A method and device for recommending advertisement placement

Abstract

The present invention relates to a method and device for recommending advertisement placement, comprising: determining a first candidate associated tag of the target object according to the correspondence between the object tag and the associated tag of the target object; determining the similarity between the target object and the historical object, and determining the similar object of the target object from all the historical objects according to the similarity; selecting the actual associated tag of the similar object as the second candidate associated tag of the target object; for each candidate associated tag of the target object, inputting the candidate associated tag and the characteristic data of the target object into a click prediction model to obtain the predicted click probability of the candidate associated tag for the target object; determining a recommended associated tag from all the candidate associated tags according to the predicted click probability; and determining the advertisement placement strategy of the target object according to the recommended associated tag. The accuracy and pertinence of the insurance product placement are increased, and the insurance product placement conversion rate is improved.

Description

Technical Field

[0001] The present invention relates to the field of big data processing of Fintech, and in particular to a method and device for recommending advertisement placement. Background Art

[0002] With the continuous development of financial technology, especially Internet technology finance, more and more technologies (such as distributed, big data, artificial intelligence, etc.) are being applied in the financial field, but the financial industry has also put forward higher requirements for technology.

[0003] Insurance products are a combination of tangible products and intangible services provided by insurance companies to the market. In a narrow sense, insurance products refer to financial instruments created by insurance companies that can be traded by customers in the insurance market; in a broad sense, they refer to all products and services provided by insurance companies to the market and available for customers to obtain, use or consume, all of which fall within the scope of insurance product services. In order to promote insurance products or provide services, it is necessary to advertise insurance products.

[0004] Currently, the distribution of insurance products is generally carried out by business personnel who select appropriate channels that match the products based on the characteristics of the insurance products. This method is highly subjective, with low accuracy and specificity, and cannot effectively improve the conversion rate of insurance product distribution. Summary of the invention

[0005] The present application provides a method and device for recommending advertisement placement, which are used to increase the accuracy and pertinence of insurance product placement and improve the conversion rate of insurance product placement.

[0006] An advertisement placement recommendation method provided by an embodiment of the present invention includes:

[0007] Determine a first candidate associated tag of the target object according to a correspondence between an object tag and an associated tag of the target object; the correspondence between the object tag and the associated tag is set according to characteristics of the target object and characteristics of the channel;

[0008] Determine the similarity between the target object and the historical object, and determine similar objects of the target object from all the historical objects according to the similarity;

[0009] Selecting the actual associated tag of the similar object as the second candidate associated tag of the target object; the actual associated tag is determined by performing point-of-view processing on the user behavior event of the delivered historical object;

[0010] For each candidate association tag of the target object, the candidate association tag and the feature data of the target object are input into a click prediction model to obtain the predicted click probability of the candidate association tag for the target object; the click prediction model is trained according to the feature data of the training sample and the actual click results to obtain corresponding model parameters;

[0011] Determine a recommended associated tag from all candidate associated tags according to the click probability;

[0012] An advertisement delivery strategy for the target object is determined according to the recommendation associated tag.

[0013] In an optional embodiment, before determining the first candidate associated tag of the target object according to the correspondence between the object tag and the associated tag of the target object, the method further includes:

[0014] Determining a preliminary associated tag of the historical object according to the feature data of the historical object;

[0015] Obtaining user behavior data during the delivery of the historical object;

[0016] According to the user behavior data, the preliminary association tag of the historical object is modified to obtain the actual association tag of the historical object.

[0017] In an optional embodiment, the click prediction model is trained in the following manner:

[0018] Acquire a training sample, wherein the training sample includes feature data of a training object, feature data of a training user, and actual click results of the training user on the training object;

[0019] Inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object;

[0020] Determining the predicted click probability of the training object according to the predicted click results of all training users for the training object;

[0021] Determining an actual click rate of the training object according to actual click results of the training user on the training object;

[0022] A loss function is calculated according to the actual click rate and the predicted click probability, and when the loss function is less than a preset threshold, parameters corresponding to the click prediction model are determined.

[0023] In an optional embodiment, the click prediction model includes a GBDT gradient boosting iterative decision tree model and an LR logistic regression model;

[0024] The step of inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object includes:

[0025] For each training user, input the feature data of the training user and the feature data of the training object into the GBDT model to construct multiple decision trees;

[0026] Using the multiple decision trees, obtaining predicted click results of the training user on the training object;

[0027] Determine the combined feature value of the GBDT model according to the predicted click results of all decision trees;

[0028] The step of determining the predicted click probability of the training object according to the predicted click results of all training users for the training object includes:

[0029] The combined feature values ​​of the GBDT model are input into the LR model to obtain the predicted click probability of the training object.

[0030] In an optional embodiment, the similarity between the target object and the historical object is determined according to the following formula:

[0031]

[0032] Among them, R u,i represents the score given by user u to target object i, R u,j represents the score given by user u to historical object j; represents the average score of target object i, It represents the average score of historical object j. sim(i,j) is the similarity between the target object and the historical object.

[0033] In an optional embodiment, the selecting the actual associated tag of the similar object as the second candidate associated tag of the target object includes:

[0034] Calculating the matching degree between each actual associated label of the similar object and the target object;

[0035] selecting the second candidate associated tag from actual associated tags of the similar objects according to the matching degree;

[0036] The degree of matching is calculated according to the following formula:

[0037]

[0038] Among them, S i,N is the similarity between the target object i and the similar object N, Ru,N is the score given by user u to similar object N.

