HeaderSIS.jpg

Difference between revisions of "IS480 Team wiki: 2012T1 6-bit Project Management UT4"

From IS480
Jump to navigation Jump to search
Line 227: Line 227:
 
* A size of a positive or negative number closer to 1 or -1 suggests strength of relationship,
 
* A size of a positive or negative number closer to 1 or -1 suggests strength of relationship,
 
* and A p value of 0 suggests an absolute lack of relationship in the input attributes.
 
* and A p value of 0 suggests an absolute lack of relationship in the input attributes.
 +
 +
It should be noted that if there is a strongly negative correlation, it may suggest a reversal in the order of recommendations in our system may eventually become an ideal model.
 +
<br>
 +
The accuracy sentimental analysis system is, in principle, how well it agrees with human judgment, despite the fact that the applied accuracy in our test is to ensure an optimal user experience and optimal sales exposure. <br>
 +
However, the limitation of the model is that human raters typically agree about 70% of the time and even if a system is 100% accurate by assumption, humans will still disagree about 30% of the time. Hence, for the purpose of our study, it may be suggested that an application with 100% human agreeableness can only be statistically justified to be 70% accurate and more sophisticated methods should be used to endorse the remaining 30% of accuracy. <br><br>
 +
After conducting the IRR Test, the results are tabulated and appended as follows.
 +
<br>
 +
[[Image: 6bituser-testing4results.png]]
 +
<br>
 +
Tester IDs that begins with M denotes male testers, while F denotes Female testers. The categorization of results will help us understand a basic level of stereotyping accuracy by gender. If need it, the results can be further drilled down to represent a more specific personality stereotype.
 +
<br><br>
 +
Summarising the test results tabulated above, the average for median p-value for male is 0.8695 and median p-value for female is 0.8675. Based on the median values, there is no significant difference in accuracy for male and females. In addition, it is satisfactorily justified that there is a strong positive correlation between the recommendations of our system and the preferences of the testers, a sample from our potential users, based on their consensus.
 +
  
 
====Click Data Analysis(User Test 1 vs. User Test 2 – Forum Functions Only)====
 
====Click Data Analysis(User Test 1 vs. User Test 2 – Forum Functions Only)====

Revision as of 23:33, 1 December 2012

6-bit logo.png
6-bit's Chapalang! is a social utility that connects people with friends and new friends
by offering a place for exchanging ideas and information on its public domain.
http://www.chapalang.com

Final Wikipage
Home Technical Overview Project Deliverables Project Management Learning Outcomes


Schedule

Planned Schedule

6-bit ScheduleDiagramOverview.png

































Meeting Minutes

Team Meeting Minutes

[|Meeting Minute 1] [|Meeting Minute 11] [|Meeting Minute 21]
[|Meeting Minute 2] [|Meeting Minute 12] [|Meeting Minute 22]
[|Meeting Minute 3] [|Meeting Minute 13] [|Meeting Minute 23]
[|Meeting Minute 4] [|Meeting Minute 14] [|Meeting Minute 24]
[|Meeting Minute 5] [|Meeting Minute 15] [|Meeting Minute 25]
[|Meeting Minute 6] [|Meeting Minute 16]
[|Meeting Minute 7] [|Meeting Minute 17]
[|Meeting Minute 8] [|Meeting Minute 18]
[|Meeting Minute 9] [|Meeting Minute 19]
[|Meeting Minute 10] [|Meeting Minute 20]

Supervisor Meeting Minutes

|Meeting Minute 1
|Meeting Minute 2
|Meeting Minute 3
|Meeting Minute 4
|Meeting Minute 5
|Meeting Minute 6
|Meeting Minute 7
|Meeting Minute 8
|Meeting Minute 9
|Meeting Minute 10
|Meeting Minute 11

Testing

Test Cases

|Test Cases

Test Plans

Test Plan 1 on 17 September 2012
Test Plan 2 on 28 September 2012
Test Plan 3 on 19 October 2012
Test Plan 4 on 4 November 2012

User Testing

User Testing 1 User Testing 2 User Testing 3 User Testing 4

User Testing 4

6bituser-testing4.png

Test Description

The objective of User Test 4 is on scalability, performance and analytics testing of the system. This is a 2-part test session, firstly on scalability and performance which does not require physical testers, and secondly on inter-rater reliability which requires rating judges.

