Data analysis of information flow audience portrait. If you can understand it, I lose!

I have been doing advertising optimization for a long time and have seen many audience portraits in the advertising backend. Generally speaking, their reference value for advertising data analysis and effect optimization is limited, but it is better than nothing.

The reason is that the audience portrait data in many advertising backgrounds only tells us what the people who saw the advertisements are like, but lacks the portrait data of the users who converted . There are two main reasons:

First, in most advertising processes , the front-end and back-end data are separated. In other words, the media can know who saw the advertisement you paid for, but generally do not know who converted. Party A can know which part of the converted users are through its own monitoring. If the monitoring is done well enough, it can also know the population portrait of this part. However, the judgment criteria of the population portrait may differ from that of the media, the statistical caliber is inconsistent, and the data cannot be manually connected.

Second, the media is unwilling to disclose so much data, and even the audience portrait itself has certain problems.

Nowadays, information flow optimization has become a hot topic in the industry. Optimizing creativity and targeting has become a commonplace, but data analysis of audience portraits is rarely mentioned, so there is still room for exploration. Today I would like to take this opportunity to share with you an idea for audience data analysis.

It should be emphasized that the following advertising data analysis has a basic premise: assuming that the data provided by the media and the data monitored by Party A are true and accurate. Below, the editor will use a real case and data ( Toutiao , home improvement category) to introduce to you how to use the naive Bayes algorithm to conduct data mining and analysis on the audience portrait of Toutiao, so as to achieve conversion rate prediction under precise targeting.

Every time I mention Bayes' theorem, I feel a sense of reverence in my heart, not because the theorem is so profound, but because it is particularly useful. This theorem solves a problem that we often encounter in real life: given a certain conditional probability, how to get the probability of two events being exchanged, that is, how to obtain P(B|A) when P(A|B) is known. For example, I know that among users who convert, 36% are women. So when a female user sees my ad, what is the probability that she will convert?

Here we first explain what conditional probability is:

P(A|B) represents the probability of event A occurring under the premise that event B has already occurred. It is called the conditional probability of event A under the occurrence of event B. The basic solution formula is:

The reason why Bayes' theorem is useful is that we often encounter this kind of problem in life: we can directly derive P(A|B) without much effort, but P(B|A) is difficult to derive directly. However, we actually care more about P(B|A). At this time, Bayes' theorem provides us with a way to obtain P(B|A) from P(A|B).

The proof process is omitted below, and Bayes' theorem is given directly. I believe that friends who still remember high school mathematics should be familiar with this formula:

The following is a simple example to introduce the data mining principle of Naive Bayes. Although the sample size is not large, it is enough to illustrate the principles and ideas.

Here is a set of audience portrait data, with a total of 20 data, representing 20 UVs. There are 9 entries with the value of 1 in this form field, which means that 9 users have converted.

Table 1

Then, we make a distribution of the proportions of other fields except ID (just a number, which has no value for mining) and provincial region (because they are all Guangdong, which has no value for mining), as shown in the figure:

Table 2

Suppose, I want to know the conversion rate of targeting X=(prefecture-level city="Foshan", gender="male", age="18-23 years old", interest category="real estate"), that is, I want to find: P(fill in the form="1" | X).

It cannot be calculated directly. Let’s go back to the Naive Bayes mentioned above. It is specifically designed to solve this kind of problem. I only need to know P(X | fill in the form = "1"), and I can get P(fill in the form = "1" | X) through the formula.

The specific direct formula is:

P(filled out form="1" | X) = P(X | filled out form="1") * P(filled out form="1") / P(X)

Similarly,

P(filled form = "0" | X) = P(X | filled form = "0") * P(filled form = "0") / P(X)

Another important formula needs to be introduced here. P(A,B) represents the probability of events A and B occurring at the same time.

When events A and B occur independently, P(A,B) = P(A|B) * P(B) = P(A)P(B).

