Coveo Machine Learning Features

The Coveo™ Machine Learning (Coveo ML) service (see Coveo Machine Learning) currently offers the features described in the following sections.

  • By default, models are built for each combination of language, hubs, and tabs since these attributes normally define different type of users and use cases. Thus, out of the box, Coveo ML models do not deliver recommendations or suggestions based on user behavior in another search hub, search interface, or language.

    Query suggestions that were recommended based on your internal search interface logged events are not recommended in your external search interface.

  • When you want the relevance on one search page to influence other search interfaces for a unified experience, you can set the filterFields custom model parameter value accordingly. If the parameter value only contains the desired search page, the model will provide recommendations or suggestions based on user behavior on that specific search page even if the model is active on a search interface in another hub. Before modifying the value, it is strongly recommended to consult your Coveo Customer Success Manager (CSM) or Coveo Support for appropriate guidance. Moreover, you should test changes thoroughly in a sandbox environment before deploying in production.

For further information on Coveo ML, see the Coveo Machine Learning FAQ section.

Automatic Relevance Tuning (ART) Feature

In short, the ART feature learns what search users seek and delivers it.

RevealARTSchema1b

In more detail, ART analyzes user behavior patterns from many usage analytics search visit actions (such as query reformulation, clicked results, if a support case was submitted) to understand which clicked results and content lead to successful outcomes such as self-service success, and automatically adjusts future search results so that the best performing content always rises to the top.

ART excels with popular and ambiguous queries where users enter only one or two terms, as well as with paragraph-sized queries expressing long descriptions. ART is robust to common typographical errors and learns implicit synonyms. When your Coveo index content includes secured items, ART queries the index to ensure to only recommend items 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. Coveo Cloud organization administrators and relevance managers can configure and activate ART in just a few clicks (see Adding and Managing Coveo Machine Learning Models and Built-In Group Privileges).

By default, ART model recommendations are based on the language of the user’s query as well as the search hub and search tab (interface) in which the query was performed. One model is made per search hub/search tab/language combination.

Items are boosted only if they were clicked in the same language, hub, and interface as the current query.

You can change this default behavior by modifying the filterFields custom model parameter value with the guidance of Coveo Support.

You can use the JavaScript Search Framework Debug Panel to temporarily highlight and therefore easily identify search results promoted by ART (see Using the JavaScript Search Debug Panel).

Query Suggestions Feature

The Coveo ML Query Suggestions feature recommends significantly more relevant queries to users as they type in the search box. The original usage analytics query suggestions are limited to top queries in which the typed characters exactly match a suggested query part.

Reveal-QuerySuggestionsEx3

The Coveo ML Query Suggestions feature:

  • Identifies typed characters exact, partial, or fuzzy matches anywhere in any individual keyword appearing in any order.

  • Stems query suggestion keywords to remove duplicates.

  • Offers the most relevant recommendations by ranking query suggestions considering:

    • The number of times the query was performed.

    • The degree of matching.

    • The query performance based on the Relevance Index and Click-Through usage analytics metrics.

  • Only considers queries performed and followed by clicks on search results enough times to eliminate outliers. The number of times a query must be performed and followed by a click varies based on the number of queries by language in the model. The following table lists diverse query counts and the respective minimum number of times a specific query must be performed and followed by a click for this query to be selected has a potential suggestion.

    The minimum number of times a query must be performed and followed by a click is calculated per language, thus more popular languages will have a higher minimum number of times. You can review the minimum for each language in the model information panel (see Reviewing Coveo Machine Learning Model Information).

Query count Minimum number of times a query must be performed and followed by a click
0 0
1 000 1
40 000 2
150 000 3
400 000 4
900 000 5
1 700 000 6
  • While typing in a Coveo JavaScript search box, pressing the Tab key automatically fills the search box with the first query suggestion.

In the end, suggested queries are surprisingly tolerant to typos, and get better as your usage analytics data set size increases.

