Thank you!

[angelhsu05]

Hi Lisa, thank you for putting together this game. I had the pleasure of testing it out with some grad students. A few tweaks I made (and please let me know if I was off):

• I didn't have 1.5/2 hours for them to do the exercise, so I shortened it to 5 (instead of 10) topics/types of produce. Some of the students asked whether the produce types could overlap, I said "yes", since it mirrors LDA in the wild. I think making that clarification might be helpful.
• I told the "MALLET"ers (in my case, since I had them implement LDA using R, I called them the "topic modelers" but I'm sure there's a more elegant way of describing their role) that they were to come up independently with the topics without necessarily looking at what the shoppers were doing. This wasn't clear in the instructions, but it was interesting to see ways in which the modelers and the shoppers diverged/converged.
• I let some of the shoppers come up with their own produce ... some of them got very specific into different varietals of produce. In the end, it ended up in some very small word probabilities.
• I had them calculate: topic-word probabilities - what's the likelihood a word appears in a topic? This required the team to label each produce (aka "assign" them) to a topic and then calculate; document-topic probabilities - what's the likelihood a document reflects a certain topic? document-term probabilities - what's the likelihood a word appears in a document. I liked the questions you provided but since I was having them calculate these specific probabilities implementing LDA in R, I wanted them to get a better intuition for how these metrics were calculated.