[0039] The embodiment of the present invention further provides an advertisement placement recommendation device, comprising:

[0040] A determination unit, configured to determine a first candidate associated tag of the target object according to a correspondence between an object tag and an associated tag of the target object; the correspondence between the object tag and the associated tag is set according to characteristics of the target object and characteristics of the channel;

[0041] A selection unit is used to determine the similarity between the target object and the historical object, and determine the similar object of the target object from all the historical objects according to the similarity; select the actual associated tag of the similar object as the second candidate associated tag of the target object; the actual associated tag is determined by burying the user behavior event of the delivered historical object;

[0042] A calculation unit is used to input the candidate association tag and the feature data of the target object into a click prediction model for each candidate association tag of the target object, so as to obtain the predicted click probability of the candidate association tag for the target object; the click prediction model is trained according to the feature data of the training sample and the actual click result to obtain the corresponding model parameters;

[0043] An association unit, used to determine a recommended association tag from all candidate association tags according to the predicted click probability;

[0044] The delivery unit is used to determine the advertisement delivery strategy for the target object according to the recommendation associated tag.

[0045] In an optional embodiment, the selection unit is further used to:

[0046] Determining a preliminary associated tag of the historical object according to the feature data of the historical object;

[0047] Obtaining user behavior data during the delivery of the historical object;

[0048] According to the user behavior data, the preliminary association tag of the historical object is modified to obtain the actual association tag of the historical object.

[0049] In an optional embodiment, a training unit is further included, which is used to train the click prediction model according to the following method:

[0050] Acquire a training sample, wherein the training sample includes feature data of a training object, feature data of a training user, and actual click results of the training user on the training object;

[0051] Inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object;

[0052] Determining the predicted click probability of the training object according to the predicted click results of all training users for the training object;

[0053] Determining an actual click rate of the training object according to actual click results of the training user on the training object;

[0054] A loss function is calculated according to the actual click rate and the predicted click probability, and when the loss function is less than a preset threshold, parameters corresponding to the click prediction model are determined.

[0055] In an optional embodiment, the click prediction model includes a GBDT (gradient boosted iterative decision tree) model and an LR (logistic regression) model;

[0056] The training unit is further used for:

[0057] For each training user, input the feature data of the training user and the feature data of the training object into the GBDT model to construct multiple decision trees;

[0058] Using the multiple decision trees, obtaining predicted click results of the training user on the training object;

[0059] Determine the combined feature value of the GBDT model according to the predicted click results of all decision trees;

[0060] The combined feature values ​​of the GBDT model are input into the LR model to obtain the predicted click probability of the training object.

[0061] In an optional embodiment, the selection unit is specifically configured to determine the similarity between the target object and the historical object according to the following formula:

[0062]

[0063] Among them, R u,i represents the score given by user u to target object i, R u,j represents the score given by user u to historical object j; represents the average score of target object i, It represents the average score of historical object j. sim(i,j) is the similarity between the target object and the historical object.

[0064] In an optional embodiment, the selection unit is specifically used to:

[0065] Calculating the matching degree between each actual associated label of the similar object and the target object;

[0066] selecting the second candidate associated tag from actual associated tags of the similar objects according to the matching degree;

[0067] The degree of matching is calculated according to the following formula:

[0068]

[0069] Among them, S i,N is the similarity between the target object i and the similar object N, R u,N is the score given by user u to similar object N.

[0070] An embodiment of the present invention further provides an electronic device, including:

[0071] at least one processor; and,

[0072] a memory communicatively connected to the at least one processor; wherein,

[0073] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described above.

[0074] An embodiment of the present invention further provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method as described above.

[0075] In an embodiment of the present invention, when formulating an advertisement delivery strategy for a target object, a first candidate associated tag for the target object is determined based on the correspondence between the existing object tag and the associated tag of the target object. On the other hand, a certain number of historical objects are selected, the similarity between the target object and each historical object is determined, a similar object of the target object is selected from the historical objects based on the similarity, and a second candidate associated tag for the target object is selected from the actual associated tags of the similar objects. In this way, the candidate associated tags for the target object include a first candidate associated tag and a second candidate associated tag. For each candidate associated tag of the target object, the characteristic data of the candidate associated tag and the target object are input into a click prediction model to obtain the predicted click probability of the candidate associated tag for the target object. Here, the click prediction model is trained based on the characteristic data of the training sample (i.e., the characteristic data of the target object) and the actual click result to obtain the corresponding model parameters. Furthermore, all candidate associated tags can be sorted based on the predicted click probability, the recommended associated tag of the target object can be determined from all candidate associated tags, and the advertisement delivery strategy for the target object can be determined based on the recommended associated tags. In the embodiment of the present invention, the feature data of the collected historical objects are used as training samples to train the click prediction model, predict the predicted click probability of the candidate associated tags of the target object, select the recommended associated tags from all the candidate associated tags of the target object according to the predicted click probability, and then determine the advertising delivery strategy of the target object according to the recommended associated tags, based on the real historical objects, thereby increasing the pertinence of advertising delivery and improving the accuracy of delivery, making the advertising delivery of insurance products more accurate and appropriate, and effectively improving the delivery conversion rate of insurance products. Furthermore, the actual associated tags of historical objects with high similarity to the target object are used as candidate associated tags of the target object, which fully considers the characteristics of the target object, analyzes and recommends the advertising delivery strategy of the target object from multiple aspects, thereby expanding the scope of advertising delivery and further improving the advertising delivery conversion rate of insurance products. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0077] Figure 1 A schematic diagram of a possible system architecture provided by an embodiment of the present invention;

[0078] Figure 2 A flowchart of a method for recommending advertisement delivery provided by an embodiment of the present invention;

[0079] Figure 3 A schematic diagram of a corresponding relationship between a product label and a user label provided by an embodiment of the present invention;

[0080] Figure 4 A schematic diagram of the structure of an advertisement placement recommendation device provided by an embodiment of the present invention;

[0081] Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] like Figure 1 As shown, a system architecture applicable to an embodiment of the present invention includes an advertisement delivery system 101 and a delivery channel system 102. The advertisement delivery system 101 can be a network device such as a computer, an independent device, or a server cluster formed by multiple servers. Preferably, the advertisement delivery system 101 can use cloud computing technology for information processing.

