Cold start is one of the critical problems faced in the recommender systems framework, when it comes to recommending new items or dealing with new users. Traditionally, this has been solved by a content-based filtering approach based on various item features like—metadata attributes, textual content, or visual representation of the item. However, this method of finding the neighborhood of the item based on its context, disregards the preferences information exhibited by the user in the past. On the other hand, a collaborative filtering approach only considers the users’ past preferences/responses to provide a warm recommendation, thereby lacking the ability to tackle the cold start problems. We propose a novel technique which combines the best of both worlds—where we have the rich item level attributes of the items and make use of the past user- item response data to understand the user feedback-determined- similarity of the items. We employ a Siamese network, where the input will be the tuples of items’ content-based features and learns the similarity/dissimilarity between two items determined by the collaborative filtering based latent item features. We then compare this solution against the content features-based neighborhood approach and evaluate the performance on three different cases: by using the content-based item features generated from–movie metadata, Word2Vec embeddings from movie synopsis and the AlexNet features from movie trailers.