Advertisement recommendation method and apparatus

[0029] figure 1 An application environment diagram for the operation of the advertisement recommendation method in an embodiment, such as figure 1 As shown, the application environment includes a terminal 110 and a server 120, and the terminal 110 and the server 120 communicate through a network.

[0030] The terminal 110 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc., but is not limited thereto. The server 120 may respond to the request sent by the terminal 110 .

[0031] In one embodiment, figure 1 The internal structure of the server 120 in figure 2 As shown, the server 120 includes a processor, a storage medium, a memory, and a network interface connected through a system bus. Wherein, the storage medium of the server 120 stores an operating system, a database, and a device for advertising recommendation. The database is used to store data, such as storing advertising resources, user information, and user tags. 120 methods of advertising recommendations. The processor of the server 120 is used to provide computing and control capabilities to support the operation of the entire server 120 . The internal memory of the server 120 provides an environment for the operation of the advertisement recommendation device in the storage medium. The network interface of the server 120 is used to communicate with the external terminal 110 through a network connection, such as receiving a request sent by the terminal 110 and returning data to the terminal 110 .

[0032] like image 3 As shown, in one embodiment, a method for advertising recommendation is provided, which is applied to the server in the above application environment as an example, including the following steps:

[0033] Step S210 , receiving an advertisement recommendation request, and obtaining a user label corresponding to the user in the advertisement recommendation request corrected according to feedback information, the feedback information including positive feedback information and negative feedback information.

[0034] Specifically, when a user needs to open a web page with an advertisement space, the terminal will send an advertisement recommendation request to the server, and the server receives the advertisement recommendation request. The advertisement recommendation request contains user information, such as a user identifier used to uniquely identify a user, The user identifier may be a user name for logging in to the application, a terminal hardware identifier or an IP address. If the terminal sends an advertisement recommendation request through a social network application, the terminal will automatically obtain the currently logged-in user name, generate an advertisement recommendation request and send it to the server, and the server determines the user through the user name included in the advertisement recommendation request. For the advertisement recommendation request sent when the application is not logged in, the terminal hardware identifier or IP address may be used to identify the user, and the server determines the user through the hardware identifier or IP address included in the advertisement recommendation request.

[0035] Each user has a corresponding user label, which is used to mark the user's interest, behavior and other characteristics, which is obtained by analyzing the user's basic information and click behavior data. The user's basic information such as age, gender, region, occupation, etc. , click behavior data includes positive feedback information and negative feedback information. Positive feedback information refers to the information returned after the user clicks on the recommended advertisement. Positive feedback information includes the number of positive clicks, the user ID of the currently clicked recommended advertisement, and the location of the currently clicked recommended advertisement page, the ad tab of the currently clicked recommended ad, etc. Negative feedback information refers to negative emotion-related information fed back by users, such as clicking on buttons that reflect negative emotions, such as buttons that are not interested in advertisements, blocking buttons, or buttons that have nothing to do with me, including the number of negative clicks and the user ID of the current button that clicks on negative feedback. , the page where the currently clicked negative feedback button is located, the advertisement label of the recommended advertisement to which the currently clicked negative feedback button belongs, etc. The user tag corresponding to the user will be continuously revised according to the feedback information, so that the user's needs can be judged more accurately, and more accurate candidate recommended advertisements can be screened out for subsequent use of the corrected user tag.

[0036] In step S220, candidate recommended advertisements are screened out according to the corrected user tags.

[0037] Specifically, advertising resources are stored in the advertisement pool, and each advertisement has a corresponding advertisement label, which is used to describe the category or delivery range of the advertisement. For example, the label of a child seat advertisement is "less than 5 years old", and the label of a dress advertisement is The labels are "girls" and "clothing", and the advertisements that match the user labels and advertisement labels are selected to obtain candidate recommended advertisements.

[0038] Step S230, calculate the user's estimated click probability for each candidate recommended advertisement on the current display page according to the estimated click model, and calculate the user's estimated non-interest probability for each candidate recommended advertisement on the current display page according to the estimated non-interest model .

