CN115330467A - Marketing advertisement click prediction method - Google Patents
Marketing advertisement click prediction method Download PDFInfo
- Publication number
- CN115330467A CN115330467A CN202211244719.0A CN202211244719A CN115330467A CN 115330467 A CN115330467 A CN 115330467A CN 202211244719 A CN202211244719 A CN 202211244719A CN 115330467 A CN115330467 A CN 115330467A
- Authority
- CN
- China
- Prior art keywords
- historical
- advertisement
- timestamp
- click rate
- strategy
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/02—Marketing; Price estimation or determination; Fundraising
- G06Q30/0241—Advertisements
- G06Q30/0242—Determining effectiveness of advertisements
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Business, Economics & Management (AREA)
- Mathematical Analysis (AREA)
- Development Economics (AREA)
- General Engineering & Computer Science (AREA)
- Strategic Management (AREA)
- Software Systems (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- Life Sciences & Earth Sciences (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Marketing (AREA)
- Entrepreneurship & Innovation (AREA)
- Economics (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Algebra (AREA)
- Game Theory and Decision Science (AREA)
- Databases & Information Systems (AREA)
- General Business, Economics & Management (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to the technical field of data processing, in particular to a method for predicting marketing advertisement clicks. The method comprises the following steps: acquiring a comprehensive characteristic vector, historical delivery strategy sequences corresponding to advertisements and historical click rate sequences corresponding to the historical delivery strategy sequences; obtaining a delivery probability vector of the corresponding advertisement at each timestamp in a preset time period according to each historical delivery strategy sequence; obtaining an overall click rate vector of the corresponding advertisement in a preset time period according to the historical click rate sequence corresponding to each historical release strategy sequence; and training the advertisement click rate prediction network according to the comprehensive characteristic vector, each historical putting strategy sequence, the corresponding historical click rate sequence, the putting probability vector and the total click rate vector to obtain the trained advertisement click rate prediction network, and predicting the predicted click rate sequence of the planned putting strategy sequence corresponding to the comprehensive characteristic vector to be predicted. The invention improves the accuracy of the network for predicting the click rate of the advertisement.
Description
Technical Field
The invention relates to the technical field of data processing, in particular to a method for predicting marketing advertisement clicks.
Background
With the rapid development of the domestic internet, more and more services provide services to users in a network form; the advertisement is used as an important income means in the media industry, and the market scale of the advertisement keeps growing at a high speed along with the development of the internet. Unlike traditional advertising, internet advertising is not fixed in time and location, and is targeted to different users; therefore, it is necessary to reasonably distribute the advertisement according to the user information, the category of the advertisement itself, the advertisement delivery environment, and the like, so as to predict the profit of the advertisement.
The most important way to predict the advertisement profit is to predict the advertisement click-through rate and then predict the profit of the advertisement which is not delivered. The artificial neural network model is an effective means for predicting the click rate of the advertisement, but the existing neural network algorithm does not consider the relation between the memory of a user and the advertisement putting strategy, only takes a timestamp as the characteristic of data to directly train the network, and cannot enable the network to learn the inherent influence of the putting strategy on the click rate, so that the accuracy of the prediction result of the network is low, and the accuracy of the advertisement click rate prediction is low.
Disclosure of Invention
In order to solve the problem of low accuracy of advertisement click rate prediction in the prior art, the invention aims to provide a method for predicting marketing advertisement click rate, which adopts the following technical scheme:
the invention provides a method for predicting marketing advertisement clicks, which comprises the following steps:
acquiring a comprehensive characteristic vector, each historical advertisement putting strategy sequence corresponding to the comprehensive characteristic vector and a historical click rate sequence corresponding to each historical advertisement putting strategy sequence; the comprehensive characteristic vector comprises a user characteristic vector, an advertisement characteristic vector and a delivery environment characteristic vector; the historical delivery strategy sequence comprises the advertisement delivery amount of the corresponding advertisement at each target timestamp in a preset time period;
according to the historical delivery strategy sequences, delivery probability vectors of the advertisements corresponding to the comprehensive characteristic vectors at the time stamps in a preset time period are obtained; obtaining a total click rate vector of the advertisement corresponding to the comprehensive characteristic vector in a preset time period according to the historical click rate sequence corresponding to each historical release strategy sequence;
training an advertisement click rate prediction network according to the comprehensive characteristic vector, the historical click rate sequences corresponding to the historical click rate sequences, the release probability vectors at the time stamps and the overall click rate vector to obtain a trained advertisement click rate prediction network;
and inputting the comprehensive characteristic vector to be predicted and the corresponding plan delivery strategy sequence into the trained advertisement click rate prediction network, and predicting the predicted click rate sequence corresponding to the plan delivery strategy sequence.
Preferably, the historical release strategy sequence and the corresponding historical click rate sequence have the same target timestamp corresponding to the element at the same position; the target timestamp is a timestamp in which the advertisement putting amount in each timestamp is not 0.
