Les data scientists doivent travailler avec d’énormes quantités de données. Comment pouvez-vous donner un sens à tout cela ?

When working with patient data, data scientists must establish clear problem definitions that align with clinically relevant causal relationships. 🧪🩺 Consider this simple scenario: Simply having lungs isn't the cause of bacterial pneumonia (example 1); however, bacteria in the lungs can lead to pneumonia (example 2). In patient's with cystic fibrosis, inherited changes in the lung cells may be part of the causal pathway of bacterial pneumonia (example 3). Example 1 should not be conducted. Example 2 is a common causal pathway. Example 3 is a more complicated causal pathway that may require more advanced methods and data. It is crucial to define clinically relevant causal pathways to ensure our research is valid and impactful. 📊🏥

1)-Understand your data. This is key. What data are you looking at, do you have domain knowledge to comprehend what you are staring at. Then based on the problem dive into the domain knowledge driven data 2)-Do heuristics, and basic analysis of the data. This can only give you a sense of what is coming at you. 3)-Clean the data with caution. Based on statistics, the outliers you just removed, what consequence does it bring to the overall learning system that is being built. 4)-Analyze the data after cleansing, and see if there are any major differences. Normalize the data also whenever possible, algorithms love that. 5)-Visualize the output. Remember the consumer of this will be the business, so the output needs to be communicable.

Data can be overwhelming at times especially if you don’t know what you are looking for. Defining a clear objective prior to analysing your data will guide you towards finding the correct tools to utilise. Objectives need to provide useful information, if you are analysing data just for the sake of analysing, your answers will likely not have any merit.

In the realm of data analysis, grappling with an ill-defined problem is akin to searching for a black cat in a dark room, only to realize the cat isn't even there. The significance of well-defined problems cannot be overstated—they essentially pave the way for half of the analysis journey. This concept mirrors the crucial phase in engineering design where neglecting the precision of problem definition can lead to substantial costs in terms of money, time, and human resources. Once that foundational step is overlooked, it becomes an unrecoverable overhead in subsequent stages of data analysis, and there's no going back to square one.

If the data scientists are not already familiar with the domain they must first acquire knowledge of the domain, and then they must participate in the data acquisition process. Often it is assumed that the businesses already have plethora of data that contains what is needed to achieve the goal set by the business leaders. If that were true, then part of data preprocessing must already have been done during acquisition, and the features of interest were already known, but we all experienced data scientists know that it is often not the case. Therefore even before the data exploration begins, the domain knowledge must be acquired. Then depending on the goals set by the business, the data acquisition strategy must be defined.

In my experience, A/B testing plays a huge role in data exploration. It gives you a mathematical result that shows whether the division is statistically meaningful or not. Python has libraries that perform A/B testing; go check them out if you haven't already. That will be worth your time.

When dealing with a new dataset no matter the size, I always find it useful to observe its basic statistical properties. What does the distribution look like, ie. does it have a single mode or multiple ones? Is the distribution relatively broad or narrow? Moreover, are there a lot of apparent outliers in the data? For time series data, a key question would be does the data exhibit seasonalities? Also, for time series, does the data mean-revert? Only by first understanding these basic statistical properties would one then be able to consider the best way to transform (eg. using rate of change instead of absolute levels) and/or analyse the data.

Usually, people proceed to the visualisation step as it looks more pleasing but one should spend time on going through chunks of dataset . Most data in companies is of bronze quality which has to be thoroughly analyzed and dealt with to make it suitable for business use cases. One can opt for statistical tests such as hypothesis testing and other clustering techniques at initial level on chunks of data by breaking down the original data. This will allow individual analysis of each chunk making it more optimal to identify critical information which could be easily overlooked while handling data as a whole.

Facing the massive amount of available data, the tasks of gathering, exploring, cleaning, and processing usually take up more than 80% of the budget estimated based on traditional methods, leading to substantial delays and running the risk of analyzing it superficially - which would be a colossal mistake detrimental to the reputation of the team. In reality, these first crucial steps shall account for more than half of the total time spent on the project, keeping in mind that these are only the preliminary steps of the entire process.

Huge amount of data can be overwhelming sometimes. It takes lots of planning and most importantly, hell lot of patience😮💨. This tasks became even harder when working with the images and/or videos 📷🎥🎞️. Because cleaning, analysing and validating the images takes more time and patience. One have to check the image in detail and make sure the image is not spoiled and can be used for training the model. For textual or numerical data, I always prefer to choose a sample of data to begin with & check it out💡. > What are the columns? > How is the data distributed? > Data types of the columns > Basic stats of the data According to the sample data, plot the needed graphs using whole data.

