This is the html version of the file https://www.opastpublishers.com/open-access-articles/a-data-centric-approach-towards-server-time-series-model-compression.pdf.
Google automatically generates html versions of documents as we crawl the web.

Location via proxy:   [ UP ]   [Report a bug]   [Manage cookies]    No cookies    No scripts    No ads    No referrer    Show this form

Page 1

Volume 2 | Issue 1 | 6

Eng OA, 2024

A Data Centric Approach Towards Server Time Series Model Compression

Research Article

Mantej S. Gill1*, Anil B2, S. Dhamodhran3 and Madhusoodhana Chari S4

1Hewlett Packard Enterprise Bangalore, India

2Hewlett Packard Enterprise Bangalore, India

3Hewlett Packard Enterprise Bangalore, India

4Hewlett Packard Enterprise Bangalore, India

*Corresponding Author

Mantej S. Gill, Hewlett Packard Enterprise Bangalore, India

Submitted: 2023, Nov 29; Accepted: 2024, Jan 05; Published: 2024, Jan 10

Citation: Gill, M. S., Anil, B., Dhamodhran, S., Chari S, M. (2024). A Data Centric Approach Towards Server Time Series

Model Compression. Eng OA, 2(1), 6-12.

Abstract

Effective model compression plays a pivotal role in mitigating the computational and interpretational challenges inherent in

the domain of time series forecasting. In this study, we introduce an innovative data-centric methodology tailored to identify a

representative data subset from the entirety of the dataset. This chosen representative segment forms the cornerstone for the training

of proficient time series forecasting models. Furthermore, our investigation unveils a compelling outcome of this approach—a

substantial reduction in the size of time series forecasting models when trained with this selected representative data segment.

This model compression strategy results in a remarkable 56.31% decrease in storage consumption, a discovery of considerable

significance for optimizing resources and enhancing scalability in time series forecasting. By distilling the dataset to its fundamental

components through our data-centric approach, we aim to enhance both computational efficiency and the interpretability of the

resultant models. This paper introduces a pioneering technique to tackle the challenges associated with data volume and model

complexity in the field of time series forecasting, offering potential pathways for more efficient and insightful modelling in this

domain.

Engineering: Open Access

Keywords: Time-Series · Model Compression · Data-Centric AI · Representative Segments

Introduction

Time-series forecasting serves as a fundamental component

in diverse domains, such as finance, healthcare, environ-

mental monitoring, and industrial processes Rasheeed et al.

[2010]. This methodology unveils historical data patterns and

empowers the anticipation of future trends, rendering it an

indispensable tool for informed decision-making. However, as

the volume of time-series data continues to grow exponentially

in our increasingly interconnected world, the computational and

interpretative challenges associated with handling these vast

datasets and the corresponding complex models have become

increasingly pronounced Ana Almeida and Pinto [2023]. One

key challenge in time-series analysis is the efficient utilization of

computational resources Shu et al. [2014]. Traditional time series

models often require significant storage and computing power,

hindering their scalability and real-time applicability Yu and

Xiao [2022]. Moreover, as these models become more complex

to capture intricate temporal dependencies, they also become less

interpretable, potentially limiting their utility in scenarios where

interpretability is essential.

Use footnote for providing further information about author

(webpage, alternative address)—not for acknowledging funding

agencies. In response to these challenges, model compression

has emerged as a critical research area aimed at reducing the

computational burden of time series analysis without sacrificing

predictive accuracy Zhu and Gupta. Existing approaches to model

compression typically focus on algorithmic or architectural

modifications. While these approaches have shown some

success, they often do not fully address the underlying issue of

excessive data volume in time-series datasets. To address this,

a data-centric approach towards time series model compression

has gained attention in recent years. A data-centric approach

towards time series model compression Yin et al. [2022] offers

a promising solution to these challenges. By prioritizing data

reduction and representation techniques, a data-centric approach

focuses on compressing the time-series data itself instead of solely

relying on modifying the models or algorithms. This approach

recognizes that the sheer volume of time-series data is a significant

contributor to the computational burden and seeks to efficiently

store and process this data while maintaining its essential

features and predictive capabilities. One proposed method for

effective storage of time-series data is through a novel approach

that reduces the computation expense Bhattarai et al. [2019].