Preferably, the obtaining, according to the historical placement strategy sequences, placement probability vectors of the advertisements corresponding to the comprehensive feature vectors at the time stamps within a preset time period includes:
counting the sum of the advertisement putting quantities at the same target timestamp in each historical putting strategy sequence corresponding to the advertisement corresponding to the comprehensive characteristic vector to obtain a putting strategy distribution histogram; the abscissa of the distribution histogram of the release strategy is a timestamp, and the ordinate is distribution probability; the distribution probability is a value obtained by normalizing the advertisement putting quantity at the timestamp;
taking all timestamps in the distribution histogram of the release strategy and the corresponding distribution probability as sample data; fitting by utilizing an EM algorithm based on the sample data to obtain a corresponding Gaussian mixture model; the Gaussian mixture model comprises a plurality of sub-Gaussian models;
and obtaining the launching probability vector corresponding to each timestamp according to the value-taking ratio of each timestamp in each sub-Gaussian model.
Preferably, a calculation formula of the value ratio of any timestamp in any sub-gaussian model is as follows:
wherein the content of the first and second substances,
the value of the nth time stamp in the a-th sub-Gaussian model is taken as a ratio,
for the nth time stamp, the time stamp is,
for the ad placement probability at the nth timestamp,
is the weight of the a-th sub-gaussian model,
for the nth time stamp
Values in the sub-Gaussian model; and the advertisement putting probability is a probability value obtained according to a Gaussian mixture model.
Preferably, the obtaining of the total click rate vector of the advertisement corresponding to the comprehensive feature vector in a preset time period according to the historical click rate sequence corresponding to each historical placement strategy sequence includes:
for any historical placement strategy sequence: multiplying the advertisement putting quantity at each target timestamp in the historical putting strategy sequence by the corresponding click rate in the corresponding historical click rate sequence to obtain the click quantity at each target timestamp corresponding to the putting strategy sequence;
accumulating the click rate of the same timestamp according to the click rate of each target timestamp corresponding to each historical release strategy sequence, and dividing the accumulated value by the total advertisement release rate of the corresponding timestamp to obtain the total click rate of each timestamp in a preset time period;
obtaining the total click rate at the timestamp mean value corresponding to each sub-Gaussian model according to the total click rate at each timestamp;
and obtaining the total click rate vector of the advertisement corresponding to the comprehensive characteristic vector in a preset time period according to the total click rate of the timestamp mean value corresponding to each sub-Gaussian model.
Preferably, the advertisement click rate prediction network is trained according to the comprehensive feature vector, the historical click rate sequences corresponding to the historical click strategy sequences, the click rate vectors at the timestamps, and the overall click rate vector, and the trained advertisement click rate prediction network has a loss function as follows:
wherein the content of the first and second substances,
for the loss function, R is the number of synthetic feature vectors input to the network,
for the number of each historical release strategy sequence corresponding to the r-th comprehensive characteristic vector,
the number of target time stamps corresponding to the kth historical delivery strategy sequence,
the kth historical putting strategy sequence corresponds to
The time stamp of each target is stored in a memory,
a launching probability vector of an nth target timestamp corresponding to a kth historical launching strategy sequence corresponding to the r-th comprehensive characteristic vector,
for the overall click rate vector corresponding to the r-th integrated feature vector,
is a transpose of the overall click-through rate vector,
for the click rate at the nth target timestamp corresponding to the kth historical release strategy sequence corresponding to the r-th comprehensive characteristic vector,
the predicted click rate at the nth target timestamp corresponding to the kth historical putting strategy sequence corresponding to the r comprehensive characteristic vector output by the network,
for the actual delivery effect at the nth target timestamp corresponding to the kth historical delivery strategy sequence corresponding to the r-th comprehensive characteristic vector,
and advertising putting quantity at the nth target timestamp corresponding to the kth historical putting strategy sequence corresponding to the r comprehensive characteristic vector.
Preferably, a calculation formula of an actual delivery effect at an nth target timestamp corresponding to a kth historical delivery policy sequence corresponding to the r-th comprehensive feature vector is as follows:
wherein the content of the first and second substances,
the kth historical putting strategy sequence corresponds to
Each eyeThe time stamp is marked on the time stamp,
the nth target time stamp and the kth target time stamp corresponding to the kth historical putting strategy sequence
The duration of the interval between individual target time stamps,
a kth history release strategy sequence corresponding to the r comprehensive characteristic vector
The amount of advertisement placement corresponding to each target timestamp,
the kth historical putting strategy sequence corresponds to
Target timestamp pair
The memory effect coefficient generated by each target timestamp.
Preferably, the memory effect coefficient is calculated by the formula:
wherein the content of the first and second substances,
for the minimum target timestamp in each historical release strategy sequence corresponding to the comprehensive characteristic vector,
for the maximum target timestamp in each historical release strategy sequence corresponding to the comprehensive characteristic vector,
is spaced between two time stamps by a time length of
The memory effect coefficient generated by the earlier time stamp to the later time stamp.
