Privacy preserving interactive editing visual analytics system for time series

Abstract

This invention provides an interactive editing and visual analysis system for time series data, belonging to the field of data privacy protection. The visual analysis system includes ranking views, flow views, radial views, pattern views, editing views, and component views. Combining domain knowledge, it performs linked detection and visual localization of privacy risks in time series data from three dimensions: time, amplitude, and pattern. Through interactive editing in each view module, it provides three types of direct manipulation editing tools: time perturbation, amplitude perturbation, and pattern replacement. This allows users to precisely modify identified sensitive segments and supports batch processing of multiple time series and fine-grained editing of series trends and residual components. The modules are closely linked through data and events, forming a complete closed loop of "identification-editing-verification," effectively masking individual privacy characteristics while maximizing the overall utility and statistical authenticity of the time series data.

Description

Technical Field

[0001] This invention relates to the field of data privacy protection technology, and in particular to a time-series interactive editing and visual analysis system for privacy protection. Background Technology

[0002] With the widespread adoption of data collection technologies in fields such as healthcare, energy, and transportation, the sharing and publication of time-series data has provided crucial support for data-driven decision-making and innovative research. However, publishing raw time-series data faces significant privacy risks, especially data involving human behavior (such as household electricity consumption and motion tracking data). The unique time-series patterns hidden within this data (such as specific daily routines, anomalous numerical fluctuations, or specific behavioral sequences) can form "behavioral fingerprints," enabling attackers to re-identify individuals through link attacks combined with external auxiliary information.

[0003] To mitigate these risks, existing technologies mainly fall into two categories: (1) Traditional manual processing: Manually delete or modify sensitive data using spreadsheets or custom code. This method is not only labor-intensive and error-prone, but also difficult to scale and highly dependent on the operator's experience.

[0004] (2) Automated algorithms: including differential privacy and data synthesis methods based on generative adversarial networks (GANs). Differential privacy often introduces excessive noise, severely compromising the temporal relevance and analytical utility of the data. Although synthesis methods based on GANs or diffusion models can generate new data, the process is usually "black box," lacking interpretability and controllability. Experts find it difficult to incorporate domain knowledge and cannot make precise modifications to specific high-risk segments. However, these technologies are significantly insufficient in effectively identifying complex and diverse temporal privacy risks and in providing intuitive and controllable risk mitigation methods without compromising data utility.

[0005] Visual analytics systems are widely used in deep learning fields, including time series prediction models. Existing visual analytics work for time series prediction models combines model interpretation analysis techniques to intuitively visualize the model's internal structural parameters, inputs, and outputs.

[0006] For example, line graphs have become the most classic method for visualizing time data, and variations based on this method, such as time curves, also exist. When visualizing multiple time series, existing solutions can be divided into two main categories: shared space (e.g., simple line graphs and stacked area graphs) and segmented space (e.g., multiplier graphs and horizontal graphs). Systems such as TimeSearcher, GeoChron, and KD-Box widely adopt shared space layouts because they outperform segmented space layouts in identifying overall trends and patterns in large-scale time series data. However, the dense overlapping of lines in a shared space layout can sometimes obscure finer details and introduce visual clutter.

[0007] To address these challenges, dimensionality reduction and clustering techniques have become powerful alternatives for visualizing large-scale time series data. Many advanced visualization and analytics systems have embedded dimensionality reduction and clustering views to aid data analysis, enabling users to explore patterns, trends, and correlations more efficiently. However, these tools are not specifically designed to discover privacy breach patterns through visualization techniques. Summary of the Invention

[0008] The purpose of this invention is to provide a time-series interactive editing and visual analysis system for privacy protection. By constructing a time-series privacy risk classification system and combining it with multi-view visual analysis technology, it helps users quickly identify privacy leakage risks from three dimensions: time, magnitude, and pattern. It also provides an interactive editing tool inspired by the idea of ​​direct manipulation, enabling users to perform precise targeted editing of high-risk data segments. This addresses the problems of existing automated methods lacking controllability and interpretability, and manual methods being inefficient. It can effectively mitigate privacy risks while preserving the statistical characteristics and downstream analytical utility of the data to the greatest extent.

