* Suggesting "H5N1" when users search for "bird flu" in text
* Identifying the merchant that is the "common point of compromise" from the transaction history of credit card owners reporting loss
* Suggesting keywords relating to stock symbol $ATI for an automated news classifier
* Spotting the fraudulent doctor who is diagnosing more than his fair share of whiplash injuries
* Spotting the tire manufacturer who has a disproportionate number of blow-outs
In all these cases the terms being selected are not simply the most popular terms in a set.
They are the terms that have undergone a significant change in popularity measured between a _foreground_ and _background_ set.
If the term "H5N1" only exists in 5 documents in a 10 million document index and yet is found in 4 of the 100 documents that make up a user's search results
that is significant and probably very relevant to their search. 5/10,000,000 vs 4/100 is a big swing in frequency.
==== Single-set analysis
In the simplest case, the _foreground_ set of interest is the search results matched by a query and the _background_
set used for statistical comparisons is the index or indices from which the results were gathered.
When querying an index of all crimes from all police forces, what these results show is that the British Transport Police force
stand out as a force dealing with a disproportionately large number of bicycle thefts. Ordinarily, bicycle thefts represent only 1% of crimes (66799/5064554)
but for the British Transport Police, who handle crime on railways and stations, 7% of crimes (3640/47347) is
a bike theft. This is a significant seven-fold increase in frequency and so this anomaly was highlighted as the top crime type.
The problem with using a query to spot anomalies is it only gives us one subset to use for comparisons.
To discover all the other police forces' anomalies we would have to repeat the query for each of the different forces.
This can be a tedious way to look for unusual patterns in an index
==== Multi-set analysis
A simpler way to perform analysis across multiple categories is to use a parent-level aggregation to segment the data ready for analysis.
Example using a parent aggregation for segmentation:
This example uses the `geohash_grid` aggregation to create result buckets that represent geographic areas, and inside each
bucket we can identify anomalous levels of a crime type in these tightly-focused areas e.g.
* Airports exhibit unusual numbers of weapon confiscations
* Universities show uplifts of bicycle thefts
At a higher geohash_grid zoom-level with larger coverage areas we would start to see where an entire police-force may be
tackling an unusual volume of a particular crime type.
Obviously a time-based top-level segmentation would help identify current trends for each point in time
where a simple `terms` aggregation would typically show the very popular "constants" that persist across all time slots.
.How are the scores calculated?
**********************************
The numbers returned for scores are primarily intended for ranking different suggestions sensibly rather than something easily understood by end users.
The scores are derived from the doc frequencies in _foreground_ and _background_ sets. The _absolute_ change in popularity (foregroundPercent - backgroundPercent) would favour
common terms whereas the _relative_ change in popularity (foregroundPercent/ backgroundPercent) would favour rare terms.
Rare vs common is essentially a precision vs recall balance and so the absolute and relative changes are multiplied to provide a sweet spot between precision and recall.
The significant_terms aggregation can be used effectively on tokenized free-text fields to suggest:
* keywords for refining end-user searches
* keywords for use in percolator queries
WARNING: Picking a free-text field as the subject of a significant terms analysis can be expensive! It will attempt
to load every unique word into RAM. It is recommended to only use this on smaller indices.
.Use the _"like this but not this"_ pattern
**********************************
You can spot mis-categorized content by first searching a structured field e.g. `category:adultMovie` and use significant_terms on the
free-text "movie_description" field. Take the suggested words (I'll leave them to your imagination) and then search for all movies NOT marked as category:adultMovie but containing these keywords.
You now have a ranked list of badly-categorized movies that you should reclassify or at least remove from the "familyFriendly" category.
The significance score from each term can also provide a useful `boost` setting to sort matches.
Using the `minimum_should_match` setting of the `terms` query with the keywords will help control the balance of precision/recall in the result set i.e
a high setting would have a small number of relevant results packed full of keywords and a setting of "1" would produce a more exhaustive results set with all documents containing _any_ keyword.
**********************************
[TIP]
============
.Show significant_terms in context
Free-text significant_terms are much more easily understood when viewed in context. Take the results of `significant_terms` suggestions from a
free-text field and use them in a `terms` query on the same field with a `highlight` clause to present users with example snippets of documents. When the terms
are presented unstemmed, highlighted, with the right case, in the right order and with some context, their significance/meaning is more readily apparent.
The above examples show how to select the _foreground_ set for analysis using a query or parent aggregation to filter but currently there is no means of specifying
a _background_ set other than the index from which all results are ultimately drawn. Sometimes it may prove useful to use a different
background set as the basis for comparisons e.g. to first select the tweets for the TV show "XFactor" and then look
for significant terms in a subset of that content which is from this week.
If there is the equivalent of a `match_all` query or no query criteria providing a subset of the index the significant_terms aggregation should not be used as the
top-most aggregation - in this scenario the _foreground_ set is exactly the same as the _background_ set and
so there is no difference in document frequencies to observe and from which to make sensible suggestions.
Another consideration is that the significant_terms aggregation produces many candidate results at shard level
that are only later pruned on the reducing node once all statistics from all shards are merged. As a result,
it can be inefficient and costly in terms of RAM to embed large child aggregations under a significant_terms
aggregation that later discards many candidate terms. It is advisable in these cases to perform two searches - the first to provide a rationalized list of
significant_terms and then add this shortlist of terms to a second query to go back and fetch the required child aggregations.
The counts of how many documents contain a term provided in results are based on summing the samples returned from each shard and
as such may be:
* low if certain shards did not provide figures for a given term in their top sample
* high when considering the background frequency as it may count occurrences found in deleted documents
Like most design decisions, this is the basis of a trade-off in which we have chosen to provide fast performance at the cost of some (typically small) inaccuracies.
However, the `size` and `shard size` settings covered in the next section provide tools to help control the accuracy levels.
Terms that score highly will be collected on a shard level and merged with the terms collected from other shards in a second step. However, the shard does not have the information about the global term frequencies available. The decision if a term is added to a candidate list depends only on the score computed on the shard using local shard frequencies, not the global frequencies of the word. The `min_doc_count` criterion is only applied after merging local terms statistics of all shards. In a way the decision to add the term as a candidate is made without being very _certain_ about if the term will actually reach the required `min_doc_count`. This might cause many (globally) high frequent terms to be missing in the final result if low frequent but high scoring terms populated the candidate lists. To avoid this, the `shard_size` parameter can be increased to allow more candidate terms on the shards. However, this increases memory consumption and network traffic.
The parameter `shard_min_doc_count` regulates the _certainty_ a shard has if the term should actually be added to the candidate list or not with respect to the `min_doc_count`. Terms will only be considered if their local shard frequency within the set is higher than the `shard_min_doc_count`. If your dictionary contains many low frequent words and you are not interested in these (for example misspellings), then you can set the `shard_min_doc_count` parameter to filter out candidate terms on a shard level that will with a resonable certainty not reach the required `min_doc_count` even after merging the local frequencies. `shard_min_doc_count` is set to `1` per default and has no effect unless you explicitly set it.
Setting `shard_min_doc_count` too high will cause significant candidate terms to be filtered out on a shard level. This value should be set much lower than `min_doc_count/#shards`.
