Submit an Academic Petition or Appeal

Petition Letter: Dear K. Deluzio and whom it may concern,

This is Benjamin Schaab, #20273389. I am requesting a regrade on my final exam for CISC/CMPE 251. The exam was a written assessment split into two parts online expected to take 30 minutes each. The parts were weighted at 15% and 20% of my final grade. On both parts I feel that I was graded too harshly based on what was asked in the questions and unclear guidelines. My grades for both parts were 1/15 and 3/20 which I feel are far too low for a written piece of work that arguably doesn’t have right and wrong answers. Considering that the written ups were only supposed to take around 30 minutes I think I wrote more than enough to give sufficient answers. I’ve copy-pasted the exam questions and my responses at the bottom of this letter for reference.

Looking at my answers on the first exam, the questions being asked were:

“What attributes would you like them to collect for you? Over what time periods would you like the data collected? What kind of data analytic techniques(s) would you apply (and in what order)? Why are these the right techniques? What results do you expect to see? How would you validate your results?”

I answered these questions one by one in short paragraphs. My responses may not have been exactly what the professor was looking for, but I think the parts I was missing should have been more clearly indicated in the question as the question posed felt very open ended, without a clear consensus of what was being graded.

As for the second exam, I feel that I followed the advice the professor gave after the first exam, but still didn’t have success. The question was looking for two answers being “(a) the predictor is performing poorly, or (b) there is only a weak connection between laptop configuration and price because manufacturers basically make up a price.” Maybe I was a bit brief in my answer to part b, but I think I brought up plenty of points to assess the performance of the predictor in my write-up.

Some examples from my text being:

“Compare the prices of the laptops with laptops that have similar components, but prices that vary, it will be important to eliminate outliers in this step as they will negatively affect the accuracy of the predictions”

“The main components to identify and test in the dataset are likely going to be computer specifications such as the CPU, graphics card, display quality, and any other key features that are major selling points to consumers. These laptop details are going to be the driving factors in establishing a strong connection with the price prediction.”

“A simple way for the McConsey company to see if the predictor is performing poorly is to do some cross-validation with laptops that have a similar description to the output of the prediction model and see if the prices make sense. This should be done with laptops that were not part of the dataset because this will make it so that the data is new, and the predictor wouldn't have been built on it.”

“The prediction model should also be visualized on a scatter plot to access its performance. It's important to look out for patterns in the visualized data as any linearity to access its success. Overall, the main way for the company to create a successful predictor for the laptop prices if going to be by selecting the data rows of laptop information that play the largest role when setting the price and building the predictor around them while eliminating the outlying unimpactful computer specifications.”

In general, I think I answered the exams sufficiently given what is being asked in the question. I understand that writing can be difficult to grade because of how opinion based it can be, but I put effort into this exam and giving me a such low marks does not reflect those efforts. I’m not asking for an amazing mark on this, but I think a higher grade is in order as my current grade gives off the impression that I only wrote 3 sentences or something, which is of course not the case.

Exam 1 Question:
“A mobile phone company notices that 1% of its customers leave them each month for another service provider. The average cost of gaining a new customer is $15. They decide that it would be better to send a $10 coupon to customers who are likely to leave soon, in the hope of retaining them.

You have been retained as an independent consultant to help them solve this problem. Describe the steps you would take, including at least answers to the following questions:

* what attributes would you like them to collect for you?
* over what time periods would you like the data collected?
* what kind of data analytic techniques(s) would you apply (and in what order)?
* why are these the right techniques?
* what results do you expect to see?
* how would you validate your results?”
Exam 1 Answer:
“If I was the independent consultant for this company there are many different attributes I would want to have collected on my behalf. First off, I'd want to know the demographic of the customers such as their age, gender, locations, and income. More specifically income as this can have a major affect on apparent value of the $10 coupon to the customer. If the customer is wealthy, the coupon won't matter as much to them as they are more concerned with the service that the phone company provides. Other attributes worth collecting include customer feedback. This includes any interaction customers have with customer support. Analysing this feedback can give insight into what customers appreciate and what they dislike about to the phone company to better understand why users are leaving in the first place. The last attribute I'd look to collect would be the coupon usage rate. This would be the number of customers issued this coupon and decided to use it. This also relates to the data associated with the number of customers not using the coupon and cancelling their subscription. As well, as information about the success of any previous coupons the company issued to customers.