The invention has the following beneficial effects:
firstly, acquiring a comprehensive characteristic vector, and each historical advertisement putting strategy sequence and a corresponding historical click rate sequence of an advertisement corresponding to the comprehensive characteristic vector; then according to the historical putting strategy sequences and the corresponding historical click rate sequences, obtaining putting probability vectors of the advertisements corresponding to the comprehensive characteristic vectors at the time stamps and total click rate vectors of the advertisements corresponding to the comprehensive characteristic vectors in a preset time period; and finally, training the advertisement click rate prediction network according to the comprehensive characteristic vector, the historical click rate sequences corresponding to the historical release strategy sequences, the release probability vectors at the time stamps and the overall click rate vector to obtain a trained advertisement click rate prediction network, and predicting the click rate at each target time stamp in the planned release strategy sequence corresponding to the comprehensive characteristic vector to be predicted by utilizing the trained advertisement click rate prediction network to obtain a corresponding predicted click rate sequence. The invention trains the network by combining the relation between the memory of the user and the advertisement putting strategy, thereby improving the accuracy of the network for predicting the advertisement click rate.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions and advantages of the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.
Fig. 1 is a flowchart of a method for predicting a marketing advertisement click according to the present invention.
Detailed Description
To further illustrate the technical means and functional effects of the present invention for achieving the predetermined object, the following detailed description of a method for predicting a marketing advertisement click according to the present invention is provided with reference to the accompanying drawings and preferred embodiments.
Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
The following describes a specific scheme of the prediction method of the marketing advertisement click provided by the present invention in detail with reference to the accompanying drawings.
The embodiment of the prediction method of the marketing advertisement click comprises the following steps:
as shown in fig. 1, the method for predicting a marketing advertisement click of the present embodiment includes the following steps:
s1, acquiring a comprehensive characteristic vector, historical delivery strategy sequences of advertisements corresponding to the comprehensive characteristic vector and historical click rate sequences corresponding to the historical delivery strategy sequences; the comprehensive characteristic vector comprises a user characteristic vector, an advertisement characteristic vector and a delivery environment characteristic vector; the historical placement strategy sequence includes advertisement placement amounts of corresponding advertisements at each target timestamp within a preset time period.
In order to predict the click rate of the advertisement, the embodiment constructs an advertisement click rate prediction network; and then training the network by combining the user information, the advertisement information, the delivery environment information, the historical delivery strategy information and the corresponding historical click rate information to obtain a trained advertisement click rate prediction network. The embodiment next analyzes the advertisement click-through rate prediction network training process.
Acquiring user information (namely information of an advertisement putting object), advertisement information, putting environment information, historical putting strategy information and historical click rate information corresponding to different advertisements, specifically: in this embodiment, the user information includes information such as age, sex, and occupation of the user, and the advertisement information includes content of advertisement,Title, industry and the like, and the delivery environment comprises information such as advertisement delivery position and the like. The information includes discrete category values (such as occupation, gender and advertisement content category number) and continuous values (such as age and title word number); in this embodiment, one-hot encoding is performed on the discrete value, so as to obtain a user feature vector, an advertisement feature vector and a delivery environment feature vector; then, according to the user characteristic vector, the advertisement characteristic vector and the delivery environment characteristic vector, obtaining a comprehensive characteristic vector, namely the comprehensive characteristic vector
Wherein, in the step (A),
is a feature vector of the user, and is,
in the form of a feature vector for an advertisement,
a characteristic vector of the launching environment is obtained; for any of the synthetic feature vectors: the comprehensive characteristic vector corresponds to a user, an advertisement and a delivery environment; that is, the same advertisement may be delivered to different users in different delivery environments. In this embodiment, an advertisement corresponding to any comprehensive feature vector is taken as an example for analysis.
In this embodiment, one advertisement placement strategy represents the advertisement placement amount of an advertisement at each timestamp within a preset time period, where the preset time period is one month (that is, 30 days), and may be specifically set according to actual needs; in this embodiment, a time stamp corresponding to one day is set, and there are thirty time stamps; a time stamp in each time stamp whose advertisement placement amount is not 0 is designated as a target time stamp.
In this embodiment, a plurality of historical advertisement delivery policy sequences corresponding to an advertisement corresponding to the comprehensive feature vector and a historical click rate sequence corresponding to each historical advertisement delivery policy sequence are obtained according to historical data, where the historical advertisement delivery policy sequences are used to show delivery policy information in a corresponding historical time period, specifically:
randomly selecting a proper number of delivery strategies in historical time (namely one delivery strategy corresponds to one month), and acquiring a delivery time sequence (namely a target timestamp sequence) of the advertisement in each month, and advertisement delivery amount and click rate at the corresponding time; obtaining a historical putting strategy sequence corresponding to each month (namely, one putting strategy corresponds to one historical putting strategy sequence) according to the putting time sequence of the advertisement in each month and the advertisement putting quantity in the corresponding time; the historical placement strategy sequence comprises the advertisement placement amount of the advertisement at each target timestamp in a corresponding month, the abscissa of the sequence is the target timestamp (sorted from morning to evening), and the value of each element in the sequence is the advertisement placement amount at each corresponding target timestamp.
