The other thing that inexperienced Kagglers do is to ask if method/model X can work in a given competition. My answer to this is always, "Try it and see if it works or not." People often miss that machine learning is an experimental science. In order to build good models, one must follow the scientific method:

• Make a hypothesis (e.g., adding this feature, or adding this NN layer, will improve pipeline performance)
• Run an experiment to test the hypothesis (train the modified pipeline)
• Analyze experiment results (is CV score better than before? Where is it better? Where is it worse?)

Each experiment should be done so that it can confirm or reject a hypothesis. For this, an experiment should change only one thing at a time. Often, inexperienced people change many things, then cannot conclude what worked or not.