Machine learning in the textile industry

La Paz. 2021

The project began with a question: can machine learning improve decision-making in the textile industry? We chose this sector because of its importance to Bolivia and the challenges identified by the team, including competition from imports and informal trade. Our aim was to explore productivity prediction using a publicly available dataset.

Illegal merchandise entering the country
Used clothes
Chinese garments

Because we saw these problems and believe that Bolivia can improve its competitiveness, we chose the textile industry sector to apply Machine Learning.

DEVELOPMENT

We found the dataset on Kaggle, it had the following variables (they are translated into Spanish next to it):

date – Date()

smv – average value per job

day – Day (Mon-Sun)

wip – queued jobs

quarter – 5 periods / month

over time

department – department

incentive – incentive

teamno – Team #

idle time – idle times

no of workers – # of workers

idleman – # of idle workers

no of style change – # of changes

current productivity – current productivity

targeted_productivity – expected productivity

After exploring the variables in the Kaggle dataset, we found missing values in SMV. We examined its distribution and chose to replace those values with the mean to prepare the data for modelling.

We initially framed the task as regression, but the early results were unsatisfactory. We decided to turn the target into categories and try classification models. We normalised the data and examined the correlation matrix shown below.

We then tested several classification models. AdaBoost achieved the best predictive result in our evaluation.

We achieved an accuracy of 0.82 which was simply incredible. After seeing how other methods did not even reach 0.50, we decided to try several additional models such as Random Forest, but the precision was lower (not by much).

In the end we stayed with Adaboost and did an excellent job.

CONCLUSIONS

With the regression models according to the performance (scores of 0.5) calculated, they do not fit the proposed dataset, a treatment was carried out on the target to become a classification problem.
The classification models applied to the dataset gave favorable results, especially Adaboost with a score of 0.82.
The process helped us learn how to reformulate a regression problem as classification. Metrics from the two tasks are not directly comparable.

Saturdays.AI

Written by