Increased
CTR
CTR
4.5%
Boosted engagement
29%
Increased
bookings
bookings
4.9%
Goals
Batch aimed to increase conversions and drive revenue by improving how users discover and book group travel experiences.
Short-term, the team focused on launching a personalized recommendation feed to surface relevant accommodations and experiences, reducing user drop-off during planning.
Long-term, Batch wanted to build a smarter event planning flow that dynamically recommends experiences & accommodations- such as lodging and activities- based on the event type, attendee preferences, and trip context.
Short-term, the team focused on launching a personalized recommendation feed to surface relevant accommodations and experiences, reducing user drop-off during planning.
Long-term, Batch wanted to build a smarter event planning flow that dynamically recommends experiences & accommodations- such as lodging and activities- based on the event type, attendee preferences, and trip context.
The challenge
Before integrating Shaped, Batch relied on a static recommendation system that lacked personalization and contextual relevance. Two major challenges limited its effectiveness:
1. Poor Relevance Matching
The legacy system struggled to identify meaningful connections between user preferences and related experiences. Without a robust content understanding or behavioral modeling, recommendations often felt generic or misaligned.
2. Gender Misalignment in Recommendations
A core issue was the system’s inability to account for gender alignment. For example, users who attended a bachelor or birthday party were frequently shown bachelorette-style experiences—resulting in mismatched and irrelevant suggestions. This problem was exacerbated by data imbalances: only ~10% of experiences were categorized as masculine, making it even harder to maintain relevance for users interested in masculine experiences.
These challenges highlighted the need for a more intelligent, behavior-driven recommendation engine that could infer user intent, handle data sparsity, and deliver personalized, gender-aware experiences at scale.
1. Poor Relevance Matching
The legacy system struggled to identify meaningful connections between user preferences and related experiences. Without a robust content understanding or behavioral modeling, recommendations often felt generic or misaligned.
2. Gender Misalignment in Recommendations
A core issue was the system’s inability to account for gender alignment. For example, users who attended a bachelor or birthday party were frequently shown bachelorette-style experiences—resulting in mismatched and irrelevant suggestions. This problem was exacerbated by data imbalances: only ~10% of experiences were categorized as masculine, making it even harder to maintain relevance for users interested in masculine experiences.
These challenges highlighted the need for a more intelligent, behavior-driven recommendation engine that could infer user intent, handle data sparsity, and deliver personalized, gender-aware experiences at scale.
Our solution
Phase 1: Building the Foundation
The engagement began with building a baseline recommender using static datasets focused on event categorization (e.g., bachelor, birthday, party-boat). This phase established a core personalization layer grounded in historical user interactions and experience types in just 2 weeks.
Phase 2: Gender Alignment via Feature Engineering
One key challenge was the lack of explicit user gender data, which sometimes led to mismatches—such as recommending bachelorette-focused experiences to male users. To address this, we engineered features to infer "gender alignment" for each experience using keyword semantics and historical participation trends. This proxy allowed for more context-aware filtering and prioritization without requiring direct demographic data.
Phase 3: Hybrid Modeling Approach
To power more accurate and relevant recommendations, we deployed a hybrid modeling approach combining:
- ML-driven category and interaction-based ranking
- Deterministic rules to prioritize and re-rank based on inferred gender alignment
- Dual modeling of key attributes as both structured categories and unstructured text
- Embedding space generation to identify hyper-similar experiences using Shaped’s content similarity capabilities
Deployment timeline
Initial model development: ~2 weeks
Mature model deployment: ~6 weeks (including tuning and iteration)
Event Weighting Schema
To guide model training and relevance scoring, user engagement events were weighted as follows:
Impressions: 0 (Neutral)
Favorites: 5 (Medium signal)
Booked Experiences: 10 (High signal)
Implementation Approach
All data transformations and model input pipelines were built using SQL interfaces, enabling fast prototyping, easy iteration, and minimal engineering overhead during model tuning.
The engagement began with building a baseline recommender using static datasets focused on event categorization (e.g., bachelor, birthday, party-boat). This phase established a core personalization layer grounded in historical user interactions and experience types in just 2 weeks.
Phase 2: Gender Alignment via Feature Engineering
One key challenge was the lack of explicit user gender data, which sometimes led to mismatches—such as recommending bachelorette-focused experiences to male users. To address this, we engineered features to infer "gender alignment" for each experience using keyword semantics and historical participation trends. This proxy allowed for more context-aware filtering and prioritization without requiring direct demographic data.
Phase 3: Hybrid Modeling Approach
To power more accurate and relevant recommendations, we deployed a hybrid modeling approach combining:
- ML-driven category and interaction-based ranking
- Deterministic rules to prioritize and re-rank based on inferred gender alignment
- Dual modeling of key attributes as both structured categories and unstructured text
- Embedding space generation to identify hyper-similar experiences using Shaped’s content similarity capabilities
Deployment timeline
Initial model development: ~2 weeks
Mature model deployment: ~6 weeks (including tuning and iteration)
Event Weighting Schema
To guide model training and relevance scoring, user engagement events were weighted as follows:
Impressions: 0 (Neutral)
Favorites: 5 (Medium signal)
Booked Experiences: 10 (High signal)
Implementation Approach
All data transformations and model input pipelines were built using SQL interfaces, enabling fast prototyping, easy iteration, and minimal engineering overhead during model tuning.
8 weeks
Phase 1: Building
the Foundation
the Foundation
Developing a Base Recommender
.svg)
Phase 2: Gender Alignment via Feature Engineering
Infer gender alignment with keyword semantics and historical participation trends.
Phase 3: Hybrid Modeling Approach
ML-driven ranking and
re-ranking model
Results & outcomes
By replacing their in-house ranking system with Shaped's personalized recommendations, Batch saw a 4.5% increase in click-through rates, a 29% lift in favorites and 4.9% increase in bookings.
An A/B test run over just 2 weeks showed a clear boost in user engagement, users interacted more with listings that felt more relevant to them, signaling stronger interest and intent.
An A/B test run over just 2 weeks showed a clear boost in user engagement, users interacted more with listings that felt more relevant to them, signaling stronger interest and intent.
Expanded use cases
Following the success of the initial implementation, Shaped has continued to power several on-going use cases, including:
1. Personalized accommodation recommendations
Lodging recommendation options that better match user’s individual preferences and behaviors.
2. Context-aware suggestions
For example, if someone is attending a specific type of experience (like a festival or outdoor activity), they can receive tailored accommodation recommendations that complement that experience.
3. Smarter search results
Better search experience with ranking and filtering, taking into account both user preferences and real-time context to show the most relevant results first.
1. Personalized accommodation recommendations
Lodging recommendation options that better match user’s individual preferences and behaviors.
2. Context-aware suggestions
For example, if someone is attending a specific type of experience (like a festival or outdoor activity), they can receive tailored accommodation recommendations that complement that experience.
3. Smarter search results
Better search experience with ranking and filtering, taking into account both user preferences and real-time context to show the most relevant results first.