Death of an Algorithm Engineer

In the wave of digital transformation, algorithm engineers are highly expected and are seen as the key to solving complex business problems. However, reality is often more cruel than ideal. This article uses a series of real cases to reveal the challenges and difficulties faced by algorithm engineers in traditional enterprises.

"Our algorithm engineers are too poor and can't solve the problem at all!" As a second-party who often deals with traditional enterprises, Mr. Chen has heard too many complaints like this and has seen too many similar tragic scenes. Today we will talk about it systematically.

Is the model amazing? It’s amazing! ChatGPT is almost better than humans. How can it not be amazing? Therefore, many companies grit their teeth and stomp their feet, paying high salaries to hire algorithm engineers, data mining engineers, and data modelers from large Internet companies, hoping that they can make super-powerful models. "As long as you can make accurate predictions, I will definitely be able to do it well" is their mantra.

It just so happened that many large companies had laid off employees in recent years, and a group of people thought they could take over traditional companies under the banner of "former senior algorithm engineer of ByteDance/Ali/Tencent", and from then on, they could turn from black chickens into phoenixes and reach the peak of their lives. The two hit it off. The tragedy began from here...

01 Not considering the business, taking the blame

Case 1: A traditional enterprise wanted to build a product recommendation model to accurately match user needs. However, the algorithm hired was fired after only half a year. The reason for the dismissal was that the recommendation was not accurate and interfered with normal sales. The head of the marketing department of Party A said disdainfully: Alibaba’s recommendation algorithm is not that good.

After carefully studying the business scenarios, I found that: Dear, it is not Alibaba that has a problem, but your company is not Alibaba. Alibaba is a platform, and there are countless products waiting to be promoted on the platform.

But when it comes to your company, you will find:

1. Some products are hot-selling products that can sell well even without promotion.

2. Some products are the heart and soul of the business. If there is any problem, they will be chopped into pieces.

3. Some products are inherently flawed, with poor functions and unreasonable pricing. They are simply no match for their competitors, so recommendation algorithms are useless.

4. Some products are of good quality, but they have low internal political status, cannot obtain resources, or have unreasonable pricing, which leads to acquired shortcomings.

The previous algorithm guy didn’t consider the business competition and directly put the model into practice. He mixed all the products together for recommendation (using collaborative filtering, without considering the user stickiness of the enterprise and the amount of user behavior data). As a result, the main product declined, and the sales and marketing departments joined forces to put the blame on him. In the end, he was not only kicked out, but also became infamous.

After carefully analyzing these backgrounds, an optimization plan was developed (as shown below):

First do a good product analysis, select the small product category with short legs, find the department that endorses it, and then you can start working.

Sure enough, the first wave of promotion took effect immediately.

So Party A happily took over and went back to optimize and iterate.

02 If you don’t refine the scene, you will have a lot of trouble

Case 2: A chain store wanted to build a model to accurately predict the sales of fish balls, rice rolls, rice balls, bread, etc. in each store, down to the specific SKU, so that the store would not waste food due to overstocking, nor miss sales due to shortages. As a result, the seven modeling guys worked hard for half a year but the predictions were not accurate enough. Four of them resigned, and the remaining three were dejected. How can it be 100% accurate?

If you think about this problem scenario carefully, you will find it very funny: If they really have the ability to predict fish balls and sausages with 100% accuracy, why would these seven guys just do shitty jobs and go and trade futures?

After careful research, we found that the so-called "out of stock and missed sales" is just empty talk. This is because there is no formal out of stock registration system (many companies have it, but this one does not). However, the loss rate caused by backlog is very high, so an optimization plan was developed (as shown below).

After running for two months, the loss rate dropped significantly, and we could see the cost reduction. At the same time, although some people complained, "Oh, some stores are out of stock." But where is the evidence? Where is the evidence? Where is the evidence? Without data, empty words are not believed! So we successfully turned the situation around.

Not surprisingly, Party A took over and continued the optimization (yes, Party A just didn’t like to sign the second or third phase, and thought they could handle the rest themselves. Of course, that’s a story for later, haha).

