Food web analysis through informatics approach- A deep learning implementation in Ecology

Abstract

Machine learning (ML) methods can build predictive models for the networks of trophic links in food webs to explaining ecosystem structure and dynamics. Implementation of machine learning for food web problems is not new: Logic-based machine learning has been successfully used to build insect food webs and others using ANN with either MLR or SOM approaches. However, the observations on previous work’s implementations of ML in the food web is that it is difficult to track their models, its architecture, internal parameters and, the number of hidden layers. The ANN/ML were presented, as a black box and this approach is now more trackable with the evolved knowledge of deep learning (DL). Which we have demonstrated in our implementation using deep learning with TensorFlow. Thinking the future-ecological big data, we have demonstrated why to consider deep learning in Ecology research. Our work explored insight into the future of the big data from various fossil sites that will help to identify species present, dead, or already extinct in environmental samples by sequencing. As of deep Learning in the recent trend of big data in Genomics, it is also, the next break in ecological research that is expected to result in the big data era of Ecological data. The need of implementing various protocols in deep learning for handling ecological big data is the motivation for this work where more ecological problems can be tackled with an informatics approach. We demonstrate how theoretically correct is the implantation of deep learning to the food web ecology problem of network model building. In our deep learning, we incorporated prior knowledge to enhance our deep machine learning to be robust in extrapolation prediction in an unsupervised manner. Prior knowledge was combined from Globi meta-data base. The trait-based approach computation space was achieved by ML forward and backward propagation of gradient descent, this gave ease to the mathematical intractable of having an effective combination of traits. The best positioning presented in food web trait-based was achieved by the ML parameters twerking. Theoretically, we generally assumed the position that most problem in the food web is similar to a graph network problem. We broadly classify node to taxa; classification link to trophic links; Prediction-direct or undirected edges (predation, herbivory, detritivores, parasitism, cannibalism). Features are relationships between trophs or species. NN input layer encodes the adjacency matrix values for the nodes. The output layer encodes the probability of species relationship type and function group learned by the machine from network data. Hidden layers are functions of the input and are used to efficiently encode through forward-backwards propagation and NN activation function: hyperbolic tangent activation function and tanh. Results obtained so far, showed greater ecological realism in representing community can be realized in DL.