page 1

CSC 558 Data Mining and Predictive Analytics II, Spring 2020

Dr. Dale E. Parson, Assignment 2, Classification of audio data samples from assignment 1 for predicting numeric white-noise amplification level for the signals’ generators.1 We will also investigate discretizing the white-noise target attribute (class) and other non-target attributes. DUE By 11:59 PM on Wednesday March 4 via make turnitin on acad. The standard 10% per day deduction for late assignments applies. There will be one in-class work session for this assignment. You may attend in person or on-line. I encourage attending the work session in person. Start early and come prepared to ask questions. Perform the following steps to set up for this project. Start out in your login directory on csit (a.k.a. acad). cd $HOME mkdir DataMine # This may already be there. cd ./DataMine cp ~parson/DataMine/whitenoise558sp2020.problem.zip whitenoise558sp2020.problem.zip unzip whitenoise558sp2020.problem.zip cd ./whitenoise558sp2020 This is the directory from which you must run make turnitin by the project deadline to avoid a 10% per day late penalty. If you run out of file space in your account and you took csc458, you can remove prior projects. See assignment 1’s handout for instructions on removing old projects to recover file space, increasing Weka’s available memory, and transferring files to/from acad. You will see the following files in this whitenoise558sp2020 directory: README.txt Your answers to Q1 through Q16 below go here, in the required format. csc558wn10Ksp2020.arff The handout ARFF file for assignment 2, wn means white noise. makefile Files needed to make turnitin to get your solution to me. checkfiles.sh makelib ALL OF YOUR ANSWERS FOR Q1 through Q16 BELOW MUST GO INTO THE README.txt file supplied as part of assignment handout directory whitenoise558sp2020. You will lose an automatic 20% of the assignment if you do not adhere to this requirement. 1. Open csc558wn10Ksp2020.arff in Weka’s Preprocess tab. This is the same dataset used for

assignment 1, with AddExpression’s derived attributes already in place, and with tosc and tid removed; tagged numeric attribute tnoign, which is the gain on the white-noise generator, is the class (a.k.a. target attribute) of assignment 2. Where assignment 1 had a nominal attribute as the class, this assignment has tnoign as a numeric class attribute.

Here are the attributes in csc558wn10Ksp2020.arff.

1 See Assn1AudioOverview http://faculty.kutztown.edu/parson/spring2020/CSC558Audio1_2020.html and in-class discussion on the Zoom archives.

page 2

centroid Raw spectral centroid extracted from the audio .wav file. rms Raw root-mean-squared measure of signal strength extracted from the audio .wav file. roll25 Raw frequency where 25% of the energy rolls off, extracted from the audio .wav file. roll50 Raw frequency where 50% of the energy rolls off, extracted from the audio .wav file. roll75 Raw frequency where 75% of the energy rolls off, extracted from the audio .wav file. amplbin1 through amplbin19 Normalized amplitudes of 1st through 19th overtones of the fundamental. Filter RemoveUseless has removed amplbin0 because of its constant value of 1.0. Raw indicates an attribute that you normalized in assignment 1 to the reference fundamental frequency or amplitude. Attributes centrfreq, roll25freq, roll50freq, roll75freq, nc, n25, n50, n75, and normrms are Derived Attributes we created in assignment 1. Even though they are redundant with attributes from which they derive, they turn out to be useful for fine-tuning classifiers. We are keeping them for now. There are 34 attributes in the ARFF data of this assignment. tnoign Target white noise signal gain passed to the audio generator in the range [0.0, 1.0]. Except for the five tnoign=0.0 samples that we will remove, the signal generator for this dataset generates tnoign in the range [0.1, 0.25). Note the Weka Preprocess

statistics for tnoign below.

Figure 1: Class attribute tnoign in the handout dataset.

