Transforming Business Problem into Advanced Analytic Problem

Why data science projects fail:

First reason: failing to transform business problems into ADVANCED analytical problems:

Here is an example of a data science project objective. I will take you through different steps of how experienced analytic person transform the problem to ensure project success:

1 )The client needs a program to forecast the items I expose to customers. So it increases my sales and customer base. Also, to retain loyal current customers. (Awful, very bad objective. wrong terminology, vague success criteria. Almost restating what the client says)

let's check a better version :

2) We need an AI model to predict the items we show to customers, so to increase sales, customer base and also to retain loyal current customers. (still bad, AI model, AI model can goes from tic tac toe to auto driving?, three different unrelated and measurable success criteria?)

let's continue:

3)We need a machine learning algorithm that learn from current data the set of items each customer interested in. So, to increase sales, number of customers and reduce the number of churned customers. ( better, at least we know it is a machine learning and we need data about items and personalization non-clearly touched. At the end, this is how we approach all projects. We need a machine learning algorithm! in).

4) We need a machine learning algorithm that explores current item properties and customer behaviour in order to recommend the best items to them. So, to increase our sales, customer base and also to retain our loyal customers . ( Better, at least , we touch personalization and we want to make sure the data of customer behaviour is required. If we do not have it, you will wait in vain for project success. Notice here, all the refinement is about the data and model but success critiera is the hardest thing for novice analytics to determine feasibility. How one model can ensure increasing of sales, or increasing of customers or retaining the current customers. What functions should be plugged in the algorithm to learn how to optimize on these things. At the end, this is how business people think. They need a silver bullet that transform them from bankruptcy to AI-first company)

5 ) we need a recommendation system that recommend items labeled as expensive to VIP customer segments. The purpose is to increase the click rate on those items. ( Love it. It is a recommendation system after all. We need only to target a segment of the customers , not all and a set of items. Clearly, we care about this segment and expensive item set as those the one that can drive my business to get more money. Success criteria is quantitative , direct and measurable: number of click through. Great, not increasing customer base or sales as it is unfair to say that a model is failure because our customer base does not increase by 30% or our revenue does not increase by millions. Data science is not magic. It is just a tool that fulfil a purpose. However, one issue is here. Are really want to target all expensive items and all VIP segment? What if our model obtained improve on the items that the client is not interested on or on VIP customers that the client observe that they already has a good history of purchase?

Let’s refine it more

6) We need to build a recommendation system that target items (last purchased was 3months ago or more for VIP customer that their purchase has been dropped 20% in the last three months and they have been our customer for more than a year of now. The purpose is to increase the click rate on these items and also to retain VIP. (Very good and concise. So the algorithm will not collect a lot of noise irrelevant data. The algorithm will not address items that has a good amount of transactions. This is good and can reduce the number of data in mega datasets that has billions of records. Further, specifying the required segment of VIP can help the clients to see the contribution of this model on the right segment). Again, one issue here: How many recommendation items we should give to each target customer: 10, 20, 100, how many? When the algorithm should give recommendation: daily, weekly, monthly. What about if some items are seasonal?

7) We need to build a recommendation system that recommend three at maximum items (last purchased was 3months ago or more for VIP customer that their purchase has been dropped 20% in the last three months and they have been our customer for more than a year of now. The algorithm should recommend the items that are probably will be purchased the following month. The purpose is to increase the click rate on these items and also to retain VIP. (Very good, not all items are supposed to be recommended but again that customers probably will purchase the next month but one issue here what about those items that only offered in specific regions. You do not want to show people who are in KSA items that only offered within North America. At the end, you do not want to disturb customers with recommendations list that you can not deliver to them.)

8) We need to build a recommendation system that recommend three at maximum items (last purchased was 3months ago or more for VIP customer that their purchase has been dropped 20% in the last three months and they have been our customer for more than a year of now. The algorithm should recommend the items that are probably will be purchased the following month and delivered to the customer location. The purpose is to increase the click rate on these items and also to retain VIP. [ Perfect. This project is surely will have a high success and good impact on business. All the previous will be marked as failure because business fail to see the contribution of the model on their business operations. The previous objectives either miss the target or provide false improvements that the client does not feel].

The 8th refinement of the objective surely will mark as a successful story. Refinement does not of course stop here. There are many further enhancements and refinements could be done to clearly capture the client's objective and his problem. It differs from one company to another, from one client to another and from one scenario to another. But the 8th refinement is the least that you should come up with.