Expectation Maximization-Log Likehood interpretation I am using EM algorithm in weka for genomic data, get the results in the images, but a don't know how interpret the log likehood index. Is better when is higher or lower, negative or positive. How interpreted?
 A: EM is a maximum-likelihood estimation (MLE) algorithm, a higher 
likelihood means that the data are more likely to be generated by the the parameters estimated.
Higher likelihood values are not always better because of overfitting, which is usually a problem of MLE methods.
A: EM is an optimization procedure, optimization procedures often run into problems so the Log likelihood gives you an idea of how good your estimate is relative to other estimates - if you're optimizing the same model with the same data multiple times. 
A: In Expectation-Maximization, you want to find the Maximum-Likelihood Estimator of the parameter $\theta$, that is, the one that maximizes the probability (the log-probability, actually, which is a similar problem) of your observed data $X$:
\begin{align}
argmax_{\theta} \ln p(X | \theta)
\end{align}
The problem is, you don't always know how to maximize this function. What you do in EM is augmenting the model with some latent data $Z$. If you knew the value of this data, then you could maximize the function.
So, you do
\begin{align}
\ln p(X | \theta) = \ln \int_{Z}  p(X, Z | \theta)
\end{align}
and then try to maximize this second function. Yet, we are still blocked here. So we do a little trick:
\begin{align}
\ln p(X | \theta) = \ln \int_{Z}  p(X, Z | \theta) = 
\ln \int_{Z}  \frac{p(X, Z | \theta)}{q(Z)} q(Z) = \ln \mathbb{E}_Z[\frac{p(X, Z | \theta)}{q(Z)}]
\end{align}
where $q(Z)$ is any probability distribution. Now, Jensen's inequality tells us that the log of an expectation is greater or equal to the expectation of the log. Then we have
\begin{align}
\ln p(X | \theta) =  \ln \mathbb{E}_Z[\frac{p(X, Z | \theta)}{q(Z)}] \geq  \mathbb{E}_Z[\ln \frac{p(X, Z | \theta)}{q(Z)}] = \mathbb{E}_Z [\ln p(X, Z | \theta)] - \mathbb{E}_Z(\ln q(Z)]
\end{align}
where we just created a lower bound of our target function $\ln p(X | \theta)$. The equality holds (the lower bound touches the target function) when $q(Z)$ is actually the posterior of $Z$ given your observed data and the parameter $\theta$ (because this makes the ratio inside the log equal to $p(X | \theta)$, which does not depend on $Z$).
So, the EM trick is to maximize $\theta$ in the lower bound (M-step). But then we break the equality and are again below the target:  we have to re-compute the expectation (E-step) of $Z$ since the current expectation was given the old $\theta$!). 
Thus, the every time we re-compute the expectation according to the last $\theta$, the value of our lower bound is actually the value of $\ln p(X | \theta)$. This is the log-likelihood value that you observe in the screen. The higher the log-likelihood, the closer you are to the solution of the problem.
And when the algorithm converges, you have a local (global, if you get lucky or if the function is concave) maximum for $\ln p(X | \theta)$.
