Figure 10

Ternary scatter plot of normalized absolute error for feature group importance for different genres. For each dot, the three values are the normalized absolute error generated by Learning to Place with only the corresponding feature group. We color each book based on the actual sale category of that book, where low is the lower 30th percentile, middle is between 30th to 80th percentile and high is the top 20th percentile. For all ternary plots of (A) fiction genres and (B) nonfiction genres, the densest area is the top corner, meaning that with only book feature the model generates the highest error for those books, implying author and publisher information are very important. The second densest area is the left corner, meaning that publisher feature is not sufficient for accurate prediction. The third densest area is the middle of the triangle. Interestingly, we see that most dots in the middle area are books with high sales, meaning that for high-selling books, the importance of three feature groups are rather balanced