Mohamed Ashour

In my line of work, I found the following techniques useful for model optimization: 1- Regularization of the data sructures used. 2- If data structures are widely different, then create a model for each format. 3- Allow for feature(s)/parameter(s) selection without the model 4- Follow a trial an error process when dealing with outliers. 5- Split the datasets into a number of subsets to identify further where outliers sit. 6- Try a number of sub-models besides your main model of choice. 7- In case you are doing multi-output regression, use the end product of multiple regression models after measuring their accuracy.

There are other non technical factors that impact the predictive analysis process as follows: 1- Predictive modelling can be quite useful in construction industry. 2- The client/end-user ,most probably, would not fully understand the predictive analysis process. 3- Some clients require the implementation of a very sophisticated predictive analysis model without investing in the technological infrastructure. 4- In the construction industry, there could be a predictive analysis guru who is awaited to provide results without having any colleagues understanding the process followed. 5- Lots of organizations think wrongly that predictive analysis is the key to all of their problems without having the right data infrastructure in place.

In my line of work, I follow the following steps to test the predictive model. 1- Split the dataset into 3 parts: training, testing & validation. 2- Run predictions by the model on both the testing & validation datasets and align the results together. 3- Agree with the management on the required accuracy levels (i.e. 65% or 70% ...etc). 4- Adjust the training dataset (by removing outliers) 5- Try 2-3 different prediction models with the adjusted training dataset and then run them by both the testing and validation datasets. 6- Repeat steps 4 &5 until the required accuracy is reached. 7- Note down all the adjustment steps for proper documentation of the process.

In my experience, the design of a predictive model is an iterative process. The steps that I ususally follow are: 1- Determine the case at hand to know the potential data sources and build a rough idea about what model would be adopted. 2- Gather information from the relevant data sources 3- Undertake data profiling and cleansing measures. 4- Remove obvious outliers skewing the end result. 5- Choose 2-3 predictive analysis models relevant to the case at hand. 6- Train the model on the data and predict results. 7- Compare predictions to actual results to determine the model(s) accuracy. 8- Refine the model(s) by removing/adding different data points or choosing other models. 9- Repeat the steps above until maintaining the required accuracy.

Predictive models are algorithms that try to find patterns in a given dataset and then predict or classify an outcome by following the same pattern in another dataset. In my experience, predictive models have a wide variety of use cases including inflation prediction (time series regression), predicting carbon emissions of certain projects based on their quantities, cost prediction of a specific type of projects, sentiment analysis for a customer sect...etc