[0039]Specifically, the estimated click model is used to calculate the probability that the user may click on the recommended advertisement. Substituting the characteristics of the current display page where the candidate recommended advertisement is located, the user characteristics, and the characteristics of the candidate recommended advertisement into the estimated click model can calculate the user's preference for the candidate recommended advertisement. The estimated click probability on the currently displayed page. A higher probability indicates that the user is more likely to click on this candidate recommended advertisement. The estimated non-interest model is used to calculate the probability that the user is not interested in the recommended advertisement. Substituting the characteristics of the current display page where the candidate recommended advertisement is located, the user characteristics, and the characteristics of the candidate recommended advertisement into the estimated non-interest model can calculate the user's interest in the candidate advertisement. The probability that the recommended advertisement is not interested in the currently displayed page. The higher the probability, the more likely the user is not interested in the candidate recommended advertisement. Among them, the characteristics of the currently displayed page refer to the characteristics of the page where the recommended advertisement is located, such as "women's shopping page", etc., and the user characteristics refer to the characteristics related to the user. User tags can be directly used to express user characteristics, such as "particularly interested in clothing". Not interested in toys", etc. The characteristics of the candidate recommended advertisement are used to describe the characteristics of the advertisement itself, such as "clothes" and "toys", etc., and the advertisement label can be directly used to represent the characteristics of the candidate recommended advertisement.

[0040] Step S240, according to the estimated click probability and the estimated non-interest probability, select the target recommended advertisement from candidate recommended advertisements.

[0041] Specifically, the estimated click probability and the estimated non-interest probability can be weighted or filtered by a custom algorithm to obtain the target recommended advertisement from candidate recommended advertisements. Since not only the estimated click probability but also the estimated non-interest probability are taken into consideration during screening, the recommended advertisements are more in line with user needs and the accuracy of advertisement recommendation is improved.

[0042] In this embodiment, by receiving the advertisement recommendation request, the corrected user label corresponding to the user in the advertisement recommendation request is obtained according to the positive feedback information and the negative feedback information, and the candidate recommended advertisement is screened out according to the corrected user label, and according to the estimated The click model calculates the user's estimated click probability for each candidate recommended advertisement on the current display page, and calculates the user's estimated non-interest probability for each candidate recommended advertisement on the current display page according to the estimated non-interest model, and according to the estimated click probability In the process of recommending advertisements, not only the positive feedback information but also the negative feedback information are taken into consideration. After using the corrected positive feedback information and negative feedback information, Candidate recommended advertisements are obtained from user tags, which improves the accuracy of candidate recommended advertisements. At the same time, when calculating the user's estimated click probability and estimated non-interest probability for candidate recommended advertisements, the current display page is comprehensively considered, so that the estimated click probability The accuracy of the estimated click probability and estimated non-interest probability is higher, and the estimated click probability and estimated non-interest probability can make the screening more accurate and further improve the accuracy of advertisement recommendation.

[0043] In one embodiment, such as Figure 4 As shown, before step S210, it also includes:

[0044] Step S310, obtaining feedback information returned by the user on historically recommended advertisements.

[0045] Specifically, the feedback information returned by the historically recommended advertisement user, including positive feedback information and negative feedback information, will be stored corresponding to the user and used to correct the user label. The user label can be corrected in real time according to the feedback information, or can be corrected periodically.

[0046] Step S320, counting the number of positive clicks by the user on each advertisement label according to the positive feedback information, and counting the number of negative clicks by the user on each advertisement label according to the negative feedback information.

