Application of a Meta - Learning (Zero , Few Shots Learning) framework for solving the User & Item - cold start problems in the realm of Recommender System
This presentation aims to find a novel solution framework for one of the critical problems for any form of recommender systems – Cold Start, when It comes to recommending the new items or dealing with new users. Traditionally such problems have been tackled by using content based filtering approach based on various user's or item's context features.
For any specific user or item, the cold start problem may create a serious bottlenecks in the recommender system by limiting its ability to generalize its recommendations for any unforeseen Users or Items. In other words, "Item cold start solution" helps in marketing the new product so that it reaches the right customers quickly. "User cold-start solution" helps in providing a personalized experience to the new users without having to wait for them to exhibit their preferences explicitly over a period of time.
This is where our proposed approach offers a solution based on meta – learning framework. Our solution framework explores methods like Zero (ZSL) or Few Shot(FSL) Learning to address this User and Item cold problem in the realm of any recommender system . FSL/ZSL paradigm aims at recognizing new categories of instances in absence/in presence of minimal amount of training instances. It tries to do that by providing a high-level description of the new categories that relate them to categories previously learned by the algorithm. This learning paradigm is inspired by the way any human being is able to identify a new object by just reading its description, exploiting the similarities between the description of the new object and previously learned concepts. So , in other words, FSL/ZSL makes an attempt to classify the data with very little or no labelled instances. This trait of FSL/ZSL makes it a very powerful strategy to deal with the classic cold start problem.
In our solution framework, we use a rich set of item attributes and user attributes along with a historic interaction(rating) matrix, to learn the semantic relationship between the users and the items respectively.
We intend to use deep network(s) to learn the user-item ratings based on these rich item/user attributes. This networks' weights inherently learn the latent preferential mapping that a typical user profile may exhibit towards any item in a generalized manner. When a new item or a new user is scored, we infer the best user-item mapping based on these learned networks, so that we can recommend the new item/user with no prior data. To showcase the efficacy of this novel approach we would compare the performance of this using standard evaluation metrics like Recall@K, NDCG(Net Discounted cumulative gain)@K, MAP(Mean of the Average Precision) against the benchmark SOTA (State-of-the-art) algorithms.
Outline/Structure of the Talk
- Introduction of the problem statement and overview of the dataset: In this section, we will share our POV about the problem we are aiming to solve and share an overview of the dataset being used.
- Basic overview of the cold start problem: Will share our perspective about cold start problem and why it is one of the major problems in the realm of recommender system.
5-7 mins in total
- Application of the Meta - Learning framework: In this section we will talk about our architecture of the meta - learning framework (like Zero-Shot/Few Shots learning) that we intend to use to solve this problem.
- Improvement over SOTA (State-of-the art) Architecture: We will present a comparison of the performance of our model against the SOTA architecture.
- Results and key findings: We will present a summarized version of the outcome and key findings from this area of research.
10-12 mins in total
- Conclusion: We will conclude with our POV about its application to solve the cold start problem in realm of Financial recommender system and future plan for further advancement.
3-5 mins in total
This methods offers a single window solution towards both the User and Item specific cold start problems in the domain of recommender system. This solution can be leveraged across various spectrum of the recommender system.This turns out to be quite effective for online learning paradigm where we don't have the luxury of collecting any data elements for training the ML models.
Anyone with specific interest in the realm of AI for Finance/Marketing, Recommender System, Application of Meta Learning
Prerequisites for Attendees
Basic understanding of Machine Learning, Rudimentary idea about recommender systems, idea about cold start problem. Basic understanding about the Deep neural network.