For the advertisement
The historical placement strategy sequence can be expressed as:
wherein, in the step (A),
is as follows
The sequence of the historical putting strategies is determined,
is as follows
The advertisement placement amount at the 1 st target timestamp in the sequence of historical placement strategies,
is as follows
The ad placement volume at the 2 nd target timestamp in the sequence of historical placement strategies,
is as follows
The advertisement placement amount at the nth target timestamp in the sequence of historical placement strategies. For example, an arbitrary placement strategy for the ad: and the number 5 putting amount is 5, the number 15 putting amount is 6, the number 20 putting amount is 7 in one month, the first target timestamp is number 5, the second target timestamp is number 15, and the third target timestamp is number 20, so that the corresponding historical putting strategy sequence is {5,6,7}. The nth target timestamp in different historical placement strategy sequences is not necessarily the same, i.e., is not necessarily the same timestamp.
Acquiring historical click rate sequences corresponding to historical click rate strategy sequences corresponding to the advertisements according to the advertisement delivery time sequences and click rates at corresponding time in each month, wherein the sequence abscissa of each historical click rate sequence is a target timestamp, and the value of each element in each sequence is the click rate at each corresponding target timestamp; the historical click rate sequence corresponds to each element in the corresponding historical release strategy sequence, namely the click rate on each target timestamp corresponds to the advertisement release amount one by one; that is, the first element in the historical placement strategy sequence is the advertisement placement amount at the first target timestamp, the first element in the corresponding historical click rate sequence is the click rate at the first target timestamp, and the first target timestamps in the two corresponding sequences are the same timestamp.
For the advertisement
The historical click rate sequence corresponding to each historical placement strategy sequence can be expressed as:
wherein, in the step (A),
is as follows
A historical click rate sequence corresponding to each historical putting strategy sequence,
is as follows
Click rate at the 1 st target timestamp in the historical click rate sequence corresponding to each historical release strategy sequence,
is as follows
Click rate at the 2 nd target timestamp in the historical click rate sequence corresponding to each historical release strategy sequence,
is as follows
And click rate at the nth target timestamp in the historical click rate sequence corresponding to the historical release strategy sequence.
Thus, in this embodiment, each historical placement strategy sequence of the advertisement corresponding to the comprehensive feature vector and the historical click rate sequence corresponding to each historical placement strategy sequence are obtained.
S2, according to the historical release strategy sequences, release probability vectors of the advertisements corresponding to the comprehensive characteristic vectors at the time stamps in a preset time period are obtained; and obtaining the total click rate vector of the advertisement corresponding to the comprehensive characteristic vector in a preset time period according to the historical click rate sequence corresponding to each historical putting strategy sequence.
Next, in this embodiment, each historical placement strategy sequence of the advertisement corresponding to the comprehensive feature vector obtained in step S1 and the historical click rate sequence corresponding to each historical placement strategy sequence are preprocessed.
Firstly, integrating each historical delivery strategy sequence, and acquiring the distribution characteristics of the advertisement delivery quantity of the advertisement at each timestamp (namely, the distribution condition of the advertisement delivery quantity at each timestamp in a preset time period), specifically: counting each historical putting strategy sequence corresponding to the advertisement corresponding to the comprehensive characteristic vector (namely counting the sum of the advertisement putting quantities at the same target timestamp in each historical putting strategy sequence to obtain the advertisement putting quantity at each timestamp), and obtaining a putting strategy distribution histogram; the distribution histogram of the delivery strategy is used for counting the distribution probability of the advertisement delivery quantity at each timestamp in a preset time period, namely, the advertisement delivery quantity at each timestamp is normalized based on the total advertisement delivery quantity of each historical delivery strategy sequence corresponding to the advertisement corresponding to the comprehensive characteristic vector, and the normalized result is used as the distribution probability (namely, the distribution probability of the advertisement delivery quantity at each timestamp is obtained), so that the distribution histogram of the delivery strategy is obtained; the abscissa of the distribution histogram of the release strategy is the timestamp, and the ordinate is the distribution probability.
Taking all timestamps in the distribution histogram of the release strategy and the corresponding distribution probability as sample data, and then fitting by utilizing an EM algorithm based on the sample data to obtain a corresponding Gaussian mixture model; the number of sub-Gaussian models in the Gaussian mixture model is
（
The value of (c) is specifically set according to actual needs). The present embodiment describes, by using the gaussian mixture model, a probability of placing an advertisement at each timestamp for an arbitrary placement strategy, which is denoted as an advertisement placement probability; this probability is given by
Multiplying the calculation result of the sub-Gaussian model by the corresponding weight to obtainTo obtain the product
The sub-Gaussian models are arranged from front to back according to the time sequence of the corresponding timestamp mean value and are respectively marked as serial numbers 1,2, …, N, and for a newly input timestamp
(i.e., the nth timestamp in the preset time period) is expressed as follows:
wherein, the first and the second end of the pipe are connected with each other,
for the nth time stamp, the time stamp is,
for the advertisement placement probability at the nth timestamp (i.e. the probability value obtained from the gaussian mixture model),
is the weight of the a-th sub-gaussian model,
for the nth time stamp
And taking values in the sub-Gaussian models, wherein N is the number of the sub-Gaussian models corresponding to the Gaussian mixture model. In this embodiment, fitting is performed on data by using an EM algorithm, and a process of obtaining a gaussian mixture model is the prior art, and is not described in detail herein.
