Dataset

This dataset was created by Mohanad Hazem Qabil

Contents

Dirty Retail Store Sales Dataset

Overview

The Dirty Retail Store Sales dataset contains 12,575 rows of synthetic data representing sales transactions from a retail store. The dataset includes eight product categories with 25 items per category, each having static prices. It is designed to simulate real-world sales data, including intentional "dirtiness" such as missing or inconsistent values. This dataset is suitable for practicing data cleaning, exploratory data analysis (EDA), and feature engineering.

File Information

• File Name: retail_store_sales.csv
• Number of Rows: 12,575
• Number of Columns: 11

Columns Description

Categories and Items

The dataset includes the following categories, each containing 25 items with corresponding codes, names, and static prices:

Electric Household Essentials

Furniture

Sample data for exercises in Further Adventures in Data Cleaning.

Dataset

This dataset was created by Martin Kanju

Released under Other (specified in description)

Contents

Dirty Cafe Sales Dataset

Overview

The Dirty Cafe Sales dataset contains 10,000 rows of synthetic data representing sales transactions in a cafe. This dataset is intentionally "dirty," with missing values, inconsistent data, and errors introduced to provide a realistic scenario for data cleaning and exploratory data analysis (EDA). It can be used to practice cleaning techniques, data wrangling, and feature engineering.

File Information

• File Name: dirty_cafe_sales.csv
• Number of Rows: 10,000
• Number of Columns: 8

Columns Description

Data Characteristics

1. Missing Values:

   • Some columns (e.g., Item, Payment Method, Location) may contain missing values represented as None or empty cells.

2. Invalid Values:

   • Some rows contain invalid entries like "ERROR" or "UNKNOWN" to simulate real-world data issues.

3. Price Consistency:

   • Prices for menu items are consistent but may have missing or incorrect values introduced.

Menu Items

The dataset includes the following menu items with their respective price ranges:

Use Cases

This dataset is suitable for: - Practicing data cleaning techniques such as handling missing values, removing duplicates, and correcting invalid entries. - Exploring EDA techniques like visualizations and summary statistics. - Performing feature engineering for machine learning workflows.

Cleaning Steps Suggestions

To clean this dataset, consider the following steps: 1. Handle Missing Values: - Fill missing numeric values with the median or mean. - Replace missing categorical values with the mode or "Unknown."

1. Handle Invalid Values:

   • Replace invalid entries like "ERROR" and "UNKNOWN" with NaN or appropriate values.

2. Date Consistency:

   • Ensure all dates are in a consistent format.
   • Fill missing dates with plausible values based on nearby records.

3. Feature Engineering:

   • Create new columns, such as Day of the Week or Transaction Month, for further analysis.

License

This dataset is released under the CC BY-SA 4.0 License. You are free to use, share, and adapt it, provided you give appropriate credit.

Feedback

If you have any questions or feedback, feel free to reach out through the dataset's discussion board on Kaggle.

Access and clean an open source herbarium dataset using Excel or RStudio.

In this project, we work on repairing three datasets:

• Trials design: This dataset was obtained from the European Union Drug Regulating Authorities Clinical Trials Database (EudraCT) register and the ground truth was created from external registries. In the dataset, multiple countries, identified by the attribute country_protocol_code, conduct the same clinical trials which is identified by eudract_number. Each clinical trial has a title that can help find informative details about the design of the trial.
• Trials population: This dataset delineates the demographic origins of participants in clinical trials primarily conducted across European countries. This dataset include structured attributes indicating whether the trial pertains to a specific gender, age group or healthy volunteers. Each of these categories is labeled as (`1') or (`0') respectively denoting whether it is included in the trials or not. It is important to note that the population category should remain consistent across all countries conducting the same clinical trial identified by an eudract_number. The ground truth samples in the dataset were established by aligning information about the trial populations provided by external registries, specifically the CT.gov database and the German Trials database. Additionally, the dataset comprises other unstructured attributes that categorize the inclusion criteria for trial participants such as inclusion.
• Allergens: This dataset contains information about products and their allergens. The data was collected from the German version of the `Alnatura' (Access date: 24 November, 2020), a free database of food products from around the world `Open Food Facts', and the websites: `Migipedia', 'Piccantino', and `Das Ist Drin'. There may be overlapping products across these websites. Each product in the dataset is identified by a unique code. Samples with the same code represent the same product but are extracted from a differentb source. The allergens are indicated by (‘2’) if present, or (‘1’) if there are traces of it, and (‘0’) if it is absent in a product. The dataset also includes information on ingredients in the products. Overall, the dataset comprises categorical structured data describing the presence, trace, or absence of specific allergens, and unstructured text describing ingredients.