[0084] The delivery channel system 102 may be a network device such as a computer, an independent device, or a server cluster formed by multiple servers. Preferably, the delivery channel system 102 may use cloud computing technology for information processing.

[0085] The advertisement delivery system 101 is connected to the delivery channel system 102 via a wired or wireless network. Optionally, the wireless network or wired network uses standard communication technology and / or protocol. The network is usually the Internet, but it can also be any network, including but not limited to a local area network (LAN), a metropolitan area network (MAN), a wide area network (WAN), a mobile, wired or wireless network, a dedicated network or any combination of a virtual private network). In some embodiments, technologies and / or formats including Hyper Text Mark-up Language (HTML), Extensible Markup Language (XML), etc. are used to represent data exchanged through the network. In addition, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), Internet Protocol Security (IPsec), etc. can also be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can also be used to replace or supplement the above data communication technologies.

[0086] In an embodiment of the present invention, a method for recommending advertisement delivery is provided. When formulating an advertisement delivery strategy for a target object, the advertisement delivery system 101 determines a first candidate associated tag for the target object based on the correspondence between the existing object tag and the associated tag of the target object. On the other hand, the advertisement delivery system 101 selects a certain number of historical objects, determines the similarity between the target object and each historical object, selects a similar object of the target object from the historical objects based on the similarity, and selects a second candidate associated tag for the target object from the actual associated tags of the similar objects. In this way, the candidate associated tags of the target object include a first candidate associated tag and a second candidate associated tag. For each candidate associated tag of the target object, the characteristic data of the candidate associated tag and the target object are input into a click prediction model to obtain a predicted click probability of the candidate associated tag for the target object. Furthermore, the advertisement delivery system 101 can sort all candidate associated tags according to the predicted click probability, determine the recommended associated tag of the target object from all candidate associated tags, and determine the advertisement delivery strategy for the target object based on the recommended associated tag. The advertisement delivery strategy here includes the delivery channel, delivery time, etc. of the target object. In this way, the advertisement delivery system 101 can send the target object to the delivery channel system 102 according to the advertisement delivery strategy.

[0087] Based on the above architecture, an embodiment of the present invention provides a method for recommending advertisement delivery, such as Figure 2 As shown, the advertisement delivery recommendation method provided by the embodiment of the present invention includes the following steps:

[0088] Step 201: Determine a first candidate associated tag of the target object according to a correspondence between an object tag and an associated tag of the target object.

[0089] Step 202: Determine the similarity between the target object and the historical objects, and determine similar objects of the target object from all historical objects according to the similarity.

[0090] Step 203: Select the actual associated tag of the similar object as the second candidate associated tag of the target object.

[0091] Step 204: For each candidate associated tag of the target object, input the candidate associated tag and the characteristic data of the target object into a click prediction model to obtain a predicted click probability of the candidate associated tag for the target object.

[0092] The candidate association tags of the target object include a first candidate association tag and a second candidate association tag of the target object.

[0093] The click prediction model is trained according to the characteristic data of the training samples and the actual click results to obtain corresponding model parameters.

[0094] Specifically, a candidate association tag and the feature data of the target object are input into the click prediction model, and the predicted click result of the candidate association tag for each feature data is calculated. Then, the predicted click results of the candidate association tag for all feature data are merged, and the predicted click probability of the candidate association tag for the target object is calculated based on the merged predicted click results.

[0095] Step 205: Determine a recommended associated tag from all candidate associated tags according to the predicted click probability.

[0096] Step 206: Determine an advertisement delivery strategy for the target object according to the recommendation associated tags.

[0097] In an embodiment of the present invention, when formulating an advertising delivery strategy for a target object, a first candidate associated tag for the target object is determined based on the correspondence between the existing object tag and the associated tag of the target object. The correspondence between the object tag and the associated tag is mainly selected and set by the staff according to the characteristics of the target object and the channel characteristics. On the other hand, a certain number of historical objects are selected, the similarity between the target object and each historical object is determined, a similar object of the target object is selected from the historical objects according to the similarity, and a second candidate associated tag for the target object is selected from the actual associated tags of the similar objects. In this way, the candidate associated tags of the target object include a first candidate associated tag and a second candidate associated tag. For each candidate associated tag of the target object, the characteristic data of the candidate associated tag and the target object are input into a click prediction model to obtain the predicted click probability of the candidate associated tag for the target object. Here, the click prediction model is trained according to the characteristic data of the training sample and the actual click result to obtain the corresponding model parameters. Furthermore, all candidate associated tags can be sorted according to the predicted click probability, the recommended associated tag of the target object can be determined from all candidate associated tags, and the advertising delivery strategy of the target object can be determined according to the recommended associated tag. In the embodiment of the present invention, the feature data of the collected historical objects are used as training samples to train the click prediction model, predict the predicted click probability of the candidate associated tags of the target object, select the recommended associated tags from all the candidate associated tags of the target object according to the predicted click probability, and then determine the advertising delivery strategy of the target object according to the recommended associated tags, based on the real historical objects, thereby increasing the pertinence of advertising delivery and improving the accuracy of delivery, making the advertising delivery of insurance products more accurate and appropriate, and effectively improving the delivery conversion rate of insurance products. Furthermore, the actual associated tags of historical objects with high similarity to the target object are used as candidate associated tags of the target object, which fully considers the characteristics of the target object, analyzes and recommends the advertising delivery strategy of the target object from multiple aspects, thereby expanding the scope of advertising delivery and further improving the advertising delivery conversion rate of insurance products.