For a time stamp
The corresponding advertisement putting probability is determined by
The proportion of sub-gaussian models (i.e. the ratio of the timestamp to the value in the a-th sub-gaussian model) is calculated as follows:
wherein the content of the first and second substances,
the value ratio of the nth timestamp in the a-th sub-Gaussian model is obtained; namely, it is
。
The advertisement placement probability of the advertisement corresponding to the integrated feature vector at any timestamp can be decomposed and expressed by a Gaussian mixture model, and then the placement probability vector corresponding to the timestamp (i.e. the placement probability vector of the advertisement corresponding to the integrated feature vector at each timestamp), i.e. the placement probability vector corresponding to the integrated feature vector at each timestamp, is obtained
Wherein, in the step (A),
for the delivery probability vector corresponding to the nth timestamp,
the value of the nth timestamp in the 1 st sub-Gaussian model is taken as a ratio,
the value of the nth timestamp in the 2 nd sub-gaussian model is taken as the ratio,
and comparing the value of the nth timestamp in the nth sub-Gaussian model.
For any historical advertisement delivery strategy sequence corresponding to the comprehensive feature vector, since the historical advertisement delivery strategy sequence corresponds to the target timestamps of the historical click rate sequence one by one, the advertisement delivery amount at each target timestamp in the historical advertisement delivery strategy sequence is multiplied by the corresponding click rate in the historical click rate sequence to obtain the click rate at each target timestamp corresponding to the advertisement delivery strategy sequence.
Accumulating the click rate of the same timestamp according to the click rate of each target timestamp corresponding to each historical advertisement delivery strategy sequence corresponding to the comprehensive characteristic vector, and dividing the accumulated value by the total advertisement delivery rate of the corresponding timestamp to obtain the total click rate of each timestamp; the total click rate at each timestamp can be obtained according to the above process, and the total click rate at the nth timestamp is
. The total click rate at each timestamp within any preset time period can be obtained.
According to the constructed Gaussian mixture model, obtaining the total click rate at the timestamp mean value corresponding to each sub-Gaussian model, and recording as the total click rate
In which
Is a first
The mean value of the time stamps of the sub-gaussian models,
is as follows
Overall click rate at timestamp mean of sub-gaussian model. According to the total click rate of the timestamp mean value corresponding to each sub-Gaussian model, the advertisement can be obtained within a preset time period (namely any time period)Put strategy) of the total click-through rate vector (i.e., one comprehensive feature vector corresponds to one total click-through rate vector), and is recorded as
Wherein
Is as follows
The overall click rate at the timestamp mean corresponding to the sub-gaussian model,
is a first
The overall click rate at the timestamp mean corresponding to the sub-gaussian model,
is as follows
The overall click rate at the timestamp mean corresponding to the sub-gaussian model,
is the overall click rate vector.
And S3, training the advertisement click rate prediction network according to the comprehensive characteristic vector, the historical click rate sequences corresponding to the historical click rate sequences, the release probability vectors at the time stamps and the overall click rate vector to obtain the trained advertisement click rate prediction network.
Next, in this embodiment, an advertisement click-through rate prediction network is constructed, where the input of the advertisement click-through rate prediction network is the comprehensive feature vector and the corresponding delivery policy sequence, and the network output is the click-through rate sequence corresponding to the predicted delivery policy sequence.
In this embodiment, a training data set is obtained, where the training data set includes a plurality of comprehensive feature vectors (one comprehensive feature vector corresponds to one advertisement), historical delivery strategy sequences corresponding to the comprehensive feature vectors, and historical click rate sequences corresponding to the comprehensive feature vectors; acquiring an advertisement delivery probability vector and a corresponding overall click rate vector at each timestamp corresponding to each comprehensive characteristic vector according to the process of the step S2; training the advertisement click rate prediction network by utilizing a training data set, wherein a loss function in the training process is as follows:
wherein the content of the first and second substances,
for the loss function, R is the number of ads input to the network (i.e., the number of synthetic feature vectors input to the network),
for the number of each historical releasing strategy sequence corresponding to the r-th comprehensive characteristic vector,
the number of target timestamps (i.e. the number of advertisement placement volumes) corresponding to the kth historical placement strategy sequence,
the kth historical putting strategy sequence corresponds to
The time stamp of each target is stored in a memory,
a delivery probability vector of an nth target timestamp corresponding to a kth historical delivery strategy sequence corresponding to the r comprehensive characteristic vector,
for the overall click rate vector corresponding to the r-th integrated feature vector,
is a transpose of the overall click-through rate vector,
for the click rate (i.e. the true click rate value) at the nth target timestamp corresponding to the kth historical release strategy sequence corresponding to the r-th comprehensive feature vector,
the click rate (marked as predicted click rate) at the nth target timestamp corresponding to the kth historical putting strategy sequence corresponding to the r-th comprehensive characteristic vector output by the network,
for the actual delivery effect at the nth target timestamp corresponding to the kth historical delivery strategy sequence corresponding to the r-th comprehensive characteristic vector,
advertising release amount at the nth target timestamp corresponding to the kth historical release strategy sequence corresponding to the r comprehensive characteristic vector; wherein
Is a numerical value.
