提供异常反馈Provide anomaly feedback

用户反馈是异常检测系统中发现缺陷的重要方法之一。User feedback is one of the most important methods to discover defects within the anomaly detection system. 在这里,我们为用户提供了一种直接在时序上标记错误检测结果的方法,并立即应用反馈。Here we provide a way for users to mark incorrect detection results directly on a time series, and apply the feedback immediately. 这样,用户可以指示异常检测系统如何通过主动交互对特定时序执行异常检测。In this way, a user can teach the anomaly detection system how to do anomaly detection for a specific time series through active interactions.

备注

目前,反馈只会通过“智能检测”来影响异常检测结果,而不会通过“硬阈值”和“变化阈值”来影响异常检测结果 。Currently feedback will only affect anomaly detection results by Smart detection but not Hard threshold and Change threshold.

如何提供时序反馈How to give time series feedback

可以在任何序列上的指标详细信息页提供反馈。You can provide feedback from the metric detail page on any series. 只需选择任意点,你就会看到以下反馈对话框。Just select any point, and you will see the below feedback dialog. 此对话框将显示所选序列的维度。It shows you the dimensions of the series you've chosen. 你可以重新选择维度值,甚至可以删除其中一些维度值来获取一批时序数据。You can reselect dimension values, or even remove some of them to get a batch of time series data. 选择时序后,可选择“添加”按钮添加反馈,你可以提供四种反馈。After choosing time series, select the Add button to add the feedback, there are four kinds of feedback you could give. 若要追加多个反馈项,请在完成批注后,选择“保存”按钮。To append multiple feedback items, select the Save button once you complete your annotations.

具有蓝线和在不同点具有红点的时序数据关系图。红色框围绕一个点,并包含文本:选择任意点

“添加反馈”对话框,其中包含两个维度以及“选择或删除维度”和“添加反馈”选项。

标记异常点类型Mark the anomaly point type

如下图所示,“反馈”对话框将自动填充所选点的时间戳,但你可以编辑此值。As shown in the image below, the feedback dialog will fill the timestamp of your chosen point automatically, though you can edit this value. 然后选择是否要将此项标识为 AnomalyNotAnomalyAutoDetectYou then select whether you want to identify this item as an Anomaly, NotAnomaly, or AutoDetect.

具有“Anomaly”、“NotAnomaly”和“AutoDetect”选项的下拉菜单

该选择会将你的反馈应用于同一序列的未来异常情况检测处理。The selection will apply your feedback to the future anomaly detection processing of the same series. 已处理的点将不会重新计算。The processed points will not be recalculated. 这意味着,如果你将 Anomaly 标记为 NotAnomaly,未来我们将取消类似的异常,如果你将 NotAnomaly 点标记为 Anomaly,未来我们会将类似点检测为 AnomalyThat means if you marked an Anomaly as NotAnomaly, we will suppress similar anomalies in the future, and if you marked a NotAnomaly point as Anomaly, we will tend to detect similar points as Anomaly in the future. 如果选择 AutoDetect,则未来将忽略同一点上以前的任何反馈。If AutoDetect is chosen, any previous feedback on the same point will be ignored in the future.

为多个连续点提供反馈Provide feedback for multiple continuous points

如果希望同时为多个连续点提供异常反馈,请选择要为其添加批注的点组。If you would like to give anomaly feedback for multiple continuous points at the same time, select the group of points you want to annotate. 当你提供异常反馈时,你将看到所选的时间范围已自动填充。You will see the chosen time-range automatically filled when you provide anomaly feedback.

异常反馈菜单,其中选择了异常和特定的时间范围

若要查看个别点是否受异常反馈的影响,请在浏览时序时选择单个点。To view if an individual point is affected by your anomaly feedback, when browsing a time series, select a single point. 如果该点的异常情况检测结果已根据反馈更改,工具提示将显示“受反馈影响: true”。If its anomaly detection result has been changed by feedback, the tooltip will show Affected by feedback: true. 如果显示“受反馈影响: false”,这意味着已对此点执行异常反馈计算,但不应更改异常情况检测结果。If it shows Affected by feedback: false, this means an anomaly feedback calculation was performed for this point, but the anomaly detection result should not be changed.

工具提示显示,其中文本:“受反馈影响: true”已用红框突出显示

在以下情况下,我们不建议提供反馈:There are some situations where we do not suggest giving feedback:

  • 异常是由假日导致的。The anomaly is caused by a holiday. 建议使用预设事件来解决此类误报,这样会更精确。It's suggested to use a preset event to solve this kind of false alarm, as it will be more precise.
  • 异常是由已知数据源更改导致的。The anomaly is caused by a known data source change. 例如,当时上游系统发生更改。For example, an upstream system change happened at that time. 在这种情况下,预期会发出异常警报,因为系统不知道导致值更改的原因以及何时会再次发生类似的值更改。In this situation, it is expected to give an anomaly alert since our system didn't know what caused the value change and when similar value changes will happen again. 因此,建议不要将此类问题批注为 NotAnomalyThus we don't suggest annotating this kind of issue as NotAnomaly.

