Interactive dimension reduction result understanding method based on comparison

Abstract

The invention provides an interactive dimension reduction result understanding method based on comparison, and belongs to the technical field of data processing. The method specifically comprises the following steps: step 1, constructing a target group and a reference group based on a region of interest and a selected region in a scatter diagram; 2, calculating a semantic vector connecting the center of the target group and the center of the reference group according to a semantic understanding result of the data points, projecting the data points onto the semantic vector, and constructing a projection space by utilizing the semantic vector; and step 3, when data interaction is carried out, a user operates the semantic vector through the auxiliary visual component of the semantic vector in the projection space, and visual analysis of the region of interest is carried out. According to the method, the semantic features of the target area can be accurately presented based on a comparison thought, and an analyst can be helped to quickly understand and explain the visual structure in the dimension reduction space.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular relates to a comparison-based interactive dimensionality reduction result understanding method. Background Art

[0002] The semantics in the scatter plot are distributed nonlinearly, that is, there are no two orthogonal, continuous coordinate vectors encoding interpretable attributes in the space like in the traditional scatter plot. Therefore, it is difficult to clarify the semantics of a certain area in the nonlinear semantic space, explain the reasons for the formation of data patterns, and explore the changes of semantics in space.

[0003] Regarding the technical problem of difficulty in understanding semantic space, the implementation methods in the existing technologies can be divided into three categories: comparison-based methods, layout optimization-based methods, and multi-scale scaling-based methods. However, all existing technologies only focus on revealing the semantics of the area selected by the user and do not support the exploration of the distribution and transition of semantics in space, which is an important means to quickly obtain a structural understanding of nonlinear semantic space. In other words, existing technologies cannot provide accurate semantic interpretation when processing nonlinear data structures. Summary of the invention

[0004] The purpose of the present invention is to provide a comparison-based interactive dimensionality reduction result understanding method to solve the technical problem that the prior art cannot provide accurate semantic interpretation when processing nonlinear data structures.

[0005] In order to solve the above technical problems, the present invention provides a comparison-based interactive dimensionality reduction result understanding method, which is characterized by specifically comprising:

[0006] The target group and the reference group are constructed based on the regions of interest and selected regions in the scatter plot;

[0007] Calculating a semantic vector connecting the centers of the target group and the reference group according to the semantic understanding results of the data points, projecting the data points onto the semantic vector, and constructing a projection space using the semantic vector;

[0008] When performing data interaction, the user operates the semantic vector through the auxiliary visual component of the semantic vector in the projection space to perform visual analysis of the region of interest.

[0009] A further technical solution is that the region of interest is determined according to the selection result of the user.

[0010] A further technical solution is that the semantic vector is a vector connecting the centers of the target group and the reference group.

[0011] A further technical solution is that the semantic vector is divided into a unipolar semantic vector and a bipolar semantic vector.

[0012] A further technical solution is to construct a projection space using the semantic vector, specifically including:

[0013] Constructing a projection space according to the distribution of data points corresponding to the semantic vector;

[0014] The projection space is provided with a filter brush with adjustable width and freely sliding along the vector in the horizontal direction, and the distribution of the semantic changes described by the semantic vector in the projection space is revealed by sliding the filter brush.

[0015] A further technical solution is that the auxiliary visual component includes a projection space, a dimensionality reduction space, a weight editor and a parallel attribute vector.

[0016] The present invention can accurately present the semantic features of the target area based on the comparison idea, and can help analysts quickly understand and interpret the visual structure in the dimensionality reduction space. The specific beneficial effects are as follows:

[0017] (1) The interaction of the present invention is intuitive and simple, which can improve the analysis efficiency: the traditional method of understanding the dimensionality reduction results usually lacks interactivity, and the analyst needs to rely on static visualization results for analysis, which is inefficient. The present invention introduces an interactive exploration technology - semantic vector, which enables analysts to explore data more flexibly and focus on areas of interest, thereby improving the efficiency of data analysis. Such an interactive method can help analysts discover potential laws and features in the data more quickly, thereby speeding up the analysis process. By introducing the idea of ​​comparison to construct semantic vectors, analysts can understand data in a semantically guided manner in the dimensionality reduction space. Compared with the traditional method of understanding dimensionality reduction results, this method can help analysts understand the semantic structure and features of the data more deeply, thereby improving the depth of data understanding. By comparing the differences between the target group and the reference group, the analyst can construct an interpretable semantic vector, further deepening the understanding of the data. The present invention not only includes basic interactive methods such as circle selection, but also has functions such as filter brushes that dynamically display semantic distribution, which conforms to the analyst's thinking habits and improves the analyst's experience and work efficiency.

