SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb

arXiv:2610.07810 · cs.LG · Submitted 2026-10-06 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb".

Jane: The gist The SIFT model replaces hand-engineered features with transformer-generated user-sequence embeddings in a multi-task framework to personalize filter ranking at Airbnb.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper now called "SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb." It’s about how they replaced those old hand-engineered features with something learned directly from the user's raw behavior.

Jane: Right, so instead of giving the system pre-calculated summaries of what a guest likes, SIFT uses a transformer to understand the sequence of events—the raw clicks and searches—to figure out what the guest wants right now. That sounds like it tackles that feature brittleness problem they mentioned earlier.

Lu: Exactly. The core idea is moving from those aggregated features that collapse everything into a single number to learning preferences directly from the sequence of events themselves, which preserves order and recency >

Meng: But how do you actually build this representation if you don't have these pre-aggregated features? What's the actual mechanism for turning those raw events into something useful?

Lalam: The paper says they use a journey encoder, which consumes the raw event history as a single chronological sequence >

Tom: A sequence of events. So it takes everything—every search, every view—and feeds it into this encoder to create one contextual user embedding, e u >

Jane: And that embedding is what they say is taken from the final position of that sequence, which they compute offline on a daily cadence so serving stays fast >

Lu: That's key because it means the representation is computed ahead of time rather than at request time to keep latency low >

Meng: Okay, so that sounds like you're doing heavy computation beforehand. What about the tasks? They aren't just building one embedding, right?

Tom: No, they use this unified representation to feed three different prediction tasks simultaneously in a multi-task learning framework >

Jane: That’s where it gets interesting because they have these three specific heads: one for booking probability, one for amenity engagement, and another set for ordinal values like the number of bedrooms or bathrooms >

Lu: Specifically, the booking head uses position-discount sample weights w i equal to one plus one over log of pos i plus one to upweight bookings that are more plausibly attributable to the filter itself > <ref:2610.07810#pg1>

Meng: That weighting scheme sounds like they're trying to solve the credit assignment problem for those booking predictions, which is a common headache in these types of systems >

Tom: Right, and they also have this amenity engagement head predicting P (E f equal one u q) for each filter f independently using a multi-label sigmoid layer and summing per-label BCE losses > <ref:2610.07810#pg1>

Jane: And then for those numeric capacity filters, they aren't modeling them as just classes, they use an ordinal decomposition to predict a threshold like "at least two bedrooms" instead of just saying it's three bedrooms >

Lu: They achieve that ordinal decomposition by training k k minus one binary sub-classifiers for each filter k using the Frank and Hall ordinal decomposition > <ref:2610.07810#pg1>

Title and authors: Meng: So they are tackling the numeric filters differently than the binary ones, which makes sense since capacity isn't just on or off; it has a range >

Tom: And that’s how they handle unmodeled numeric capacity filters by folding them into these multi-task heads to get them into the system >

Jane: The training setup is also pretty specific too. They use a joint multi-task loss L equal to lambda B L B plus lambda E L E plus sigma k lambda O k L O k >

Lu: With the weights set as lambda B equal to one point zero, and then a much larger weight for amenity engagement at ten point zero, and one point zero for each numeric filter head > <ref:2610.07810#pg1>

Meng: The training data construction is also interesting because they use a rolling three hundred sixty-five-day window of daily generated user journeys, ending in a one day snapshot with seven day forward-attributed booking labels >

Tom: And to make sure they aren't cheating by looking into the future, they apply a forward-attribution scheme scanning forward from the retained event within a seven day window to attribute bookings plausibly to the filter >

Jane: That attribution method sounds like it’s designed specifically to avoid session-level label leakage, which is a major issue when you have time-based data >

Lu: The empirical results show offline evaluation improved booking PR AUC by plus fifty-one point nine percent and amenity engagement PR AUC by plus sixty-two point eight percent over the production MLP baseline >

Meng: Those are some big numbers for the offline stuff, but what about the online impact? Does this actually make a difference when millions of people are using it?

