microsoft / fost

FOST is a general forecasting tool, which demonstrate our experience and advanced technology in practical forecasting domains, including temporal, spatial-temporal and hierarchical forecasting. Current general forecasting tools (Gluon-TS by amazon, Prophet by facebook etc.) can not process and model structural graph data, especially in spatial domains, also those tools suffer from tradeoff between usability and accuracy. To address these challenges, we design and develop FOST and aims to empower engineers and data scientists to build high-accuracy and easy-usability forecasting tools.

• Fost
• Framework of Fost
• Quick Start
  • Installation
  • Train with FOST
• Data Format
• Examples
• Contact Us

FOST

FOST(Forecasting open source tool) aims to provide an easy-use tool for spatial-temporal forecasting. The users only need to organize their data into a certain format and then get the prediction results with one command. FOST automatically handles the missing and abnormal values, and captures both spatial and temporal correlations efficiently.

Framework of FOST

Following is the framework of FOST, basically it contains 4 main components:

Module name	Description
Preprocessing	Preprocessing module aims at handle varies data situation, currently FOST designed sub-module to handle issues such as missing value, unalignment timestamp and feature selection.
Modeling	FOST contains implements for different mainstream deep learning models such as RNN, MLP and GNN, for better performance on varies custom data. Further model implements such as Transformer, N-beats are in progress.
Fusion	Fusion module aims at automatically select and ensemble model predictions.
Utils	There are many other utils in FOST, such as neural-network trainer and predictor, result plotter and so on.

Quick Start

Installation

Installation of dependency packages

1. Prerequisites

This project relies on pytorch >= 1.8 and torch-geometric >= 1.7.2

• torch installation reference link：https://pytorch.org/get-started/previous-versions/

• torch-geometric installation reference link: https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html

2. Installation

You can install fost with pip:

pip install fostool

Train with FOST

1. Import forecasting pipeline

from fostool.pipeline import Pipeline

2. Setting data path and lookahead

You need to pass your train.csv and graph.csv for model training, see dataformat for data preparing.

train_path = '/path/to/your/train.csv'
graph_path = '/path/to/your/graph.csv' # graph_path is alternative
lookahead = 7 # Forward steps you would like to predict.

3. Fit and predict

We provide a default config file in config/default.yaml. You could use your config file through config_path augment.

fost = Pipeline(lookahead=lookahead, train_path=train_path, graph_path=graph_path)
fost.fit()
result = fost.predict()

4. Plot results

fost.plot(result)

Data Format

You can fetch sample data on /examples

1. train.csv

3 columns are required for train.csv:

• Node: node name for current data
• Date: date or timestamp for current data
• TARGET: target for prediction

A valid format may look like:

Node	Date	TARGET
Alaska	1960-01-01	800592.0
Alaska	1961-01-01	933600.0
Alabama	1960-01-01	10141633.0
Alabama	1961-01-01	9885992.0
Alabama	1962-01-01	10497917.0

Columns except above will be regarded as feature columns.

2. graph.csv (option)

graph.csv should only contains 3 columns:

• node_0: node name for fist node, node name should align with node name in train.csv.
• node_1: node name for second node, node name should align with node name in train.csv.
• weight: weight on connection for node_0 to node_1.

If graph.csv is not provided, identity graph will be used.

Examples

We prepared several examples on examples:

1. Predict simulation cosine curve
2. Predict States Energy Data
3. Save and load model

Contact Us

• If you have any issues, please create issue here or send messages in gitter.
• For other reasons, you are welcome to contact us by email([email protected]).

fost's Issues

Undeclared dependency

It appears that torch-sparse is also required.

fost.plot data miss-ordered

if the train.csv data is ordered in descending order by date, fost.plot will not give the correct result. I got something like this with lookahead=lookback_size=300 (day):

config.yaml documentation

Looks like I need a config file somewhere?

unhashable type

when I run the code:
#predict future 10 steps
lookahead = 10
fost = Pipeline(lookahead=lookahead, train_path='train.csv')

in the 1. Cosine prediction.ipynb
I meet some error
TypeError: unhashable type: 'pandas._libs.tslibs.offsets.Day'

how could I solve it

Could you please attach an example about spatial-temporal forecasting

According to the introduction, FOST has the function of spatiotemporal forecasting, but only time series forecasting is listed in the example.

这个模型能用于时序分类任务吗

需要修改哪些

Not support dataframe as a train_path

In the docs of class Pipeline, train_path can be str or DataFrame. While there is no method to load dataframe as an input.

prerequisites missing some packages

torch_sparse, torch_scatter should be added as prerequisites

support for python 3.10

add support for python 3.10 ?

Is there a roadmap of FOST?

Is there a roadmap of FOST?
such as support more models, datasets and more features ?

predictor not support 'cpu'

The predictor does not support 'cpu', the current version will throw an exception:
https://github.com/microsoft/FOST/blob/main/fostool/pipeline.py#L281)

gconv.py

Are there any papers to describe principles of the sentence 'Build an edge-attribute-aware graph convolutional layer' in class EGNNConv?

a error that FOST lost a node information when it was matching the node in graph.

PyYAML >= 5.1 required