A Review of Modern Fashion Recommender Systems
- Key Notes
- Recommender systems have grown to be an essential part of all large Internet retailers, driving up to 35% of Amazon sales [103] or over 80% of the content watched on Netflix [31].
- Localizing fashion items
- Determining their category and attributes
- Degree of similarity to other products
- Product-to-product relationships
- Product-to-user uncertainties
- Fashion item compatibility - associated image and text data is then used to learn to generalize to stylistically similar products
- The fashion item recommendation task, similar to the classical recommendation problem, focuses on suggesting individual fashion items (clothing), that match users’ preferences.
- Fashion pair and outfit recommendation: Fashion outfits are sets of 𝑁 items that are worn together, e.g., for an outdoor wedding, graduation party, baby shower, and so forth
- Modeling outfits as a sequence. to take advantage of the representation of order-aware models such as LSTMs
- Fashion Item Relevancy network (FIR) learns the compatibility of fashion items and learns garment item relevance embeddings
- Physical body-related features. The easiest way to make effective sizing recommendations is to use data from certain parts of the body [58, 60] such as bust, waist, and hip
- User-item fit feedback. To provide personalized size recommendations, the interaction between the user and the item is essential
- Color. The most common means to identify how one looks is achieved via colors, materials, and silhouettes on the body
- Brand. Product brands are a critical feature users consider when deciding among items.
- Texture. The texture describes the body and surface of a garment.
- Context = image + text. In addition to images, users may also include words (textual descriptions) to aid in the recommendation process
- Context = image. Images are an important visual tool for users to communicate with a fashion recommender system
Toward Explainable Fashion Recommendation
- Influence of the itemfeature pair, which we call its Item-Feature Influence Value (IFIV)
- CNNs trained for generic image recognition are used to extract features for their respective purposes.
Fashion Recommendation and Compatibility Prediction Using Relational Network
- Learning compatibility between "tops" and "bottoms" Treating outfits as a sequence and using an LSTM-based mode
Single-Item Fashion Recommender: Towards Cross-Domain Recommendations
- Category: Defines the main category of an image, such as top, bottom, footwear, and jewelry.
- Subtype: Defines subtypes of the same category, such as boots, high heels, college, and slippers.
- Fabric/Texture: Shows the main fabric or garment’s texture, such as denim, leather, smooth, and shiny.
- Color: Defines the dominant color of the item, such as red, green, blue, yellow.
- Variety: The number of novel items (different category, subtype, or color). Almost on the opposite side of the other criteria, because the higher the variety score is, the lower other scores will be.
- Details: The number of results that follow fine details, such as necklines, zipper, pockets, and design.
- Shape Difference: The number of items that do not follow the outline of the query item, such as images with different angles, different perspectives, rotations, flips
Key Concepts
- Data generation
- Embedding generation
Similarity
- Cosine
- Euclidean
Data Size Reduction
- SVD
- NMF
Good Read - System Design for Recommendations and Search
The tradeoff between batch vs realtime
- What is computed prior?
- What is used in real time to adjust prior recommendations?
- Offline - creating embeddings for catalog items, and building an approximate nearest neighbors (ANN)
- Online - converting the input item or search query into an embedding, followed by candidate retrieval and ranking
Let's explore how recsys & search are often split into:
— Eugene Yan (@eugeneyan) June 30, 2021
• Latency-constrained online vs. less-demanding offline environments
• Fast but coarse candidate retrieval vs. slower and more precise ranking
Examples from Alibaba, Facebook, JD, DoorDash, etc.https://t.co/zTsfElLw1z
Transitioning to a real-time serving system has been made possible by two products: Feature Store and Online Inference Platform
Keep Exploring!!!
No comments:
Post a Comment