的示例性应用,参见图6,图6为本申请实施例提供的特征有效性评估方法的流程示意图,在一些实施例中,该特征有效性评估方法可由终端实施,也可由服务器实施,还可由服务器及终端协同实施。
下面以服务器及终端协同实施为例,如通过图1中的终端100-1及服务器300协同实施,结合图6示出的步骤,对本申请实施例提供的特征有效性评估方法的实现进行说明。对于下文各步骤的说明中未尽的细节,可以参考上文而理解。本申请实施例提供的特征有效性评估方法可包括以下步骤:
在步骤601中,终端向服务器发起内容获取请求。
这里,终端对应的用户通过终端的界面触发内容获取指令,终端响应于内容获取指令,生成内容获取请求,并向服务器发送内容获取请求。
在步骤602中,服务器接收到内容获取请求后,构建包括多个特征组合方式的特征组合方式集合。
这里,特征组合方式为针对待推荐内容的原始特征的组合方式。在构建特征组合方式集合时,可以获取多个原始特征,即为用户或者待推荐内容的相关原始特征,比如“用户标识、内容标签”等,将获取的多个原始特征进行两两组合,以形成多个特征组合方式,从而根据所获取的多个原始特征及多个特征组合方式,构建特征组合方式集合。
在实际实施时,也可以选取两个以上的原始特征进行组合,以形成多个特征组合方式。
在步骤603中,服务器获取各特征组合方式对应的特征值集合。
这里,特征值集合包括各组合特征的特征取值,该组合特征可基于相应的特征组合方式对原始特征进行特征组合得到。其中,组合特征的特征取值可以是基于历史日志数据提取得到的,也可以是对历史日志数据进行采样得到的部分历史日志数据。
在步骤604中,服务器确定组合特征的各特征取值对应的正样本统计量,以及组合特征的各特征取值对应的负样本统计量。
这里,正样本,表征在待推荐内容的展示过程中,所展示的多个待推荐内容中被点击的内容对应的样本数据;负样本,表征在待推荐内容的展示过程中,所展示的多个待推荐内容中未被点击的内容对应的样本数据。
在步骤605中,服务器基于各特征取值对应的正样本统计量与负样本统计量,得到相应的组合特征的权重值。
在一些实施例中,服务器还可以采用权重计算模型的方式确定各组合特征的权重值,在实际实施时,将各组合特征的特征取值输入至权重计算模型中,得到权重计算模型输出的对应各组合特征的权重值。
在步骤606中,服务器基于各组合特征的权重值,构建相应的特征组合方式的权重 值集合。
这里,特征组合方式的权重值集合中包括对应各组合特征的权重值,该组合特征可基于相应的特征组合方式对原始特征进行组合得到。
在步骤607中,服务器将权重值集合中的所有组合特征的权重值进行加权,得到对应特征组合方式的评分。
这里,特征组合方式的评分,用于表征基于相应的特征组合方式得到的组合特征进行内容推荐时,待推荐内容被用户点击的可能性大小。
在步骤608中,服务器将各特征组合方式的评分与目标评分进行比较,得到对应各特征组合方式的比较结果。
这里,可通过AUC、Logloss等精度指标,计算各特征组合方式的评分与目标评分的比较结果。
在步骤609中,服务器基于比较结果,确定对应各特征组合方式的有效性。
这里,特征组合方式的有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
在步骤610中,服务器基于各特征组合方式的有效性的排序,从特征组合方式集合中筛选得到目标数量的特征组合方式作为目标特征组合方式。
这里,服务器在确定了各特征组合方式的有效性后,还可基于各特征组合方式的有效性的排序,从特征组合方式集合中筛选得到目标数量的特征组合方式作为第一候选特征组合方式;基于第一候选特征组合方式及原始特征,生成多个第二候选特征组合方式;从多个第二候选特征组合方式中,选取符合筛选条件的特征组合方式作为目标特征组合方式。
在步骤611中,服务器基于目标特征组合方式,对原始特征进行特征组合,得到目标组合特征。
在步骤612中,服务器基于目标组合特征确定目标推荐内容,并将目标推荐内容返回至终端。
这里,服务器可通过如下方式基于目标组合特征确定目标推荐内容:将目标组合特征作为输入特征,输入至点击率预估模型中,得到目标用户对待推荐内容的点击率;基于点击率,从待推荐内容中选取目标数量的内容作为目标推荐内容。
在步骤613中,终端将目标推荐内容进行呈现。
应用本申请实施例的技术方案,对于给定的包括多个特征组合方式的特征组合方式集合,通过获取各特征组合方式对应的特征值集合,这里特征值集合包括各组合特征的特征取值,基于各组合特征的特征取值,确定相应的组合特征的权重值,从而基于各组合特征的权重值,构建相应的特征组合方式的权重值集合,如此,基于各特征组合方式的权重值集合,就可以确定各特征组合方式的有效性,无需进行耗时的实验和模型训练,能够实现在短时间内评估大量的特征组合方式的有效性,提高对大量的特征组合方式的有效性进行评估的效率,进而提升推荐系统的推荐效果。
本申请实施例提供的特征有效性评估方法可应用在以下场景:
一种使用场景如下:某APP从零搭建推荐系统,以所用的输入特征(待推荐内容的原始特征)为用户ID、内容ID、内容标签、用户所在城市,以此预测目标用户的点击率。通常每个原始特征本身都难以单独预测点击率,但由各个原始特征进行组合得到的组合特征有可能对点击率提供有效信息。可通过本申请实施例提出的特征有效性评估方法,对所有可能的特征组合方式(也可以根据从业人员经验,预先挑选出一些可能有效的特征组合方式,如“用户ID-内容标签”,“用户所在城市-内容ID”等)进行有效性评估,并最终确定用于预测点击率的最有效的特征组合方式(即目标特征组合方式,例 如“用户ID-内容标签”)。
另一种使用场景如下:在某新闻类APP的推荐系统中,已经存在的原始特征包括用户ID、内容ID、内容标签、用户所在城市,目前新增了一个原始特征“发文媒体”,对于该新增的原始特征与之前存在的原始特征如何组合能够更有效地反映用户的兴趣是至关重要的。可能的特征组合方式包括“发文媒体-用户ID”,“发文媒体-内容ID”,“发文媒体-内容标签”,“发文媒体-用户所在城市”,“发文媒体-用户ID-内容标签”,“发文媒体-用户所在城市-内容标签”等,可通过本申请实施例提出的特征有效性评估方法,确定上述这些特征组合方式的评分,基于确定的各特征组合方式的评分与目标评分的比较结果,选择最有效的若干个目标特征组合方式(例如,“发文媒体-用户所在城市”),以用于该新闻类APP的点击率预估。
另一种使用场景如下:某APP的推荐系统具有一个特征自动化筛选的工具,可以从待推荐内容的一些原始特征,比如用户ID、内容ID、内容标签、用户所在城市等中挑选出可能有效的特征组合方式,针对这些特征组合方式,可通过本申请实施例提出的特征有效性评估方法,进一步评估各特征组合方式的有效性,并筛选出目标数量的最有效的目标特征组合方式,以用于该APP的点击率预估。
接下来继续对本申请实施例提供的特征有效性评估装置255的软件实现进行说明。以上述实施本申请实施例的特征有效性评估方法的电子设备20中的存储器250所包括的软件模块为例进行说明,对于下文关于模块的功能说明中未尽的细节,可以参考上文本申请方法实施例的描述而理解。如图3所示,本申请实施例提供的特征有效性评估装置255可以包括:
第一构建单元2551,配置为构建包括多个特征组合方式的特征组合方式集合,所述特征组合方式为针对待推荐内容的原始特征的组合方式;获取单元2552,配置为获取各所述特征组合方式对应的特征值集合,所述特征值集合包括各组合特征的特征取值,所述组合特征基于相应的特征组合方式对所述原始特征进行特征组合得到;第一确定单元2553,配置为基于各所述组合特征的特征取值,确定相应的组合特征的权重值;第二构建单元2554,配置为基于各所述组合特征的权重值,构建相应的特征组合方式的权重值集合;第二确定单元2555,配置为基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,所述有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
在一些实施例中,第一确定单元包括:
第一确定子单元,配置为确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量;
第二确定子单元,配置为基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值。
在一些实施例中,就第一确定子单元确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量来说,可以采用以下方式实现:
当所述组合特征的特征取值为一个时,确定包括所述特征取值的正样本统计量,为所述特征取值对应的正样本统计量,以及
确定包括所述特征取值的负样本统计量,为所述特征取值对应的负样本统计量。
在另一些实施例中,就第一确定子单元确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量来说,可以采用以下方式实现:
确定包括所述特征取值的正样本中组合特征的特征取值统计量,以及包括所述特征取值的负样本中组合特征的特征取值统计量;
基于包括所述特征取值的正样本中组合特征的特征取值统计量,确定所述特征取值对应的正样本统计量;
基于包括所述特征取值的负样本中组合特征的特征取值统计量,确定所述特征取值对应的负样本统计量。
在另一些实施例中,就第一确定子单元确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量来说,可以采用以下方式实现:
确定包括所述特征取值的正样本为训练样本时对应的第一预测精度,以及包括所述特征取值的负样本为训练样本时对应的第二预测精度;
基于所述第一预测精度,确定所述特征取值对应的正样本统计量;
基于所述第二预测精度,确定所述特征取值对应的负样本统计量;
其中,所述第一预测精度表征基于所述特征取值的正样本进行推荐的准确度;所述第二预测精度表征基于所述特征取值的负样本进行推荐的准确度。
在另一些实施例中,就第一确定子单元确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量来说,可以采用以下方式实现:
确定第一采样样本的权重值及第二采样样本的权重值;
