A method and device for expanding similar groups of people based on computational advertising
By obtaining and processing samples in the context of computing advertising, building and adjusting machine learning models, and combining rule filtering, the problem of advertisers identifying high-potential users and balancing accuracy and scale is solved, efficient target population screening is achieved, and advertising is improved Effect.
Patent Information
- Application Number
- CN202110295616.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-03-19
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2041-03-19
AI Technical Summary
In the context of computing advertising, it is difficult for advertisers to effectively identify high-potential users and balance accuracy with scale. The existing Lookalike technology has the problems of high cost of manual participation, insufficient accuracy and low sensitivity to seed users.
By obtaining positive and negative samples, performing hierarchical sampling and feature processing, building a machine learning model and adjusting weight parameters, combining rule filtering, filtering out similar groups in fuzzy scoring areas, reducing the reach scale, and improving the accuracy of screening target groups.
It has achieved the reduction of the population size on the basis of reaching the target population as much as possible, improving the accuracy of target population screening, reducing advertising costs, and enhancing advertising effectiveness.
Smart Images

Figure CN112907295B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information processing technology, and in particular to a method and device for expanding similar groups of people based on the background of computational advertising. Background Art
[0002] In Internet commercial applications, many advertisers encounter problems such as difficulty in identifying high-potential groups and difficulty in balancing costs and scale when "searching for potential customers". Against this background, Lookalike (similar group expansion) population targeting technology was created. This technology can use the advertiser's first-party data to expand the user group similar to the seed based on a small number of seed users through big data analysis and machine learning. These expanded similar groups are likely to become the advertiser's target customers in the end, thereby improving advertising effectiveness.
[0003] Mainstream methods of Lookalike technology: Method 1, display positioning, advertisers select the target population based on labels. This method is simple and intuitive. Advertisers use user portrait labels to filter gender, age, preference and other labels to directly filter the target population. However, this method requires a lot of manual trial and error by advertisers, and this method has certain limitations. Manual labels cannot fully summarize all the attributes of the target population. For example, it is not necessarily that men do not care about skin care products. The display positioning method is long-term, costly, and difficult to be universal.
[0004] Method 2: Implicit positioning, which uses machine learning to model seed users. Implicit positioning requires almost no involvement from advertisers. It only requires advertisers to provide the characteristics of the target population (i.e., seed users). Through machine learning, similar populations are automatically found based on seed data, effectively avoiding the problems faced by custom labels.
[0005] Difficulties of Lookalike technology: Difficulty 1: It is difficult to find high-potential users and it is difficult to strike a balance between precision and scale. These are the two main problems faced by advertisers. The core lies in effectively reaching a large number of potential users. It is relatively difficult to achieve the "Pareto optimality" (the most ideal state) between effect and scale. Specifically, if advertisers want to reach potential target customers as much as possible, they will need to reach a large number of potential users. The focus of the crowd will inevitably decrease gradually, causing the proportion of non-target people to increase with the increase in traffic, increasing advertising costs. However, if advertisers reduce the scale of reach, a part of the target population will not be reached, affecting the advertising effect.
[0006] Difficulty 2: Reducing the sensitivity of seed users: Seed users are the premise and foundation for expansion, and the quality of seed users is often the key to the effectiveness of lookalikes. However, it is difficult for advertisers to provide seed packages with a large enough and wide enough data volume. At this time, it is necessary to consider how to perform effective data preprocessing and model learning when there are a small number of seed packages and the seeds may not cover global features. Summary of the invention
[0007] The technical problem to be solved by the present invention is to provide a method and device for expanding similar groups of people based on the background of computational advertising in view of the deficiencies in the prior art.
[0008] The technical solution of the present invention to solve the above technical problems is as follows: a similar group expansion method based on the background of computational advertising, comprising:
[0009] Obtain positive samples consisting of seed populations, negative samples consisting of non-seed populations, preset negative sample sampling conditions, advertising front-end monitoring data, third-party label data, and preset rule labels;
[0010] Performing stratified sampling processing on the negative samples according to the preset negative sample sampling conditions to obtain sampled negative samples;
[0011] Performing feature processing on the advertising front-end monitoring data and the third-party label data respectively, and constructing models respectively, to obtain a first model and a second model respectively;
[0012] In the first model and the second model, the weight parameters of the positive samples are adjusted respectively, and the third model and the fourth model are obtained accordingly;
[0013] The third model and the fourth model are integrated and scored to obtain crowd scoring information;
[0014] Selecting a fuzzy scoring area in the crowd scoring information;
[0015] A model is further constructed for the fuzzy scoring area to filter and screen to obtain a first similar population;
[0016] The first similar group of people is screened and filtered according to preset rule tags and the advertising front-end monitoring data to obtain a similar group of people.