[0047] Specifically, each recommended advertisement has a corresponding advertisement label, and multiple advertisements may have the same advertisement label, and the number of positive clicks and negative clicks on each advertisement label by the user is counted. If the number of positive clicks on an advertisement label is large, It indicates that the user is interested in the advertisement corresponding to the advertisement label, and if the number of negative clicks on an advertisement label is large, it indicates that the user is not interested in the advertisement corresponding to the advertisement label. In one embodiment, the user's positive click probability for each advertisement label is counted according to the positive feedback information, and the user's negative click probability for each advertisement label is counted according to the negative feedback information. The positive click probability is clicked by all recommended advertisements corresponding to each advertisement label The number of times is divided by the number of occurrences of all recommended ads corresponding to each ad label. The negative click probability is obtained by dividing the number of negative clicks of all recommended advertisements corresponding to each advertising label by the number of occurrences of all recommended advertisements corresponding to each advertising label. The number of negative clicks refers to the number of buttons clicked to respond to negative information feedback.

[0048] Step S330 , generate corresponding interested ad user tags for the ad tags whose number of positive clicks exceeds the preset threshold, and generate corresponding non-interested ad user tags for the ad tags whose negative clicks exceed the preset threshold.

[0049] Specifically, the preset threshold can be customized according to needs. For example, if the number of positive clicks clicked by users on the advertisement label "Children" exceeds the preset threshold, "Children" can be used as a user label of the user, such as the advertisement label "Children". If the number of negative clicks clicked by the user on the "category" exceeds the preset threshold, the "category of uninterested children" can be used as a user label of the user. In one embodiment, an advertisement tag whose positive click probability exceeds a preset threshold is generated into a corresponding interested advertisement user tag, and an advertisement tag whose negative click probability exceeds a preset threshold is generated into a corresponding non-interested advertisement user tag.

[0050] In this embodiment, both positive feedback information and negative and positive feedback information are taken into account to modify the user label, so that the user's note can change according to the user's behavior, and the accuracy of the user label is improved.

[0051] In one embodiment, such as Figure 5 As shown, before step S210, it also includes:

[0052] Step S340, obtaining feedback information returned by different users on historically recommended advertisements.

[0053] Specifically, the feedback information returned by different users for historical recommended advertisements, including positive feedback information and negative feedback information, will be stored corresponding to the users.

[0054] Step S350, obtain the first user characteristics, first display page characteristics, and first historical recommended advertisement characteristics corresponding to different users corresponding to the positive feedback information, and combine the first user characteristics, first display page characteristics, and first historical recommended advertisement characteristics The corresponding combination generates the first feature vector.

[0055] Specifically, the user feature vector u corresponding to different users corresponding to the positive feedback information, the first display page feature vector d, and the first historical recommended advertisement feature vector a are combined into a feature vector x T , User characteristics may include multiple values, such as "female" and "children", each corresponding to a numerical value, and the combination of these numerical values ​​generates the first user characteristic vector u. The features of the first display page also include multiple features, such as "shopping page" and "girls' clothing page", which correspond to a value respectively. The combination of these values ​​generates the feature vector d of the first display page, and the feature vector of the first historical recommended advertisement also includes multiple , such as "children", "over 3 years old", etc., respectively correspond to a value, and these values ​​are combined to generate the first historical recommended advertisement feature vector a, which is a corresponding relationship when generating each feature vector, that is, the current first historical recommended advertisement The feature of the first display page where the current first historical recommended advertisement is located corresponds to the user feature of the user whose current first historical recommended advertisement is clicked.

[0056] Step S360, obtaining the second user characteristics, second display page characteristics, and second historical recommended advertisement characteristics corresponding to different users corresponding to the negative feedback information, and combining the second user characteristics, second display page characteristics, and second historical recommended advertisement characteristics The corresponding combination generates a second feature vector.

[0057] Specifically, the user feature vector u1 corresponding to different users corresponding to the negative feedback information, the second display page feature vector d1, and the second historical recommended advertisement feature vector a1 are combined into a feature vector x 1 T , User characteristics may include multiple values, such as "female" and "children", each corresponding to a numerical value, and the combination of these numerical values ​​generates the second user characteristic vector u1. The second display page feature also includes multiple features, such as "shopping page", "girl clothing page", etc., each corresponding to a numerical value, and the combination of these values ​​generates the second display page feature vector d1, and the second historical recommended advertisement feature vector also includes multiple Each, such as "children", "over 3 years old", etc., respectively correspond to a value, these values ​​are combined to generate the second historical recommended advertisement feature vector a1, which is a corresponding relationship when generating each feature vector, that is, the current second historical recommended advertisement The features of the second display page where the current second historical recommended advertisement is located correspond to the user features of the clicked user.