其中,所述第一采样样本,表征从包括所述特征取值的正样本中抽取的第一比例的样本;所述第二采样样本,表征从包括所述特征取值的负样本中抽取的第二比例的样本;
基于所述第一采样样本的权重值,确定所述特征取值对应的正样本统计量;
基于所述第二采样样本的权重值,确定所述特征取值对应的负样本统计量。
在另一些实施例中,就第一确定子单元确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量来说,可以采用以下方式实现:
确定第一部分样本中包括所述特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,所述第一采样样本,表征从归属于所述第二部分样本且包括所述特征取值的正样本中抽取的第一比例的样本;
确定所述第一部分样本中包括所述特征取值的负样本的权重值,以及所述第二部分样本中第二采样样本的权重值,所述第二采样样本,表征从归属于所述第二部分样本且包括所述特征取值的负样本中抽取的第二比例的样本;
基于所述第一部分样本中包括所述特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,确定所述特征取值对应的正样本统计量;
基于所述第一部分样本中包括所述特征取值的负样本的权重值,以及所述第二部分样本中第二采样样本的权重值,确定所述特征取值对应的负样本统计量。
在一些实施例中,就第二确定子单元基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值来说,可以采用以下方式实现:
基于各所述特征取值对应的正样本统计量与负样本统计量,采用如下公式,得到相应的所述组合特征的权重值:
Figure PCTCN2020123721-appb-000057
其中,w F,j为组合特征的权重值;F为特征组合方式;j为该特征组合方式F下所 对应的各组合特征的特征取值;
Figure PCTCN2020123721-appb-000058
为特征取值j对应的正样本统计量;
Figure PCTCN2020123721-appb-000059
为特征取值j对应的负样本统计量。
在另一些实施例中,就第二确定子单元基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值来说,可以采用以下方式实现:
Figure PCTCN2020123721-appb-000060
小于λ 1时,所述组合特征的权重值为零;
Figure PCTCN2020123721-appb-000061
大于λ 1时,所述组合特征的权重值为差值与第一和值的比值,所述差值为
Figure PCTCN2020123721-appb-000062
与λ 1的差值,所述第一和值为N′与λ 2的和值;
Figure PCTCN2020123721-appb-000063
小于-λ 1时,所述组合特征的权重值为第二和值与所述第一和值的比值,所述第二和值为
Figure PCTCN2020123721-appb-000064
与λ 1的和值;
其中,λ 1和λ 2均为非负实数;N′根据乘积值与第三和值的比值得到;所述乘积值为所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的乘积值;所述第三和值为所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的和值;
Figure PCTCN2020123721-appb-000065
根据所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的比值得到。
这里,正样本表征在所述待推荐内容的展示过程中,所展示的多个待推荐内容中被点击的内容对应的样本数据;负样本表征在所述待推荐内容的展示过程中,所展示的多个待推荐内容中未被点击的内容对应的样本数据。
在一些实施例中,就第一确定单元基于各所述组合特征的特征取值,确定相应的组合特征的权重值来说,可以采用以下方式实现:
将各所述组合特征的特征取值输入至权重计算模型中,得到所述权重计算模型输出的各所述组合特征的权重值。
在一些实施例中,就第二确定单元基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性来说,可以采用以下方式实现:
将所述权重值集合中的所有所述组合特征的权重值进行加权,得到对应所述特征组合方式的评分;
基于各所述特征组合方式的评分,分别确定各所述特征组合方式的有效性。
在一些实施例中,就第二确定单元基于各所述特征组合方式的评分,分别确定各所述特征组合方式的有效性来说,可以采用以下方式实现:
将各所述特征组合方式的评分与目标评分进行比较,得到对应各所述特征组合方式的比较结果;
基于所述比较结果,确定对应各所述特征组合方式的有效性。
在一些实施例中,特征有效性评估装置还包括:
第一筛选单元,配置为基于各所述特征组合方式的有效性的排序,从所述特征组合方式集合中筛选得到目标数量的特征组合方式作为目标特征组合方式;
