In rapidly growing world of indi’s digital payment ecosystem. UPI(*Unified Payment Interface *— It is an instant payment system developed in india) has experienced unprecedented growth. In 2017, UPI has seen 100Mn transactions with value worth of 67 Bn INR. UPI has witnessed exponential surge in use. Fast forwarding to 2023, the total volume of UPI transactions made are 83.75 Billion with value worth in trillions. With thousands of transactions taking place everyday, it becomes important to ensure that each transaction that is taking place is reliable and consistent. Ensuring confidence of users in system.

Source: National Informatics Centre(NIC)

ACID properties play a crucial role in upholding the consistency & realibity of these transactions in this context. These ACID properties are Atomicity, Consistency, Isolation, & Durability. Lets delve into each property along with an example.

1. Atomicity: Single Indivisible Unit

When first time using an online payment system have you ever pondered the scenario where funds are debuted from your acount but it’s not credited in recipient due of some error? To prevent such situation, atomicity dictates that every transaction must be treated as single indivisible unit of work.

which simply means that, the transer of funds from one account to another will be treated as one transaction. However if you can see, this one transaction can be further broken down to more indivisible single process. The first transaction involves debiting funds from your account, and the second transaction would be crediting funds to recepient]s account. This breakdown highlights the very nature of atomicity, ensuring the execution of each individual step within the entire transaction process.

So the system that follows the atomicity, should either commit entire transaction or no transaction will occur at all. In our case either both crediting & debiting process will takes place or none of them will occur, ensuring no partial or incomplete operation.

2. Consistency: Preserving the Integrity of the System

Consistency preserves the databases integrity by enforcing rules and constraints, thereby preventing any inconsistencies.

In UPI, an individual can’t transfer funds more than 1 lakh Rupees in a day(it changes from bank to bank, but max limit is 1 lakh). this is a constraint placed inorder to protect user from possible threat of having big financial losses in case of compromised account or fraudulent activities.

if in case anyone attempts to transfer more than 1 lakh Rupees, inorder to protect the consistency of the database, the system must enforce the above mentioned rule. Consquently any such attempts of transfering funds more than 1lakhs Rupess must be terminated. This will result in maintaining the consistent state in the database, protecting against any deviations from predefined criteria.

3. Isolation: Preventing Transaction Interference

Isolation ensures that concurrent transactions don’t interfere with each other and should remain independent, preventing potential issues like conflicting reads and updates.

Daily thousands of transactions are made and it becomes important to isolate of each of the transaction. Lets understand it through a real world example. Consider your in the shop or the mall and there are numerous people who are engaged in shopping or paying for groceries. The shop has multiple QR code linking to one account. Now, imagine you have paid to shopkeeper after your shopping by scanning same QR code used by other. However your payment is not reflecting in the shopkeepers account.

This may happen because other customers simultaneously making payment to same account. The system is accessing and modifying simultaneously the same data i.e, shopkeepers account. This may lead to override of your transaction details before they are commited on shopkeepers account. This situation is occured because someone else is making a payment exactly at the same time as you do. In UPI the changes from one transaction won’t be visible to the other until they are committed, thereby avoiding anyconflicts and maintaining data consistency

4. Durability: Ensuring Permanent Changes

Durability assures that once a transaction is committed, its changes are permanent and survive any system failures.

Think that you received your monthly salary of 20k via UPI, that very day you decided host party and invited all of your collegues for party. You and your friends are enjoying the party and living your moments. Everything is going well until you attempted to make payment at the end of party, only to discover that transaction has failed because of insufficient balance. So you you check your balance and it is showing ZERO.

This is an example of Non-Durable system. In such system even after commiting transaction like you receiving your salary, they do not become permanent or these transactions mayn’t survive the system failure. In above example the record of your salary wasn’t retained for some reason, resulting in disappearnace of your funds.

In case of Durable system, once the transaction is committed successfully, it will retain it permanently. durability ensures that the changes are stored persistently and will not be lost for any reason, even if system is crashed.

The ACID properties form the bedrock of database systems, offering the necessary assurance of reliability, consistency, and data integrity, especially in managing multiple concurrent transactions. Understanding and implementing these principles is essential for maintaining a resilient and trustworthy database environment.

Thank you for taking the time to read the article. If you enjoyed the article, hit the “clap” button and stay connected with me by following me on Linkedin, GitHub, and Medium. Visit: thejagdish.com.

When we develop a ML model, it becomes important to measure the performance of the model. After measuring its performance and rebuilding the model to advance it, repeat the process until it becomes better and better.

Here, **metrics **help us track the performance of our model. so that we can know if the change made to our model improves it or not.