[0009] To achieve the above-mentioned objectives, an embodiment provides a privacy-preserving time-series interactive editing and visual analysis system, comprising: The ranking view is used to dynamically calculate the risk score of a time series and sort and display users according to the risk score to guide users to identify high-risk users. The flow view, with a background of a data distribution flow graph and an overlay of outliers in the foreground, is used to identify and visualize outliers in time series based on locked high-risk users. It supports interactive hovering to link with other views and present risk levels. The radial view displays each user's time series in the form of concentric rings and has a color mapping function. It highlights individual trends and group distribution through time slices and is used to analyze privacy risks related to time and amplitude. The pattern view, which includes a symbol overview and a card list, is used to automatically detect and visualize key patterns in time series, and supports interactive filtering and linking to view specific pattern instances. The edit view, based on the privacy risk classification in the time series, provides three types of editing operations: time perturbation, amplitude perturbation, and pattern replacement. It supports direct manipulation and real-time feedback on the selected time series to mitigate privacy risks. The component view is used to treat multiple time series as independent levels, supports unified editing operations to be applied in batches to selected levels, and introduces time series decomposition to achieve component-based editing.

[0010] In one embodiment, the risk score in the ranking view is calculated based on a weighted average of outlier values ​​in the flow view and key patterns in the pattern view, using the following formula: , in, For users in the flow view The outlier count, P = {peaks, plateaus, steps, periodic behavior} is the set of four key patterns detected in the pattern view. For users In mode Instance count on, and The weighting coefficients for each item are determined to incorporate domain knowledge.

[0011] In one embodiment, the streaming view uses the following steps to generate and identify outliers: for each user, aggregate time series data for all available weeks by a day of the week; perform time dimension downsampling on the aggregated time series data; for each timestamp, calculate the statistical percentile and interquartile range of the time series data, and identify and highlight time series data points that exceed the user's adjustable threshold as outliers.

[0012] In one embodiment, the color mapping function in the radial view is implemented by: performing initial color mapping using a normalization method based on global data extrema; providing an interactive range slider that allows users to dynamically set the upper and lower boundaries of the range of values ​​of interest; and recalculating and constraining the color mapping intensity based on the upper and lower boundaries of the range of values ​​of interest set by the user, in order to suppress the interference of extreme outliers on the overall color distribution.

[0013] In one embodiment, key patterns include peak, plateau, step, and periodic behavior.

[0014] In one embodiment, the symbol overview in the pattern view is encoded in the following way: each pattern instance is represented by a unique symbol, the horizontal position of the symbol encodes the time when the pattern occurs, and the vertical position of the symbol encodes the amplitude or intensity of the corresponding pattern, so that the user can quickly perceive the distribution of a specific pattern in time and amplitude. The list of cards in the pattern view displays a thumbnail line graph of each pattern instance, providing visual evidence and contextual information for the detected patterns.

[0015] In one embodiment, privacy risks in time series are categorized as time risk, magnitude risk, and pattern risk. Among them, time risk refers to the time location of an anomaly or the moment when the event occurs; amplitude risk refers to statistical outliers, including extreme single-point spikes or continuous numerical plateaus; pattern risk refers to unique sequence shapes or behavioral fingerprints, including periodic patterns or specific waveforms.

[0016] In one embodiment, the three types of editing operations in the edit view—time perturbation, amplitude perturbation, and pattern replacement—specifically include: The time perturbation editing operation includes horizontal translation and time scaling; the horizontal translation allows users to horizontally translate selected time series data segments along the time axis, thereby adjusting the time positioning of the time series data segments while maintaining the original value distribution; the time scaling is used to stretch or compress selected time series data segments to blur the duration of events and time nodes. The amplitude perturbation editing operation includes vertical displacement and curve mapping. The vertical displacement allows the user to adjust the size of the selected time series data segment vertically along the numerical axis based on the selected time series data segment. The curve mapping provides a mapping curve with several draggable control points. The user defines a non-linear mapping function between the original value and the target value by adjusting the position of the control points. The non-linear mapping function is applied to all numerical points of the selected time series data segment, and the numerical values ​​are non-linearly remapped by adjusting the curve control points. The pattern replacement editing operation includes cloning and removal. Cloning is used by the user to select a normal time series data segment from the source time series, copy the time series data of this segment and cover the target sensitive period, and automatically perform smooth transition processing at the splicing boundary to ensure the continuity of the sequence. Removal is used to automatically retrieve all candidate time series data segments with the same start and end values ​​as the target sensitive period in the dataset, exclude time series data segments from the same user, and sort the candidate time series data segments according to the degree of morphological difference with the original time series data segments for the user to select or automatically replace.