[0098] The conversion rate of an ad is the ratio of the number of completed conversions to the total number of clicks on the promotion information. The calculation formula is: conversion rate = (number of conversions / number of clicks) × 100%. For example: 10 target users saw the insurance ad, 5 of them clicked on the target URL, and then 2 of them left their information and made subsequent conversions. Then, the conversion rate of this promotion result is (2 / 5) × 100% = 40%.

[0099] When the insurance product in the embodiment of the present invention is first launched, a preliminary launch strategy will be formulated. The preliminary launch strategy is for business personnel to select a suitable strategy that matches the product and channel for launch based on the characteristics of the insurance product and the channel characteristics.

[0100] Specifically, the launch of insurance products is mainly based on two aspects, namely the characteristics of the insurance products themselves and the characteristics of the launch channels. Therefore, it is necessary to manage and enter some information about products and channels so that relevant information can be traced and updated when necessary.

[0101] In addition to entering the basic information of insurance products, such as product name, coverage details, claims terms and procedures, etc., insurance products also need to be labeled so that they can be laterally classified and compared. Currently used product labels mainly include insurance types, special insurance targets, product target groups (mainly gender, age), etc. The target object in the embodiment of the present invention is the insurance product here, and the object label is the product label.

[0102] Channel management aims to input relevant information of each delivery channel, such as channel name, channel rate, channel user characteristics and audit points, and to specify the audience characteristics of the channel itself as detailed and clear as possible. This information needs to be determined after in-depth communication and understanding with the channel. At the same time, it will also include the channel's audit requirements for delivery, to provide a reference for subsequent delivery strategies in the channel, so that the delivery is legal and compliant, and can effectively pass the channel audit.

[0103] In a specific embodiment, the target object may be a target product, and the associated tag may be a user tag.

[0104] Furthermore, determining the advertising delivery strategy for the target object according to the recommendation associated tag includes:

[0105] A distribution channel for the target product is determined according to the recommended user tag.

[0106] Specifically, the connection between distribution channels and user tags is relatively close. Generally speaking, a specific distribution channel is associated with users with specific user tags. For example, there are more users aged 15-35 on website A. If the user tag corresponding to a certain insurance product is "age 20-30 years old", website A can be selected as the distribution channel for the insurance product.

[0107] Therefore, in an embodiment of the present invention, after the insurance product advertisement is first placed according to the preliminary placement strategy, the placed insurance product advertisement is tracked to obtain relevant data, and then the preliminary placement strategy is modified using these data to obtain a revised advertising placement strategy, so that the insurance product advertisement placement is more accurate and appropriate.

[0108] Furthermore, before determining the first candidate associated tag of the target object according to the correspondence between the object tag and the associated tag of the target object, the method further includes:

[0109] Determining a preliminary associated tag of the historical object according to the feature data of the historical object;

[0110] Obtaining user behavior data during the delivery of the historical object;

[0111] According to the user behavior data, the preliminary association tag of the historical object is modified to obtain the actual association tag of the historical object.

[0112] The preliminary association tags of the historical objects in the embodiment of the present invention are determined according to the characteristics of the historical objects and the characteristics of the channels. After the historical objects are delivered, the user behavior events of the delivered historical objects are processed, and the user behavior data for the historical objects is reported in real time. Then, the data analysis system recalculates and counts the data every hour to achieve the purpose of real-time monitoring.

[0113] Specifically, the front-end tracking technology is to call the tracking SDK function by manually writing code, call the interface at the business logic function location where tracking is required, and report the tracking data. The historical object, that is, the advertising of insurance products, mainly focuses on two types of indicators: data monitoring and performance monitoring.

[0114] Data monitoring is to monitor user information and behavior. Common data monitoring items include: PV (page view), which refers to the number of page views or clicks; UV (unique visitor), which refers to the number of people with different IP addresses who visit a site or click on a piece of news; the time users stay on each page, etc.

[0115] Performance monitoring refers to monitoring the performance of the front-end, mainly including monitoring the user experience of web pages or products. Common performance monitoring items include: first screen loading time under different users, different models and different systems; response time of http waiting requests; overall download time of static resources; page rendering time, etc.

[0116] In the embodiment of the present invention, the following three correspondences can be generated based on the user behavior data of the historical object, which are used as the basis for revising the initial delivery strategy. The correspondences include: 1) product-user tag, 2) product-channel-delivery time, and 3) product-material AB test results.

[0117] Among them, 1) product-user tag matching

[0118] The advertising delivery system can obtain all user tags and the user information behind them from the user portrait library. In the process of formulating the advertising delivery strategy for insurance products, a 1-to-N relationship will be formed between an insurance product and different user tags. This correspondence will count the relationship between product-user tag-conversion rate, and perform real-time sorting based on the CR (critical ratio) value of each user tag, and gradually select the optimal user tag that can maximize the CR value for the product.

[0119] 2) Product-channel-launch time.