In the above formula
The smaller the predicted value of the network is, the closer the predicted value is to the true value, and the smaller the corresponding Loss is; when in use
The smaller the corresponding Loss is.
The actual putting effect obtaining process comprises the following steps:
for any comprehensive characteristic vector and any corresponding historical release strategy sequence:
considering that the user may focus on the advertisement before using the advertisement, but actually click after thinking about the advertisement before seeing the advertisement; this process illustrates that the advertisement delivered at the previous moment may be the reason why the advertisement was clicked at the subsequent moment, and therefore, the actual delivery effect of the delivery strategy may be different from the delivery strategy itself under the influence of the memory effect. Therefore, in this embodiment, the actual delivery effect of each target timestamp in the historical delivery policy sequence is calculated by combining the interval distance between each target timestamp in the historical delivery policy sequence, that is:
wherein the content of the first and second substances,
the kth historical putting strategy sequence corresponds to
The time stamp of each target is compared with the time stamp of each target,
the nth target time stamp and the kth target time stamp corresponding to the kth historical putting strategy sequence
The duration of the interval between the target time stamps,
a kth history release strategy sequence corresponding to the r comprehensive characteristic vector
The amount of advertisement placement corresponding to each target timestamp,
the kth corresponding to the kth historical delivery strategy sequence
Target timestamp pair
The memory effect coefficient generated by each target timestamp.
The calculation formula of the memory effect coefficient is as follows:
wherein the content of the first and second substances,
for the minimum target timestamp in each historical release strategy sequence corresponding to the comprehensive characteristic vector,
for the maximum target timestamp in each historical release strategy sequence corresponding to the comprehensive characteristic vector,
is spaced apart by a time duration of
The memory effect coefficient generated by the earlier time stamp to the later time stamp;
has the functions of
Are normalized, i.e.
Are normalized values.
According to the above formula when
The closer to 0, the smaller the memory effect is; when the temperature is higher than the set temperature
The closer to 1, the greater the memory effect.
So far, according to the above process, a trained advertisement click-through rate prediction network can be obtained.
And S4, inputting the comprehensive characteristic vector to be predicted and the corresponding plan delivery strategy sequence into the trained advertisement click rate prediction network, and predicting the predicted click rate sequence corresponding to the plan delivery strategy sequence.
In the embodiment, a trained advertisement click rate prediction network is obtained according to the step S3; then, acquiring a comprehensive characteristic vector to be predicted (an advertisement to be predicted) and a delivery strategy vector of an advertisement plan (marked as a plan delivery strategy sequence) corresponding to the comprehensive characteristic vector to be predicted; and inputting the comprehensive characteristic vector to be predicted and the corresponding planned delivery strategy sequence into a trained advertisement click rate prediction network, wherein the network can predict a click rate sequence (recorded as a predicted click rate sequence) corresponding to the planned delivery strategy sequence, namely the click rate of the corresponding advertisement under the planned delivery strategy.
The method comprises the steps of firstly obtaining a comprehensive characteristic vector, and each historical advertisement putting strategy sequence and corresponding historical click rate sequence of advertisements corresponding to the comprehensive characteristic vector; then according to the historical releasing strategy sequences and the corresponding historical click rate sequences, releasing probability vectors of the advertisements corresponding to the comprehensive characteristic vectors at the time stamps and total click rate vectors of the advertisements corresponding to the comprehensive characteristic vectors in a preset time period are obtained; and finally, training the advertisement click rate prediction network according to the comprehensive characteristic vector, the historical click rate sequences corresponding to the historical release strategy sequences, the release probability vectors at the time stamps and the overall click rate vector to obtain a trained advertisement click rate prediction network, and predicting the click rate at each target time stamp in the planned release strategy sequence corresponding to the comprehensive characteristic vector to be predicted by utilizing the trained advertisement click rate prediction network to obtain a corresponding predicted click rate sequence. The embodiment trains the network by combining the relation between the memory of the user and the advertisement putting strategy, thereby improving the accuracy of the network to the advertisement click rate prediction.
It should be noted that: the above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (8)
1. A method for predicting a marketing advertisement click, the method comprising the steps of:
acquiring a comprehensive characteristic vector, each historical advertisement putting strategy sequence corresponding to the comprehensive characteristic vector and a historical click rate sequence corresponding to each historical advertisement putting strategy sequence; the comprehensive characteristic vector comprises a user characteristic vector, an advertisement characteristic vector and a delivery environment characteristic vector; the historical release strategy sequence comprises the advertisement release amount of the corresponding advertisement at each target timestamp in a preset time period;
obtaining delivery probability vectors of the advertisements corresponding to the comprehensive characteristic vectors at each timestamp in a preset time period according to the historical delivery strategy sequences; obtaining a total click rate vector of the advertisement corresponding to the comprehensive characteristic vector in a preset time period according to the historical click rate sequence corresponding to each historical putting strategy sequence;
training an advertisement click rate prediction network according to the comprehensive characteristic vector, the historical click rate sequences corresponding to the historical click rate sequences, the release probability vectors at the time stamps and the overall click rate vector to obtain a trained advertisement click rate prediction network;
and inputting the comprehensive characteristic vector to be predicted and the corresponding plan delivery strategy sequence into the trained advertisement click rate prediction network, and predicting the predicted click rate sequence corresponding to the plan delivery strategy sequence.