Since this assignment is about predicting white noise gain tagged as attribute tnoign, it is important to review the definition of white noise. As linked from Assn1AudioOverview, “White noise is a random signal having equal intensity at different frequencies, giving it a constant power spectral density…In discrete time, white noise is a discrete signal whose samples are regarded as a sequence of serially uncorrelated random variables with zero mean and finite variance.2” This white noise signal is distinct from the Sine, Triangle, Square, Sawtooth, and Pulse wave signals that were the focus of assignment 1, added into the composite signal with a random gain in the range [0.5, 0.75]. The dataset of assignments 1 and 2 add white noise with a random gain in the range [0.1, 0.25] to each signal-record in the dataset, with 5 exceptions that you will remove in step 2 below. Compare the frequency domain plot of the noiseless

2 Wikipedia page on white noise https://en.wikipedia.org/wiki/White_noise , quotation checked for accuracy.

page 3

1000 Hz training sine wave of assignment 13 with the 1001 Hz sine wave with a tnoign= 0.139453694281 used as a noise-bearing training instance in assignment 14. Both peak at about 1000 Hz, but the signal without white noise loses most of its strength after that. The signal with tnoign= 0.139453694281 white noise falls off considerably less, maintaining an almost constant signal strength all the way to the Nyquist frequency of 22050 Hz. The contribution of white noise at each frequency is random and seemingly small, but the net contribution of white noise is to add signal strength evenly across the frequency spectrum. We are trying to determine that contribution based strictly on audio data in the WAV files in this assignment. Important points to note include the following.

• Most of the frequency spectrum in the range [0, 22050] Hz lies above the non-noise signal generation (sine, triangle, etc.) fundamental frequency of [100, 2000] Hz. White noise spans the [0, 22050] Hz range. While the non-sine waves contribute harmonics that push measures such as centroid and the rolloff frequencies higher than the fundamental frequency, white noise pushes these measures even further up the frequency spectrum because it spans the [0, 22050] Hz range.

• White noise contributes additional power beyond the non-noise signals across the wave + white noise signal. Attribute rms is the measure of power across the time-varying, time-domain signal. Unlike the normalized fundamental frequency of amplbin0, which represents only the strongest frequency component of a signal, rms integrates signal strength across the frequency spectrum. The five tnoign=0.0 samples illustrated in Figure 1 are outliers in relation to the other 10,000 instances. 2. Use Weka’s Unsupervised -> Instance -> RemoveWithValues Preprocess filter to remove the five

outlying instances with tnoign=0.0. Use the attributeIndex to select tnoign, use the splitPoint to select a value for this attribute above which OR below which instances will be discarded, using invertSelection if necessary to change the direction of the split. Successful application of RemoveWithValues to tnoign results in 10,000 instances with tnoign in the range [0.1, 0.25], which is the range of white noise gain for the signal generator. SAVE THIS 10000-INSTANCE DATASET OVER TOP OF csc558wn10Ksp2020.arff, replacing the original csc558wn10Ksp2020.arff file.

Q1: What is your exact RemoveWithValues command line from the top of Weka’s Preprocess tab? RemoveWithValues –S 0.09 –C last -Lfirst-last The "cut point" has to be between the 0-noise instances and the others. It must be > 0 and <= 0.1. 3. Run Classify -> Functions -> LinearRegression on this 10,000-instance dataset, for which you should

get approximately the following results. THIS IS THE MODEL: tnoign = 0.5689 * centroid + 9.3174 * rms + -0.155 * roll25 + 0.0823 * roll50 + 0.0776 * roll75 + -0.0417 * amplbin1 + -0.0911 * amplbin2 +

3 http://faculty.kutztown.edu/parson/spring2020/lazy1_SinOsc_1000_0.9_0.0_0.FREQ.png 4 http://faculty.kutztown.edu/parson/spring2020/lazy1_SinOsc_1001_0.500235007566_0.139453694281_615143.FREQ.png

page 4

0.0149 * amplbin3 + -0.0528 * amplbin4 + 0.0117 * amplbin5 + -0.0172 * amplbin6 + 0.0692 * amplbin7 + -0.0745 * amplbin8 + 0.1115 * amplbin9 + 0.0514 * amplbin10 + 0.074 * amplbin11 + 0.0254 * amplbin12 + 0.0166 * amplbin13 + 0.0525 * amplbin14 + 0.0734 * amplbin16 + 0.1784 * amplbin18 + 0.1485 * amplbin19 + 0 * centrfreq + -0 * roll25freq + 0 * roll50freq + 0.0191 * n25 + -0.0075 * n50 + 0.0033 * n75 + 0.4098 * normrms + -0.4479 THESE ARE THE RESULTS: Correlation coefficient 0.7964 Mean absolute error 0.0205 Root mean squared error 0.0263 Relative absolute error 54.5234 % Root relative squared error 60.4761 % Total Number of Instances 10000 I have highlighted using bold-underline the two strongest contributing attributes from LinearRegression’s perspective, and I have highlighted using bold other contributors with an absolute value for the coefficient of at least 0.1. Note that negative coefficients are still important contributors in predicting a target numeric attribute; the sign means simply that they have a negative correlation. The following two Figures show the correlation between pairs centroid<->tnoign and rms<-> tnoign in Weka graphical form, and via running LinearRegression using on these pairs of attributes, using one pair at a time.

