Applications of Matrix Methods in Artificial Intelligence and Machine Learning (AMAIML) Session 2

Time and Date: 16:50 - 18:30 on 12th June 2018

Room: 0.3

Chair: Kourosh Modarresi

521 Determining Adaptive Loss Functions and Algorithms for Predictive Models [abstract]
Abstract: We consider the problem of training models to predict sequential processes. We use two econometric datasets to demonstrate how different losses and learning algorithms alter the predictive power for a variety of state-of-the-art models. We investigate how the choice of loss function impacts model training and find that no single algorithm or loss function results in optimal predictive performance. For small datasets, neural models prove especially sensitive to training parameters, including choice of loss function and pre-processing steps. We find that a recursively-applied artificial neural network trained under L1 loss performs best under many different metrics on a national retail sales dataset, whereas a differenced autoregressive model trained under L1 loss performs best under a variety of metrics on an e-commerce dataset. We note that different training metrics and processing steps result in appreciably different performance across all model classes and argue for an adaptive approach to model fitting.
Kourosh Modarresi and Michael Burkhart
522 Adoptive Objective Functions and Distance Metrics for Recommendation Systems [abstract]
Abstract: We describe, develop, and implement different models for the stan-dard matrix completion problem from the field of recommendation sys-tems. We benchmark these models against the publicly available Netflix Prize challenge dataset, consisting of ratings on a 1-5 scale for (user,movie)-pairs. We used the 99 million examples to develop individual models, built ensembles on a separate validation set of 1 million examples, and tested both individual models and ensembles on a held-out set of over 400,000 examples. While the original competition concentrated only on RMSE, we experiment with different objective functions for model training, ensemble construction, and model/ensemble testing. Our best-performing estimators were (1) a linear ensemble of base models trained using linear regression (see ensemble e1, RMSE: 0.912) and (2) a neural network that aggregated predictions from individual models (see ensemble e4, RMSE: 0.912). Many of the constituent models in our ensembles had yet to be developed at the time the Netflix competition con-cluded in 2009. To our knowledge, not much research has been done to es-tablish best practices for combining these models into ensembles. We con-sider this problem, with a particular emphasis on the role that the choice of objective function plays in ensemble construction. For a full list of learned models and ensembles, see Tables 1 and 2.
Kourosh Modarresi and Michael Burkhart
60 An Early Warning Method for Basic Commodities Price Spike Based on Artificial Neural Networks Prediction [abstract]
Abstract: Basic commodities price spike is a serious problem for food security and can carry wide effect and even social unrest. Its occurrences should always be anticipated early enough because government needs sufficient time to form anticipatory policies and proactive actions to overcome the problem. According to law regarding food in Indonesia, the government should develop an integrated information system on food security, which includes an early warning function. This study proposes an early warning method based on Multi-Layer Perceptron predictive model with Multiple Input Multiple Output (MIMO). The warning status is determined based on the coefficient of variation of obtained price prediction from the government’s reference price. A great deal of attention was paid for tuning the model parameters to obtain the most accurate prediction. Model selection was conducted by time series k-fold cross-validation with the mean squared error criterion. The predictive model gives a good performance, where the average of normalized root mean squared errors of sample commodities is ranging from 9.909% to 18.046%. Importantly, the method is promising for modelling basic commodities price and may help the government to predict price spikes and to determine further anticipatory policies.
Amelec Viloria
22 Predicting Heart Attack through Explainable Artificial Intelligence [abstract]
Abstract: This paper reports a novel classification technique by implementing a genetic al-gorithm (GA) based trained ANFIS to diagnose heart diseases. The performance of the proposed system was investigated by evaluation functions including sensi-tivity, specificity, precision, accuracy and also Root Mean Squared Error (RMSE) between the desired and predicted outputs. It was shown that the sug-gested model is reliable and suggests high values of evaluation functions. Also, a novel technique was proposed which provides explainability graphs based on the predicted results for the patients, automatically. The reliability and explainability of the system was the main aim of this paper and was proved by providing dif-ferent criteria. Additionally, the importance of the different symptoms and fea-tures in diagnosis of heart disease was investigated by defining an importance evaluation function and it was shown that some features have key role in predic-tion of the heart disease.
Mehrdad Aghamohammadi, Manvi Madan, Jung Ki Hong and Ian Watson