There will be a lunch for participants and industry mentors in ESB 2012 from 12:00 to 13:00. Afterwards, students will present their work to industry mentors in a 20 minute presentation (plus 10 minutes for questions) in ESB 2012. Please see the table below for a schedule.
Participants: Please send your project titles (and abstracts) to
aberk@math.ubc.ca before 18:00 on Thursday June 7.
| Start | End | Group | Project title |
|---|---|---|---|
| 13:00 | 13:30 | CloudPBX | Quality Analysis of CloudPBX VoIP Calls |
| 13:30 | 14:00 | Comm100 | Predicting Message Paths and Determining Conversation Intent from Live Chat Transcripts |
| 14:00 | 14:30 | SNC-Lavalin | Interpolating Ship Paths |
| 14:30 | 15:00 | –break– | |
| 15:00 | 15:30 | St. Paul’s | St. Paul Shenanigans - Or a deep look into cytokines |
| 15:30 | 16:00 | SSR Mining | Exploratory Analysis and Failure Prediction of SSR Mining Trucks |
CloudPBX is a “Voice over IP” telecommunication service provider. In this work, we address the central problem of evaluating the quality of their VoIP calls. By exploring the call data, we noticed that the current industry standard, MOS, was not sensitive enough to accurately represent the spectrum of call quality - therefore, we devised a new metric. Equipped with this new metric, we explored the correlations between call quality and relevant geographic and technological factors. These findings provide future direction for CloudPBX’s data collection, research, and business partnerships.
We are given a data set from Comm100 and with this data set we predict the intent of conversations as well as cluster each message. We do this based on key word counts and other engineered features. We use the resulting clusters as states to build a Markov Chain transition matrix which models paths of messages between clusters, especially examining the visits to undesirable message states (angry users). We also implement an algorithmic (non-random) decision tree to categorically classify the conversation. We defined eight key topic categories based on the FAQ page from the Comm100 website.
The Port Metro Vancouver aims to reduce air emissions caused by ships, trucks, and equipment. In response to this request, SNC-Lavalin investigates, in particular, the emission patterns caused by using ship engines. This requires an extensive data frame including variables such as the current location, velocity and acceleration over time. This is useful when calculating the pollution and total emission of a ship as they can be computed partly from the velocity and acceleration of the vessel along the path. However, the location of ships are tracked irregularly and the data is corrupted and missing.
In the PIMS BC Data Science Workshop, we aim to approximate the ship path from the measurement. The simplest model (interpolation of the path from consecutive measurements) does not provide the accuracy needed as it does not consider the noise and time between measurements. During this workshop, we developed a method based on the Kalman filter to resolve these issues. The model shows good performance both on current data with dense measurements and old data with sparser measurements. The next stage of the project is to cluster the paths in order to apply the knowledge gathered from densely sampled paths to sparsely sampled paths. This brings new challenges like defining a good metric for paths and number of clusters, which is still ongoing.
Sepsis is the leading cause of death in the intensive care unit worldwide. Despite having many environmental factors, septic shock can be significantly attributed to cytokines which are specialized proteins that regulate inflammation in the body. Understanding the correlation between the cytokine levels, as well as genes that code for them, is crucial in reducing the rates of mortality. We performed genome-wide association studies (GWAS) to determine which SNPs are correlated with mortality of septic patients. We also identified SNPs which are correlated with serum concentrations of various cytokines, determined the cytokines which are correlated with patient mortality, and trained various machine learning classifiers. We have identified several SNPs that correlate with mortality rates including one in the gene of a protein (PP1) that is found to regulate cytokine levels.
We explored a data set of 50 GB split into many categories and sub-components in the SSR mining company’s daily operations. This data records massive amounts of sensor information on the machine components in Hitachi haul trucks as well as weather data, location, job role, load value among many others. With all of this, we were tasked with exploring the alarm types and predicting under what conditions a costly breakdown would occur. Our analysis is multi-staged: first we considered the time series over the sensor data, next we compare when critical alarms signal against their GPS coordinates, lastly we sorted the critical alarms and performed a Principal Component Analysis (PCA) over the state of the truck to determine what the causes of the alarm were. Ultimately, we put together all of the these graphics to suggest a viable approach to build a predictive model for determining when a break-down will occur and what needs to be fixed to prevent this.