[0018] (2) The computation of the present invention is simple and efficient: it is applicable to large-scale data sets and can respond instantly even to data of millions. Traditional data analysis methods face the problems of high computational complexity and slow response speed when processing large-scale data, but the present invention adopts efficient computational algorithms and optimization techniques to quickly process large-scale data sets. This simple and efficient computational feature ensures that analysts can obtain results in a shorter time, quickly conduct data exploration and analysis, and improve work efficiency and analysis efficiency.

[0019] (3) The present invention supports multi-attribute data analysis: The present invention is not only applicable to the understanding of dimensionality reduction results of document data, but can also be applied to the analysis of multi-attribute data. The dimensionality reduction result understanding and multi-attribute sorting are unified in a context-consistent interactive environment, realizing a visual analysis system that supports multi-attribute data analysis, allowing analysts to complete a variety of data analysis tasks in the same interface. By introducing visual components such as weight editors and parallel attribute vectors, analysts can flexibly adjust attribute weights and compare data performance on different attributes, supporting more complex multi-attribute data analysis tasks. Such an analysis method can help analysts understand the data more comprehensively and discover hidden information in the data.

[0020] (4) The present invention has strong applicability and flexibility: the present invention does not rely on a specific linear or nonlinear dimensionality reduction algorithm, but provides a general interactive exploration framework that can be combined with various dimensionality reduction algorithms. Analysts can select a suitable dimensionality reduction algorithm according to specific needs and embed it into the framework of the present invention for data analysis and visualization. Since the design concept of the present invention is universal, it can be seamlessly integrated with existing dimensionality reduction algorithms and data processing tools, whether it is the classic PCA (Principal Component Analysis), LLE (Locally Linear Embedding) or the popular t-SNE (t-Distributed Stochastic Neighbor Embedding) and other algorithms, they can be integrated and applied with the framework of the present invention. The present invention is not only applicable to dimensionality reduction algorithms, but can also be applied to other data analysis and visualization tasks, such as clustering, classification, association rule mining, etc., and has wide applicability and application prospects.

[0021] Other features and advantages will be described in the following description, and partly become apparent from the description, or understood by practicing the invention. The purpose and other advantages of the invention are realized and obtained by the structures particularly pointed out in the description and the drawings.

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other features and advantages of the present invention will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings.

[0024] Figure 1 It is a flowchart of a comparison-based interactive dimensionality reduction result understanding method;

[0025] Figure 2 A diagram of understanding data patterns in the semantic space of scatter plots;

[0026] Figure 3 A basic principle diagram of a semantic vector according to the present invention;

[0027] Figure 4 A visual design diagram of a semantic vector presented according to the present invention;

[0028] Figure 5 is one of the types of semantic vectors: single-level semantic vector;

[0029] Figure 6 One of the types of semantic vectors: two-level semantic vectors;

[0030] Figure 7 A multi-attribute data visual analysis system with semantic vectors as the core is designed according to the present invention. DETAILED DESCRIPTION

[0031] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that the present invention will be comprehensive and complete and fully convey the concepts of the example embodiments to those skilled in the art. The same reference numerals in the figures represent the same or similar structures, and thus their detailed description will be omitted.

[0032] The terms "a", "an", "the", and "said" are used to indicate the presence of one or more elements / components / etc.; the terms "comprising" and "having" are used to express an open-ended inclusive meaning and mean that additional elements / components / etc. may be present in addition to the listed elements / components / etc.

[0033] In today's data-driven world, visualization has become an important method and means of data science due to its intuitiveness and flexibility. Among the many visualization charts, scatter plots are one of the most versatile and multifunctional inventions. They are not only widely used in scientific research, but also play an important role in various fields such as business and society.

[0034] The widespread use of scatter plots can be attributed to three main reasons. First, scatter plots have a simple structure and strong visual expression. Both professionals and non-professionals can intuitively perceive data patterns such as clusters, outliers, trends, and correlations in large-scale data. Second, scatter plots have strong extension capabilities and can be combined with a variety of other charts and designs such as regression lines, contour lines, color mapping, and visual abstract schemes. This further enhances its visual expression and information transmission capabilities. Finally, scatter plots can break through the limitations of two-dimensional space. By using dimensionality reduction technology, scatter plots can be used to analyze and explore high-dimensional data, expanding its application areas.