[0017] In one embodiment, the edit view also has redo, undo, and semi-automation assistance functions; wherein the semi-automation assistance functions include: when importing time series data into the edit view, analyzing the data source and data characteristics and providing editing tools; and, after a user manually edits a single layer, the user can choose to apply the same transformation to all unedited layers in batches.

[0018] In one embodiment, time series decomposition in the component view includes: decomposing the original time series into low-frequency trend items and high-frequency residual items, allowing users to select in the edit view to apply editing operations only to the trend items or only to the residual items, so as to achieve targeted modifications to different information levels of the time series data.

[0019] Traditional methods often involve a one-size-fits-all approach to modifying the entire time series, while different components of the time series (such as low-frequency trends and high-frequency residuals) typically carry different informational values ​​and privacy risks. Therefore, decomposing the raw data into these editable, independent components allows for highly precise modifications: for example, users can focus on adjusting high-frequency residuals to mask specific short-term behaviors while preserving the main trends to ensure the overall usability of the dataset.

[0020] Compared with the prior art, the beneficial effects of the present invention include at least the following: This invention provides a privacy-preserving interactive editing and visual analysis system for time series data, including a ranking view, a flow view, a radial view, a pattern view, an editing view, and a component view. Combining domain knowledge, it performs linked detection and visual localization of privacy risks in time series data from three dimensions: time, amplitude, and pattern. Through interactive editing in each view module, it provides three types of direct manipulation editing tools: time perturbation, amplitude perturbation, and pattern replacement. This allows users to precisely modify identified sensitive segments and supports batch processing of multiple time series and fine-grained editing of series trends and residual components. The modules are closely linked through data and events, forming a complete closed loop of "identification-editing-verification," effectively masking individual privacy characteristics while maximizing the overall utility and statistical authenticity of the time series data. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0022] Figure 1 A schematic diagram of the structure of a privacy-preserving time-series interactive editing visual analytics system.

[0023] Figure 2 This is a schematic diagram of the ranking view.

[0024] Figure 3 This is a schematic diagram of the flow view.

[0025] Figure 4 This is a schematic diagram of a radial view.

[0026] Figure 5 This is a schematic diagram of the pattern view.

[0027] Figure 6 This is a schematic diagram of the edit view.

[0028] Figure 7 This is a diagram illustrating the effects of three types of editing operations.

[0029] Figure 8 This is a schematic diagram of the component view.

[0030] Figure 9 A schematic diagram illustrating the interactive operation of a case study on privacy protection of household electricity load data.

[0031] Figure 10 This is a schematic diagram illustrating the interactive operation of a case study on multivariate analysis and batch processing of motion tracking data. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the accompanying drawings and... The embodiments further illustrate the present invention in detail. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0033] This embodiment provides a privacy-preserving time-series interactive editing and visual analysis system. By constructing a time-series privacy risk classification system and combining it with multi-view visual analysis technology, it helps users quickly identify privacy leakage risks from three dimensions: time, magnitude, and pattern. It also provides an interactive editing tool inspired by direct manipulation, enabling users to precisely target and edit high-risk data segments. This addresses the problems of existing automated methods lacking controllability and interpretability, and manual methods being inefficient. It effectively mitigates privacy risks while maximizing the preservation of the data's statistical characteristics and downstream analytical utility. Figure 1 As shown, the time-series interactive editing visual analysis system provided in the embodiment includes a ranking view, a flow view, a radial view, a pattern view, an editing view, and a component view.

[0034] Ranking View: Used to dynamically calculate risk scores for time series data, and to sort and display users based on these risk scores to guide users to identify high-risk users.