ISSN: 2993-8643

Volume 2 | Issue 1 | 7

Eng OA, 2024

This method leverages data compression techniques to reduce

the storage requirements of time-series data, allowing for more

efficient utilization of computational resources Ryabko [2012]. By

compressing the time-series data, this approach enables faster data

retrieval and analysis, leading to improved scalability and real-

time applicability of time-series models Bhattarai et al. [2019].

Additionally, this data-centric approach also acknowledges the

importance of preserving the interpretability of time series models.

By reducing the data volume while retaining the essential features

of the time series, interpretable insights can still be derived

from the compressed data. Furthermore, a data-centric approach

recognizes that time series analysis often involves working with

specific temporal ranges of data rather than the entire dataset Barez

et al. [2023]. To address this, compression and deletion strategies

can be implemented to store and retrieve only the relevant portions

of the time-series data. One example of a data-centric approach

towards time series model compression is the proposed symbolic

representation of time series. This symbolic representation allows

for a reduction in dimensionality and numerosity, effectively

compressing the time-series data.

Continuing with this concept, in this paper, we present a data-

centric approach designed to tackle the challenges associated

with model compression in the realm of time series analysis. Our

methodology is rooted in the idea that not all data points in a time-

series dataset are equally informative or necessary for training

effective models. Instead, we advocate for the identification of

a representative data segment from the complete dataset, which

serves as the cornerstone for training time series models. Our

research introduces an intriguing outcome of this data-centric

approach: a substantial reduction in the size of time series models

when trained with the selected representative data segment.

By distilling the dataset to its essential components, we aim to

significantly reduce storage consumption. In our experimental

results, we observe a remarkable 56.31% reduction in storage

requirements, a finding of profound significance for resource

optimization and scalability in time series analysis. Furthermore,

our data-centric approach not only addresses the computational

challenges but also enhances the interpretability of the resultant

time series models. By focusing on the most informative data, we

aim to simplify model complexity, making it easier to understand

the underlying patterns and relationships within the data.

In summary, this paper presents a pioneering method to tackle

the issues of data volume and model complexity in time series

analysis. Our data-centric approach promises more efficient and

insightful modelling in this domain, offering potential avenues for

improving both computational efficiency and interpretability in

time-series analysis. Through a comprehensive exploration of our

methodology and experimental results, we aim to demonstrate the

transformative potential of this approach in the field of time-series

analysis.

Related Works

Previous research has explored various strategies for addressing

the challenges of data volume and model complexity in time

series analysis. One common approach is pruning, which involves

removing unnecessary connections or parameters from a time series

model Wielgosz [2020]. This approach aims to reduce the size and

complexity of the model by selectively removing components that

have little impact on the overall performance. Another strategy is

compressing larger models, which involves applying compression

algorithms to reduce the storage requirements of the model

without compromising its performance. While these approaches

have shown some success in reducing the size of time series

models, they often focus on model-centric methods rather than

data-centric approaches. Data-centric approaches towards time

series model compression include lightweight data compression

methods based on data statistics and deviation Vidhi Agrawal.

Another method involves using real-world data compressors

for time series forecasting, where multiple data compressors are

combined into one forecasting method with the automatic selection

of the best algorithm for the input data K.S. Chirikhin [2019].

Additionally, a two-level compression model has been proposed

that selects a proper compression scheme for each individual

point, capturing diverse patterns at a fine granularity Marjai et al.

[2021]. Furthermore, a data-centric approach has been proposed

for detecting anomalous data points in time series data, achieving

100% performance in correctly identifying anomalies Yanjun

Zhou [2021]. Finally, a data compressing apparatus has been

developed that performs a polygonal line approximation process

on time series data Adrián Gómez-Brandón a [2021]. In line with

these existing approaches, our research focuses on developing a

data-centric approach towards time series model compression.

