Datawhale dry goods
Author: Student W, Datawhale Excellent Learner
written in front
Hello everyone, I am a classmate W from Xijiao Jiaotong University. Both my undergraduate and master’s degrees are not computer science classes. When I was looking for a summer internship in the spring of 2023, I got good results. I got offers from several Internet companies including Alibaba, Baidu, and Mihayou. , and finally decided to do a summer internship with Ali, and I would like to share my experience with you.
Compared with other people, my preparation time started relatively late. It was around February that I completely gave up my plan to study for a Ph. D. and began to devote myself to the army of job hunting. time began. So from the perspective of test-taking and quick-learning, there are indeed some dry goods and experiences that can be shared. But in terms of long-term accumulation, I have not done enough. I will propose some directions and I will work hard with everyone. This experience post is mainly divided into two parts, dry knowledge and experience sharing.
dry knowledge
In many interviews, the interviewer revealed a message that the requirements of the internship interview will be much lower than that of the autumn recruitment interview. For example, the answer to the scene question is very simple, and some interviewers will say nothing. For example, if Likou didn’t tear it out by hand, some interviewers will say that this question is indeed a bit difficult for interns, so students should not put too much pressure on their preparations, just review the content on their resumes Let’s start posting, the content of ML/DL can’t be finished, learn as you go, and even prepare for lifelong learning.
So based on the above content, I think that among the following three dry goods, the most important thing is dry goods 1, which is the mastery of basic knowledge points. Dry goods 2 and dry goods 3 can only be said to be extra points or some information to chat with the interviewer.
Dry goods 1: Iron still needs to be hard
If you want to apply for a job algorithm internship, here is the first place to pay attention to. Algorithm interviews have some fairly frequent questions, so some knowledge must be mastered, including but not limited to projects written in the resume, stereotyped questions in machine learning and deep learning, and scenario questions that examine the ability to understand actual business. Check out the code questions.
Here are some frequently asked knowledge points and some recommended information:
Machine learning knowledge: overfitting, underfitting, optimizer, feature engineering, evaluation indicators, common ML algorithms (classification/regression/decrease), the models that are often asked are GBDT, XGBoost, and they will be asked Algorithm principles, tree building process, advantages and disadvantages, improvement methods, etc. Here I recommend Baimian machine learning , Niuke.com algorithm engineer question bank , and statistical learning methods .
In-depth learning knowledge: The specific content of the examination is related to the job interview. The promotion search will ask about the twin-tower model, single-objective DIN, DIEN, multi-objective MMOE, and ESSM. Generally speaking, I will examine some basic DL models, such as CNN, RNN, and LSTM. I will also examine content such as Attention, Transformer, and RL. I will ask some network models, backpropagation, and internship interviews will not be too difficult. At most, I will write LR by hand. The loss function and gradient, but I saw in other people's interviews that it seems that the difficulty of school recruitment will increase to handwriting SVM and CNN. I recommend reading Baimian deep learning and classic papers in the field of your interview position .
Scenario questions : I will ask some questions related to the actual business, such as ab testing, the reasons for the inconsistency between online and offline, and sometimes the interviewer will also describe to you the problems they are currently solving, and let you provide some ideas and solutions , come to inspect (for nothing) you, and recommend our Datawhale's fun-rec open source tutorial .
Li button: Hand tearing the code has basically become the repertoire of every interview, and it is also one of the basis for measuring whether the interview is passed. You can use tag classification, routine classification, and high-frequency question list to brush. The force button in the algorithm post will not be too difficult, and it will be around medium. According to the above method, you can brush about 150-200 questions. question.
Here I suggest that you don’t just understand a model when you are preparing, but try to retell the knowledge points, because it is an online video interview, you can’t really give the interviewer a blank sheet of paper to push these models, or Writing pseudo-code is either dictation, so it is very important to have a clear idea and be detailed. For example, when answering XGBoost, first answer the two points of building a tree based on the residual and being greedy, and then look at the situation later or follow the interviewer’s questions. Expand the second-order taylor and supplement the loss function.
Dry goods 2: Focus on cutting-edge technology sharing
Being able to track cutting-edge technologies in the industry is also an ability that an algorithm engineer should have. From the perspective of an internship interview, the interviewer will hope that an intern who understands what he is doing now can help him share some tasks after he comes in. During the interview, he will also ask questions about what he is doing, because this is what he knows best. As interns, we also need to pay attention to what the industry is doing.
I learned about Meituan technology sharing (pdf manual), Taobao technology sharing (pdf manual), and Xiaohongshu technology sharing (video at station b). The general direction of the industry as a whole is the integration of search and push, the integration of different modes, the integration of rough and fine sorting, and large models. steps.
Dry goods 3: Pay attention to the development context between models
This is also one of the points of investigation, which is the development trajectory of the model. Several multi-objective models, hard-parm-sharing, to mmoe, to ple, to esmm, to esm2, why should they be optimized in this way? In actual business, each optimization is to solve some problems, or actually use Among them, which one is better between the two parallel models.
These questions are relatively difficult, and they are also open questions. It is a process of mutual discussion between the interviewer and the person being interviewed. If you answer these questions well, the interviewer may think that you know the whole picture of the business, and you will have a high score. Add points.
Experience Sharing
Experience 1: Multi-investment and multi-faceted
Every interview has many uncertain elements, such as whether you can chat with the interviewer, whether the department still has HC, and whether it matches the position. Uncertainty can never be eliminated, and the way to seek certainty is to invest in multiple perspectives, spend time on different companies, and put eggs in different baskets.
Experience 2: Active communication during the interview process
The focus of the interview is job matching, and the core is self-presentation. Actively communicating with the interviewer will not only let him understand your technical level better, but also make him think that he can work smoothly with you in the future. Try to avoid the form of one question and one answer, but be careful not to do it by grabbing words, but by adding appropriate supplements and examples. For example, if the interviewer asks you question a, you can’t answer it, but you can say that you have learned about the more relevant question b before, what is its thinking, and you can use your strengths to avoid weaknesses.
Lesson 3: Prepare for a protracted war
The interview process is generally procrastinating, and the few who can solve the battle within a week are considered fast. During this time, maintain your own rhythm, eat, sleep, and exercise. If the result is not good enough, it must be because the final point has not been reached, and if you persist, a bowl of simple chicken soup will be given to everyone.
I hope that those who read this article can gain something
thank you
Finally, I would like to thank the fourth brother and Ran Geha of the Datawhale employment team. I participated in the mock interview organized by Datawhale on March 12, 2023. I did not ask boring stereotypes, but put forward corresponding questions based on the background of the interviewer. , Let me realize a lot of problems before the official start of the interview, and I would like to express my sincere thanks to them.
It's not easy to organize, so I like it three times ↓