数据方向
实习项目
Fraud investigation of employees’ behaviors on imbalanced transaction data using Python, SQL, R and RapidMiner
Extracted~10TB data and performed texts vectorization with word2vee model
Explored effects of bootstrap resampling，synthetic minority oversampling (SMOTE) and cluster centroid sampling algorithms on logistic regression and decision tree (publication under review by International Journal of Managing Information Technology)
Implemented variable clustering and principal component analysis (PCA) to reduce data dimensionality
Displayed the topology of the high-dimensional data using topological data
analysis (TDA)
Created new features based on results from TDA and clustering analysis that improved accuracy by 0.05
Explored the characteristics of the fraud employees using k-means clustering analysis
Measured the effectiveness of the proposed model using Kolmogorov-Smirnov test, lift charts and cumulative gains charts.
Customer usage patterns and anomalous activities detection using Python, SQL
Analyzed over 3GB workforce management data to identify customer usage patterns and anomalous activities using advanced SQL Querying and Python.
Performed exploratory data analysis on numerical, categorical and time series data using Matplotlib and Seaborn.
Constructed multivariance Time Series Clustering model, DTW with Hierarchical clustering, on 120 features to group customers, segment market and described the clustering centroid by DBA (DTW Barycenter Averaging).
Trained both Anomaly Detection Models including isolation forest, robust PCA and Regression Models including linear models, random forest to build customer’ anomalous activities alert system.
Built dashboards to describe the customer’s usage activities via Tableau and weekly reported analysis results to the team and manager.
Assembled an Apache Flume service on a private cloud infrastructure to ingest100+TB/day log files from cloud computing applications and dispatch tagged log events to downstream services all in real-time
Used Kafka as a primary data storage solution to hold short term log data reducing data retrieval latency from 3s to 0.5s
Used Hadoop Distributed File System as a secondary data storage solution to persist PB-level historical log data
Built an alerting module using Spark Streaming to catch error logs from Kafka and notify the internal management system of cloud computing applications achieving a total latency of less than ls
Built an analysis module using Spark Streaming to consume and aggregate log data based on severity levels and generate health reports of all cloud computing applications on a daily basis, improving 5x efficiency.
Performed data crawling using Python to collect 114687 images and trained, built a Faster-RCNN with Wider Face dataset to detect and extract 94% faces using PyTorch.
Trained and applied a SRGAN model to augment and standardize picture resolution from less than 100x100 to 256x256 using PyTorch.
Implemented, trained and applied an 8-layer StyleGAN with PyTorch framework to generate dynamic style of cartoon faces according to input image and commands
Added effects with mixing the styles of the generated images by controlling the latent model spaces
Implemented a diffusion model(ddpm)and trained with 10881 face images.
Constructed GPU infrastructure and implemented parallel computation to speed up 12x run-time performance.
Established an NLP system to summarize the text through TensorFlow in the environment built by Nvidia Docker.
Analyzed and processed 2TB text summarization datasets from THUCnews. LCSTS.CSL news headlinescontexts and judicial summaries by NLTK
Benchmarked the performance of Point-Generator, WoBERT, Nezha, and T5 on the above datasets to obtain proper title/ abstract, guideline and summaries.
Applied BERT WoBERT to improve prediction accuracy by 5.6% and 4.5% respectively
Expedited the inference of these transformer-based models by1.55xfaster via Turbo Transformer
Accelerated the modeling run-time performance by 8xwith parallel computation using CUDA based on GPUs.
Applied the Bert-of-Theseus method to distill WoBERT shrank model size to 50%.
汇聚1300+
在职面试官
Senior级、 大厂、 在职面试官 ，而想要加入导师库，只满足这3个条件还远远不够。
如同同学求职一般，我们的导师也会历经多达5轮的“面试考核”，通过背景经历、技术能力、面试经验、导师sense等维度的层层筛选下，入选率仅10%。有导师戏称，“并不比一次大厂面试简单”。
为了保证导师库水准，我们从未因考核成本而降低标准。辅导内容随着科技届趋势不断迭代更新，对导师的培养和考核标准也会只增不减。
全球各大一线大厂
专属导师1对1直播授课- 精准匹配头部公司在职面试官
- 成功签约全职OFFER为止
- 无OFFER承诺退款
- 简历项目双修双审
- 一线大厂List，在职员工内推
- 附赠真实实习项目
直通硅谷成立于2015年3月，由北大计算机系师兄联合MIT、前百度网络科技产品经理、Harvard高级学者、香港上市公司联席董事共同创立，心之所向，是壮大全球华人力量。 凭借在求职辅导中积累的丰富经验，我们不断研发顺应科技界求职趋势的学练结合课程，组建富有实战经验的国内外名企导师团队，已成功帮助超过8000+学员进入全球一线大厂。
- 热门ALLinONE求职计划
- 软件工程师求职1v1定制计划
- 数据方向求职1v1定制计划
- 人工智能求职1v1定制计划
- 金融数据求职1v1定制计划