[0035] In the embodiments, such as Figure 2 As shown, after the operator uploads the dataset, the system automatically calculates the risk score and displays the user IDs in a list sorted by the risk score, guiding users to prioritize high-risk individuals and quickly identify high-risk users in the ranking view.

[0036] The risk score is calculated by weighting outliers in the flow view and key patterns in the pattern view, and is used to quantify the degree of risk. The calculation formula is as follows: , in, For users in the flow view Outlier count, ={Peaks, Plateaus, Steps, Periodic Behaviors} represents the set of four key patterns detected in the pattern view. For users In mode Instance count on, and The view incorporates various weighting coefficients determined to incorporate domain knowledge. Its core design philosophy prioritizes scalability. By highlighting high-scoring users, it serves as a starting point for analysis, helping operators quickly locate the most potentially valuable key information within massive amounts of time-series data.

[0037] Stream View: Its background is a flow graph of data distribution, with outliers overlaid in the foreground. It is used to identify and visualize outliers in time series based on locked high-risk users. It supports interactive hovering to link with other views and present the risk level.

[0038] In the embodiments, such as Figure 3 As shown, the flow view is designed to identify privacy risks presented as outliers that may expose vulnerabilities related to the "size" of anomalous time-series data or specific "patterns" exhibiting outlier characteristics. The view employs an improved boxplot design, with a flow graph of the data distribution as the background and outliers overlaid in the foreground, allowing users to control risk thresholds by adjusting detection sensitivity.

[0039] Specifically, in order to effectively visualize the entire dataset, a two-stage data preprocessing is first performed: In the first stage, in order to capture representative daily behaviors, the data is aggregated by a day of the week, and the average is calculated for each time point of all available weeks for each individual; In the second stage, time downsampling is applied by averaging every k consecutive points, which reduces visual clutter and adapts to display limitations.

[0040] This condensed and aggregated data, inspired by box plots, provides an intuitive visualization based on risk levels. For each timestamp, key statistical percentiles are first calculated. The underlying data distribution is then presented as a manifold with two nested shaded bands: the darker inner band represents data between the 25th and 75th percentiles (interquartile range, IQR), while the lighter outer band shows the wider range of the 10th to 90th percentiles. Standard box plot methods are used to identify outliers. Data points outside the range defined by the multiplier are marked as outliers. To provide users with analytical flexibility, this multiplier can be customized (default value 1.5), allowing them to adjust detection sensitivity based on different privacy risk thresholds. These detected outliers are then overlaid on the manifold as individual data points, immediately highlighting them. Furthermore, outliers belonging to the same user and workday are correlated over time, helping to simultaneously highlight both outliers and persistent biases.

[0041] Radial view: Displays each user's time series in concentric rings, with color mapping function, highlighting individual trends and group distribution through time slices, used to analyze privacy risks related to time and amplitude; In the embodiments, such as Figure 4 As shown, the radial view is designed to analyze privacy risks related to the magnitude of time features. The aggregated time series of each individual is presented in a radially stacked form of concentric rings. The core interactive mechanism is a dual highlighting function that supports multi-view analysis: hovering the mouse over any time period will simultaneously highlight (yellow) the complete ring data of the corresponding user (a longitudinal view of their individual trends) and the distribution of all users in the same time slice (a cross-sectional view of group behavior).

[0042] To enable more detailed user profile analysis, users can double-click the circular data to switch to individual mode, where each concentric ring represents that user's daily data. By mapping time to angular position, this view helps identify periodic fluctuations and time trends.

[0043] The color mapping function in the radial view is implemented in the following ways: the initial color mapping is performed using a normalization method based on the global data extrema; an interactive range slider (range 0-1.7) is provided, allowing users to dynamically set the upper and lower boundaries of the range of values ​​of interest; the color mapping intensity is recalculated and constrained according to the upper and lower boundaries of the range of values ​​of interest set by the user, in order to suppress the interference of extreme outliers on the overall color distribution.

