Retail product listing escalation event detection
The systems and methods provided for automatic retail product listing escalation event detection. The method includes aggregating content items from online resources, determining a topic classification of a content item based on a first machine learning (ML) model, determining an escalation classification of the content item based on a second ML model, generating a keyword list including a plurality of keywords from the content item and scores associated with each of the plurality of keywords based on the topic classification and the escalation classification of the content item, selecting a list of top keywords based on performing a keyword frequency analysis on a plurality of keyword lists associated with the plurality of content items, and providing a detected escalation event.
1 . A system for retail product listing escalation event detection, the system comprises:
a network interface for accessing a wide-area network;
a user interface device; and
a control circuit coupled to the network interface and the user interface device, the control circuit executes computer-readable instructions to:
aggregate a plurality of content items from the wide-area network from a plurality of online resources, including social media services, wherein the plurality of content items comprises social media content;
determine a topic classification of a content item based on a first machine learning (ML) model;
determine an escalation classification of the content item based on a second ML model;
generate, with an ML explanation engine, a keyword list comprising a plurality of keywords from the content item and scores associated with each of the plurality of keywords based on the topic classification from the first ML model and the escalation classification from the second ML model;
select, with a keyword analysis module, a list of top keywords based on performing a keyword frequency analysis on a plurality of keyword lists associated with the plurality of content items, including a plurality of social media content items; and
provide for display, via the user interface device, an escalation event detected by the second ML model, a topic associated with the escalation event detected by the first ML model, and one or more keywords associated with the topic determined based on the list of top keywords;
wherein the ML explanation engine uses a perturbation layer that provides perturbation as input to at least one of the first ML model and the second ML model based on an embedded layer, and
wherein the keyword list is determined based on model outputs from the perturbations generated by the perturbation layer.
2 . The system of claim 1 , wherein content data retrieved from the wide-area network via the network interface is processed through stemming and lemmatization to normalize text and remove stop words to generate the content item.
3 . The system of claim 1 , wherein content data retrieved from the wide-area network via the network interface is cleaned by replacing uniform resource locators (URLs), numbers, and/or emojis with text strings to generate the content item.
4 . The system of claim 1 , wherein the first ML model comprises an ML model for multiclass classification trained on content items with topics as categorization.
5 . The system of claim 1 , wherein the second ML model comprises a deep learning model for binary classification trained on content items with escalation classification as binary categorization.
6 . The system of claim 1 , wherein the ML explanation engine uses a normalized linear Gaussian perturbation layer for ML model interpretation.
7 . The system of claim 6 , wherein the ML explanation engine includes an embedded layer between input and the normalized linear Gaussian perturbation layer to provide additional context to the ML explanation engine.
8 . The system of claim 6 , wherein a score for a keyword has a positive value when the keyword is positively associated with escalation, and wherein the score for the keyword has a negative value when the keyword is negatively associated with escalation.
9 . The system of claim 1 , wherein the control circuit is further configured to:
provide, via the user interface device, an escalation review interface, wherein feedback received from the escalation review interface is used to further train the first ML model and/or the second ML model.
10 . The system of claim 1 , wherein the control circuit is further configured to:
identify at least one product based on the list of top keywords and product characteristics stored in a product database; and
generate an alert for removal of the at least one product via the user interface device.
11 . A method for retail product listing escalation event detection, the method comprises:
aggregating, via network interface for accessing a wide-area network, a plurality of content items from a plurality of online resources including social media services, wherein the plurality of content items comprises social media content;
determining, with a control circuit, a topic classification of a content item based on a first machine learning (ML) model;
determining, with the control circuit, an escalation classification of the content item based on a second ML model;
generating, with an ML explanation engine executed on the control circuit, a keyword list comprising a plurality of keywords from the content item and scores associated with each of the plurality of keywords based on the topic classification from the first ML model and the escalation classification from the second ML model;
selecting, with a keyword analysis module executed on the control circuit, a list of top keywords based on performing a keyword frequency analysis on a plurality of keyword lists associated with the plurality of content items, including a plurality of social media content items; and
providing for display, via an user interface device, an escalation event detected by the second ML model, a topic associated with the escalation event detected by the first ML model, and one or more keywords associated with the topic determined based on the list of top keywords;
wherein the ML explanation engine uses a perturbation layer that provides perturbation as input to at least one of the first ML model and the second ML model based on an embedded layer, and
wherein the keyword list is determined based on model outputs from the perturbations generated by the perturbation layer.
12 . The method of claim 11 , wherein content data retrieved from the wide-area network via the network interface is processed through stemming and lemmatization to normalize text and remove stop words to generate the content item.
13 . The method of claim 11 , wherein content data retrieved from the wide-area network via the network interface is cleaned by replacing URLs, numbers, and/or emojis with text strings to generate the content item.
14 . The method of claim 11 , wherein the first ML model comprises an ML model for multiclass classification trained on content items using topics as categorization.
15 . The method of claim 11 , wherein the second ML model comprises a deep learning model for binary classification trained on content items using escalation classification as binary categorization.
16 . The method of claim 11 , wherein the ML explanation engine uses a normalized linear Gaussian perturbation layer for ML model interpretation.
17 . The method of claim 16 , wherein the ML explanation engine includes an embedded layer between input and a perturbation layer to provide additional context to the ML explanation engine.
18 . The method of claim 16 , wherein a score for a keyword has a positive value when the keyword is positively associated with escalation, and wherein the score for the keyword has a negative value when the keyword is negatively associated with escalation.
19 . The method of claim 11 , further comprising:
providing, via the user interface device, an escalation review interface, wherein feedback received from the escalation review interface is used to further train the first ML model and/or the second ML model.