Tom: The online A/B testing showed a substantial lift in engagement with recommended filters by plus twenty point zero percent, overall filter usage among searchers by plus zero point seven two percent, and usage of newly supported filters like bedroom, bathroom, and bed filters >

Jane: And they even extended it to a novel hotel-intent filter, which drove a plus three point eight percent lift in uncancelled hotel bookings and a plus zero point seven six percent lift in overall marketplace bookings online > <ref:2610.07810#pg1>

Lu: The real strength here is the extensibility; they validate that SIFT can generalize intent signals to new surfaces, like hotel promotions, just by extending the multi-label engagement head with a new label >

Meng: That means you don't need a whole new feature pipeline every time you want to support a different kind of filter dimension; you just add a head >

Tom: It’s all about that unified representation feeding multiple prediction tasks rather than building separate systems for each one >

Jane: So, if we put it all together, this SIFT: Search Intent-to-Filter Transformer paper shows they've successfully replaced brittle manual features with a model that learns directly from raw sequences to personalize filter ranking across three different prediction goals >

Title and authors: Lu: It moves the focus from aggregate features which suffer from temporal blindness and ordinal collapse to a sequence model that preserves order and allows for fine-grained intent modeling >

Meng: From my side, the main thing I'm watching is how robust this representation is when it hits real traffic at scale, especially given that the training data construction relies on a specific three hundred sixty-five-day window >

Tom: Yeah, and we still have that limitation they admit in section nine—the system struggles with same-day sessions because their daily-refreshed embeddings can't keep up with the speed of those interactions >

Jane: That’s a fair point. So, while the results are strong for historical data patterns, it might not capture the immediate intent from very short user sessions >

Lu: The authors plan to pursue more frequent updates for that specific segment, so they see it as a clear next step in improving the system >

Meng: It’s good they acknowledged that limitation upfront; deploying something that fails on real-time needs is a bigger problem than having a great offline score >

Tom: So, to wrap up, the SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb paper successfully demonstrates how sequence modeling can outperform hand-engineered features by unifying guest preferences into one representation that powers booking and amenity prediction tasks >

Jane: It shows the power of learning directly from raw behavioral sequences to lift conversion rates through better filter ranking, even across different types of filters like numeric capacity ones >

Lu: The shift from aggregate features to sequence modeling is what unlocks the ability to model temporal drift and ordinal values in a way that was previously impossible with traditional methods >

Meng: For deployment, it’s a solid architecture because the inference is decoupled, meaning the embedding materialization doesn't block the request path at all, aiming for sub-10ms latency > <ref:2610.07810#pg1>

Tom: It's definitely a solid architecture. We’ve seen offline improvements of plus fifty-one point nine percent on booking PR AUC and plus sixty-two point eight percent on amenity engagement PR AUC compared to the MLP baseline >

Jane: And online, the A/B tests confirm that this translates into real user actions, with a twenty percent lift in engagement with recommended filters and a boost in using specific filters like bathroom or bedroom filters >

Lu: The generalization to hotel intent is another point; the system can handle new surfaces just by adding a new head without needing a complete overhaul of the feature engineering pipeline >

Meng: So, it’s less about building more pipelines and more about building smarter heads on top of a single representation >

Tom: That’s exactly what they’re doing. They're using the same learned user embedding to score everything, which is what makes it scalable for millions of guests >

Jane: So that's the picture with SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb >

The paper's summary: Tom: So, SIFT is taking all those raw user events—the sequence of clicks and searches—and instead of summarizing them into some static profile for a guest, it’s using a transformer to learn what the guest actually wants at that exact moment >

Jane: Right. The core idea is moving away from those old "hand-engineered" features that just sit there in a database, and instead building this unified representation directly from the raw sequence of user behavior >

Tom: Yeah, so they call it a journey encoder, which basically consumes the entire event history as one long chronological string >

Jane: And they take that long string and compress it into a single contextual embedding vector for each guest >

Tom: It’s computed offline on a daily batch job to keep the serving fast, which is smart because you don't want to do heavy transformer inference every time someone refreshes their search results >