Volume 2 | Issue 1 | 8

Eng OA, 2024

3 Problem Statement

3.1 Definitions

• Time Series Data: Let D represent the time series dataset, consisting of N data points:

D = 1x1,x2,...,xN l

where xi represents the ith data point in the time series.

• Training Dataset: The training dataset, denoted as Dtrain, comprises a subset of D used for model training:

Dtrain = 1x1,x2,...,xntrain l

where ntrain is the number of data points in the training dataset.

• Validation Dataset: The validation dataset, denoted as Dval, is a subset of D used for model selection and

tuning:

Dval = 1xntrain+1,xntrain+2,...,xntrain+nval l

where nval is the number of data points in the validation dataset.

• Test Dataset: The test dataset, denoted as Dtest, is employed to assess the performance of the time series

forecasting model:

Dtest = 1xntrain+nval+1,xntrain+nval+2,...,xN l

where ntest is the number of data points in the test dataset.

• Time Series Forecasting Model: A time series forecasting model, M, is a function that maps a sequence of

past data points to predict future data points:

M : [x1,x2,...,xt] → ˆxt+1

where ˆxt+1 represents the predicted data point at time t + 1.

3.2 Mathematically Represented Problem

The problem at hand can be mathematically formulated as follows: Given a time series dataset D, we aim to find a

3.2 Mathematically Represented Problem

The problem at hand can be mathematically formulated as follows: Given a time series dataset D, we aim to find a

representative data segment S∗ from D that minimizes the Root Mean Square Error (RMSE) in time series forecasting.

Formally, our objective is to find S∗ such that:

S∗ = arg min

S⊂D

RMSE(M(Dtrain,S),Dval)

where Dtrain,S represents the training dataset created using the selected segment S, and RMSE(·) denotes the Root

Mean Square Error between the predicted values using M and the actual values in the validation dataset Dval.

Our objective is to find the segment S∗ that optimally balances the trade-off between data reduction and forecasting

accuracy, thus improving the efficiency of time series forecasting models while preserving their predictive capabilities.

4 Proposed Solution

4.1 Data Split and Segmentation

To initiate our approach, we employ a structured data splitting and segmentation strategy. The time series dataset D is

divided into three distinct subsets: the training dataset (Dtrain), the validation dataset (Dval), and the test dataset (Dtest).

This division is conducted with a predetermined proportion of 60%, 20%, and 20%, respectively. Within the training

dataset, we further partition the data based on a parameter selected by the user. In our experimentation, we segment

on a weekly basis, creating distinct segments labeled as week1, week2, week3, and so forth. The validation dataset

assumes a crucial role in the process of identifying candidate segments, ultimately leading us to select the representative

data segment. The test dataset, reserved for model evaluation, allows us to assess the performance of the selected

representative data segment.

Problem Statement

Data Split and Segmentation

Volume 2 | Issue 1 | 9

Eng OA, 2024

Figure 1: Pictorial representation of segments derived for training in the iterations

Figure 1: Pictorial Representation of Segments Derived for Training in The Iterations

Iterative Model Training and Selection

Our methodology is characterized by an iterative model training

and selection process, which is instrumental in narrowing down

the choice of the representative data segment. This process

encompasses the following key steps:

First Iteration: In the initial iteration, as illustrated in Figure 1

4.2, we initiate the process by creating two data segments from

the training dataset. The first segment is formed by excluding the

data from the last week, and the second segment excludes the

data from the first week. We proceed to train two distinct time

series forecasting models using these two segments. Subsequently,

we compare the Root Mean Square Error (RMSE) observed on

the validation dataset for these models. The segment associated

with the model exhibiting the lower RMSE is selected for further

iterations.

