I suggest that:

Modify the table like this :

1- Compute the average of attr1 rows (4 rows) in new field attr1_avg for each class.

2- Compute the average of attr2 rows in new field attr2_avg.

3- Compute the average of attr3 rows in new field attr3_avg. ... etc

Now construct new table with single row classification.

You can now apply any classification technique with traditional methods.

You can spread rows horizontally by adding new fields/attributes/columns to make each class represented by only one row and apply any traditional classification methods.

For a basic classification using Lp metric (e.g. Euclidean distance) you can compute a mean value for each class during the training phase and then compare a test sample with these mean values (cluster centroids). If you are using a neural network then while feeding each sample value for a specific class keep the target output 1 for that class during the training phase. After all the samples of all classes are fed in, this would constitute a single iteration. Repeat the iteration until convergence of the network i.e. cluster memberships do not change further. Now feed in the test samples using the final weights obtained during the training phase.

This is a multivariate classification task in Machine Learning. Thus, the following Machine Learning techniques/models are most suitable for multivariate classification tasks, viz:

1) Decision Tree

2) Random Forest

3) K-Nearest Neighbor

4) Multilayer Perceptron (MLP)

Note that Machine Learning techniques/models like Linear Regression, Logistics Regression, SVM, etc are most suitable for binary classification tasks; however, your problem is actually a multivariate classification task

I must miss something. In both cases you have a two-class problem with three features (that you call attributes). The only difference is: in the first case you have 4 examples (3 in class A and one in class B), while in the second case you have 8 examples in class A and 4 in class B. From an algorithmic point of view, it makes no difference at all. It is a perfectly standard multivariable classification problem, which can be solved by any standard method such as those listed in Bonaventure Molokwu's answer above, and many others such as Support Vector Machines (SVM), Naïve Bayes, logistic regression, to name a few. Now if your next question is "Which is the best method?", the answer is "None". No classifier can be uniformly superior to all others in all situations (this is known as the "no free lunch" theorem - 1997).

There is, however, a fundamental difference between the two cases. In the first case, you have four examples in the three dimensional feature space: the number of examples is smaller than twice the number of features, which, from Cover's theorem (1965), means that virtually any dichotomy of the examples into the classes will be linearly separable. Therefore there is no point in using a nonlinear classifier; any linear classifier (linear SVM, logistic regression, ...) will do the job. In the second case, the number of examples is larger than twice the number of features, so that a nonlinear classifier may be needed. Note however, that in both cases the number of examples is much too small to obtain any significant result, because you must save some of the examples as a test set for testing the classifier.

I strongly suggest that you stop worrying about computer tools, and first read and understand a good textbook on machine learning and classification. In order to find the right tool, you must understand the problem that you are trying to fix. I suggest Chris Bishop's book "Pattern recognition and machine learning" (2011), for instance.

@ Gérard Dreyfus [ in both cases the number of examples is much too small to obtain any significant result, because you must save some of the examples as a test set for testing the classifier. ] Obviously, this is not the whole set and doesn't represent the actual data. I was describing the problem with some false data. Maybe I couldn't explain it properly.

The question is regarding second case "Target label depends on whole 4 consecutive examples together". Let me give an example: say, a pen with green color, and 3 inch in size is from the company X. This can be designed with one row (two features and a target).

But imagine, two machines (A,B) are producing pens. I am considering only one feature(color) here. If first pen is red and next one is green, it's from machine A. It first pen is red and second one is also red, then it's from machine B. Now, I can't decide without looking at consecutive two pens. Both rows represent one target

A bird is flying or a bird is eating. (fly and eat are two targets). These classification depends one multiple features over several frames. This is the problem.

YES, I lack the knowledge of full machine learning since I am beginner and worked with a fixed set of problems. That's why asked a question in this platform to find a quick fix.

@Bonaventure Molokwu Thank you for the term multivariate classification task . I will get on with it.

Atanu Shome You mention: "But the problem occurs when classification label / target label depends on multiple rows such as the second table from the attached file. Consecutive N rows makes one category."

Why do you consider this a problem? You want many observations to have the same label.

I think the problem you're running into is thinking you can't partition your data into train/test because of its initial structure. If this is your problem, it's not really a problem if your classes (categories) are fairly balanced. Simply shuffle the rows and split a train/test or learn on a single data frame (keeping train/test in tact) with appropriate resampling methods (cross validation, ...)

Machine learning is extremely variable depending on domain problem, sometimes it's acceptable to train/test from the same dataset depending on what sampling parameters are used. Not sure what your case is, but try that out.