N.B: Each '.zip' file contains a set of 5 '.csv' files which are part of the afro-mentioned datasets:

• "{dataset_name}_train.csv": samples used for the ML-model training. (e.g "allergens_train.csv")
• "{dataset_name}_test.csv": samples used to test the the ML-model performance. (e.g "allergens_test.csv")
• "{dataset_name}_golden_standard.csv": samples represent the ground truth of the test samples. (e.g "allergens_golden_standard.csv")
• "{dataset_name}_parker_train.csv": samples repaired using Parker Engine used for the ML-model training. (e.g "allergens_parker_train.csv")
• "{dataset_name}_parker_train.csv": samples repaired using Parker Engine used to test the the ML-model performance. (e.g "allergens_parker_test.csv")

All data are prone to error and require data cleaning prior to analysis. An important example is longitudinal growth data, for which there are no universally agreed standard methods for identifying and removing implausible values and many existing methods have limitations that restrict their usage across different domains. A decision-making algorithm that modified or deleted growth measurements based on a combination of pre-defined cut-offs and logic rules was designed. Five data cleaning methods for growth were tested with and without the addition of the algorithm and applied to five different longitudinal growth datasets: four uncleaned canine weight or height datasets and one pre-cleaned human weight dataset with randomly simulated errors. Prior to the addition of the algorithm, data cleaning based on non-linear mixed effects models was the most effective in all datasets and had on average a minimum of 26.00% higher sensitivity and 0.12% higher specificity than other methods. Data cleaning methods using the algorithm had improved data preservation and were capable of correcting simulated errors according to the gold standard; returning a value to its original state prior to error simulation. The algorithm improved the performance of all data cleaning methods and increased the average sensitivity and specificity of the non-linear mixed effects model method by 7.68% and 0.42% respectively. Using non-linear mixed effects models combined with the algorithm to clean data allows individual growth trajectories to vary from the population by using repeated longitudinal measurements, identifies consecutive errors or those within the first data entry, avoids the requirement for a minimum number of data entries, preserves data where possible by correcting errors rather than deleting them and removes duplications intelligently. This algorithm is broadly applicable to data cleaning anthropometric data in different mammalian species and could be adapted for use in a range of other domains.

Netflix is a popular streaming service that offers a vast catalog of movies, TV shows, and original contents. This dataset is a cleaned version of the original version which can be found here. The data consist of contents added to Netflix from 2008 to 2021. The oldest content is as old as 1925 and the newest as 2021. This dataset will be cleaned with PostgreSQL and visualized with Tableau. The purpose of this dataset is to test my data cleaning and visualization skills. The cleaned data can be found below and the Tableau dashboard can be found here .

Data Cleaning

We are going to: 1. Treat the Nulls 2. Treat the duplicates 3. Populate missing rows 4. Drop unneeded columns 5. Split columns Extra steps and more explanation on the process will be explained through the code comments

--View dataset

SELECT * 
FROM netflix;

--The show_id column is the unique id for the dataset, therefore we are going to check for duplicates
                                  
SELECT show_id, COUNT(*)                                                                                      
FROM netflix 
GROUP BY show_id                                                                                              
ORDER BY show_id DESC;

--No duplicates

--Check null values across columns

SELECT COUNT(*) FILTER (WHERE show_id IS NULL) AS showid_nulls,
    COUNT(*) FILTER (WHERE type IS NULL) AS type_nulls,
    COUNT(*) FILTER (WHERE title IS NULL) AS title_nulls,
    COUNT(*) FILTER (WHERE director IS NULL) AS director_nulls,
    COUNT(*) FILTER (WHERE movie_cast IS NULL) AS movie_cast_nulls,
    COUNT(*) FILTER (WHERE country IS NULL) AS country_nulls,
    COUNT(*) FILTER (WHERE date_added IS NULL) AS date_addes_nulls,
    COUNT(*) FILTER (WHERE release_year IS NULL) AS release_year_nulls,
    COUNT(*) FILTER (WHERE rating IS NULL) AS rating_nulls,
    COUNT(*) FILTER (WHERE duration IS NULL) AS duration_nulls,
    COUNT(*) FILTER (WHERE listed_in IS NULL) AS listed_in_nulls,
    COUNT(*) FILTER (WHERE description IS NULL) AS description_nulls
FROM netflix;

We can see that there are NULLS. 
director_nulls = 2634
movie_cast_nulls = 825
country_nulls = 831
date_added_nulls = 10
rating_nulls = 4
duration_nulls = 3 

The director column nulls is about 30% of the whole column, therefore I will not delete them. I will rather find another column to populate it. To populate the director column, we want to find out if there is relationship between movie_cast column and director column