03 Failure to cope with changes leads to unjust death

Case 3: A large channel dealer wanted to build a model to accurately predict the sales of mobile phones and tablets to avoid backlogs. He changed 5 model builders in succession, but none of them were satisfactory! The feedback from the business was: the prediction was not accurate enough, which led to wrong decisions.

After careful research, we found that the problem was not the prediction at all, but the business side's repeated jumps. The evaluation of the model effect is based on the total sales volume, but after the total sales volume is allocated to each channel manager, there are always people who jump out to ask for an increase or decrease in the volume. Moreover, if the sales volume is good in the first two weeks, they will try their best to increase the volume, resulting in a backlog. If the sales volume is poor in the first two weeks, they don't want to do anything and will just give up if they can. In the end, the overall data has a large deviation, and they blame the algorithm for inaccurate predictions.

Knowing how these idiots acted, an optimization plan was developed. After the optimization, the effect was immediate: 90% of the so-called inaccurate predictions were caused by the business side’s own unreliable negotiations, predictions, and tricky operations. Not only did it successfully get out of the situation, but it also helped the five innocent people who died earlier to clear their grievances (as shown in the following figure).

04 Poor data quality, very frustrating

Case 4: A large enterprise wanted to build intelligent customer service, so they hired a young man at a high salary. However, he found that not only was the original data chaotic, but also the most basic classification labels: consultation, complaint, and suggestion were all messed up because the customer service training was too poor. As a result, he left the company after working for half a year without any results.

Case 5: A large enterprise wanted to build a "content recommendation algorithm similar to TikTok" and hired a young man at a high salary. However, he found out that there were no content classification labels in the company and the labels given by users were all garbage, 90% of which were empty... The boss also said, "I've paid you so much money, why can't you do it? Why do you need your help? Don't you see that TikTok is all done by algorithm engineers?"

╮(╯▽╰)╭

Yes, the more people believe in algorithm models, the less they value data construction. They all say, “You already have algorithms, why do you need data? Isn’t data elementary???”

By the way, some students may have noticed that the cycle of these failures is half a year. Why? Because many algorithm positions are mascots in Internet companies, in order to prove that the company is on the "road of artificial intelligence" and maintain the stock price. Therefore, the assessment in Internet companies is far less strict than that in physical enterprises. In physical enterprises, if you don't achieve results in half a year, you will be fired.

05 The root cause of the problem

The essence of the problem is that data modeling is essentially fighting against inefficiency. It helps people solve problems that are difficult to handle due to the large number of operational variables, complicated manual calculations, and so on. It is a method of calculation, not a mysterious power with wisdom higher than that of ordinary people, nor a hermit with a fairy-like appearance and bones. The best application areas of data modeling are not diagnosing business problems, but relatively objective areas such as image recognition and speech conversion.

The problems faced by traditional enterprises are:

• There are many unexpected situations: the weather forecast predicts rain, so we have less stock, but it suddenly stops raining, and there is not enough stock to sell...
• Unclear goals: A certain product was put on the market because the boss liked it, but the boss ended up making a mistake...
• Poor business ability: inaccurate prediction, emotional, accepting kickbacks from customers and suppliers, trying to please the boss and take credit

These messy situations are more suitable to be solved by data analysis methods. Data analysis essentially fights against uncertainty. It is through careful data collection, business process sorting, business problem diagnosis, and data testing. Put subjective assumptions into a cage. Replace "I think" with "I am sure". Therefore, when encountering complex business management problems, the best approach is to do a good job of data collection, build analysis models, and accumulate analysis experience bit by bit. Instead of expecting an alpha dog to bark and clear the clouds and usher in the return of spring.

So we can see that as long as the complex scenarios are sorted out and the messy factors are removed, the model can solve the business problems to a certain extent. Unfortunately, from the articles in the circle of friends, to the inner thoughts of the management, to the keyboard of the guy who is adjusting the parameters, all the voices are:

Algorithms beat humans again!

The algorithm knows you better than you know yourself!

The algorithm achieved 99% super accurate prediction!

So this kind of tragedy will continue to happen, and as a large number of companies accelerate digitalization, there will be more and more tragic ones. Let's wait and see.