Core Methods in Educational Data Mining HUDK4050 Fall 2015

Assignment B8 I know there were some technical glitches in the assignment – Not to mention a key ambiguity – sorry about that! Let’s go over the answers

Question 1: Set up a RapidMiner process using Read CSV and the GSP operator (Generalized Sequential Patterns – not the WEKA version W- GeneralizedSequentialPatterns). What should your customer id be? anonid obsnum behavior affect

Question 1: Set up a RapidMiner process using Read CSV and the GSP operator (Generalized Sequential Patterns – not the WEKA version W- GeneralizedSequentialPatterns). What should your customer id be? anonid obsnum behavior affect

Question 2: What should your time attribute be? anonid obsnum behavior affect

Question 2: What should your time attribute be? anonid obsnum behavior affect

Did anyone have trouble with this one? Question 3: Set min support = 0.6, window size = 0.0, max gap = 5.0, min gap = 0.0, positive value = 1. Which of these association rules has the highest support? behavior-ontask  behavior-ontask behavior-offtask  behavior-offtask behavior-ontask  affect-concentrating affect-concentrating  affect-concentrating  affect- concentrating  affect-concentrating  behavior- ontask AND affect-concentrating

Did anyone have trouble with this one? Question 4: If you set window size = 2.0, what is the association rule with the highest support that now is created (but was not created in question 3’s settings)? affect-concentrating  behavior-ontask  behavior- ontask AND affect-concentrating affect-concentrating  behavior-ontask  affect- concentrating affect-concentrating  behavior-ontask  affect- concentrating  affect-concentrating behavior-ontask  behavior-ontask  behavior- ontask AND affect-concentrating

Did anyone have trouble with this one? Question 5: Set window size back to 0.0. Set max gap to 1.0. Which is the rule with the most items? behavior-ontask AND affect-concentrating  behavior-ontask AND affect-concentrating affect-concentrating  behavior-ontask  affect- concentrating behavior-ontask  behavior-ontask  behavior- ontask AND affect-concentrating affect-concentrating  affect-concentrating

Question 6: Which of these is a reason why you might want to create a window size above 0? Related events may be linked but separated by a few seconds Unrelated events may be separated by a few seconds Related events may occur at exactly the same time Unrelated events may occur at exactly the same time

Question 6: Which of these is a reason why you might want to create a window size above 0? Related events may be linked but separated by a few seconds Unrelated events may be separated by a few seconds Related events may occur at exactly the same time Unrelated events may occur at exactly the same time

Question 7: How many students had the sequential (and immediate, max gap = 1) rule behavior-ontask -> affect-concentrating at least once? (Hint: RapidMiner may not be the easiest tool to compute this with)

Does anyone want to see this calculation? Question 7: How many students had the sequential (and immediate, max gap = 1) rule behavior-ontask -> affect-concentrating at least once? (Hint: RapidMiner may not be the easiest tool to compute this with) 49

Question 8: What is the confidence for sequential rule behavior-ontask -> affect- concentrating? Give three digits after the decimal point, round to nearest number. (Hint: RapidMiner may not be the easiest tool to compute this with)

0.726

Question 9: What is the cosine for sequential rule behavior-ontask -> affect-concentrating? Give three digits after the decimal point, round to nearest number. (Hint: RapidMiner may not be the easiest tool to compute this with)

0.671

Question 10: What is the lift for sequential rule behavior-ontask -> affect-concentrating? Give three digits after the decimal point, round to nearest number. (Hint: RapidMiner may not be the easiest tool to compute this with) Did anyone get the answer in the system? I think there might have been a bug.

Question 11 (ungraded due to technical issues) Question 11: Would Merceron & Yacef say that this is an interesting association rule? Yes, because cosine is over threshold Yes, because lift is over threshold Yes, because both lift and cosine are over threshold No, because cosine is over threshold No, because lift is over threshold No, because cosine is below threshold No, because lift is below threshold

Sorry again about technical glitches Any content-related technical questions or comments?

ARM vs SPM What are the differences between – Association Rule Mining – Sequential Pattern Mining

Any questions about GPS algorithm?

Perera et al. (2009) What were the three ways that Perera et al. (2009) used sequential pattern mining? What did they learn, and how did they use the information?