[0120] Since product characteristics and channel characteristics are different, and the acceptance of channel customers for such products is unknown, horizontal comparison of the matching degree between products and channels becomes an important reference indicator for subsequent delivery. At the same time, the time of advertising is an important indicator that affects the effectiveness of advertising. The wrong time of delivery will greatly reduce attention and lower the conversion rate. This correspondence will count the CR values ​​of products in different channels and at different delivery times, and sort them. The matching and sorting relationship of product-product label-channel-delivery time is formed, which is used as the basis for subsequent modification of advertising delivery strategies.

[0121] 3) Product-Material AB Testing

[0122] Since advertising conversion is a process of attracting customers' attention in a short period of time and successfully converting them, materials (i.e. advertising pages, wording, etc.) are important facades and tools for improving conversion. Usually, AB testing is used to determine which advertising interface is more attractive and more effective in converting customers.

[0123] That is to say, in the embodiment of the present invention, by collecting data of the historical objects that have been placed, the corresponding relationship of "product-user tag" and the corresponding relationship of "product-channel" of the historical objects are generated. Therefore, the advertising placement strategy of the historical objects is modified, and the preliminary associated tags therein are modified to obtain the actual associated tags.

[0124] Furthermore, in the embodiment of the present invention, the candidate association tags of the target object include, in addition to the first candidate association tags initially selected according to the characteristics of the target object and the channel characteristics, second candidate association tags, which are determined from actual association tags of historical objects with greater similarity to the target object.

[0125] Among them, the similarity between the target object and the historical object is determined according to the following formula:

[0126]

[0127] Among them, Ru,i represents the score given by user u to target object i, R u,j Represents the score given by user u to historical object j. represents the average score of target object i, It represents the average score of historical object j. sim(i,j) is the similarity between the target object and the historical object.

[0128] The average value of user ratings. Some users may tend to give high scores, while others may tend to give low scores. This method eliminates the influence of user rating habits by subtracting the average value of user ratings. The closer the value of sim(i,j) is to 1, the higher the similarity between the target object and the historical object.

[0129] After similar objects of the target object are determined according to the similarity, second candidate associated tags are determined from actual associated tags of the similar objects.

[0130] Preferably, the selecting the actual associated label of the similar object as the second candidate associated label of the target object comprises:

[0131] Calculating the matching degree between each actual associated label of the similar object and the target object;

[0132] selecting the second candidate associated tag from actual associated tags of the similar objects according to the matching degree;

[0133] The degree of matching is calculated according to the following formula:

[0134]

[0135] Among them, S i,N is the similarity between the target object i and the similar object N, R u,N is the score given by user u to similar object N.

[0136] The scores of the objects that user u has scored are weighted and summed, with the weight being the similarity between each similar object and the target object, and then the sum of all similarities is averaged to calculate the score of user u on the target object i. The higher the score, the higher the matching degree of the user to the target object i (candidate association label).

[0137] Specifically, all actual associated tags of similar objects can be used as the second candidate associated tags of the target object, or a part of all actual associated tags of similar objects can be selected as the second candidate associated tags of the target object, and the selection method is to calculate the matching degree between each actual associated tag of the similar objects and the target object, so as to select the corresponding actual associated tag as the second candidate associated tag according to the matching degree. In addition, the embodiment of the present invention does not limit the number of similar objects corresponding to the target object, that is, the similar object can be one or more.

[0138] For example, if Figure 3 As shown in the figure, product C is the target object, and the historical objects are product A and product B. Formula 1 is used to calculate the similarity between product C and product A, and the similarity between product C and product B. According to the calculation results, it is determined that the similarity between product A and product C meets the conditions, and product A is used as the similar object of product C. Since the product labels of product A are 1, 2, and 3, and the corresponding user labels are a, b, c, d, e, and g, respectively, user labels a, b, c, d, e, and g are used as the second candidate association labels of product C. In addition, since the first candidate association labels of product C are user labels b, c, e, and g, the candidate association labels of product C are a, b, c, d, e, and g.

[0139] After obtaining the candidate associated tags of the target object, the click prediction model is used to output the predicted click probability corresponding to each candidate associated tag, that is, the candidate associated tags of the target object are input into the click prediction model, and the predicted click probability of each candidate associated tag is calculated by the click prediction model. Therefore, the recommended associated tags can be selected from all candidate associated tags according to the predicted click probability. The association relationship between the target object and the recommended associated tags is established, that is, the association relationship between the target product and the user tag is established, and the channel for placing the target product is determined according to the user tag.

[0140] Preferably, in the embodiment of the present invention, the click prediction model is trained according to the following method:

[0141] Acquire a training sample, wherein the training sample includes feature data of a training object, feature data of a training user, and actual click results of the training user on the training object;

[0142] Inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object;

[0143] Determining the predicted click probability of the training object according to the predicted click results of all training users for the training object;

[0144] Determining an actual click rate of the training object according to actual click results of the training user on the training object;

[0145] A loss function is calculated according to the actual click rate and the predicted click probability, and when the loss function is less than a preset threshold, parameters corresponding to the click prediction model are determined.

[0146] Among them, the click prediction models include GBDT (Gradient Boosting Decision Tree) model and LR (Logistic, logistic regression) model.

[0147] The step of inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object includes:

[0148] For each training user, input the feature data of the training user and the feature data of the training object into the GBDT model to construct multiple decision trees;

[0149] Using the multiple decision trees, obtaining predicted click results of the training user on the training object;

[0150] Determine the combined feature value of the GBDT model according to the predicted click results of all decision trees;

[0151] The step of determining the predicted click probability of the training object according to the predicted click results of all training users for the training object includes:

[0152] The combined feature values ​​of the GBDT model are input into the LR model to obtain the predicted click probability of the training object.

