Rongzheng Bian,Yumeng Xue,Liang Zhou,Jian Zhang,Baoquan Chen,Daniel Weiskopf,Yunhai Wang
Abstract:We propose a visualization method to understand the effect of multidimensional projection on local subspaces, using implicit function differentiation. Here, we understand the local subspace as the multidimensional local neighborhood of data points. Existing methods focus on the projection of multidimensional data points, and the neighborhood information is ignored. Our method is able to analyze the shape and directional information of the local subspace to gain more insights into the global structure of the data through the perception of local structures. Local subspaces are fitted by multidimensional ellipses that are spanned by basis vectors. An accurate and efficient vector transformation method is proposed based on analytical differentiation of multidimensional projections formulated as implicit functions. The results are visualized as glyphs and analyzed using a full set of specifically-designed interactions supported in our efficient web-based visualization tool. The usefulness of our method is demonstrated using various multi- and high-dimensional benchmark datasets. Our implicit differentiation vector transformation is evaluated through numerical comparisons; the overall method is evaluated through exploration examples and use cases.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that in multi - dimensional data projection, traditional point - based visualization methods can only display the density distribution of data points, while ignoring the information of local linear structures. Specifically, after performing multi - dimensional data dimension reduction, traditional methods only encode the position information of data points, without showing the direction and shape information of the local linear subspaces around these points. This leads to an insufficient in - depth understanding of the global structure of the data, especially having limitations in identifying local patterns and trends in the data.
To overcome this problem, the author proposes a new visualization method to analyze and visualize the local subspaces after multi - dimensional data projection through implicit function differentiation techniques. This method can not only display the positions of data points, but also represent the characteristics of local linear subspaces through the shapes and directions of glyphs, thereby providing a more in - depth understanding of the global structure of the data. In this way, users can more easily identify local patterns and trends in the data, such as the separation between different categories of data points, the trends of data streams, etc.
### Main Contributions
1. **Visualization Model of Local Subspaces**: Proposed a visualization model that can infer the global trends of data through the perception of local subspaces.
2. **Vector Transformation Method Based on Implicit Function Differentiation**: Developed an efficient vector transformation method for accurately projecting the local subspace basis vectors in multi - dimensional space onto the low - dimensional visualization space.
3. **Glyph Representation**: Designed a glyph representation that can display the geometric shapes of the transformed local subspace basis vectors in the traditional scatter plot domain.
### Method Overview
1. **Extract Local Linear Subspaces**: Use the k - nearest neighbor (kNN) method and local principal component analysis (PCA) to determine the local linear subspaces around each data point.
2. **Consistent Subspace Projection**: Through implicit function differentiation techniques, convert the basis vectors of local subspaces to the low - dimensional visualization space, ensuring that this conversion is consistent with the dimension - reduction technique applied to data points.
3. **Visualize Subspace Information**: Use glyphs constructed by closed B - spline curves to visualize the projected local subspace information, and these glyphs can capture the characteristics of basis vector deformation.
### Mathematical Model
- **Projection of Multi - Dimensional Data Points**: Define a multi - dimensional projection function \(\pi: \mathbb{R}^D\rightarrow\mathbb{R}^d\), which maps a point \(P\in\mathbb{R}^D\) in the high - dimensional space to a point \(p\in\mathbb{R}^d\) in the low - dimensional space.
- **Local Subspaces and Their Vector Transformations**: A local subspace \(S\) is spanned by a set of basis vectors \(V = \{V_i\}_{i = 1}^L\). Through implicit function differentiation techniques, calculate the transformations \(v=\frac{\partial\pi(P)}{\partial P}V\) of these basis vectors in the low - dimensional space.
- **Implicit Vector Transformation**: Utilize the implicit function theorem to derive the formula for vector transformation \(\frac{\partial\pi(P)}{\partial P}=-\left(\frac{\partial^2f(P,\pi(P))}{\partial\pi(P)^2}\right)^{-1}\left(\frac{\partial^2f(P,\pi(P))}{\partial P\partial\pi(P)}\right)\), where \(f(P,\pi(P))\) is the optimization objective function.
### Visualization and Interaction
- **Glyph Generation**: Use glyphs constructed by closed B - spline curves to represent the projected local subspaces, which can effectively reduce visual clutter and support the visualization of multiple basis vectors.
- **User Interaction**: Provide a series of dedicated user interaction functions, such as changing the transparency and size of glyphs, and sorting according to the glyph area, to reduce overlap and improve the visualization effect.
Through these methods and techniques, the paper proposes a new multi - dimensional data visualization framework that can more comprehensively display the local and global structures of data.