Skip to main content

Choose a model policy

This guide provides an overview of models available for training in Shaped, helping you choose the right model for your use case.

Model types​

You can train two types of models in Shaped:

  • Embedding models, which learn dense vector representations for users and items for search and similarity
  • Scoring models, which take rich feature sets for a (user, item, context) pair and output a relevance score to rank candidates

These models can be configured in the training block of your config, under models:

training: 
  models: 
    - name: my_elsa_score_model
      policy_type: elsa
      strategy: early_stopping

You don't need a model policy!​

You can get a lot of functionality out of an engine with no trained embedding or scoring models.

Add an interaction table: Any engine with an interaction_table (which must have user_id, item_id, label, and created_at columns) will rank items by popularity, trending, and so on. The items table will have columns that expose the rank of these items compared to other items: Derived columns

Use huggingface embeddings for similarity search: If you create an embedding in the index config, you can use that to find items that are similar to a single item, or items that are complementary to a set of items: Complement items

Use value modeling: All engines support value modeling by default, so you can begin by scoring items with expressions. Here's an example that uses popular rank with a time decay: ORDER BY score(expression='(1000 - item._derived_popular_rank) / ((((now_seconds() - item.published_at) / 3600) + 2) ** 1.8)', input_user_id='$user_id', input_interactions_item_ids='$interaction_item_ids')

Embedding models​

SVD (Singular Value Decomposition)​

Model Summary​

SVD implements matrix factorization using iterative algorithms inspired by singular value decomposition (like Funk SVD or SVD++). It can incorporate user/item biases and is typically trained using Stochastic Gradient Descent (SGD).

When to Use This Model​

  • You have explicit feedback data (ratings, reviews with scores)
  • You want to model user and item biases (mean-centering tendencies)
  • You prefer SGD-based optimization over ALS
  • You need a straightforward matrix factorization approach
  • You're working with medium-sized datasets

When to Not Use This Model​

  • You have only implicit feedback (ALS is typically better for this)
  • You need to leverage item content features (use Two-Tower or BeeFormer)
  • You want the most scalable solution (ELSA may be better)
  • You need to model sequential patterns (use sequential models)

Sample Use Cases / Item Types​

  • Movie recommendations with explicit ratings (e.g., MovieLens)
  • Product recommendations with review scores
  • Restaurant or service recommendations with ratings
  • Any domain with explicit user feedback

ELSA (Efficient Latent Sparse Autoencoder)​

Model Summary​

ELSA is a scalable shallow linear autoencoder for implicit feedback collaborative filtering that learns item-item relationships by reconstructing user interaction vectors. Unlike EASE, it uses a factorized hidden layer structure (low-rank plus sparse) to improve scalability.

When to Use This Model​

  • You have large-scale datasets and need scalability
  • You're working with implicit feedback data
  • You want better scalability than EASE while maintaining similar performance
  • You need efficient item-item similarity computation
  • You want a modern autoencoder approach for collaborative filtering

When to Not Use This Model​

  • You have very small datasets (EASE might be simpler)
  • You need to incorporate item content features (use Two-Tower or BeeFormer)
  • You want to model sequential patterns (use sequential models)
  • You have explicit feedback only (SVD might be more appropriate)

Sample Use Cases / Item Types​

  • Large e-commerce platforms with millions of products
  • Content streaming services with extensive catalogs
  • Social media feed recommendations
  • Any large-scale implicit feedback recommendation system

EASE (Embarrassingly Shallow Autoencoder)​

Model Summary​

EASE is a simple linear autoencoder for item-based collaborative filtering that uses a closed-form solution, avoiding iterative optimization for computational efficiency.