Because the occurrence of fields (or directions) such as prefecture-level city, gender, and age can be understood as being independent of each other, P(X | fill in the form = "1") = P(X) * P(fill in the form = "1"), and P(X | fill in the form = "1") = P(prefecture-level city = "Foshan", gender = "male", age = "18-23 years old", interest category = "real estate" | fill in the form = "1") = P(prefecture-level city = "Foshan" | fill in the form = "1") * P(gender = "male" | fill in the form = "1") * P(age = "18-23 years old" | fill in the form = "1") * P(interest category = "real estate" | fill in the form = "1"), at this time, P(X | fill in the form = "1"), which seems to be equally impossible to obtain directly, is split into the product of the probabilities of 5 seemingly simpler events.

Substitute the specific values ​​and calculate:

P(fill in the form = "1" | X) = P(X | fill in the form = "1") * P(fill in the form = "1") / P(X) = (3/9 * 6/9 * 2/9 * 1/9)*0.45 / P(X) = 0.002469 / P(X)…………………………………………①

P(fill in the form = "0" | X) = P(X | fill in the form = "0") * P(fill in the form = "0") / P(X) = (2/11 * 5/11 * 2/11 * 2/11) * 0.55 / P(X) = 0.0015026 / P(X)………………………………………………②

Next, we encounter a problem: we don’t know what P(X) is! But it doesn’t matter. When a user targeting X comes in, he or she will either convert or not.

P (fill in the form = "1" | X) + P (fill in the form = "0" | X) = 1………………………………………………③

Combining ①②③, we finally get:

P(filled form="1" | X) = 62.2%

P(filled form = "0" | X) = 37.8%

Therefore, when the orientation is X, the Naive Bayes data mining model believes that the conversion rate of this type of users is 62.2%.

Mainstream data mining algorithms, such as neural networks, decision trees, etc. Most of them rely on the data shown in Table 1. Each field represents a different dimension of the user, and each row represents the data of an independent user. However, in the actual optimization process, it is impossible for the media to provide such detailed audience portrait data. However, Naive Bayes is different. The requirements for original data are slightly lower. It only needs to provide the proportions under different dimensional combinations without having to break down the situation of each user.

1) Raw data and preprocessing

After simple processing, the data we obtained from the Toutiao advertising backend is as follows:

A total of 13,339 clicks and 37 conversions.

Table 3

2) Calculate the importance of fields and determine the input fields

Because all fields are categorical fields (as opposed to numeric fields), here is a more general algorithm for evaluating the importance of all possible input fields to the output field.

The formula is interpreted as follows:

(Note: ABS function is used to find the absolute value)

The importance of all possible input fields to the output field is calculated as follows:

As a general rule of thumb, fields with an importance less than 0.1 may not be included in the data mining model.

Therefore, based on the limited data currently, the "gender" field is not very helpful in determining whether a user converts. Therefore, in the following data mining model, the input fields include: prefecture-level city, age, and interest classification.

Start doing data mining. The specific principles will not be repeated here, and the results will be given directly.

. . . (The middle part is too long, so it is omitted)

We can see that data mining shows that the maximum probability of conversion to 1 is 1.51%, and the directional condition at this time is "Guangzhou & (over 50 years old) - mother and child". The overall conversion rate of the sample data is 37/13339 = 0.28%.

The figure below is the probability distribution of conversion to 1. It can be seen that about 25% of the data is greater than 0.28%. In other words, based on user data mining based on Naive Bayes, about a quarter of the 324 directional combinations were found to make the estimated conversion rate higher than the average level of 0.28%. As the amount of data increases, the model will continue to improve, and the conversion rate prediction performance of precise targeting combinations will become better and better, allowing limited advertising costs to be spent on users who are most likely to convert.

Next, consider how to apply this insight to advertising and create a higher ROI. For example, creating targeted creatives, raising bids, etc. You are all veterans in this area, so I won’t go into details here.

Finally, I would like to emphasize that data mining of audience portraits needs to meet certain conditions, that is, it must be able to distinguish between converted and non-converted users.

The above is user data mining based on Naive Bayes for your reference.