37 might include description information, taxon information, price information, brand or designer information, or sales history of the item. The embedding model 33 generates and outputs an embedding 37 from the image data 20 and optional additional information. In one example, taxon information 38 (“T1”) is also represented in the embedding 37. The embedding 37 is represented by a one-dimensional array [S₁, S₂, . . . S_N]. For each item in inventory 19, an embedding is generated and stored in the storage system 18.

FIG. 4 is a diagram that shows operation of the outfit recommender system 10. A customer 14 accesses an electronic commerce platform associated with the provider entity 15. In this example, the electronic commerce platform is an online website or mobile application that communicates with the application server system 12 over the network 13 and is accessed via the user device 11. The user device 11 is not shown in FIG. 4. The online website or mobile application associated with the electronic commerce platform includes various product detail pages 16 allowing the customer 14 to view and purchase any items of interest. In this example, the customer 14 selects to view the image for a selected item 20. The outfit recommender 17 receives the request and performs various novel steps (labeled “A” through “E”) to render a PDP 21 showing the image of the selected item 20 along with recommended complementary items 25.

First, an identifier (“ID”) associated with selected item 20 is received onto the outfit recommender 17 (labeled as “A”). The identifier is a unique key that is used to identify and track items internally within the application server system 12. The identifier is also referred to as a “product ID”, “item ID”, or “ID” depending on the implementation. In this example, the identifier is received onto the outfit recommender 17 as a HTTP GET request in response to the customer 14 selecting to view selected item 20.

Next, an embedding associated with the selected item 20 is obtained (labeled as “B”). Each embedding is a low-dimensional vector that is generated using an embedding model. The embedding model reduces dimensionality of image data and any other input data associated with the item. The embedding represents image information, taxon, price, designer or source information associated with the item, description or characteristic information of the item, sale information, market characteristics associated with the item (e.g. time on marketplace or supply-chain information), filter information (e.g. customer profile characteristics), or other information useful in grouping or classifying items.

In one embodiment, embeddings 36 for each item in available inventory 19 are generated beforehand and are stored in the storage system 18 as explained in connection with FIG. 3. Each image of available inventory 19 shown in FIG. 4 represents an embedding. The storage system 18 need not store actual images. Instead, the images represent data embeddings and are only shown throughout the diagrams for purposes of easing explanation. The embedding for an item is accessed using the ID for the item. In this example, embedding 37 for selected item 20 is stored in the storage system 18. In another embodiment, embeddings for each item in available inventory 19 are generated dynamically rather than being generated and stored before the request.

Next, a preconfigured item most visually similar to the selected item 20 is obtained (labeled as “C”). The preconfigured item has associated preconfigured complementary items that together form a preconfigured outfit. Embeddings 26 for preconfigured outfits are stored in the storage system 18. Each image of the preconfigured outfits shown in FIG. 2 represents an embedding. Each embedding is grouped together with other embeddings that together form a preconfigured outfit. In this example, the preconfigured item 27 is determined to be most visually similar to the selected item 20.

In this embodiment, visual comparison is performed by computing cosine similarity values between the embedding 37 of the selected item 20 to each embedding 26 of the preconfigured outfits having the same taxon as the selected item 20. The cosine similarity values are sorted and ranked to obtain a most visually similar item. In the instant example, the selected item 20 is a taxon of type “sweater” labeled “T1”. As such, the embedding 37 of the selected item 20 is compared to embeddings 26 in the preconfigured outfits also having the taxon of type “sweater” labeled “T1”. This reduces the number of comparison operations that must be performed and improves operation and efficiency of system 10. The embedding 34 of preconfigured item 27 is determined to have the greatest cosine similarity value in this example. The embeddings 39 of preconfigured complementary items of matching outfit #4 will be used in generating the recommended complementary items.

Next, complementary items in available inventory most visually similar to the preconfigured complementary items are obtained (labeled as “D”). In this example, each embedding 39 of the preconfigured complementary items is compared via cosine similarity to embeddings 36 of available inventory having the same taxon. Computed cosine similarity values are ranked and sorted to obtain a most visually similar complementary item available in inventory 19. This process is performed for each preconfigured complementary item in the matching outfit. For example, an embedding 32 for sunglasses in matching outfit #4 is compared to each embedding in available inventory 19 having a taxon of type “sunglasses” labeled as “T3”. This process is repeated for each of the remaining preconfigured complementary items 28-31 until the most visually similar complementary items in available inventory 19 are identified.

