Quelles sont les erreurs courantes à éviter lors de la sélection de modèles pour le nettoyage des données de Machine Learning ?

Before choosing the model for data cleaning, the things we should look for are: 1. Feature engineering-Not considering the importance of features can lead to the removal of critical information or introduce bias 2. Domain knowledge-Without looking for domain can results of inappropriate data cleaning techniques or the exclusion of important information 3. Missing values and outliers handling-Before removing or considering the missing values and outliers always look for the domain according to the problem 4.Too much relying on automation-Automated tools with human intelligence to ensure a comprehensive and problem-aware data cleaning process 5. Validation-Before generalizing the model, validation is important to remove misguided assumptions

Things to consider are: - Bias in regards to sensitive topics such as race, gender, religion, etc. - Nulls or missing values - Personally Identifiable Information (PII) - Proprietary, classified, or privileged information - Total size of data (too little affects proper analysis, and too much will require processing time) - Format of existing data (CSV, SQL, manual, etc.) - Entity relationships, views, structure of tables - Last updated date (is it fresh?)

Skipping EDA can lead to poorly informed model choices that don’t address specific data issues. For instance, a retail company may ignore unusual sales spikes during holidays if they don’t examine data trends, leading to inaccurate sales forecasts. A practical approach involves using simple statistical tools and visualizations like histograms or box plots to spot these anomalies or trends. This initial analysis helps tailor cleaning techniques, such as adjusting for seasonal effects or treating outliers, ensuring the data accurately reflects the underlying trends.

One thing that I found helpful is to make sure that your models are valid for real world data. In many cases we use synthetic data in order to approximate real world data. For example street traffic images generated in a lab than the once’s in real life. You need to make sure that your testing/validation data fall the same population as the real world data that you will deploy your model against. So even if you use a hold back set , a validation set ,and cross validation. But this is out of the the actual data in you real world application population, then your accuracy may decrease significantly. You need to watch out for this issue especially if you are doing data augmentation

Assuming data cleaning models are always correct can mislead future analysis. For example, a model might inaccurately fill missing income data in a survey, skewing results towards average values. To avoid such pitfalls, it's crucial to validate models using techniques like cross-validation or hold-out sets. This process involves periodically checking the model’s effectiveness on unseen data, ensuring it performs well consistently and not just on specific subsets. Regular validation helps detect overfitting, underfitting, or biases, allowing adjustments to model parameters for more reliable outcomes.

Ignoring the specific context of data can lead to ineffective cleaning. For instance, in healthcare, data on patient symptoms needs careful treatment. Automatic removal of seemingly redundant data, like repeated symptom entries, might overlook the significance of symptom persistence in diagnosing chronic conditions. It’s vital to adapt data cleaning methods to the unique aspects of each field. Engaging with domain experts to understand nuances and setting tailored cleaning rules ensure that important information is retained and accurately represented, enhancing the quality and reliability of data analysis.

Using Jupyter notebooks allows a data scientist to include markdown text alongside executable code. These can be easily shared since the results are saved inside the document itself. Writing out what the assumptions, steps, and process that went into your analysis is an important part of any scientific endeavor.

Lack of documentation in data cleaning leads to confusion and errors down the line. For example, if a financial analyst cannot access the rationale behind data adjustments in a loan application dataset, they might make decisions based on flawed assumptions. Proper documentation includes detailed notes on the methods used for cleaning, parameters set, and reasons for changes. This makes the process transparent, allowing future users to understand, replicate, or improve the cleaning methods. Documentation also facilitates easier audits and collaboration, ensuring that adjustments to the data are clearly justified and traceable.

Failing to update data cleaning models can compromise data relevance. Take e-commerce, where product categorizations evolve. Old models might misclassify new products or miss trends, leading to poor stock decisions and marketing strategies. Regularly monitoring data shifts and testing models on recent data ensures they adapt to changes like new product types or customer behavior patterns. Retraining models with updated data helps maintain accuracy, ensuring the cleaning process remains effective and relevant, aligning with current data realities and business needs. This proactive approach prevents the gradual decline in data quality that could affect analytical outcomes and business insights.

Relying on manual methods for data cleaning can be especially problematic in fast-paced industries like social media, where data volumes and formats change rapidly. For instance, manually filtering inappropriate content from user posts can't keep up with the influx of data, leading to lapses in content moderation. Automating data cleaning with advanced frameworks ensures processes like anomaly detection and content filtering are consistent and swift, adapting to new trends without constant human oversight. This not only saves time but also reduces errors that can occur with manual handling, making the data management process scalable and more efficient.

It hasnt been mentioned here but business or survey errors of data are equally important as the most complex of the methodologies. For example transcription or communication errors. The team must always speak to each other to minimize them.