-- Below, we find out if some directors are likely to work with particular cast

WITH cte AS
(
SELECT title, CONCAT(director, '---', movie_cast) AS director_cast 
FROM netflix
)

SELECT director_cast, COUNT(*) AS count
FROM cte
GROUP BY director_cast
HAVING COUNT(*) > 1
ORDER BY COUNT(*) DESC;

With this, we can now populate NULL rows in directors 
using their record with movie_cast 

UPDATE netflix 
SET director = 'Alastair Fothergill'
WHERE movie_cast = 'David Attenborough'
AND director IS NULL ;

--Repeat this step to populate the rest of the director nulls
--Populate the rest of the NULL in director as "Not Given"

UPDATE netflix 
SET director = 'Not Given'
WHERE director IS NULL;

--When I was doing this, I found a less complex and faster way to populate a column which I will use next

Just like the director column, I will not delete the nulls in country. Since the country column is related to director and movie, we are going to populate the country column with the director column

--Populate the country using the director column

SELECT COALESCE(nt.country,nt2.country) 
FROM netflix AS nt
JOIN netflix AS nt2 
ON nt.director = nt2.director 
AND nt.show_id <> nt2.show_id
WHERE nt.country IS NULL;
UPDATE netflix
SET country = nt2.country
FROM netflix AS nt2
WHERE netflix.director = nt2.director and netflix.show_id <> nt2.show_id 
AND netflix.country IS NULL;


--To confirm if there are still directors linked to country that refuse to update

SELECT director, country, date_added
FROM netflix
WHERE country IS NULL;

--Populate the rest of the NULL in director as "Not Given"

UPDATE netflix 
SET country = 'Not Given'
WHERE country IS NULL;

The date_added rows nulls is just 10 out of over 8000 rows, deleting them cannot affect our analysis or visualization

--Show date_added nulls

SELECT show_id, date_added
FROM netflix_clean
WHERE date_added IS NULL;

--DELETE nulls

DELETE F...

This dataset includes all of the data downloaded from GBIF (DOIs provided in README.md as well as below, downloaded Feb 2021) as well as data downloaded from SCAN. This dataset has 2,808,432 records and can be used as a reference to the verbatim data before it underwent the cleaning process. The only modifications made to this datset after direct download from the data portals are the following:

1) for GBIF records, I renamed the countryCode column to be "country" so that the column title is consistent across both GBIF and SCAN 2) A source column was added where I specify if the record came from GBIF or SCAN 3) Duplicate records across SCAN and GBIF were removed by identifying identical instances "catalogNumber" and "institutionCode" 4) Only the Darwin core columns (DwC) that were shared across downloaded datasets were retained. GBIF contained ~249 DwC variables, and SCAN data contained fewer, so this combined dataset only includes the ~80 columns shared between the two datasets

For GBIF, we downloaded the data in three separate chunks, therefore there are three DOIs. See below:

GBIF.org (3 February 2021) GBIF Occurrence Downloadhttps://doi.org/10.15468/dl.6cxfsw GBIF.org (3 February 2021) GBIF Occurrence Downloadhttps://doi.org/10.15468/dl.b9rfa7 GBIF.org (3 February 2021) GBIF Occurrence Downloadhttps://doi.org/10.15468/dl.w2nndm

The dissertation_demo.zip contains the base code and demonstration purpose for the dissertation: A Conceptual Model for Transparent, Reusable, and Collaborative Data Cleaning. Each chapter has a demo folder for demonstrating provenance queries or tools. The Airbnb dataset for demonstration and simulation is not included in this demo but is available to access directly from the reference website. Any updates on demonstration and examples can be found online at: https://github.com/nikolausn/dissertation_demo