[0035] In recent years, based on the above three capabilities, with the growth of data scale and the popularity of nonlinear dimensionality reduction algorithms, using nonlinear dimensionality reduction algorithms to construct scatter plots to analyze the hidden patterns in the data has become the most popular technical path for exploring large-scale high-dimensional data. The scatter plots constructed in this way describe a visible semantic space, which can quickly discover and understand data features by following the exploration mode from overview to details. In addition, compared with the scatter plots created by linear dimensionality reduction algorithms, the visual structure in the scatter plots constructed by nonlinear dimensionality reduction algorithms is often clearer and more significant, and can adapt to higher-dimensional data sets. However, the semantics in such scatter plots are nonlinearly distributed, that is, there are no two orthogonal, continuous, coordinate vectors encoding interpretable attributes in the space like in traditional scatter plots. Therefore, it is difficult to clarify the semantics of a certain area in a nonlinear semantic space, explain the reasons for the formation of data patterns, and explore the changes in semantics in space.

[0036] In order to solve the problem of difficulty in understanding semantic space, researchers have developed many methods and tools. According to different interaction strategies, existing methods can be divided into three categories: comparison-based methods, layout optimization-based methods, and multi-scale scaling-based methods.

[0037] The contrast-based method is to clarify the semantics of the local area by creating a reference group and a target group. This method is similar to the control experiment in experimental science, which reveals the characteristics of the region of interest by comparing the characteristics of the two groups of data. AxisSketcher is one of the typical methods based on explicit selection of the reference group. It uses theme-river style visualization to reflect the linear transition of the attributes of each segment. Users can arbitrarily circle two groups of points in the semantic space and reveal their semantic differences by comparing the density distribution curves of the two groups of points. The t-viSNE method allows users to understand the causes behind the pattern by drawing curves that run through the data pattern of interest. It emphasizes that this interaction is suitable for explaining the data distribution pattern of the line shape and searches for the neighborhood of each segment point based on the distance in 2D space. This type of method provides an intuitive interactive interface. Users can define the contrast relationship through simple operations, helping users quickly understand the differences between data. Although the interaction of this type of method is simple, it is relatively difficult to understand the results of the interaction. Users need to know the specific calculation process. In addition, due to the nonlinearity of the semantic space, constructing multi-segment semantic axes is often meaningless and only increases the difficulty of understanding.

[0038] Methods based on layout optimization focus on solving local or global distortion problems in data visualization by interactively adjusting the layout of the semantic space. iVisClustering is based on the LDA (Latent Dirichlet Allocation) topic model, and optimizes document clustering results by interactively adjusting the weights of topic description words, modifying document topic attribution, and merging, splitting, and deleting topics. This interactive method makes the final semantic space more consistent with the user's cognition. UTOPIAN is also an interactive system based on the topic model, but it uses a semi-supervised topic model of non-negative matrix decomposition to avoid the shortcomings of LDA in terms of convergence and result stability. These methods provide rich interactive means, allowing users to directly participate in the construction and adjustment of the semantic space, so as to better understand and grasp the semantic information of the data. However, this type of method is suitable for scenarios where there is a need to transform the original scatter layout, such as merging and splitting clusters, but is not suitable for more common and simple scenarios where only the visual features in the scatter plot, such as clusters and curves, need to be understood. Therefore, this type of method is more complicated to operate and requires some understanding of the explored data, which is more suitable for professional data analysts.

[0039] The multi-scale approach allows users to switch freely between high-level overview and detailed exploration as needed, supporting a multi-level understanding of the data. TopicLens allows users to dynamically explore the semantic space using a lens similar to a centrifugal marker. The documents covered by the lens are divided into several more fine-grained topics in real time, and the coordinates of the document are updated to adapt to the results of the topic division, maintaining the visual Figure 1 Consistency. TexTonic gradually discloses the semantic space through three levels to control the amount of information managed by the user. At the highest level, the system only displays the terms with the highest weight in the cluster, providing an initial overview of the entire dataset; at the middle level, the system displays more intermediate terms to help users focus on specific areas of interest; at the lowest level, the system displays additional terms and fragments of the source text to provide the most detailed context. Users transition between these three levels through zooming operations. These methods provide multi-level data presentation, and users can freely switch between overall overviews and local details as needed to meet the needs of information of different granularity. However, such methods must be combined with automatic label layout algorithms, and are usually computationally complex.

