Decoding TV Drama Success: Machine Learning Insights from Japanese Shows

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Dec 7 2023

In a paper published in the journal PLOS One, researchers examined the correlation between Television (TV) drama and social trends using machine learning to predict drama ratings based on various metadata, including broadcast details, genre, cast, and even poster facial features.

Analyzing 800 Japanese TV dramas aired from 2003 to 2020, the study employed different machine-learning classifiers. Notably, incorporating facial features in poster imagery significantly improved the accuracy of predicting ratings. The findings offered valuable insights into the relationship between drama metadata and the dissemination of social information, showcasing how these factors influence audience engagement and viewership patterns.

Background

TV dramas, explored in prior studies, stand as versatile storytelling mediums bridging radio, cinema, and television to portray human relationships through dialogue and action. This blend of immersive radio elements and visual storytelling caters to the human appetite for engaging narratives. These dramas became a predominant indoor entertainment choice for stress relief, particularly evident during societal stress like the Coronavirus Disease (COVID-19) pandemic.

Their depiction of real-life scenarios fosters audience empathy, subtly reflecting collective societal values. Japanese dramas, focusing on familial and workplace dynamics, offer nuanced insights into daily social contexts and ideologies. The economic implications tied to drama ratings further highlight their role as commercial benchmarks, influencing advertising revenues and potentially prompting early show cancellations if ratings fall short of expectations.

Methodology Overview: TV Drama Analysis

Data collection involved gathering metadata and poster images from Japanese TV dramas broadcast during prime time from 2003 to 2020. The metadata underwent quantification and normalization to serve as classification features. Poster images were utilized both for facial information extraction, combined with metadata, and independently for rating prediction comparison. High and low ratings were classified using clustering methods as classification labels.

The TV drama metadata included broadcast year, season, day of the week, time slot, and TV station. These elements aimed to capture audience leisure habits, linking broadcast timing with viewers' intentions. Additionally, the dataset encompassed genre, cast, screenwriter information, and original work or sequel status, which is pivotal in gauging audience preferences and potential brand advantages.

Microsoft Azure Face extracted facial features from drama posters, generating nine key features. The dataset encompassed various genres, each providing distinct thematic insights, and preprocessing involved normalization techniques like label encoding and one-hot encoding for discrete features to enhance classifier readability.

Four classifiers - Naïve Bayes (NB), Artificial Neural Networks (ANN), Support Vector Machine (SVM), and Random Forest (RF) - were employed for numerical prediction models, each with distinct methodologies. NB operates on the feature independence assumption, making it robust to noise. ANN utilizes input, hidden, and output layers for classification. SVM derives a generalized hyperplane for classification, and RF combines decision trees.

The image prediction model utilized Convolutional Neural Networks (CNNs) such as AlexNet, Inception V3, Residual Network (ResNet) 152, and DenseNet 201 to analyze correlations between image compositions from drama posters and ratings. CNNs use convolutional layers for feature extraction, pooling layers for feature reduction, and fully connected layers for final classification results.

Drama Ratings: Insights and Analysis

Analyzing Drama Ratings: The study analyzed drama ratings using a clustering threshold of 12.66% to differentiate between high and low-rated TV dramas. Observations showcased a peak in highly rated dramas in 2005, while since 2015, the number has consistently remained below 10. This trend continued across seasons, with spring showing the highest numbers and summer the lowest. Specific time slots, such as TBS TV Sunday at 9 pm, consistently presented a higher ratio of highly rated dramas, and specific screenwriters and actors, like Yasushi Fukuda and Takuya Kimura, contributed significantly to this group.

Genre Impact and Predictive Accuracy: Among the highly rated dramas, medical dramas stood out with a remarkable 50.00% share, while original productions notably surpassed adaptations, comprising 129 instances compared to 102. The predictive models achieved varying accuracies, with single features averaging around 70.00%. Combining all features in the RF classifier produced the highest accuracy of 77.10%, highlighting the collective importance of these features in predicting ratings.

Impactful Features and Viewing Habits: Features like scheduled slots aligned with viewer habits significantly impacted ratings. Japanese audience adherence to particular time slots influenced the stability of drama ratings. Additionally, trends over 18 years indicated a declining number of highly rated dramas, suggesting a potential challenge in achieving higher ratings in future productions.

Influential Factors: Actors, Writers, and Genres: Leading actors and screenwriters like Takuya Kimura and Yasushi Fukuda significantly contributed to highly rated dramas. Medical dramas particularly resonated due to their relatability and the aging demographic's heightened interest in health-related narratives.

Image Predictions and Comparative Analysis: The study utilized CNNs trained on drama posters, achieving 71.70% accuracy. Compared with similar literature on movie poster analysis, the study demonstrated a similar challenge in achieving higher accuracy. Despite this, drama posters showcased a meaningful impact on enhancing predictive accuracy.

Conclusion

To summarize, this study leveraged drama metadata and posters to gauge audience resonance with characters and the synchronization of dramas with social trends. Predicting ratings based on numerical and facial features from 800 Japanese TV dramas revealed the significance of facial information, achieving 77.10% accuracy with RF and 71.70% with DenseNet201.

The analysis unearthed factors influencing ratings, shedding light on cultural connotations. Future directions might explore quantifying elements like voices and subtitles and budget considerations to enhance rating predictions and guide the creation of compelling dramas.