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

1 Introduction

• Guests’ initial queries state only coarse intent, so recommended filters make unstated preferences explicit; prior work found proactive filter recommendations drove up to +2.4% booking conversion [1].The paper frames filter recommendation as a lower-funnel ranking problem.Source: §1 · p. 1
• The production MLP uses nightly, hand-engineered aggregate features, and adding filters or contextual dimensions requires new pipelines; SIFT uses raw events and needs a new head or token feature instead.Aggregate features also lose event order and recency, while the incumbent models booking and binary amenity engagement but not ordinal capacity filters.Source: §1 · pp. 1–2
• Offline, booking PR-AUC improved +51.9% and amenity-engagement macro PR-AUC +62.8% over the production MLP baseline; online, Recommended Filter Clickers rose +20.0% and Searchers With Filters +0.72%.The introduction also reports +3.8% uncancelled hotel bookings and +0.76% overall marketplace bookings for the hotel-intent extension.Source: §1 · pp. 2–3
• SIFT jointly trains a journey encoder and shared task heads, ranking filters by booking and engagement probabilities while supporting binary and ordinal filter tasks.The framework was extended to a previously unmodeled hotel-intent class.Source: §1 · p. 2

Faceted and filter-based search

• Earlier facet methods optimize informative partitions or navigational cost, while Airbnb’s supervised filter-ranking work optimizes downstream conversion [1]; SIFT instead personalizes slates using guest history.Guests with identical queries and result sets can receive different filter slates.Source: §2 · p. 3
• Facet research asks which facets help users navigate collections, while SIFT shifts the objective to which filters a particular guest is likely to apply and convert under.This reframes filter selection as personalization rather than result-set summarization.Source: §2 · p. 3

Sequence-based recommendation models

• SASRec [8] and BERT4Rec [9] show sequence encoders can recover order and recency discarded by aggregates; Airbnb’s JourneyFormer [2] models long-horizon events alongside recent page views.Industrial ranking systems have also adopted sequence encoders as components of larger architectures [7].Source: §2 · p. 3

Temporal Drift in Filter Preferences

• A majority of guests changed their single most-used filter between two disjoint, season-matched years, and the pattern held for their top-five filter sets.Table 1 reports guest-level preference drift across those periods.Source: §3 · p. 3
• Filter preference changed more rapidly than population-level conversion value across the same two disjoint years: unweighted average relative FCR change was 35.5%, versus 24.2% volume-weighted, with the largest single swing at 88.3%.Residual conversion instability was concentrated in a subset of low-volume filters.Source: §3 · p. 3, Table 2 · p. 3
• Some guests’ capacity thresholds rose across years or a previously unused amenity became their top filter; ordered histories can distinguish recent adoption from long-term steady use.A crib filter could become a guest’s most-used filter after a year with no use, and sequence recency distinguishes searches begun three days ago from ten months of steady searches.Source: §3 · pp. 3–4

• Candidate filters are ranked by a score combining booking probability given the guest, query, and filter with engagement probability given the guest and query.The formulation treats a filter’s utility as its contribution to booking conversion, rather than engagement alone.Source: §4 · p. 4
• The three tasks predict booking probability, per-amenity engagement, and threshold probabilities for numeric filters such as at least 2 bedrooms.The ordinal task predicts whether a numeric filter value meets each threshold.Source: §4 · p. 4
• For a numeric threshold filter, engagement probability is recovered by subtracting adjacent cumulative threshold probabilities.For threshold v, the formulation subtracts P(V_k ≥ v+1 | u,q) from P(V_k ≥ v | u,q).Source: §4 · p. 4

5 Model Architecture

• A transformer produces a daily materialized guest embedding, which feature assembly combines with query and candidate-filter signals before three task-specific head types score filters.The heads predict booking, engagement, and ordinal capacity.Source: §5 · p. 4, Figure 2 · p. 5

5.1 Architecture

• A jointly trained transformer processes the chronological history and projects its final sequence position into a 32-dimensional guest embedding for the task heads.The architecture uses one self-attention layer with four heads and model dimension 128.Source: §5.1 · p. 4

5.2 SIFT Encoder (User Sequence Modeling)

• Searches, reservation requests, and booking-related events share a fixed-width, null-padded token schema, with event order and positional encoding carrying recency.Shared and event-specific attributes let one encoder process different event types without separate input branches.Source: §5.2 · p. 4

5.3 Feature Assembly

• All heads receive the guest embedding and query features, while only booking also receives candidate-filter features, decoupling personalization from query and filter signals.This design removes the need for pre-aggregated user-feature pipelines.Source: §5.3 · pp. 4–5

