哪款打印机耐用又好用 DashCLIP: Leveraging multimodal models for generating semantic embeddings for DoorDash

Next, we illustrate the integration of the projected embeddings of the product e​m​b​e​dpembed_{p}, query e​m​b​e​dqembed_{q}, and their derived features with the ranking model. Our model architecture is shown in Figure 3: a DCN tower that behes the same as the current production model, and a new tower that takes the product and query embeddings from DashCLIP and their derived features as input. The intermediate outputs from the two towers are concatenated and passed through a few fully connected layers before the final sigmoid layer that produces the output prediction in the range from 0 to 1. In the offline experiment stage, we iterated over different architecture designs and found that this approach, which separates the existing dense and sparse features from the new embedding-related features using the two-towers, achieved the best performance. This architecture promotes the crossing between the different embeddings before interacting them with the existing features. The derived features are of two types: the similarity between embeddings and representations of user engagement history on the platform. The former is the cosine similarity of the embeddings such as c​o​s​i​n​e​(e​m​b​e​dp,e​m​b​e​dq)cosine(embed_{p},embed_{q}) and aims to capture the present user intent. One particular instance of the latter is the c​o​n​s​u​m​e​r​_​p​84​d​_​p​u​r​c​h​a​s​e​d​_​p​r​o​d​u​c​tconsumer\_p84d\_purchased\_product which is a list of product ids that the user purchased in the last 84 days on the platform. Then it retrieves the product embeddings using the product ids as indices from the pre-computed embedding table. A mean pooling erages the retrieved embeddings, and then pooled vector is fed to the model. Similarly, c​o​s​i​n​e​(e​m​b​e​dp,p​o​o​l​(p​u​r​c​h​a​s​e​d​_​p​r​o​d​u​c​t))cosine(embed_{p},pool(purchased\_product)) captures how relevant the product is to the user’s purchase profile.