System and method for providing personalized item recommendations with personalized title
Systems and methods for providing personalized item recommendations with personalized title are disclosed. In some embodiments, each user and each item of a website is represented as an embedding in a same heterogeneous graph. Based on an anchor embedding representing an anchor item in the heterogeneous graph, a set of items are determined. At least one feature aspect is selected from a set of feature aspects based on at least one implicit intent of a user, each feature aspect characterizing a respective aspect of features common to the set of items. From the set of items, a ranked list of items are generated as recommended items based on the at least one feature aspect. A recommendation title describing the at least one feature aspect is generated for the recommended items. The recommended items and the recommendation title are transmitted to be displayed together with the anchor item to the user.
1 . A system, comprising:
a processor; and
a non-transitory memory storing instructions that, when executed, cause the processor to:
determine a set of items based on an anchor item to be displayed to a user,
determine a set of feature aspects each characterizing a respective aspect of features shared by the set of items,
select at least one feature aspect from the set of feature aspects,
generate, from the set of items, a ranked list of recommended items based on the at least one feature aspect,
generate a trained machine learning model based on historical recommendation data,
execute the trained machine learning model to generate, for the ranked list of recommended items, a recommendation list title describing the at least one feature aspect shared by the ranked list of recommended items,
re-train the trained machine learning model based on the recommendation list title, wherein the re-training updates parameters of the trained machine learning model,
store the parameters of the re-trained machine learning model in a database, and
transmit the ranked list of recommended items and the recommendation list title to be displayed together with the anchor item to the user.
2 . The system of claim 1 , wherein:
the anchor item is determined based on an interaction of the user with a website.
3 . The system of claim 1 , wherein the at least one processor is configured to:
represent each user of a website as a user embedding in a heterogeneous graph based on the trained machine learning model;
represent each item offered on the website as an item embedding in the heterogeneous graph based on the trained machine learning model; and
build, for each respective item having a respective item embedding in the heterogeneous graph, at least one nearest neighbor index indicating at least one nearest neighbor item whose item embedding is closest to the respective item embedding.
4 . The system of claim 3 , wherein the set of items is determined based on:
representing the anchor item as an anchor embedding in the heterogeneous graph based on the trained machine learning model;
searching in the heterogeneous graph, based on the anchor embedding, to identify K nearest neighbor items for the anchor item, wherein K is a predetermined positive integer; and
determining the set of items based on the K nearest neighbor items.
5 . The system of claim 4 , wherein the set of feature aspects is determined based on:
obtaining user aspect affinity data for the user, wherein the user aspect affinity data is obtained based on at least one item that is either interacted with by the user or nearest to a user embedding of the user in the heterogeneous graph; and
determining the set of feature aspects based on the user aspect affinity data and the set of items, wherein the set of feature aspects are top feature aspects shared by the set of items.
6 . The system of claim 5 , wherein the at least one feature aspect is selected from the top feature aspects based on at least one implicit intent of the user and the user aspect affinity data.
7 . The system of claim 6 , wherein:
the recommended items are ranked based on at least one of: their distances to the anchor item in the heterogeneous graph, or their relatedness's to the at least one feature aspect; and
a relatedness of each recommended item to the at least one feature aspect is measured based on a frequency that the at least one feature aspect is mentioned in user reviews of the recommended item.
8 . The system of claim 3 , wherein the trained machine learning model is trained based on:
positive and negative samples that are generated based on historical user transaction data and expert inputs;
user item affinity data indicating which user prefers which item based on historical user interaction data with respect to the website; and
item aspect relation data including feature aspects frequently mentioned in user reviews and/or item descriptions for each item.
9 . The system of claim 1 , further comprising:
execute the re-trained machine learning model to generate, for a second ranked list of recommended items, a second recommendation list title describing at least one feature aspect shared by the second ranked list of recommended items, and
transmit the second ranked list of recommended items and the second recommendation list title for display.
10 . A computer-implemented method, comprising:
determining a set of items based on an anchor item to be displayed to a user;
determining a set of feature aspects each characterizing a respective aspect of features shared by the set of items;
selecting at least one feature aspect from the set of feature aspects;
generating, from the set of items, a ranked list of recommended items based on the at least one feature aspect;
generating a trained machine learning model based on historical recommendation data;
executing the trained machine learning model to generate, for the ranked list of recommended items, a recommendation list title describing the at least one feature aspect shared by the ranked list of recommended items; and
re-training the trained machine learning model based on the recommendation list title, wherein the re-training updates parameters of the trained machine learning model,
storing the parameters of the re-trained machine learning model in a database,
transmitting the ranked list of recommended items and the recommendation list title to be displayed together with the anchor item to the user.
11 . The computer-implemented method of claim 10 , further comprising:
representing each user of a website as a user embedding in a heterogeneous graph based on the trained machine learning model;
representing each item offered on the website as an item embedding in the heterogeneous graph based on the trained machine learning model; and
building, for each respective item having a respective item embedding in the heterogeneous graph, at least one nearest neighbor index indicating at least one nearest neighbor item whose item embedding is closest to the respective item embedding.
12 . The computer-implemented method of claim 11 , wherein determining the set of items comprises:
representing the anchor item as an anchor embedding in the heterogeneous graph based on the trained machine learning model;
searching in the heterogeneous graph, based on the anchor embedding, to identify K nearest neighbor items for the anchor item, wherein K is a predetermined positive integer; and
determining the set of items based on the K nearest neighbor items.
13 . The computer-implemented method of claim 12 , wherein determining the set of feature aspects comprises:
obtaining user aspect affinity data for the user, wherein the user aspect affinity data is obtained based on at least one item that is either interacted with by the user or nearest to a user embedding of the user in the heterogeneous graph; and
determining the set of feature aspects based on the user aspect affinity data and the set of items, wherein the set of feature aspects are top feature aspects shared by the set of items.
14 . The computer-implemented method of claim 13 , wherein the at least one feature aspect is selected from the top feature aspects based on at least one implicit intent of the user and the user aspect affinity data.
15 . The computer-implemented method of claim 14 , wherein:
the recommended items are ranked based on at least one of: their distances to the anchor item in the heterogeneous graph, or their relatedness's to the at least one feature aspect; and
a relatedness of each recommended item to the at least one feature aspect is measured based on a frequency that the at least one feature aspect is mentioned in user reviews of the recommended item.
16 . The computer-implemented method of claim 11 , wherein the trained machine learning model is trained based on:
positive and negative samples that are generated based on historical user transaction data and expert inputs;
user item affinity data indicating which user prefers which item based on historical user interaction data with respect to the website; and
item aspect relation data including feature aspects frequently mentioned in user reviews and/or item descriptions for each item.
17 . A non-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by at least one processor, cause at least one device to perform operations comprising:
determining a set of items based on an anchor item to be displayed to a user;
determining a set of feature aspects each characterizing a respective aspect of features shared by the set of items;
selecting at least one feature aspect from the set of feature aspects;
generating, from the set of items, a ranked list of recommended items based on the at least one feature aspect;
generating a trained machine learning model based on historical recommendation data;
executing the trained machine learning model to generate, for the ranked list of recommended items, a recommendation list title describing the at least one feature aspect shared by the ranked list of recommended items; and
re-training the trained machine learning model based on the recommendation list title, wherein the re-training updates parameters of the trained machine learning model,
storing the parameters of the re-trained machine learning model in a database,
transmitting the ranked list of recommended items and the recommendation list title to be displayed together with the anchor item to the user.
18 . The non-transitory computer readable medium of claim 17 , wherein the instructions, when executed by the at least one processor, further cause the at least one device to perform operations comprising:
representing each user of a website as a user embedding in a heterogeneous graph based on the trained machine learning model, wherein the anchor item is determined based on an interaction of the user with the website;
representing each item offered on the website as an item embedding in the heterogeneous graph based on the trained machine learning model; and
building, for each respective item having a respective item embedding in the heterogeneous graph, at least one nearest neighbor index indicating at least one nearest neighbor item whose item embedding is closest to the respective item embedding.
19 . The non-transitory computer readable medium of claim 18 , wherein determining the set of items comprises:
representing the anchor item as an anchor embedding in the heterogeneous graph based on the trained machine learning model;
searching in the heterogeneous graph, based on the anchor embedding, to identify K nearest neighbor items for the anchor item, wherein K is a predetermined positive integer; and
determining the set of items based on the K nearest neighbor items.
20 . The non-transitory computer readable medium of claim 19 , wherein determining the set of feature aspects comprises:
obtaining user aspect affinity data for the user, wherein the user aspect affinity data is obtained based on at least one item that is either interacted with by the user or nearest to a user embedding of the user in the heterogeneous graph; and
determining the set of feature aspects based on the user aspect affinity data and the set of items, wherein the set of feature aspects are top feature aspects shared by the set of items, wherein:
the at least one feature aspect is selected from the top feature aspects based on at least one implicit intent of the user and the user aspect affinity data,
the recommended items are ranked based on at least one of: their distances to the anchor item in the heterogeneous graph, or their relatedness's to the at least one feature aspect, and
a relatedness of each recommended item to the at least one feature aspect is measured based on a frequency that the at least one feature aspect is mentioned in user reviews of the recommended item.