Bridging Search and Recommendation in Generative Retrieval

Marco

Ali

Enrico

Hugues

Gustavo

Retrieval task

For a given textual query retrieve K items out of the entire corpus.

“workout music”

query

Item collection

Model

1

2

K

…

Recommendation task

​

​

Given the previous history of item interactions from a user retrieve K items out of the entire corpus.

Item collection

Model

1

2

K

…

user

“item_1→ item_3→ item_4”

Dense vs generative approaches

Dense models: bi-encoder (search)

bi-encoder

Item collection

Items ↓

→ Embedding dim (e.g. 512)

​

Can be done offline

Dense models: bi-encoder (search)

bi-encoder

Item collection

Items ↓

→ Embedding dim (e.g. 512)

​

“workout music”

bi-encoder

query

Embedding dim (e.g. 512) →

online/cache

Dense models: bi-encoder (search)

bi-encoder

Item collection

Items ↓

→ Embedding dim (e.g. 512)

​

“workout music”

bi-encoder

query

Find closest items

Embedding dim (e.g. 512) →

Dense models: bi-encoder (search)/two-tower (rec.)

bi-encoder

two-tower

Item collection

Items ↓

→ Embedding dim (e.g. 512)

​

“workout music”

“Gustavo Penha”

(previous interactions: item1, item2, etc..)

Items ↓

→ Embedding dim (e.g. 512)

​

bi-encoder

two-tower

query

user

Find closest items

Find closest items

Embedding dim (e.g. 512) →

Embedding dim (e.g. 512) →

​

Recommendation pipeline can be quite similar

Generative retrieval and recommendation

“workout music”

Gen retrieval

Gen recs

query

Items ( |item collection| )

bi-encoder

two-tower

Item collection

Items ↓

→ Embedding dim (e.g. 512)

​

Items ↓

→ Embedding dim (e.g. 512)

​

Items ( |item collection| )

user

Not needed with generative retrieval

“Gustavo Penha”

(previous interactions: item1, item2, etc..)

“Extreme classification problem”

Generative retrieval and recommendation

“workout music”

Gen retrieval

Gen recs

query

Items ( |item collection| )

bi-encoder

two-tower

Item collection

Items ↓

→ Embedding dim (e.g. 512)

​

Items ↓

→ Embedding dim (e.g. 512)

​

Items ( |item collection| )

user

This work:

joint learning

“Gustavo Penha”

(previous interactions: item1, item2, etc..)

Joint generative model

Generative retrieval

LM

“bossa nova music”

query

Tokenizer

item₁

relevant item

ID strategy

labels

<item_1>

bossa nova music

(4 tokens)

(1 token)

In this work we used ‘atomic IDs’, so each item has its own token

new_vocab_size = vocab_size + num_items

Generative recommendation

Tokenizer

LM

item₃→ item₄→ item₂

user

ID strategy

labels

<item_3> <item_4> <item_2>

<item_1>

(3 tokens)

(1 token)

item₁

relevant item

‘atomic ids’

Generative retrieval and recommendation

LM

“bossa nova music”

query

Tokenizer

item₁

relevant item

item₃→ item₄→ item₂

user

ID strategy

labels

labels

<item_3> <item_4> <item_2>

<item_1>

<item_1>

bossa nova music

(4 tokens)

(3 tokens)

(1 token)

(1 token)

item₁

relevant item

Joint model: same LM and multi-task learning

But why would joint training improve effectiveness?

“Multi-task learning increases effectiveness over single-task models if tasks are similar”

But what does it mean for two tasks to be similar?

Let’s try to be more specific for search and recommendation…

Hypothesis 1 Regularization of item’s popularity estimation.

Hypothesis 2 Regularization of item’s latent representations.

Let’s see some examples to motivate the hypotheses next…

H1 Regularization of item’s popularity estimation

H1 Regularization of item’s popularity estimation

What if item₃ should be retrieved more than what the search train data shows ?

Due to distribution shifts, incomplete sample of dataset, problems with current search stack, retrievability bias, or other reasons …

H1 Regularization of item’s popularity estimation

The distribution from the recommendation data might regularize the output distribution and fix this gap !

Search data

Test query: q₁

Rel: item₃, item₅, item₁

H2 Regularization of item’s latent representations

​

​

Let’s say for this test query the item₅ is not very close to the query q₁ using the models learned representation, so it would not rank it highly.

Search data

Recommendation data

Test query: q₁

Rel: item₃, item₅, item₁

User1: item₃→ item₅→ item₁

User2: item₃→item₉ → item₅

User3: item₅→ item₃

H2 Regularization of item’s latent representations

​

​

But if we have recommendation data where users often interact with both item₃ and item₅ we can fix this!

Gen_S

Gen_S+R

H2 Regularization of item’s latent representations

​

​

The joint model would correctly place item₅ close to q₁ , as it would be close to item₃!

Content-based and collaborative-filtering information

Recommendation ‘similar due to co-interaction’

​

• Users that interact with different items can indicate how they are related in other aspects besides their content.

Search ‘similar due to semantics’

​

• Broad queries that are relevant to different items can indicate how the content of two items are related.

​

​

​

“adventure-filled audiobooks”

is relevant to

co-occurs in user profile with

Experiments

Simulated datasets

💡Let’s control variables in simulated search and recommendation datasets to train and test generative models.

​

​

H1: Control the divergence between popularity distributions.�

H2: Control co-occurence of items between queries and users.

Let’s look closer into one simulated dataset …

Simulated dataset for H2

S→R

​

Five clusters of items C0, C1, C2, C3, and C4.

​

Users only interact with items from a single cluster in the recommendation data.

​

Simulated dataset for H2

S→R

​

Five clusters of items C0, C1, C2, C3, and C4.

​

Users only interact with items from a single cluster in the recommendation data.

�In the search data, queries related to items from a single cluster match according to the variable we are controlling.

​

Queries for which

the item is relevant to

Results for SIM2

Gen_R

Effectiveness

Control variable →

0%

25%

50%

75%

100%

If queries are randomly distributed across item clusters (<50%) they just add noise to the task-specific model

Gen_R+S

Results for SIM2

Gen_R

Gen_R+S

Effectiveness

Control variable →

With more queries matching inside each cluster we have effectiveness improvements!

0%

25%

50%

75%

100%

Results for SIM2

The search data helps to learn better representations, which are aligned with the clusters used in this simulation

Real-world datasets

ML: Based on MovieLens 25M. Genres and tags as queries. Ratings to movies from ML25M.

​

MPD: Based on Million Playlist Dataset. Tracks inside a playlist used as user interactions and playlist titles as queries.

​

Podcasts: Based on a sample of Spotify search logs (focused on broad queries) and sequential interactions with podcast shows.

movies

music

podcasts

Effectiveness (R@30) for head entities of joint model vs task-specific models

Analyses of model predictions (Tables 6 and 7 of paper)

H1 Popularity

​

The distribution of the predicted entities did not change much when comparing the task-specific vs joint model:

-1.12 % on average in popularity increase across the 3 datasets

H2 Latent representations

​

The co-occurrences statistics changed significantly when comparing the task-specific vs joint model:

204.63 % on average in co-occurrence increase across the 3 datasets

This indicates that H2 (regularization of item’s representation) might have a bigger impact on the joint model than H1 (regularization of item’s popularity).

Thank you