Abstract: Methods, systems, and computer products are herein provided for lazy evaluation of input data by a machine learning (ML) framework. An ML pipeline receives input data and compiles a chain of operators into a chain of dataviews configured for lazy evaluation of the input data. Each dataview in the chain represents a computation over data as a non-materialized view of the data. The ML pipeline receives a request for column data and selects a chain of delegates comprising one or more delegates for one or more dataviews in the chain to fulfill the request. The ML pipeline processes the input data with the selected chain of delegates. The ML pipeline performs delegate chaining on a dataview. A feature value for a feature column of the dataview is determined based on the delegate chaining and provided to an ML algorithm to predict column data.

Abstract: Embodiments provide a machine learning framework that enables developers to author and deploy machine learning pipelines into their applications regardless of the programming language in which the applications are structured. The framework may provide a programming language-specific API that enables the application to call a plurality of operators provided by the framework. The framework provides any number of APIs, each for a different programming language. The pipeline generated via the application is represented as an execution graph comprising node(s), where each node represents a particular operator. When a pipeline is submitted for execution, calls to the operators are detected, and nodes corresponding to the operators are generated for the execution graph.

Abstract: Embodiments provide a machine learning framework that enables developers to author and deploy machine learning pipelines into their applications regardless of the programming language in which the applications are structured. The framework may provide a programming language-specific API that enables the application to call a plurality of operators provided by the framework. The framework provides any number of APIs, each for a different programming language. The pipeline generated via the application is represented as an execution graph comprising node(s), where each node represents a particular operator. When a pipeline is submitted for execution, calls to the operators are detected, and nodes corresponding to the operators are generated for the execution graph.

Abstract: An efficient, streaming-based, lazily-evaluated machine learning (ML) framework is provided. An ML pipeline of operators produce and consume a chain of dataviews representing a computation over data. Non-materialized (e.g., virtual) views of data in dataviews permit efficient, lazy evaluation of data on demand regardless of size (e.g., in excess of main memory). Data may be materialized by DataView cursors (e.g., movable windows over rows of an input dataset or DataView). Computation and data movement may be limited to rows for active columns without processing or materializing unnecessary data. A chain of dataviews may comprise a chain of delegates that reference a chain of functions. Assembled pipelines of schematized compositions of operators may be validated and optimized with efficient execution plans. A compiled chain of functions may be optimized and executed in a single call. Dataview based ML pipelines may be developed, trained, evaluated and integrated into applications.

Abstract: Sparse data handling and/or buffer sharing are implemented. Data may be buffered in reusable buffer arrays. Data may comprise fixed or variable length vectors, which may be represented as sparse or dense vectors in a values array and indices array. Data may be materialized from a dataview comprising a non-materialized view of data in a machine-learning pipeline by cursoring over rows of the dataview and calling delegate functions to compute data for rows in an active column. A buffer and/or its set of arrays storing a first vector may be reused for a second and additional vectors, for example, when the length of buffer arrays is equal to or greater than the length of the second and additional vectors, which may be selectively stored as sparse or dense vectors to fit the array set. Shared buffers may be passed as references between delegate functions for reuse.