[0017] Furthermore, the steps of performing feature processing on the advertising front-end monitoring data and the third-party tag data respectively and constructing models respectively, and correspondingly obtaining the first model and the second model, include:
[0018] Performing numerical coding processing on the advertisement front-end monitoring data to obtain numerically coded advertisement front-end monitoring data;
[0019] Performing numerical binning processing on the numerically encoded advertising front-end monitoring data to obtain customer relationship management features;
[0020] A model is constructed for the customer relationship management features to obtain a first model.
[0021] Furthermore, the steps of performing feature processing on the advertising front-end monitoring data and the third-party tag data respectively and constructing models respectively, and correspondingly obtaining the first model and the second model, include:
[0022] Performing word embedding processing on the third-party label data to obtain embedded third-party label data;
[0023] A model is constructed for the embedded third-party tag data to obtain a second model.
[0024] Furthermore, the step of selecting the fuzzy scoring area in the crowd scoring information includes:
[0025] The area with a score between 0.5 and 0.7 in the crowd scoring information is set as a fuzzy scoring area.
[0026] Furthermore, the respectively adjusting weight parameters of the positive samples is respectively allocating and adjusting the weights of the positive samples according to the time attenuation coefficient.
[0027] The beneficial effect of the present invention is that through the method of machine learning and rule package filtering, the scale of the target population is reduced while ensuring that the target population is reached as much as possible, thereby improving the accuracy of target population screening.
[0028] In addition, the present invention also provides a similar group expansion device based on the background of computational advertising, which includes:
[0029] An acquisition device is used to acquire positive samples consisting of a seed population, negative samples consisting of a non-seed population, preset negative sample sampling conditions, advertising front-end monitoring data, third-party label data, and preset rule labels;
[0030] A processing device, used for performing stratified sampling processing on the negative samples according to the preset negative sample sampling conditions to obtain sampled negative samples;
[0031] The processing device is further used to perform feature processing on the advertising front-end monitoring data and the third-party label data respectively, and to construct models respectively, to obtain a first model and a second model respectively;
[0032] The processing device is further used to adjust the weight parameters of the positive samples in the first model and the second model respectively, so as to obtain the third model and the fourth model accordingly;
[0033] The processing device is further used to perform fusion scoring on the third model and the fourth model to obtain crowd scoring information;
[0034] The processing device is further used to select the fuzzy scoring area in the crowd scoring information;
[0035] The processing device is further used to further construct a model for filtering and screening the fuzzy scoring area to obtain a first similar group of people;
[0036] The processing device is further used to screen and filter the first similar group of people according to preset rule tags and the advertising front-end monitoring data to obtain similar groups of people.
[0037] Furthermore, the processing device is also used to perform digital coding processing on the advertising front-end monitoring data to obtain digitally coded advertising front-end monitoring data;
[0038] The processing device is further used to perform numerical binning processing on the numerically encoded advertising front-end monitoring data to obtain customer relationship management features;
[0039] The processing device is further used to construct a model for the customer relationship management characteristics to obtain a first model.
[0040] Furthermore, the processing device is further used to perform word embedding processing on the third-party label data to obtain embedded third-party label data;
[0041] The processing device is also used to build a model for the embedded third-party tag data to obtain a second model.
[0042] Furthermore, the processing device is also used to set the area with a score between 0.5 and 0.7 in the crowd scoring information as a fuzzy scoring area.
[0043] Furthermore, the respectively adjusting weight parameters of the positive samples is respectively allocating and adjusting the weights of the positive samples according to the time attenuation coefficient.
[0044] The beneficial effect of the present invention is that through the method of machine learning and rule package filtering, the scale of the target population is reduced while ensuring that the target population is reached as much as possible, thereby improving the accuracy of target population screening.
[0045] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 One of the schematic flow charts of the similar group expansion method based on the computational advertising background provided by the embodiment of the present invention.