Jane: The real genius, though, is how they use that one representation to feed three separate prediction tasks at once >

Tom: Exactly. They have this multi-task learning framework where the same embedding feeds a booking head, an amenity engagement head, and even a set of heads for those numeric capacity filters >

Jane: That’s what makes it powerful because it learns what a guest likes for different things simultaneously rather than optimizing one thing in isolation >

Tom: The booking head uses position-discount sample weights to make sure the bookings it predicts are actually plausible based on the filter being applied >

Jane: And that ordinal decomposition they use for filters like bedrooms isn't just saying "yes" or "no," it's predicting a threshold, like "at least two bedrooms" >

Tom: That’s how they handle those numeric filters; they use K k minus one binary sub-classifiers to predict that minimum requirement >

Jane: The results are pretty strong. Offline, the booking prediction improvement was plus fifty-one point nine percent and amenity engagement was plus sixty-two point eight percent compared to their old baseline model >

Tom: And online, they showed a twenty point lift in engagement with recommended filters and a boost in using things like bathroom or bedroom filters >

Jane: But there's one big caveat. They admit that the daily embeddings are stale for very fast, same-day sessions because they can’t keep up with the speed of those interactions >

Tom: That means if a guest makes a rapid series of changes in one session, their embedding might not reflect it correctly until the next day's batch runs >

Jane: So, while this system is fantastic for long-term patterns and general intent, it has to work with these immediate sessions that happen in seconds >

Tom: Exactly. It’s a powerful representation learning approach that replaces brittle aggregates with sequence models, but you still have to manage those real-time session staleness issues >

Jane: So the paper shows how you can get a much richer signal from raw behavior, but the next challenge is making that signal instantly relevant for every single click >

The paper's improvements: Tom: So, beyond just building this system from scratch, SIFT actually suggests some specific ways to make it better >

Jane: Right, so they aren't just handing us a black box; they’re giving us advice on how to tune the training and how to deploy it >

Tom: One thing they focus on is that credit assignment for those booking predictions, which we talked about earlier >

Jane: They suggest using an attribution-discounted objective where each booking gets reweighted based on its rank in the search results >

Tom: That means they upweight conversions more if the filter actually caused them to appear at a better spot, rather than just saying "this filter led to a booking" >

Jane: It’s trying to make sure the model learns from causal effects instead of just correlation between filters and bookings >

Tom: Then there’s the way they handle those numeric capacity filters by using ordinal decomposition, which we touched on earlier >

Jane: They suggest training K k minus one binary sub-classifiers for each filter to predict a threshold, like "at least two bedrooms," instead of just treating it as an unordered class >

Tom: That’s a cleaner way to model those continuous ranges in the system and keeps the multi-task heads more manageable >

Jane: And they really push for decoupling the embedding materialization, which is what makes serving this thing at scale feasible >

Tom: They want a daily incremental batch job that only refreshes embeddings for guests with new events and writes them to a key-value store >

Jane: That keeps the heavy transformer computation off the request path entirely, aiming for sub-ten milliseconds inference latency on every single search query >

Tom: I like that architecture because it keeps the encoder inference totally separate from what people are actually doing in real time >

Jane: It moves the heavy lifting to an offline stage, which is necessary for models this complex, but the trade-off is still those same-day sessions being a bit stale >

Tom: Right. So they’re optimizing for high performance and generalization across different filter types while acknowledging that real-time session speed is still a hurdle >

Jane: It shows that even with these limitations, using raw sequences as the input representation can significantly boost conversion and engagement metrics across the board >

Tom: It's not just about better accuracy numbers; it’s about how this structure lets you extend the system to new surfaces, like hotel intent, without having to rebuild all your entire feature pipeline >

Jane: Exactly. The unified representation is the real asset here because it means adding a new filter type is just adding a new head or token feature, not building a whole new ETL pipeline from scratch >

Tom: So SIFT’s main implication is that you can build one powerful representation and plug different prediction tasks into it, which makes scaling personalization much more modular >