[0058] Step S370: Calculate the positive click probability of historically recommended advertisements when different users correspond to the first feature vector according to the positive feedback information, and calculate the negative click probability of historically recommended advertisements when different users correspond to the second feature vector according to the negative feedback information.

[0059] Specifically, the positive feedback information includes the number of positive clicks, the user ID of the currently clicked recommended advertisement, the page where the currently clicked recommended advertisement is located, the advertisement label of the currently clicked recommended advertisement, etc., and the statistics of each first recommended advertisement corresponding to the first feature vector The number of clicks is compared with the total number of times each recommended advertisement appears on the first display page corresponding to the first feature vector to obtain the positive click probability of the historical recommended advertisements corresponding to the first feature vector by P(y=1|x T ), where y represents the user’s positive click behavior y ∈ {0, 1}, where 0 means no click and 1 means click. Negative feedback information includes the number of negative clicks, the user ID of the currently clicked negative feedback button, the page where the currently clicked negative feedback button is located, the ad label of the recommended advertisement to which the currently clicked negative feedback button belongs, etc., and the statistics of each second recommended advertisement corresponding to the second The number of negative clicks when the feature vector is compared with the total number of times of each second recommended advertisement on the second display page corresponding to the second feature vector, the negative click probability of each historical recommended advertisement on each display page is obtained by where y1 represents the user’s negative click behavior y1 ∈ {0,1}, where 0 means no click and 1 means click.

[0060] In step S380, an estimated click model is obtained by solving according to the positive click probability and the first eigenvector, and an estimated non-interest model is obtained by solving according to the negative click probability and the second eigenvector.

[0061] Specifically, the logistic regression model is used to represent the predicted click model and the predicted non-interest model, and P(y=1|x T ) and x T Substitute into the formula P(y=1|x T ) = exp(x T w)/(1+exp(x T w)), according to the user's historical positive click behavior records Use gradient stochastic descent (SGD) to solve the model parameter w, and get the estimated click model. For a new user ad recommendation request, pass p 0 =exp(x 0 T w)/(1+exp(x 0 T w)) Predict the probability p of the user's positive click behavior in this request 0 ,in It is the user feature vector of the target user corresponding to the candidate recommended advertisement for which the estimated click probability needs to be calculated, the feature vector formed by displaying the page feature vector, and the candidate recommended advertisement feature vector. P(y1=1|x 1 T ) and x 1 T Substitute into the formula P(y1=1|x 1 T ) = exp(x 1 T w1)/(1+exp(x 1 T w1)), according to the user's historical negative click behavior records Use gradient stochastic descent (SGD) to solve the model parameter w1, and get the estimated non-interest model. For new user advertisement recommendation requests, pass p 1 =exp(x 0 T w1)/(1+exp(x 0 T w1)) predicts the probability p of the user’s negative click behavior in this request 1.

[0062] In one embodiment, such as Image 6 As shown, step S240 includes:

[0063] Step S241, weighting the estimated click probability and estimated non-interest probability corresponding to each candidate recommended advertisement to obtain a quality score.

[0064] Specifically, the weighting coefficient can be customized according to the needs. The weighting coefficient of the estimated click probability is a positive value, and the weighting coefficient of the estimated non-interest probability is a negative value. For example, through the formula q=(α*p 0 +β*p 1 ) to get the quality score, where α is the weighting coefficient of the estimated click probability, α>0, p 0 is the estimated click probability, β is the weighting coefficient of the estimated non-interest probability, β<0, p 1 is the estimated probability of non-interest.

[0065] Step S242, screening candidate recommended advertisements according to quality scores to obtain target recommended advertisements.