特征组合单元,配置为基于所述目标特征组合方式,对所述原始特征进行特征组合,得到目标组合特征;
内容推荐单元,配置为基于所述目标组合特征进行内容推荐。
在另一些实施例中,特征有效性评估装置还包括:
第二筛选单元,配置为基于各所述特征组合方式的有效性的排序,从所述特征组合方式集合中筛选得到目标数量的特征组合方式作为第一候选特征组合方式;
生成单元,配置为基于所述第一候选特征组合方式及所述原始特征,生成多个第二候选特征组合方式;
第三筛选单元,配置为从所述多个第二候选特征组合方式中,选取符合筛选条件的特征组合方式作为目标特征组合方式;
特征组合单元,配置为基于所述目标特征组合方式,对所述原始特征进行特征组合,得到目标组合特征;
内容推荐单元,配置为基于所述目标组合特征进行内容推荐。
在一些实施例中,就内容推荐单元基于所述目标组合特征进行内容推荐来说,可以采用以下方式实现:
将所述目标组合特征作为输入特征,输入至点击率预估模型中,得到目标用户对所述待推荐内容的点击率;
基于所述点击率,从所述待推荐内容中选取目标数量的内容作为目标推荐内容;
将所述目标推荐内容返回至所述目标用户。
本申请实施例还提供了一种电子设备,所述电子设备包括:
存储器,配置为存储可执行指令;
处理器,配置为执行存储器中存储的可执行指令时,实现本申请实施例提供的上述特征有效性评估方法。
本申请实施例还提供了一种存储介质,存储有可执行指令,可执行指令被处理器执行时,用于实现本申请实施例提供的上述特征有效性评估方法。
在一些实施例中,存储介质具体可为计算机可读存储介质,例如可以是铁电随机存取存储器(FRAM,Ferromagnetic Random Access Memory)、ROM、PROM、可擦除可编程只读存储器(EPROM,Erasable Programmable Read-Only Memory)、电可擦除可编程只读存储器(EEPROM,Electrically Erasable Programmable Read-Only Memory)、快闪存储器(Flash Memory)、磁表面存储器、光盘或只读光盘(CD-ROM,Compact Disc Read-Only Memory)等存储器;也可以是包括上述存储器之一或任意组合的各种设备。
在一些实施例中,可执行指令可以采用程序、软件、软件模块、脚本或代码的形式,按任意形式的编程语言(包括编译或解释语言,或者声明性或过程性语言)来编写,并且其可按任意形式部署,包括被部署为独立的程序或者被部署为模块、组件、子例程或者适合在计算环境中使用的其它单元。
作为示例,可执行指令可以但不一定对应于文件系统中的文件,可以可被存储在保存其它程序或数据的文件的一部分,例如,存储在超文本标记语言(HTML,Hyper Text Markup Language)文档中的一个或多个脚本中,存储在专用于所讨论的程序的单个文件中,或者,存储在多个协同文件(例如,存储一个或多个模块、子程序或代码部分的文件)中。
作为示例,可执行指令可被部署为在一个计算设备上执行,或者在位于一个地点的多个计算设备上执行,又或者,在分布在多个地点且通过通信网络互连的多个计算设备上执行。
综上所述,本申请实施例的技术方案具有以下有益效果:
1)无需进行耗时的实验和模型训练,能够在短时间内快速评估出大量的特征组合方式的有效性,提高对大量的特征组合方式的有效性进行评估的效率。
2)由于基于各特征组合方式的权重值集合确定各特征组合方式的有效性的计算速度快,因此,从业人员可以尝试尽可能多的特征组合方式,进而提升推荐系统的性能。
3)能够帮助开发人员快速持续向推荐系统中加入有效的组合特征,该组合特征可为基于各特征组合方式的有效性确定的目标特征组合方式,进行组合得到的目标组合特征,基于目标组合特征进行内容推荐,能够提升推荐系统的推荐效果。
4)本申请实施例的方案的原理简明且易实现,具有较强的工程价值,提高特征工 程的效率。
以上所述,仅为本申请的具体实施方式,但本申请的保护范围并不局限于此,任何熟悉本技术领域的技术人员在本申请揭露的技术范围内,可轻易想到变化或替换,都应涵盖在本申请的保护范围之内。因此,本申请的保护范围应以所述权利要求的保护范围为准。
工业实用性
本申请实施例中电子设备构建包括多个特征组合方式的特征组合方式集合,所述特征组合方式为针对待推荐内容的原始特征的组合方式;获取各所述特征组合方式对应的特征值集合,所述特征值集合包括各组合特征的特征取值,所述组合特征基于相应的特征组合方式对所述原始特征进行特征组合得到;基于各所述组合特征的特征取值,确定相应的组合特征的权重值;基于各所述组合特征的权重值,构建相应的特征组合方式的权重值集合;基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,所述有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度;如此,基于各特征组合方式的权重值集合,就可以确定各特征组合方式的有效性,无需进行耗时的实验和模型训练,能够实现在短时间内评估大量的特征组合方式的有效性,提高对大量的特征组合方式的有效性进行评估的效率,进而提升推荐系统的推荐效果。

