CMSE11459 Data Mining 1

Is The Reduction Of Features Beneficial To Improving Performance?

Many variables will influence success in many situations. Reduced functionality is one cause. Data mining and WEKA analysis was used in this project to see if reducing features will help increase performance. A provided data set contains not only performance-related characteristics, but also attribute data. Missing attributes, outliers, and attribute count reductions are addressed using several data pre-processing methods in WEKA, resulting in a clean final data collection. Proceed to the data review phase after the data has been pre-processed. This is the stage. To perform analysis on training data sets, various methods such as naive Bayes, J48, and support vector machines are used. The findings of the evaluation data set are first applied to the training data set. These effects are compared to determine the model's accuracy.

Conversion

First Database

The research made use of a text data directory. The information presented can be classified into two categories: business and sports. To import data into Weka, the analytics environment, the first approach was to use the TextDirectoryLoader feature. Notice that the data is then separated into two categories: text (which includes text information) and @@ class @@ (which includes the original text information class).... To reduce the preprocessing of a single data set (including just converting text data to word vectors, removing only the simple punctuation defined in Weka, and setting the word to be preserved), the original data was split into two data sets, one of which was set to include more predictive features.

The first data conversion involved transforming text data from a string to a word vector for both attributes. This means that word vectors can be used to connect various text data to a target class. A word vector is a numeric vector that represents the context of a word in reality. Weka used the StringToWordVector function, which converts text data to a word vector and is included in the attribute's unsupervised filter.

However, you found during the initial conversion that the data had other features, such as punctuation and stop words, which you needed to remove from the data set after studying the effects of punctuation and making observations about how to improve the data mining model's efficiency. As a result, "0123456789 |# percent &$ * - + =., Delete a punctuation mark that usually includes punctuation, such as '"()?!" In the StringToWordVector tokenizer module, use the delimiter choices. Stop terms were also omitted from the dataset. A survey on how deleting a stop word affects various data mining models that data mining model success is being recorded told the decision to delete a stop word. The "Stop-word removal dataset of traditional sentiment classifier from raw dataset," which claims to affect efficiency, has been increased to the "Stop-word removal dataset of traditional sentiment classifier from raw dataset," so the multistopwords stopwordsremover function used in the stopwordsHandler function of the StringToWordVector function has been used to remove the stopwords.

Second Database

Additionally, StringToWordVector has a method for determining the number of words to hold after the filter is applied. We decided to hold about 3000 words for each second of data in the current sample. Our goal in maintaining up to 3,000 words was to provide as much information as possible while reducing computing resources including time and memory used to pre-process data and enforce the model. No stemming, however, was done. That is, you used the Nullstemmer option in the StringToWordVector function's steamer function.

Cleaning noise data will aid in data quality improvement. All characters in the data should be converted to lowercase, punctuation should be eliminated, and stop words and typos should be removed. To convert all words to lowercase, we used the lowerCaseTokens choice of the StringToWordVector feature.

Naive Bayes

1. We can see that the class that begins using Naive Bayes is untouched now that you have the attributes.
2. J48 is included in the second half of the decision tree. The following characteristics are left:
3. Then you use filtering to figure out which attributes to retain and which to get rid of. The first move is to use the CfsSubsetEval filtering system. This method leaves us with a few characteristics.
4. CorrelationAttributeEval is another filtering process. When you apply this filter to a dataset, the following attributes, in order of class effect, are left:
5. Based on this filter, use the.04 threshold. As a result, attributes with a value of less than.04 in a ranked list are overlooked.

Decision-tree (J48)

The data collection is analysed in the final stages of pre-processing to ensure that it is as reliable as a class forecast. The WEKA Categories tab is where you do this. You will see that the training data set is not very reliable in class prediction after running the classifier in J48 with it. Just 48.04 percent of cases were properly categorised. This is less than half, which does not indicate that the data collection is accurate. Gets the attribute to see if it's low, but the number is going down.

We first agreed to get rid of the gender attribute. This resulted in a 1% reduction in the figure. As a result, the gender of the respondents was not required to be included in the data collection. The usable band was the second feature to be deleted. The percentage fell by around 0.5 percent after dropping this attribute, which is also a positive indication that you don't need to use it in your dataset. An area has been deleted because of the following attribute. This had little impact on the number of correctly categorised cases, so deleting them was the right decision. Imd band is the last attribute to be removed. This did not have a significant impact on our overall classified percentage.

Since there are very few correctly categorised instances in this dataset, it is not appropriate for instruction. We conclude that using the final attributes and functions would result in certain repetitive results. These characteristics can be used to decide whether a function can help boost results, according to this data collection.

When there are so many input variables, the performance of machine learning algorithms will suffer. The input variable is a column that is fed as input to the model to predict the target variable whether data is expressed using rows and columns, as in a spreadsheet. Functions are another name for input variables. n-Dimensional Characteristics A column of data may be thought of as a dimension of space, and a row of data as a point in that space. The data collection can be interpreted geometrically in this way. Characteristics The presence of several dimensions in a space may indicate that the volume of the space is very high. Furthermore, the points in that space (data rows) are often thin, non-representative samples. This can have a big effect on the success of machine learning algorithms that have to fit into data with a lot of different input options, also known as "dimensional curses." As a result, reducing the number of input features is often beneficial. The term "Dimension Reduction" comes from the fact that it decreases the number of dimensions in the function space.