[0044] By default, the ring structure in the radial view is arranged sequentially from the inside out by user ID. This random distribution hinders users from discovering aggregation patterns and periodic features. To mitigate the information imbalance caused by the difference in inner and outer ring radii, the Revert operation allows users to switch the order of the rings. This layout optimization helps identify clusters with similar temporal patterns and clearly reveals periodic recurring trends.

[0045] Pattern View: Includes a symbol overview and card list, used to automatically detect and visualize key patterns in time series, and supports interactive filtering and linking to view specific pattern instances.

[0046] In the embodiments, such as Figure 5 As shown, the pattern view is designed to automatically detect and visualize privacy risks based on "patterns." The system is pre-configured to identify four common time patterns: peaks, plateaus, steps, and periodic behavior.

[0047] This view includes a symbol-based overview (summarizing all detected patterns) and a scrollable list of cards (for detailed examination) for the operator to filter. In the symbol overview, each detected pattern instance is represented by a unique symbol. The horizontal position of the symbol encodes the time of pattern occurrence, while the vertical position indicates the amplitude or intensity of the corresponding pattern. This allows users to quickly perceive the distribution of specific patterns, such as whether high-amplitude peaks cluster together. The card list below displays a thumbnail line graph of each instance, providing visual evidence and contextual information about the detected patterns.

[0048] Edit View: Based on the privacy risk classification in the time series, it provides three types of editing operations: time perturbation, amplitude perturbation, and pattern replacement. It supports direct operation on the selected time series and real-time feedback to mitigate privacy risks.

[0049] In this embodiment, privacy risks in time series are defined as three types: time risk, amplitude risk, and pattern risk. Among them, time risk refers to the abnormal time location or the moment of event occurrence (such as atypical activity time); amplitude risk refers to statistical outliers (such as extreme single-point spikes or continuous numerical plateaus); and pattern risk refers to unique sequence shapes or behavioral fingerprints (such as periodic patterns or specific waveforms).

[0050] Interactive editing is performed based on these three risk types, such as... Figure 6 As shown, the edit view is the core workspace of the visual analytics system, providing a toolset inspired by the concept of direct manipulation interaction. It supports direct manipulation of time series data to mitigate privacy risks. For detailed instructions, please refer to [link to documentation / reference]. Figure 7 : The time perturbation editing operations include horizontal translation (Move-x) and time expansion (Expand); Move-x allows users to horizontally translate selected time series data segments along the time axis, adjusting the time positioning of the time series data segments while maintaining the original value distribution; Expand stretches or compresses selected time segments to cover the entire time axis, used to blur the duration of events and time nodes; The amplitude perturbation editing operation includes vertical displacement (Move-y) and curve mapping (Curve). Vertical displacement allows users to adjust the size of a selected time series data segment vertically along the numerical axis based on the selected time series data segment. Curve mapping provides a mapping curve with several draggable control points. Users define a non-linear mapping function between the original value and the target value by adjusting the position of the control points. The non-linear mapping function is applied to all numerical points of the selected time series data segment, and the numerical values ​​are non-linearly remapped by adjusting the curve control points. Pattern replacement editing operations include cloning and removal. Cloning allows users to select a normal time series data segment from the original time series, copy the data of this segment to cover the target sensitive period, and automatically perform smooth transition processing at the splicing boundary to ensure the continuity of the series. Removal automatically retrieves all candidate data segments in the dataset that have the same start and end values ​​as the target sensitive period. After excluding time series data segments from the same user, the candidate time series data segments are sorted according to the degree of morphological difference from the original time series data segments for users to select or for automatic replacement.

[0051] The edit view also features redo, undo, and semi-automation assistance functions. Among these, the semi-automation assistance functions include: analyzing the data source and data characteristics when importing time series data into the edit view and providing editing tools; and allowing users to manually edit a single layer and then choose to apply the same transformation to all unedited layers in batches.

[0052] Component View: Used to treat multiple time series as independent levels, supports unified editing operations applied in batches to selected levels, and introduces time series decomposition to achieve component-based editing.

