Search Engine

Based on work done freshman summer. Used GCP to train a model that implements product search with Tensorflow API. Used data from online stores as training and testing data. Used embeddings, feature extraction and sentence encoding as input to the Tensorflow model.

Below is an extremely high level view of the model.

Github

https://github.com/rah4927/search

How it Works

Embeddings

• Used Word2Vec and Doc2Vec to get vector embeddings of products and product metadata
• Used Universal Sentence Encoding to get embeddings of product raw data and user queries
• Used VGG 16 to extract features of product images

Model

Here’s a very simplified picture of the mode. Created a neural network N(q, w) that takes a query q and a product w and outputs a score based on the likelihood of them being correlated.

Ranking

For a given query q, sort N(q, w) over all products w. This is our initial search engine. ::: :::::