Clustering Device Sparse Data Inner Product Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The Lloyd method for clustering large-scale sparse data faces inefficiencies in memory capacity and calculation time due to the requirement for inner product calculations between object feature vectors and mean feature vectors, which are not effectively addressed in existing technologies.
Innovation Solution
A clustering device that stores transposed files associating cluster identifiers with non-zero feature quantities of mean feature vectors and calculates distances using sparse representations of object feature vectors, reducing memory usage and calculation time by only multiplying non-zero feature quantities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the Lloyd method is applied to clustering of sparse data using conventional data structures, then the clustering algorithm can be implemented, but the memory capacity and calculation time for inner product calculation are not efficient
Solution Approach 1:
The patent segments the inner product calculation by identifying and processing only non-zero feature quantities in sparse data. Instead of calculating products for all features, the method divides the calculation into non-zero and zero components, computing only the necessary non-zero products to reduce calculation time while maintaining accuracy.
Solution Approach 2:
The patent extracts and stores non-zero feature quantities of mean feature vectors in a transposed file data structure. By separating non-zero elements from zero elements and storing them in an optimized structure, the system enables faster access and computation during inner product calculations, directly improving clustering efficiency.
2Productivity
If the Lloyd method is applied to clustering of sparse data using conventional data structures, then the clustering algorithm can be implemented, but the memory capacity required is not optimized
Solution Approach 1:
The patent extracts only the non-zero feature quantities from mean feature vectors and stores them in a transposed file structure. This extraction eliminates storage of redundant zero values, significantly reducing memory capacity requirements while preserving all necessary information for accurate inner product calculation during clustering.
Solution Approach 2:
The patent inverts the conventional data storage approach by creating a transposed file where non-zero features are organized by feature index rather than by sample index. This inversion allows efficient memory utilization by storing only necessary data in an optimized layout that matches the computation pattern of sparse matrix operations.
3Measurement precision
If conventional inner product calculation methods are used for sparse data, then the calculation can be performed, but the calculation time and memory usage are suboptimal
Solution Approach 1:
The patent segments the inner product calculation into non-zero and zero feature components. By identifying which features are non-zero and processing only those segments, the method maintains precise distance calculation accuracy while eliminating wasted computations on zero features, thereby reducing overall calculation time.
Solution Approach 2:
The patent applies local quality optimization by using different data structures and calculation strategies for non-zero and zero features. Non-zero features are stored and processed with full precision in the transposed file structure, while zero features are implicitly handled without storage or computation, optimizing both accuracy and performance locally for each feature type.
Data Source
AI summary
A clustering device includes a storage means configured to store a transposed file including array elements in a storage unit, the array elements storing a pair in which a cluster identifier and a first non-zero feature quantity of a mean feature vector are associated with each feature identifier, a calculation means configured to calculate a distance between the object feature vector in sparse representation indicating the object feature vector in a pair in which a second non-zero feature quantity of the object feature vector and a feature identifier of the second non-zero feature quantity are associated with each other, and the mean feature vector of each cluster, and an assignment means configured to assign the object feature vector to one of the plurality of clusters on the basis of the distance, and the calculation means calculates a product of a second non-zero feature quantity and a first non-zero feature quantity stored in an array element of an index based on a feature identifier corresponding to the second non-zero feature quantity among array elements constituting the transposed file, to calculate an inner product necessary for the calculation of the distance.