Next, the identified complementary items are presented to the customer along with the selected item 20 (labeled as “E”). In this example, the selected item 20 is presented to the customer 14 on the PDP 21 along with the images of recommended complementary items 25. Customer 14 is able to view PDPs for each complementary item 25 to view additional information, purchase the item, or view other complementary items on the respective PDP for the complementary item of interest. Upon selecting a complementary item, a PDP is presented along with new complementary items identified via the novel process described in the present disclosure.

FIG. 5 is a diagram showing how the outfit recommender 17 identifies an item of the preconfigured outfits most visually similar to the image of the selected item 20. The selected item 20 has a unique identifier (“ITEM_1” ID) 40, taxon information 38 (“T1” identifies “sweaters”), and embedding 37 ([S₁, S₂, . . . S_N]). The embedding 37 for the selected item 20 is compared to each embedding 26 of the preconfigured outfits of the same taxon to obtain a most visually similar preconfigured item of the preconfigured outfits. In this embodiment, embeddings are compared to each other via cosine similarity. The cosine similarity values are ranked and sorted. The largest cosine similarity value indicates the most visually similar preconfigured item. In this example, embedding 34 corresponding to preconfigured item 27 of outfit #4 yields the greatest cosine similarity value of “0.83”, indicating that preconfigured item 27 is the most visually similar to selected item 20. Preconfigured items of different taxons are not compared to the selected item 20, thereby reducing utilization of computational resources and increasing operating speed. In other embodiments, techniques other than cosine similarity are used to compare image or embedding data to identify most visually similar items.

FIG. 6 is a diagram showing how the outfit recommender 17 identifies complementary items in available inventory 19 most visually similar to preconfigured complementary items 39. After determining that preconfigured item 27 is the most visually similar item in the preconfigured outfits to selected item 20, each of the preconfigured complementary items 39 associated with the preconfigured item 27 in outfit #4 is used to obtain complementary items in available inventory that complement the selected item 20. The visual comparison for each preconfigured complementary item is shown in FIG. 6 and described below.

Embedding 41 corresponding to preconfigured complementary item 28 (“shoes”) of outfit #4 is compared to each embedding 36 of available inventory 19 having the same taxon (“T4”). Embedding 72 corresponding to item 73 yields the greatest cosine similarity value of “0.93” as compared to other cosine similarity computations for other available items of the same taxon (“T4”). This indicates that item 73 in available inventory 19 is the most visually similar to preconfigured complementary item 28.

Embedding 42 corresponding to preconfigured complementary item 29 (“bag”) of outfit #4 is compared to each embedding 36 of available inventory 19 having the same taxon (“T2”). Embedding 74 corresponding to item 75 yields the greatest cosine similarity value of “0.86” as compared to other cosine similarity computations for other available items of the same taxon (“T2”). This indicates that item 75 in available inventory 19 is the most visually similar to preconfigured complementary item 29.

Embedding 43 corresponding to preconfigured complementary item 30 (“earrings”) of outfit #4 is compared to each embedding 36 of available inventory 19 having the same taxon (“T5”). Embedding 76 corresponding to item 77 yields the greatest cosine similarity value of “0.85” as compared to other cosine similarity computations for other available items of the same taxon (“T5”). This indicates that item 77 in available inventory 19 is the most visually similar to preconfigured complementary item 30.

Embedding 44 corresponding to preconfigured complementary item 31 (“pants”) of outfit #4 is compared to each embedding 36 of available inventory 19 having the same taxon (“T6”). Embedding 78 corresponding to item 79 yields the greatest cosine similarity value of “0.89” as compared to other cosine similarity computations for other available items of the same taxon (“T6”). This indicates that item 79 in available inventory 19 is the most visually similar to preconfigured complementary item 31.