Subsequent Iterations: The iterative cycle continues, with the

segment selected from the previous iteration serving as the

foundation for the subsequent one. In each iteration, two new

segments are created, one excluding the data from the last week

and the other excluding the data from the first week. Once more,

two models are trained using these segments, and the RMSE on

the validation dataset is compared. The segment corresponding to

the model with the lower RMSE is chosen to advance to the next

iteration.

This iterative process persists until a minimum of two weeks’

worth of data is available for segment creation. Subsequently,

the data segment from the iterations that exhibits the lowest error

rate on the validation data is identified as the representative data

segment essential for training an effective time series forecasting

model.

Innovative Data-Centric Approach

The approach delineated above is both novel and data-centric

in its nature. It seeks to identify the representative data segment

that optimally balances the trade-off between data reduction

and forecasting accuracy, ultimately improving the efficiency of

time series forecasting models while preserving their predictive

capabilities. Our approach offers a pioneering solution to the

challenge of selecting a representative data segment within a time-

series dataset, fostering a more efficient and insightful approach to

time series forecasting.

Experiments

In our pursuit to validate the effectiveness of the proposed data-

centric approach in addressing the challenges of model compression

for time series forecasting, we conducted comprehensive

experiments. These experiments utilized power utilization time

series data collected from 9 servers over 18 weeks, spanning from

November 1, 2021, to March 30, 2022. The dataset records the

average input power consumed by each server at hourly intervals.

Experimental Setup

For our experiments, we employed Meta’s Prophet algorithm to

train time series forecasting models. To assess the efficacy of our

proposed approach, we conducted two comparative studies:

Comparison with Standard 80-20 Split

The first study involved comparing the performance of time series

models trained using the proposed approach with a standard 80-20

data split. The Root Mean Square Error (RMSE) was used as the

evaluation metric. The results of this comparison are summarized

in Table 11.

Volume 2 | Issue 1 | 10

Eng OA, 2024

System ID

RMSE (Standard

80-20)

RMSE (Representative

Data)

Reduction in Error

Rate (%)

server0

29.37

29.31

0.2%

server1

21.10

19.84

5%

server2

40.52

22.10

45.45%

server3

18.77

18.42

1.86%

server4

21.29

26.40

-24%

server5

16.79

18.15

-8.1%

server6

43.20

29.91

30.7%

server7

56.48

41.74

26.09%

server8

27.37

41.33

-51%

Table 1: Comparing Error Rates of Models Trained Using the Standard 80-20 Split Approach Against the Proposed Approach

For 9 Different Servers.

Figure 2: Plot of Trainset, Test set And Forecasted Data Using Model Trained Using Standard 80-20 Split

Figure 3: Plot of Representative Data Segment, Test Set and Forecasted Data Using Model Trained Using Proposed Approach.

Comparison with Models Trained with 60% Data

In the second study, we compared the performance of models

trained with the proposed approach against models trained using

only 60% of the data. Again, RMSE served as the evaluation

metric, and the results are captured in Table 22.

Results

Table 11 presents the comparison of error rates between models

trained using the standard 80-20 split approach and the proposed

data-centric approach for the 9 different servers. Notably, 6 out of

9 time-series models trained using the representative data segment

exhibited lower error rates compared to the models trained using

the standard 80-20 split approach. This observation underscores the

effectiveness of our proposed approach in enhancing forecasting

accuracy. In Table 22, we compare the error rates of models trained

using 60% of the data against those trained using the proposed

data-centric approach for the same set of 9 servers. Encouragingly,

7 out of 9 time-series models trained with the representative data

segment demonstrated lower error rates compared to models

trained with only 60% of the data. This reaffirms the effectiveness

of our approach, even in comparison to the utilization of a larger

data portion.