[0066] Specifically, according to the quality score, the candidate recommended advertisements within the preset score range are selected as the target recommended advertisements, or the candidate recommended advertisements are sorted according to the quality scores, and the preset number of candidate recommended advertisements are selected according to the sort order as the target recommended advertisements . In one embodiment, the preset advertisement placement information of the merchant is acquired, and the candidate recommended advertisements are screened according to the quality score and the advertisement placement information to obtain the target recommended advertisements. The merchant's preset advertisement delivery information is the information related to the advertisement delivery of the merchant, such as the bid for the advertiser, the popularity of the merchant, etc., according to the specific amount of the advertiser's bid or the level of popularity of the custom merchant, by a custom algorithm Get the final score of each candidate recommendation advertisement, and filter the candidate recommendation advertisements within the preset score range according to the score as the target recommendation advertisement, or sort each candidate recommendation advertisement according to the score, and filter the preset number of candidate recommendation advertisements according to the sorting order Ads serve as targeted recommendation ads. In one embodiment, the final score of each candidate recommended advertisement is obtained by the formula q1=q*bidprice. Wherein, q is the quality score calculated in step S241, and bidprice is the bid amount of the advertiser.

[0067] In one embodiment, such as Figure 7 As shown, step S240 includes:

[0068] Step S243, filtering candidate recommended advertisements whose estimated non-interest probability is greater than or equal to a preset threshold.

[0069] Specifically, the preset threshold can be customized according to conditions such as the number of advertisement slots, and filter candidate recommended advertisements whose estimated non-interest probability is greater than or equal to the preset threshold, so that advertisements that the user is not interested in can be quickly excluded.

[0070] In step S244, the target recommended advertisement is obtained by filtering the filtered candidate recommended advertisements according to the estimated click probability.

[0071] Specifically, the candidate recommended advertisements whose estimated click probability is greater than the preset threshold can be used as the target recommended advertisements, or the candidate recommended advertisements can be sorted according to the estimated click probability, and the preset number of candidate recommended advertisements can be selected according to the sorting order as the target recommended advertisements. advertise. In one embodiment, the merchant's preset advertisement delivery information is obtained, and the target recommended advertisement is obtained by screening the filtered candidate recommended advertisements according to the estimated click probability and the merchant's preset advertisement delivery information. The merchant's preset advertisement delivery information is the information related to the advertisement delivery of the merchant, such as the bid for the advertiser, the popularity of the merchant, etc., according to the specific amount of the advertiser's bid or the level of popularity of the custom merchant, by a custom algorithm Get the final score of each candidate recommended advertisement, filter the filtered candidate recommended advertisements within the preset score range according to the high and low score as the target recommended advertisement, or sort each candidate recommended advertisement according to the score, and filter the preset number according to the sorting order The filtered candidate recommended advertisements are used as target recommended advertisements. In one embodiment, the final score of each filtered candidate recommended advertisement is obtained by the formula q2=p0*bidprice. Among them, p0 is the estimated click probability of the filtered candidate recommended advertisement, and bidprice is the bid amount of the advertiser.

[0072] like Figure 8 As shown, it is a schematic diagram of a software architecture model implemented by an advertisement recommendation method in an embodiment. The user label module 420 continuously receives the positive feedback information and negative feedback information for recommended advertisements returned by the user corresponding to the terminal 410 sent by the terminal 410, and Correct user labels based on positive and negative feedback information. When the page display module 450 receives the advertisement recommendation request sent by the terminal 410, it sends the request to the advertisement delivery system 440, and the advertisement delivery system 440 screens out candidate recommended advertisements from the advertisement pool 430 according to the revised user tags, and passes the positive feedback information The estimated click model 470 and estimated non-interest model 460 established with the negative feedback information combine the estimated click probability and estimated non-interest probability calculated by the characteristics of the current display page content to obtain the target recommended advertisement, and place the target recommended advertisement In the advertisement area corresponding to the page display module 450, the target recommended advertisement receives the positive click or negative click of the terminal, and generates positive feedback information and negative feedback information for further correcting the user label and updating the estimated click model 470 and estimated non-interest Model 460.