Embedding 45 corresponding to preconfigured complementary item 32 (“sunglasses”) of outfit #4 is compared to each embedding 36 of available inventory 19 having the same taxon (“T3”). Embedding 80 corresponding to item 81 yields the greatest cosine similarity value of “0.91” as compared to other cosine similarity computations for other available items of the same taxon (“T3”). This indicates that item 81 in available inventory 19 is the most visually similar to preconfigured complementary item 32. Complementary items 73, 75, 77, 79, and 81 are presented to the customer 14 as available complementary items on the PDP 21.

FIG. 7 is an equation 52 for computing cosine similarity between two embeddings. Equation 52 measures the similarity between two vectors. Embeddings are compared to each other using equation 52. In equation 52, “A” represents a first embedding and “B” represents a second embedding. The resulting computation of equation 52 indicates how similar two embeddings are to each other. It is appreciated that cosine similarity is but one technique for comparing similarity between two embeddings and that other embodiments use different techniques to compare embeddings.

FIG. 8 is a detailed diagram of the outfit recommender 17. The outfit recommender 17 includes a network interface 53, a processor 54, a memory 55, and a communication link or bus 56. The memory 55 includes an amount of computer readable instructions 57 and a routing table 58. The instructions 57 perform the novel method to obtain complementary items from dynamically changing available inventory based on preconfigured outfits. The routing table 58 is used to receive and process representational state transfer application programming interface (REST API) requests received over a network. In one embodiment, various components of the outfit recommender 17 are realized as cloud computing resources, such as Google Cloud Platform (GCP), Amazon Web Services (AWS), or Microsoft Azure, that operate over a distributed network.

FIG. 9 is a detailed diagram of storage system 18. The storage system 18 comprises an available inventory data structure 59, a preconfigured outfit data structure 60, and an elasticsearch engine 61. The available inventory data structure 59 stores information for available inventory including an item identifier (ID), taxon information, and embedding information for each item. Data in the available inventory data structure 59 changes dynamically as items are sold or added to inventory 19. The preconfigured outfit data structure 60 stores information for preconfigured outfits including a preconfigured item identifier (ID), an outfit identifier (ID), taxon information, and embedding information for each preconfigured item. The preconfigured outfit data structure 60 tends to remain static but can be modified to add new outfit configurations or to remove or modify existing outfits. The elasticsearch engine 61 provides an API accessible by the outfit recommender 17 to obtain and compare embeddings for items in available inventory and preconfigured outfits. The elasticsearch engine 61 is a distributed, RESTful search and analytics engine available from elastic.co/elasticsearch.

FIG. 10 is a detailed diagram of a routing table 58. The routing table 58 indicates how REST API requests are to be handled by the outfit recommender 17. The routing table 58 shown in FIG. 10 includes one entry for an HTTP (HyperText Transfer Protocol) GET request. When a product ID is received onto the outfit recommender 17 via an HTTP GET request, the outfit recommender 17 processes the request and returns complementary items via the novel process.

FIG. 11 is a flowchart of a method 100 in accordance with at least one novel aspect. In a first step (step 101), an identifier (ID) for a selected item is received onto an outfit recommender. For example, in FIG. 2, the PDP 21 for the selected item 20 accessed by the customer 14 sends the corresponding ID to an outfit recommender 17 via a network 13.

In a second step (step 102), an embedding of the selected item is obtained. For example, in FIG. 4, the embedding 26 for the selected item 20 is stored and accessed via a storage system 18.

In a third step (step 103), a preconfigured item most visually similar to the selected item is obtained. The preconfigured item has associated preconfigured complementary items that, together, form a preconfigured outfit. For example, in FIG. 4, an embedding 26 of the selected item 20 is compared (e.g., via cosine similarity) to embeddings of each preconfigured item.

In a fourth step (step 104), complementary items in available inventory most visually similar to the preconfigured complementary items are obtained. For example, in FIG. 4, each embedding 29 of the preconfigured complementary items is compared (e.g., via cosine similarity) to embeddings of each available item.

In a fifth step (step 105), the recommended complementary items 25 are presented along with the selected item 20. For example, in FIG. 4, recommended complementary items 25 are presented on the PDP 21 to customer 14.

