Machine learning model for click through rate predication using three vector representations
An online concierge system uses a machine learning click through rate model to select promoted items based on user embeddings, item embeddings, and search query embeddings. Embeddings obtained by an embedding model may be used as inputs to the click through rate model. The embedding model may be trained using different actions to score the strength of a customer interaction with an item. For example, a customer purchasing an item may be a stronger signal than a customer placing an item in a shopping cart, which in turn may be a stronger signal than a customer clicking on an item. The online concierge system generates a ranking of candidate promoted items based on the search query and using the click through rate model. Based on the ranking, the online concierge system displays promoted items along with the organic search results to the customer.
1. A method comprising:
at an online concierge system comprising one or more processors and one or more computer-readable media:
receiving, by the online concierge system, a search query from a user;
retrieving, by the online concierge system, offering embeddings for offerings, wherein the offering embeddings are generated using a machine learning embedding model, wherein an offering is offered by the online concierge system to the user;
calculating, by the online concierge system, a Hadamard product of a search query embedding for the search query and a user embedding for the user, wherein the user embedding is generated using the machine learning embedding model;
selecting, by the online concierge system, a set of offering embeddings based on distances of the offering embeddings in the set from the Hadamard product;
retrieving, by the online concierge system, a set of candidate offerings corresponding to the set of offering embeddings;
ranking, by the online concierge system, the set of candidate offerings using a second machine learning model; and
transmitting, by the online concierge system and based on the ranking, a subset of the set of candidate offerings to a user device for display to the user.
2. The method of claim 1 , further comprising training the machine learning embedding model, wherein training the machine learning embedding model comprises:
using historical data from a search log comprising triplets, each triplet including a search query, a user, and an item, wherein each triplet is labeled based on an action performed by the user on the item.
3. The method of claim 1 , wherein the user embeddings and the offering embeddings are learned by calculating an inner product of the search query embeddings, the user embeddings, and offering embeddings, and performing a learning to rank pairwise loss function.
4. The method of claim 1 , further comprising calculating, by the online concierge system, a predicted click through rate for each of the subset of the set of candidate offerings.
5. The method of claim 1 further comprising calculating, by the online concierge system, an effective cost per mile for each of the subset of the set of candidate offerings.
6. The method of claim 1 , wherein calculating the product comprises calculating a Hadamard product of the search query embeddings and the user embeddings.
7. The method of claim 1 , further comprising retrieving the search query embeddings from a search ranking module, wherein the search query embeddings remain fixed while the user embeddings and the offering embeddings are learned by the machine learning embedding model.
8. The method of claim 1 , wherein the set of candidate offerings comprises one or more items offered by the online concierge system.
9. The method of claim 1 , further comprising rotating the user embeddings and the offering embeddings into a latent space of the search queries.
10. A method comprising:
at an online concierge system comprising one or more processors and one or more computer-readable media:
receiving, by the online concierge system, a search query from a user;
retrieving, by the online concierge system, offering embeddings for offerings, wherein the offering embeddings are generated using a machine learning embedding model, wherein an offering is offered by the online concierge system to the user;
calculating, by the online concierge system, a matrix operation product of a search query embedding for the search query and a user embedding for the user, wherein the user embedding is generated using the machine learning embedding model;
selecting, by the online concierge system, a set of offering embeddings based on distances of the offering embeddings in the set from the matrix operation product;
retrieving, by the online concierge system, a set of candidate offerings corresponding to the set of offering embeddings;
ranking, by the online concierge system, the set of candidate offerings using a second machine learning model; and
transmitting, by the online concierge system and based on the ranking, a subset of the set of candidate offerings to a user device for display to the user.
11. The method of claim 10 , further comprising training the machine learning embedding model, wherein training the machine learning embedding model comprises:
using historical data from a search log comprising triplets, each triplet including a search query, a user, and an item, wherein each triplet is labeled based on an action performed by the user on the item.
12. The method of claim 10 , wherein the user embeddings and the offering embeddings are learned by calculating an inner product of the search query embeddings, the user embeddings, and offering embeddings, and performing a learning to rank pairwise loss function.
13. The method of claim 10 , further comprising calculating, by the online concierge system, a predicted click through rate for each of the subset of the set of candidate offerings.
14. The method of claim 10 further comprising calculating, by the online concierge system, an effective cost per mile for each of the subset of the set of candidate offerings.
15. The method of claim 10 , wherein calculating the product comprises calculating a Hadamard product of the search query embeddings and the user embeddings.
16. The method of claim 10 , further comprising retrieving the search query embeddings from a search ranking module, wherein the search query embeddings remain fixed while the user embeddings and the offering embeddings are learned by the machine learning embedding model.
17. The method of claim 10 , wherein the set of candidate offerings comprises one or more items offered by the online concierge system.
18. The method of claim 10 , further comprising rotating the user embeddings and the offering embeddings into a latent space of the search queries.
19. The method of claim 10 , wherein the matrix operation product comprises a Hadamard product.