page 5

Q2: Given the fact that centroid is more closely correlated with tnoign than rms is correlated with tnoign, illustrated by both the approximate slopes of these graphs (we will discuss in class) and their individual correlation coefficients, why does rms have a coefficient 9.3174 that is much higher than centroid’s coefficient of 0.5688 in the complete LinearRegression model? The answer lies within Figures 2 & 3. Because with a mean value of .01 and a range [.006, .017], rms is much weaker in magnitude than centroid with a mean value of .37 and a range [.194, .465]; rms, even though it correlates less

page 6

significantly with tnoign than centroid does, requires more magnification in order to have an effect in the linear regression formula. Q3: Why does signal centroid correlate positively with white noise gain tnoign?

• Most of the frequency spectrum in the range [0, 22050] Hz lies above the non-noise signal generation (sine, triangle, etc.) fundamental frequency of [100, 2000] Hz. White noise spans the [0, 22050] Hz range. While the non-sine waves contribute harmonics that push measures such as centroid and the rolloff frequencies higher than the fundamental frequency, white noise pushes these measures even further up the frequency spectrum because it spans the [0, 22050] Hz range. Q4: Run Classify -> Trees -> M5P model tree on this 10,000-instance dataset, and record the Results (not the Model) for Q4. How do the M5P Results (correlation coefficient and error measures) compare with those of LinearRegression for this dataset? Correlation coefficient ? Mean absolute error ? Root mean squared error ? Relative absolute error ? % Root relative squared error ? % Total Number of Instances 10000 Correlation coefficient 0.922 Mean absolute error 0.0125 Root mean squared error 0.0168 Relative absolute error 33.2354 % Root relative squared error 38.7192 % Total Number of Instances 10000 M5P has higher correlation coefficient and lower error measures than LinearRegression. Q5. Run the instance-based (lazy) classifier IBk repeatedly with its default configuration parameters, increasing the KNN (number of nearest neighbors) parameter on each run until its performance begins to degrade, inspecting correlation coefficient for its peak. What value of KNN gives the most accurate result? Shows its Results. KNN = N Correlation coefficient n.n? Mean absolute error n.n? Root mean squared error n.n ? Relative absolute error n.n? % Root relative squared error n.n? % Total Number of Instances 10000 KNN = 9 Correlation coefficient 0.8817 Mean absolute error 0.0157 Root mean squared error 0.021 Relative absolute error 41.6986 % Root relative squared error 48.43 % Total Number of Instances 10000

page 7

Q6: Run the instance-based (lazy) classifier IBk one more time with its KNN as determined in Q5, then run it again after changing the nearest neighbor search algorithm from LinearNNSearch to KDTree with default parameters, and run it again using BallTree instead of KDTree. What change in behavior or performance do you notice compared to using the default LinearNNSearch nearest neighbor search algorithm? KDTree and BallTree run faster but give the same Results. Some students reported longer times for KDTree and BallTree compared to LinearNNSearch; that was for time to build the model, not to run it; that answer is OK. KDTree and BallTree take longer to build because they build trees; LinearNNSearch does not build a complicated search data structure. In preparation for the next steps, run Preprocess filter Unsupervised -> Attribute -> Discretize on the target attribute tnoign, making sure to set the ignoreClass configuration parameter to true. Leave the useEqualFrequency parameter at false, leave bins at 10, and check tnoign before and after using the filter to make sure its distribution histograms look similar, and that it is not numeric after discretization. Do NOT discretize any numeric attributes other than tnoign. Check in the Preprocess tab to make sure no other attributes are discretized. Q7: Now, run Preprocess filter Unsupervised -> Attribute -> Discretize on all remaining attributes with useEqualFrequency parameter at the default false and bins at 10. Inspect some of them in the Preprocess tab. Run classifiers rule OneR, tree J48, BayesNet, and instance (lazy) classifier IBk with the KNN parameter found in Q5 and nearest neighbor search algorithm of KDTree, and give their Results as outlined below, preceding each Result with the name of its classifier. Correctly Classified Instances N N.N % Incorrectly Classified Instances N N.N % Kappa statistic N.N Mean absolute error N.N Root mean squared error N.N Relative absolute error N.N % Root relative squared error N.N % Total Number of Instances 10000 OneR Correctly Classified Instances 1887 18.87 % Incorrectly Classified Instances 8113 81.13 % Kappa statistic 0.0976 Mean absolute error 0.1623 Root mean squared error 0.4028 Relative absolute error 90.1496 % Root relative squared error 134.2755 % Total Number of Instances 10000 J48 Correctly Classified Instances 2733 27.33 % Incorrectly Classified Instances 7267 72.67 % Kappa statistic 0.1924 Mean absolute error 0.1512 Root mean squared error 0.3242 Relative absolute error 83.9821 %

