Efficient instance-based learning on data streams

Q: What future works have the authors mentioned in the paper "Efficient instance-based learning on data streams" ?

There are various directions for further research. As mentioned previously, however, it is not immediately clear how such techniques can be used in a streaming application with tight time and resource constraints.

Question

1. What are the contributions in "Efficient instance-based learning on data streams" ?

2. What future works have the authors mentioned in the paper "Efficient instance-based learning on data streams" ?

3. Why is the algorithm unable to handle discrete attributes with relatively small domains?

4. What fields of computer science have so-called data streams attracted considerable attention in recent years?

Accepted Answer

This paper considers the problem of classification on data streams and develops an instance-based learning algorithm for that purpose.. The experimental studies presented in the paper suggest that this algorithm has a number of desirable properties that are not, at least not as a whole, shared by currently existing alternatives.

Accepted Answer

There are various directions for further research.. As mentioned previously, however, it is not immediately clear how such techniques can be used in a streaming application with tight time and resource constraints.

Accepted Answer

Due to the need to maintain statistics for attribute combinations, the algorithm can only handle discrete attributes with relatively small domains.

Accepted Answer

In recent years, so-called data streams have attracted considerable attention in different fields of computer science, such as database systems, data mining, and distributed systems.

Accepted Answer

First of all, in order to guarantee that results are always up-to-date, it is necessary to analyze the incoming data in an online manner, tolerating not more than a constant time delay.

Accepted Answer

As an important advantage of synthetic data let us note that it allows for conducting experiments in a controlled way and, hence, to answer specific questions concerning the performance of a method and its behavior under particular conditions.

Accepted Answer

The main reason for the poor performance of the other algorithms is the fact that the corresponding parameters, and therefore the size of the case bases, are not optimized for the high drift rate.

Accepted Answer

As a possible explanation for the poor performance of these approaches, note that the distribution of the (input) data is quite skewed and far from uniform; considerable changes of the stock rates occur rather rarely, but in case they do, there is usually a high influence on the other stocks.

Accepted Answer

To this end, 1,000 extra test instances are generated at random according to a uniform distribution, and the classification accuracy for this test set is derived by using the current case base; this is done for every 10 time points.

Accepted Answer

real-world and benchmark streaming data is currently not available in a form that is suitable for conducting systematic experiments.

Accepted Answer

In this case, their strategy of removing examples at random pays off, since it ensures that the deletion process does not focus on a small subregion of the data space only.

Accepted Answer

In this connection, it seems reasonable to select or delete examples on the basis of several indicators like temporal relevance, spatial relevance, and consistency (even though these are of course not requirements per se).

Accepted Answer

By changing the characteristics of the streams the authors will then show that IBL-DS is more stable in the sense of still producing good results for the same parameter configuration, whereas the performance of the other algorithms strongly deteriorates if the parameters are set in a suboptimal way.

Accepted Answer

In fact, it is not useful to store all examples that support the current concept; this only makes the model less flexible with regard to the next concept shift but does not lead to a higher accuracy.

Efficient instance-based learning on data streams

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What are the contributions in "Efficient instance-based learning on data streams" ?

2. What future works have the authors mentioned in the paper "Efficient instance-based learning on data streams" ?

3. Why is the algorithm unable to handle discrete attributes with relatively small domains?

4. What fields of computer science have so-called data streams attracted considerable attention in recent years?

5. How do the authors ensure that the data is always up to date?

6. What is the advantage of synthetic data?

7. What is the main reason for the poor performance of the other algorithms?

8. What is the reason for the poor performance of these approaches?

9. How many extra test instances are generated at random?

10. Why is the data currently not available in a form that is suitable for conducting systematic experiments?

11. What is the strategy for removing examples?

12. What are the main criteria for selecting or deleting examples?

13. What is the difference between IBL-DS and the other algorithms?

14. What is the way to store all the examples that support the current concept?

Figures

Citations

A survey on data preprocessing for data stream mining

Learning under Concept Drift: an Overview

Learning under Concept Drift: an Overview

Handling drifts and shifts in on-line data streams with evolving fuzzy systems

Batch-incremental versus instance-incremental learning in dynamic and evolving data

References

Data Mining: Practical Machine Learning Tools and Techniques

Data mining

Instance-Based Learning Algorithms

Models and issues in data stream systems

Case-Based Reasoning

Related Papers (5)

Instance-Based Learning Algorithms

Learning with Drift Detection

Mining high-speed data streams

Mining time-changing data streams

Learning in the presence of concept drift and hidden contexts