Claims (34)

  1. 一种特征有效性评估方法,所述方法由电子设备执行,所述方法包括:
    构建包括多个特征组合方式的特征组合方式集合,所述特征组合方式为针对待推荐内容的原始特征的组合方式;
    获取各所述特征组合方式对应的特征值集合,所述特征值集合包括各组合特征的特征取值,所述组合特征基于相应的特征组合方式对所述原始特征进行特征组合得到;
    基于各所述组合特征的特征取值,确定相应的组合特征的权重值;
    基于各所述组合特征的权重值,构建相应的特征组合方式的权重值集合;
    基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,所述有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
  2. 如权利要求1所述的方法,其中,所述基于各所述组合特征的特征取值,确定相应的组合特征的权重值,包括:
    确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量;
    基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值。
  3. 如权利要求2所述的方法,其中,所述确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量,包括:
    当所述组合特征的特征取值为一个时,确定包括所述特征取值的正样本统计量,为所述特征取值对应的正样本统计量,以及
    确定包括所述特征取值的负样本统计量,为所述特征取值对应的负样本统计量。
  4. 如权利要求2所述的方法,其中,所述确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量,包括:
    确定包括所述特征取值的正样本中组合特征的特征取值统计量,以及包括所述特征取值的负样本中组合特征的特征取值统计量;
    基于包括所述特征取值的正样本中组合特征的特征取值统计量,确定所述特征取值对应的正样本统计量;
    基于包括所述特征取值的负样本中组合特征的特征取值统计量,确定所述特征取值对应的负样本统计量。
  5. 如权利要求2所述的方法,其中,所述确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量,包括:
    确定包括所述特征取值的正样本为训练样本时对应的第一预测精度,以及包括所述特征取值的负样本为训练样本时对应的第二预测精度;
    基于所述第一预测精度,确定所述特征取值对应的正样本统计量;
    基于所述第二预测精度,确定所述特征取值对应的负样本统计量;
    其中,所述第一预测精度表征基于所述特征取值的正样本进行推荐的准确度;所述第二预测精度表征基于所述特征取值的负样本进行推荐的准确度。
  6. 如权利要求2所述的方法,其中,所述确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量,包括:
    确定第一采样样本的权重值及第二采样样本的权重值;
    其中,所述第一采样样本,表征从包括所述特征取值的正样本中抽取的第一比例的 样本;所述第二采样样本,表征从包括所述特征取值的负样本中抽取的第二比例的样本;
    基于所述第一采样样本的权重值,确定所述特征取值对应的正样本统计量;
    基于所述第二采样样本的权重值,确定所述特征取值对应的负样本统计量。
  7. 如权利要求2所述的方法,其中,所述确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量,包括:
    确定第一部分样本中包括所述特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,所述第一采样样本,表征从归属于所述第二部分样本且包括所述特征取值的正样本中抽取的第一比例的样本;
    确定所述第一部分样本中包括所述特征取值的负样本的权重值,以及所述第二部分样本中第二采样样本的权重值,所述第二采样样本,表征从归属于所述第二部分样本且包括所述特征取值的负样本中抽取的第二比例的样本;
    基于所述第一部分样本中包括所述特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,确定所述特征取值对应的正样本统计量;
    基于所述第一部分样本中包括所述特征取值的负样本的权重值,以及所述第二部分样本中第二采样样本的权重值,确定所述特征取值对应的负样本统计量。
  8. 如权利要求2所述的方法,其中,所述基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值,包括:
    基于各所述特征取值对应的正样本统计量与负样本统计量,采用如下公式,得到相应的所述组合特征的权重值:
    Figure PCTCN2020123721-appb-100001
    其中,w F,j为组合特征的权重值;F为特征组合方式;j为该特征组合方式F下所对应的各组合特征的特征取值;
    Figure PCTCN2020123721-appb-100002
    为特征取值j对应的正样本统计量;
    Figure PCTCN2020123721-appb-100003
    为特征取值j对应的负样本统计量。
  9. 如权利要求2所述的方法,其中,所述基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值,包括:
    Figure PCTCN2020123721-appb-100004
    小于λ 1时,所述组合特征的权重值为零;
    Figure PCTCN2020123721-appb-100005
    大于λ 1时,所述组合特征的权重值为差值与第一和值的比值,所述差值为
    Figure PCTCN2020123721-appb-100006
    与λ 1的差值,所述第一和值为N′与λ 2的和值;
    Figure PCTCN2020123721-appb-100007
    小于-λ 1时,所述组合特征的权重值为第二和值与所述第一和值的比值,所述第二和值为
    Figure PCTCN2020123721-appb-100008
    与λ 1的和值;