The coverage of the scalability and performance test is focused on the bottleneck functions, which are the discussion forums and marketplace. The terms “performance” and “scalability” are commonly used interchangeably, but the two are distinct: performance measures the speed with which a single request can be executed, while scalability measures the ability of a request to maintain its performance under increasing load.

Additionally, Inter-Rater Reliability Test is performed on the Personalized Dashboard to determine the concordance of the personalized results and actual personality of user stereotypes.

Testers Background

Scalability & Performance Testing

As the test does not require physical testers, the following appends the test environment.
6-bit ut4spec.png

Inter-Rater Reliability Test

Testers will assume the role of raters or judges for our Inter-Rater Reliability Test, represented by a total of 20 people with a 50-50 male is to female ratio. Testers are stratified from a diverse background, intended to represent personality stereotypes designed.

Personality stereotypes include characteristics such as gender, age group, education, personality traits, online activity, mobility and interested topics.

Test Groups

There is no test grouping employed in this test.

Test Procedures

Scalability & Performance Testing

The first part of the test is on scalability and performance testing. Chapalang Benchmark is configured to measure the time taken for each controller method before it starts and as soon as it ends. The results will be used to study the performance of the system and application at different scales of operations.

Additionally, custom application is used to perform a series of activities on a forum and marketplace, simulating an arbitrary number of concurrent users on the system, or the load.

Subsequently, we will study the benchmark timing data to understand the performance differences under different load.

Inter-Rater Reliability Test

An Inter-Rater Reliability (IRR) is the degree of agreement among raters. It gives a score of how much homogeneity or consensus there is in the ratings given by judges.

The first rater is the system itself, which will generate a list of 10 products and 10 discussion topics recommendations in descending order of relevance to a target user. Every recommendation is tied to a specific and distinct order number.

The second rater is a human tester, who will be provided with the same list of 10 products and 10 discussion topics generated by the system in relevance to him or herself. To mitigate the effects of Experimenter’s Bias, the order of each item is unordered and randomized without any intended logic. The second rater is expected to reorder the items according to his or her preferences in descending order.

Subsequently, we will make use of Spearman’s Rank Correlation Coefficient to understand the reliability of our personalized dashboard which features product and topic recommendations.

Test Instruction

Inter-Rater Reliability Test

This is a sample output of the first rater, for a product recommendation test.
6-bit ut4a.png
This is a sample input sheet for the second rater, on product recommendation test.
6-bit ut4b.png
In a descending order, 1 represents the most relevant item while 10 represents the least relevant item.

Test Results

Scalability & Performance Test

File:6bituser-testing4resultsfigure.png

Inter-Rater Reliability Test

In order to evaluate the test results, we rely on the statistical model called Spearman’s Rank Correlation Coefficient (SRCC). The model is appended below:
6bituser-testing4formula.png
In short, the SRCC model takes into account the rank rating from 2 different raters, represented by xi and yi respectively. In conventional Correlation of determination model, it takes in absolute data instead to find out statistical and data-driven correlation between 2 inputs.

However, we are interested in the consensus between human judgment on the data, therefore SRCC is a suitable model of analysis. The SRCC model assumes that the rating scale is ordinal, or basically serial scale of rating. This assumption is aligned with our 1 – 10 rating score which is incremental and serial. Additionally, the SRCC model considers only relative position of the ratings. For example, (1, 2, 1, 3) is considered perfectly correlated with (2, 3, 2, 4). This is acceptable in our test because each rating in our test is distinct and exhaustive, where no repeats or unused score is allowed.

The following is a sample of data tabulation in visual form.
6bituser-testing4z.png
With 6bituser-testing4di2.png found, we can add them to find 6bituser-testing4sumdi2.png 28. The value of n is 10. So these values can now be substituted back into the equation, we evaluate that p = 0.83.
Based on the nature of correlation coefficient value p, it has a range between -1 to 1. Here are some of the characteristics of the p value.

  • A negative number suggests negative relationship,
  • A positive number suggests positive relationship,
  • A size of a positive or negative number closer to 1 or -1 suggests strength of relationship,
  • and A p value of 0 suggests an absolute lack of relationship in the input attributes.

