update model with xval

Jepse

Hi everyone 

i'm currently try to predict sentiment within text. Therefor i use k-nn witch takes a classified dataset (with label) to create a model. Then i apply the model on unclassified data.When i validate the model with xval i archive a accuracy depends between 60 and 70. Not so bad for now!

But how could i train (update) the model out of the xval mistakes?

SebastianLoh

Hi Jepse,

Jepse wrote:
But how could i train (update) the model out of the xval mistakes?

What do you mean by that? Could you clreafy this?

Ciao Sebastian

Jepse

Let's say i usually train a model by providing a classified dataset (with label) to a learner (knn). As a result i get a model witch i can use for predicting unclassified dataset.

When i use the xval in combination with performance i get as a result the accuracy of my model. In other words, the probability of how the model works on my classified data, a self-test, right?

To calculate the performance the xval- and performance-operator probably compares the labeled with the predicted data to get the accuracy-probability. Wouldn't it be the next step to learn (update the model) out of the wrong predicted data?

steffen

Jumping in ...

what you describe is called "Boosting" (see wikipedia for an detailed explanation).

In Rapidminer you may find interesting
- example process under Samples -> processes -> 10_Boosting
- the operators under Modeling -> Classification and Regression -> Meta Modeling

greetings,

steffen

Jepse

Hey Steffen,

that's exactly what i was looking for. Unfortunately it takes some time to create and apply the model. But the accuracy went up from 70% to 76%.

What would be the right approach to keep training the model. Do we have to classify new training set and update the model? Or do we have to correct random wrong predicted sentences and passing them to the update model operator?

Cheers Philippe

charly

Hello ,, Whats the new with update model ??