專題演講 主講人：花文妤博士（Amazon machine learning scientist）
題 目：Similarity Recommendation based on the Attention Mechanism
主講人：花文妤博士（Amazon machine learning scientist）
Item-to-item similarity has been long used for building recommender systems in industrial settings, owing to its interpretability and real-time computational productivity. In this work, we have developed a new embedding representation to the similarity-based recommendation system. The proposed solution enhances the information to both text embedding and image embedding. First of all, we have successfully improved the text embedding in two ways: 1) add item description and bullet points on top of the title along with some key attributes to enlarge the text information; 2) apply topic modeling on the description and bullet points to get key topics and keywords, and compare the performance between Word2Vec model and pre-trained fine-tuned Bidirectional Encoder Representations from Transformers (BERT) model on the text attributes. Moreover, we have tested product image embeddings with different settings and compare the performance with two settings: 1) apply max-pooling on a ResNet50 with triplet loss model to get 205-dimension embeddings; 2) apply PCA on the same ResNet50 model to reduce the dimension. Based on the experiment results with different text and image embeddings, we propose a better solution which outperforms the baseline result  with increased 20% precision on a fixed recall (0.05). The contribution of this work includes 1) the most comprehensive ASIN catalog information to the text model is used; 2) the best combination of text and image embedding is found. The result shows smaller distance in terms of k-nearest neighbors (KNN) Euclidean measurement and significant precision increased on a down-stream click and purchase task; 3) this framework is not limited to a specific use case, and can be easily adapted to different product categories and marketplaces.