Knowledge Hub

As the name says itself, Data Splitting is about dividing data into multiple datasets which can be for multiple departments, for multiple regions, for multiple time periods etc. but here we will discuss how it’s used for data science?

Data Splitting for data science activities is an important aspect of creating, processing, validating, and testing machine learning models.

In Data Science, mostly 2 approaches are adopted to split datasets.

• Approach#1: Dataset are split in 2 subsets i.e., 1) Training 2) Testing
• Approach#2: Dataset are split in 3 subsets i.e., 1) Training 2) Validating 3) Testing

Let’s understand with an example, consider a dataset contains 100,000 bank customers, now a created model’s aim is to tag each customer under three categories e.g., Silver, Gold or Premium based on their credit history, buying behavior, transaction values etc.

Now let’s see how both approaches on with above example.

This Feature is not that traditional feature 😊. Here, we will be discussing Feature which is an integral part of Data Science activity. I am sure non-data scientists or non-data citizens might not be aware of this term.

We know what Data Science is, it is about running algorithms and machine learning models (ML models) on structured, semi and unstructured data sets for analytics.

For any ML model there is a basic requirement of data, right? And we all know Data Scientists love to play with Raw Data. Now, Raw Data means irrelevant data as well, so Data Scientists need Raw BUT relevant Raw Data. In Big Data, we know data sizes are huge and to train models, one doesn’t need all Raw Data. So, what data scientists do, they themselves or ideally data engineers clean and filter the data, in other words do Data Wrangling, before feeding into ML models.

For example, there is a table with 1 million rows containing data around the globe with all kinds of customers, but ML models is supposed to predict customers buying behavior for a particular country and for a particular set of products. Means, we don’t need to feed 1 million rows to our ML model so now we will create another table (FEATURE) where we will only keep data for that country and for required set of products and now make that table as source for our ML models.

We touched base with Data Mining and Data Science, it will be ideal to discuss Data Dredging, Snooping, p-hacking, and Fishing. New terms right 😊?

Data Dredging, Snooping and Fishing all refer to the same behavior of data analysis BUT without proper hypothesis and relationship among datasets.

Data mining finds results based on the correlation of data in large data sets, but Data Dredging, Snooping, p-hacking, and Fishing find results based on chance methodology. In other words, Data Dredging, Snooping, p-hacking, and Fishing share the results which require more investigation. These sometimes bypass Data Mining techniques and come up with immature conclusions. On occasions, it shows more details about something than it contains.

There are two types of Fact Findings in any analysis that ultimately assists in DSS i.e., 1) known facts 2) unknown facts. Data mining techniques extract Known Facts as it’s done only on structured datasets and mostly, we know what's there in structured data like relationships between tables and relationships of data within the same table. But that's Not True, sometimes, even in structured data there are Unknown Facts and that’s where Data Dredging, Snooping and Fishing techniques come into the action.

Data Anomalies are about any kind of unusual behavior which can degrade the quality of data. Please note, data quality is not only about dirty data but also about not being able to insert, update and delete required data.

Type of Data Anomalies

• Insertion: If one is unable to enter quality data in database.
• Update: If one is unable to update relevant data or able to update irrelevant data.
• Delete: If one is unable to delete relevant data or able to delete irrelevant data.
• Redundancy: If there are duplicate records in a table.

Data Normalization and database constraints play a vibrant role to make sure data anomalies doesn’t occur. To start with, let’s say there is a table with employee ID, name, department, and salary details. First wrong approach we can noticed i.e., this table is in de-normalized form, by right there should be separate tables for department and salary data using normalization process. Normalization and Denormalization are covered, in detail, in data modelling topic.

By now, we all know what Semantic Layer is, IT defines data elements for business ease. Then there came the era for Business folks defines data elements themselves based on their own understanding and now the latest approach is going towards Machine Learning Models define data element based on open-ended requirements.

A major difference between first two approaches vs Augmented Analytics is, first two were derived by humans either by IT or Business folks so there is always a factor of human era and at the same time human can be biased base on knowledge and information they possess about the business domain. On the other hand, Augmented Analytics is based on automated Machine Learning which learns the data elements while new data sets keep coming in.

The first approach was static as requirements were always coming from business and implemented by IT folks, so any change had a round robin effect.

There are many ways to achieve data privacy like data generalization, data swapping, data masking, data encryption etc., all these techniques change original data which of course is not desired by most. All these are done to ensure data should not fall into wrong hands.