Furthermore, we investigated the reduction in model sizes when

trained with the representative data segment. Table 4 details the

comparison of time series model sizes for the 9 servers when trained

using the standard 80-20 split approach against our proposed

approach. Strikingly, there was a total reduction of 56.31% in

space consumption, demonstrating the significant efficiency

gains associated with our data-centric approach. In summary,

our experiments reveal that the proposed data-centric approach

leads to improvements in time series forecasting accuracy, with

substantial reductions in model sizes. This approach not only

enhances computational

Volume 2 | Issue 1 | 11

Eng OA, 2024

System ID

RMSE (60% Data)

RMSE (Representative

Data)

Reduction in Error Rate

(%)

server0

64.37

29.31

54.46%

server1

32.39

19.84

38.74%

server2

41.22

22.10

46.38%

server3

26.86

18.42

31.42%

server4

38.31

26.40

31.08%

server5

22.35

18.15

18.8%

server6

112.07

29.91

73.31%

server7

37.91

41.74

-10.10%

server8

37.05

41.33

-11.55%

System ID

Representative Data Segment

Date Range

Representative Data

Segment Size (in weeks)

Reduction in Error Rate

(%)

server0

14th Dec 2020 – 21st Feb 2021

9

54.46%

server1

14th Dec 2020 – 7th Feb 2021

7

38.74%

server2

4th Jan 2021 – 31st Jan 2021

4

46.38%

server3

25th Nov 2020 – 7th Feb 2021

10

31.42%

server4

1st Feb 2021 – 7th Feb 2021

1

31.08%

server5

2nd Jan 2021 – 21st Feb 2021

8

18.8%

server6

7th Dec 2020 – 14th Feb 2021

9

73.31%

server7

1st Feb 2021 – 7th Feb 2021

1

-10.10%

server8

25th Nov 2020 – 21st Feb 2021

12

-11.55%

Table 2: Comparing Error Rates of Models Trained Using 60% Of the Data Against The Proposed Approach For 9 Different

Servers.

Table 3: Data Range and Size of Representative Data Segment for Each Server.

efficiency but also contributes to the interpretability and scalability

of time series forecasting models. The results presented here

validate the transformative potential of our approach in the realm

of time series analysis.

Conclusion

In this paper, we introduced a novel data-centric approach

designed to address the challenges of model compression in time

series analysis. Our methodology, grounded in the idea that not

all data points in a time-series dataset are equally informative,

advocated for the identification of a representative data segment

as the foundation for training effective time series models.

Through a comprehensive series of experiments, we explored the

effectiveness of this approach. Our results clearly demonstrate

the significant advantages of our data-centric approach. When

compared to the standard 80-20 split, our approach consistently

exhibited lower root mean square error (RMSE) on the test set

for the majority of the servers, illustrating its superior predictive

performance. Moreover, the reduction in error rate was most

pronounced when compared to models trained with only 60% of

the data, highlighting the efficiency of our proposed approach.

Not only did our approach improve predictive accuracy, but it also

addressed the issue of model size. The models trained with the

representative data segment were notably more compact, resulting

in a 56.31% reduction in model size on average. This reduction is

of paramount importance for resource optimization, particularly

in scenarios where storage space and computational resources are

at a premium. Furthermore, our data-centric approach also aligns

with the need for model interpretability. By focusing on the most

informative data segments, we simplified model complexity,

making it easier to understand the underlying patterns and

relationships within the data. This improvement in interpretability

has practical implications for decision-making and forecasting in

various fields. In summary, our research offers a pioneering method

to address the challenges of data volume and model complexity in

time series analysis. By focusing on a representative data segment,

we have demonstrated the potential for more efficient and

insightful modeling, both in terms of predictive performance and

resource optimization. We believe that our data-centric approach

opens up new possibilities for enhancing computational efficiency

and interpretability in the realm of time series analysis, making it

a valuable tool for a wide range of applications across different

domains.

Volume 2 | Issue 1 | 12

Eng OA, 2024

System ID

Size of Model (Standard 80-

20, Bytes)

Size of Model (Representative

Data, Bytes)

Reduction in Model

Size (%)

server0

454,615

256,557

43.56%

server1

454,247

189,232

58.34%

server2

454,201

138,036

69.60%

server3

454,309

277,887

38.83%

server4

454,210

37,988

91.63%

server5

454,276

248,255

45.35%

server6

454,137

254,901

43.87%

server7

454,222

38,018

91.63%

server8

454,258

345,105

24.02%

Total

4,088,475

1,785,979

56.31%

Table 4: Comparison of Time Series Model Sizes Of 9 Different Servers Trained Using the Standard 80-20 Split Approach

Against the Proposed Approach.

