This repository focuses on exploring and enhancing time series analysis through various embedding techniques, including the Time2Vec approach, and evaluating their impact on model performance.
- Build EDA (Exploratory Data Analysis) to thoroughly understand the dataset.
- Introduce methods to build embeddings for time series data, adding valuable features for models.
- Explore the Time2Vec approach as a novel method of time series representation.
- Consider alternatives to Time2Vec for comparative analysis.
- Provide models for experimentation:
- v1: Model trained on original data.
- v2: Model trained on data enriched with embeddings.
- Compare metrics to assess the impact of embeddings on model performance.
- Implement an LSTM model to further leverage the time series embeddings.
- Further compare metrics post-LSTM implementation.
The EDA is designed to offer deep insights into the dataset, helping us understand underlying patterns, anomalies, and correlations. It's the first step in ensuring the robustness of subsequent analyses and model training.
Various techniques for embedding time series data are introduced, providing diverse ways to represent time-dependent features effectively.
We delve into the Time2Vec approach, evaluating its methodology and effectiveness in capturing temporal patterns in time series data.
The implementation details of the Time2Vec model are provided, illustrating how it transforms time series data into a more informative representation for predictive models.
We enrich our time series dataset with the generated embeddings, aiming to enhance the dataset's informational value and predictive power.
Two model versions are trained and evaluated:
- Linear regression on original data.
- Linear regression on data enriched with embeddings. The performance of both models is then compared using standard metrics.
An LSTM model is implemented to utilize the embedded data effectively, harnessing its potential in capturing sequential dependencies in time series.
Post-LSTM, we again compare metrics to gauge the effectiveness of embeddings when used in conjunction with more complex models like LSTM.
The Time2Vec model, as it is right now, requires a lot of refinement.
Theoretically, its ability to capture the periodic aspects of data is advantageous, but practically, this involves a series of steps:
- Identifying appropriate periodic functions for creating representations.
- Determining the optimal dimensionality for embeddings.
- Undertaking substantial efforts in dimensionality reduction, which is currently lacking.
- Addressing the issue of rapidly increasing dimensionality in more complex models.
However, on a brighter note:
- There are possibilities with other modeling approaches.
- There still exists some space for experimentation.