"Our algorithm engineers are too poor and can't solve the problem at all!" As a Party B who often deals with traditional enterprises, Mr. Chen has heard too many complaints, and seen too many similar painful pictures. Today we talk about the system.
The model is not great, great! Look at Alpha Big Dog biting the genius Ke Jie, can it be great? As a result, many companies gritted their teeth and stomped their feet and offered high salaries, hiring algorithm engineers, data mining engineers, and data modelers from major Internet companies, hoping that they could make super powerful models. "As long as you can predict accurately, then I will definitely be able to get in the water" is their mantra.
It just so happened that a group of algorithm engineers who were mixed into the so-called Internet giants around 2017 were laid off, thinking that they could harvest a wave of traditional companies under the banner of "the advanced algorithm engineer of the front". From then on, the black chicken became a phoenix and embarked on life. peak. The two hit it off. The tragedy begins here...
1. Regardless of business
Death case 1: A traditional company wants to build a product recommendation model to accurately match user needs. As a result, the algorithm was fired in just six months. Reasons for speculation: Recommendations are not accurate, which interferes with normal sales. The head of the marketing department of Party A said with disdain: Ali's recommendation algorithm is not good.
After studying the business scenario carefully, I found out: Dear, it's not that Ali is the problem, but your company is not Ali. Ali is the platform side, and there are countless products waiting to be promoted on the platform. But specific to your company, you will find:
1. Some of the products are popular products that are easy to sell without pushing them.
2. Some products are the heart and soul of the business, as long as there is a problem, it will be a thousand cuts.
3. Some products are inherently short-legged, have poor functions, and are unreasonably priced. They can't do competing products at all, and the recommendation algorithm is useless.
4. Some products are of good quality, but their internal political status is not high, resources are not available, or pricing is unreasonable, resulting in short legs.
The previous little brother of the algorithm, regardless of these business battles, went straight to the model. All products are stewed in one pot to make recommendations (collaborative filtering is still used, without considering the company's user stickiness and user behavior data issues). As a result, the main product declined, and the sales department and the marketing department joined forces to throw the pot on him. The ending was not only driven out, but also notorious.
After careful analysis of these backgrounds, an optimization plan was released (as shown below):
Do a good job of product analysis first, choose the short-legged small category, and find the department that endorses it. Then you can start doing it. Sure enough, the first wave of promotion took effect immediately. So Party A happily took over and went back to optimize the iteration.
Second, if you don’t refine the scene, you will die in trouble
Death case 2: A chain store hopes to build a model to accurately predict the fish eggs, rice rolls, rice balls, breads of each store... specific to the sales of each SKU, so that the store will not waste ingredients due to the backlog, nor will it Missed sales because of out of stock. As a result, the seven modeling brothers were not accurate enough after half a year. Four of them resigned, leaving three depressed. How exactly is it 100% accurate!
If you think about this problem scene carefully, you will find it very funny: you really have the ability to predict fish ball sausages with 100% accuracy. These seven brothers are still doing a job and go directly to the futures. After careful study, I found that the so-called "out of stock missed sales" is simply empty talk. Because there is no formal out-of-stock registration system (many companies do, but this one does not). However, the attrition rate caused by the backlog is sturdy and high, so an optimized plan was released (as shown in the figure below).
After running in this way for two months, the wastage rate has dropped significantly, and the cost reduction has really been seen. At the same time, although some people complained: "Oh, some stores are out of stock." But what about the evidence? What about the evidence? The evidence! There is no data, nothing but a vernacular, a ghost letter! So the situation was reversed smoothly. Not surprisingly, Party A took over by himself and continued to optimize (Yes, Party A just doesn't like signing the second and third phases, and thinks that he can deal with the latter, of course, this is a later story, haha).
3. Do not respond to changes, including unjust death
Case of death 3: A large-scale channel company hopes to build a model to accurately predict the sales of mobile phones and tablets to avoid backlogs. I changed 5 models one after another, but I was not satisfied! The feedback from the business is that the forecast is not accurate enough, which leads to decision-making errors.
After careful study, it was found that the problem was not in the forecast at all, but in the repeated horizontal jumps of the business side. To evaluate the effect of the model, we look at the total sales volume, but after the total sales volume is allocated to the person in charge of each channel, someone always jumps out and asks for an increase or decrease. Moreover, it is often seen that the first two weeks of buying are good, and then desperately increasing, resulting in a backlog. I don't want to do it for the first 2 weeks, and I just want to. In the end, the overall data has a large deviation, but it turns back to blame the algorithm for inaccurate prediction.
Knowing what the grandsons were doing, an optimization plan came out. After optimization, the effect is immediate: 90% of the so-called inaccurate predictions are caused by unreliable negotiations, pre-judgments, and harassment by the business side. Not only got away smoothly, but also helped the five unjust dead in front of them to clean up their grievances (pictured below).