This paper encourages further exploration of data-centric

approaches and their application in addressing the computa-

tional and interpretative challenges of time series analysis. As the

volume of time series data continues to grow, these innovative

methodologies are crucial for staying ahead in the realm of data-

driven decision-making and forecasting.

References

1. Rasheed, F., Alshalalfa, M., & Alhajj, R. (2010). Efficient

periodicity mining in time series databases using suffix trees.

IEEE Transactions on Knowledge and Data Engineering,

23(1), 79-94.

2. Almeida, A., Brás, S., Sargento, S., & Pinto, F. C. (2023).

Time series big data: a survey on data stream frameworks,

analysis and algorithms. Journal of Big Data, 10(1), 83.

3. Shu, W., Wang, W., & Wang, Y. (2014). A novel energy-

efficient resource allocation algorithm based on immune clonal

optimization for green cloud computing. EURASIP Journal

on Wireless Communications and Networking, 2014(1), 1-9.

4. Yu, D., & Xiao, J. (2022). Multilevel Information Granule

Construction Model Based on Large Data Environment and

Its Application in Time Series. Mobile Information Systems,

2022.

5. Michael H. Zhu and Suyog Gupta. pruning for model

compression - arxiv.org.

6. Yin, X. X., Miao, Y., & Zhang, Y. (2022). Time series-based

data explorer and stream analysis for anomaly prediction.

Wireless Communications and Mobile Computing, 2022.

7. Bhattarai, B. P., Paudyal, S., Luo, Y., Mohanpurkar, M.,

Cheung, K., Tonkoski, R., ... & Zhang, X. (2019). Big

data analytics in smart grids: state‐of‐the‐art, challenges,

opportunities, and future directions. IET Smart Grid, 2(2),

141-154.

8. Ryabko, D., & Mary, J. (2012). Reducing statistical time-

series problems to binary classification. Advances in Neural

Information Processing Systems, 25.

9. Barez, F., Bilokon, P., & Xiong, R. (2023). Benchmarking

Specialized Databases for High-frequency Data. arXiv

preprint arXiv:2301.12561.

10. Pietron, M., & Wielgosz, M. (2020). Retrain or not retrain?-

Efficient pruning methods of deep CNN networks. In

Computational Science–ICCS 2020: 20th International

Conference, Amsterdam, The Netherlands, June 3–5, 2020,

Proceedings, Part III 20 (pp. 452-463). Springer International

Publishing.

11. Dhananjoy Dey Vidhi Agrawal, Gajraj Kuldeep. Near lossless

time series data compression methods using statistics.

12. Chirikhin, K. S., & Ryabko, B. Y. (2019). Application of data

compression techniques to time series forecasting. arXiv

preprint arXiv:1904.03825.

13. Marjai, P., Lehotay-Kéry, P., & Kiss, A. (2021). The use of

template miners and encryption in log message compression.

Computers, 10(7), 83.

14. Zhou, Y., Ren, H., Li, Z., Wu, N., & Al-Ahmari, A. M. (2021).

Anomaly detection via a combination model in time series

data. Applied Intelligence, 51, 4874-4887.

15. Gómez-Brandón, A., Paramá, J. R., Villalobos, K.,

Illarramendi, A., & Brisaboa, N. R. (2021). Lossless

compression of industrial time series with direct access.

Computers in Industry, 132, 103503.

Copyright: ©2024 Mantej S. Gill. This is an open-access article distributed

under the terms of the Creative Commons Attribution License, which

permits unrestricted use, distribution, and reproduction in any medium,

provided the original author and source are credited.

https://opastpublishers.com