writing report(2 pages) about Data mining using Weka software

STAT390-14B (Ham): Directed Study Project

Individual Project Focus: Work vs. Play

Project co-ordinator: Associate Professor David Bainbridge

Process the weather data for Auckland in January and February in the given

dataset (10 minute readings) and experiment with various data mining

techniques to see if a model can be generated that predicts power

consumption for Monday-Friday (work), Saturday, and Sunday (play). Is it

easier to predict the power usage for one of time periods? Trial having

Saturday and Sunday represented as a single entity (i.e. the weekend) and as

separate days.

The aim of this directed study project is combine the programming skills learnt in COMP5002 (BoPP)

with the Data Mining MOOCs that were studied earlier in the semester at Waikato, and the

JavaScript skills for web use taught in COMP 223 (TGA) in the A semester of this year.

The central theme to this project—shared across all the projects being run in this course—is to

investigate the relationship between power usage in New Zealand, and chronological data (the time

of day and the time of year) and meteorological data (the weather!) to see if any patterns exist;

more specifically, to see whether the latter information helps predict the former. Each project

investigates a separate aspect within this theme, applying Data Mining techniques to publically

available data produced by both Transpower and the National Institute of Water and Atmospheric

Research (NIWA), from which a range of visualizations will be generated.

The key steps to the project are:

1. Undertake data cleaning and processing of a rich dataset containing information that

captures power consumption and climatic conditions recorded across various regions of

New Zealand.

2. Feed the processed data into Weka to undertake a data modelling task.

3. Produce a set of visualizations that provide insight into the generated data

Two types of visualization will be produced: the first is focused on showing how well the predictive

modelling is performing; the second is a more open-ended task, with the aim of showing “something

interesting” in the data related to the project’s focus. An example of “something interesting” could

be a time-based geographical map showing power usage in the different regions of New Zealand

enriched with what is happening in terms of temperature in the different regions. For further ideas

see Prof Apperley’s Data Visualization slides, available through the STAT390 web site:

www.cs.waikato.ac.nz/~davidb/stat390/

The dataset provided for this project (also available for download through the same web site) is in

the form of a set of Comma Separated Value (CSV) files. The files span a mixture of years and

locations within New Zealand. While each project is different, there is one common dimension to

how the data is to be used:

 Data pertaining to the years 2011 and 2012 form the basis for training the Data Mining

models;

 Data pertaining to the year 2013 forms the basis for establishing the accuracy of the models

developed.

We will now go through and detail what is involved in the three keys steps to the project. The

schedule (see below) allows 1 week for each of these steps, although it should be noted there is

some flexibility around this, as long as the final deadlines—a presentation and a report, due in the

final week—are met. If at any point during the project you wish to go back to an earlier step and

revise/adjust what you have done, this is not only permissible it is actively encouraged (!), as it

reflects an increased level of understanding. At the end of each week, a 2–3 page “mini” report is

requested describing the work you have done that corresponds to the relevant step in the schedule.

The intention of each mini report is to help you develop a section of the final report. Feedback on

mini reports submitted according to the schedule will be given to assist you in developing the final

report.

Step 1: Data processing and cleaning

One of the first things you will need to do in this project is to process the provided dataset into a

more amenable form, reducing it down to just the data values that are meaningful to your project.

Example C code is provided on the course web site for reading in CSV files, breaking each line into

individual fields, and then writing out a selection of those fields.

The code you need to write needs to go beyond this. The fields you select will be motivated by what

type of data you have been directed to focus on for the Data Mining step. You will also need to

develop ways of controlling which lines of the CSV files make it through to the next stage of

processing: filtered, for example by time, or location—the exact details again will be determined by

the task you have been assigned in your project.

There are also undefined values to be aware of. These are typically represented as a hyphen (-) in

the CSV files. Sometimes you might find an entire column will consist of hyphens (for the particular

lines of the data you have filtered down to), other times most of the values will be there, with only

the occasional hyphen.

In preparing the way for the Data Mining step, something you might consider doing is to merge data:

fields (either in rows or columns). For example, 12 power readings taken every 5 minutes could be

combined to provide an hourly figure instead, which would fit more nicely with weather data

reported every hour.

