Groupwise Anomaly Detection with SQL

Overview

I’m going to demonstrate a straightforward anomaly detection procedure using a limited toolset: SQLite. Given the limited toolset, the algorithm must use nothing more than fundamental statistics (avg, variance, z-score) to identify atypical, anomalous events. We will be taking a step-by-step approach to build up the SQL query incrementally.

So, imagine you are monitoring some thing, taking multiple observations of various metrics. This could be:

• Account balances
• Heart rate and blood pressure of a patient
• CPU and memory utilization of a server

The goal is to identify when a anomaly, an outlier, an oddball occurs.

The dataset

This exact method imposes some requirements on the dataset. The dataset we’ll be using is stored in a SQLite table with the following columns:

• ts
• group_name
• metric
• value

ts is for timestamp, and it is expressed in unix time. That is, the number of seconds since January 1, 1970. This allows for easy date arithmetic and efficient date comparisons.

In this dataset, there are two groups and two metrics, with an observation recorded in the events table every five minutes. There is one row for each observation, so four rows for each observation period.

SELECT
  datetime(ts, 'unixepoch', 'localtime') as ts_local,
  ts,
  group_name,
  metric,
  value
FROM events

Top 12 rows:

If you’d like, skip the explanation; get to the query.

Essential statistics

To compute the Z-score for each point, we will need the window average and variance. We will build up to these stats using the window count mov_n, window sum mov_sum, and window sum squared mov_sum_sq.

Since SQLite lacks functions for these window operations, they will instead be calculated after performing a self-join of the table on the ts column. Here you can define the window size, in seconds, by subtracting a number of seconds from the timestamp of each point. Each point then has a unique set of observations from which to calculate the average and variance for. In this example, the window size is 10800 seconds, or 3 hours.

SELECT
  t1.ts,
  t1.group_name,
  t1.metric,
  t1.value,
  count(t2.value) AS mov_n,
  sum(t2.value) AS mov_sum,
  sum(t2.value*t2.value) AS mov_sum_sq
FROM events t1 LEFT JOIN events t2
  ON t1.group_name = t2.group_name
  AND t1.metric = t2.metric
  AND t2.ts >= t1.ts - 10800
  AND t2.ts < t1.ts
GROUP BY t1.ts,
  t1.group_name,
  t1.metric,
  t1.value

Top 16 rows:

Notice how the first few rows are missing some values. Also, mov_n, the number of observations in the moving window, is 0. This is because each observation computes the stats for a window of observations that occured before it. Starting out, there are no past events to aggregate. As we proceed, the window grows incrementally larger until a “complete” window is obtained. See the last 12 observations:

With the building blocks for average and variance, we compute them and derive the Z-score for each point.

mov_sum/mov_n AS mov_avg

( mov_sum_sq - (mov_sum*mov_sum)/mov_n ) / (mov_n - 1) AS mov_var

((value - (mov_sum/mov_n)) * (value - (mov_sum/mov_n))) /   -- (value - mov_avg)^2 /
  ((mov_sum_sq - (mov_sum*mov_sum/mov_n)) / (mov_n - 1))    -- mov_var

CASE
  WHEN
    ((value - (mov_sum/mov_n)) * (value - (mov_sum/mov_n))) /
    ((mov_sum_sq - (mov_sum*mov_sum/mov_n)) / (mov_n - 1)) > 3*3
    THEN 1
  ELSE 0
END is_anomaly

The query

All together, the query starts by performing the non-equi self-join of itself, the window size is 10800 seconds (3 hrs). From this, it calculates the essential stats, the number of observerations mov_n, moving \(\text{sum}\) mov_sum, moving \(\text{sum}^2\) mov_sum_sq. Built atop that are statements that use these stats to return the moving average mov_avg, moving variance mov_var, and moving \(z^2\) mov_z_sq. Finally, the threshold value 3 is squared and compared to mov_z_sq to determine if the current point is atypical from previous points in the moving window.

SELECT
 ts,
 group_name,
 metric,
 value,
 mov_n,
 mov_sum / mov_n AS mov_avg,
 (mov_sum_sq - (mov_sum*mov_sum/mov_n)) / (mov_n - 1) AS mov_var,
 ((value - (mov_sum/mov_n)) * (value - (mov_sum/mov_n))) / ((mov_sum_sq - (mov_sum*mov_sum/mov_n)) / (mov_n - 1)) AS mov_z_sq,
 CASE
    WHEN
     ((value - (mov_sum/mov_n)) * (value - (mov_sum/mov_n))) /
      ((mov_sum_sq - (mov_sum*mov_sum/mov_n)) / (mov_n - 1)) > 3*3 THEN 1
    ELSE 0
 END is_anomaly
FROM (
 SELECT
     t1.ts,
     t1.group_name,
     t1.metric,
     t1.value,
     count(t2.value) AS mov_n,
     sum(t2.value) AS mov_sum,
     sum(t2.value*t2.value) AS mov_sum_sq
 FROM events t1 LEFT JOIN events t2
  ON t1.group_name = t2.group_name
 AND t1.metric = t2.metric
 AND t2.ts >= t1.ts - 10800
 AND t2.ts < t1.ts
 GROUP BY t1.ts,
          t1.group_name,
          t1.metric,
          t1.value
) as t

Top 12 rows:

View the results

Create a view of the query above to easily return anomalous records.

