First Coursera Course (Recommender systems)

This fall I signed up for Introduction to Recommender Systems on Coursera.  I'm really appreciating the review/introduction because my thesis is generally related to the field of recommender systems.

While working on one of the first assignments, I remember feeling anxious because I knew there was a "peer evaluation" component with a specific window of time in which I would have to look at peer assignments.  I hunted all around the site and I couldn't find the dates to mark my calendar to remind myself that I would have to pay attention.  I concluded that the dates hadn't been announced yet.  "Oh well," I thought, "I'm sure they'll send me an email or something."  So, I submitted my assignment and I trusted that someone would tell me when I had to come back to do the peer evaluation part.  Big mistake!  I logged in recently only to discover that I missed the deadline.  So, my assignment score was lowered because I did not evaluate any peer assignments.  Gah!  That is discouraging. 

Oh well.  I'm not taking the course for credit anyway.  ;)

Here are some of my notes on Module 1.

• recommender systems, use of "persistent preferences"
• "information retrieval" 
• where content base is relatively static, invest in indexing
• "term frequency inverse document frequency"(weighting measure)
• how often the term appears in a specific document vs. the rest of the corpus  
• we don't care about a term if it appears everywhere, while we care a lot about a term if it's rare
• "information filtering"
• preferences are static but the content is dynamic
• invest in creating user models
• "collaborative filtering"
• where preferences are more complex than just keywords
• keywords, annotations, referrals
• recommendation vs. prediction 
• one is an actual calculation of something we expect they will like
• vs. an estimate of what their rating would be on a new item.  So a prediction will quantify ANY item, even if it looks like they will dislike it a lot; the recommendation has an implication that they will like it.
• "ephemeral personalization" is based on current browsing activity 
• explicit vs organic
• refers to the presentation of the items, i.e. "Recommendations for You" (explicit) vs. strategically placed items on the page or adjusted navigation (organic)

How is this related to my area?

• "preferences" are akin to the "user model"
• I haven't been looking at keywords at all -- only annotations, referrals, outcomes.  However, I do consider prerequisite relationships.
• When I create a recommendation of sequences, this could either be used explicitly (i.e we suggest first you do this time, then this, then this) vs. organically (here is a trail we have laid ahead, feel free to go in that general direction or not...)

Here are some of my notes on Module 2.

• we did an assignment in Excel where we we calculated "movies that most often occur with a given movie, M" using the "x + y / x" method.
• For example, to calculate the relative occurrence of another movie N, with movie N, just count the number of people who saw BOTH movies and divide by the number of people who saw the first movie. 
• gives a number to represent how much the items "occur together".
• banana trap
• what was that again?  Making a useless (but correct)  recommendation because everyone buys that anyway?