Hundreds, thousands, or even millions of input variables may be considered high dimensionality. Since there are no input dimensions, the machine learning model has fewer parameters and a simplified form, which is referred to as degrees of freedom. Models with so many degrees of freedom are more likely to overfit the training data collection, making them unsuitable for new data. A well-generalized basic model and input data with few input variables are ideal. This is particularly valid for linear models, which have a close relationship between the number of inputs and the model's degrees of freedom. The wrapper approach wraps a machine learning model, fits, and evaluates the model using various subsets of input features, and selects the subset with the best model output. A wrapper function selection system like RFE is an example. The filtering approach selects the most stable subset of input features using scoring measures such as the association between a function and a target variable. Pearson correlations and chi-squared experiments are two examples. These methods, often referred to as "manifold learning," are used to project high-dimensional data onto low-dimensional projections. This is often used to visualise details. The aim of projection is to construct a low-dimensional representation of the data set while preserving the data's incredible structure and relationships.

There is no one-size-fits-all solution to dimension reduction, and you can't map the solution to the problem. Instead, rigorous structured testing is the best choice for finding the best performing dimensional reduction tool in your data set when paired with your chosen model. Linear algebra and manifold learning usually presume that all input functions have the same size or distribution. If the size or units of the input variables vary, it's a safe idea to normalise or normalise the data when using these approaches. Too many missed values in a data column are unlikely to provide any valuable information. As a result, data columns with a missing value higher than the given threshold can be deleted. The more aggressive the reduction, the higher the threshold. Data columns with no change in data, similarly to the previous system, contain little detail. As a result, all data columns with a deviation below a certain threshold are discarded. Variance is determined by the range. As a result, prior to using this technique, normalisation is necessary. Very similar details can be found in data columns with very similar patterns. Just one of them is adequate to feed the machine learning algorithm in this situation. The Pearson product factor and the Pearson chi-square value are used to measure the correlation coefficient between the numeric and nominal columns, respectively. Column pairs with a connection number greater than the threshold are limited to one. Notice that correlations are scale dependent. As a result, for valid correlation comparisons, column normalisation is needed. The decision tree ensemble, also known as a random forest, is a powerful classifier that also assists with function selection. One method for reducing dimension is to create a broad, carefully built set of trees for target attributes and then use the utilisation statistics for each attribute to identify the most useful subset of characteristics. You will, for example, produce a huge number of very shallow trees (2000). (2 levels). Each tree is taught a small subset of the total number of attributes (3). If a function is often selected as the best division, it could be worth keeping. The score derived from a random forest's attribute consumption statistics is measured by comparing it to other attributes, which are the most predictable.

Conclusion

This data set acknowledges that it isn't a "healthy data set" for predicting whether reducing features would increase efficiency. The consistency of the data collection was not known until the project pre-processing began. The proposal must receive final approval before it can go forward. Provided certain characteristics, it was possible to determine whether a student passed or failed the class after pre-processing and data interpretation on this specific data collection. The final attributes chosen from a dataset were significant enough that one of them could estimate a class score. The support vector machine algorithm produced the best results with the training data set, but naive Bayes had the highest number of correctly categorised instances with the test data set. The highest percentage of accurately categorised instances was 48.56 percent overall. This also leads to the conclusion that the number of correctly identified instances in the dataset is very limited.

Finally, we will guarantee you that reducing functionality can help you improve efficiency. We ought to use a stronger dataset as a training and evaluation dataset because this was the main issue we had. Another option for data pre-processing is to use a dataset with a far higher, well-organized case.

References

Celeux, G. and Govaert, G. (2012). A classification EM algorithm for clustering and two stochastic versions. Computational statistics and data analysis, 14:315–332

Hans-Peter Kriegel, Peer Kröger, Jörg Sander, Arthur Zimek (2011). "Density-based Clustering". WIREs Data Mining and Knowledge Discovery 1 (3): 231–240. doi:10.1002/widm.30.

Ian, W. H. (2013, September 22). Retrieved April 10, 2021, from https://www.youtube.com/watch?v=l7R9NHqvI0Y

Luhaybi, Mashael Al, et al. “The Prediction of Student Failure Using Classification Methods: A Case Study.” Computer Science & Information Technology , 2018, doi:10.5121/csit.2018.80506.

Pang-Ning, T., Michael, S., & Vipin, K. (2006, March 25). Introduction to Data Mining. Retrieved from https://www-users.cs.umn.edu/~kumar001/dmbook/firsted.php

Need to wrap up assignments on time? Stringent deadlines getting the better of you? Our in-house academic papers writers are available round the clock to work on your assignments and share the same much ahead of the deadline. From offering Finance assignment help to backing you up with Law assignment help in London, we are right here to assist you through the thick and thin of assignment stringencies. So, the next time you would worry about a narrow deadline or wonder, “Can I pay someone to do my assignment on time?” count on us and never look back.