[0047] Figure 2 The second schematic flow chart of the similar population expansion method based on the computational advertising background provided by the embodiment of the present invention.
[0048] Figure 3 A schematic structural block diagram of a similar population expansion device based on the computational advertising background provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0049] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.
[0050] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a similar group expansion method based on the background of computational advertising, which includes:
[0051] S1. Obtain positive samples consisting of seed populations, negative samples consisting of non-seed populations, preset negative sample sampling conditions, advertising front-end monitoring data, third-party label data, and preset rule labels;
[0052] S2. Performing stratified sampling processing on the negative samples according to the preset negative sample sampling conditions to obtain sampled negative samples;
[0053] S3, performing feature processing on the advertising front-end monitoring data and the third-party label data respectively, and constructing models respectively, to obtain a first model and a second model respectively;
[0054] S4. In the first model and the second model, respectively adjust the weight parameters of the positive samples to obtain a third model and a fourth model accordingly;
[0055] S5, fusing and scoring the third model and the fourth model to obtain crowd scoring information;
[0056] S6, selecting a fuzzy scoring area in the crowd scoring information;
[0057] S7, further constructing a model for filtering and screening the fuzzy scoring area to obtain a first similar group of people;
[0058] S8. Screen and filter the first similar group of people according to preset rule tags and the advertising front-end monitoring data to obtain similar groups of people.
[0059] Lookalike, or similar group expansion, is a technology that uses a certain algorithm evaluation model based on seed users to find more similar groups of people with potential correlations.
[0060] Seed population, that is, target population.
[0061] Logistics regression is a supervised statistical learning method, which is mainly used to classify samples.
[0062] GBDT (Gradient Boosting Decision Tree) is an iterative decision tree algorithm that consists of multiple decision trees, and the conclusions of all trees are accumulated to get the final answer.
[0063] The embodiment of the present invention uses rule labels for unified filtering, specifically: step 11, select positive and negative samples, and sample negative samples; step 12, perform feature processing on the advertising front-end monitoring data and the first and third-party label data respectively, and build a model; step 13, adjust the weight parameters of the positive samples, and assign weights according to the time attenuation coefficient; step 14, model fusion scoring; step 15, further build a model for filtering and screening for the fuzzy scoring area; step 16, use rule labels for unified filtering, and screen out active people in recent months, and finally circle out high-potential people.
[0064] Among them, 1. Positive samples are selected from the seed population, and negative samples are sampled in a stratified manner. For example, in the scenario of expanding users who have retained information, negative samples can be stratified and sampled from samples of those who have arrived at the station but have not retained information and those who have not arrived at the station. Then, the pu-learning (positive-unlabeled learning, a semi-supervised binary classifier with only positive samples) idea is used to train the model multiple times to remove high-scoring samples from negative samples and screen out the final negative samples. Since most of the target population is included in the recently active population, the recently active population is selected for both positive and negative samples.
[0065] 2. Since the data dimension of front-end advertising monitoring is high and it is difficult to adopt a tree model, the front-end advertising monitoring data is numerically encoded and binned, and RFM (Recency-Frequency-Monetary, customer relationship management) features are constructed. After the features are flattened, the model is constructed using the logistics regression method. The word2vec (unique hot encoding) method is used for word embedding of one- and three-party label data, and then the GBDT (Gradient Boosting Decision Tree) method is used to build the model.
[0066] 3. Since recent user behaviors are more valuable for mining high-potential groups (similar groups), the weights of positive samples are allocated in the model according to the time decay coefficient, and the weights of recent positive samples are greater than those of early positive samples.
[0067] 4. Use the stacking method to fuse the LR model (Logistic Regression) and the GBDT (Gradient Boosting Decision Tree) model, and finally output the Quanling crowd score.
[0068] 5. We grouped and rated the population and found that the model divided the large population into high-potential groups, and a large number of them were concentrated between scores of 0.5-0.7. We defined it as a fuzzy scoring area, and further constructed a model for filtering and screening for this part of the samples.
[0069] 6. Finally, use commonly used rule tags and combine them with the active population monitored by the advertising front-end monitoring data in recent months to uniformly screen and filter all high-potential populations, and finally circle out the high-potential population.