[0153] Using the GBDT model alone is prone to overfitting. Therefore, in the embodiment of the present invention, the GBDT model is combined with the LR model for training. The training process of the GBDT model and the LR model is specifically introduced below.

[0154] GBDT is an iterative decision tree algorithm, which consists of several decision trees. The gradient boosting tree model is a model that combines decision trees with boosting methods (such as XGBoost). Its core idea is that in the process of building a series of decision trees, the next decision tree learns the conclusions and residuals of all previous decision trees. When the residual of the next decision tree is less than the set threshold or reaches the number of iterations, the model training is terminated. By continuously fitting the residuals of the previous decision trees, several decision trees are finally obtained. For the sample to be tested, the output result of the gradient boosting tree model is the sum of the output results of k decision trees. In this process, the selection of samples is random, and the selection of features is also random, which means that some samples in the total training set may appear in the training set of a tree multiple times, or may never appear in the training set of any tree. The main steps are as follows: Apply resampling technology to randomly extract n sample sets from m model training samples, and construct n decision trees from them. During the growth of each decision tree, each node randomly selects F features from all features as the subset of the current node split. When constructing a decision tree, the minimum mean square error is usually used as the splitting criterion to select the best splitting method. The n decision trees are combined into the final GBDT model.

[0155] In the embodiment of the present invention, the corresponding user can be determined for the target product, and then the advertisement can be pushed to the determined user; the corresponding target product can also be determined for a specific user. Here, taking the determination of the corresponding user for the target product as an example, the training process of the click probability model is introduced.

[0156] First, obtain training samples. The training sample format in the embodiment of the present invention is {(X1, Y1), (X2, Y2), …, (X n ,Y n )}, where X n is the multidimensional feature data of the nth training object. For example, X1 is the multidimensional feature data of the first training object, and X2 is the multidimensional feature data of the second training object. The form of multidimensional feature data is (x n1 , x n2 , x n3 , …, x ni ), where i is the i-th feature of the n-th training object. For example, x n1 represents the age of the user corresponding to the training object, x n2 Can represent the gender of the user corresponding to the training object. n Indicates whether the user recommends the nth training object. If so, Y n is 1; if not, then Y n is 0.

[0157] Input the above training samples into the GBDT model to generate n decision trees and form a regression tree f(x).

[0158] During the training process, the above training samples are input into the GBDT model. Specifically, in the input space where the training data set is located, each region is recursively divided into two sub-regions and the output value on each sub-region is determined to construct a two-branch decision tree;

[0159] (1) Select the optimal segmentation variable j and segmentation point s, and solve the following formula:

[0160]

[0161] (2) Traverse the variable j, scan the split point s for the fixed split variable j, and select the value (j, s) that minimizes Formula 3.

[0162] Divide the region by the selected (j, s) and determine the corresponding output value according to the following formula:

[0163] R1(j,s)={x|x (j) ≤s},R2(j,s)={x|x (j) ≤s}

[0164]

[0165] (3) Continue to call steps (1) and (2) for the two sub-regions until the stopping condition is met;

[0166] (4) Divide the input space into M regions R1, R2, …, R M , generate the first decision tree:

[0167]

[0168] (5) Generate the second decision tree. Construct the training sample data of the second decision tree, and use the residual to represent the target variable of each training sample. The details are as follows. The CART (Classification And Regression Tree) algorithm is also used to generate the second decision tree. The second decision tree is defined as DT2, DT2(X i ) represents the second decision tree for the training sample X i The predicted value of the second decision tree is input as follows:

[0169] {(X1,Y1-DT1(X1)),(X2,Y2-DT1(X2)),…,(X n ,Y n -DT1(X n ))}

[0170] ...Formula 6

[0171] Initialize the weak learner:

[0172]

[0173] For the number of iterations m = 1, 2, ..., M, we have:

[0174] For each training sample i = 1, 2, ..., N, the negative gradient, i.e., the residual, is calculated according to the following formula:

[0175]

[0176] The obtained residual is used as the new true value of the sample, and the data (x i ,r im )(i=1,2,…,N) is used as the training data for the next decision tree to obtain a new regression tree f m (x). Its corresponding leaf node area is R jm ,(j=1,2,…,J). Where J is the number of leaf nodes of regression tree t.

[0177] The best fit value is calculated for the leaf region (j=1,2,…,J) according to the following formula:

[0178]

[0179] Update the strong learner according to the following formula:

[0180]

[0181] The strong learner is obtained according to the following formula:

[0182]

[0183] (6) And so on, the mth decision tree is generated. The training sample data for constructing the mth decision tree is as follows:

[0184] {(X1,DT m-2 (X1)-DT m-1 (X1)),(X2,DT m-2 (X2)-DT m-1 (X2)),…,(X n ,DT m-2 (X1)-DT m-1 (X n ))}.

[0185] In this way, the CART algorithm is used to generate m decision trees, where the mth decision tree is defined as DT m , DTm (X i ) represents the mth decision tree for the training sample X i The predicted click result of each decision tree is added up to get the predicted value of the GBDT model, which is in the form of (0110111…). Among them, 1 means that the predicted click result is recommended, and 0 means not recommended.

[0186] The leaf nodes of the decision tree of the training sample in the GBDT model are taken as output and input into the LR model to obtain the predicted click probability of each user label for the training sample.

[0187] Finally, the loss function is calculated based on the actual click rate of the user label for the training sample and the above predicted click probability. When the loss function is less than the preset threshold, the parameters corresponding to the click prediction model are determined, such as the split point of each decision tree in the regression tree. In the embodiment of the present invention, the parameters corresponding to the click prediction model change according to the change of the training sample, and as the number of training samples input into the click prediction model increases, the parameters of the click prediction model are more appropriate, and the prediction results of the click prediction model are more accurate.

