Deep transfer learning with multimodal embedding to tackle cold-start and sparsity issues in recommendation system.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Recommender systems (RSs) are crucial for personalized suggestions and managing information overload.
• Traditional RSs and deep learning-based recommender systems (DLRSs) struggle with data sparsity and cold-start problems, limiting their effectiveness.

Purpose of the Study:

• To propose a novel multistage model to enhance personalized product recommendations.
• To address and overcome the persistent challenges of data sparsity and cold-start issues in recommender systems.

Main Methods:

• A two-phase approach utilizing deep learning for image feature extraction and multimodal data embedding for dense user/item vectors.
• Offline phase: learning hidden features from images and creating feature vectors to combat item and user cold starts.
• Online phase: employing similarity matrices from the offline phase to generate top-N relevant item recommendations.

Main Results:

• The proposed model achieved Mean Absolute Error (MAE) of 0.5882 and Root Mean Square Error (RMSE) of 0.4011 on a Brazilian E-commerce dataset.
• These results indicate superior accuracy compared to baseline recommender systems.
• The model successfully minimized common cold-start and data sparseness issues.

Conclusions:

• The proposed multistage recommender system effectively mitigates data sparsity and cold-start problems.
• The integration of multimodal data embedding and deep transfer learning offers improved accuracy and personalization in recommendations.
• This approach provides a robust solution for enhancing e-commerce recommendation engines.