Cleaning data is foundational in data science. It involves handling inconsistencies, errors, and missing values to ensure accurate analysis. Steps include removing duplicates, addressing missing data, standardizing formats, and handling outliers. This process sets the stage for effective analysis, enhancing the reliability and validity of insights derived from the dataset.

It's important to thoroughly examine your data after developing your problem statement. Gaining understanding of its structure, integrity, content, and general pattern is necessary for this. To clearly show and analyze your data, make use of descriptive statistics, visual aids, and exploratory data analysis techniques. You can identify trends, abnormalities, gaps, and possible inaccuracies that may affect the accuracy of your study by using this procedure.

One of the crucial steps in any Data Science pipeline is to prepare the data diligently, which includes assessing data quality, filling in missing data, cleaning the data, removing outliers, and transforming/scaling the data based on the available attributes. We employ different cleaning mechanisms for structured and unstructured data. Handling and processing unstructured data is more cumbersome than dealing with structured data.

Dealing with missing values in a customer dataset required thoughtful imputation strategies to ensure the integrity of the analysis. Cleaning data is often time-consuming but crucial for accurate results. Addressing outliers, handling missing values, and ensuring consistency are essential steps. Creating reusable data-cleaning scripts helps automate the process and maintain consistency across different datasets.

Visualizing raw data is a powerful technique in the data analysis process, particularly when it comes to data cleaning. It guides the data cleaning process by highlighting areas that need to be addressed, helps in deciding on data transformation strategies and continuous visualization throughout the cleaning process ensures that data integrity is maintained and the dataset's original meaning is not lost.

It typically involves examining the information you have and looking for patterns, relationships, or trends within the data. This process often includes using statistical methods, visualization tools, and, in the context of machine learning, applying algorithms to extract meaningful insights or make predictions. Effective data analysis can help in making informed decisions and understanding the underlying patterns in the information you have.

The modelling phase builds various models based on major Statistics & ML techniques using the prepared data, including: 1. Statistics & ML techniques and algorithms: Sklearn, StatsModels 2. Generate test design: split the data into training, test, and validation sets. 3. Build model: An object-oriented approach to building scalable and reusable codes maintains isolation between different parts of the code, making it easier to understand and maintain. 4. Assess model: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE). Multiple models are competing against each other, and the data scientist needs to interpret the model results based on domain knowledge, the pre-defined success criteria, and the test design.

Applying machine learning algorithms to predict customer churn led to actionable insights for implementing retention strategies. Choosing the right analysis methods depends on the nature of the problem. Regression, clustering, and classification are common techniques. Regularly validating and tuning models ensures their relevance and effectiveness in real-world scenarios.

There are many different ways to analyse data and there is not the one solution. You will need to continuously train yourself to stay up to date. However, the most crucial one is to have a repetorie of different methods which can be adapted to different situations. You will need to tools to comb through sentiments in social media, you will need to understand statistics to go through big data sets, you will need to understand when it is senseful to aggregate and disaggregate data and when to visualize it. So best for you would be to get signed up for different classes, google methods and stay up-to-date.

Data analysis plays a pivotal role across diverse domains, providing valuable insights and solutions to varied challenges. Examples include customer segmentation for targeted marketing, predictive maintenance using historical equipment data, sentiment analysis of text data for product or brand perception, financial fraud detection through transaction analysis, healthcare analytics to improve patient outcomes, supply chain optimization for enhanced efficiency, A/B testing to compare product versions, and energy consumption forecasting for efficient resource planning. These applications demonstrate the versatility and impact of data analysis in informing strategic decisions and addressing specific needs.

A data analyst must be a powerful person both in technical and soft skills. He must be a communicated person and could present his knowledge with other teammates, collaborate with them, lead them and etc. Such specifications enable the project to be matured as soon as possible.

To better communicate through your dashboards, start learning data storytelling. Avoid using unnecessary queries in your dashboards. Don't show them the visuals you want. Just show them what they need. Don't use more than 8 visuals per canvas. Keep it minimal and purposeful.

Presenting findings in a clear and non-technical manner facilitated better understanding among stakeholders. Effective communication is as important as the analysis itself. Use visuals, storytelling, and layman's terms to convey complex results. Creating concise and visually appealing dashboards can be a powerful tools for conveying insights to non-technical audiences.

