Data cleaning is the most critical task during the data gathering and preparation phase of the AI system development process. The principle of "garbage in, garbage out" (GIGO) is fundamental to AI and machine learning; the quality of the input data directly dictates the performance, accuracy, and reliability of the resulting model. Raw data is often incomplete, inconsistent, or contains errors. Without a rigorous cleaning process to address these issues, any subsequent steps, including model training and validation, will be built on a flawed foundation, leading to a biased or ineffective AI system.

A. Stratifying the data: This is a sampling technique used to ensure representative data subsets. It is an important step but is secondary to ensuring the fundamental quality and accuracy of the data itself.

B. Isolating the system: This relates to security or operational deployment (e.g., sandboxing) and is not a task within the data gathering and preparation phase of the AI development lifecycle.

D. Training the system: Training is a distinct, subsequent phase in the AI lifecycle that relies on the successful completion of data gathering, cleaning, and preparation.

1. Zhang

Z.

Zhao

Z.

Zheng

Y.

Cui

Z.

& Chen

E. (2023). A Survey on Data Quality for Trustworthy AI. IEEE Transactions on Knowledge and Data Engineering

1-21. DOI: https://doi.org/10.1109/TKDE.2023.3324511. The paper emphasizes that high-quality data is a cornerstone for building trustworthy AI

and data cleaning is a primary method to improve data quality.

2. Bishop

C. M. (2006). Pattern Recognition and Machine Learning. Springer. In Chapter 1

the text introduces the machine learning workflow

highlighting the essential nature of a "pre-processing" stage (p. 4) to transform raw data into a suitable format

which includes cleaning and normalization

before any model training can occur.

3. Hastie

T.

Tibshirani

R.

& Friedman

J. (2009). The Elements of Statistical Learning: Data Mining

Inference

and Prediction. Springer. The book implicitly and explicitly discusses the need for data preparation and handling issues like missing values throughout its chapters

establishing it as a foundational step before applying learning algorithms.