Data cleaning is one of the most important but time-consuming tasks for data scientists. The data cleaning task consists of two major steps: (1) error detection and (2) error correction. The goal of error detection is to identify wrong data values. The goal of error correction is to fix these wrong values. Data cleaning is a challenging task due to the trade-off among correctness, completeness, and automation. In fact, detecting/correcting all data errors accurately without any user involvement is not possible for every dataset. We propose a novel data cleaning approach that detects/corrects data errors with a novel two-step task formulation. The intuition is that, by collecting a set of base error detectors/correctors that can independently mark/fix data errors, we can learn to combine them into a final set of data errors/corrections using a few informative user labels. First, each base error detector/corrector generates an initial set of potential data errors/corrections. Then, the approach ensembles the output of these base error detectors/correctors into one final set of data errors/corrections in a semi-supervised manner. In fact, the approach iteratively asks the user to annotate a tuple, i.e., marking/fixing a few data errors. The approach learns to generalize the user-provided error detection/correction examples to the rest of the dataset, accordingly. Our novel two-step formulation of the error detection/correction task has four benefits. First, the approach is configuration free and does not need any user-provided rules or parameters. In fact, the approach considers the base error detectors/correctors as black-box algorithms that are not necessarily correct or complete. Second, the approach is effective in the error detection/correction task as its first and second steps maximize recall and precision, respectively. Third, the approach also minimizes human involvement as it samples the most informative tuples of the dataset for user labeling. Fourth, the task formulation of our approach allows us to leverage previous data cleaning efforts to optimize the current data cleaning task. We design an end-to-end data cleaning pipeline according to this approach that takes a dirty dataset as input and outputs a cleaned dataset. Our pipeline leverages user feedback, a set of data cleaning algorithms, and a set of previously cleaned datasets, if available. Internally, our pipeline consists of an error detection system (named Raha), an error correction system (named Baran), and a transfer learning engine. As our extensive experiments show, our data cleaning systems are effective and efficient, and involve the user minimally. Raha and Baran significantly outperform existing data cleaning approaches in terms of effectiveness and human involvement on multiple well-known datasets.

Dataset

This dataset was created by Shiva Vashishtha

Contents

Dataset

This dataset was created by Hussein Al Chami

Released under MIT

Contents

The dataset has been obtained by web scraping a Wikipedia page for which code is linked below: https://www.kaggle.com/amruthayenikonda/simple-web-scraping-using-pandas

This dataset can be used to practice data cleaning and manipulation for example dropping of unwanted columns, null vales, removing symbols etc

This CNVVE Dataset contains clean audio samples encompassing six distinct classes of voice expressions, namely “Uh-huh” or “mm-hmm”, “Uh-uh” or “mm-mm”, “Hush” or “Shh”, “Psst”, “Ahem”, and Continuous humming, e.g., “hmmm.” Audio samples of each class are found in the respective folders. These audio samples have undergone a thorough cleaning process. The raw samples are published in https://doi.org/10.18419/darus-3897. Initially, we applied the Google WebRTC voice activity detection (VAD) algorithm on the given audio files to remove noise or silence from the collected voice signals. The intensity was set to "2", which could be a value between "1" and "3". However, because of variations in the data, some files required additional manual cleaning. These outliers, characterized by sharp click sounds (such as those occurring at the end of recordings), were addressed. The samples are recorded through a dedicated website for data collection that defines the purpose and type of voice data by providing example recordings to participants as well as the expressions’ written equivalent, e.g., “Uh-huh”. Audio recordings were automatically saved in the .wav format and kept anonymous, with a sampling rate of 48 kHz and a bit depth of 32 bits. For more info, please check the paper or feel free to contact the authors for any inquiries.

CodeParrot 🦜 Dataset Cleaned

  What is it?

A dataset of Python files from Github. This is the deduplicated version of the codeparrot.

  Processing

The original dataset contains a lot of duplicated and noisy data. Therefore, the dataset was cleaned with the following steps:

Deduplication Remove exact matches

Filtering Average line length < 100 Maximum line length < 1000 Alpha numeric characters fraction > 0.25 Remove auto-generated files (keyword search)

For… See the full description on the dataset page: https://huggingface.co/datasets/codeparrot/codeparrot-clean.

Instagram data-download example dataset

In this repository you can find a data-set consisting of 11 personal Instagram archives, or Data-Download Packages (DDPs).

How the data was generated

These Instagram accounts were all new and generated by a group of researchers who were interested to figure out in detail the structure and variety in structure of these Instagram DDPs. The participants user the Instagram account extensively for approximately a week. The participants also intensively communicated with each other so that the data can be used as an example of a network.

The data was primarily generated to evaluate the performance of de-identification software. Therefore, the text in the DDPs particularly contain many randomly chosen (Dutch) first names, phone numbers, e-mail addresses and URLS. In addition, the images in the DDPs contain many faces and text as well. The DDPs contain faces and text (usernames) of third parties. However, only content of so-called `professional accounts' are shared, such as accounts of famous individuals or institutions who self-consciously and actively seek publicity, and these sources are easily publicly available. Furthermore, the DDPs do not contain sensitive personal data of these individuals.

Obtaining your Instagram DDP

After using the Instagram accounts intensively for approximately a week, the participants requested their personal Instagram DDPs by using the following steps. You can follow these steps yourself if you are interested in your personal Instagram DDP.