Your 2-3 page “mini” report for this step of the project should detail the decisions you made in how

the data needed to be processed, and how that was accomplished.

Step 2: Data Mining with Weka

The second step to the project is to load the processed data into Weka and start experimenting with

the data to develop a model that can predict power usage. To reiterate what was stated above, use

data from 2011 and 2012 to train your models, and then run it on the 2013 values (test data) to

establish how accurate the predictions are. While a technique such as 10-fold cross-validation is a

quick and convenient way to gauge how well a model is performing (in general)—and you may very

well use this in early stages of testing—the needs of this project is to produce a model that can go on

to be used to make predictions on other (previously unseen) data.

It is anticipated that the Explorer tool will be the most likely sub-system you will work with in Weka,

and within that the Classifier section; however, there are no hard-and-fast rules here. Use what you

have learned in the Data Mining MOOCs wisely. When data is “flying around” at speed it is easy to

overlook important details that turns what would otherwise be a highly successful model to

garbage! Similarly, accidentally including the feature you wish to learn in the set of attributes used

to train a model is a mistake that is easy enough to make if you are not careful. In such cases it leads

to results that are amazingly high. If the result you are getting seem “too good to be true” … that

might very well be exactly what is going on! In short, “know your data.”

The key ability for this stage of the project is to be able to train a model on the 2011 and 2012 data,

from which a run can be made against a test set (2013) with the predictions made saved in a

machine readable form, ready for processing by Step 3 (Data Visualization)

The “mini” report for this step of the project should provide an overview of the different methods

you experimented with, along with the one that you established performed the best, and the

reasons for why that was.

Step 3: Data Visualization

There are two parts to the data visualization step.

 The first is to produce a set of visualizations using Google Charts which shows how well the

chosen Data Mining model is performing in making predictions about power usage;

 The second is a more “open ended” task and may use other visualization software if desired:

the key requirement for this part of the project is to visually show something interesting

about the cleaned up and processed data that has been produced.

Data Model Accuracy visualized with Google Charts

Google Charts (https://developers.google.com/chart/) is a web-based technology for presenting data

in a variety of forms. There are over 25 standard forms to choose from. See Google’s web site for

extensive documentation, and the course web site for some selected examples that are more closely

aligned with the needs of the Directed Study project.

Produce “Something Interesting”

For this final part of the project you might choose to visualize something interesting that has been

produced as a result of your experimentation using Weka, but equally it might be something that is

already present in the data produced in Step 1 of the project (no Data Mining required).

The overall intention for this part of the project is to think back to (and look back at, since the slides

are on the course web site!) the Data Visualization examples given by Prof Apperley in the first week

of the project, and be inspired by this to produce a visualization that shows “something interesting”

in the dataset you have been working with.

To achieve this, the scope for this part of the project can be widened. For example, if the focus of

the project had been to compare extremes of latitude, using Auckland and Invercargill as the two

extremes, then in the “something interesting” visualization it is permissible to broaden this to other

centres: the visualization produced could be, say, a map of New Zealand showing power-usage and

temperature data across all the main centres of population in the country, over time. Going further,

if the visuals drawn on the map per centre make more sense if normalized by city population, then

this information too can be added in to the dataset used to produce the visualization.

Given such an open ended brief, if you are at all unsure what to attempt for this final part of the

project then please consult with me for guidance as to what is a reasonable expectation.

As a final remark, for this visualization do not feel constrained to working with Google Charts

(although that is a valid option). There are several interactive Data Visualizations resources on-line,

such as IBM’s ManyEyes, (http://www.ibm.com/manyeyes) that allow you to upload datasets to

their web-site from which you can then develop your visualizations.

Schedule

 Wed 22/10/14: Progress reports by each student in class (4-5pm G1.15)

 Fri 24/10/14: Mini-report on Step 1 submitted

 Mon 27/10/14: Feedback on mini-reports can be collected from department office

 Wed 29/10/14: Progress reports by each student in class (4-5pm G1.15)

 Fri 31/10/14: Mini-report on Step 2 submitted

 Mon 03/11/14: Feedback on mini-reports can be collected from department office

Deadlines

 Wed 05/10/14: Presentations on Data Visualizations produced (3-5pm G1.15)

 Fri 07/10/14: Final Reports due (5pm)