Designing for Digital Wellbeing

1. Introduction

These massive datasets may be employed with machine learning algorithms that have been developed by researchers in statistics and AI. These algorithms are widely used to predict outcomes, such as classifying a person into a certain group based on the information or predicting an event that will happen in the future. In these circumstances, strategies to prevent overfitting, or the finding of fake ←13 | 14→patterns in a dataset that do not transfer to other datasets, are typically integrated with machine learning algorithms (Hamaker, & Wichers, 2017).

Researchers have employed machine learning algorithms to solve a variety of prediction-focused psychological research concerns using the kinds of datasets mentioned above. Social media data have been used to identify significant predictors for a number of mental health disorders, including depression, posttraumatic stress disorder, suicidal thoughts, and schizophrenia (Schwartz et al., 2014; Ismail et al., 2020; De Choudhury et al., 2016; Mitchell et al., 2015). Analysis of usage of smartphones and wearable sensor data has shown predictors of current and future cognitive states, including emotional states, with a focus on mood disturbance (Mehrotra et al., 2017; Rachuri et al., 2010; Canzian & Musolesi, 2015; Mehrotra et al., 2016; Saeb et al., 2015).

Using big, readily available clinical datasets, clinical psychology, psychiatry, and neuroscience have all predicted how to identify, diagnose, and treat mental illness (Dwyer et al., 2018).

Studies that concentrate on prediction, like the ones we’ve just discussed, can assist guide efforts at intervention, prevention, and therapy in real-world situations. By highlighting key variables and interactions that may then be looked into as potential causative elements underlying the phenomena being examined, a prediction-focused approach can also contribute to psychological theory.

To obtain correct predictions, however, psychologists frequently have to carefully pick relevant variables to include in machine learning models; this process is known as feature engineering in machine learning (Zheng & Casari, 2018). Without a priori understanding of the phenomena being studied, which is commonly absent in complicated datasets with many variables, it is difficult to develop accurate forecasting models. This is because feature engineering is required. In fact, it’s possible that the challenge of feature engineering had a significant role in the fact that no one strategy consistently produced high predicted accuracy in the research mentioned above.

2. What is Feature Engineering

It could be important to create and train better features for enhancing machine learning’s effectiveness in psychological forecasting. Any quantifiable input that may be employed in a predictive model is referred to as a feature. Feature engineering (Turner, Fuggetta, Lavazza, & Wolf, 1999) represents the process of employing statistical or machine learning techniques to transform unprocessed observations into desired features.

←14 | 15→The craft of creating meaningful features from existing data in accordance with the aim to be learnt and the used machine learning model is known as feature engineering (Kuhn, & Johnson, 2019). It entails changing data into formats that are more closely related to the underlying goal that must be taught. When feature engineering is done correctly, it may increase the value of current data and boost the effectiveness of the proposed machine learning models. On the other side, if researchers use poor features, they might need to create models that are far more complicated to obtain the same level of performance (Dong, & Liu, 2018). A solid understanding of the business issue and the data sources at hand is the foundation for effective feature engineering. Thus, researchers gain a deeper comprehension of the data and gain more insightful knowledge by developing new features. Feature engineering is one of the most useful data science approaches when done properly, but it is also one of the most difficult (Duboue, 2020).

Feature engineering types include:

• • Facilitating learning and enhancing the comprehension of the results, scaling and normalizing include changing the range and centering of the data.
• • Filling up missing values entails substituting null values using machine learning algorithms, heuristics, or expert knowledge. Due to the difficulties of gathering complete datasets as well as mistakes made during the data collection process, real-world datasets can contain missing values.
• • Removing elements that are unnecessary, redundant, or plain ineffective for learning is referred to as feature selection. There are occasions when researchers have too many features and should have less.
• • In feature coding, many categories are represented by a collection of symbolic values. Concepts can be recorded in many columns, each of which represents a single value and has a true or false in each field, or they can be recorded in a single column that has numerous values. A new feature or features are produced through feature building from one or more previous features. For instance, you might incorporate a function that shows the day of the week using the date. With this additional knowledge, the algorithm can learn that some outcomes are more likely to occur on Mondays or weekends.
• • Moving from low-level characteristics that are inappropriate for learning, getting bad testing results, to higher-level features that are advantageous for learning is known as feature extraction. When converting specialized data formats, such as photos or text, to a tabular row-column, example-feature structure, feature extraction is frequently useful.

←15 | 16→Scaling just changes the range of the data. Normalization is a more significant alteration. The process of normalization involves converting your observations into data that can be compared to a normal distribution (Webb‐Robertson, et al., 2011).

Imputation relates to the treatment of missing value data. While deleting items without particular values is one way to solve this issue, doing so might result in the loss of some crucial information. Imputation is advantageous in this circumstance (Musil, et al., 2002). It may be loosely classified into two categories. Specifically, missing numerical values are commonly imputed using the mean of the equivalent value in other entries, and missing categorical values are typically substituted with the value that appears the most frequently in other entries.

Missing values are one of the most prevalent issues when it comes to preparing your data for machine learning (Fernando, et al., 2021). Missing values may be the result of a variety of factors, such as human error, interruptions in the data flow, privacy concerns, and others. Missing values have an impact on how effectively machine learning models operate. Imputation’s major objective is to deal with these missing values (García, Luengo, & Herrera, 2015). Two categories of imputation exist: numerical imputation and categorical imputation.