Technology

Deploying complex infrastructure with heterogeneous data raises significant technical problems. From storage to real-time processing and batch analysis, all the steps of a data pipeline are made tedious when dealing with sparse and weakly structured data sets. The ideal solution to address this problem is condensing all the available data into compact, meaningful representations: embeddings.

An embedding is a low-dimensional vector of continuous numbers learned to store the relevant information in a compact and convenient form. Embeddings are designed to reduce the dimensionality of both categorical and continuous features.

The expressive power and convenience of embeddings is witnessed by their rising role in the state-of-the-art of the machine learning literature.

Typical Standard Machine Learning Methods

Most data-driven companies are using a combination of multiple DBMSs to store the data about their consumers: SQL, document-based, key-value stores, etc. None of these traditional systems provide any tool to efficiently process the information using high level, intuitive concepts.

Our Methods

We leverage our interdisciplinary expertise, backed by our academic experience, to generate and deploy state-of-the-art machine learning embeddings that serves a wide variety of business purposes.

Deep Collaborative filtering allows to understand and present complex patterns of human behavior. Whether it is through grouping data or referencing patterns, it is imperative to gather as much user information as possible from all available sources to acquire usable results. Simple models fail to learn simultaneously global trends and rare individual preferences from the long tail. Deep Collaborative Filtering scales to settings surpassing traditional boundaries, such as cross-domain or unbalanced data sets.

Typical Standard Machine Learning Methods

Most other solutions use “Matrix Factorization” which reduces an individual's preferences to a simplistic linear model.

Our Methods

Deep Collaborative Filtering instead focuses on pinpointing complex patterns that our algorithms then use to determine user tastes.

Recommending new items is a common challenge for businesses. More generally, any item without enough history of interactions from users won't be recommended by traditional methods. A novel solution to this cold-start problem is to extract deep information from text or image data using advanced natural language processing and creating a neural network of associations. Deep Content Extraction allows our software to recommend items no one has interacted with yet because the algorithm is able to understand the genre of the content by automatically extracting information from items such as a movie poster, synopsis, or reviews.

Typical Standard Machine Learning Methods

Other solutions commonly patch their recommender system with hand-crafted rules to diminish the cold-start problem. However this requires expensive iterations and does not scale.

Our Methods

Using Deep Content Extraction, our solution can generate accurate recommendations of items as soon as they are available in the data set.

Semantic Graphing makes the goal of finding correlations between seemingly unrelated data simpler by using variables such as metadata, labels, tags, genre, actors and more. By doing this we can make sense of semantic data in the same way a human mind would. In a nutshell, if we know user likes a certain item, and this item is connected to a second item, for instance if two movies share certain key actors, then our algorithm can consider the second item as a promising candidate to recommend to our user.

Typical Standard Machine Learning Methods

Most other solutions do not include any graph databases for their recommendations.

Our Methods

Our algorithm leverages all the available information to increase the accuracy of the recommendations, enabling users to discover hidden gems.