    其中,λ 1和λ 2均为非负实数;N′根据乘积值与第三和值的比值得到;所述乘积值为所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的乘积值;所述第三和值为所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的和值;
    Figure PCTCN2020123721-appb-100009
    根据所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的比值得到。
  10. 如权利要求2至9任一项所述的方法,其中,
    所述正样本,表征在所述待推荐内容的展示过程中,所展示的多个待推荐内容中被点击的内容对应的样本数据;
    所述负样本,表征在所述待推荐内容的展示过程中,所展示的多个待推荐内容中未被点击的内容对应的样本数据。
  11. 如权利要求1所述的方法,其中,所述基于各所述组合特征的特征取值,确定相应的组合特征的权重值,包括:
    将各所述组合特征的特征取值输入至权重计算模型中,得到所述权重计算模型输出的各所述组合特征的权重值。
  12. 如权利要求1所述的方法,其中,所述基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,包括:
    将所述权重值集合中的所有所述组合特征的权重值进行加权,得到对应所述特征组合方式的评分;
    基于各所述特征组合方式的评分,分别确定各所述特征组合方式的有效性。
  13. 如权利要求12所述的方法,其中,所述基于各所述特征组合方式的评分,分别确定各所述特征组合方式的有效性,包括:
    将各所述特征组合方式的评分与目标评分进行比较,得到对应各所述特征组合方式的比较结果;
    基于所述比较结果,确定对应各所述特征组合方式的有效性。
  14. 如权利要求1所述的方法,其中,所述方法还包括:
    基于各所述特征组合方式的有效性的排序,从所述特征组合方式集合中筛选得到目标数量的特征组合方式作为目标特征组合方式;
    基于所述目标特征组合方式,对所述原始特征进行特征组合,得到目标组合特征,以基于所述目标组合特征进行内容推荐。
  15. 如权利要求1所述的方法,其中,所述方法还包括:
    基于各所述特征组合方式的有效性的排序,从所述特征组合方式集合中筛选得到目标数量的特征组合方式作为第一候选特征组合方式;
    基于所述第一候选特征组合方式及所述原始特征,生成多个第二候选特征组合方式;
    从所述多个第二候选特征组合方式中,选取符合筛选条件的特征组合方式作为目标特征组合方式;
    基于所述目标特征组合方式,对所述原始特征进行特征组合,得到目标组合特征,以基于所述目标组合特征进行内容推荐。
  16. 如权利要求14或15所述的方法,其中,所述基于所述目标组合特征进行内容推荐,包括:
    将所述目标组合特征作为输入特征,输入至点击率预估模型中,得到目标用户对所述待推荐内容的点击率;
    基于所述点击率,从所述待推荐内容中选取目标数量的内容作为目标推荐内容;
    将所述目标推荐内容返回至所述目标用户。
  17. 一种特征有效性评估装置,所述装置包括:
    第一构建单元,配置为构建包括多个特征组合方式的特征组合方式集合,所述特征组合方式为针对待推荐内容的原始特征的组合方式;
    获取单元,配置为获取各所述特征组合方式对应的特征值集合,所述特征值集合包括各组合特征的特征取值,所述组合特征基于相应的特征组合方式对所述原始特征进行特征组合得到;
    第一确定单元,配置为基于各所述组合特征的特征取值,确定相应的组合特征的权重值;
    第二构建单元,配置为基于各所述组合特征的权重值,构建相应的特征组合方式的权重值集合;
    第二确定单元,配置为基于各所述特征组合方式的权重值集合,分别确定各所述特征组合方式的有效性,所述有效性,用于表征基于相应的特征组合方式组合得到的特征进行内容推荐的准确度。
  18. 如权利要求17所述的装置,其中,所述第一确定单元包括:
    第一确定子单元,配置为确定所述组合特征的各特征取值对应的正样本统计量,以及所述组合特征的各特征取值对应的负样本统计量;
    第二确定子单元,配置为基于所述各特征取值对应的正样本统计量与负样本统计量,得到相应的所述组合特征的权重值。
  19. 如权利要求18所述的装置,其中,所述第一确定子单元,还配置为:
    当所述组合特征的特征取值为一个时,确定包括所述特征取值的正样本统计量,为所述特征取值对应的正样本统计量,以及
    确定包括所述特征取值的负样本统计量,为所述特征取值对应的负样本统计量。
  20. 如权利要求18所述的装置,其中,所述第一确定子单元,还配置为:
    确定包括所述特征取值的正样本中组合特征的特征取值统计量,以及包括所述特征取值的负样本中组合特征的特征取值统计量;
    基于包括所述特征取值的正样本中组合特征的特征取值统计量,确定所述特征取值对应的正样本统计量;
    基于包括所述特征取值的负样本中组合特征的特征取值统计量,确定所述特征取值对应的负样本统计量。
  21. 如权利要求18所述的装置,其中,所述第一确定子单元,还配置为:
    确定包括所述特征取值的正样本为训练样本时对应的第一预测精度,以及包括所述特征取值的负样本为训练样本时对应的第二预测精度;
    基于所述第一预测精度,确定所述特征取值对应的正样本统计量;
    基于所述第二预测精度,确定所述特征取值对应的负样本统计量;
    其中,所述第一预测精度表征基于所述特征取值的正样本进行推荐的准确度;所述第二预测精度表征基于所述特征取值的负样本进行推荐的准确度。
  22. 如权利要求18所述的装置,其中,所述第一确定子单元,还配置为:
    确定第一采样样本的权重值及第二采样样本的权重值;
    其中,所述第一采样样本,表征从包括所述特征取值的正样本中抽取的第一比例的样本;所述第二采样样本,表征从包括所述特征取值的负样本中抽取的第二比例的样本;
    基于所述第一采样样本的权重值,确定所述特征取值对应的正样本统计量;
    基于所述第二采样样本的权重值,确定所述特征取值对应的负样本统计量。
  23. 如权利要求18所述的装置,其中,所述第一确定子单元,还配置为:
    确定第一部分样本中包括所述特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,所述第一采样样本,表征从归属于所述第二部分样本且包括所述特征取值的正样本中抽取的第一比例的样本;