Data Synthetic is something which will really solve the problem statement like anonymize data before sharing with developers, data science teams for data models etc.

Data Synthetic is about generating new dataset using computer simulations. This dataset contains fake data with the right properties or attributes. For example, customer information is required in almost every application or system, but it can’t be shared to protect individual identity. Data Synthetic tools create a dummy dataset for customers with name, email, contact number, family details etc. but nothing related to original data. This way, developers can build the data pipelines and data scientists can build their data models without thinking of breaching any regulatory or compliance clauses.

Data Vectorization is something very common nowadays, especially since the inception of Big Data or Hadoop. It’s not like it was not in-use in the past, but I would say it was not famous.

All of us must have heard about MPP (Massive Parallel Processing) right? But do we know how it works at the back end? Data Vectorization is about enabling parallel processing to fetch data.

To understand MPP, let’s refer to four types of instructions to pull data in computer architecture i.e., SISD, SIMD, MISD and MIMD (all explained in separate topic). In this topic, we will be referring to SISD and SIMD only.

Traditionally, a computer or machine or a server works in SISD mode i.e., Single Instruction Stream and Single Data Stream means each instruction fetches required data one by one.

In Data Vectorization, we change our approach to MPP mode i.e., computer or machine or server starts working in SIMD mode i.e., Single Instructure and Multiple Data, means if one query is executed and data is residing in multiple data nodes, data from all nodes will be pulled in parallel making computation must faster as compared to SISD.

Data Drift concept is used in Data Science practice, and it is one of the most expected behaviors in every data modelling. One must have automated ways to keep tracking and identifying Data Drifts as after Data Drift, there are high chances the model will not produce or predict the expected accuracy which was planned. Wrong prediction means wrong decisions.

We all know what a Drift is. Drift is about shifting or moving from one point to another but not in a standard approach. This term is mostly referred in Data Science practice, to evaluate trained vs actual., performances of Machine Learning (ML) Models.

Data Drift is also known as covariate shift, virtual drift, or virtual concept drift. When we create and run a ML model, just like any traditional query, it inputs a dataset and outputs a dataset. In ` ML modelling, we select certain data elements to build its logic. Data Drift is about change of data within same select data element.

Concept Drift is used in Data Science practice, and it is mostly unexpected behavior in data. One must have automated ways to keep tracking and identifying Concept Drifts as after Concept Drift, the model will totally not produce or predict the output which was planned. Wrong prediction means wrong decisions.

We all know what a Drift is. Drift is about shifting or moving from one point to another but not in a standard approach. This term is referred in Data Science practice to evaluate tested vs actual performance of Machine Learning (ML) Models.

Concept Drift is also known as class drift, real concept drift, or posterior probability shift. As we discussed in Data Drift, the data elements do not change rather data changes. But in Concept Drift, data elements themselves change.

For example, referring to the same example from Data Drift, model is trained on products but at the same time buying behavior of those products is dependent on external factors like weather. Let’s say, those products are Cold Drinks and it’s obvious in the region where it’s been sold, the weather is expected to be hot and based on the ML prediction, a production of 1M Cold Drinks are planned on weekly basis. Now, what if, due to any reason the weather sees a sudden change and for a week it rains, and the temperature drops unexpectedly, THIS IS CALLED CONCEPT DRIFT, then the whole model’s behavior will be impacted. Due to pleasant whether people will only consume 25% of 1M Cold Drinks which means the ML model must quickly change its training data and predict the production behavior.

Data Skew Issue is one of the main hidden culprits for query performance 😊. It is normally referred to as Skewed data distribution which is used in Data Science domain.

Have you ever noticed, your query is correct, your joins are correct, your all indexes are enabled, you have high data quality, you have data standardization in place but still your query performance is not as expected? You noticed when you loaded new datasets, queries were flying but as time passed, the performance has gone down. You have added more CPU, more RAM, more nodes but NO performance is not like newly added data. It’s because of the Data Skew Issue.

Data Skew Issue is about when data is not distribution evenly. Uneven distribution of data at cluster level causes this issue. It heavily degrades the performance of the queries especially those with joins.

We all have faced a situation where query run exceptionally good till 99% and then hangs for 3-4 hours right lolz yeah, I can understand your pain 😊. Let me tell you the reason, it’s because of Data Skew Issue.