Complementary product recommendation systems
Generally, the present disclosure relates to methods and systems for generating complementary product recommendations. In some example aspects, human-identified complementarity of a subset of products can be used to train a neural network, which is in turn used to generate pairwise complementarity values for items. Based on such values, complementary items can be identified.
1. A method of generating a set of one or more complementary items associated with an item, the method comprising:
receiving data identifying one or more pairs of complementary items from within an item collection the received data including an item description for each complementary item of the one or more pairs;
generating a model of item complementarity based on:
performing a vectorization of a pair of item descriptions from the received data to generate text embeddings that are indicative of similarity between the pair of item descriptions, wherein performance of the vectorization produces vectorized pairs of item descriptions; and
performing classification of the vectorized pairs of item descriptions to determine an extent of complementarity by scoring the vectorized pairs of item descriptions using a Siamese network, the Siamese network receiving as its input the generated text embeddings and having been previously trained on known pairs of complementary items and non-complementary items, wherein the Siamese network projects the generated text embeddings into an N-dimensional vector space where the generated text embeddings for pairs of complementary items are located closer in space than generated text embeddings for pairs of non-complementary items;
receiving a selection of a seed item from within the item collection; and
in response to selection of the seed item, identifying one or more modeled complementary items,
wherein at least one of the seed item or the one or more modeled complementary items is not included in the one or more pairs of complementary items included in the received data.
2. The method of claim 1 , further comprising:
generating complementarity scores between the seed item and each of a plurality of items within the item collection; and
based on a complementarity score between the seed item and an item from among the plurality of items being above a predetermined complementarity threshold, including the item from among the plurality of items within the one or more modeled complementary items.
3. The method of claim 1 , wherein the Siamese model utilizes a Contrastive Loss function according to the following:
L ( W ,( Y,X 1 ,X 2) i )= Yi ( DWi )2+(1− Yi )(max(0,1− DWi ))2,
wherein (Y, X1, X2)′ is a labeled sample pair, Yi is a label value of the sample pair, and Di is a Euclidean distance between the pair of points.
4. The method of claim 3 , wherein the Yi label value is a selected value of 0 for known non-complementary items and 1 for known complementary items.
5. The method of claim 1 , further comprising outputting a listing of the one or more modeled complementary items.
6. The method of claim 5 , wherein outputting the listing comprises providing the listing to a retail web server.
7. The method of claim 5 , wherein outputting the listing comprises generating a user interface including a display of the one or more modeled complementary items and the seed item from a retail web site server.
8. A system for generating a set of one or more complementary items associated with an item, the system comprising:
a computing system comprising a processor operatively connected to a memory, the memory storing computing instructions which, when executed by the processor, cause the computing system to:
generate a model of item complementarity based at least in part on data identifying one or more pairs of complementary items and item descriptions of each of the items included in an item collection, the one or more pairs of complementary items being a subset of all items included in the item collection, the received data including an item description for each complementary item of the one or more pairs, the generating of the model of item complementarity including:
performance of a vectorization of a pair of item descriptions to generate text embeddings that are indicative of similarity between the pair of item descriptions, wherein performance of the vectorization produces vectorized pairs of item descriptions; and
performance of a classification of the vectorized pairs of item descriptions to determine an extent of complementarity by scoring the vectorized pairs of item descriptions using a Siamese network, the Siamese network receiving as its input the generated text embeddings and having been previously trained on known pairs of complementary items and non-complementary items, wherein the Siamese network projects the generated text embeddings into an N-dimensional vector space where the generated text embeddings for pairs of complementary items are located closer in space than generated text embeddings for pairs of non-complementary items;
receive a selection of a seed item from within the item collection; and
in response to selection of the seed item, identify one or more modeled complementary items,
wherein at least one of the seed item or the one or more modeled complementary items is not included in the one or more pairs of complementary items included in the received data.
9. The system of claim 8 , wherein the computing system comprises a plurality of computing devices including a retail web site server.
10. The system of claim 9 , wherein the instructions comprise a complementary products generation application accessible from the retail web site server via an application programming interface (API).
11. The system of claim 10 , wherein the complementary products generation application includes instructions which cause the computing system to:
generate complementarity scores between the seed item and each of a plurality of items within the item collection; and
based on a complementarity score between the seed item and an item from among the plurality of items being above a predetermined complementarity threshold, include the item from among the plurality of items within the one or more modeled complementary items.
12. The system of claim 10 , wherein the retail web site server is further configured to generate a user interface displaying the one or more modeled complementary items.
13. The system of claim 10 , wherein the retail web site server is configured to provide, to the complementary products generation application via the API, an item identifier of the seed item.
14. A system for generating a set of one or more complementary items associated with an item, the system comprising:
a computing system comprising a plurality of computing devices, the plurality of computing devices including a retail web site server and a complementary item computing device, the computing system storing instruction which, when executed cause the computing system to:
generate a model of item complementarity based at least in part on data identifying one or more pairs of complementary items and item descriptions of each of the items included in an item collection, the one or more pairs of complementary items being a subset of all items included in an item collection, the received data including an item description for each complementary item of the one or more pairs, the generating of the model of item complementarity including:
performance of a vectorization of a pair of item descriptions to generate text embeddings that are indicative of similarity between the pair of item descriptions, wherein performance of the vectorization produces vectorized pairs of item descriptions;
performance of a classification of the vectorized pairs of item descriptions to determine an extent of complementarity by scoring the vectorized pairs of item descriptions using a Siamese network, the Siamese network receiving as its input the generated text embeddings and having been previously trained on known pairs of complementary items and non-complementary items, wherein the Siamese network projects the generated text embeddings into an N-dimensional vector space where the generated text embeddings for pairs of complementary items are located closer in space than generated text embeddings for pairs of non-complementary items;
receive a user input identifying a seed item from within the item collection;
provide the selection of the seed item to the complementary item computing device via an API;
in response to receiving of the seed item, identify, at the complementary item computing device, one or more modeled complementary items, at least one of the seed item or the one or more modeled complementary items not being included in the one or more pairs of complementary items included in the received data; and
generate, at the retail web site server, a user interface displaying the one or more modeled complementary items.
15. The system of claim 14 , wherein the user interface further displays the seed item.