When to Use This Model​

  • You want the simplest possible autoencoder approach
  • You need fast training with a closed-form solution
  • You have implicit feedback data
  • You're working with medium-sized datasets
  • You want a quick baseline or prototype

When to Not Use This Model​

  • You have very large-scale datasets (ELSA scales better)
  • You need to leverage item content features (use Two-Tower or BeeFormer)
  • You want to model sequential patterns (use sequential models)
  • You need the highest accuracy (more complex models may perform better)

Sample Use Cases / Item Types​

  • Medium-sized e-commerce sites
  • Content recommendation systems
  • Product similarity for "customers who bought this also bought"
  • Quick prototypes and baselines

Two-Tower​

Model Summary​

The Two-Tower architecture separates user and item computation into two distinct neural networks that output embeddings in the same vector space. Item embeddings can be pre-computed offline, enabling efficient Approximate Nearest Neighbor (ANN) search at inference time.

When to Use This Model​

  • You have rich item metadata (text descriptions, categories, images)
  • You need efficient large-scale vector search and retrieval
  • You want to leverage both collaborative and content signals
  • You need production-ready embeddings for real-time recommendations
  • You want the best performance for general item similarity tasks

When to Not Use This Model​

  • You have only interaction data without item features (use ALS/ELSA)
  • You need to model strict sequential patterns (use SASRec/BERT4Rec)
  • You have very limited compute resources
  • You need a simple baseline model

Sample Use Cases / Item Types​

  • E-commerce with product descriptions, categories, images (e.g., clothing, electronics)
  • Content platforms with rich metadata (articles, videos with descriptions)
  • Job recommendations with job descriptions and requirements
  • Real estate with property descriptions and features
  • Any domain where items have rich textual or categorical attributes

BeeFormer​

Model Summary​

BeeFormer fine-tunes pre-trained sentence Transformers using user-item interaction data to bridge semantic similarity (from language models) and interaction-based similarity (from collaborative patterns).

When to Use This Model​

  • You have rich text content (product descriptions, titles, reviews)
  • You want to combine semantic understanding with behavioral signals
  • You need good cold-start performance for new items
  • You want to leverage pre-trained language model knowledge
  • You have items with detailed textual descriptions

When to Not Use This Model​

  • You have minimal or no text content for items
  • You only have interaction data without item descriptions
  • You need the fastest training possible
  • You have very limited text data quality
  • You want a pure collaborative filtering approach

Sample Use Cases / Item Types​

  • E-commerce with detailed product descriptions (fashion, furniture, electronics)
  • Content platforms (articles, blog posts, research papers)
  • Job boards with detailed job descriptions
  • Real estate with property descriptions
  • Books, movies, or media with synopses and reviews
  • Any domain where semantic understanding of text content matters

Item2Vec​

Model Summary​

Item2Vec adapts the Word2Vec algorithm (CBOW or Skip-gram) to learn item embeddings from user interaction sequences, capturing co-occurrence patterns within a context window.

When to Use This Model​

  • You have sequential interaction data (user sessions, browsing history)
  • You want to capture item co-occurrence patterns
  • You need a simple, efficient sequential embedding approach
  • You're working with session-based recommendations
  • You want to model "items frequently viewed together"

When to Not Use This Model​

  • You need to model strict temporal order and dependencies (use SASRec/BERT4Rec)
  • You want to leverage item content features (use Two-Tower or BeeFormer)
  • You only have non-sequential interaction data
  • You need the highest accuracy for sequential recommendations

Sample Use Cases / Item Types​

  • E-commerce browsing sessions (items viewed in same session)
  • Music playlists and song sequences
  • Video watching sequences
  • Shopping cart co-occurrence patterns
  • Session-based web recommendations

SASRec (Self-Attentive Sequential Recommendation)​

Model Summary​

SASRec utilizes the Transformer architecture's self-attention mechanism to model user interaction sequences, capturing both short-term and long-range dependencies to predict the next item. It's unidirectional, processing sequences from past to future.

When to Use This Model​

  • You have sequential interaction data with clear temporal order
  • You need to predict the "next item" in a sequence
  • You want to capture both short-term and long-term patterns
  • You need state-of-the-art sequential recommendation performance
  • You're building session-based or sequential recommendation systems

When to Not Use This Model​

  • You don't have sequential or temporal data
  • You want general item similarity rather than next-item prediction
  • You need bidirectional context understanding (use BERT4Rec)
  • You have very short sequences (Item2Vec might be simpler)
  • You want to leverage item content features primarily (use Two-Tower)

Sample Use Cases / Item Types​

  • E-commerce next-item prediction (what to buy next)
  • Video streaming next-video recommendations
  • Music playlist continuation
  • News feed article sequences
  • Gaming item progression recommendations
  • Any domain with clear sequential user behavior

BERT4Rec​

Model Summary​

BERT4Rec adapts the bidirectional Transformer architecture (BERT) for sequential recommendation, using bidirectional self-attention to learn context-aware item representations by considering both past and future context.

When to Use This Model​

  • You have sequential data and want bidirectional context understanding
  • You need richer item representations than unidirectional models
  • You want to leverage both past and future context in sequences
  • You're working with sequences where context matters in both directions
  • You need state-of-the-art sequential recommendation with bidirectional attention

When to Not Use This Model​

  • You need real-time next-item prediction (unidirectional is more natural)
  • You have very long sequences (computational complexity)
  • You want the simplest sequential model (Item2Vec or SASRec)
  • You don't have sequential data
  • You primarily need general item similarity (use Two-Tower or ALS)

Sample Use Cases / Item Types​

  • Context-aware sequential recommendations
  • Playlist generation with full context
  • Reading sequences where future context matters
  • Educational content sequences
  • Any sequential recommendation where bidirectional understanding helps

GSASRec (Generalized Self-Attentive Sequential Recommendation)​

Model Summary​

GSASRec is an enhancement of SASRec designed to mitigate overconfidence issues from negative sampling by improving prediction calibration while retaining the core self-attention mechanism.

When to Use This Model​

  • You're experiencing overconfidence issues with SASRec
  • You need better calibrated predictions for sequential recommendations
  • You want improved negative sampling strategies
  • You need the benefits of SASRec with better training stability
  • You're working on production systems where calibration matters

When to Not Use This Model​

  • You don't have sequential data
  • You want the simplest sequential model (use SASRec or Item2Vec)
  • You need general item similarity (use Two-Tower or ALS)
  • You're just starting with sequential recommendations (SASRec is simpler)
  • You want to leverage item content primarily (use Two-Tower or BeeFormer)

Sample Use Cases / Item Types​

  • Production sequential recommendation systems
  • E-commerce next-item prediction with calibrated scores
  • Video streaming with improved prediction confidence
  • Any sequential recommendation where prediction calibration is critical

Embedding Models - Quick Decision Guide​

For general item similarity and vector search:

  • Best choice: Two-Tower (if you have item features) or ALS/ELSA (if you only have interactions)

For text-rich items:

  • Best choice: BeeFormer (combines semantic + behavioral) or Two-Tower (if you have other features too)

For sequential recommendations:

  • Best choice: SASRec (unidirectional) or BERT4Rec (bidirectional), or Item2Vec (simpler co-occurrence)

For large-scale implicit feedback:

  • Best choice: ELSA (scalable) or ALS (simple and effective)

For explicit feedback:

  • Best choice: SVD (with biases) or ALS (can work with explicit too)

Scoring models​

LightGBM​

Model Summary​

LightGBM is a highly efficient Gradient Boosting Decision Tree (GBDT) framework optimized for speed and memory usage. It builds ensembles of decision trees with histogram-based splitting, leaf-wise growth, and techniques like Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB), making it well-suited for large, sparse, and high-cardinality feature spaces. It supports regression, classification, and learning-to-rank objectives such as LambdaRank/LambdaMART.

When to Use This Model​

  • You have large-scale tabular datasets with many categorical and numerical features
  • You need fast training and inference for ranking or CTR prediction workloads
  • You want a learning-to-rank objective (e.g., LambdaRank) for slate or search result ranking
  • You need a production-ready, memory-efficient GBDT implementation as a strong baseline
  • You want to combine rich behavioral, content, and affinity features into a single scoring model

When to Not Use This Model​

  • You primarily work with small or medium-sized datasets and prefer more conservative defaults (use XGBoost)
  • You want to explicitly model high-order feature interactions via deep networks (use DeepFM or Wide & Deep)
  • You have only sparse interaction data and no rich features (use ALS/ELSA or other embedding models)
  • You need sequence-aware modeling of user behavior (use SASRec, BERT4Rec, or other sequential models)
  • You only need a simple trending or heuristic baseline (use Rising Popularity or value-model expressions)

Sample Use Cases / Item Types​

  • E-commerce CTR prediction and ranking for product search and browse results
  • Feed ranking and homepage personalization using mixed behavioral and content features
  • Ad ranking and sponsored content placement with learning-to-rank objectives
  • Re-ranking of retrieved candidates in multi-stage recommendation architectures

XGBoost​

Model Summary​

XGBoost is a widely used GBDT framework known for its performance, rich regularization options, and robustness on structured data. It builds ensembles of decision trees with level-wise growth, supports various objectives (classification, regression, ranking), and includes regularization (L1/L2) and advanced handling of missing values. It is often favored when you need strong baselines and interpretability on small to medium-sized tabular datasets.

When to Use This Model​

  • You work with structured/tabular data and want a strong, regularized baseline model
  • You have small to medium-sized datasets where XGBoost’s conservative defaults work well
  • You need robust handling of noisy data, missing values, and mixed feature types
  • You want a well-understood, battle-tested framework for CTR prediction or ranking
  • You care about interpretability via tree-based feature importance and decision paths

When to Not Use This Model​

  • You need maximum training and inference speed on very large datasets (use LightGBM)
  • You want neural models that can learn complex high-order feature interactions (use DeepFM or Wide & Deep)
  • You primarily have sparse interaction-only signals without many engineered features (use ALS/ELSA or other embedding models)
  • You need sequence-aware modeling of user behavior (use SASRec, BERT4Rec, or other sequential models)
  • You mostly care about non-personalized trending content (use Rising Popularity)

Sample Use Cases / Item Types​

  • CTR prediction for recommendation or advertising on moderate-scale datasets
  • Ranking search and browse results using handcrafted and learned features
  • Scoring candidates in multi-stage recommendation pipelines where robustness matters
  • Rapid prototyping and A/B testing of feature sets before deploying more complex models

Wide & Deep​

Model Summary​

Wide & Deep jointly trains a wide linear model and a deep neural network to combine memorization and generalization. The wide part consumes raw sparse features and manually engineered cross-product features, excelling at memorizing frequent co-occurrence patterns, while the deep part consumes dense embeddings and continuous features to learn non-linear, high-level representations. Their outputs are combined to produce the final score.

When to Use This Model​

  • You want to combine memorized feature interactions with generalization to unseen combinations
  • You can invest in engineering high-value cross-product features for the wide part
  • You have rich categorical features that can be embedded and fed into a deep network
  • You are building recommendation or search ranking similar to app stores or content feeds

When to Not Use This Model​

  • You do not want to maintain manual feature engineering for a wide component (use DeepFM instead)
  • You prefer tree-based models with simpler feature workflows (use LightGBM or XGBoost)
  • You primarily need retrieval-stage embeddings rather than point-wise scoring (use Two-Tower, ALS, or BeeFormer)
  • You care mainly about sequential order in interactions (use SASRec, BERT4Rec, or other sequential models)

Sample Use Cases / Item Types​

  • App recommendation and app store ranking (apps, games, extensions)
  • Personalized content ranking for homepages, feeds, and carousels
  • Search ranking where memorized query–item patterns and generalization both matter
  • CTR and conversion modeling in environments with rich, hand-crafted cross features

DeepFM​

Model Summary​

DeepFM combines a Factorization Machine (FM) component and a deep neural network (DNN), sharing the same input embeddings. The FM component efficiently models low-order feature interactions (linear and pairwise), while the deep component captures high-order, non-linear interactions. This removes the need for manual feature crossing as in Wide & Deep, while still covering a wide spectrum of interactions.

When to Use This Model​

  • You need a strong CTR or conversion prediction model that captures both low- and high-order feature interactions
  • You have many sparse categorical features and continuous features, and want to avoid manual feature crossing
  • You care about modeling interaction terms between user, item, and context features
  • You want a neural scoring model that can outperform linear or pure FM-based approaches

When to Not Use This Model​

  • You have relatively simple feature sets and want a fast, tree-based baseline (use LightGBM or XGBoost)
  • You do not have the infrastructure to train and serve neural models in production
  • You mainly need retrieval-stage embeddings rather than point-wise scoring (use Two-Tower, ALS, or BeeFormer)
  • You primarily need to capture sequential patterns in user behavior (use SASRec, BERT4Rec, or other sequential models)

Sample Use Cases / Item Types​

  • CTR prediction for feeds, recommendations, and advertising with many categorical IDs
  • Ranking items in e-commerce or content platforms using rich user, item, and context features
  • Scoring candidates in ranking stages after retrieval from embeddings or lexical/vector search
  • Personalization tasks where complex feature interactions drive performance

Ngram​

Model Summary​

The Ngram policy is a simple, frequency-based sequential model. It estimates probabilities of the next item given the preceding (n-1) items by counting n-grams in historical interaction sequences, with optional Laplace smoothing to handle unseen patterns. It focuses on short-term co-occurrence and locality in sequences rather than long-range dependencies.

When to Use This Model​

  • You want a simple, interpretable baseline for sequential or session-based recommendations
  • You have sparse interaction data where complex neural sequence models may overfit
  • You mostly care about short-term co-occurrence patterns in the latest interactions
  • You need a fast, low-compute sequential model with minimal training complexity

When to Not Use This Model​

  • You need to capture long-range dependencies or rich temporal structure (use SASRec, GSASRec, or BERT4Rec)
  • You have dense sequential data and can benefit from more expressive neural models
  • You primarily need general item similarity or retrieval embeddings (use Two-Tower, ALS, or Item2Vec)
  • You mainly focus on global trending content rather than per-user sequences (use Rising Popularity)

Sample Use Cases / Item Types​

  • Session-based recommendations on small or medium-sized sites
  • Next-click or next-view prediction from a short recent interaction history
  • Baseline models for evaluating more sophisticated sequential architectures
  • Educational or low-traffic environments where interpretability and simplicity are preferred

Rising Popularity​

Model Summary​

Rising Popularity is a rule-based policy that ranks items based on recent changes in engagement, emphasizing momentum rather than absolute volume. It compares engagement in a recent window against a previous baseline to surface items that are rapidly gaining traction, making it useful for trending and “what’s hot” experiences.

When to Use This Model​

  • You want to surface items that are rapidly gaining engagement or momentum
  • You need a non-personalized trending or “Top rising” experience
  • You are handling cold-start or anonymous users where you lack rich history
  • You want a simple, robust baseline for time-sensitive content without training a complex model

When to Not Use This Model​

  • You need highly personalized rankings driven by user history (use ALS, ELSA, Two-Tower, or sequential models)
  • You care more about long-term relevance than short-term spikes (use LightGBM, XGBoost, DeepFM, or Wide & Deep)
  • Your catalog or usage patterns change slowly and recent engagement is not a strong signal
  • You need fine-grained control over feature interactions and objectives (use learning-to-rank or neural scoring models)

Sample Use Cases / Item Types​

  • Trending or “Top rising” sections on homepages and category pages
  • Surfacing viral or breaking news articles, videos, or social posts
  • Cold-start experiences for new or logged-out users before personalization is available
  • Fallback or blended signals in multi-signal ranking strategies (e.g., mixing trending with personalized scores)

Scoring Models - Quick Decision Guide​

For general CTR and ranking on tabular data:

  • Best choice: LightGBM or XGBoost, depending on dataset size and performance needs

For rich feature interaction modeling without manual crosses:

  • Best choice: DeepFM (or Wide & Deep if you are already invested in feature engineering)

For simple or resource-constrained sequential setups:

  • Best choice: Ngram (baseline), or reuse sequential embedding models in a scoring stage

For trending and cold-start experiences:

  • Best choice: Rising Popularity, often blended with personalized scores from other models