These attributes and data should be collected over 12 months so that the customer basis can be analysed seasonally. It may be that more customers change providers around the holidays because of sales or better coupons offered by competitors.

Once data is collected it must be analysed using a variety of techniques. To begin the data should be organized the better understand the churn rate of customers based on the demographic and usage patterns. The next step would be to create a predictive model in KNIME or another software to use machine learning to better identify the customers at risk of leaving over the 12 month period. This data should then be segmented based on characteristics of customers to see what factors lead to the highest churn rate.

Based on previous customer loyalty expectancies, customers should have a value set to them related to how much the company could profit in the future. This can be used to determine how much more money the company should be willing to allocate to these groups to retain them as customers and prioritize retention efforts on the high-value customers. Finally, a form of bucket testing should be done to evaluate the effectiveness of the $10 coupon. Randomly assign eligible customers to either receive the coupon or not and compare their churn rates.

These are the correct techniques because it starts off by getting an initial understanding of the data and the situation the company is in before diving into deeper data analysis strategies. Predictive modeling should be used as it is an efficient and effective way of determining the customers at risk of leaving. The bucket testing is a simple way to validify the $10 coupon and see if it has any affect on customer retention.

Once data analysis has been done what I'd want to learn from the data is confirmation on who the high-risk customers are who might leave the phone company. As well as the insight into the effectiveness of the $10 coupon on reducing the customers leave rate.
Results can be validated by collecting more data in the future to see what worked well and what didn't. This can be combined with a feedback loop where results are monitored with adjustments being made to the coupon and the company based on continuous data analysis. The validate the predictive model the predicted churn rates can simply be compared with the real churn rates given some time.”
Exam 2 Question:
“You may have noticed that the prices of laptops often do not seem to have much relationship with the specifications of each laptop's hardware and software. The McConsey consulting company wants to explore this issue.

They gather a large dataset where each record describes one laptop product, available in 2023, labelled with its best price.

They want to build a predictor for prices from hardware and software configuration, but they are concerned that they won't be able to tell if (a) the predictor is performing poorly, or (b) there is only a weak connection between laptop configuration and price because manufacturers basically make up a price.

Discuss how you would help them address this dilemma, using other aspects of data analytics.”
Exam 2 Answer:
“After analyzing McConsey company's situation, it looks like the dataset will be analyzed using clustering and a prediction model by selecting specific attributes from the dataset that have the largest effect on the price of the laptop. Before making a clustering model for the data we'll want to compare the prices of the laptops with laptops that have similar components, but prices that vary, it will be important to eliminate outliers in this step as they will negatively affect the accuracy of the predictions. This elimination process may also help identify trends in laptop brands that may be over/under pricing their products. I believe DBSCAN is going to be a good clustering option for the data because it is effective at identifying outliers which are likely to appear when comparing many laptops from multiple brands. This will show whether the data have a weak connection between the laptop specifications and the price and show which brands are outliers in the clustering and are making their own prices for their products as the manufacturers. The main components to identify and test in the dataset are likely going to be computer specifications such as the CPU, graphics card, display quality, and any other key features that are major selling points to consumers. These laptop details are going to be the driving factors in establishing a strong connection with the price prediction. This will also help identify trends in laptop brands that may be over/under pricing their products. The best prediction model to be used alongside these clusters is likely going to be SVM. I believe SVM will be most effective because there should be a clear margin of separation between the rows of the dataset containing the specifications and descriptions of the laptops. Some specifications will have a much larger impact on the price of the laptop as previously mentioned. A simple way for the McConsey company to see if the predictor is performing poorly is to do some cross-validation with laptops that have a similar description to the output of the prediction model and see if the prices make sense. This should be done with laptops that were not part of the dataset because this will make it so that the data is new, and the predictor wouldn't have been built on it. If they find that the prediction varies heavily then there are likely still outliers in the data, negatively affecting the prediction model. The prediction model should also be visualized on a scatter plot to access its performance. It's important to look out for patterns in the visualized data as any linearity to access its success. Overall, the main way for the company to create a successful predictor for the laptop prices if going to be by selecting the data rows of laptop information that play the largest role when setting the price and building the predictor around them while eliminating the outlying unimpactful computer specifications.”