[0188] Still Figure 3 For example, the candidate user labels a, b, c, d, e, and g of product C are input into the above click prediction model, and the predicted click probabilities are ranked as c, b, e, d, a, and g. Therefore, the corresponding channels can be selected for advertising delivery according to the above ranking.

[0189] In addition, the embodiments of the present invention can also deliver advertisements from the user dimension, that is, determine the corresponding target products for specific users. The user dimension is similar to the product dimension in principle. The basic idea is that if in the historical delivery, user A matches the product label a, user B matches the product labels a, b, c, and user C matches the product labels a and c, then user A is considered similar to users B and C because they all match a. Users A, B, and C will have a common user label for a in the engine, such as "love outdoor sports". The user who matches a also matches c, so c is recommended to user A, and this c is the product label "sports accident insurance". Since the user labels are very large, it will be very time-consuming and labor-intensive to calculate the entire set, which is not efficient enough. Therefore, instead of directly calculating the global set, the user labels are first filtered according to the intention and goal. Not only that, the product labels can also be limited synchronously to further narrow the scope of the content library. According to the needs of the delivery or the characteristics of the insured group that the insurance product is targeting (input during product management), the content is truncated, and a small part of the content that meets the target is efficiently screened out from a large content library, turning a massive and uncontrollable content library into a relatively small and controllable content library, and then entering the recommendation model. This effectively balances the computing cost and effect. For example, this delivery is mainly aimed at female users between 30 and 40 years old with children. Then the user tags that can be determined will be: "around 30 years old", "under 40 years old", "mom", "female", "health-conscious". After narrowing down the determined user tags, the intelligent engine calculates and outputs the matching product tags (that is, the possible product preferences of this type of user), delivery channels and recommended delivery ratios, and the best delivery time for the channel for users under this tag (historical conversion rate is high). If there is further need, the product tags can be selected twice, such as "women's health insurance", "children's accident insurance", etc., to make more detailed calculations on the recommendation results. The final conclusion may be that customers with this type of label are more willing to invest in their children than to pay attention to their own health (the engine calculates that for users with this type of label, children's insurance products have higher scores and are ranked higher). This type of user is more active in a certain channel and is more likely to go online in the channel after 20:00 in the evening, see promotional ads and convert.

[0190] The embodiment of the present invention also provides a device for recommending advertisement delivery, such as Figure 4 As shown, including:

[0191] A determination unit 401 is used to determine a first candidate associated tag of the target object according to a correspondence between an object tag and an associated tag of the target object;

[0192] A selection unit 402 is used to determine the similarity between the target object and the historical object, and determine the similar object of the target object from all the historical objects according to the similarity; and select the actual associated tag of the similar object as the second candidate associated tag of the target object;

[0193] A calculation unit 403 is used for inputting the candidate association tag and the feature data of the target object into a click prediction model for each candidate association tag of the target object to obtain a predicted click probability of the candidate association tag for the target object; the click prediction model is trained according to the feature data of the training sample and the actual click result to obtain corresponding model parameters;

[0194] An association unit 404, configured to determine a recommended associated tag from all candidate associated tags according to the predicted click probability;

[0195] The delivery unit 405 is used to determine the advertisement delivery strategy for the target object according to the recommendation associated tag.

[0196] In an optional embodiment, the selection unit 402 is further configured to:

[0197] Determining a preliminary associated tag of the historical object according to the feature data of the historical object;

[0198] Obtaining user behavior data during the delivery of the historical object;

[0199] According to the user behavior data, the preliminary association tag of the historical object is modified to obtain the actual association tag of the historical object.

[0200] In an optional embodiment, a training unit 406 is further included, which is used to train the click prediction model according to the following method:

[0201] Acquire a training sample, wherein the training sample includes feature data of a training object, feature data of a training user, and actual click results of the training user on the training object;

[0202] Inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object;

[0203] Determining the predicted click probability of the training object according to the predicted click results of all training users for the training object;

[0204] Determining an actual click rate of the training object according to actual click results of the training user on the training object;

[0205] A loss function is calculated according to the actual click rate and the predicted click probability, and when the loss function is less than a preset threshold, parameters corresponding to the click prediction model are determined.

[0206] In an optional embodiment, the click prediction model includes a GBDT (gradient boosted iterative decision tree) model and an LR (logistic regression) model;

[0207] The training unit 406 is further used for:

[0208] For each training user, input the feature data of the training user and the feature data of the training object into the GBDT model to construct multiple decision trees;

[0209] Using the multiple decision trees, obtaining predicted click results of the training user on the training object;

[0210] Determine the combined feature value of the GBDT model according to the predicted click results of all decision trees;

[0211] The combined feature values ​​of the GBDT model are input into the LR model to obtain the predicted click probability of the training object.

[0212] In an optional embodiment, the selection unit 402 is specifically configured to determine the similarity between the target object and the historical object according to the following formula:

[0213]

[0214] Among them, R u,i represents the score given by user u to target object i, R u,j Represents the score given by user u to historical object j. represents the average score of target object i, It represents the average score of historical object j. sim(i,j) is the similarity between the target object and the historical object.

[0215] In an optional embodiment, the selection unit 402 is specifically configured to:

[0216] Calculating the matching degree between each actual associated label of the similar object and the target object;

[0217] selecting the second candidate associated tag from actual associated tags of the similar objects according to the matching degree;

[0218] The degree of matching is calculated according to the following formula:

[0219]

[0220] Among them, S i,Nis the similarity between the target object i and the similar object N, R u,N is the score given by user u to similar object N.

