Massachusetts Institute of Technology

Nike, Inc.

Harvard College

Universidade Do Porto

Celect, Inc.

Devavrat D Shah

Ying-zong Huang

Vishal Doshi

Vasudha Shivamoggi

Jay Kumar Sundararajan

Muriel Medard

Hari Balakrishnan

Ritesh Madan

Balaji Rengarajan

Jonathan Perry

Anish Agarwal

Dennis Shen

Michael D Mitzenmacher

Joao Barros

Szymon Jakubczak

Vivek Francis Farias

Srikanth Jagabathula

Ammar Tawfiq Ammar

Vinayak Ramesh

Muhammad Amjad

Tauhid Rashed Zaman

George Chen

Vighnesh Sachidananda

Parastoo Sadeghi

Sai Kiran Burle

Joshua Lee

Quan Li

A computer-implemented method includes: identifying, from first and second data, interventions common to a target unit and one or more of a plurality of donor units as filtered donor units, the first data corresponding to the target unit under one or more interventions, the second data corresponding to the plurality of donor units each under one or more interventions; identifying, from the first data, third data corresponding to the target unit under the common interventions; identifying, from the second data, fourth data corresponding to the filtered donor units under the common interventions; identifying, from the second data, fifth data corresponding to the filtered donor units under a subject intervention; generating, from the third and fourth data, a learned model representing to a relationship between the target unit and the filtered donor units; applying the learned model to the fifth data to generate the synthetic data; and outputting the synthetic data.

Systems and methods for synthetic interventions

Computer-implemented prediction methods, in accordance with various embodiments, involve storing key-value pairs each including a key tuple and an associated value in a data store, building a prediction model associated with at least a subset of the key-value pairs, and in response to receipt of a prediction query that specifies a key tuple, generating a prediction response based on the prediction model that includes a predicted value for the specified key tuple. In one example application, the key tuples each include an operation specifying a purchase and a two-part identifier specifying a customer and a product.

Event prediction

A system for event prediction is configured to ingest at least partially incomplete unstructured data from at least one data source, convert at least a subset of the ingested data into a prediction model in accordance with a received model definition, receive a prediction query, and in response, supply an event prediction based on the prediction model. In some embodiments, ingesting the data involves transforming the unstructured data into a universal structured format in which the data is represented as key-value pairs each comprising a key tuple and an associated value, the key tuple specifying an operation and at least one identifier; the prediction query includes a key tuple specifying an operation and at least one identifier; and the event prediction generated with the prediction model includes a value associated with the key tuple that is missing from the transformed data.

A system and method model a time series from missing data by imputing missing values, denoising measured but noisy values, and forecasting future values of a single time series. A time series of potentially noisy, partially-measured values of a physical process is represented as a non-overlapping matrix. For several classes of common model functions, it can be proved that the resulting matrix has a low rank or approximately low rank, allowing a matrix estimation technique, for example singular value thresholding, to be efficiently applied. Applying such a technique produces a mean matrix that estimates latent values, of the physical process at times or intervals corresponding to measurements, with less error than previously known methods. These latent values have been denoised (if noisy) and imputed (if missing). Linear regression of the estimated latent values permits forecasting with an error that decreases as more measurements are made.

Model agnostic time series analysis via matrix estimation

A system includes at least one connector configured to gather at least partially incomplete data from at least one data source. The gathered data is communicated to a model definition module that converts at least a subset of the gathered data into a prediction model in accordance with a received definition. A prediction module receives a prediction query and, in response, supplies an event prediction based on the prediction model.

Methods and apparatus for predicting unknown values given a data set of known values. A prediction engine ingests a variety of data sets. Once the dataset has been ingested, requests for predictions may be presented to the prediction engine. The engine responds with a prediction as well as a confidence score based on the ingested information using a variety of techniques. Some of these techniques identify witness values in the ingested data that are similar to the requested prediction, other techniques compute witness values from ingested data, etc. These witness values are aggregated to yield an answer and a confidence level in a way that permits the user to review the underlying witness values.

Prediction using fusion of heterogeneous unstructured data

A method is disclosed including: receiving a graph-based program that identifies a bipartite graph and one or more update function sets, the bipartite graph including a plurality of graph nodes and a plurality of edges, such that each graph node corresponds to one of the update function sets; associating each of a plurality of computing units with a different respective one of the graph nodes; instantiating, by a Publisher Subscriber platform, a plurality of channels, the plurality of channels defining a topology that matches a topology of the bipartite graph; and executing the graph-based program based on the plurality of channels to produce a result.

Method and apparatus for graph-based computing

Systems and methods for tracking at least one entity. Systems according to various embodiments may include an interface for at least receiving training data representing movements of a plurality of training entities and a query regarding a test entity's movement, a memory, and at least one processor executing instructions stored on the memory to create a plurality of tracks, generate an index of the plurality of tracks, execute the received query, and output a feature value with respect to the test entity.

Entity tracking

An engine and method for tracking the influence of an entity operating within a social network are presented. A query containing social network content is received. A database for entries referencing the social network content is searched, and interactions between an entity participating within the social network and the social network content are identified. A dynamic interaction network (DIN) of a plurality of the entities is identified and a dynamic influence score for an entity in the query specific DIN is calculated.

Dynamic influence tracking engine and method

Described herein are new error-correction (channel) codes: permute codes, iterative ensembles of permute and spinal codes, and graphical hash codes. In one aspect, a wireless system includes an encoder configured to encode data using one of the aforementioned channel codes. The wireless system also includes a decoder configured to decode the encoded data.

Permute codes, iterative ensembles, graphical hash codes, and puncturing optimization

A system and method for providing a personalized recommendation from a series of partial preferences is presented. A preference distribution of a population including a plurality of weighted ranked lists is identified. A revealed preference of a user is compared to the plurality of ranked lists. An affinity weight between the user and each of the plurality of ranked lists is assigned, and a weighted average of each of the affinity weights is taken.

System and method for providing personalized recommendations

A system and method for determining a rank aggregation from a series of partial preferences is presented. A distribution is learned over preferences from partial preferences with sparse support. A computer receives a plurality of partial preferences selected from two or more preference lists. Weights are assigned to each of said plurality of partial preferences, resulting in multiple ranked lists.

System and method for finding mood-dependent top selling/rated lists

In one aspect, a wireless system includes an encoder configured to encode data using a spinal code which uses a function over the message bits to directly produce a sequence of constellation symbols for transmission. The wireless system also includes a decoder configured to decode the spinal code. The function may be at least one of a non-linear function and a hash function.

Growing community of inventors

Newton, MA, United States of America

Devavrat D Shah