Abstract:
Recommendation systems are widely used in many applications to enhance user experience by reducing the impact of the information overload caused by the rapid evolution of technology. However, such systems often face the cold start problem, where new users or items lack sufficient data to generate accurate recommendations.
This research proposes a hybrid framework that combines an online clustering model with a sequential neural network to address this issue. The clustering model groups users based on demographics and preferences captured through the initialization phase, while the neural network captures the behavioural patterns from sequential interactions. The final recommendations are dynamically generated by filtering the predictions from both models based on the cosine similarity between the user preference vector and item category vector. This allows the framework to generate dynamic recommendations, adapting to shifts in user preferences.
Experiments performed on the MovieLens 1M dataset demonstrate the framework’s effectiveness, achieving higher precision (0.74) and accuracy (0.82) compared to selected baselines, significantly improving recommendations for cold-start users.
