Using the data created for this demonstration, let's find out the features of the winners in the Wide Learning™ Trial Tool.

In order to use the Wide Learning™ Trial Tool to predict the outcome of an election, the following two data sets are necessary. Please download the linked files and use them.

The characters, organizations, names, etc. that are used here are fictitious and have nothing to do with actual entities.

Election Case Study - Data for learning: Data for Wide Learning™ to learn from (election_train.csv)

Election Case Study - Data for prediction: Data for Wide Learning™ to use for prediction (election_test.csv)

The data for learning contains various information in the form of a table, including that of 219 winners and 456 losers in a fictitious election, for a total of 675 people. If the solution label, which is the "Elected" column, is 1, it means that person was elected, and if it is 0, it means that person lost.

For example, the first line of the data for learning contains the following information about Taro Fujitsu.

The detailed content of the study data can be found in the "1 -3. Check the data for learning" section of the Wide Learning™ Trial Tool.

In order to use Wide Learning™, data for learning must be input into the AI. Upload the data for learning for the election case study you just downloaded to the Wide Learning™ Trial Tool to see what it contains.

When you upload data for learning, the contents of the data are displayed in a table. One row represents the data for one candidate.

You can check the statistical information of the data for learning from "1 -3. Check the data for learning". For details, see "Let's look at the statistics" in Identifying Mammals.

Click the "Generate combinations of important items" button. (Generation takes some time)

When generation is complete, Wide Learning™ will display the combinations of items that have been determined to be important (useful in determining a candidate's election prospects) in a table.

Each line displays the details of a single combination. The leftmost column, "Combinations of important items", indicates the combinations of important items found by Wide Learning™.

Let's get the AI to learn from the combinations of key items we have found. Press the "Weight important combinations" button.

When learning is complete, you can see how much weight is attached to the important combinations you have identified.

Here, "Weight" is a numerical expression of how important a combination of items is in determining whether or not a candidate is elected.

POS (= Positive) combinations are an important basis for determining the reasons a candidate has won.

On the other hand, NEG (= Negative) combinations are an important basis for determining the reasons a candidate did not win (= Lost).

Check that POS is selected in the upper right corner of the table in "3 -1 Overall Learning Results".

In POS, you can check the combinations of features that contribute to candidates winning. What combinations of items are there that are POS?

The combination "Alma Mater_University C ∧ Faculty_Economics" shows that many people who attended the Economics faculty of University C advance into politics.

The combination "Advocates_Healthcare_Yes ∧ Advocates_Welfare/Nursing_Yes ∧ Advocates_Pension_Yes" suggests that voters are interested in issues related to the elderly and pensions after retirement.

Also, the combination "Age>= 60.0 ∧ Number of Times Elected>= 3.0" indicates veterans are often winners.

By looking at all the combinations using AI, you can also find features that consist of combinations of seemingly unrelated items.

Click "NEG" at the top right of the table. Next, you can look at the combinations that contribute to the determination of NEG.

You can see the combination "Block_North Kanto ∧ Advocates_Child Support_No". This may be due to voters in the North Kanto block having a strong interest in children and child support.

"Block_Tohoku ∧ Party_B Party" shows that the approval rating of the B Party is low in the Tohoku region.

There are also many items with the form "Block_Shikoku ∧ zzzz", such as "Block_Shikoku ∧ Advocates_Differing Family Names_No", "Block_Shikoku ∧ Advocates_Healthcare_Yes", and "Block_Shikoku ∧ Number of Times Elected>= 3.0". This suggests that the winning and losing patterns in the Shikoku region may be different from those in other regions.

Click "3 -2. Individual Learning Results".

Here, for each piece of the data for learning, you can check what kinds of combinations contribute to POS and NEG and whether they prediction a win or a loss in an election.

Let's take a look at the results of Mr. Hiroto Tanimoto. Hiroto Tanimoto has actually won the election, and the AI computation results consist only of pink items, which represent the positive factors. So let's check the combinations of important items that were the basis of the prediction.

Hiroto Tanimoto is expected to win the election because of his important combinations of "Age>= 60.0 ∧ Number of Times Elected>= 3.0" and Advocates_Healthcare_Yes ∧ Advocates_Welfare/Nursing_Yes ∧ Advocates_Pension_Yes". These combinations are actually the important combinations for winning the election that we confirmed earlier.

Next, let's take a look at the results of Takahiro Sano. Takahiro Sano has lost the election, and you can see he has numerous blue combinations, which signify NEG. Now, click "NEG" and look at the combinations of important items that will help you predict failure. The important NEG combination of "Block_North Kanto ∧ Advocates_Child Support_No", which we confirmed earlier, is in the list.

In this way, Wide Learning™ can be used to identify combinations of important items that can be used as a basis for winning or losing an election.

Please check the reasons for other candidates who won or lost.

You can review the results and scores for candidates in the prediction data in "5 -1. Individual prediction results". When the score exceeds 0.5, the result is POS (The candidate is likely to win the election).

In this case which involves Kanagawa Prefecture, as high chance of winning the election = order of high scores, Katsumi Ishihara and Sawa Taniguchi have the highest odds of winning the election. By selecting each candidate in the same way as in the data for learning, you can confirm the combinations of features that will be the basis of the prediction of winning or losing.

In this case study, we used data that was created specifically for this case, but it is also possible to perform the same analysis on real data. A prediction of the winners and losers of the 2019 Japanese House of Councilors election performed using Wide Learning™ can be viewed at the following link.

A proposal of an AI-based national election forecasting framework