The final, crucial phase of data analysis involves communicating your findings to stakeholders, clients, or your audience. This step requires presenting the insights you've derived in a clear, concise, and compelling manner. Effective communication often involves the use of visualizations, reports, dashboards, or stories to make the data accessible and understandable. Tools and libraries like matplotlib, seaborn, plotly, or bokeh are invaluable for creating impactful data visualizations. The aim is to effectively convey your message, elucidate the methods used, and substantiate your conclusions, ensuring that your analysis has the intended impact and drives informed decision-making.

It's not just about the numbers but how you bring them to life through storytelling. Treating data as a narrative, we make complex information accessible and engaging for all audiences. Keeping things simple and avoiding jargon ensures our message resonates. Tailoring our approach to the audience's needs and encouraging interactive engagement is crucial in making our findings genuinely impactful. Our role as data scientists goes beyond analysis; it's about bridging the gap between raw data and actionable insights. As data scientists, our role is not just uncovering insights but sharing them in a way that inspires and drives change.

Embracing an iterative approach is pivotal in the data science workflow. It involves continuously revisiting and refining each step—data collection, preprocessing, analysis, and model building. Iteration allows for the incorporation of new insights, adjustments to models, and improvements in methodologies, resulting in more accurate and impactful outcomes.

In the realm of data science, iteration is the heartbeat of progress. Each cycle offers a chance to refine problems, unearth insights from new data, and enhance methodologies. Embracing this iterative nature empowers practitioners to adapt to evolving challenges, fostering a culture of perpetual improvement. As feedback and results unfold, the iterative process becomes a compass, guiding data scientists through the maze of complexity towards more robust solutions. It's not merely a journey from A to B but a continuous exploration, where learning, experimentation, and refinement weave the fabric of success in data science.

Business process change every day and so does underlying data, change and adaptability is key to success of any data science project.

Iterar es clave como proceso de entrega de valor rápido y recurrente, garantizar que en el proceso se encuentren con herramientas o pasos de feedback constantes optimiza tiempos y reduce costos

In my experience, I combined machine learning and other statistical analyses in order to mitigate their technical limits by making them complementary. If your method isn't helping you with meaningful outputs, just try to fill those gaps with other methods that can provide additional features. Always try very different parameters when setting up a method, there are ad-hoc tools that can perform the parameters exploration for you.

My basic answer to this question is follow the steps: Data scientists manage large datasets by employing a variety of techniques: 1. Data Cleaning: Removing inconsistencies and errors to ensure data quality. 2. Data Transformation:Converting data into a format suitable for analysis. 3. Exploratory Data Analysis (EDA):Using statistics and visualizations to understand data trends and patterns. 4. Feature Engineering:Creating new variables from existing data to enhance models. 5. Machine Learning Models:pplying algorithms to predict or classify data. 6. Data Visualization:Presenting data in graphical form for easier interpretation. 7. Big Data Technologies: Utilizing tools like Hadoop or Spark for handling very large datasets.

In addition to all, 1. Extracting / Filtering required data: Extracting important data out of all the data available is one step that I have often done. To give an example from aviation industry. Problem : Detect turns of an aircraft flight Data : A time series data from a aircraft flight recording. Solution : After cleaning and visualizing, feeding complete data to a model will just add more inbalanced data, but filtering segments from the data and then feeding it to model considerably increases its accuracy.

Trabalhar com dados costuma ter melhores resultados e mais rápidos quando se trabalha de forma colaborativa! A área técnica e a área de negócios conversando, entendendo as ideias do outro, contribuindo e evoluindo juntos. Por isso, considere um ambiente de desenvolvimento que possa ter regras diferentes em relação às trilhas comuns de desenvolvimento de software (Dev/Test/Hom/Prod).

If you are starting from a "data swamp" where most of the data is completely disorganised, it can be a daunting task. Even if you come up with a comprehensive framework to catalogue the data, it may still require a monumental effort to implement it. Moreso if the team's data literacy is uneven. My suggestion is to start with a few key categories of data that are the most valuable for your business, on a manageable scale, and forget about the rest for the time being. Once you manage to get a rudimentary system going, the buy-in will increase and you can scale up your efforts over time. In short, aim high and far but start small, don't aim for perfection. Instead, go for steady momentum, show positive results, rinse and repeat.

If you can find the person the result will impact and ask for his wisdom. He will help you to make sense of the problem and data. Also he can be helpful to find a solution.