page 8

Root relative squared error 108.0764 % Total Number of Instances 10000 BayesNet Correctly Classified Instances 1908 19.08 % Incorrectly Classified Instances 8092 80.92 % Kappa statistic 0.1005 Mean absolute error 0.1631 Root mean squared error 0.3182 Relative absolute error 90.6288 % Root relative squared error 106.0662 % Total Number of Instances 10000 IBk with KNN = 9 and KDTree Correctly Classified Instances 2494 24.94 % Incorrectly Classified Instances 7506 75.06 % Kappa statistic 0.1657 Mean absolute error 0.1582 Root mean squared error 0.2892 Relative absolute error 87.9063 % Root relative squared error 96.3879 % Total Number of Instances 10000 Q8: Execute Preprocess -> Undo once, then check to make sure that only class tnoign is still Discretized. All other attributes except tnoign should be numeric. Now, run Preprocess filter Supervised -> Attribute -> Discretize on all remaining attributes (not tnoign). Inspect some of them in the Preprocess tab. Run classifiers rule OneR, tree J48, BayesNet, and instance (lazy) classifier IBk with the KNN parameter found in Q5 and nearest neighbor search algorithm of KDTree, and give their Results as in Q7, preceding each Result with the name of its classifier. Which classifiers became BETTER as measured by “Correctly Classified Instances” when compared with Q7, and which became WORSE. Just write BETTER or WORSE behind their classifier names. OneR BETTER Correctly Classified Instances 1911 19.11 % Incorrectly Classified Instances 8089 80.89 % Kappa statistic 0.1005 Mean absolute error 0.1618 Root mean squared error 0.4022 Relative absolute error 89.8829 % Root relative squared error 134.0768 % Total Number of Instances 10000 J48 WORSE Correctly Classified Instances 2194 21.94 % Incorrectly Classified Instances 7806 78.06 % Kappa statistic 0.1324 Mean absolute error 0.1597 Root mean squared error 0.331 Relative absolute error 88.7249 % Root relative squared error 110.3212 % Total Number of Instances 10000

page 9

BayesNet BETTER Correctly Classified Instances 2074 20.74 % Incorrectly Classified Instances 7926 79.26 % Kappa statistic 0.1189 Mean absolute error 0.1598 Root mean squared error 0.3217 Relative absolute error 88.7992 % Root relative squared error 107.2456 % Total Number of Instances 10000 IBk with KNN = 9 and KDTree WORSE Correctly Classified Instances 2298 22.98 % Incorrectly Classified Instances 7702 77.02 % Kappa statistic 0.1438 Mean absolute error 0.1608 Root mean squared error 0.2925 Relative absolute error 89.3439 % Root relative squared error 97.4961 % Total Number of Instances 10000 Q9: Execute Preprocess -> Undo once, then check to make sure that only class tnoign is still Discretized. All other attributes except tnoign should be numeric. Run classifiers rule OneR, tree J48, BayesNet, and instance (lazy) classifier IBk with the KNN parameter found in Q5 and nearest neighbor search algorithm of KDTree, and give their Results as in Q8, preceding each Result with the name of its classifier. Which classifiers became BETTER as measured by “Correctly Classified Instances” when compared with Q8, and which became WORSE. Just write BETTER or WORSE behind their classifier names. OneR WORSE Correctly Classified Instances 1589 15.89 % Incorrectly Classified Instances 8411 84.11 % Kappa statistic 0.065 Mean absolute error 0.1682 Root mean squared error 0.4101 Relative absolute error 93.4609 % Root relative squared error 136.7193 % Total Number of Instances 10000 J48 BETTER Correctly Classified Instances 3012 30.12 % Incorrectly Classified Instances 6988 69.88 % Kappa statistic 0.2235 Mean absolute error 0.1407 Root mean squared error 0.3524 Relative absolute error 78.1794 % Root relative squared error 117.4591 % Total Number of Instances 10000 BayesNet SLIGHTLY WORSE Correctly Classified Instances 2059 20.59 % Incorrectly Classified Instances 7941 79.41 %