[0040] These three classification methods have different focuses. The contrast-based method focuses on revealing the local semantic differences of data through explicit or implicit contrast; the layout optimization-based method solves the problem of visualization distortion by adjusting the layout of the semantic space to improve the user's understanding depth; the multi-scale-based method supports appropriate data exploration by providing views of different granularities. However, all existing methods only focus on revealing the semantics of the area selected by the user and do not support the exploration of the distribution and transition of semantics in space, which is an important means to quickly obtain a structural understanding of nonlinear semantic space.

[0041] In order to overcome the problem that existing methods cannot provide accurate semantic interpretation when processing nonlinear data structures, the present invention proposes a comparison-based interactive dimensionality reduction result understanding method. Due to the challenges of poor interactivity, difficulty in understanding, and high computational complexity of previous methods, the present invention aims to provide a more intuitive and accurate semantic interpretation, while reducing the computational complexity and difficulty of parameter selection, and helping analysts better understand the semantics of dimensionality reduction results through comparative analysis.

[0042] Compared with previous methods, the present invention solves the problem of inaccurate interpretation of nonlinear data structures. By interactively allowing analysts to circle two groups of data points as the target group and the reference group, semantic vectors are automatically generated and displayed on the visual interface. Analysts can observe the position changes of data points in the reduced dimensionality space in real time by dragging filters. This comparative analysis method can help analysts represent and interpret the abstract concepts implicit in the local or global features of the reduced space, thereby more accurately understanding the semantics of the dimensionality reduction results and using them as ranking metrics for defining concepts.

[0043] In addition, the present invention also highlights the ability to intuitively understand data patterns. By introducing multiple semantic vectors and displaying the semantic content and changes of data in different directions, analysts can intuitively understand data patterns. By designing a simple, easy-to-use, lightweight interactive interface, analysts can easily explore data and perform multi-attribute sorting tasks. At the same time, the understanding of dimensionality reduction results and multi-attribute sorting tasks are integrated in a unified exploration environment, which can improve work efficiency. In addition, the understanding of dimensionality reduction results and multi-attribute sorting tasks are integrated in a unified exploration environment, which enables analysts to complete a variety of data analysis tasks in the same interface, improving work efficiency. The invention is applicable to any dimensionality reduction algorithm because it focuses on the semantic content and changes of the data, rather than the implementation details of a specific algorithm, which enhances applicability.

[0044] Example 1

[0045] In order to solve the above technical problems, Figure 1 As shown, the present invention provides a comparison-based interactive dimensionality reduction result understanding method, which is characterized by specifically comprising:

[0046] The target group and the reference group are constructed based on the regions of interest and selected regions in the scatter plot;

[0047] Furthermore, the region of interest is determined according to a selection result of the user.

[0048] Calculating a semantic vector connecting the centers of the target group and the reference group according to the semantic understanding results of the data points, projecting the data points onto the semantic vector, and constructing a projection space using the semantic vector;

[0049] Furthermore, the semantic vector is a vector connecting the centers of the target group and the reference group.

[0050] It should be noted that constructing the projection space using the semantic vector specifically includes:

[0051] Constructing a projection space according to the distribution of data points corresponding to the semantic vector;

[0052] The projection space is provided with a filter brush with adjustable width and freely sliding along the vector in the horizontal direction, and the distribution of the semantic changes described by the semantic vector in the projection space is revealed by sliding the filter brush.

[0053] Furthermore, the projection space is constructed through semantic vectors to display the projection of data points on the semantic vectors, and the dimensionality reduction space is the result of dimensionality reduction processing of the data points.

[0054] Furthermore, the parallel attribute vector is divided into three parts: a filtering result row, an attribute row and a weighted sorting row.

[0055] (1) Calculation of semantic vector:

[0056] like Figure 2 As shown, specifically, in the reduced space, the analyst can select the region of interest and another set of points as the target group and the reference group respectively. Then the high-dimensional vector connecting the centers of the two groups is calculated. In particular, if the vector has an absolute value significantly larger than that of other dimensions in some dimensions, and the combination of these dimensions can be interpreted by the analyst as a reasonable concept, this vector is called a semantic vector.