[0221] Based on the same principle, the present invention also provides an electronic device, such as Figure 5 As shown, including:

[0222] The device comprises a processor 501, a memory 502, a transceiver 503, and a bus interface 504, wherein the processor 501, the memory 502 and the transceiver 503 are connected via the bus interface 504;

[0223] The processor 501 is configured to read the program in the memory 502 and execute the following method:

[0224] Determine a first candidate associated tag of the target object according to a correspondence between the object tag and the associated tag of the target object;

[0225] Determine the similarity between the target object and the historical object, and determine similar objects of the target object from all the historical objects according to the similarity;

[0226] Selecting the actual associated label of the similar object as the second candidate associated label of the target object;

[0227] For each candidate association tag of the target object, the candidate association tag and the feature data of the target object are input into a click prediction model to obtain the predicted click probability of the candidate association tag for the target object; the click prediction model is trained according to the feature data of the training sample and the actual click results to obtain corresponding model parameters;

[0228] Determine a recommended associated tag from all candidate associated tags according to the predicted click probability;

[0229] An advertisement delivery strategy for the target object is determined according to the recommendation associated tag.

[0230] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0231] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0232] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0233] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0234] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. An advertisement placement recommendation method, characterized in that: include: Determine a first candidate associated tag of the target object according to a correspondence between the object tag and the associated tag of the target object; The correspondence between the object tag and the associated tag is set according to the characteristics of the target object and the channel characteristics; Determine the similarity between the target object and the historical object, and determine similar objects of the target object from all the historical objects according to the similarity; Selecting the actual associated tag of the similar object as the second candidate associated tag of the target object; the actual associated tag is determined by performing point-of-view processing on the user behavior event of the delivered historical object; For each candidate association tag of the target object, the candidate association tag and the feature data of the target object are input into a click prediction model to obtain the predicted click probability of the candidate association tag for the target object; the click prediction model is trained according to the feature data of the training sample and the actual click results to obtain corresponding model parameters; Determine a recommended associated tag from all candidate associated tags according to the predicted click probability; An advertisement delivery strategy for the target object is determined according to the recommendation associated tag.

2. The method according to claim 1, characterized in that Before determining the first candidate associated tag of the target object according to the correspondence between the object tag and the associated tag of the target object, the method further includes: Determining a preliminary associated tag of the historical object according to the feature data of the historical object; Obtain user behavior data during the delivery of the historical object, According to the user behavior data, the preliminary association tag of the historical object is modified to obtain the actual association tag of the historical object.

3. The method according to claim 1, characterized in that The click prediction model is trained in the following manner: Acquire a training sample, wherein the training sample includes feature data of a training object, feature data of a training user, and actual click results of the training user on the training object; Inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object; Determining the predicted click probability of the training object according to the predicted click results of all training users for the training object; Determining an actual click rate of the training object according to actual click results of the training user on the training object; A loss function is calculated according to the actual click rate and the predicted click probability, and when the loss function is less than a preset threshold, parameters corresponding to the click prediction model are determined.

4. The method according to claim 3, characterized in that The click prediction model includes a GBDT gradient boosting iterative decision tree model and an LR logistic regression model; The step of inputting the characteristic data of the training object and the characteristic data of the training user into the click prediction model to obtain the predicted click result of the training user for the training object includes: For each training user, input the feature data of the training user and the feature data of the training object into the GBDT model to construct multiple decision trees; Using the multiple decision trees, obtaining predicted click results of the training user on the training object; Determine the combined feature value of the GBDT model according to the predicted click results of all decision trees; The step of determining the predicted click probability of the training object according to the predicted click results of all training users for the training object includes: The combined feature values ​​of the GBDT model are input into the LR model to obtain the predicted click probability of the training object.

5. The method according to any one of claims 1 to 4, characterized in that: The similarity between the target object and the historical object is determined according to the following formula: Among them, R u,i represents the score given by user u to target object i, R u,j represents the score given by user u to historical object j; represents the average score of target object i, represents the average score of historical object j; sim(i,j) is the similarity between the target object and the historical object.

6. The method according to claim 1, characterized in that The selecting the actual associated label of the similar object as the second candidate associated label of the target object comprises: Calculating the matching degree between each actual associated label of the similar object and the target object; selecting the second candidate associated tag from actual associated tags of the similar objects according to the matching degree; The degree of matching is calculated according to the following formula: Among them, S i,N is the similarity between the target object i and the similar object N, R u,N is the score given by user u to similar object N.

7. An advertisement placement recommendation device, characterized in that: include: A determination unit, configured to determine a first candidate associated tag of the target object according to a correspondence between an object tag and an associated tag of the target object; The correspondence between the object tag and the associated tag is set according to the characteristics of the target object and the channel characteristics; A selection unit is used to determine the similarity between the target object and the historical object, and determine the similar object of the target object from all the historical objects according to the similarity; select the actual associated tag of the similar object as the second candidate associated tag of the target object; the actual associated tag is determined by burying the user behavior event of the delivered historical object; A calculation unit is used to input the candidate association tag and the feature data of the target object into a click prediction model for each candidate association tag of the target object, so as to obtain the predicted click probability of the candidate association tag for the target object; the click prediction model is trained according to the feature data of the training sample and the actual click result to obtain the corresponding model parameters; An association unit, used to determine a recommended association tag from all candidate association tags according to the predicted click probability; The delivery unit is used to determine the advertisement delivery strategy for the target object according to the recommendation associated tag.

8. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute any one of claims 1 to 6.

Images