[0073] In one embodiment, such as Figure 9 As shown, a device for advertisement recommendation is provided, including:

[0074] The obtaining module 510 is configured to receive an advertisement recommendation request, and obtain a user label corresponding to the user in the advertisement recommendation request corrected according to feedback information, the feedback information includes positive feedback information and negative feedback information.

[0075] Candidate recommended advertisement screening module 520, configured to screen out candidate recommended advertisements according to the corrected user tags.

[0076] The calculation module 530 is used to calculate the estimated click probability of the user for each candidate recommended advertisement on the current display page according to the estimated click model, and calculate the estimated non-interest rate of the user for each candidate recommended advertisement on the current displayed page according to the estimated non-interest model. probability of interest.

[0077] Target recommended advertisement screening module 540, configured to screen candidate recommended advertisements to obtain target recommended advertisements according to estimated click probability and estimated non-interest probability.

[0078] In one embodiment, such as Figure 10 As shown, the device also includes:

[0079] The user label correction module 550 is used to obtain the feedback information returned by the user to the historically recommended advertisements, count the number of positive clicks by the user on each advertisement label according to the positive feedback information, and count the number of negative clicks by the user on each advertisement label according to the negative feedback information, and the The advertisement tags whose number of positive clicks exceeds the preset threshold generate corresponding interested advertisement user tags, and the advertisement tags whose negative clicks exceed the preset threshold generate corresponding non-interested advertisement user tags.

[0080] In one embodiment, such as Figure 11 As shown, the device also includes:

[0081] The modeling module 560 is configured to obtain the feedback information returned by different users to historical recommended advertisements, obtain the first user characteristics, first display page characteristics, and first historical recommended advertisement characteristics corresponding to different users corresponding to the positive feedback information, and The first user feature, the first display page feature and the first historical recommended advertisement feature are correspondingly combined to generate a first feature vector, and the second user feature, the second display page feature, and the second user feature corresponding to different users corresponding to the negative feedback information are obtained. Two historical recommended advertisement features, and correspondingly combine the second user feature, the second display page feature and the second historical recommended advertisement feature to generate a second feature vector, and calculate the first feature corresponding to different users according to the positive feedback information The vector is the positive click probability of historically recommended advertisements, and the negative click probability of different users corresponding to the second eigenvector is calculated according to the negative feedback information, and is obtained by solving according to the positive click probability and the first eigenvector The estimated click model is solved according to the negative click probability and the second feature vector to obtain an estimated non-interest model.

[0082] In one embodiment, such as Figure 12 As shown, the target recommended advertisement screening module 540 includes:

[0083] The scoring unit 541 is configured to weight the estimated click probability and estimated non-interest probability corresponding to each candidate recommended advertisement to obtain a quality score.

[0084] The first screening unit 542 is configured to screen candidate recommended advertisements according to quality scores to obtain target recommended advertisements.

[0085] In one embodiment, such as Figure 13 As shown, the target recommended advertisement screening module 540 includes:

[0086] The filtering unit 543 is configured to filter candidate recommended advertisements whose estimated non-interest probability is greater than or equal to a preset threshold.

[0087] The second screening unit 544 is configured to screen the filtered candidate recommended advertisements to obtain target recommended advertisements according to the estimated click probability.

[0088] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the programs can be stored in a computer-readable storage medium, as described in the present invention. In an embodiment, the program may be stored in a storage medium of a computer system, and executed by at least one processor in the computer system, so as to implement the processes of the embodiments including the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM), and the like.

[0089] The technical features of the above-mentioned embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, should be considered as within the scope of this specification.

[0090] The above-mentioned embodiments only express several implementation modes of the present invention, and the descriptions thereof are relatively specific and detailed, but should not be construed as limiting the patent scope of the invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention should be based on the appended claims.