Assignment 4 -Jewelina 

1.b What information is being captured by Perusall in this dataset?

The Perusall dataset captures detailed interactions of students with course materials, including comments, responses, and highlighted texts. It includes identifiers for comments and documents, personal identifiers for students (such as names and IDs), and engagement metrics like word count, upvotes, and replies. This information allows for the analysis of student engagement, the effectiveness of materials, and peer interaction within the educational platform

i. Are there missing data?

The histograms provided do not directly show missing data. however. i found missing values in specific columns such as "Last Name" and "In Response to Comment ID."

ii. Are there more comments or questions?

From the 'Comment vs. Question'  bar plot, we can see the distribution of comment is more frequent.

iii. Which papers receive more comments?

The "Introduction to Learning Analytics" and "Introduction to Information Visualization, Chapter 1, Introduction" receive the most comments, indicating high engagement or possibly contentious topics in these papers.

iv. In which pages most of the comments happen?

Most comments occur on the first few pages, specifically from pages 1 to 6.40, indicating that students are most engaged or have more queries and discussions at the beginning of the documents.

v. Who were the most prolific commentators?

"Maximo Perez" is the most prolific commentator, followed by "SarahD_NA" and "Jane_Li".

i. What new columns are added to the dataset?

"Tokenized Text": This column appears to be the new column added to the dataset, which contains each word from the "Submission" column broken down into separate tokens (words).

ii. The structure of the data is wide or it is long?

Long Structure: The data structure has transformed into a long format because each word from the original "Submission" is now placed in its own row under the "Tokenized Text" column, while repeating other identifying information (like "document_id", "sentence_id", "id", etc.) for each word.

iii. Could you rebuild the original text from the new columns? How?

Yes, i could potentially rebuild the original text. To do so, i  would group the rows by a combination of "document_id", "sentence_id", "id", or any other identifier that ensures the grouping of words that belong to the same original sentence or submission. Then, concatenate the words in the "Tokenized Text" column in their order of appearance. Here’s a conceptual way to do it using Python:

iv. What are the most frequent words?

Top Frequent Word:that

Other Frequent Words: Other notable words with high frequencies are visible throughout the chart. Words like "data," "process," and "students" appear to be among the more frequently occurring, indicated by relatively taller bars compared to the rest.

d. Which are the most frequent words now?

The chart shows that certain words like "data," "students," "information," and "learning" are among the most frequent. These words have taller bars, indicating higher frequency relative to other terms in the dataset.

e. Does this list make sense given the topic of our course?

Words like "data" and "students" are central to educational analytics or any course that deals with data and learning environments. The prominence of these terms aligns well with such topics, indicating the text analysis is reflecting key course themes.

c. i. Analyzing bi- and tri-grams

From the pivot table for bigrams (gram=2), the most common bigrams include:

data_collect (56 occurrences)

data_analysi (38 occurrences)

learn_analyt (81 occurrences)

predict_model (42 occurrences)

These terms suggest frequent discussions around data collection, data analysis, learning analytics, and predictive modeling, which are typical in data-driven or analytics-focused courses.

c.ii. What are the most common trigrams?

From the pivot table for trigrams (gram=3), the most common trigrams include:

use_learn_analyt (6 occurrences)

learn_analyt_softwar and type_learn_analyt (5 occurrences each)

learn_analyt_tool, social_media_platform (5 occurrences each)

These trigrams reflect more specific topics or methods within learning analytics, such as the use of learning analytics software, types of learning analytics, and tools used, indicating a deeper layer of discussion or analysis.

c.iii. Are bigrams and trigrams common compared with simple words?

The total counts in the pivot tables show that bigrams and trigrams, while informative, appear less frequently than simple words. The bigram table shows a total of 334 occurrences, and the trigram table shows 46 occurrences. This suggests that while n-grams capture more nuanced or specific information, individual terms or simple words occur more frequently overall.

This frequency distribution is common in text data, where simple words are more abundant but n-grams provide richer context or more specific information, which can be crucial for in-depth analysis.

d.i. What are we the most common bigrams and trigrams with the word student?

bigrams:student_data

trigrams: student_learn_progress

6.vi.Are the most relevant words, used in the comments, relevant for each one of the papers? Why?

Contextual Significance: The tokens are likely specific to the themes or topics covered in each paper. For instance, terms like collect_done and survey_think suggest topics related to data collection and survey methodologies, which would be relevant if the paper discusses these areas.

Specificity and Uniqueness: The TF-IDF metric ensures that the words are not only frequent within a particular paper but also relatively rare across other papers. This uniqueness enhances the relevance of these tokens to their respective papers, as they capture key concepts or discussions unique to each.