[0053] In the embodiments, such as Figure 8 As shown, this view directly addresses the challenge of applying uniform modifications across multiple time series with similar privacy risks. In this view, each time series is treated as an independent tier. Users can select multiple tiers simultaneously, for example, grouping all series exhibiting similar high-risk patterns. Crucially, any operation performed in the edit view is uniformly applied to all selected tiers. This feature significantly improves editing efficiency and ensures temporal consistency when processing large datasets.

[0054] The component view also introduces time series decomposition to enable component editing, including: decomposing the original time series ( Figure 8 The original data is broken down into low-frequency trend items and high-frequency residual items, allowing users to select to apply editing operations only to trend items or only to residual items in the edit view, so as to achieve targeted modification of different information levels of the data and more refined privacy protection.

[0055] The following section, using two real-world examples (REFIT power load dataset and Capture-24 activity tracking dataset) and accompanying figures, details the specific implementation process of this invention.

[0056] Case 1: Privacy Protection of Household Electricity Load Data This embodiment, based on the REFIT power load measurement dataset, demonstrates how the system can detect and mitigate privacy risks associated with anomalous temporal distributions and regular patterns. The specific process corresponds to... Figure 9 As shown.

[0057] 1. Identification and editing of abnormal nighttime electricity usage patterns (1-1) Risk Identification: The operator first focuses on the user with the highest risk score in the ranking view. After selecting this user, as follows... Figure 9 As shown in A1, a series of significant outliers were observed in the flow view. Hovering over the outliers to examine the details revealed that these outliers belonged to user 8. Figure 9 As shown in A2, the detailed curves reveal an extremely high peak in electricity consumption between 5:00 AM and 7:30 AM, while consumption is almost zero during the typical evening hours. This anomaly is represented by highly saturated dark areas in the radial view (see Figure 1). Figure 9 A3 in the model view is classified as a "spiky" pattern and has the largest deviation (see A3 in the model view). Figure 9 (A4 in the text). These features together constitute the "behavioral fingerprint" that poses a high risk of identity theft.

[0058] (1-2) Interactive editing: Time Disruption: To obscure the time characteristics, the operator used the Move-x tool to shift the entire abnormal power consumption segment of user 8 from 5:00 to 7:30 to a more regular time of around 8:00 AM (see [link]). Figure 9 (C1 in the middle).

[0059] Amplitude perturbation: Observing that the value was still too high after shifting, the operator used the Move-y tool to vertically compress the peak value by about 50% (see...). Figure 9 (C2 in the middle).

[0060] Data Generation and Repair: To address the data gaps left from the original time period and fill the evening downtime, the operator used the Clone tool to copy the normal low-power mode from midnight to 5:00 AM to the 6:00 PM time period (see...). Figure 9 (C3 in the original text); then the Removal tool is used to mask the original sensitive segment (see...). Figure 9 (C4 in the middle).

[0061] (1-3) Effectiveness Evaluation: After editing, the view is updated synchronously, and User 8's abnormal pattern no longer shows significant deviation (see...). Figure 9 (E1 in the middle).

[0062] 2. Decomposition and Editing of Weekend Routine Behaviors (2-1) Risk Identification: To analyze the differences in electricity consumption patterns between weekdays and weekends, the operator switched the navigation bar to compare weekday and weekend patterns, and found obvious outliers in the flow view that only appeared on weekend afternoons (see...). Figure 9 (See B1 in the details). Details show that user 10 experiences a consistent, sharp surge in electricity consumption every Saturday afternoon (see B1 in the details). Figure 9 B2 in the text suggests fixed patterns of family activities, which may pose pattern-based privacy risks.

[0063] (2-2) Fine-grained editing based on decomposition: To remove privacy risks while preserving data utility, the operator enabled the time series decomposition function, breaking down the original series into low-frequency "trend terms" and high-frequency "residual terms" (see...). Figure 9 (D1 in the middle).

[0064] Trend Item Editing: The operator only used the Move-y tool on the trend item to reduce the peak power from over 6000 watts to a more typical 4000 watts (see...). Figure 9 (D2 in the middle).

[0065] Residual term editing: The operator only uses the Curve tool on the residual terms, applying a "half-value" template to smooth high-frequency fluctuations (see...). Figure 9 D3 in ).