[0057] The analyst selects a set of data points of interest As the target group, another group of points As a reference group, the high-dimensional vector passing through the centers of these two groups of data points is the semantic vector, that is:

[0058]

[0059] Where T is the number of data points in the target group, C is the number of data points in the reference group, N is the number of attributes, and ω∈R 1×N is the attribute weight vector, which is used to modify the weights of each dimension to satisfy The default value is ⊙ represents Hadamard multiplication, also known as pairwise multiplication.

[0060] like Figure 3 As shown in , the semantics of the semantic vector comes from those attributes that have obvious numerical differences between the target group and the reference group. It can also be seen as a sorting criterion based on abstract semantics, where data points are sorted according to their projection positions on the semantic vector. Figure 4 As shown in , this can be explained as follows: in high-dimensional space, as we move along a certain attribute vector, the strength of the semantics represented by the attribute increases or decreases monotonically. The semantic vector is a linear combination of the attribute vectors, so similarly, as we move along the semantic vector, the semantics it represents will also change monotonically in strength or concept. Then, after projecting the data point p onto the semantic vector, its projection position relative to other data points is Describes the strength of the point in the current semantics.

[0061] (2) Visualization of semantic vectors

[0062] The semantic vector component consists of two parts: the reduced dimension space for viewing data distribution and constructing semantic vectors, and the projected space for understanding the constructed semantic vectors and observing the projected distribution of data points on the vectors. In the reduced dimension space, the selection of the target group and the reference group is completed by the lasso tool. The circular spotlight is not used because the shape of the visual structure of interest is not fixed. The data nodes selected as the target group and the reference group are colored red and green respectively. An arrow pointing from the center of the target group to the center of the reference group expresses the metaphor of the vector. In the projected space, the projected distribution of data points is visualized by force-directed Beeswarm. The horizontal direction of the node encodes the projection position, and the radius is proportional to the weighted score. The rectangles on both sides of the vector are called attribute rectangles, and their heights are proportional to the absolute values ​​of the corresponding attribute values. Therefore, they express the semantics carried by the semantic vector. The horizontal position of the attribute rectangle is determined by the positive and negative signs and the absolute value of the corresponding attribute value. This arrangement order highlights the feature attributes and forms a sharp contrast at both ends of the vector, so it can help analysts quickly understand the semantics of the semantic vector.

[0063] (3) Types of semantic vectors

[0064] When constructing a semantic vector, the choice of the target group and the reference group is crucial to the type of vector. The target group is usually a visual structure that the analyst wants to examine, which may be a cluster, an outlier, or a line in the data. The reference group can be the majority of points in the remaining data, or another visual structure with specific characteristics. Based on this, semantic vectors can be divided into unipolar semantic vectors and bipolar semantic vectors.

[0065] like Figure 5 As shown in the figure, the unipolar semantic vector has only one semantic, and the feature attributes above and below it describe the advantages and disadvantages of the target group, respectively. When constructing a unipolar semantic vector, the analyst first selects a target group and a reference group. These selections may be based on the visualization results of the data, where the target group is usually the area of ​​interest to the analyst, while the reference group may be other parts of the data. By comparing the differences between the target group and the reference group, an interpretable semantic vector can be constructed. For example, if the target group is a data point within a specific cluster and the reference group is other data points, then one end of the constructed unipolar semantic vector can represent "belonging to the cluster" and the other end represents "not belonging to the cluster".

[0066] like Figure 6As shown, the bipolar semantic vector whose control group is another cluster. In this case, each end acts as a reference group for the other end. Therefore, each end has a unique semantics that represents the characteristics of the corresponding cluster. The semantics are gradually transferred along the axis. It is worth noting that in this case, the projected position of the data point reflects the relative difference in semantic performance between the two ends rather than the absolute difference. Therefore, the nodes close to one end are the nodes when the semantic performance of the current end is much stronger than that of the other end, and the nodes in the middle of the axis are the nodes with similar semantic performance at both ends.