2. The method of claim 1, wherein the historical placement strategy sequence and the corresponding historical click rate sequence have the same target timestamp for the co-located element; the target timestamp is a timestamp with an advertisement placement amount of 0 in each timestamp.
3. The method of claim 1, wherein obtaining placement probability vectors of the advertisement corresponding to the integrated feature vector at each timestamp within a preset time period according to the historical placement strategy sequences comprises:
counting the sum of the advertisement putting quantities at the same target timestamp in each historical putting strategy sequence corresponding to the advertisement corresponding to the comprehensive characteristic vector to obtain a putting strategy distribution histogram; the abscissa of the distribution histogram of the release strategy is a timestamp, and the ordinate is distribution probability; the distribution probability is a value obtained by normalizing the advertisement putting quantity at the timestamp;
taking all timestamps in the distribution histogram of the release strategy and the corresponding distribution probability as sample data; fitting by utilizing an EM algorithm based on the sample data to obtain a corresponding Gaussian mixture model; the Gaussian mixture model comprises a plurality of sub-Gaussian models;
and obtaining the launching probability vector corresponding to each timestamp according to the value-taking ratio of each timestamp in each sub-Gaussian model.
4. The method of claim 3, wherein the calculation formula of the value ratio of any timestamp in any sub-Gaussian model is as follows:
wherein the content of the first and second substances,
the value of the nth time stamp in the a-th sub-Gaussian model is taken as a ratio,
for the nth time stamp, the time stamp is,
for the ad placement probability at the nth timestamp,
is the weight of the a-th sub-gaussian model,
for the nth time stamp
Values in the sub-Gaussian model; the advertisement putting probability is a probability value obtained according to a Gaussian mixture model.
5. The method of claim 3, wherein obtaining the total click rate vector of the advertisement corresponding to the integrated feature vector within a preset time period according to the historical click rate sequences corresponding to the historical placement strategy sequences comprises:
for any historical placement strategy sequence: multiplying the advertisement putting quantity at each target timestamp in the historical putting strategy sequence by the corresponding click rate in the corresponding historical click rate sequence to obtain the click quantity at each target timestamp corresponding to the putting strategy sequence;
accumulating the click rate of the same timestamp according to the click rate of each target timestamp corresponding to each historical release strategy sequence, and dividing the accumulated value by the total advertisement release rate of the corresponding timestamp to obtain the total click rate of each timestamp in a preset time period;
obtaining the total click rate at the timestamp mean value corresponding to each sub-Gaussian model according to the total click rate at each timestamp;
and obtaining the total click rate vector of the advertisement corresponding to the comprehensive characteristic vector in a preset time period according to the total click rate of the timestamp mean value corresponding to each sub-Gaussian model.
6. The method of claim 1, wherein an advertisement click-through rate prediction network is trained according to the comprehensive feature vector, the historical click-through strategy sequences, click-through rate sequences corresponding to the historical click-through strategy sequences, the click-through probability vectors at the timestamps, and the total click-through rate vector, and a loss function of the trained advertisement click-through rate prediction network is:
wherein the content of the first and second substances,
for the loss function, R is the number of synthetic feature vectors input to the network,
for the number of each historical release strategy sequence corresponding to the r-th comprehensive characteristic vector,
the number of target time stamps corresponding to the kth historical delivery strategy sequence,
the kth historical putting strategy sequence corresponds to
The time stamp of each target is stored in a memory,
a launching probability vector of an nth target timestamp corresponding to a kth historical launching strategy sequence corresponding to the r-th comprehensive characteristic vector,
for the overall click rate vector corresponding to the r-th integrated feature vector,
is a transpose of the overall click-through rate vector,
for the click rate at the nth target timestamp corresponding to the kth historical release strategy sequence corresponding to the r-th comprehensive characteristic vector,
the predicted click rate at the nth target timestamp corresponding to the kth historical release strategy sequence corresponding to the r-th comprehensive characteristic vector output by the network,
for the actual delivery effect at the nth target timestamp corresponding to the kth historical delivery strategy sequence corresponding to the r-th comprehensive characteristic vector,
and advertising putting quantity at the nth target timestamp corresponding to the kth historical putting strategy sequence corresponding to the r comprehensive characteristic vector.
7. The method of claim 6, wherein the actual placement effect at the nth target timestamp corresponding to the kth historical placement strategy sequence corresponding to the r-th integrated feature vector is calculated by the following formula:
wherein the content of the first and second substances,
the kth historical putting strategy sequence corresponds to
The time stamp of each target is stored in a memory,
the nth target time stamp and the kth target time stamp corresponding to the kth historical putting strategy sequence
The duration of the interval between individual target time stamps,
a kth history release strategy sequence corresponding to the r comprehensive characteristic vector
The amount of advertisement placement corresponding to each target timestamp,
the kth historical putting strategy sequence corresponds to
Target timestamp pair
Memory generated by individual target time stampThe coefficient of effect.
