Understanding Network Evidence
A Bayesian network stores conditional probabilities for connected variables. Each row describes one local choice, such as a node state under a parent state. When observed data arrives, the model receives evidence counts. Log likelihood measures how strongly the chosen probabilities explain those observations.
Why Logs Are Used
Raw likelihood multiplies many probabilities. The product can become extremely tiny. Computers may round it to zero. Logs solve this problem. They turn multiplication into addition. This calculator sums count times log probability for every entered row.
What The Result Means
A larger log likelihood is better when the same data and network structure are compared. Values are often negative, because probabilities are below one. A value closer to zero usually means the model fits the observed evidence more closely. Do not compare unrelated datasets without care. Different sample sizes change the scale.
Advanced Inputs
The observed count may be an integer or a decimal weight. Decimal weights are useful for expected counts from expectation maximization. The probability column accepts values from zero to one. When zero appears, smoothing can prevent undefined logs. The smoothing field adds a small amount before the log is taken. This helps test rare events while still showing warnings.
Using The Calculator
Enter each node, parent configuration, observed state, probability, and count. Choose the log base you need. Natural logs are common for statistical learning. Base two gives bits. Base ten gives decimal log units. Press calculate to view the full summary above the form.
Practical Checks
Probabilities for sibling states should normally sum to one under the same parent setting. This tool does not force that rule, because users may enter only observed rows. Review the example table before entering your own data. Export the result as CSV for spreadsheets. Use the PDF button for reports, notes, or classroom work.
Common Mistakes
Avoid entering percentages like twenty five unless you convert them to 0.25. Keep counts non negative. Separate parent settings clearly, so later review is easy. If a row has zero count, it adds nothing to the score. If a row has high count and low probability, it strongly lowers the total score. That pattern often signals a weak model assumption.