    确定所述第一部分样本中包括所述特征取值的负样本的权重值,以及所述第二部分样本中第二采样样本的权重值,所述第二采样样本,表征从归属于所述第二部分样本且包括所述特征取值的负样本中抽取的第二比例的样本;
    基于所述第一部分样本中包括所述特征取值的正样本的权重值,以及第二部分样本中第一采样样本的权重值,确定所述特征取值对应的正样本统计量;
    基于所述第一部分样本中包括所述特征取值的负样本的权重值,以及所述第二部分样本中第二采样样本的权重值,确定所述特征取值对应的负样本统计量。
  24. 如权利要求18所述的装置,其中,所述第二确定子单元,还配置为:
    基于各所述特征取值对应的正样本统计量与负样本统计量,采用如下公式,得到相应的所述组合特征的权重值:
    Figure PCTCN2020123721-appb-100010
    其中,w F,j为组合特征的权重值;F为特征组合方式;j为该特征组合方式F下所对应的各组合特征的特征取值;
    Figure PCTCN2020123721-appb-100011
    为特征取值j对应的正样本统计量;
    Figure PCTCN2020123721-appb-100012
    为特征取值j对应的负样本统计量。
  25. 如权利要求18所述的装置,其中,所述第二确定子单元,还配置为:
    Figure PCTCN2020123721-appb-100013
    小于λ 1时,所述组合特征的权重值为零;
    Figure PCTCN2020123721-appb-100014
    大于λ 1时,所述组合特征的权重值为差值与第一和值的比值,所述差值为
    Figure PCTCN2020123721-appb-100015
    与λ 1的差值,所述第一和值为N′与λ 2的和值;
    Figure PCTCN2020123721-appb-100016
    小于-λ 1时,所述组合特征的权重值为第二和值与所述第一和值的比值,所述第二和值为
    Figure PCTCN2020123721-appb-100017
    与λ 1的和值;
    其中,λ 1和λ 2均为非负实数;N′根据乘积值与第三和值的比值得到;所述乘积值为所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的乘积值;所述第三和值为所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的和值;
    Figure PCTCN2020123721-appb-100018
    根据所述特征取值对应的正样本统计量,与所述特征取值对应的负样本统计量的比值得到。
  26. 如权利要求18至25任一项所述的装置,其中,
    所述正样本,表征在所述待推荐内容的展示过程中,所展示的多个待推荐内容中被点击的内容对应的样本数据;
    所述负样本,表征在所述待推荐内容的展示过程中,所展示的多个待推荐内容中未被点击的内容对应的样本数据。
  27. 如权利要求17所述的装置,其中,所述第一确定单元,还配置为:
    将各所述组合特征的特征取值输入至权重计算模型中,得到所述权重计算模型输出的各所述组合特征的权重值。
  28. 如权利要求17所述的装置,其中,所述第二确定单元,还配置为:
    将所述权重值集合中的所有所述组合特征的权重值进行加权,得到对应所述特征组合方式的评分;
    基于各所述特征组合方式的评分,分别确定各所述特征组合方式的有效性。
  29. 如权利要求28所述的装置,其中,所述第二确定单元,还配置为:
    将各所述特征组合方式的评分与目标评分进行比较,得到对应各所述特征组合方式的比较结果;
    基于所述比较结果,确定对应各所述特征组合方式的有效性。
  30. 如权利要求17所述的装置,其中,所述装置还包括:
    第一筛选单元,配置为基于各所述特征组合方式的有效性的排序,从所述特征组合方式集合中筛选得到目标数量的特征组合方式作为目标特征组合方式;
    特征组合单元,配置为基于所述目标特征组合方式,对所述原始特征进行特征组合,得到目标组合特征;
    内容推荐单元,配置为基于所述目标组合特征进行内容推荐。
  31. 如权利要求17所述的装置,其中,所述装置还包括:
    第二筛选单元,配置为基于各所述特征组合方式的有效性的排序,从所述特征组合 方式集合中筛选得到目标数量的特征组合方式作为第一候选特征组合方式;
    生成单元,配置为基于所述第一候选特征组合方式及所述原始特征,生成多个第二候选特征组合方式;
    第三筛选单元,配置为从所述多个第二候选特征组合方式中,选取符合筛选条件的特征组合方式作为目标特征组合方式;
    特征组合单元,配置为基于所述目标特征组合方式,对所述原始特征进行特征组合,得到目标组合特征;
    内容推荐单元,配置为基于所述目标组合特征进行内容推荐。
  32. 如权利要求30或31所述的装置,其中,所述内容推荐单元,还配置为:
    将所述目标组合特征作为输入特征,输入至点击率预估模型中,得到目标用户对所述待推荐内容的点击率;
    基于所述点击率,从所述待推荐内容中选取目标数量的内容作为目标推荐内容;
    将所述目标推荐内容返回至所述目标用户。
  33. 一种电子设备,所述电子设备包括:
    存储器,配置为存储可执行指令;
    处理器,配置为执行所述存储器中存储的可执行指令时,实现如权利要求1至16任一项所述的特征有效性评估方法。
  34. 一种存储介质,存储有可执行指令,所述可执行指令被处理器执行时,用于实现如权利要求1至16任一项所述的特征有效性评估方法。
PCT/CN2020/123721 2020-01-03 2020-10-26 特征有效性评估方法、装置、电子设备及存储介质 WO2021135562A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/703,822 US20220215296A1 (en) 2020-01-03 2022-03-24 Feature effectiveness assessment method and apparatus, electronic device, and storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010007053.1A CN111242310B (zh) 2020-01-03 2020-01-03 特征有效性评估方法、装置、电子设备及存储介质
CN202010007053.1 2020-01-03

