In last article, we showed you how Prediction One learned which keywords are most effective for improving click-through rates,CTR, based on the data we were given.

In this article, we will use the learning results to predict the click-through rate,CTR, for new keyword combinations.

Needless to say, this software won’t find the magic keywords that will increase your click through rate.

Using the CSV data, It is only to find out what keyword combinations that would contribute to a higher CTR.

Now, let’s prepare a CSV file with the actual keywords we want to look at.
In lines 2, 3, and 4, I put three different keyword combinations that I think will contribute to a higher click-through rate.

The number of clicks and the number of displays is left blank because there is no data at all.

Let’s see if Prediction One can predict in this situation.

Simply drag this file into the “Open File ファイルを開く” field in the “New Forecast　新規予測” section and the forecast results will appear.
(in advance, clicking on the “Add Reason for Prediction　予測理由を追加” button).

And here is the result. ↓

Keyword samples that increase the click probability are displayed.

To be honest,……The results are not clear.

In my case, when I clicked on it, there were only about 110 posts and about 300 keywords.We randomly selected keywords with a high contribution rate and ran the prediction.

In general, it is said that machine learning requires thousands to tens of thousands of sample data.
In this case, the amount of data was insufficient to obtain significant results.

In addition, it cannot be said that the sample preprocessing (quality) was appropriate because the keywords that we wanted to predict were simply extracted, the contents were not taken into consideration, and there were also data deficiencies.

Vice versa, machine learning requires a certain number of good quality sample data, and data pre-processing that is appropriate for machine learning.

I’m sure some of you will be a little disappointed to hear this.

Still, it was good to be able to actually understand part of machine learning, especially the number of data and the importance of data preprocessing.

In machine learning, as is often said, “Garbage In Garbage Out.”

You can only get useful results from useful data.

Moreover, if the data is not preprocessed properly, you will not get useful results.

In this example, the view / click count for a single keyword plus the view / click count for a combination of multiple keywords was essential.

Also, You can predict useful keyword combinations by adding keyword data that you have never seen before.

Unfortunately, Prediction One doesn’t seem to be able to predict keywords that you didn’t use in your training. Therefore, if you want to predict freeword clicks, you need to add a sample of all the keywords you can find.

However, there seems to be no other software that can perform AI analysis in such a simple way, so I hope Sony will do its best to release “Prediction Two” for free.