MBTI personality types: impact on advertising perception
Hi Habr! Analyzing user data for marketing and advertising needs, we decided to investigate the influence of the user's type of personality on how he reacts to an advertisement. They decided to take, perhaps, the most popular typology of MBTI (Myers-Briggs Type Indicator), known since the middle of the 20th century. Many large western companies use MBTI tests when hiring or forming a team to work on projects.
But we are interested, of course, not with the user's readiness for teamwork, but with the influence of his personality type on the desire to click on the banner. Therefore, the question we explored is: “Can personality types influence CTR in advertising campaigns?”
In this article I will talk about how we did it.
')
Myers-Briggs typology includes four characteristics.
The combination of the characteristics described above gives the designation of one of 16 types, for example: JTSE.
To determine the type of user's identity, we built four classifiers, one for each of the characteristics.
For the training sample, we used information from one of our partner sites; on this site, the user fills in a questionnaire, which determines his personality type. The sample size is about ten thousand users.
The dependent variable in each classifier is a class of characteristics. For example, to characterize EI: I is a positive class (1), E is a negative class (0).
As independent variables, we used the history of user behavior on the Internet (pages visited) for the period preceding the test at the partner site. The address of each visited page is represented in the form of tokens: words are from three to ten characters long. For example, the address habrahabr.ru/company/dca/blog/260845 is converted to the following set of tokens: ['http', 'habrahabr', 'company', 'dca', 'blog'].
After that, all data are randomly divided into a training sample (37.5%), a sample for feature engineering (37.5%) and a test sample (25%).
The process of feature engineering is similar to the one we use for the hierarchical classification of sites, which is described in our previous article , however, I will describe it below.
For each token in the sample for feature engineering, we calculate the following characteristics
Next, we consider the metric of distributed grade (dg) for each token. We select twenty tokens with the highest value of this metric for each class. The result is forty signs. Attribute values are conditional probabilities of a token to belong to a class: true_number to total_number and false_number to total_number.
For classification, we used the Gradient Boosting Classifier from the scikit-learn library. To assess the quality of the classifiers, we analyzed the area under the ROC curve. The ROC curve is a graphic characteristic of the quality of a binary classifier. The curve reflects the dependence of TPR (true positive rate) on FPR (false positive rate).
where TP is true positive, FP is false positive, FN is false negative, TN is true negative.
The area under the AUC ROC curve (Area Under Curve) is a characteristic of the classification quality: the larger the AUC value, the better the classification model.
In the process of selecting the parameters, we managed to increase this metric from 0.63 to 0.77 using the grid search for the parameters n_estimators (number of trees) and max_depth (depth of trees). Table 1 shows the total values of the area under the ROC curve for each of the classifiers. And in the figure below, the ROC curves themselves are plotted.
Well, we got to the most interesting. Namely, before answering the question “Can personality types influence CTR in advertising campaigns?”.
To understand this, we analyzed data from one of our advertising campaigns. At the same time, no restrictions were imposed on the audience of this campaign. In total, over 89 million impressions were made as part of this campaign. For the analysis, we used about 30,000 unique users who clicked on the banner and 300,000 unique users to whom the banner was displayed at least once.
Further, we estimated the probability of a user to belong to one of the classes for each characteristic of our typology. Then, 10% of users with the highest probabilities and 10% with the lowest probabilities were taken. In each group, we estimated the CTR and built 95% confidence intervals for it using the Wilson score interval.
where n is the sample size,
where k is the number of clicks 
- alpha quantile of the standard normal distribution.
As can be seen from table 2 for the characteristics of EI and TF, the difference in ctr is more than 20% and is statistically significant. While for SN and JP characteristics, the difference in CTR is not statistically significant. Thus, there are advertising campaigns in which personality types affect CTR.
It is worth noting that this is a result for a specific advertising campaign, which is largely determined by what we advertised within it, the appearance of banners, etc. Perhaps in other campaigns, other characteristics would have a significant difference in CTR.
The difference in CTR in different classes by 20% inspires us to use the knowledge about personality types in advertising campaigns. In the near future, we plan to make eight user segments available for targeting advertising campaigns - two for each of the characteristics. In addition, we are faced with the task of learning to understand, prior to the launch of an advertising campaign, which type of personality information from it will be most interesting.
But this is not the only application we received segments. Information about personality types can be used in almost all areas of business.
For example, when conducting a large campaign for hiring employees (usually this is done by fast food chains, manufacturing enterprises and retail chains), you can target ads on people with certain characteristics. Thus, the conversion funnel is narrowed down at the entrance and the budget is significantly saved.
From the less obvious - knowing the psycho-type of users of the site, the company can adapt its appearance and content to better meet expectations and, as a result, increase sales.
In fact, any business whose work is based on interaction with a large number of customers through digital channels will be able to find the use of this knowledge to optimize their processes and build more personalized communications.
» Here is information for those who want to get acquainted with ROC curves more closely.
» Here you can learn more about the confidence interval we used (wilson score interval)
"And here you can read about the Myers-Briggs typology.
But we are interested, of course, not with the user's readiness for teamwork, but with the influence of his personality type on the desire to click on the banner. Therefore, the question we explored is: “Can personality types influence CTR in advertising campaigns?”
In this article I will talk about how we did it.
')
Myers-Briggs Typology
Myers-Briggs typology includes four characteristics.
