Find Patterns and Predict

Pattern Finding Agent

The pattern finding agent can be used to quickly search for interesting patterns, distributions and anomalies in your data.

To use it, open the Pattern Finding menu in the agents drawer and add any feature nodes of interest. When you click What's in My Data?, conode will run a suite of tests and algorithms on the data graph to search for patterns and anomalies. The results will be displayed in a new view, showcasing the most interesting results discovered automatically.

This view contains a hierarchy of discovered insights. At the top level, you will see headings including:

• Inferred Distributions: Every numerical feature in your data is tested against a set of distributions to find the best fit. Where a good fit is discovered, that feature is annotated with the name of the distribution. The uncovered distributions are included in this view as successors of this node.
• Anomalies: This section looks for anomalies in several ways, including searching through individual features for numerical outliers, and looking at combinations of features to find "Combined Anomalies". The successors of this node walk you through the discovered anomalies.
• Correlations: This section looks at strong correlations between features. It separates these into perfect correlations - which are likely to represent duplicated information, highlighting that it may be worth considering merging these features - and strong correlations, which could represent interesting relationships in the data.

Nodes referencing features will be connected via edges to the relevant features, so you can easily navigate to the feature node(s) to explore further. For example, try plotting pairs of correlated features or single features with inferred distributions.

Nodes referencing data will be connected via edges to the relevant data. This allows you to quickly select the data and cross-highlight it between views.

Note, nodes referencing both features and data be connected to both! When navigating to successors, ensure that you select only the relevant features that you are interested in before plotting.

Fitting A Prediction Model

Conode makes it easy to run fast predictions on your data.

The "Fit" agent will fit either a linear regression or logistic regression depending on the target data type. It assumes all given data is training data. Upon completion, a set of views explaining the model performance and score will be opened.

Both linear and logistic regression will open 3 new views. A summary view, a weights view and an evaluation view.

• The summary view just shows the model input and out. It is worth highlighting that this view contains a header node with the evaluation of the model on the training data.
• The weights view shows the weight associated with each input feature vector towards the result.
• The evaluation view is used to evaluate the model performance. It is a plot of true vs predicted target feature vector.

If predicting categorical data, the model used is a multinomial logistic regression. Therefore, there is a weight per input feature and output category. For example, predicting the shirt size (S, M, L, XL) given the height of a person, the model will produce a weight for each target category S, M, L, XL.