5.4 Task-Specific MLP Heads

• The booking head predicts booking probability with binary cross-entropy and weights bookings more when the booked listing ranked near the top of results.The position-discount weight is w_i = 1 + 1/ln(pos_i + 1).Source: §5.4 · p. 5, Figure 3 · p. 5
• Amenity engagement uses one sigmoid output per filter, while each numeric capacity filter uses Frank–Hall ordinal binary classifiers for its value thresholds.The ordinal labels are monotone across thresholds by construction; the heads are two-layer MLPs with swish activations.Source: §5.4 · p. 5, Figure 4 · p. 6

6.1 Training Data and Label Construction

• Training uses a rolling 365-day corpus of daily journeys, with each instance representing a search session and its historical sequence truncated to 384 tokens.Each journey ends in a one-day snapshot of search events with seven-day forward-attributed booking labels.Source: §6.1 · p. 6
• To limit session-level label leakage, the training pipeline keeps one representative search per session and attributes a booking only when its listing appeared in that search’s results within seven days.The retained event is chosen to prefer one with filter interactions when multiple searches exist in a session.Source: §6.1 · p. 6

6.2 Cold-Start Optimization

• Guests with no prior history use a learned UNKNOWN_ACTION embedding, providing a data-driven cold-state representation instead of an arbitrary zero vector.Source: §6.2 · p. 6

7.1 Serving

• Daily offline inference refreshes guest embeddings in a key-value store, so online scoring fetches the embedding and combines it with query features without running the encoder.Only guests with a new eligible event since the previous run require recomputation.Source: §7.1 · p. 6
• Engagement and ordinal heads score all candidate filters in one query-level pass, while the booking head is evaluated per candidate filter.Source: §7.1 · p. 6

7.2 Latency and Throughput

• The serving design targets sub-10ms inference at over 1000 QPS, using a colocated TensorFlow Serving sidecar and shared feature computation.The offline/online split leaves only the three MLP heads to run per request.Source: §7.2 · pp. 6–7

8.1 Offline Evaluation

• +51.9% booking PR-AUC and +62.8% amenity-engagement macro PR-AUC, with +35.5% amenity-engagement micro PR-AUC, versus the production baseline.SIFT also reports PR-AUC for bedroom, bed, and bathroom capacity filters, which the baseline cannot model; its three head types are well calibrated for direct use as ranking scores.Source: §8.1 · p. 7

8.2 Online A/B Tests

• Online A/B tests used identity-randomized traffic splits, tracking filter use, newly supported capacity-filter use, bookings, and low-inventory search rate as a guardrail.Low-inventory search rate measures searches returning too few results, which could be affected by overly aggressive filter recommendations.Source: §8.2 · p. 7

8.3 Ablation Study

• Removing the journey encoder drops booking PR-AUC from 0.2108 to 0.1410 (−33.1% relative) and degrades every other head by a comparable or larger margin.The ablation scores heads from candidate query features alone, with encoder parameters excluded from training.Source: §8.3 · p. 7

8.4 Cross-Domain Generalization: Hotel-Intent Case Study

• SIFT added a Hotel filter class to its engagement head and computed a personalized, query-conditioned hotel-intent score from guest history without new feature engineering or ETL pipelines.The offline evaluation used the same cohort targeted by the heuristic; for confirmed hotel bookings, model precision was 2.07× the heuristic (0.0085 vs. 0.0041) at threshold 0.11, with 34% recall.Source: §8.4 · p. 8, Table 5 · p. 8
• +0.76% overall marketplace bookings accompanied favorable hotel-surface engagement and booking metrics in the SIFT-gated treatment.The authors report that the marketplace gain was additive to hotel-surface gains, unlike the heuristic-gated treatment.Source: Table 6 · p. 8, §8.4 · p. 8

9 Discussion and Limitations

• ∼30–40% of Airbnb bookings complete within a day of the first search, too quickly for daily-refreshed embeddings to reflect those sessions’ activity.At serving time, these guests are scored with the previous day’s embedding, not treated as cold-start.Source: §9 · p. 8
• Shortening the refresh cadence could improve online performance, but requires additional streaming infrastructure.The paper identifies same-day sessions as the segment with the most to gain from more frequent updates.Source: §9 · p. 8
• The practical boundary is stale rather than absent signal: same-day predictions do not incorporate that day’s searches or bookings.The paper describes this as the daily-refresh/latency trade-off also adopted by JourneyFormer [2].Source: §9 · p. 8

10 Conclusion

• SIFT replaces hand-engineered aggregated features with transformer-generated user-sequence embeddings in a multi-task framework.Source: §10 · p. 8
• +51.9% booking PR-AUC and +62.8% amenity-engagement PR-AUC offline; online, uncancelled hotel bookings rose +3.8% and marketplace bookings +0.76%.Online results also include +20% Recommended Filter Clickers and +0.72% Searchers With Filters.Source: §10 · p. 8
• SIFT is fully deployed, serving scalable personalization to millions of guests.Source: §10 · p. 8