- E — I - orientation of vital energy:
E (Extraraversion, extraversion) - to the outside world;
I (Introversion, Introversion) - on the inner world; - S — N - way of orientation in a situation:
S (Sensing, common sense) - targeting specific information;
N (iNtuition, intuition) - orientation to the generalized information; - T — F - decision making basis:
T (Thinking, thinking) - rational weighting of alternatives;
F (Feeling, feeling) - making decisions on an emotional basis; - J — P - a method of preparing solutions:
J (Judging) - preference to plan and pre-arrange information;
P (Perception, perception) - preference to act without detailed preliminary preparation, focusing more on the circumstances.
The combination of the characteristics described above gives the designation of one of 16 types, for example: JTSE.
Definition of personality type
To determine the type of user's identity, we built four classifiers, one for each of the characteristics.
For the training sample, we used information from one of our partner sites; on this site, the user fills in a questionnaire, which determines his personality type. The sample size is about ten thousand users.
The dependent variable in each classifier is a class of characteristics. For example, to characterize EI: I is a positive class (1), E is a negative class (0).
As independent variables, we used the history of user behavior on the Internet (pages visited) for the period preceding the test at the partner site. The address of each visited page is represented in the form of tokens: words are from three to ten characters long. For example, the address habrahabr.ru/company/dca/blog/260845 is converted to the following set of tokens: ['http', 'habrahabr', 'company', 'dca', 'blog'].
After that, all data are randomly divided into a training sample (37.5%), a sample for feature engineering (37.5%) and a test sample (25%).
The process of feature engineering is similar to the one we use for the hierarchical classification of sites, which is described in our previous article , however, I will describe it below.
Feature engineering
For each token in the sample for feature engineering, we calculate the following characteristics
- how many times the token met (total_number);
- how many times the token met, provided that the user belongs to a positive class (true_number);
- how many times the token met, provided that the user belongs to a negative class (false_number);
Next, we consider the metric of distributed grade (dg) for each token. We select twenty tokens with the highest value of this metric for each class. The result is forty signs. Attribute values are conditional probabilities of a token to belong to a class: true_number to total_number and false_number to total_number.
Characteristics of the classifiers obtained
For classification, we used the Gradient Boosting Classifier from the scikit-learn library. To assess the quality of the classifiers, we analyzed the area under the ROC curve. The ROC curve is a graphic characteristic of the quality of a binary classifier. The curve reflects the dependence of TPR (true positive rate) on FPR (false positive rate).
where TP is true positive, FP is false positive, FN is false negative, TN is true negative.
The area under the AUC ROC curve (Area Under Curve) is a characteristic of the classification quality: the larger the AUC value, the better the classification model.
In the process of selecting the parameters, we managed to increase this metric from 0.63 to 0.77 using the grid search for the parameters n_estimators (number of trees) and max_depth (depth of trees). Table 1 shows the total values of the area under the ROC curve for each of the classifiers. And in the figure below, the ROC curves themselves are plotted.
| EI classifier | 0.763 |
| SN classifier | 0.793 |
| TF classifier | 0.768 |
| JP classifier | 0.768 |
Check on real data
Well, we got to the most interesting. Namely, before answering the question “Can personality types influence CTR in advertising campaigns?”.
To understand this, we analyzed data from one of our advertising campaigns. At the same time, no restrictions were imposed on the audience of this campaign. In total, over 89 million impressions were made as part of this campaign. For the analysis, we used about 30,000 unique users who clicked on the banner and 300,000 unique users to whom the banner was displayed at least once.
Further, we estimated the probability of a user to belong to one of the classes for each characteristic of our typology. Then, 10% of users with the highest probabilities and 10% with the lowest probabilities were taken. In each group, we estimated the CTR and built 95% confidence intervals for it using the Wilson score interval.
where n is the sample size,
where k is the number of clicks - alpha quantile of the standard normal distribution.As can be seen from table 2 for the characteristics of EI and TF, the difference in ctr is more than 20% and is statistically significant. While for SN and JP characteristics, the difference in CTR is not statistically significant. Thus, there are advertising campaigns in which personality types affect CTR.
| Extraversion Introversion | 8.7 11.4 | (8.4, 9.0) (11.1, 11.8) |
| Sensing Intuition | 10.2 10.0 | (9.9, 10.6) (9.7, 10.4) |
| Thinking Feeling | 9.5 12.5 | (9.1, 9.8) (12.1, 12.9) |
| Judging Perception | 10.0 10.7 | (9.6, 10.3) (10.3, 11.1) |
It is worth noting that this is a result for a specific advertising campaign, which is largely determined by what we advertised within it, the appearance of banners, etc. Perhaps in other campaigns, other characteristics would have a significant difference in CTR.
What's next
The difference in CTR in different classes by 20% inspires us to use the knowledge about personality types in advertising campaigns. In the near future, we plan to make eight user segments available for targeting advertising campaigns - two for each of the characteristics. In addition, we are faced with the task of learning to understand, prior to the launch of an advertising campaign, which type of personality information from it will be most interesting.
But this is not the only application we received segments. Information about personality types can be used in almost all areas of business.
For example, when conducting a large campaign for hiring employees (usually this is done by fast food chains, manufacturing enterprises and retail chains), you can target ads on people with certain characteristics. Thus, the conversion funnel is narrowed down at the entrance and the budget is significantly saved.
From the less obvious - knowing the psycho-type of users of the site, the company can adapt its appearance and content to better meet expectations and, as a result, increase sales.
In fact, any business whose work is based on interaction with a large number of customers through digital channels will be able to find the use of this knowledge to optimize their processes and build more personalized communications.
useful links
» Here is information for those who want to get acquainted with ROC curves more closely.
» Here you can learn more about the confidence interval we used (wilson score interval)
"And here you can read about the Myers-Briggs typology.
Source: https://habr.com/ru/post/263729/
All Articles