Metrics for Regression

• Mean Absolute Error(MEA)

• Mean Squared Error(MSE)

• Root Mean Squared Error(RMSE)

• Root Mean Squared Log Error(RMSLE)

• R-Squared or R²

• Adjusted R²

Mean Absolute Error(MAE):

It is also called an “L1 error.” It calculates the average of the absolute difference between actual and predicted values. It only measures the magnitude of the error, not its direction.

The MAE value increases linearly with an increase in errors, and the input and output values have the same unit*.*

The MAE value is between 0 and infinity. where 0 indicates a good-performing model and higher values indicate a poor-performing model.

Following is the formula for mean absolute error:

Here, true value is the actual value of the data point, and prediction is the predicted value of that data point. In MAE, all individual errors have the same weight.

In the above graph, red dots are individual data points, the straight blue line is predicted values, and the green vertical line is the magnitude of the difference between the actual data point and the predicted data point.

When should MAE be used?

• It is used because it is intuitive and simple to understand for the stakeholder’s.

• MAE penalises the larger error.

Drawback of MAE:

• As you can see from Fig. 2, the graph of MAE is not differentiable at 0.

Mean Squared Error(MSE):

MSE is one of the most common loss functions used for regression problems. It is the sum of the squared difference between actual and predicted values.

It is also known as L2 loss. Unlike MAE, the MSE value increases exponentially with an increase in error.

Here, all individual errors have different weights. Squaring the error yields a higher weight for the outlier (the larger error) and a lower weight for the smaller error.

Following graph shows the parabola. Beacuse MSE value is quadratic, therefore it is also called as Quadratic loss. Since it is a parabola, it has only one global minima.

Benefits of MSE:

• As it has only one global minima and doesn’t have any local minima, we can use it in gradient descent.

• It is helpful for dealing with an outlier, as it penalises a larger error.

Drawbacks of MSE:

• If our model has a single large error, the error value increases. because of squaring.

• MSE is scale-dependent.

Root Mean Squared Error (RMSE):

It is nothing more than the source of MSE.It shows the standard deviation of prediction errors (i.e., residuals). It tells you how the data is scattered around the best-fit line.

It also penalises larger errors with a high weight, but comparatively less than MSE. Input and output data are stored in the same unit.

Benefits of RMSE:

• Because the input and output data are in the same unit.It is very intuitive to understand.

• It is differentiable, so we can use an optimization algorithm.

Drawbacks of RMSE:

• Like MSE, RMSE is also scale-dependent.

• It is not robust to outliers.

Root Mean Squared Log Error (RMSLE):

Here we take the log of actual and predicted values in order to penalise errors in predictions of large values more than smaller values. It is used in problems where the target variable has a* skewed distribution* or when the model’s prediction is used on an exponential scale. It is not affected by the scale of the values being predicted.

It does not penalise high errors due to the log. Underestimation* *is penalised more than overestimation.

Benefits of RMSLE:

• It is also differentiable, which is good for an optimization algorithm.

• It is scale-invariant, which means it is not affected by the scale of the values being predicted.

Drawbacks of the RMSLE:

• There is a biassed penalty that gives more weight to underestimation compared to overestimation.

R-square(R²):

It is also called the “coefficient of determination” or “goodness of fit.” The residual sum of squares (SSres) is compared to the total sum of squares (SStot). It defines to what extent the variance of independent variables explains the variance of the dependent variable.

It either increases or doesn’t change, but it never decreases when new data is added. which means we are not penalised for an unimportant feature. every time we add data to it. it increases, even if the variable is insignificant.

It is assumed that every independent variable explains variations in the dependent variable.

The R2 value ranges from 0–1. where 1 indicates that the independent variables explain 100% of the variation in the target variable and 0 indicates that the independent variables don’t explain any of the variation in the target variable.

Drawback of R2:

• As I said earlier, when a new variable is added to R2, it either increases or doesn’t change without detecting the significance of the newly added variable. which means an insignificant variable will also be added to the R2 value.

Adjusted R²(adj-R²):

It measures the proportion of variations explained by only those independent variables that really help explain the dependent variable.

Unlike R2, the adjusted R2 punishes you for adding such independent variables that don’t help in predicting the dependent variable. Adjusted R² is ≤ R².

Where,

• N = no of observations

• p = no of independent variable.

Note: R-squared has many flaws to consider before using it. I never use it to compare models or to determine the goodness of fit. So I suggest that you reconsider using it, and if you want to use it, then read these articles: “8 Tips for Interpreting R-Squared,” written by Tim Bock, and “Is R-Squared Useless?” written by Clay Ford.

Thank you for taking the time to read the article. To get more updates following me on Linkedin, GitHub, and Medium. Visit: https://thejagdish.com