[0067] (4) Interaction of semantic vectors

[0068] There are two ways to construct semantic vectors: active and passive. Active construction refers to the analyst directly adjusting the height of the attribute rectangle in the projection space to express the existing semantics. Passive construction refers to the automatic generation of semantic vectors in the projection space after the target group and the reference group are circled in the dimensionality reduction space, and the semantics are unknown before construction. In addition, there is a filter brush with adjustable width in the projection space that can slide freely along the vector in the horizontal direction, and the data points covered by it are highlighted in orange in the dimensionality reduction space. By slowly moving the filter brush from one end of the vector to the other end and carefully observing the spatial distribution changes of the highlighted data points in the dimensionality reduction space at the same time, the analyst can understand the distribution and transition of the semantics defined by the semantic vector in space. Here, the importance of this component and interaction is emphasized. It and the circle selection interaction are specific ways to implement the interactive paradigm of the rotation execution mode-oriented exploration and semantic-oriented exploration in the multi-attribute semantic space of the present invention. Among them, the circle selection interaction supports mode-oriented exploration based on visual structures such as clusters and outliers, and the filter brush promotes semantic-oriented exploration by providing distribution and filtering of local semantics in the circled area.

[0069] When performing data interaction, the user operates the semantic vector through the auxiliary visual component of the semantic vector in the projection space to perform visual analysis of the region of interest.

[0070] like Figure 7 As shown, the visualization system of the present invention is a multi-attribute data visual analysis system based on semantic vectors, which includes auxiliary visual components of semantic vectors, mainly including projection space, dimensionality reduction space, weight editor and parallel attribute vectors.

[0071] Projection space and reduced dimension space: These two components are the core parts of the semantic vector. The projection space is constructed through the semantic vector and is used to show the projection of data on the semantic vector, thereby helping analysts understand the semantic structure and characteristics of the data. The reduced dimension space is the result of the original data after dimensionality reduction processing. Through the reduced dimension space, analysts can observe the distribution and structure of the data in the reduced dimension space. Through interactive operations such as filter brushing, analysts can freely explore data in the projection space and reduced dimension space, thereby deeply understanding the semantics of the data.

[0072] Weight Editor (eg Figure 7 The left column in the figure is the weight editor): The weight editor allows analysts to adjust the weight of each attribute in order to build new sorting criteria based on analysis requirements. Analysts can adjust the attribute weight by dragging one side of the attribute corresponding rectangle in the editor and observe the changes in the sorting results in real time. This interactive method is simple and intuitive, greatly improving operation efficiency and analyst experience.

[0073] Parallel attribute vectors: Parallel attribute vectors are designed to provide specific values ​​and rankings of data points in each attribute and support multi-attribute sorting tasks. Parallel attribute vectors are divided into three parts: filter result row, attribute row, and weighted sort row. In the attribute row, the position of the data point in each row represents its value on the corresponding attribute, while the filter result row shows the data points that meet multiple filter conditions at the same time. The weighted sort row is used to show the sorting results of data points based on weighted scores, thereby helping analysts track changes in the sorting.

[0074] In summary, the visualization system of the present invention organically combines semantic vectors with components such as weight editors and parallel attribute vectors through the design of auxiliary visual components to form a complete multi-attribute data visual analysis environment. Multiple visual components and interactions enhance the adaptability to complex real-world scenarios. For example, the weight editor is used to refine attributes in multi-attribute ranking and reduce spatial reconstruction; ranking rows supports combined filtering and detailed inspection, allowing analysts to further verify and gain insights.

[0075] Specifically, the auxiliary visual components include a projection space, a dimension reduction space, a weight editor and a parallel attribute vector.

[0076] The beneficial effects of the present invention are:

[0077] (1) The interaction of the present invention is intuitive and simple, which can improve the analysis efficiency: the traditional method of understanding the dimensionality reduction results usually lacks interactivity, and the analyst needs to rely on static visualization results for analysis, which is inefficient. The present invention introduces an interactive exploration technology - semantic vector, which enables analysts to explore data more flexibly and focus on areas of interest, thereby improving the efficiency of data analysis. Such an interactive method can help analysts discover potential laws and features in the data more quickly, thereby speeding up the analysis process. By introducing the idea of ​​comparison to construct semantic vectors, analysts can understand data in a semantically guided manner in the dimensionality reduction space. Compared with the traditional method of understanding dimensionality reduction results, this method can help analysts understand the semantic structure and features of the data more deeply, thereby improving the depth of data understanding. By comparing the differences between the target group and the reference group, the analyst can construct an interpretable semantic vector, further deepening the understanding of the data. The present invention not only includes basic interactive methods such as circle selection, but also has functions such as filter brushes that dynamically display semantic distribution, which conforms to the analyst's thinking habits and improves the analyst's experience and work efficiency.

