程序代写代做代考 COMP3430 / COMP8430 Data wrangling – cscodehelp代写
COMP3430 / COMP8430 Data wrangling
Lab 6: Evaluation for Record Linkage
Objectives of this lab
● Today’s lab is the fourth in a series of five labs during which we will gradually build a complete record linkage system.
● We will be working with different evaluation measures and learn how they work and why they are important in the RL process.
● Completion of the evaluation module in the overall system.
Outline of this lab
● Learn how different evaluation measures work
● Implement different evaluation measures
● Explore and experiment with different evaluation measures
● Summary
Preliminaries
● Before you begin, aim to review lecture 19 if you have not already viewed them.
● Go back over the work from lab 5 and remind yourself what we were doing and how the overall program is structured.
● You can download the classification module with sample solutions in week 8 and use it with your RL program if you find difficulties implementing the required classification techniques.
What is evaluation?
● This week we focus on the next step in the linkage process, evaluation.
● The aim of a evaluation metric is to measure the performance of a RL process and see how well it has linked the data sets.
● Why do you think we need different evaluation measures and are they equally important?
How to evaluate a linkage process
● Before we begin let us see how different evaluation metrics work. The evaluation measures are described in lecture 19.
Predicted Matches
Predicted Non-matches
True Matches
1,000
400
True Non-matches
600
8,000
Predicted Matches
Predicted Non-matches
True Matches
1,200
200
True Non-matches
800
7,800
● See if you can calculate the following measures from the above two confusion matrices:
1. Accuracy
2. Precision 3. Recall
How to evaluate a linkage process
● Accuracy= (TP+TN)/(TP+FP+FN+TN) = 1000 + 8000 / 10000
= 0.9
● Precision = TP / (TP + FP)
= 1000 / (1000 + 600)
= 0.625
● Recall = TP/(TP+FN)
= 1000 / (1000 + 400) = 0.7143
Predicted Matches
Predicted Non-matches
True Matches
1,000 (TP)
400 (FN)
True Non-matches
600 (FP)
8,000 (TN)
How to evaluate a linkage process
● Accuracy= (TP+TN)/(TP+FP+FN+TN) = 1200 + 7800 / 10000
= 0.9
● Precision = TP / (TP + FP)
= 1200 / (1200 + 800)
= 0.6
● Recall = TP/(TP+FN)
= 1200 / (1200 + 200) = 0.8571
Predicted Matches
Predicted Non-matches
True Matches
1,200 (TP)
200 (FN)
True Non-matches
800 (FP)
7,800 (TN)
Implement different evaluation measures
● Now start looking at evaluation.py and explore how the evaluation functions work (inputs, return values, etc.).
● We have already provided two evaluation functions, accuracy() and reduction_ratio().
● Run the RL program with different settings and see what the output of these two functions look like and how they perform.
● Now try to implement the other evaluation metrics as required in the lab tutorial document.
Questions to consider
● Are there any measures that are not useful, either because they are always extremely high, or low, or difficult to calculate, etc?
● What is the impact of the data quality on the linkage results? Does this vary depending on which functions you use for the blocking, comparison, and classification steps?
● What effect do the different blocking techniques have on the final record linkage results?
● Extra task – Run the RL program with different data sets provided.
Summary
● In this lab we implemented different evaluation measures and learnt how they can be used in the RL program to evaluate it performs.
● Make sure to complete any unfinished work in this module before you come to the next lab.
● In the next lab we will be conducting experiments with more data sets with different sizes and data quality.
