Loading and Understanding the Data

There are a number of variables included in this dataset:

• Last name - the last name or last initial of the student
• First name - the first name of the student
• Student ID - the student’s ID number
• Submission - the text content of the post that was submitted
• Type - label indicating if the submission was a comment or question
• Score - the automatically assigned score by Perusall from 0-2
• Created - the date and time of the submission
• Last edited at - the data and time of the last edit of the submission
• Number of replies - the number of submissions in the same thread after the submission
• Number of upvoters - the number of times the submission received a check mark or upvote
• Status - if the submission was submitted before the deadline or not, on-time or invalid
• Page number - the page number of the paper where the submission can be found
• Range - the highlighted portion connected with the submission, from the start of the highlight to the end
• Paper - which paper the submission is for

Examining the Data Distribution

Within this data set, there is some missing data such as Student ID in a few instances. There are also inconsistencies in the formatting of Student ID (eg. jz1220 vs N19945522). There were also inconsistencies in the formatting of Last name where it was sometimes the initial, the full last name, or even a middle initial with the full last name. Fortunately, much of what we did with the data did not require these elements so there should be little detrimental effect.

According to the data set, we completed 11 readings all together with McNamara having the most submissions and Saqr having the fewest. Interestingly, the Saqr was the reading due before Easter weekend and the McNamara was the reading due after Easter weekend.

Not surprisingly, more discussion took place beyond the questions being asked. Out of the 830 submissions, 654 of them were comments while 176 of them were questions.

The submissions would consistently occur most frequently in the first seven pages of the reading and dropped off significantly after.

Tokenizing the Text

After tokenizing the text per the instructions, we found some additional columns, specifically document_id, Token, and sentence_id. After completing this operation, the data became exceedingly longer, now providing an individual data entry for each word in each submission. If it came to rebuilding those submissions, you would mostly be able to aside from information such as punctuation and capitalization that is removed during the operation.

After this operation, the most frequent words which many of the filler words such as “the”, “to”, “of”, “and”, “a”, etc.

Cleaning the Bag of Words

After cleaning the bag of words further with the next operation, many of those filler words had since been removed. The left many of words that dealt with the content for the course including “data”, “student”, and “learn”. This is not surprising given a course of “Learning Analytics” where we specifically discuss students, data, and learning. Without seeing exactly words were removed, I assume that we didn’t lose significant information from this process. However, if there was significant content related to one of those removed words, it would be much more problematic. (What if it were a course about language with an emphasis on prepositions?)

Creating Bi-grams and Tri-grams

By processing the data for bi-grams and tri-grams, we see some changes in the table. Specifically, the Token column has been replaced by token with the different permutations of monograms, bi-grams, tri-grams. There is also a new gram column to indicated the number of “grams” in the token for each data entry.

We now see that the the most common bi-grams are learn_analyt, predict_model, and make,sens.

We also find that the most common tri-grams are design_base_research, use_predict_model, and educ_data_mine.

Bi-grams and tri-grams are significantly less common than monograms with the largest monogram having 362 entries, the largest bi-gram having 62 entries, and the largest tri-gram having only 4 entries.

“Student”, a common word encountered in our course has a wide range of bi-grams and tri-grams associated with it throughout the papers.

Important Words for Papers and Authors

The content of each paper had its topic made pretty clear once the analysis was completed. Brook discussed prediction models, Carolan dealt with social network analysis and so forth.

Looking at my own comments, I was surprised while also unsurprised with the frequency in which I used the word “language” in my comments. I suspect that would tie relating the course content to my experience as a language teacher.

Cluster the Authors

Despite being in three dimensions, the graph does make sense in you can expect some similarities between different authors while other authors stick out. Of note, you can see that Xavier is far displaced from the rest of the group. Also, Elizabeth and I are super near eachother suggesting that we share commonalities in how we interact and communicate through our comments.

When placed clustered and placed into two dimensions, we see some interesting groupings.

• Xavier is far from the students, which is unsurprising given that he’s the instructor and will interact in the forums much differently than us.
• Elizabeth and I are clustered as well, as seen in the above graph.
• Maria is also far from nearly everyone else.
• There is a cluster including Taniya, Evgeniya, Ailin, and Collin. This is an interesting cluster that’s hard for me to describe. I could see commonalities in their personality, but I’m not sure how that would manifest in their forums interactions.
• The final cluster has the majority of students. Again, it’s hard to pin-point the exact similarity between the group, but maybe there is a cultural link that has an impact in the language that they write. This is conjecture and I really can’t be certain.

Sentiment Analysis

I ran sentiment analysis on the data set as well and found some interesting things. Firstly, our distribution trended positive as a class overall.

The most positive submission across all students and papers was Ruby in the Ochoa writing: > I think that having a personalized prediction to guide a student’s learning could be efficient in the way that the student’s learning are tailored to their individual needs and not comparing with others. Doing so presents a more “efficient” way of learning when comparing the student without personalized learning to the same student with personalized learning.

The most negative submission across all students and papers was Ana in the Selwyn writing: > It is also possible that data become an excuse or scapegoat for police failures.

Given these examples, I could see how the calculation would determine that a submission was positive (“efficient”, “not comparing”, “personalized”) or negative (“excuse”, “scapegoat”, “failures”). I think that this analysis doesn’t go further than determining the positivity or negativity of a message, but does not determine the author’s positivity, negativity, or anything else meaningful. It’s just the language being used.