更改点Change points

有时,数据的趋势变化会影响异常情况检测结果。Sometimes the trend change of data will affect anomaly detection results. 在决定某点是否为异常点时,应将历史数据的最新窗口纳入考虑范围。When a decision is made as to whether a point is an anomaly or not, the latest window of history data will be taken into consideration. 当时序发生趋势变化时,可以标记确切的更改点,这将有助于我们以后分析异常探测器。When your time series has a trend change, you could mark the exact change point, this will help our anomaly detector in future analysis.

如下图所示,你可以选择 ChangePoint 作为反馈类型,然后从下拉列表中选择 ChangePointNotChangePointAutoDetectAs the figure below shows, you could select ChangePoint for the feedback Type, and select ChangePoint, NotChangePoint, or AutoDetect from the pull-down list.

“更改点”菜单,其中具有包含 ChangePoint、NotChangePoint 和 AutoDetect 选项的下拉列表

备注

如果数据不断变化,只需将一个点标记为 ChangePoint,因此,如果你标记了 timerange,我们将自动填充最后一个点的时间戳和时间。If your data keeps changing, you will only need to mark one point as a ChangePoint, so if you marked a timerange, we will fill the last point's timestamp and time automatically. 在这种情况下,批注只会影响 12 个点后的异常情况检测结果。In this case, your annotation will only affect anomaly detection results after 12 points.

季节性Seasonality

对于季节性数据,执行异常情况检测时,需要估算时序的周期(季节性),并将其应用于异常情况检测阶段。For seasonal data, when we perform anomaly detection, one step is to estimate the period(seasonality) of the time series, and apply it to the anomaly detection phase. 有时,很难确定精确的周期,而且周期也可能发生变化。Sometimes, it's hard to identify a precise period, and the period may also change. 定义错误的周期可能会对异常情况检测结果产生副作用。An incorrectly defined period may have side effects on your anomaly detection results. 你可以在工具提示中查找当前周期,其名称为 Min PeriodYou can find the current period from a tooltip, its name is Min Period.

工具提示被“周期”一词覆盖,并用红框红色突出显示了数字 7。

你可以提供周期反馈来修复这种异常情况检测错误。You can provide feedback for period to fix this kind of anomaly detection error. 如图所示,你可以设置一个周期值。As the figure shows, you can set a period value. 单位 interval 表示一个粒度。The unit interval means one granularity. 在这里,零间隔表示数据是非季节性的。Here zero intervals means the data is non-seasonal. 如果你想取消之前的反馈,并让管道自动检测周期,也可以选择 AutoDetectYou could also select AutoDetect if you want to cancel previous feedback and let the pipeline detect period automatically.

备注

设置周期时,无需指定时间戳或时间范围,该周期将影响从你提供反馈的那一刻起整个时序上的未来异常情况检测。When setting period you do not need to assign a timestamp or timerange, the period will affect future anomaly detections on whole timeseries from the moment you give feedback.

“自动检测”周期设置为 28 且间隔设置为 0 以指示非季节性的菜单。

提供注释反馈Provide comment feedback

还可添加注释以批注数据,并为数据提供上下文。You can also add comments to annotate and provide context to your data. 若要添加注释,请选择时间范围并添加注释文本。To add comments, select a time range and add the text for your comment.

可以设置时间范围和框以添加基于文本的注释的菜单

时序批反馈Time series batch feedback

如前文所述,使用反馈模式,你可以重新选择或删除维度值,以获取由维度筛选器定义的一批时序。As previously described, the feedback modal allows you to reselect or remove dimension values, to get a batch of time series defined by a dimension filter. 还可以通过单击左侧面板中反馈的“+”按钮,然后选择维度和维度值来打开此模式。You can also open this modal by clicking the "+" button for Feedback from the left panel, and select dimensions and dimension values.

在“反馈”一词旁用红框突出显示蓝色加号的菜单

具有 Dim1 和 Dim2 指示的两个维度的“添加反馈”菜单

如何查看反馈历史记录How to view feedback history

可通过两种方式查看反馈历史记录。There are two ways to view feedback history. 可在左侧面板中选择“反馈历史记录”按钮,随后你就将看到反馈列表模式。You can select the feedback history button from the left panel, and will see a feedback list modal. 该列表将列出你之前为单个序列或维度筛选器提供的所有反馈。It lists all the feedback you've given before either for single series or dimension filters.

反馈列表菜单

查看反馈历史记录的另一种方法是从序列进行查看。Another way to view feedback history is from a series. 每个序列的右上角显示多个按钮。You will see several buttons on the upper right corner of each series. 选择“显示反馈”按钮,该行将从显示异常点切换为显示反馈条目。Select the show feedback button, and the line will switch from showing anomaly points to showing feedback entries. 绿色标志表示更改点,蓝色点为其他反馈点。The green flag represents a change point, and the blue points are other feedback points. 你还可以选择它们,并将获得一个反馈列表模式,其中将列出对这些点提供的反馈的详细信息。You could also select them, and will get a feedback list modal that lists the details of the feedback given for this point.

反馈历史记录图

具有两个维度的反馈列表菜单

备注

任何有权访问该指标的人都可以提供反馈,因此你可能会看到其他数据源所有者提供的反馈。Anyone who has access to the metric is permitted to give feedback, so you may see feedback given by other datafeed owners. 如果你和其他人编辑了同一个点,你的反馈将覆盖以前的反馈条目。If you edit the same point as someone else, your feedback will overwrite the previous feedback entry.

后续步骤Next steps