Queries-for-queryRecommendations

Table of Contents

Data tips
Boosting recommendations
Tuning tips
The similar queries collection
The query pipeline

Queries-for-query recommendations are queries performed by other users who also performed the current query. For example, when a user searches for "madonna", your app may also suggest searches for "evita", "a league of their own", "lady gaga", and so on.

The "Suggested Searches" shown below are one example of queries-for-query recommendations:

Queries-for-query examples

To enable queries-for-query recommendations, use the Query-to-Query Session-Based Similarity , which is session-based.

The job generates two sets of recommendations based on the two approaches described below, then merges and de-duplicates them to present unique query-recommender pairs.

Similar queries based on documents clicked	Similar queries based on co-occurrence in sessions
Queries are considered for recommendation if two queries have similar sets of document IDs clicked according to the signals data. This is directly implemented from the similar queries portion of the Synonym Detection job. This approach can work on both raw and aggregated signals.	Queries are considered for recommendation if two queries have co-occurred in the same session based on the assumption that users search for similar items in a single search session (this may or may not hold true depending on the data). This approach, based on session/user IDs, needs raw signals to work.

Similar queries based on documents clicked

Similar queries based on co-occurrence in sessions

Queries are considered for recommendation if two queries have similar sets of document IDs clicked according to the signals data. This is directly implemented from the similar queries portion of the Synonym Detection job.

This approach can work on both raw and aggregated signals.

Queries are considered for recommendation if two queries have co-occurred in the same session based on the assumption that users search for similar items in a single search session (this may or may not hold true depending on the data).

This approach, based on session/user IDs, needs raw signals to work.

Query-to-Query Session-Based Similarity job dataflow

A default Query-to-Query Session-Based Similarity job (COLLECTION_NAME_query_recs) and a dedicated collection and pipeline are created when you enable recommendations.

At a minimum, you must configure these:

an ID for this job
the input collection containing the signals data, usually COLLECTION_NAME_signals
the data format, usually solr
the query field name, usually query_s
the document ID field name, usually doc_id_s
optionally, the user session ID or user ID field name

If this field is not specified, then the job generates click-based recommendations only, without session-based recommendations.

Data tips

Running the job on other types of data than signals is not recommended and may yield unexpected results.
To get about 90% query coverage with the query pipeline, we recommend a raw signals dataset of about ~170k unique queries. More signals will generally improve coverage.
On a raw signal dataset of about 3 million records, the job finishes execution in about 7-8 minutes on two executor pods with one CPU and about 3G of memory each. Your performance may vary depending on your configuration.

Boosting recommendations

Generally if a query and recommendation has some token overlap, then they’re closely related and we want to highlight these. Therefore, query-recommendation pair similarity scores can be boosted based on token overlap. This overlap is calculated in terms of the number or fractions of tokens that overlap.

For example, consider the pair (“a red polo shirt”, “red polo”). If the minimum match parameter is set to 1, then there should be 1 token in common. For this example there is 1 token in common and therefore it is boosted. If it is set to 0.5, then at least half of the tokens from the shorter string (in terms of space separated tokens) should match. Here, the shorter string is “red polo” which is 2 tokens long. Therefore, to satisfy the boosting requirement, at least 1 token should match.

Tuning tips

These tuning tips pertain to the advanced Model Tuning Parameters:

Special characters to be filtered out. Special characters can cause problems with matching queries and are therefore removed in the job.

Only the characters are removed, not the queries, so a query like ps3$ becomes ps3.
Query similarity threshold. This is for use by the similar queries portion of the job and is the same as that used in the Synonym Detection job.
Boost on token overlap. This enables or disables boosting of query recommendation pairs where all or some tokens match. How much match is required to boost can be configured using the next parameter.

For example, if this is enabled, then a query-recommendation pair like (playstation 3, playstation console) is boosted with a similarity score of 1, provided the minimum match is set to 1 token or 0.5.
Minimum match for token overlap. Similar to the mm param in Solr, this defines the number/fraction of tokens that should overlap if boosting is enabled. Queries and recommendations are split by “ “ (space) and each part is considered a token. If using a less-than sign (<), it must be escaped using a backslash.

The value can be an integer, such as 1, in which case that many tokens should match. So in the previous example, pair is boosted because the term “playstation” is common to both and the mm value is set to 1.

The value can also be a fraction, in which case that fraction of the shorter member of the query and recommendation pair should match. For example, if the value is set to 0.5 and query is 4 tokens long and recommendation is 6 tokens long then there should be at least 2 common tokens between query and recommendation.

Here the stopwords specified in the list of stopwords are ignored while calculating the overlap.
Query clicks threshold. The minimum number of clicked documents needed for comparing queries.
Minimum query-recommendation pair occurrence count. Minimum limit for the number of times a query-recommendation pair needs to be generated to make it to the final similar query recommendation list. Default is set to 2. Higher value will improve quality but reduce coverage.

The similar queries collection

The following fields are stored in the COLLECTION_NAME_queries_query_recs collection:

query_t
recommendation_t
similarity_d, the similarity score
source_s, the approach that generated this pair, one of the following: SessionBased or ClickedDocumentBased
query_count_l, the number of times the query occurred in signals
recommendation_count_l, the number of times recommendations occurred in signals
pair_count_l, the number of instances of the pair generated in the final recommendations using either of the approaches
type_s, always set to similar_queries

The query pipeline

When you enable recommendations, a default query pipeline, COLLECTION_NAME_queries_query_recs. is created.