Inferential Analysis: Without the actual content of the papers, a precise validation of relevance is challenging. However, the TF-IDF approach typically identifies terms that are critically tied to the main arguments or content of the papers. For example, terms forming n-grams like data_collect_done or survey_type_interact are likely central to the discussions within those papers, indicating methodological discussions or specific case studies.

6.iv.1 Can you recognize the most important words that you have used?

the most i used is cours_design

6.iv.2 What is the most common word that you have used, that is least used by the rest of your classmates?  Does it make sense to you?

no，yes it make sense to me

7.f i. Does this graph make sense?

Analysis of the Graph:

Dimensional Representation: The graph plots authors in a three-dimensional space, where each axis (X, Y, Z) corresponds to one of the three main components derived from the singular value decomposition (SVD) of the TF-IDF matrix. This reduction technique helps in visualizing high-dimensional data in a more comprehensible form.

Clustering of Authors: The positioning of the points (authors) in the plot suggests how similar or dissimilar they are in terms of their word usage. Authors whose points are closer together are likely to use similar significant words in their comments, whereas those further apart use different sets of words.

Isolation and Grouping: Some points are isolated, and others are grouped. Isolated points could represent authors with unique styles or subjects of discussion that are distinct from their peers. Conversely, grouped points indicate authors with similar thematic or stylistic content in their messages.

Does the Graph Make Sense?

Yes, this graph makes sense as it provides a visual representation of the relationships between authors based on their lexical choices. By reducing the dimensionality to three principal components, we can see patterns that might not be obvious in the original high-dimensional space.

The spread and clustering can help identify which authors have similar interests or writing styles, which can be valuable for further analysis such as grouping authors into communities or understanding the diversity of topics discussed within a group.

7.h.Similarities with the Previous Graph

Preservation of Similarities: The 2D plot using Principal Component 1 (PC1) and Principal Component 2 (PC2) appears to preserve some of the groupings observed in the 3D scatterplot. Clusters of authors that were close in the 3D space are likely represented by proximity in the 2D space as well.

Cluster Formation: The clustering algorithm has identified distinct groups, labeled as Cluster 1, Cluster 2, and Cluster 3, which suggest that despite the reduction from three dimensions to two, the essential relationships and distances between points (authors) are maintained to a reasonable extent.

7.h.Interpretation of the Graph

Distribution of Clusters: The graph shows that most authors fall within a central cluster with slight variations along the PC2 axis, which could represent a commonality in terms of the vocabulary used across many authors. Two outliers (Clusters 1 and 3) indicate unique usage patterns distinct from the main group.

Insight into Variance: The axes percentages (33.3% for PC1 and 33.3% for PC2) suggest that these two components capture a significant portion of the variance in the dataset, which is good for interpreting the clustering results as they represent a substantial part of the data's structure.

Implications for Further Analysis:

Authors in the central cluster likely discuss similar themes or use similar language, which could be further explored to understand common topics or sentiment.

The outliers may represent authors who either focus on niche topics or have a distinctive style, which might warrant separate detailed analysis to understand their unique contributions or perspectives.

8.b.i. Which paper receive the most positive / negative comments?

Most Positive Comments: The paper titled "Srinivasa 2021" seems to receive the most positive comments overall, as indicated by the significant proportion of green bars (high positive sentiment).

Most Negative Comments: The paper titled "Benoit 2019" appears to have received the most negative comments, as it shows the most extensive part of the bar in red, indicating negative sentiments.

8.b.ii. Which student in average was the most positive / negative according to the analysis?

Most Positive Student: "Jewelina, Wen" (hhhh that is me !) exhibits predominantly positive sentiments across her comments, as evidenced by the long green bar in her section.

Most Negative Student: "Christina, Rodriguez" shows a considerable portion of red in her sentiment bar, suggesting she has the most negative sentiment on average.

8.b.iii. Do you think the sentiment analysis is accurate? Why?

Reliability and Limitations: Sentiment analysis algorithms, like the one presumably used to generate these charts, typically rely on the polarity of words to determine sentiment. This method can be effective in general contexts but may not always capture the nuances of sentiment accurately, especially in academic or technical discussions where the connotation of words might differ from common usage.

Context and Subjectivity: The accuracy of automated sentiment analysis can be limited by the context and subjectivity of language. For instance, technical discussions might use words that have neutral connotations in general language but carry positive or negative sentiments within the specific academic context.

Cultural and Situational Factors: Sentiment analysis may not account for cultural differences in expression or the situational context that might affect the interpretation of sentiments.