[0066] The edited results show that although user 10's electricity consumption on Saturday afternoon was still relatively high, it no longer constituted an outlier (see [link]). Figure 9 D4 and Figure 9 In E2), privacy protection was achieved.

[0067] Case Study 2: Multivariate Analysis and Batch Processing of Motion Tracking Data This embodiment is based on the Capture-24 dataset, which contains 61,920 triaxial (x, y, z) acceleration sampling points, demonstrating how the system processes multiaxial acceleration data and performs batch editing. The specific workflow corresponds to... Figure 10 .

[0068] (1) Risk Identification: In the radial view, the operator adjusts the color mapping to highlight areas close to zero. Users 72 and 97 exhibit the same white band (i.e., persistent zero value) during non-sleep periods, suggesting that the device was intentionally removed (see [link]). Figure 10 A1 and Figure 10 This finding (A2 in the original text) is confirmed in the "Plateau" category of the pattern view, where the symbol distribution shows a clear clustering of low-amplitude, low-variance plateau periods. Brushing this clustered area filters the card list, revealing specific periods of disconnection, ultimately leading to the same conclusion.

[0069] Batch editing: The operator drags the data for user 72 and user 97 into the edit view and observes that there is continuous acceleration on all three axes (see...). Figure 10 B1 and Figure 10 (See B2 in the original text), therefore it was decided to exclude these periods of disconnected data. For user 97, the operator used the Expand tool to stretch the valid activity data segments before and after it to cover the non-wearing gaps in between (see...). Figure 10 (C2 in the middle).

[0070] To improve efficiency, the operator uses the batch application function in the component view to apply the same Expand operation logic to user 72's layer with a single click (see [link]). Figure 10 (C1 in the middle).

[0071] Results Verification: After synchronization, all abnormal white stripes in the radial view were successfully eliminated (see [link]). Figure 10 (A3 in the text).

[0072] In summary, the privacy-preserving time-series interactive editing and visual analysis system constructed in this invention brings the following technical effects: (1) Full-process coverage and human-machine collaboration: This invention establishes a complete workflow from risk identification and targeted editing to effect evaluation. Compared with fully automatic "black box" algorithms, this system introduces expert domain knowledge to determine which patterns are sensitive, thereby avoiding data utility loss caused by overprotection.

[0073] (2) Multi-dimensional risk discovery capability: Through multiple coordinated views such as flow view, radial view and pattern view, the system can help users discover complex privacy risks that are difficult to detect from a single perspective, such as abnormal behavior that occurs at a specific time or regular patterns across cycles.

[0074] (3) Intuitive and precise editing control: Inspired by the idea of ​​direct manipulation and interaction, the system provides a set of intuitive editing tools, which lowers the learning threshold and enables users to make pixel-level precise modifications to the time series and preview the impact of the modifications on the statistical characteristics of the data in real time.

[0075] (4) Efficient batch processing: Component view and automatic batch processing function allow users to quickly apply an editing strategy for a pattern to all similar sequences when processing large datasets, which significantly improves editing efficiency and ensures processing consistency.

[0076] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A privacy-aware time-series interactive editing visual analytics system, characterized in that, include: The ranking view is used to dynamically calculate the risk score of a time series and sort and display users according to the risk score to guide users to identify high-risk users. The flow view, with a background of a data distribution flow graph and an overlay of outliers in the foreground, is used to identify and visualize outliers in time series based on locked high-risk users. It supports interactive hovering to link with other views and present risk levels. The radial view displays each user's time series in the form of concentric rings and has a color mapping function. It highlights individual trends and group distribution through time slices and is used to analyze privacy risks related to time and amplitude. The pattern view, which includes a symbol overview and a card list, is used to automatically detect and visualize key patterns in time series, and supports interactive filtering and linking to view specific pattern instances. The edit view, based on the privacy risk classification in the time series, provides three types of editing operations: time perturbation, amplitude perturbation, and pattern replacement. It supports direct operation on the selected time series and real-time feedback to mitigate privacy risks. The component view is used to treat multiple time series as independent levels, supports unified editing operations to be applied in batches to selected levels, and introduces time series decomposition to achieve component-based editing.