SELECT *
FROM v_anomaly
WHERE is_anomaly=1

Visualized

Now that our original dataset is augmented with valuable statistics and a classification, we can visualize the anomaly detection process. For this, we will need to switch to R and import the packages RSQLite and ggplot2. Also, since we’re now in R, we have a sqrt() function; sweet. Note that, while R is among the best tools for analytics, the anomaly detection process is being done entirely in SQL; we’re just using R for the viz.

In the plot below, each group is plotted in a pane. Each group metric is plotted as a colored line. The window average is plotted as a black dashed line within each group metric. The light gray lines illustrate the define threshold which determines the classification. Notice how the threshold increases after an anomaly occurs, then drops back down after the window size (3 hrs) elapses.

library(RSQLite)
library(ggplot2)
library(data.table)

### Connect to and retrieve teh data

db <- dbConnect(RSQLite::SQLite(), db_name)

df_anom <- RSQLite::dbGetQuery(db, 'SELECT * FROM v_anomaly')

RSQLite::dbDisconnect(db)

### Quick prep

df_anom <- as.data.table(df_anom) # makes working with dataframes nicer.

df_anom[is.na(mov_avg), mov_avg := value]
df_anom[is.na(mov_var), mov_var := 0]

z_thresh <- 3 # define the threshold to visualize it. The classification has already been made.

df_anom[, ':=' (thresh_high = mov_avg + (sqrt(mov_var) * z_thresh),
                thresh_low = mov_avg - (sqrt(mov_var) * z_thresh),
                is_anomaly = as.logical(is_anomaly))]

### Visualize

ggplot(df_anom, aes(x=as.POSIXct(ts, origin='1970-01-01'), y=value)) +
        geom_line(aes(group=metric, color=metric)) +
  
        geom_line(aes(group=metric, y=mov_avg), color='black', alpha=0.8, linetype='dashed') + 
        geom_line(aes(group=metric, y=thresh_high), color='gray', alpha=0.5) + 
        geom_line(aes(group=metric, y=thresh_low), color='gray', alpha=0.5) +
        geom_point(data=df_anom[is_anomaly==TRUE,], color='red', shape='O', size=5) +
  
        facet_grid(rows=vars(group_name)) +
        scale_x_datetime(date_labels='%H:%M',
                        breaks=unique(df_anom[ts %% 3600 == 0,]$ts) %>% as.POSIXct(origin='1970-01-01')) +
        scale_colour_manual(values=c('blue', 'purple')) +
        labs(x=NULL, y=NULL) +
        theme_bw() +
        ylim(0, NA) +
        theme(axis.text.x = element_text(angle = 90),
              legend.title = element_blank(),
              panel.grid = element_blank())

Another Approach

The above approach implements a non-equi self-join on the ts column with a window size defined in seconds. I felt this was the most suitable way to describe this method of anomaly detection. However, a faster query using windowing functions is shown below. Here, the window size is not defined by time, but by the number of preceding points, so it assumes periodic observations within the dataset. The clause ROWS BETWEEN 36 PRECEDING AND 1 PRECEDING defines the window size to be 35 preceding observations, not including the current observation. At 5 minute intervals, this equates to a window size of \(35\ points*5\ minutes*60\ seconds=10500\ seconds\). The query above has a window size of 10800 seconds, but the upper bound is non-inclusive; the current point is not considered. We could have made the upper bound <= 10500 seconds and achieve the same results. Long story short, the query below will produce the same results as the query above, but much faster, though with more assumptions and less flexibility.

SELECT
  ts,
  group_name,
  metric,
  value,
  avg(value) OVER w as mov_avg,
  case
    when ((value - (avg(value) OVER w)) * (value - (avg(value) OVER w))) /
            (( (sum(value * value) OVER w) - ((sum(value) OVER w) * (avg(value) OVER w))) /
                ( (count(*) OVER w) - 1))
            > 3*3 THEN 1
    ELSE 0
  END as is_outlier
FROM events
WINDOW w AS (PARTITION BY group_name, metric ORDER BY ts ROWS BETWEEN 36 PRECEDING AND 1 PRECEDING)
ORDER by ts asc

Parting words

While the methods are fairly simple, this process will not detect outliers that occur within the defined threshold. Tools such as R and Python offer libraries that are better suited for more advanced anomaly detection tasks. However, I hope the approach described in this post serves to demystify the fundamental aspects of anomaly detection by using a simple toolset and elementary statistics.

Until next time,
Donald