FIG. 12 is a diagram showing how recommended complementary items are optionally filtered based on filter information 120 and profile information 130. In one embodiment, the storage system 18 stores filter information 120 and profile information 130 in addition to the embedding information 26 of preconfigured outfits and embedding information 36 of available inventory 19. The filter information 120 and profile information 130 are used by the outfit recommender 17 to determine items to include or exclude from the filtered complementary items 111. The filter information 120 is generated by analyzing individual customer behavior, consumer trends, price data, provider or source data, sales data, market trends, or other information usable to drive sales. The profile information 130 is generated by the outfit recommender 17, by customer input, or by a combination of both. For example, a customer may optionally select to view complementary items within a certain price range or by certain designers.

After customer 14 selects to view details for selected item 112, PDP 113 is presented showing information in connection with selected item 112 and recommended complementary items 111. The recommended complementary items 111 presented to customer 14 include items 114 (“purse”), 115 (“earrings”), 116 (“sunglasses”), and 117 (“shoes”). Items 118 (“sunglasses”) and 119 (“shoes”) are filtered out and not presented to the customer 14, even though they were deemed by the outfit recommender 17 to have been acceptable recommended item options. Items 118 and 119 are filtered out based on the filter information 120 and/or profile information 130. The filter information 120 and/or profile information 130 indicates that item 116 is more likely to result in a sale than item 118. The outfit recommender 17 utilizes filter information 120 and/or profile information 130 in determining that item 117 is more likely to result in a sale than item 119. In one example, items 118 and 119 are filtered out because they are not consistent with the price range (“<$200”) or designer preference (“D1, D2, D3, . . . ”) set by customer 14 or determined by outfit recommender 17 to be a preference of the customer 14. In another example, items 118 and 119 are filtered out as being less likely to result in sales after considering customer behavior, consumer trends, price data, provider or source data, sales data, market trends, or other information usable to drive sales.

In one embodiment, the filter information 120 and profile information 130 are applied after the similarity analysis is performed as shown in FIGS. 5 and 6. After the recommended items are obtained, they are filtered using the filter information 120 and profile information 130 to obtain filtered complementary items 111. The filtered complementary items are presented to the customer 14. In another embodiment, the filter information 120 and profile information 130 are stored within embeddings, such as embedding information 26 and 36. The similarity analysis performed in connection with FIGS. 5 and 6 incorporates the filter and profile information 120 and 130 and yields filtered complementary items without requiring any further filtering.

FIG. 13 is a diagram of an event based personalized item generation system 1304 in accordance with another embodiment. The system 1304 receives available inventory information 1301, user information 1302, and event information 1303 to generate personalized item recommendations 1305. The information can be obtained directly or through a network 1306. For example, in one embodiment, a user enters user information 1302 about an event the user would like to attend, such as a fashion show event. The user also specifies information such as event location, time of year, the user's clothing and shoe sizes, budget, and any other relevant user information. In one embodiment, the user's information is general in nature, such as any events within a city or state. In another embodiment, the user's information is more specific, such as a particular type of event, in a particular city, within a particular month during the year.

The system 1304 accesses the event information 1303 based on the user's input to determine one or more events that meet the user's preferences. The user may select one or multiple events from all the events that meet the user's criteria. Each event that is selected has associated event information, such as location, season of year, weather conditions, and any other relevant event information.

Using the information gathered about the event, the system 1304 accesses the available inventory 1301 to discover items that are available that can be recommended to the user. In one embodiment, user information, such as user size, budget, purchase history, personal color preferences, and any other user information is used to discover available items that can be recommended to the user for the selected event(s). The available items may include clothing, shoes, hats, jewelry and any other personal items.

In one embodiment, the system 1304 ranks the discovered items and provides the top ranked items (e.g., top 5 items) to the user as recommended items 1305. For example, as illustrated in FIG. 13, a fashion show event has been selected, and based on information about the event, user, and available inventory, the system 1304 recommends the top 5 ranked dresses for the user to wear to the event. The total number of recommended items is selectable and the user can select one of the items and receive further outfit recommendations for the selected item as described above.

FIG. 14 is a detailed diagram of the event based personalized item generation system 1304. In one embodiment, the system 1304 comprises a computing instance 1401 and a storage system 1404. The computing instance 1401 includes an application program interface (API) 1402 and an event based personalized item generator 1403. The storage system 1404 includes training data 1405 and models 1406.

The generator 1403 utilizes a variety of embedding models 1406 that are trained using training data 1405. The models are used to generate embeddings of inventory 1301, user 1302, and event 1303 information that is received through the API 1402. The generator 1403 combines the embeddings to form a matched embedding that is used to determine available inventory items that match the user's event. Matching items that are discovered in inventory are ranked and then the top ranking items are provided as recommended items 1308.

FIG. 15 is a functional diagram 1500 of an event based personalized item generation system. In an embodiment, system 1500 includes an event based personalized item generator function 1501, user input function 1502, event search function 1503, event information function 1504, and item inventory function 1505.

During operation, the user function 1502 provides user information 1506 comprising one or more of an event, location, user gender, item price range, user clothing size, and user shoe size. The information 1506 provided by the user input function 1502 is not limited to the user information shown.

The event based personal item generator function 1501 receives the user information 1506 and provides some or all of the information 1507 to the event search function 1503. For example, in one embodiment, the type of event and location are provided to the event search function 1503 as indicated by path 1507. The search function 1508 generates search results 1508 based on the received inputs and provides the search results to the generator function 1501. The search result includes event results that meet the event type and location requirements.

Based on the event results, selected events are determined by the generator 1501. For example, the event results can be filtered based on user input to determine selected events of interest to the user.

The selected events 1509 are provided to the event information function 1504 which determines event information 1510 such as location, weather, season, and/or temperature at the location of the event, which are returned to the generator 1501.

Based on the event information 1510, the generator 1501 generates product categories 1511 that are processed by the inventory function 1505 to access the available inventory to obtain available inventory items 1512 in the desired product categories. The generator 1501 then ranks and selects the top ranked items as recommended items 1513 for presentation to the user.

FIG. 16 is a system architecture 1600 of an event based personalized item generation system. The architecture 1600 comprises an event model 1601, item model 1602, location model 1603, and a user model 1604. The architecture 1600 also includes a domain matching model 1605, available inventor model 1606 and ranking model 1607.

During operation, the models receive a variety of information and generate associated embeddings. For example, the models implement one or more algorithms, such as XGBOOST, support vector learning machines (SVM), visual geometry group (VGG), and/or residual neural network (RESNET) to generate embeddings from received information.

The event model 1601 receives event information that is derived from surveys, scraped, or otherwise obtained. The event model 1601 processes the received information to generate an event embedding 1613 that represents the event information that is received.

The item model 1602 receives item information, such as types of clothing, accessories, shoes and other item information. The item model 1602 processes the received information to generate an item embedding 1613 that represents the item information that is received.

The location model 1603 receives location information of events, such as city, state, climate, culture, and other location information. The location model 1603 processes the received information to generate a location embedding 1615 that represents the location information that is received.

The user model 1604 receives user information, such as preferences, colors, size, budget, preferred events, purchase history and other user information. The user model 1604 processes the received information to generate a user embedding 1616 that represents the user information that is received.

The domain matching model 1605 receives the embeddings from the event, item, location, and user models. The domain matching model 1605 processes the received embeddings to generate a domain matching embedding 1616 that represents the information that is received in the various embeddings. For example, in one embodiment, the domain matching model 1605 implements a principal component analysis (PCA) algorithm to (1) standardize the range of continuous initial variables, (2) compute the covariance matrix to identify correlations, (3) compute the eigenvectors and eigenvalues of the covariance matrix to identify the principal components, (4) create a feature vector to decide which principal components to keep, and (5) recast the data along the principal components axes.

The available inventory model 1606 receives information about available inventory, such as available clothing, shoes, hats, jewelry, and other inventor items. The available inventory model 1606 processes the received inventory information to generate an available inventory embedding 1617 that represents the inventory information that is received.

The ranking model 1607 receives the available inventory embedding 1617 and the domain matching embedding 1612