Perera et al. (2009) 1.Overall uses of collaborative tools by groups 2.Sequences of collaborative tool use by different group members 3.Sequences of access of specific resources by different group members In all cases, they found common patterns and then looked at how support differed for successful and unsuccessful groups

Perera et al. (2009): Important Findings 1.Overall uses of collaborative tools by groups – Successful groups used ticketing system more than the wiki; weaker groups used wiki more – Patterns were particularly strong for group leaders

Perera et al. (2009): Important Findings 2.Sequences of collaborative tool use by different group members – Successful groups characterized by leader opening ticket and other student working on ticket – Successful groups characterized by students other than leader opening ticket, and other students working on ticket

Perera et al. (2009): Important Findings 3.Sequences of access of specific resources by different group members – The best groups had interactions around the same resource by multiple students – The poor groups did no work on tickets before closing them

Variants

Differential Sequence Mining (Kinnebrew et al., 2013) Split data into two groups Look for differences in pattern frequencies between groups

Example (Jiang et al., 2015) Compare behaviors of – Students who had used inquiry science environment before – Students who had never used inquiry science environment before

Difference found Experienced students are more likely to read about several topics after conducting an experiment and looking at the results Novice students are more likely to read about a single topic after conducting several experiments and looking at the results

MOTIF Extraction

Motif Short, recurring pattern in a sequence of categories occurring over time

Motif in Music Short, recurring pattern of notes in a musical composition

Motif in Music What’s the motif? KIY KIY How many times does the motif occur?

Motif in Music What’s the motif? KIY KIY How many times does the motif occur? – Depends on how you define it, right? – And that’s part of the challenge…

Motif in Language Short, recurring pattern of characters in a sequence of characters occurring over time

Motif in Genetics Short, recurring pattern of genes in a sequence of genes occurring over time Typically written as letters

Goal of Motif Extraction Discern a common pattern of characters in a large corpus of characters The characters may vary slightly from case to case

Can you find the motif?

UBSWWDFKLWPRHUC JBDPXBDVEJVMBKK VBDWNLROFVUBFFW OWIFTIENDOXJXIOB AUAAOOXZAABZSBT MUAWSNTVZXSFHMI LFQRKUTFRIENDOV LOMTPOQHJVYYMFJ LWGJMVPKYOZNMSA RUPMFOHPVSPPVPT BAZXVFTPQFQJVBM HLPMOKUOXGRIENDO INUSUNSGDAAICAV XRZZWCDXOVZZJKQ VOVCROMCJTOLXYU HUVRYFREENDOBBGC AQJBVXJCAJLEMAU ONJORIFCGAUGIRN PJGCHBDQIWJJTMQ IQYQHKKBNBVDFPV JJLHWPZAYZIGGEH IGJZRMAAWJBESSS JXZFRIEMDOVZRBJY IRPWYIRJISLFVFF

How would you describe the motif? UBSWWDFKLWPRHUC JBDPXBDVEJVMBKK VBDWNLROFVUBFFW OWIFTIENDOXJXIOB AUAAOOXZAABZSBT MUAWSNTVZXSFHMI LFQRKUTFRIENDOV LOMTPOQHJVYYMFJ LWGJMVPKYOZNMSA RUPMFOHPVSPPVPT BAZXVFTPQFQJVBM HLPMOKUOXGRIENDO INUSUNSGDAAICAV XRZZWCDXOVZZJKQ VOVCROMCJTOLXYU HUVRYFREENDOBBGC AQJBVXJCAJLEMAU ONJORIFCGAUGIRN PJGCHBDQIWJJTMQ IQYQHKKBNBVDFPV JJLHWPZAYZIGGEH IGJZRMAAWJBESSS JXZFRIEMDOVZRBJY IRPWYIRJISLFVFF

Finding motifs Several algorithms

Finding motifs Variant on PROJECTION algorithm (Tompa & Buhler, 2001) used in (Shanabrook et al., 2010) – Only example of motif extraction in educational data mining so far

Big idea For each character string C that could be a motif example (e.g. all character strings of desired length) – Create a set of projections, random variations of C that vary in one or more ways

Big idea For each pair of strings C1 and C2, see how many overlaps there are between their projection matrices Take the pair with the most matches and combine into a motif – Creating multi-example motif if 3+ get added together Repeat until goal number of motifs is found, or until new motif is below criterion goodness

Motif in Education Short, recurring pattern of behaviors in a sequence of behaviors occurring over time Written as letters in Shanabrook et al. (2010)

Detail for education How do you segment student behavior? Could use student’s interaction on an entire problem, and compute letters across whole problem – Might make more sense in tutors with shorter problems Could use student’s interaction on an entire problem, and define letters differently for context within whole problem – Approach used by Shanabrook et al. (2010) Could use “sliding window” of N actions

Behaviors in Shanabrook et al. “hints (a, b, c) – Hints is a measure of the number of hints viewed for this problem. Although each problem has a maximum number of hints, the hint count does not have an upper bound because students can repeat hints and the count will increase at each repeated view. The three categories for hints are: (a) no hints, meaning that thestudent did not use the hint facility for that problem, (b) meaning the student used the hint facility, but was not given the solution, and (c) last hint solved, meaning that the student was given the solution to the problem by the last hint. As described above, this metric combines two values logged by the tutor: the count of hints seen, and an indicator that the final hint giving the answer was seen. The data could have been simply binned low, medium, high hints; however, this would have missed the significance of zero hints and using hints to reveal the problem solution.”