Jane: It proves that sequence modeling can handle the complexity of guest intent better than fixed features ever could, even if there are still some technical challenges with real-time data freshness >

Conclusion: Tom: So, to wrap up, SIFT is replacing those old hand-engineered features with a transformer that learns guest preferences directly from raw behavioral sequences across three different prediction tasks >

Jane: Right. It’s about moving away from pre-aggregated numbers and learning what a guest wants right now by looking at their entire journey of events >

Tom: The big win here is how it handles the complexity of numeric filters, which they do through that ordinal decomposition method >

Jane: And the results are solid. Offline, we saw improvements in booking prediction accuracy around plus fifty-one point nine percent and amenity engagement up by sixty-two point eight percent compared to the baseline model >

Tom: And online, those numbers translate into real lift for users with a twenty point increase in engagement and better use of specific filters like bedroom or bathroom ones >

Jane: It also shows that this architecture is flexible; you can extend it to new surfaces, like hotel intent, just by adding a new head without having to rebuild the entire feature pipeline >

Tom: The system is designed for efficiency too, with the embedding materialization happening offline so you’re aiming for sub-ten milliseconds inference latency on every request >

Jane: But they have to be careful about that same-day session problem, where the embeddings might not reflect very fast interactions because they are refreshed daily >

Tom: So the takeaway is that sequence modeling can outperform fixed features when you need to capture temporal drift and ordinal relationships in user behavior >

Jane: It shows how rich intent signals are when you process raw event history instead of just summary statistics >

Tom: The paper, SIFT: Search Intent-to-Filter Transformer for Multi-Task Personalized Filter Ranking at Airbnb, really proves that unifying representations for multiple prediction tasks is a solid way to build scalable personalization >

Jane: It sets a new direction for how we think about user modeling in marketplace settings, moving toward learning from the sequence itself >

Tom: Next up, we're going to look at some of these other papers on agents and how they are trying to get AI systems to actually do complex tasks rather than just predicting numbers >

Jane: We'll start with something called Agents’ Last Exam, which is looking at how well these models can actually perform in real-world testing scenarios >

Tom: Let's see what the benchmark paper says about those agent evaluations.

Shashank Dabriwal, Tanya Piplani, Hao Li, Yiwei Wang, Ashish Jain, Kedar Bellare, Stephanie Moyerman

Airbnb

cs.LG

Submitted: 2026-10-06

Updated: 2026-10-06

Comments: 9 pages, 5 figures, 6 tables. Accepted at GRAIL 2026: Workshop on Generative, Retrieval-augmented, and Agentic Intelligence for Personalization, co-located with CIKM 2026, Rome, Italy

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: The gist The SIFT model replaces hand-engineered features with transformer-generated user-sequence embeddings in a multi-task framework to personalize filter ranking at Airbnb.

Key concepts

Journey Encoder

This component takes the entire chronological history of a guest's search events as a single sequence. It uses a specific token schema to encode every event into a fixed-width vector, capturing the temporal flow of user interactions. This creates a comprehensive understanding of the guest's journey up to that point.

Multi-Task Learning Framework

Instead of training separate models for each goal, SIFT uses one shared guest representation to predict multiple outcomes simultaneously. It has three distinct prediction heads: one for booking probability, one for amenity engagement (liking specific filters), and one specifically for ordinal capacity thresholds like 'at least two bedrooms'.

Ordinal Decomposition

For filters with numerical ranges, like the number of bedrooms, SIFT doesn't just predict a single class. Instead, it uses an ordinal decomposition technique to predict a threshold (e.g., '2+'). This is achieved by training several binary sub-classifiers that collectively determine if the guest preference meets that specific minimum requirement.

Terminology

• Transformer — 1× in this paper · explained in 2 episodes

Summary

Improvements for AI systems

Sources

• Recommending Search Filters To Improve Conversions At Airbnb
• Improving Deep Learning For Airbnb Search
• Self-Attentive Sequential Recommendation
• BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations from Transformer

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: The Predictive Cost of ELBO Binding
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks