User Modeling | Skelter Labs
User Modeling
User modeling to train unstructured data
User modeling is the extraction of key characteristics of a target user that meets a goal. User modeling based on context and behavior can enhance the accuracy of recommendations and predictions. Skelter Labs’ recommendation and prediction solution for e-commerce, based on unique deep learning-based modeling, learns the relationship between tens of thousands of users and products, and provides recommendations and prediction results in real time through the following methods.
User Modeling
User modeling to train unstructured data
User modeling is the extraction of key characteristics of a target user that meets a goal. User modeling based on context and behavior can enhance the accuracy of recommendations and predictions. Skelter Labs’ recommendation and prediction solution for e-commerce, based on unique deep learning-based modeling, learns the relationship between tens of thousands of users and products, and provides recommendations and prediction results in real time through the following methods.

Product and User Embedding
Finding the proper features from unstructured data is called “embedding.” Skelter Labs utilizes product and user embedding methods, inspired by the word embedding method actively used in the field of natural language understanding (NLU).
Product and user embedding learns the correlation between users and products by extracting features and placing similar values close together based on interactions between users and products and users’ behavior history. Skelter Labs’ embedding model can be modeled in various ways utilizing data such as text, images and context, in addition to the structured user behavior log. Learned embeddings are utilized in the recommendation engine in various ways to enable mathematical operations. In addition, the embedding model extracts long-term user-specific features by applying style embedding, which is mainly used in the fields of speech and image recognition, and contributes to the improved performance of the recommendation engine.
Product and User Embedding
Finding the proper features from unstructured data is called “embedding.” Skelter Labs utilizes product and user embedding methods, inspired by the word embedding method actively used in the field of natural language understanding (NLU).
Product and user embedding learns the correlation between users and products by extracting features and placing similar values close together based on interactions between users and products and users’ behavior history. Skelter Labs’ embedding model can be modeled in various ways utilizing data such as text, images and context, in addition to the structured user behavior log. Learned embeddings are utilized in the recommendation engine in various ways to enable mathematical operations. In addition, the embedding model extracts long-term user-specific features by applying style embedding, which is mainly used in the fields of speech and image recognition, and contributes to the improved performance of the recommendation engine.
Prediction Model
Skelter Labs’ purchase prediction model utilizes a variety of networks and techniques used in the latest language models such as BERT. The model can understand and learn user behavior in various aspects using the sequential user behavior log, and reflect the context around the user in the learning.
User embeddings created as a result of predictions allow similar users to be clustered and used in marketing tools. Plus, the latest log, containing changes in users’ real-time behavior and preferences, is reflected in the model, resulting in high prediction accuracy based on the latest data.
Prediction Model
Skelter Labs’ purchase prediction model utilizes a variety of networks and techniques used in the latest language models such as BERT. The model can understand and learn user behavior in various aspects using the sequential user behavior log, and reflect the context around the user in the learning.
User embeddings created as a result of predictions allow similar users to be clustered and used in marketing tools. Plus, the latest log, containing changes in users’ real-time behavior and preferences, is reflected in the model, resulting in high prediction accuracy based on the latest data.