Abstract: An online concierge system trains a classification model as a domain adversarial neural network from training data labeled with source classes from a source domain that do not overlap with target classes from a target domain output by the classification model. The online concierge system maps one or more source classes to a target class. The classification model extracts features from an image, classifies whether an image is from the source domain or the target domain, and predicts a target class for an image from the extracted features. The classification model includes a gradient reversal layer between feature extraction layers and the domain classifier that is used during training, so the feature extraction layers extract domain invariant features from an image.

Abstract: A server receives a plurality of product data entries from a plurality of retailer computing systems. Each product data entry includes a product identifier uniquely identifying a corresponding physical product and descriptive data of the corresponding physical product. A subset of the plurality of product data entries having a same product identifier is determined. An embedding vector representative of a product data entry in the subset is pairwise compared with each of respective embedding vectors representative of other product data entries in the subset other than the product data entry to compute respective vector similarity metrics. A pooled semantic similarity metric for the product data entry based on the computed respective vector similarity metrics. It is determined whether the product data entry is an outlier in the subset based on the pooled semantic similarity metric. A notification is transmitted to a client device of a user based on the determination.

Abstract: An online concierge system trains a classification model as a domain adversarial neural network from training data labeled with source classes from a source domain that do not overlap with target classes from a target domain output by the classification model. The online concierge system maps one or more source classes to a target class. The classification model extracts features from an image, classifies whether an image is from the source domain or the target domain, and predicts a target class for an image from the extracted features. The classification model includes a gradient reversal layer between feature extraction layers and the domain classifier that is used during training, so the feature extraction layers extract domain invariant features from an image.

Abstract: 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.

Abstract: An online concierge system uses a co-engagement graph to assign attribute values to items for which those attribute values are uncertain. A co-engagement graph is a graph with nodes that represent items and edges that represent co-engagement between items. The online concierge system generates a co-engagement graph for a set of items based on item engagement data and item data for the items. The set of items includes items for which the online concierge system has an attribute value for a target attribute and items for which the online concierge system does not have an attribute value for the target attribute. The online concierge system identifies a node that corresponds to an unknown item and identifies a node connected to that first node that corresponds to a known item. The online concierge system assigns the attribute value for the known item to the unknown item.

Abstract: A server receives a plurality of product data entries from a plurality of retailer computing systems. Each product data entry includes a product identifier uniquely identifying a corresponding physical product and descriptive data of the corresponding physical product. A subset of the plurality of product data entries having a same product identifier is determined. An embedding vector representative of a product data entry in the subset is pairwise compared with each of respective embedding vectors representative of other product data entries in the subset other than the product data entry to compute respective vector similarity metrics. A pooled semantic similarity metric for the product data entry based on the computed respective vector similarity metrics. It is determined whether the product data entry is an outlier in the subset based on the pooled semantic similarity metric. A notification is transmitted to a client device of a user based on the determination.

Abstract: An online concierge system uses a domain-adaptive suggestion module to score products that may be presented to a user as suggestions in response to a user's search query. The domain-adaptive suggestion module receives data that is relevant to scoring products as suggestions in response to a search query. The domain-adaptive suggestion module uses one or more domain-neutral representation models to generate a domain-neutral representation of the received data. The domain-neutral representation is a featurized representation of the received data that can be used by machine-learning models in the search domain or the suggestion domain. The domain-adaptive suggestion module then scores products by applying one or more machine-learning models to domain-neutral representations generated based on those products. By using domain-neutral representations, the domain-adaptive suggestion module can be trained based on training examples from a similar prediction task in a different domain.

Abstract: A masked language model is used to predict an attribute of an object, such as a physical item or product based on the predicted value of a masked token. The masked language model may be trained on a general corpus of text for the language, such that the masked language model learns context and text token relationships. Information about the object may then be added to a query template that structures the item information in an attribute query that may be interpretable by the masked language model to provide a resulting token related to the provided information or to confirm or reject an attribute specified in the query template.

Abstract: 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.

Abstract: An online concierge system uses a domain-adaptive suggestion module to score products that may be presented to a user as suggestions in response to a user's search query. The domain-adaptive suggestion module receives data that is relevant to scoring products as suggestions in response to a search query. The domain-adaptive suggestion module uses one or more domain-neutral representation models to generate a domain-neutral representation of the received data. The domain-neutral representation is a featurized representation of the received data that can be used by machine-learning models in the search domain or the suggestion domain. The domain-adaptive suggestion module then scores products by applying one or more machine-learning models to domain-neutral representations generated based on those products. By using domain-neutral representations, the domain-adaptive suggestion module can be trained based on training examples from a similar prediction task in a different domain.

Abstract: To train an embedding-based model to determine relevance between items and queries, an online system generates training data from previously received queries and interactions with results for the queries. The training data includes positive training examples including a query and an item with which a user performed a specific interaction after providing the query. To generate negative training examples for the query to include in the training data, the online system determines measures of similarity between items with which the specific interaction was not performed and the query. The online system may weight a loss function for the embedding-based model by the measure of similarity for a negative example, increasing the effect of a negative example including a query and an item with a larger measure of similarity. In other embodiments, the online system selects negative training examples based on the measures of similarities between items and queries in pairs.

Abstract: An online concierge system selects content for presentation to a user by using a product scoring engine. The product scoring engine generates a user embedding for user data and a query embedding for query data. The product scoring engine generates an anchor embedding based on the user embedding and the query embedding, where the anchor embedding is an embedding in a product embedding space. The product scoring engine compares the anchor embedding to a set of product embeddings to score a set of products for presentation to a user.

Abstract: An online concierge system uses a domain-adaptive suggestion module to score products that may be presented to a user as suggestions in response to a user’s search query. The domain-adaptive suggestion module receives data that is relevant to scoring products as suggestions in response to a search query. The domain-adaptive suggestion module uses one or more domain-neutral representation models to generate a domain-neutral representation of the received data. The domain-neutral representation is a featurized representation of the received data that can be used by machine-learning models in the search domain or the suggestion domain. The domain-adaptive suggestion module then scores products by applying one or more machine-learning models to domain-neutral representations generated based on those products. By using domain-neutral representations, the domain-adaptive suggestion module can be trained based on training examples from a similar prediction task in a different domain.

Abstract: To simplify retrieval of items from a database that at least partially satisfy a received query, an online concierge system trains a model that outputs scores for items from the database without initially retrieving items for evaluation by the model. The online concierge system pre-trains the model using natural language inputs corresponding to items from the database, with a natural language input including masked words that the model is trained to predict. Subsequently, the model is refined using multi-task training where a task is trained to predict scores for items from the received query. The online concierge system selects items for display in response to the received query based on the predicted scores.

Abstract: 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.

Abstract: An online concierge system trains a classification model as a domain adversarial neural network from training data labeled with source classes from a source domain that do not overlap with target classes from a target domain output by the classification model. The online concierge system maps one or more source classes to a target class. The classification model extracts features from an image, classifies whether an image is from the source domain or the target domain, and predicts a target class for an image from the extracted features. The classification model includes a gradient reversal layer between feature extraction layers and the domain classifier that is used during training, so the feature extraction layers extract domain invariant features from an image.