DEV Community

Cover image for A New Way to Recommend Video Games at Amazon Part 3
Jinhoon Chung
Jinhoon Chung

Posted on

A New Way to Recommend Video Games at Amazon Part 3

This is part 3 of the whole post. Please refer to the links below.

part 1
part 2

Intro

In the previous parts, I built a new recommendation system and shown you if the system has value or not. Let's try to display a more personalized list of recommended games for a customer.

Here is a graph of the numbers of recommended games by predicted ratings.

many rating 5

This is one of the critical cases where one customer would be recommended with above 14,000 games with the predicted rating equaling 5. Amazon would love it if this customer would buy all of them, but it is not very realistic. Let's try to trim down this long list.

Similarity

One of the straight forward ways to trim down the long list is to use the game information and find similarities between different games. Examples of information are console information and company information. Unfortunately, the metadata does not contain genre information.

game_information

Do you see some familiar names? Let's pick one game.

selected_game

This is a famous game owned by Capcom. This game runs on PC, Playstation, and Xbox, but the selected game is for the Xbox console. With a matrix calculation that compares 72,000 game information to the selected game, I can get values from 0 to 1 where 1 means the perfect match, and 0 means no information match.

0_to_1

If you look at the third row from the bottom, you can see there are 16 games that have the perfect match with the selected game. However, it is actually 15 because one of 16 is the selected game. During matrix calculation, the selected game was compared to all games including itself.

Trimming the List

The following step is simple. The similarity ratios are multiplied by the predicted ratings. This is possible because each game is given a similarity ratio and a predicted rating. Even with a high predicted rating, the game would be ranked lower if the similarity ratio is low.

Let's check out the final list of the top 10 recommended games.

top_10

You can observe other series of Resident Evil. Please ignore the same name as the selected game located in the 4th row. I know that "DMC Devil May Cry", "Strider", and "Mega Man" are from the same company. Except for "Strider" and "Mega Man", all games are 3D graphics.

Here are the final predicted ratings after multiplying by similarity ratios.

final_ratings

The first six games still have the perfect predicted rating. The predicted rating starts to be smaller from the 7th game. I do think these numbers would look more interesting if the metadata has genre information because one game has multiple genres.

Final Note

I personally enjoyed this project because the chosen topic was motivating to me. There are several elements that can be improved. If all of these ideas are finalized, then the next steps are to discuss where and how to put this information for which customers and if these ideas can be applied to the different departments. This was a project completed for the data science Bootcamp program, but I have tried to work it out as if this is a real problem.

Thank you for reading.

Project GitHub Link

Top comments (0)