About Automatic Relevance Tuning (ART)
About Automatic Relevance Tuning (ART)
Automatic Relevance Tuning (ART) models learn from user search behavior to rank the most relevant results higher.
ART analyzes user behavior patterns from many Coveo Analytics search visit actions, such as query reformulation, clicked results,[1] and whether a support case was submitted. It uses this behavior to understand which clicked results and content lead to successful outcomes, such as self-service success. ART automatically adjusts future search results, so that the best-performing content always rises to the top.
The following example illustrates how ART learns from search behavior:
A consumer electronics retailer has many online community visitors seeking help configuring a popular media player console.
Using queries such as media console help, many of them found a particular article to be very helpful, and it proved successful in preventing ticket submissions.
Coveo ML ART learns and automatically boosts the relevance of this article for new visitors running similar queries.
After the company releases a new media console model that quickly becomes very popular, visitors searching for media console help find an article on the new model to be more helpful.
ART automatically learns this new trend and updates its recommendations.
ART excels with popular and ambiguous queries, in which users only enter one or two terms, as well as with paragraph-sized queries expressing long descriptions. ART can handle common typographical errors, and it learns implicit synonyms. When your Coveo index includes content from sources that index permissions, ART queries the index to ensure that it only recommends items that the user performing the query is allowed to access.
In practice, ART boosts the ranking weight of recommended items so that they appear among the top search results. ART can also take advantage of the atomic-result-badge component to highlight items that have been promoted in the results list.
Members with the required privileges can create, manage, and activate an ART model in just a few clicks.
|
Notes
By default, ART model recommendations are based on the language of the user’s query as well as the search interface in which the query is performed. ART models build a submodel for each language, and then apply filters on these submodels for each search hub and search tab to better tailor the provided recommendations to the user’s context. Items are boosted only if they were clicked in the same search interface as the current query. For more information, see Submodels. |
|
Leading practice
You can use the Relevance Inspector to identify search results promoted by ART. |
Submodels
A submodel is a smaller, specialized model that operates as part of a larger model.
An ART model automatically creates submodels to handle different combinations of language, search hub, and tabs.
A model learns separately from search visits made in interfaces offered in different languages, since the keywords used for similar searches often vary by language. Different search hubs or interfaces may also serve different purposes, meaning that users may expect different results for the same query depending on where they search. Submodels account for these differences by filtering out recommendations that don’t match the current search hub and interface combination, helping ensure that recommendations remain relevant to the context in which the search is performed.
For example:
-
If your search hub supports multiple languages, each language will have its own submodel to ensure that recommendations are relevant to users searching in that language.
-
If your model is used in multiple search hubs that all share the same language, each search hub will have its own submodel to ensure that recommendations are relevant to the context of that specific search hub.
-
If a search hub contains multiple tabs, each tab will have its own submodel to account for the different search contexts represented by those tabs.
-
If your model is used across multiple search hubs, languages, and tabs, each combination of search hub, language, and tab will have its own submodel. For example, if a search hub has English and French versions and contains Products and Documentation tabs, separate submodels are created for each combination: English-Products, English-Documentation, French-Products, and French-Documentation.
You can review the number of recommended items for each submodel in the model’s information tab.
|
|
Notes
|
Your Community search page and content are available in several languages. Most search traffic occurs on the English version of the page, while only 4% of queries are made on the Greek version.
A user on the Greek search page searches for DFT-400, a product name that’s the same across all supported languages.
Because the Greek submodel learns only from search activity on the Greek search page, an ART model can recommend relevant Greek items for DFT-400. Without separate language submodels, the much larger volume of English search activity could cause the ART model to recommend English items instead, even though those items are outside the scope of the Greek search page.
|
|
If you want user behavior from one search interface (search hub) to influence recommendations in other search interfaces for a unified experience, you can configure the Suggestion filters advanced model setting accordingly. If the parameter value specifies only a particular search hub, the model uses behavior from that search hub to generate recommendations or suggestions, including when the model is used in a different search hub. Before modifying the Suggestion filters advanced option, we strongly recommend that you consult your Account Manager or Coveo Support for appropriate guidance. Moreover, you should test any changes in a sandbox environment before deploying in production. |
1. ART models also learn from actions performed by users within a given search result, such as clicking a search result Quick view or attaching a result to a Case in a Coveo Insight Panel.