Publicação
Portuguese Television Audience Forecasting using Deep Learning Models and Specialized Neural Networks
| Resumo: | Television networks are facing growing challenges due to technological advances and the attractiveness of streaming services. In a world of rising data-driven decisions, where audience rise is directly linked to advertising revenue, understanding viewers’ behavior and pinpointing the optimal time slots for advertisements are pivotal to survive in an increasingly competitive market. Therefore, audience forecasting plays a crucial role. In Portugal, however, few studies addressed this topic, particularly with respect to advanced forecasting algorithms, as Recurrent Neural Networks and Transformers. Over the years, time series forecasting methods have evolved. Originally, started with statistical methods like ARIMA since it offered high interpretability, although it lacked flexibility. Machine Learning approaches increased accuracy by incorporating heterogeneous features. Whereas recent studies relied on neural networks to capture time dependencies. This work proposes to investigate and compare traditional machine learning regres- sion models, currently used by Markdata to forecast the TVI audience ratings, with state-of-the-art deep learning techniques, including RNN, LSTM, GRU, and Transformers implemented in TensorFlow. This dissertation evaluates both individual and hybrid mod- els, with results evidencing that the sequential models and Transformers can make more accurate forecasts than the established baseline, Lasso regression. Notably, transformer architectures demonstrated great business potential, highlighting the power of attention mechanisms and automated feature engineering. Finally, the results suggest that combining a two-Stacked GRU algorithm with both a simpler and a more complex transformer model delivers complementary insights into temporal dependencies and audience patterns. The hybrid strategy not only enables the improvement of the forecasting performance but also allows the use of a lightweight transformer model for production deployment. |
|---|---|
| Autores principais: | Pimentel, Mariana Afonso Pires |
| Assunto: | TV Audience Portugal TensorFlow Keras Time Series Forecasting Deep Learning |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Television networks are facing growing challenges due to technological advances and the attractiveness of streaming services. In a world of rising data-driven decisions, where audience rise is directly linked to advertising revenue, understanding viewers’ behavior and pinpointing the optimal time slots for advertisements are pivotal to survive in an increasingly competitive market. Therefore, audience forecasting plays a crucial role. In Portugal, however, few studies addressed this topic, particularly with respect to advanced forecasting algorithms, as Recurrent Neural Networks and Transformers. Over the years, time series forecasting methods have evolved. Originally, started with statistical methods like ARIMA since it offered high interpretability, although it lacked flexibility. Machine Learning approaches increased accuracy by incorporating heterogeneous features. Whereas recent studies relied on neural networks to capture time dependencies. This work proposes to investigate and compare traditional machine learning regres- sion models, currently used by Markdata to forecast the TVI audience ratings, with state-of-the-art deep learning techniques, including RNN, LSTM, GRU, and Transformers implemented in TensorFlow. This dissertation evaluates both individual and hybrid mod- els, with results evidencing that the sequential models and Transformers can make more accurate forecasts than the established baseline, Lasso regression. Notably, transformer architectures demonstrated great business potential, highlighting the power of attention mechanisms and automated feature engineering. Finally, the results suggest that combining a two-Stacked GRU algorithm with both a simpler and a more complex transformer model delivers complementary insights into temporal dependencies and audience patterns. The hybrid strategy not only enables the improvement of the forecasting performance but also allows the use of a lightweight transformer model for production deployment. |
|---|