[0070] The main improvements of the present invention are: 1. Stratified sampling is performed on negative samples, and pu-learning (positive-unlabeled learning, a semi-supervised binary classifier with only positive samples) is used to filter negative samples that may be positive samples, and recent activity is used. 2. Feature processing is performed on the advertising front-end monitoring data and the first- and third-party label data respectively, and the LR (Logistic Regression) and GBDT (Gradient Boosting Decision Tree) methods are used to build models respectively, and finally the stacking method is used for model fusion scoring. 3. The weights of positive samples are assigned in the model according to the time attenuation coefficient. 4. A model is further constructed for filtering and screening the low-scoring population of the high-potential population circled by the first-layer model. 5. In order to reduce the scale of the population reached by the advertisement, the high-potential users circled by the model are finally filtered in combination with the rule labels and the active population in recent months.
[0071] Furthermore, the steps of performing feature processing on the advertising front-end monitoring data and the third-party tag data respectively and constructing models respectively, and correspondingly obtaining the first model and the second model, include:
[0072] Performing numerical coding processing on the advertisement front-end monitoring data to obtain numerically coded advertisement front-end monitoring data;
[0073] Performing numerical binning processing on the numerically encoded advertising front-end monitoring data to obtain customer relationship management features;
[0074] A model is constructed for the customer relationship management features to obtain a first model.
[0075] Furthermore, the steps of performing feature processing on the advertising front-end monitoring data and the third-party tag data respectively and constructing models respectively, and correspondingly obtaining the first model and the second model, include:
[0076] Performing word embedding processing on the third-party label data to obtain embedded third-party label data;
[0077] A model is constructed for the embedded third-party tag data to obtain a second model.
[0078] Furthermore, the step of selecting the fuzzy scoring area in the crowd scoring information includes:
[0079] The area with a score between 0.5 and 0.7 in the crowd scoring information is set as a fuzzy scoring area.
[0080] Furthermore, the respectively adjusting weight parameters of the positive samples is respectively allocating and adjusting the weights of the positive samples according to the time attenuation coefficient.
[0081] The beneficial effect of the present invention is that through the method of machine learning and rule package filtering, the scale of the target population is reduced while ensuring that the target population is reached as much as possible, thereby improving the accuracy of target population screening.
[0082] like Figure 3 As shown, in addition, the present invention also provides a similar group expansion device based on the background of computational advertising, which includes:
[0083] An acquisition device is used to acquire positive samples consisting of a seed population, negative samples consisting of a non-seed population, preset negative sample sampling conditions, advertising front-end monitoring data, third-party label data, and preset rule labels;
[0084] A processing device, used for performing stratified sampling processing on the negative samples according to the preset negative sample sampling conditions to obtain sampled negative samples;
[0085] The processing device is further used to perform feature processing on the advertising front-end monitoring data and the third-party label data respectively, and to construct models respectively, to obtain a first model and a second model respectively;
[0086] The processing device is further used to adjust the weight parameters of the positive samples in the first model and the second model respectively, so as to obtain the third model and the fourth model accordingly;
[0087] The processing device is further used to perform fusion scoring on the third model and the fourth model to obtain crowd scoring information;
[0088] The processing device is further used to select the fuzzy scoring area in the crowd scoring information;
[0089] The processing device is further used to further construct a model for filtering and screening the fuzzy scoring area to obtain a first similar group of people;
[0090] The processing device is further used to screen and filter the first similar group of people according to preset rule tags and the advertising front-end monitoring data to obtain similar groups of people.
[0091] Furthermore, the processing device is also used to perform digital coding processing on the advertising front-end monitoring data to obtain digitally coded advertising front-end monitoring data;
[0092] The processing device is further used to perform numerical binning processing on the numerically encoded advertising front-end monitoring data to obtain customer relationship management features;
[0093] The processing device is further used to construct a model for the customer relationship management characteristics to obtain a first model.
[0094] Furthermore, the processing device is further used to perform word embedding processing on the third-party label data to obtain embedded third-party label data;
[0095] The processing device is also used to build a model for the embedded third-party tag data to obtain a second model.
[0096] Furthermore, the processing device is also used to set the area with a score between 0.5 and 0.7 in the crowd scoring information as a fuzzy scoring area.
[0097] Furthermore, the respectively adjusting weight parameters of the positive samples is respectively allocating and adjusting the weights of the positive samples according to the time attenuation coefficient.