2. The time- series interactive editing visual analytics system of claim 1, wherein, The risk score in the ranking view is calculated by weighting outliers in the flow view and key patterns in the pattern view, using the following formula: ， where, is the anomaly value count of the user in the flow view, is the set of four key patterns detected in the pattern view, is the instance count of the user on the pattern , and and are the respective weight coefficients determined by introducing domain knowledge.

3. The time-series interactive editing and visual analysis system according to claim 1, characterized in that, The streaming view uses the following steps to generate and identify outliers: For each user, aggregate time series data for all available weeks by a specific day of the week; perform time dimension downsampling on the aggregated time series data; for each timestamp, calculate the statistical percentile and interquartile range of the time series data, and identify and highlight time series data points that exceed the user's adjustable threshold as outliers.

4. The time-series interactive editing and visual analysis system according to claim 1, characterized in that, The color mapping function in the radial view is implemented in the following ways: the initial color mapping is performed using a normalization method based on the global data extrema; an interactive range slider is provided, allowing users to dynamically set the upper and lower boundaries of the range of values ​​of interest; the color mapping intensity is recalculated and constrained according to the upper and lower boundaries of the range of values ​​of interest set by the user, in order to suppress the interference of extreme outliers on the overall color distribution.

5. The time-series interactive editing and visual analysis system according to claim 1, characterized in that, Key patterns include peaks, plateaus, step changes, and cyclical behavior.

6. The time-series interactive editing and visual analysis system according to claim 5, characterized in that, The symbol overview in the pattern view is encoded in the following way: each pattern instance is represented by a unique symbol, the horizontal position of the symbol encodes the time when the pattern occurs, and the vertical position of the symbol encodes the amplitude or intensity of the corresponding pattern, so that users can quickly perceive the distribution of a specific pattern in time and amplitude. The list of cards in the pattern view displays a thumbnail line graph of each pattern instance, providing visual evidence and contextual information for the detected patterns.

7. The time-series interactive editing and visual analysis system according to claim 1, characterized in that, Privacy risks in time series are categorized into time risk, magnitude risk, and pattern risk. Among them, time risk refers to the time location of an anomaly or the moment when the event occurs; amplitude risk refers to statistical outliers, including extreme single-point spikes or continuous numerical plateaus; pattern risk refers to unique sequence shapes or behavioral fingerprints, including periodic patterns or specific waveforms.

8. The time-series interactive editing and visual analysis system according to claim 7, characterized in that, The three types of editing operations in the edit view—time perturbation, amplitude perturbation, and pattern replacement—specifically include: The time perturbation editing operation includes horizontal translation and time scaling; the horizontal translation allows users to horizontally translate selected time series data segments along the time axis, thereby adjusting the time positioning of the time series data segments while maintaining the original value distribution; the time scaling is used to stretch or compress selected time series data segments to blur the duration of events and time nodes. The amplitude perturbation editing operation includes vertical displacement and curve mapping. The vertical displacement allows the user to adjust the size of the selected time series data segment vertically along the numerical axis based on the selected time series data segment. The curve mapping provides a mapping curve with several draggable control points. The user defines a non-linear mapping function between the original value and the target value by adjusting the position of the control points. The non-linear mapping function is applied to all numerical points of the selected time series data segment, and the numerical values ​​are non-linearly remapped by adjusting the curve control points. The pattern replacement editing operation includes cloning and removal. Cloning is used by the user to select a normal time series data segment from the source time series, copy the time series data of this segment and cover the target sensitive period, and automatically perform smooth transition processing at the splicing boundary to ensure the continuity of the sequence. Removal is used to automatically retrieve all candidate time series data segments with the same start and end values ​​as the target sensitive period in the dataset, exclude time series data segments from the same user, and sort the candidate time series data segments according to the degree of morphological difference with the original time series data segments for the user to select or automatically replace.

9. The time-series interactive editing and visual analysis system according to claim 1, characterized in that, The time series decomposition in the component view includes: decomposing the original time series into low-frequency trend items and high-frequency residual items, allowing users to select to apply editing operations only to the trend items or only to the residual items in the edit view, so as to achieve targeted modifications to different information levels of the time series data.

Images