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/703,822 Continuation US20220215296A1 (en) 2020-01-03 2022-03-24 Feature effectiveness assessment method and apparatus, electronic device, and storage medium

Publications (1)

Publication Number Publication Date
WO2021135562A1 true WO2021135562A1 (zh) 2021-07-08

Family

ID=70876006

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/123721 WO2021135562A1 (zh) 2020-01-03 2020-10-26 特征有效性评估方法、装置、电子设备及存储介质

Country Status (3)

Country Link
US (1) US20220215296A1 (zh)
CN (1) CN111242310B (zh)
WO (1) WO2021135562A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108090032B (zh) * 2018-01-03 2021-03-23 第四范式(北京)技术有限公司 逻辑回归模型的可视化解释方法及装置
CN111242310B (zh) * 2020-01-03 2023-04-18 深圳市雅阅科技有限公司 特征有效性评估方法、装置、电子设备及存储介质
CN111274480B (zh) * 2020-01-17 2023-04-04 深圳市雅阅科技有限公司 用于内容推荐的特征组合方法及装置
CN111914172B (zh) * 2020-07-29 2021-09-10 上海梅斯医药科技有限公司 一种基于用户标签的医学信息推荐方法及系统
CN112001442B (zh) * 2020-08-24 2024-03-19 北京达佳互联信息技术有限公司 特征检测方法、装置、计算机设备及存储介质
US20220101068A1 (en) * 2020-09-30 2022-03-31 International Business Machines Corporation Outlier detection in a deep neural network using t-way feature combinations
CN111966914B (zh) * 2020-10-26 2021-01-15 腾讯科技(深圳)有限公司 基于人工智能的内容推荐方法、装置和计算机设备
CN113344147A (zh) * 2021-08-04 2021-09-03 北京世纪好未来教育科技有限公司 一种特征共享建模方法、装置、设备及其存储介质
CN116402354A (zh) * 2023-06-07 2023-07-07 北京京东乾石科技有限公司 评估参数确定方法、装置、介质与电子设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109902849A (zh) * 2018-06-20 2019-06-18 华为技术有限公司 用户行为预测方法及装置、行为预测模型训练方法及装置
CN109992710A (zh) * 2019-02-13 2019-07-09 网易传媒科技(北京)有限公司 点击率预估方法、系统、介质和计算设备
CN110119474A (zh) * 2018-05-16 2019-08-13 华为技术有限公司 推荐模型训练方法、基于推荐模型的预测方法及装置
CN111242310A (zh) * 2020-01-03 2020-06-05 腾讯科技(北京)有限公司 特征有效性评估方法、装置、电子设备及存储介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005008572A1 (en) * 2003-07-11 2005-01-27 Computer Associates Think, Inc. Method and apparatus for automated feature selection
CN108764486A (zh) * 2018-05-23 2018-11-06 哈尔滨工业大学 一种基于集成学习的特征选择方法及装置
FR3082963A1 (fr) * 2018-06-22 2019-12-27 Amadeus S.A.S. Systeme et procede d'evaluation et de deploiement de modeles d'apprentissage automatique non supervises ou semi-supervises
CN109685107A (zh) * 2018-11-22 2019-04-26 东软集团股份有限公司 特征选择方法、系统、计算机可读存储介质及电子设备
CN110276369B (zh) * 2019-04-24 2021-07-30 武汉众邦银行股份有限公司 基于机器学习的特征选择方法、装置、设备及存储介质

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110119474A (zh) * 2018-05-16 2019-08-13 华为技术有限公司 推荐模型训练方法、基于推荐模型的预测方法及装置
CN109902849A (zh) * 2018-06-20 2019-06-18 华为技术有限公司 用户行为预测方法及装置、行为预测模型训练方法及装置
CN109992710A (zh) * 2019-02-13 2019-07-09 网易传媒科技(北京)有限公司 点击率预估方法、系统、介质和计算设备
CN111242310A (zh) * 2020-01-03 2020-06-05 腾讯科技(北京)有限公司 特征有效性评估方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
CN111242310A (zh) 2020-06-05
US20220215296A1 (en) 2022-07-07
CN111242310B (zh) 2023-04-18

Similar Documents

Publication Publication Date Title
WO2021135562A1 (zh) 特征有效性评估方法、装置、电子设备及存储介质
US10958748B2 (en) Resource push method and apparatus
WO2021139325A1 (zh) 媒体信息推荐方法、装置、电子设备及存储介质
CN110781321B (zh) 一种多媒体内容推荐方法及装置
US20230153857A1 (en) Recommendation model training method, recommendation method, apparatus, and computer-readable medium
US20230102337A1 (en) Method and apparatus for training recommendation model, computer device, and storage medium
Singer et al. Detecting memory and structure in human navigation patterns using Markov chain models of varying order
CN112085205A (zh) 用于自动训练机器学习模型的方法和系统
WO2019047790A1 (zh) 生成机器学习样本的组合特征的方法及系统
WO2015188699A1 (zh) 推荐项目的方法和装置
WO2017197806A1 (zh) 基于人工智能提供智能服务的方法、智能服务系统及智能终端
US11288709B2 (en) Training and utilizing multi-phase learning models to provide digital content to client devices in a real-time digital bidding environment
CN112508609B (zh) 人群扩量的预测方法、装置、设备及存储介质
CN113139134B (zh) 一种社交网络中用户生成内容的流行度预测方法、装置
WO2016107354A1 (zh) 提供用户个性化资源消息推送的方法和装置
US20220172260A1 (en) Method, apparatus, storage medium, and device for generating user profile
CN111400603A (zh) 一种信息推送方法、装置、设备及计算机可读存储介质
CN111143684A (zh) 基于人工智能的泛化模型的训练方法及装置
WO2023050143A1 (zh) 一种推荐模型训练方法及装置
Ouyang et al. Modeling and forecasting the popularity evolution of mobile apps: A multivariate hawkes process approach
CN117194779A (zh) 基于人工智能的营销系统优化方法、装置及设备
CN113836388A (zh) 信息推荐方法、装置、服务器及存储介质
CN114298118B (zh) 一种基于深度学习的数据处理方法、相关设备及存储介质
US20240070743A1 (en) Systems and methods for recommended sorting of search results for online searching
WO2024113641A1 (zh) 视频推荐方法、装置、电子设备、计算机可读存储介质及计算机程序产品

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20908480

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20908480

Country of ref document: EP

Kind code of ref document: A1