Behaviors in Shanabrook et al. “secFirst (d, e, f) – The seconds to first attempt is an important measure as it is during this time that the student is reading the problem and formulating their response. In previous research [6], five seconds was determined to be a threshold for this metric representing gaming: students who make a first attempt in less than five seconds are considered not working on-task. We divide secFirst into three bins: (d) less than 5 sec, (e) 5 to 30 sec, (f) greater than 30 sec. (d) represents students who are gaming the system, (e) represents a moderate time to the first attempt, (f) represents a long time to the first attempt. The cut at 30 seconds was chosen because it equalizes the distribution of bins (e and f), representing a division between a moderate and a long time to the first attempt.”

Behaviors in Shanabrook et al. “secOther (g, h, i, j, k) – This variable represents actions related to answering the problem after the first attempt was made. While the first attempt includes the problem reading and solution time, subsequent solution attempts could be much quicker and the student could still be making good effort. secOther is categorized in five bins: (g) skip, (h) solved on first, (i) 0 to 1.2 sec, (j) 1.2 to 2.9 sec, (k) greater than 2.9 sec. First, there are two categorical bins, skip and solve on first attempt. These are each determined from an indicator in the log data for that problem. Skipping a problem implies only that students never clicked on a correct answer; they could have worked on the problem and then given up, or immediately skipped to the next problem with only a quick look. Solved on first attempt indicates correctly solving the problem. If neither of the first two bins are indicated in the logs, then the secOther metric measures the mean time for all attempts after the first. The divisions of 1.2 sec and 2.9 sec for the latter three bins were obtained using the mean and one standard deviation above the mean for all tutor usage; (i) less than 1.2 seconds would indicate guessing, (j) would indicate normal attempts, and (k) would indicate a long time between attempts.”

Behaviors in Shanabrook et al. “numIncorrect – (o, p, q) - Each problem has four or five possible answer choices, that we divide into three groups: (o) zero incorrect attempts, indicates either solved on first attempt, skipped problem, or last hint solves problem (defined by the other metrics); (p) indicates choosing the correct answer in the second or third attempt, and (q) obtaining the answer by default in a four answer problem or possibly guessing when there is five answer problem.”

What other constructs could be used? What other kinds of constructs could be used for the atoms of motif analyses in educational analyses? – At this grain-size (e.g. specific actions)

What other constructs could be used? What other kinds of constructs could be used for the atoms of motif analyses in educational analyses? – At other grain-sizes?

Common Motifs {adgo, adip, adiq} {aeho, afho} {ceho} {adgo, aeho} {aeiq aeho aeho aekp aeho aeiq aeho aeip aeho aeip}

Interpretations (Shanabrook et al., 2010) {adgo, adip, adiq} – gaming the system {aeho, afho} – “This student is using the tutor appropriately, but not being challenged.” {ceho} – problem is too difficult {adgo, aeho} – student is skipping problems {aeiq aeho aeho aekp aeho aeiq aeho aeip aeho aeip} – working on-task

Do you agree with interpretations? {adgo, adip, adiq} – gaming the system {aeho, afho} – “This student is using the tutor appropriately, but not being challenged.” {ceho} – problem is too difficult {adgo, aeho} – student is skipping problems {aeiq aeho aeho aekp aeho aeiq aeho aeip aeho aeip} – working on-task

How can researchers form good interpretations?

Questions? Comments?

What else? What else could sequential pattern mining and motif extraction be used for in education? – Beyond Perera et al. and Shanabrook et al.

Assignment C4 Sequential Pattern Mining II Due Tuesday, December 8 Questions?

Final Project Any questions? If you have not yet signed up for a slot on the discussion forum, please do so ASAP

Next Class Tuesday, December 8: Text Mining Readings Baker, R.S. (2015) Big Data and Education. Ch. 8, V3. Graesser, A. C., D'Mello, S. K., Craig, S. D., Witherspoon A., Sullins J., McDaniel B., Gholson, B. (2008) The Relationship between Affective States and Dialog Patterns during Interactions with AutoTutor. Journal of Interactive Learning Research, 19(2), Adamson, D., Bharadwaj, A., Singh, A., Ashe, C., Yaron, D., & Rosé, C. P. (2014). Predicting Student Learning from Conversational Cues. Proceedings of the International Conference on Intelligent Tutoring Systems, Crossley, S., McNamara, D., Baker, R.S., Wang, Y., Paquette, L., Barnes, T., Bergner, Y. (2015) Language to Completion: Success in an Educational Data Mining Massive Open Online Course. Proceedings of the 8th International Conference on Educational Data Mining,

The End