Data Science in Research? What will you explore and what not?

Credit to https://www.linkedin.com/in/egdataguy/ ‘s post.

Hi, this will be my first post. I decided to write about this for several reasons. First, I met with a person who wrote her portfolios in her blog, and it motivated me(?). Second was because I am about to switch from research to industry in a month (this article was written in Jan 2022, but I forgot to post it). So I would like to share my perspective about data science in research vs industry.

To make this more specifics, I will not be talking about data science research being done in computer science department. Because they have a huge range about research in data analytics, starting from how you improve your model performance until how you you store and get large amount of data for data science purposes efficiently(which is something done in daily basis when you work in tech industry). I will only talk those researches being done in faculties not related to computer science, such as economy, bioengineering, medical, communication etc. Those are focusing only on machine learning application on their specific field.

Data Science in Research

New Model and Improvement

Machine learning has been gaining its popularity lately, and it just keeps increasing. The reason was its application in many aspects of our lives. From getting recommendation in Spotify to detecting early cancer in a patient. Its performance has also been tested many times in many competitions, which they claimed surpasses human intelligence(from the chess match with one of the chess grand master).

Because of its popularity, other fields also started to explore its application on their specific problems at hand. For example, finance field(which in the economy faculty, I suppose) might try to develop a new model to predict stocks market prices(currently being popular for this purpose is Time Series Analysis), or bioengineering field develop a machine learning model to find a new way to construct a medicine, or machine learning to detect breast cancer from MRI.

The purpose of these researches is to explore possible new models in the hope that it might result in better performance for their particular task. They usually started with pre-defined metrics(qualitative and/or quantitative) to measure the model performance. And based on these metrics, they will try to improve the model performance. It is certainly desirable to have a model which perform excellently. However, there is a limit to how we can improve our model performance.

Benefits

• You can focus on just exploring on how to improve the model performance. Ignorance certainly is a bliss. You don’t really need to worry on how it is going to be used in real life application. It allows you to explore many different complex models, even those that require a very large memory space to store its tensor points.
• You might only be working with a very clean dataset. In research, the dataset has usually been cleaned before you use your machine learning model to predict an outcome from it. Which also a dream of many data scientist to be working with a very clean dataset.
• Frontliners! Doing research in data science, give you more opportunities to learn and contribute on the artificial intelligence frontline. This is a cool time to build your portfolios in the data science field.

Limitations

• Slow environment. One of the reason is because the data need to be obtained by different team. In my case, because I am working in bioengineering research company, the data is obtained by clinician. However, in my opinion, the prime reason is because we don’t have the need to hurry. Especially if our purpose is just to get the best performance and technical novelty. It is better to explore and get better results than rushing but get a result that is not worth to published.
• Local data storage. I might be biased about this. However I have been in two research company, and they are both very protective with their dataset. To the point they don’t use any cloud platform to store their dataset (and recently I am not even able to access my working email unless I go to the office). This makes data transfer difficult. We have to put it in a encrypted physical external hard drive and transfer it to the target computer. On top of that, this results in all of the training is also done in a local working station with GPUs.
• Deploying your model to an app might be just a dream. Most of research purposes are publications. Once it is published, it is either done, continue to be researched, being used by internal team, or patented. Being patented does not mean that your work is ready for industrial usage. Most of the time you only patent your method and not your code. This is because research institutes do not always have a platform to commercialize their research output. And even if your model is being used by internal team, they usually use it using a raw code that you wrote. That is opening an IDE and run it. So you won’t see your model to be deployed to an app/web.
• Career path is limited by the degree you have. It will be a rocky road ahead for you in research if you are just a bachelor degree holder. If you want to advance, you need to take higher degree studies and publish a lot of paper.
• It pays less :(. Because research don’t give a direct financial impact to the organization, making money out of the research you are doing is difficult.

Data Science in Industry

Fast and iterative

Different things become the focus of the data science in industry. In the common tech companies, data science focuses on fast implementation and will be improved through the iterative processes called data pipelines. Developing a new model to get better result is desirable but not the main focus of data science in industry. This is because most of tech companies have to move fast. Thousands to millions of users use their platform everyday. A delay in implementation might result in bad services which later might lead to losing customers and revenue.

Of course fast implementation cannot guarantee the high retention rate of the customers. Good performance also perform a key role to retain and attract new customers. E.g. if e-commerce recommend stuffs that users are not interested in, it might result to lower purchase rate; or if the risk model does not perform well, the company might take a bad decision which result in a big loss. That is why improving the model is still considered to be important.

To leverage the model improvement with the fast implementation environment, industry usually use data pipelines to store, get, and clean the data, make prediction, evaluation, and updating the model. This allows a model to be used and updated in real time. On top of that, new dataset coming in also helps us to evaluate the model on the unseen dataset and update the performance.

Benefits

• Fast environment. You will be involved in the development of data science product, from exploring the dataset, building the model, deploying it, until evaluating and maintaining the model. These tasks need to be done in daily basis and therefore require a fast working environment. Especially the need to ensure that the model is working in real time, make the collaboration between cross teams of data scientist, data engineer, and software engineer very important. Data scientists are also required to report to stakes holders on the model development, performance and maintenance.
• You will see your data product deployed!! Either it will be used by internal team(mostly for business purposes: predicting sales or market segment), or used by the user(e.g. recommendation system, data auto filling by optical character recognition). You might deploy it yourself, or working with software engineer to integrate it with the system. This includes making data pipelines, API, deploying the model, building dashboard for monitoring and reporting. Data pipelines is important to ensure that the data flows correctly; from ETL(extract, transform, load), model training, prediction, evaluation, maintenance, and updates. This makes the flows become efficient and neat.
• Working in a cloud base environment. Everything is done in cloud. From data query, to training. This makes our work easier and not limited by time and space.
• You can just implement whatever is state-of-the-art modelling techniques that suitable for your problem. Making your own solutions are, often time, time consuming. Using other people or organizations model (that already known to be working properly) to solve your problem will save a lot of your time. In some cases, you don’t even need to train a new model, and just need to hit a third party API suitable for your problem.
• Career path isvery flexible. You can just be a high school graduate but still make it big if you have a good skill in industry. Because in industry, they value skill to solve their problem more than degree.
• It pays good (it can be very huge if you hit the jackpot). Industries use data product to make money (or make it easier to do that). And therefore, the role to make this data product is important and pays very well.

Limitations

• Dataset in industry is huge, however it is extremely difficult to get a clean dataset. This might be caused be the non existing component of internal audit for our model prediction. We can solve this problem by putting internal audit in our data pipeline. We can use the audit result for monitoring and later use the “cleaned” data for retraining the model.
• Due to time constraint in the project timeline, exploring new possible models or solutions is hard. We often prefer a solution that we already know working. We see the tradeoff between time to explore and small improvement of 1–3% is not worthy.

Conclusion

The choice between doing data science by working in research or in industry depends on one values and priorities. They have their own benefits and limitations as listed above (or probably more than that, as I wrote this article based on my own opinion and view). What I can say for sure, whatever you choose, do your best and enjoy the journey!