Multi-model Account Sequence Recommender for Robust Opportunity Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification methods for determining the likelihood of a user making a yes/no decision based on user features are not robust enough to handle outlier noise and overfitting, particularly in complex data processing applications.
Innovation Solution
A combination of models including a reinforcement learning model and an account propensity model, along with trained recurrent neural networks (RNNs) and hidden Markov models (HMMs), is used to identify optimal sequences of individuals to contact for account/product pairs, maximizing the likelihood of a winning outcome.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If random forest classification is used for binary classification, then robustness to outlier noise and reduction of overfitting is improved, but model complexity and computational requirements increase
Solution Approach 1:
The random forest model segments the classification task into multiple independent decision trees, each trained on a different bootstrap sample of the data. Each tree partitions the feature space through binary decisions, and the final classification is obtained by averaging the probabilities from all trees. This segmentation approach reduces overfitting and increases robustness to outliers while maintaining manageable complexity through parallel independent tree training.
2Measurement precision
If reinforcement learning with Q-learning is used to determine action sequences, then optimization of cumulative reward and future state prediction is improved, but computational complexity and training time increase
Solution Approach 1:
The Q-learning algorithm performs preliminary actions by simulating many agent trajectories through each state before deploying the learned policy. During training, the system pre-computes Q function values for state-action pairs by running multiple simulated sequences, allowing the policy to be determined from pre-learned values rather than requiring complex real-time computation during actual deployment.
3Measurement precision
If multiple trained models are used to identify optimal account/product sequences, then accuracy in filtering unlikely successes is improved, but system complexity and processing time increase
Solution Approach 1:
The system merges multiple specialized models (account propensity model using random forest, reinforcement learning model for sequence optimization, and target engagement sequence generators) into a unified account/product sequence recommender. These models work together synergistically, with each contributing its strength: the random forest provides robust propensity estimation, the RL model optimizes sequencing for cumulative reward, and the engagement generators provide target sequences. This combination achieves high accuracy in filtering unlikely successes while managing processing through coordinated model execution.
Data Source
AI summary
To automatically identify a sequence of recommended account/product pairs with highest likelihood of becoming a realized opportunity, an account/product sequence recommender uses an account propensity (AP) model and a reinforcement learning (RL) model and target engagement sequence generators trained on historical time series data, firmographic data, and product data. The trained AP model assigns propensity values to each product corresponding to received account characteristics. The trained RL model generates an optimal sequence of products that maximizes the reward over future realized opportunities. The target engagement sequence generators create target engagement sequences corresponding to the optimal sequence of products. The recommender prunes the optimal sequence of products based on the propensity values from the trained AP model, the completeness of these target engagement sequences, and a desired product sequence length. The recommender uses the remaining products, validated on three models, for account/product recommendations.