Two-Tower Model

The “two-tower” model is a type of neural network model often used in recommender systems to predict the relevance or preference of items for a user based on some sort of context. The name comes from the model’s structure, which is typically visualized as two separate towers (subnetworks) that handle two different types of inputs.

Here’s a high-level breakdown of how a typical two-tower model is structured:

Query Tower: This component of the model handles user-related input. It could include information about a user’s past behavior, demographic information, user’s past purchases, browsing history, etc. The user tower processes this information and generates a representation (often a vector) that encapsulates the user’s preferences.
Product Tower: This component of the model handles product-related input. It might process information about a product’s features, historical data about how users have interacted with the item, item’s category, price, and more. Similar to the user tower, the item tower generates a representation (again, often a vector) that encapsulates the item’s characteristics.

The two vectors generated by the user and product towers are then combined in using dot product to predict that a user will buy an item. This score is used to recommend items to the user.

In essence, the tower aims to bring a user in close distance with products they may have a preference for. And, the corresponding user and product embeddings would be in close proximity to each other in the user-product vector space. We will discuss how the embeddings are used in model serving that involves the approximate nearest neighbor Index (ANN Index).

Retrieval

The retrieval model needs to be efficient in identifying a set of candidates relevant to a user given the query and history. In essence, the retrieval model needs to be “light” in the complexity of parameters and features required. This may mean that the model’s precision may decrease, but you gain the advantage of efficient serving time which is a requirement for this problem.

In terms of the choice of model, you could use a lighter version of the two-tower model which only incorporates user ID, a handful of user features (e.g. location) and search query.

Ranking

The ranker model needs to be optimized for precision. Among the ~500 product candidates retrieved, you want to order the products based on the likelihood that the user will purchase one from the list. If the most relevant product is far down in the list, there’s less chance that the user will find the product in the result.

To train such a model, you can use more complex model architecture that incorporates more signals and also includes deeper layers of neural networks.

Offline Evaluation

When evaluating the recommender system model offline, the following set of metrics you want to observe are:

Log-Loss
AUC
Precision@k
Recall@k
Mean Average Precision@k (MAP)
Normalized Discounted Cumulative Gain (NDCG)

Online Evaluation

The offline metrics alone are not sufficient in informing whether the recommender system is reliable or not. You need offline metrics. To do so, you can observe the following online metrics in the form of AB testing.

Impression Rate
Order Count
Revenue

Model Training

Unlock now by joining.