8. The method of claim 7, wherein the memory effect factor is calculated as:
wherein the content of the first and second substances,
for the minimum target timestamp in each historical release strategy sequence corresponding to the comprehensive characteristic vector,
for the maximum target timestamp in each historical release strategy sequence corresponding to the comprehensive characteristic vector,
is spaced apart by a time duration of
The memory effect coefficient generated by the earlier time stamp to the later time stamp.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211244719.0A CN115330467B (en) | 2022-10-12 | 2022-10-12 | Marketing advertisement click prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211244719.0A CN115330467B (en) | 2022-10-12 | 2022-10-12 | Marketing advertisement click prediction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115330467A true CN115330467A (en) | 2022-11-11 |
| CN115330467B CN115330467B (en) | 2022-12-20 |
Family
ID=83914968
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211244719.0A Active CN115330467B (en) | 2022-10-12 | 2022-10-12 | Marketing advertisement click prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115330467B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107018493A (en) * | 2017-04-20 | 2017-08-04 | 北京工业大学 | A kind of geographical position Forecasting Methodology based on continuous sequential Markov model |
| CN111682972A (en) * | 2020-08-14 | 2020-09-18 | 支付宝(杭州)信息技术有限公司 | Method and device for updating service prediction model |
| CN112182430A (en) * | 2020-09-22 | 2021-01-05 | 汉海信息技术(上海)有限公司 | Method and device for recommending places, electronic equipment and storage medium |
| CN112884529A (en) * | 2021-03-24 | 2021-06-01 | 杭州网易云音乐科技有限公司 | Advertisement bidding method, device, equipment and medium |
| CN115131079A (en) * | 2022-08-25 | 2022-09-30 | 道有道科技集团股份公司 | Data processing-based advertisement putting effect prediction method and device |
-
2022
- 2022-10-12 CN CN202211244719.0A patent/CN115330467B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107018493A (en) * | 2017-04-20 | 2017-08-04 | 北京工业大学 | A kind of geographical position Forecasting Methodology based on continuous sequential Markov model |
| CN111682972A (en) * | 2020-08-14 | 2020-09-18 | 支付宝(杭州)信息技术有限公司 | Method and device for updating service prediction model |
| CN112182430A (en) * | 2020-09-22 | 2021-01-05 | 汉海信息技术(上海)有限公司 | Method and device for recommending places, electronic equipment and storage medium |
| CN112884529A (en) * | 2021-03-24 | 2021-06-01 | 杭州网易云音乐科技有限公司 | Advertisement bidding method, device, equipment and medium |
| CN115131079A (en) * | 2022-08-25 | 2022-09-30 | 道有道科技集团股份公司 | Data processing-based advertisement putting effect prediction method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115330467B (en) | 2022-12-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111651722B (en) | Risk assessment method and platform for advertisement putting effect of advertiser | |
| WO2019242331A1 (en) | User behavior prediction method and apparatus, and behavior prediction model training method and apparatus | |
| CN107657486A (en) | A kind of advertisement placement method and device | |
| US20130332264A1 (en) | Method and system for determining touchpoint attribution | |
| US20080103887A1 (en) | Selecting advertisements based on consumer transactions | |
| CN103942236A (en) | System and method for serving electronic content | |
| CN109636477B (en) | Early warning method and device for advertisement delivery shortage | |
| CN111798280B (en) | Multimedia information recommendation method, device and equipment and storage medium | |
| CN106897892A (en) | Advertisement placement method and device | |
| CN104574160A (en) | Smooth advertisement traffic control method | |
| EP2359328A1 (en) | A system for providing information concerning the effectiveness of advertising | |
| CN106296247A (en) | The online sort method of network information resource and device | |
| CN109949089A (en) | A kind of method, apparatus and terminal of determining displaying rate | |
| CN112150191B (en) | Advertisement putting method and system | |
| CN111899041A (en) | Information delivery processing method, information delivery device, information delivery equipment and storage medium | |
| US10181130B2 (en) | Real-time updates to digital marketing forecast models | |
| CN108564404A (en) | A kind of method and device of prediction investment in advertising return rate | |
| CN112465549A (en) | System and method for identifying channel cheating | |
| CN112884529B (en) | Advertisement bidding method, device, equipment and medium | |
| CN115330467B (en) | Marketing advertisement click prediction method | |
| CN109191159B (en) | Data orientation method and device, computer equipment and computer readable storage medium | |
| CN110245969B (en) | Promotion information processing method and device, electronic equipment and computer readable medium | |
| CN103942194B (en) | A kind of information delivers the optimization method and device of account | |
| CN114693361A (en) | Method, device, equipment and storage medium for determining contribution degree of channel | |
| CN116109353A (en) | Mobile phone application store advertisement delivery management platform |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |