By Sonali Inamdar, Group Engineering Manager, Adaptive Intelligent Apps for Human Capital Management and Shekhar Agrawal, Director of Data Science, Adaptive Intelligent Apps for Human Capital Management
AI models are optimally useful for recommendations when they are trained against data with required characteristics and volume for the model to perform to acceptable outcomes. One of the common challenges around the uptake of AI-based recommendation solutions in production customer environments is the cold start problem.
Cold Start refers to the state of a system where the system is uninitialized. As the system completes a required initialization cycle, the system is warmed up and is optimally productive. Upon cold start, an AI-based recommendation system produces less than optimal recommendation outcomes. When the AI models complete the training phase, the system is considered warmed up and improved recommendation outcomes can be expected.
The term Cold Start is derived from cars: When the engine is cold, the car is not yet working so smoothly, but once the optimal temperature is reached, it works just fine. Similarly, in all recommender systems, the engine cannot provide best results until the conditions are not optimal for it to operate smoothly.
Refer to the following diagram for a representation of state transition from cold state to warm state for an AI-based recommendation system.
One option to improve the cold start outcomes for recommendation systems on Day 1 is to review feasible alternatives and augment the recommender system depending on the specific use-case.
For example, Oracle Adaptive Intelligence Best Candidates includes support for Enhanced Cold Start provided by the Adaptive Intelligence Cold Start Support Engine, which is a key differentiator of our Recruiting offering in implementing Best Candidates in production. This provides our customers improved recommendation benefits starting Day 1.
In the next few paragraphs, we will address two Cold Start problems and how we overcome them.
Product Cold Start is when the product is new and there are no reviews or history to measure the success of recommendations. User Cold Start is when there are relatively new users, and hence there is insufficient user history to personalize recommendations. Most companies use a popularity-based strategy to deal with user cold start problems. Popularity based recommendations use click-rank and browse-rank algorithms to measure popularity and then use it to recommend content to a new user. So, user can be recommended globally relevant content until personalization (locally relevant content) can be generated using user’s history. In Best Candidates we address both issues.
Best Candidates uses word embedding models to generate recommendations, that work by learning context from available historical data. The model is customer data agnostic as long as it has seen the relevant context. For example, if the model has seen healthcare jobs and resumes, irrespective of the source, it will be able to recommend candidates in the healthcare domain.
How did we address “Product Cold Start” problem?
Best Candidates delivers a pre-trained model based on a sufficiently large amount of available hiring data. As a result, the model is able to deliver useful recommendations with coverage of domains like technology, marketing, sales, and customer service on Day 1. Best Candidates is not dependent on customer data to deliver successful recommendations “out of the box”.
How did we address “User Cold Start” problem?
The performance of word embedding models depend on the amount and spread of context the model has been trained on. To enhance the domain coverage, the improved Best Candidates word embedding model has learned semantics and context from standard taxonomies like O*NET and ESCO, which serve as the base model for all customers. Thus, Best Candidates is able to deliver meaningful recommendations to customers in any domain from Day 1.
Cold Start Problems exists in every recommender system. Best Candidates delivers a sophisticated solution to resolve this problem. AI systems will continue to evolve and improve as they learn from additional data over a period of time. Customers can benefit from these improvements as the system transitions from a cold to warm state, providing enhanced personalized recommendations.
Explore Oracle HCM Cloud
How is your recruiting system doing in attracting the best candidates? Discover how Oracle HCM Cloud fosters a more human recruiting process with technology and speak with one of our experts tonight.
You Might Also Like