page 10

Kappa statistic 0.1172 Mean absolute error 0.1606 Root mean squared error 0.3226 Relative absolute error 89.2539 % Root relative squared error 107.5425 % Total Number of Instances 10000 IBk with KNN = 9 and KDTree BETTER Correctly Classified Instances 2995 29.95 % Incorrectly Classified Instances 7005 70.05 % Kappa statistic 0.2213 Mean absolute error 0.1486 Root mean squared error 0.2829 Relative absolute error 82.5451 % Root relative squared error 94.2932 % Total Number of Instances 10000 In general, increasing the resolution of the non-target attributes by keeping them numeric may help accuracy of prediction, since discretized non-target attributes only approximate the precision found in numeric non-target attributes. Unfortunately, precise numeric attributes may be harder for some classifiers to analyze. Bayesian analysis, for example, does its own discretization of numeric non-target attributes; this discretization may be better or worse than the Supervised Weka discretization filter at correlating non-target attributes to the target class. Q10. Try using ensemble meta-classifier Bagging, using your most accurate classifier (in terms of Correctly Classified Instances) configuration from Q9 as its base classifier. What base classifier did you select, and does it improve performance over Q9 in terms of Correctly Classified Instances by more than 2% of 100% correct of the non-bagged Result of Q9? Show your Result as before. All attributes except the target tnoign should be numeric at this point. J48, 3% improvement. Correctly Classified Instances 3373 33.73 % Incorrectly Classified Instances 6627 66.27 % Kappa statistic 0.2636 Mean absolute error 0.1456 Root mean squared error 0.2788 Relative absolute error 80.8852 % Root relative squared error 92.9235 % Total Number of Instances 10000 Q11. Try using ensemble meta-classifier AdaBoostM1, using your most accurate classifier configuration form Q9 as its base classifier. What base classifier did you select, and does it improve performance over Q9 in terms of Correctly Classified Instances by more than 2% of 100% correct of the non-boosted Result of Q9? Show your Result as before. All attributes except the target tnoign should be numeric at this point. J48, 4% improvement Correctly Classified Instances 3421 34.21 % Incorrectly Classified Instances 6579 65.79 % Kappa statistic 0.2689 Mean absolute error 0.1334 Root mean squared error 0.3272

page 11

Relative absolute error 74.1228 % Root relative squared error 109.0535 % Total Number of Instances 10000 Q12. What accounts for any performance improvements in terms of Correctly Classified Instances in Q10 and Q11 over Q9 results? Bagging “shuffles” instances in the training set to reuse them and come up with a more varied training set; it removes some instances and duplicates others during distinct test runs to create a diverse cross- product of training data. Boosting learns from its mistakes (incorrectly classified instances in previous attempts) by adding weight to previously incorrectly classified instances while learning. In preparation for the final steps, you must copy your modified, 10,000-instance csc558wn10Ksp2020.arff back into the project directory on acad and run make train to create a 100-instance training set and a 9900-instance test set as follows. $ make train 'echo "making 100 training instances in csc558wnTrain100sp2020.arff" making 100 training instances in csc558wnTrain100sp2020.arff bash -c "echo '@relation csc558wnTrain100sp2020' > csc558wnTrain100sp2020.arff" bash -c "grep @ csc558wn10Ksp2020.arff | grep -v @relation >> csc558wnTrain100sp2020.arff" bash -c "grep ^[0-9] csc558wn10Ksp2020.arff | head -100 >> csc558wnTrain100sp2020.arff" echo "making 9900 test instances in csc558wnTest9900sp2020.arff" making 9900 test instances in csc558wnTest9900sp2020.arff bash -c "echo '@relation csc558wnTest9900sp2020' > csc558wnTest9900sp2020.arff" bash -c "grep @ csc558wn10Ksp2020.arff | grep -v @relation >> csc558wnTest9900sp2020.arff" bash -c "grep ^[0-9] csc558wn10Ksp2020.arff | tail -9900 >> csc558wnTest9900sp2020.arff" This places the first 100 instances of csc558wn10Ksp2020.arff into csc558wnTrain100sp2020.arff and the remaining 9900 instances into csc558wnTest9900sp2020.arff. You could do this by hand in a text editor, but using make train is a lot less time consuming and less error prone. Q13. After bringing files csc558wnTrain100sp2020.arff and csc558wnTest9900sp2020.arff back onto your Weka machine, load csc558wnTrain100sp2020.arff in the Preprocess tab as the training set, and set csc558wnTest9900sp2020.arff to be the supplied test set in the Classify tab. Run M5P and record its Results here. How many rules (linear formulas) does M5P generate? 1 formula LM num: 1 tnoign = 1.6315 * centroid + 17.5576 * rms - 0.2291 * roll25 - 0.1245 * roll50 - 0.3147 * roll75 + 0.7223 * amplbin1 + 0.3774 * amplbin2 + 0.3496 * amplbin3 + 0.766 * amplbin5 + 0.2347 * amplbin7 + 0.3956 * amplbin12