[0098] The beneficial effect of the present invention is that through the method of machine learning and rule package filtering, the scale of the target population is reduced while ensuring that the target population is reached as much as possible, thereby improving the accuracy of target population screening.
[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A similar group expansion method based on computational advertising, characterized in that: include: Obtain positive samples consisting of seed populations, negative samples consisting of non-seed populations, preset negative sample sampling conditions, advertising front-end monitoring data, third-party label data, and preset rule labels; Performing stratified sampling processing on the negative samples according to the preset negative sample sampling conditions to obtain sampled negative samples; Performing feature processing on the advertising front-end monitoring data and the third-party label data respectively, and constructing models respectively, to obtain a first model and a second model respectively; In the first model and the second model, the weight parameters of the positive samples are adjusted respectively, and the third model and the fourth model are obtained accordingly; The third model and the fourth model are integrated and scored to obtain crowd scoring information; Selecting a fuzzy scoring area in the crowd scoring information; A model is further constructed for the fuzzy scoring area to filter and screen to obtain a first similar population; The first similar group of people is screened and filtered according to preset rule tags and the advertising front-end monitoring data to obtain similar groups of people; The step of selecting the fuzzy scoring area in the crowd scoring information includes: setting the area with a score between 0.5 and 0.7 in the crowd scoring information as the fuzzy scoring area; The respectively adjusting weight parameters of the positive samples is respectively allocating and adjusting the weights of the positive samples according to the time attenuation coefficient.
2. The similarity group expansion method based on computational advertising according to claim 1, characterized in that: The steps of performing feature processing on the advertising front-end monitoring data and the third-party label data respectively and constructing models respectively to obtain the first model and the second model respectively include: Performing numerical coding processing on the advertisement front-end monitoring data to obtain numerically coded advertisement front-end monitoring data; Performing numerical binning processing on the numerically encoded advertising front-end monitoring data to obtain customer relationship management features; A model is constructed for the customer relationship management features to obtain a first model.
3. The similarity group expansion method based on computational advertising according to claim 1, characterized in that: The steps of performing feature processing on the advertising front-end monitoring data and the third-party label data respectively and constructing models respectively to obtain the first model and the second model respectively include: Performing word embedding processing on the third-party label data to obtain embedded third-party label data; A model is constructed for the embedded third-party tag data to obtain a second model.
4. A similar group expansion device based on computational advertising, characterized in that: include: An acquisition device is used to acquire positive samples consisting of a seed population, negative samples consisting of a non-seed population, preset negative sample sampling conditions, advertising front-end monitoring data, third-party label data, and preset rule labels; A processing device, used for performing stratified sampling processing on the negative samples according to the preset negative sample sampling conditions to obtain sampled negative samples; The processing device is further used to perform feature processing on the advertising front-end monitoring data and the third-party label data respectively, and to construct models respectively, to obtain a first model and a second model respectively; The processing device is further used to adjust the weight parameters of the positive samples in the first model and the second model respectively, so as to obtain the third model and the fourth model accordingly; The processing device is further used to perform fusion scoring on the third model and the fourth model to obtain crowd scoring information; The processing device is further used to select the fuzzy scoring area in the crowd scoring information; The processing device is further used to further construct a model for filtering and screening the fuzzy scoring area to obtain a first similar group of people; The processing device is further used to screen and filter the first similar group of people according to preset rule tags and the advertising front-end monitoring data to obtain similar groups of people; Wherein, the processing device is further used to set the area with a score between 0.5 and 0.7 in the crowd scoring information as a fuzzy scoring area; The respectively adjusting weight parameters of the positive samples is respectively allocating and adjusting the weights of the positive samples according to the time attenuation coefficient.
5. The device for expanding similar groups of people based on the background of computational advertising according to claim 4, characterized in that: The processing device is further used to perform digital coding processing on the advertising front-end monitoring data to obtain digitally coded advertising front-end monitoring data; The processing device is further used to perform numerical binning processing on the numerically encoded advertising front-end monitoring data to obtain customer relationship management features; The processing device is further used to construct a model for the customer relationship management characteristics to obtain a first model.
6. The device for expanding similar groups of people based on the background of computational advertising according to claim 4, characterized in that: The processing device is further used to perform word embedding processing on the third-party label data to obtain embedded third-party label data; The processing device is also used to build a model for the embedded third-party tag data to obtain a second model.