1. Go to www.instagram.com and log in
2. Click on your profile picture, go to Settings and Privacy and Security
3. Scroll to Data download and click Request download
4. Enter your email adress and click Next
5. Enter your password and click Request download

Instagram then delivered the data in a compressed zip folder with the format username_YYYYMMDD.zip (i.e., Instagram handle and date of download) to the participant, and the participants shared these DDPs with us.

Data cleaning

To comply with the Instagram user agreement, participants shared their full name, phone number and e-mail address. In addition, Instagram logged the i.p. addresses the participant used during their active period on Instagram. After colleting the DDPs, we manually replaced such information with random replacements such that the DDps shared here do not contain any personal data of the participants.

How this data-set can be used

This data-set was generated with the intention to evaluate the performance of the de-identification software. We invite other researchers to use this data-set for example to investigate what type of data can be found in Instagram DDPs or to investigate the structure of Instagram DDPs. The packages can also be used for example data-analyses, although no substantive research questions can be answered using this data as the data does not reflect how research subjects behave `in the wild'.

Authors

The data collection is executed by Laura Boeschoten, Ruben van den Goorbergh and Daniel Oberski of Utrecht University. For questions, please contact l.boeschoten@uu.nl.

Acknowledgments

The researchers would like to thank everyone who participated in this data-generation project.

The National Health and Nutrition Examination Survey (NHANES) provides data and have considerable potential to study the health and environmental exposure of the non-institutionalized US population. However, as NHANES data are plagued with multiple inconsistencies, processing these data is required before deriving new insights through large-scale analyses. Thus, we developed a set of curated and unified datasets by merging 614 separate files and harmonizing unrestricted data across NHANES III (1988-1994) and Continuous (1999-2018), totaling 135,310 participants and 5,078 variables. The variables conveydemographics (281 variables),dietary consumption (324 variables),physiological functions (1,040 variables),occupation (61 variables),questionnaires (1444 variables, e.g., physical activity, medical conditions, diabetes, reproductive health, blood pressure and cholesterol, early childhood),medications (29 variables),mortality information linked from the National Death Index (15 variables),survey weights (857 variables),environmental exposure biomarker measurements (598 variables), andchemical comments indicating which measurements are below or above the lower limit of detection (505 variables).csv Data Record: The curated NHANES datasets and the data dictionaries includes 23 .csv files and 1 excel file.The curated NHANES datasets involves 20 .csv formatted files, two for each module with one as the uncleaned version and the other as the cleaned version. The modules are labeled as the following: 1) mortality, 2) dietary, 3) demographics, 4) response, 5) medications, 6) questionnaire, 7) chemicals, 8) occupation, 9) weights, and 10) comments."dictionary_nhanes.csv" is a dictionary that lists the variable name, description, module, category, units, CAS Number, comment use, chemical family, chemical family shortened, number of measurements, and cycles available for all 5,078 variables in NHANES."dictionary_harmonized_categories.csv" contains the harmonized categories for the categorical variables.“dictionary_drug_codes.csv” contains the dictionary for descriptors on the drugs codes.“nhanes_inconsistencies_documentation.xlsx” is an excel file that contains the cleaning documentation, which records all the inconsistencies for all affected variables to help curate each of the NHANES modules.R Data Record: For researchers who want to conduct their analysis in the R programming language, only cleaned NHANES modules and the data dictionaries can be downloaded as a .zip file which include an .RData file and an .R file.“w - nhanes_1988_2018.RData” contains all the aforementioned datasets as R data objects. We make available all R scripts on customized functions that were written to curate the data.“m - nhanes_1988_2018.R” shows how we used the customized functions (i.e. our pipeline) to curate the original NHANES data.Example starter codes: The set of starter code to help users conduct exposome analysis consists of four R markdown files (.Rmd). We recommend going through the tutorials in order.“example_0 - merge_datasets_together.Rmd” demonstrates how to merge the curated NHANES datasets together.“example_1 - account_for_nhanes_design.Rmd” demonstrates how to conduct a linear regression model, a survey-weighted regression model, a Cox proportional hazard model, and a survey-weighted Cox proportional hazard model.“example_2 - calculate_summary_statistics.Rmd” demonstrates how to calculate summary statistics for one variable and multiple variables with and without accounting for the NHANES sampling design.“example_3 - run_multiple_regressions.Rmd” demonstrates how run multiple regression models with and without adjusting for the sampling design.

Note. Where frequencies do not add up to the total sample size, data were missing. For religious affiliation, C = Christianity, J = Judaism, I = Islam, H = Hinduism, B = Buddhism, O = Other Religion, N = No Religion. No Religion includes both participants who self-identified as not having any religious affiliation and those who did not report anything to the item on religious affiliation.Description of Demographic Characteristics of Samples After Data Cleaning.