It should be noted that if there is a strongly negative correlation, it may suggest a reversal in the order of recommendations in our system may eventually become an ideal model.
The accuracy sentimental analysis system is, in principle, how well it agrees with human judgment, despite the fact that the applied accuracy in our test is to ensure an optimal user experience and optimal sales exposure.
However, the limitation of the model is that human raters typically agree about 70% of the time and even if a system is 100% accurate by assumption, humans will still disagree about 30% of the time. Hence, for the purpose of our study, it may be suggested that an application with 100% human agreeableness can only be statistically justified to be 70% accurate and more sophisticated methods should be used to endorse the remaining 30% of accuracy.

After conducting the IRR Test, the results are tabulated and appended as follows.
6bituser-testing4results.png
Tester IDs that begins with M denotes male testers, while F denotes Female testers. The categorization of results will help us understand a basic level of stereotyping accuracy by gender. If need it, the results can be further drilled down to represent a more specific personality stereotype.

Summarising the test results tabulated above, the average for median p-value for male is 0.8695 and median p-value for female is 0.8675. Based on the median values, there is no significant difference in accuracy for male and females. In addition, it is satisfactorily justified that there is a strong positive correlation between the recommendations of our system and the preferences of the testers, a sample from our potential users, based on their consensus.  

Click Data Analysis(User Test 1 vs. User Test 2 – Forum Functions Only)

Additionally, click data of each test session has also been collected and analysed. They are also being compared to that with the results of User Test 1.
6bituser-testing2cda1.png
The above box-plot represents 3 sets of data comparing the number of clicks per task, for discussion forum functions only. UT1 represents the results from User Test 1, UT2A represents the results from Group A testers of User Test 2, while UT2B represents the results from Group B testers of User Test 2. For the objective of fair comparisons, the results from User Test 2 has been drilled down to consists of data

The median number of clicks it takes per tester to accomplish a forum-related task in User Test 2 ranges from 1 to 3 clicks with 2 clicks being the median, a decrement from the median of 3 clicks, as well as a smaller variance as compared to User Test 1. Additionally, it can also be observed that there is no significant difference in the results between Group A and Group B users.

Preliminary, we can observe an improvement in the user experience for Group A users between the 2 tests. The improvement can be broadly attributed to the improvements made as well as the high learnability of the system interface design. However, this observation is not conclusive and more data is required.
6bituser-testing2cda2.png
Additional statistics were computer and observed that the median time spent to accomplish a forum-related task for Group A tester is 4 seconds, and Group B tester is 5 seconds. Again, this is a significant decrement in the time spent from User Test 1, where testers spent a median of 10 seconds between each task. Based on the finding, we can reasonably derive that there is an improved user experience between User Test 1 and User Test 2, attributing to the improvements made and high learnability of the system. In addition, the improved user experience is shared between both Group A and Group B users, possibly suggesting that the improved system does not require much training or high learning curve.

Click Data Analysis (Group A vs. Group B – Marketplace Functions)

Prospectively, we will also study the user experience difference between Group A and Group B testers on marketplace functions, based on the click data which measures the number of clicks involved per task and time taken in seconds between each task.

In the box-plot diagram above, UT2A refers to User Test 2 Group A testers, while UT2B refers to User Test 2 Group B testers. Each box-plot is represented by data of a specific group of users, and the results computed based on the number of clicks of time pertaining to forum or marketplace functions.
6bituser-testing2cda3.png
Comparing marketplace functions, both Group A and Group B testers have made a median of 2 clicks to accomplish each task. While Group B testers have a wider variance of clicks of up to 4 clicks, it can be broadly attributed to outliers, user experiments or some learning curve involved in getting used to the functions or interface objects placements.
6bituser-testing2cda4.png
The result when comparing the time taken is consistent with the preliminary conclusion when comparing the number of clicks per task. The median time taken for Group A and Group B testers for forum and marketplace functions are within the range of 4 to 5 seconds. The difference between the median records is insignificant.

Overall, the result is consistent across forums and marketplace functions, between testers from both User Tests and test groups. It is also consistent with our earlier preliminary conclusion that the improvements made between the two User Tests have resulted in improved user experience, and there is a good level of learnability in the interface design.

While there are limitations in this test, where there are other externalities such as network performance, computing habits of testers and response time of each users, the macro results of the test provide a reasonable sampling on the objective of the test.

Milestones

6-bit milestones.png

Schedule Metric


6-bit ScheduleMetric.png

Bug Metric


6-bit BugMetric.png Bug Log: |Click Here

Risk & Mitigation


6-bit RiskDiagram.png