This master's thesis addresses the development of an explainable recommendation system. The starting point is the demographic aging of the population and the need for technological solutions that support healthy and active aging while being tailored to the specific needs of older users. Existing recommender systems often operate as "black boxes," which in sensitive domains, such as elderly care, poses an obstacle due to a lack of transparency and trust. Instead of classic black boxes, we formulate the recommendation problem as a rating prediction task on tabular data. We investigate how various sets of contextual, demographic, and behavioral variables influence the prediction accuracy and the readability of the model’s explanations.
To solve the problem, we used data from the national survey of the Anton Trstenjak Institute, based on a sample of 1,047 individuals over 50 years of age, which includes validated questionnaires on health, activities, relationships, and attitudes. Based on these constructs, we build a user–activity rating matrix and six feature configurations: the baseline model (only userID, itemID), a model with current activity context, a model with context classes, a demographic model, a model with behavioral and attitudinal factors, and a model with user groups. For all configurations, we train an XGBoost regression model with a user-based split into training and test sets, measuring performance with MAE and RMSE. For prediction explanation, we use the SHAP framework, which allows for global and local explanations of individual feature contributions. Based on SHAP values, we also introduce the measure of explanation entropy to quantitatively assess how concentrated or dispersed the explanation is.
Experimental results showed that the inclusion of contextual information significantly improves the model’s prediction accuracy compared to the baseline model. The best result was achieved by the model including the current activity context, which reduced the Mean Absolute Error (MAE) from 1.1393 to 0.6031, confirming that the preferences of older users are highly situation-dependent. The analysis of explanation entropy revealed that the inclusion of behavioral and attitudinal factors leads to the most focused and thus the most humanly understandable explanations. Demographic features and broad user groups contribute to the stability of the explanations, although their impact on accuracy was smaller.
We conclude that by combining Gradient Boosting Trees (XGBoost), SHAP explanations, and explanation entropy, it is possible to build an effective recommendation system that simultaneously achieves good predictive accuracy and generates brief, understandable explanations. Such a system is suitable for supporting healthy aging, as it allows professionals, caregivers, and older users themselves to understand why a certain activity was suggested and to more easily assess the appropriateness of the recommendation.
|
