Algorithm and implementation for fast computation of content recommendations

A system comprising:

• at least one memory device and at least one processor;
• a preprocessing engine stored in the at least memory and executable by the at least one processor to pre-process documents to determine a fingerprint for each document, the determining of the fingerprint for each document including:
• determining a document vector based on features of the respective document, the document vector including keyword and relevance pairs with regard to keywords appearing in the respective document, the relevance indicating a degree of relevance of a respective keyword within the respective document, the relevance determined by dividing a number of occurrences of the respective keyword within the respective document by a total number of keywords appearing in the respective document, and the document vector including a number N-dimensions equal to the number of keywords in the respective document;
• and
• applying a hash function on the document vector to obtain a hash value for each of N-dimensions, each hash value equal to a frequency of a keyword of a respective dimension;
• a user profile module stored in the at least memory and executable by the at least one processor to determine a dynamic user profile for a user, the dynamic user profile being based on one or more fingerprints of viewed documents by the user and a set of relevant keywords collected from documents in a browsing history of the user;
• a similarity calculation module stored in the at least memory and executable by the at least one processor to determine one or more relevant documents, using a processor, based on the dynamic user profile in view of the determined document fingerprints;
• wherein the determining the one or more relevant documents comprises performing a cosine product approximation between a user fingerprint associated with the dynamic user profile and fingerprints of all available documents;
• and
• a list determination module stored in the at least memory and executable by the at least one processor to provide a list of the one or more relevant documents to the user.

A non-transitory machine-readable storage medium in communication with at least one processor, the machine-readable storage medium storing instructions which, when executed by the at least one processor, performs a method, the method comprising:

• preprocessing documents to determine a fingerprint for each document, the determining of a fingerprint for each document including:
• determining a document vector based on features of the respective document, the document vector including keyword and relevance pairs with regard to keywords appearing in the respective document, the relevance indicating a degree of relevance of a respective keyword within the respective document, the relevance determined by dividing a number of occurrences of the respective keyword within the respective document by a total number of keywords appearing in the respective document, and the document vector including a number N-dimensions equal to the number of keywords in the relevant document;
• and
• applying a hash function on the document vector to obtain a hash value for each of N-dimensions, each hash value equal to a frequency of a keyword of a respective dimension;
• determining a dynamic user profile for a user, the dynamic user profile being based on one or more fingerprints of viewed documents by the user and a set of relevant keywords collected from documents in a browsing history of the user;
• determining one or more relevant documents, using a processor, based on the dynamic user profile in view of the determined document fingerprints;
• wherein the determining the one or more relevant documents comprises performing a cosine product approximation between a user fingerprint associated with the dynamic user profile and fingerprints of all available documents;
• and
• providing a list of the one or more relevant documents to the user.