Coveo Cloud organization administrators and relevance managers can configure and activate Coveo ML Query Suggestions in a few clicks (see Adding and Managing Coveo Machine Learning Models and Built-In Group Privileges).

The JavaScript Search Omnibox can easily be configured to provide Coveo ML Query Suggestions (see Omnibox Component - enableQuerySuggestAddon). By default, the Omnibox highlights exact matches in bold, fuzzy matches in bold and italic, while a keyword completion in the search box appears in gray.

  • By default, Query Suggestions model recommendations are made based on the language of the user’s query, the search hub or the search tab (interface) in which the query is performed. Query Suggestions are only returned if the language, search hub, or search interface matches the current user context.

  • Once a Query Suggestions model is enabled, the search box provides query completion suggestions from the moment end users type their first character.

Event Recommendations Feature

Reveal-Recommentations

The Coveo ML Event Recommendations feature learns from your website user page and search navigation history to return the most likely relevant content for each user in his current session. The Recommendation service results can be included in a search page or in any web page such as in a side panel window (see Coveo Machine Learning Event Recommendations Deployment Overview).

The recommendations can be interpreted as “People who viewed this page also viewed the following pages”.

Your company offers product technical documentation and Q&A content on several public websites for customer end users, administrators, and developers. These websites are configured to send all page views to the Coveo Usage Analytics service. Your website pages include Recommended Articles side panel windows.

The recommendation algorithm is based on the co-occurrence of the events such as page views within a user session. When two events abnormally frequently co-occur within sessions, the algorithm learns that they are linked. When one event is seen, the model recommends the other.

Event Recommendations model suggestions are provided based only on the user’s language since page views are not done in a search hub or a search interface.

Dynamic Navigation Experience (DNE) Feature

Pilot Feature

The Coveo ML Dynamic Navigation Experience (DNE) feature is only available for Coveo Cloud organizations created after May 21, 2019.

FacetDne2

The Coveo ML DNE feature learns from usage analytics events to pertinently order facets and facet values according to user queries. More precisely, DNE models analyze queries and target specific user behaviors such as result clicks and facet selections to make the most relevant facets appear at the top for a given query.

Furthermore, Coveo ML DNE models also reorder facet values within a given facet to make the most popular values appear at the top. To do so, the models use the search events performed by previous users, who have selected certain facet values for a specific query.

Coveo ML DNE feature also uses its facet value ranking to boost search results. The model uses the most popular facet values ​​for a certain query and applies query ranking expressions (QREs) to boost the search results whose field values match the values of those facets.

See Deploying Dynamic Navigation Experience.

You are selling smartphones on your e-commerce website. Before enabling Coveo ML DNE, your search page, powered by the Coveo JavaScript Search Framework, displays facets in the following order when customers search for cellphone:

  • Screen size
  • Storage capacity
  • Price
  • Brand

You enable a Coveo ML DNE model. When your search interface sends a query to the Search API in order to request facets, the DNE model can now modify that query in the query pipeline. It applies insights gained from the analysis of past customer behavior and determines that users are most likely to sort search results using the Brand and the Price facets. Your search page now displays facets in the following order:

  • Brand
  • Price
  • Screen size
  • Storage capacity

Before you enabled Coveo ML DNE, the Brand facet displayed its facet values in the following order when customers searched for cellphone:

  • LG
  • Samsung
  • Apple

You enable a Coveo ML DNE model. When your search interface sends a query to the Search API in order to request facets, the DNE model can now modify that query in the query pipeline. It applies insights gained from the analysis of past customer behavior and determines that users are most likely to search for Apple and Samsung smartphones rather than for LG devices. The JavaScript Search Framework thus displays the facet values within the Brand facet in the following order:

  • Apple
  • Samsung
  • LG

Since the Coveo ML DNE model determined that customers are more likely to shop for Apple smartphones, the model modifies the user query to boost Apple smartphone result list items.