page 12

+ 0.3082 * amplbin16 + 0.2903 * amplbin18 + 0.0001 * nc - 0.6465 * normrms - 0.3665 Correlation coefficient 0.0296 Mean absolute error 0.4923 Root mean squared error 0.7583 Relative absolute error 1293.4681 % Root relative squared error 1715.7283 % Total Number of Instances 9900 Next, load your file csc558wn10Ksp2020.arff into Weka, run the Unsupervised -> Instance -> Randomize filter on it one time, with the default seed of 42, to shuffle the order of the instances. Save this as csc558wn10Ksp2020.arff, copy it back into the project directory on acad, and run make train again. Now, the first 100 instances in csc558wnTrain100sp2020.arff and the remaining 9900 instances in csc558wnTest9900sp2020.arff have been randomized with respect to order. Bring these new training and test files onto your Weka machine. Q14. After bringing randomized files csc558wnTrain100sp2020.arff and csc558wnTest9900sp2020.arff back onto your Weka machine, load csc558wnTrain100sp2020.arff in the Preprocess tab as the training set, and set csc558wnTest9900sp2020.arff to be the supplied test set in the Classify tab. Run M5P and record its Results here. How many rules (linear formulas) does M5P generate? 12 rules. Correlation coefficient 0.6744 Mean absolute error 0.0252 Root mean squared error 0.0329 Relative absolute error 66.479 % Root relative squared error 74.9903 % Total Number of Instances 9900 Q15. Before I removed tosc from your handout data, the instances were in the following order by tosc values. They remained in this order until you Randomized instance order. Note the five initial, 0-noise instances that you have deleted at the start of the current assignment in the above command output: $ grep Osc csc558lazyraw10005sp2018.arff | cut -d, -f2 |uniq -c 1 'PulseOsc' 1 'SawOsc' 1 'SinOsc' 1 'SqrOsc' 1 'TriOsc' 2000 'SinOsc' 2000 'TriOsc' 2000 'PulseOsc' 2000 'SawOsc' 2000 'SqrOsc' What accounts for the improvement in going from Q13 to Q14? Note that before Randomization, instances in file csc558wnTrain100sp2020.arff were in the same order as they are in the above csc558lazyraw10005sp2018.arff file.

page 13

The training set in Q13 was skewed by including strictly SinOsc instances, which have different frequency spectra than the other waveforms. There are 2000 SinOsc instances, so all 100 training instances in the non-Randomized training set in Q13 are SinOsc, and the test set is mostly not SinOsc. Randomizing the test and training sets helps to interlace the various waveform types, eliminating this skew. Q16. Can you improve performance of M5P further by bagging it? Give Results showing improvement, or explain why this attempt at improvement fails. Make sure to use the randomized training and set files csc558wnTrain100sp2020.arff and csc558wnTest9900sp2020.arff, with M5P as the base classifier. YES Correlation coefficient 0.759 Mean absolute error 0.0223 Root mean squared error 0.0286 Relative absolute error 58.759 % Root relative squared error 65.2712 % Total Number of Instances 9900 Each of Q1 through Q16 is worth 6% of the project, with the remaining 4% going for having a correctly Randomized csc558wn10Ksp2020.arff file in the project directory. Use make turnitin by the deadline to avoid a penalty.