[0078] (2) The computation of the present invention is simple and efficient: it is applicable to large-scale data sets and can respond instantly even to data of millions. Traditional data analysis methods face the problems of high computational complexity and slow response speed when processing large-scale data, but the present invention adopts efficient computational algorithms and optimization techniques to quickly process large-scale data sets. This simple and efficient computational feature ensures that analysts can obtain results in a shorter time, quickly conduct data exploration and analysis, and improve work efficiency and analysis efficiency.

[0079] (3) The present invention supports multi-attribute data analysis: The present invention is not only applicable to the understanding of dimensionality reduction results of document data, but can also be applied to the analysis of multi-attribute data. The dimensionality reduction result understanding and multi-attribute sorting are unified in a context-consistent interactive environment, realizing a visual analysis system that supports multi-attribute data analysis, allowing analysts to complete a variety of data analysis tasks in the same interface. By introducing visual components such as weight editors and parallel attribute vectors, analysts can flexibly adjust attribute weights and compare data performance on different attributes, supporting more complex multi-attribute data analysis tasks. Such an analysis method can help analysts understand the data more comprehensively and discover hidden information in the data.

[0080] (4) The applicability and flexibility of the present invention are strong: the present invention does not rely on a specific linear or nonlinear dimensionality reduction algorithm, but provides a general interactive exploration framework that can be combined with various dimensionality reduction algorithms. Analysts can select a suitable dimensionality reduction algorithm according to specific needs and embed it into the framework of the present invention for data analysis and visualization. Since the design concept of the present invention is universal, it can be seamlessly integrated with existing dimensionality reduction algorithms and data processing tools, whether it is the classic PCA, LLE or popular algorithms such as t-SNE, they can be integrated and applied with the framework of the present invention. The present invention is not only applicable to dimensionality reduction algorithms, but can also be applied to other data analysis and visualization tasks, such as clustering, classification, association rule mining, etc., and has wide applicability and application prospects.

[0081] In the description of this specification, the description of the terms "one embodiment", "a preferred embodiment", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0082] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A comparison-based interactive dimensionality reduction result understanding method, characterized in that: Specifically include: The target group and the reference group are constructed based on the regions of interest and selected regions in the scatter plot; Calculating a semantic vector connecting the centers of the target group and the reference group according to the semantic understanding results of the data points, projecting the data points onto the semantic vector, and constructing a projection space using the semantic vector; When performing data interaction, the user operates the semantic vector through the auxiliary visual component of the semantic vector in the projection space to perform visual analysis of the region of interest.

2. The method for understanding interactive dimensionality reduction results based on comparison according to claim 1, characterized in that: The region of interest is determined according to the selection result of the user.

3. The method for understanding interactive dimensionality reduction results based on comparison according to claim 1, characterized in that: The semantic vector is a vector connecting the centers of the target group and the reference group.

4. The method for understanding interactive dimensionality reduction results based on comparison according to claim 1, characterized in that: The semantic vector is divided into a unipolar semantic vector and a bipolar semantic vector.

5. The method for understanding interactive dimensionality reduction results based on comparison according to claim 1, characterized in that: The projection space is constructed using the semantic vector, specifically including: Constructing a projection space according to the distribution of data points corresponding to the semantic vector; The projection space is provided with a filter brush with adjustable width and freely sliding along the vector in the horizontal direction, and the distribution of the semantic changes described by the semantic vector in the projection space is revealed by sliding the filter brush.

6. The method for understanding interactive dimensionality reduction results based on comparison according to claim 1, characterized in that: The auxiliary visual components include a projection space, a dimension reduction space, a weight editor, and a parallel attribute vector.

7. The method for understanding interactive dimensionality reduction results based on comparison according to claim 6, characterized in that: The projection space is constructed through the semantic vector and is used to display the projection of the data point on the semantic vector. The dimensionality reduction space is the result of the dimensionality reduction processing of the data point.

8. The method for understanding interactive dimensionality reduction results based on comparison according to claim 6, characterized in that: The parallel attribute vector is divided into three parts: a filter result row, an attribute row and a weighted sorting row.

Images