4. Poor data quality and anxious death
Death case 4: A large company wanted to establish an intelligent customer service. He hired a high-paying brother and found out that not only was the original data confusing, the customer service training was poorly done, and even the most basic classification labels: consultation, complaint, The suggestions are all confusing. As a result, it was naturally no results after half a year of work, so I got out.
Death case 5: A large company wanted to build a “content recommendation algorithm like Douyin”. He hired a high-paying brother and found out that there were no content classification tags inside, and the tags that users typed were all rubbish. 90 % Are all empty...The leader also said: "I have given you so much money, why can't you do it, why do you need a little girl to help, do you think people do Tik Tok by algorithm engineers?? "
╮(╯▽╰)╭
It is Di, the more superstitious the algorithm model, the less emphasis on data construction, all with a face: "You have an algorithm, what do you want data for, isn't the data elementary and low-level??"
By the way, some students should have noticed that these completed cycles are half a year. Why, it is because of many algorithm positions, which 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 of Internet companies is far less strict than that of physical enterprises. No performance in the entity enterprise for half a year, so what if you don’t get out.
5. The apparent cause of the problem
The above scenario, if you change to a data mining engineer who entered the industry around 2010, it will not exist at all. Because most of the data mining engineers of that era were doing telecommunications and banking projects, they had a solid grasp of data analysis methods and were very cautious about the effective scenarios of data models. However, firstly, these people are not moved easily, and secondly, they know how to do it. Take a look at your business:
- Leaders expect too much
- Business
- Do not understand the basic principles
- Poor data base
- Anxious to produce performance
- Lack of clear goals
They will never come!
So there is the beginning picture. The wave of artificial intelligence that began in 2017 has attracted a large number of newcomers into the field of data mining and algorithms. Quite a few people have no basis in data analysis at all. They read "Watermelon Book" + "Statistical Learning Methods", and do "Titanic", "Iris", and "Boston House Price". If they encounter problems, they can use the model and do it. It's over with Aoli! In this case, the blind man is riding a blind horse, and he is in the deep pool in the middle of the night.
6. The essential cause of the problem
The essence of the problem lies in the fact that data modeling essentially combats inefficiency. It is to help people solve the problem of outdated calculation variables, complicated manual calculations, and difficult to handle. This is a calculation method, not a mysterious power with wisdom higher than that of ordinary people, and not a worldly expert with the bones of a fairy wind and crane. The best application of data modeling is not in diagnosing business problems, but in relatively objective areas such as image recognition and voice conversion.
The problems faced by traditional enterprises, such as:
- There are many emergencies: there is rain in the weather forecast, so there are fewer stocks, but suddenly there are not enough goods to sell...
- The goal is not clear: because the boss personally likes it, so he got on a certain product, but the boss looked away...
- Poor business ability: inaccurate in anticipation, emotional, received rebates from customers and suppliers, and behaved like the boss to take credit
These messy situations are more suitable to be solved by data analysis methods. Data analysis essentially fights against uncertainty. It is through earnestly collecting data, sorting out business processes, diagnosing business problems, and conducting data tests. Put your subjective assumptions in a cage. Replace "I thought" with "I'm sure." Therefore, when encountering complex business problems, the best way is to do a good job of data collection, carefully establish analysis models, and accumulate a little bit of analysis experience. Instead of expecting a big Alpha dog to bark and greet the spring.
So we see that as long as the complex scenes are sorted out clearly and the messy factors are excluded, the model can solve business problems to a certain extent. Unfortunately, from the Moments article, to the management's heart, to the keyboard of the brother who is tuning, all the voices are:
- Algorithms beat humans again!
- The algorithm knows yourself better than you!
- The algorithm has achieved 99% super accurate prediction!
So this kind of tragedy will continue to happen, and as a large number of companies accelerate their digitalization in 2020, there will be more and more tragic events. We will wait and see.
Finally, some students said: Mr. Chen, you are all examples of physical enterprises, so Internet enterprises are pure land. Ha ha! Not to mention, just talk about fresh food e-commerce, the impact of the epidemic, everyone thinks fresh food e-commerce has a future, so a bunch of algorithm engineers who have not even cooked rice and have never given birth to children are working hard to study "vegetables". The "precise recommendation" "shopping smart prediction" algorithm.
Yes! Use familiar collaborative filtering, or use familiar association rules. As for the result, we will find an opportunity to make complaints. If you are interested, follow the public account [Grounding Qi School]. In the next article, we will share a detailed modeling process, so stay tuned.
Author: Chen grounded gas, micro-channel public number: down to earth school. A data analyst with ten years of experience and CRM experience in multiple industries.