Prize

1st place - $2000

2nd place - $1500

3rd place - $1000

4th place - $750

5th place - $500

Goal

Given a set of parameters of a battery, we want to predict the risk of the battery. The risk is defined as the number of days before the battery becomes a “bad” battery. See more details in the Data section about the definition of risk.

Data

We have collected 102,223 records from 588 batteries. Each battery may have more than one record on different dates. Keep in mind there is a date associated with each record, although it has been removed from the data.

 

Each record has 18 attributes.

1. event_country_code: Country where support contact initiated.

2. batt_manufacturer: Battery manufacturer. It has been encoded to be anonymous.

3. installed_count: Number of batteries in the laptop as reported at the date of this record.

4. batt_instance: Identifies whether this battery is a primary or secondary battery in the laptop.

5. cycle_count: Number of times that battery has been discharged and recharged.

6. temperature: Temperature of the battery at the date of this record.

7. battery_current: Battery electrical current at the date of this record.

8. design_capacity: Design capacity of the battery.

9. full_charge_capacity: Full charge capacity of the battery at the date of this record.

10. remaining_capacity: Remaining battery charge at time of injection date.

11. design_voltage: Design voltage of the battery.

12. batt_voltage: Battery voltage at the date of this record.

13. cell_voltage1: Voltage of battery cell #1 at the date of this record. If the battery contains 2 cells then cell_voltage1 will be 0.  

14. cell_voltage2: Voltage of battery cell #2 at the date of this record.

15. cell_voltage3: Voltage of battery cell #3 at the date of this record.   

16. cell_voltage4: Voltage of battery cell #4 at the date of this record. If the battery contains 2 cells, then cell_voltage4=0 AND cell_voltage1=0.

17. status_register:  Status register of the battery at the date of this record.

18. risk: The risk value is defined as the number of days before this battery becomes a “bad” battery. If the battery is “bad” at the date of this record, the risk value is 0. Otherwise, we will find all its later records and look for the first date for the “bad” status. If there’s no “bad” status afterwards, the risk value is -1.

 

In order to build the training and testing set, we randomly sampled 80% of these batteries as the train set (i.e., “train.csv”), while the remaining 20% become the test set (i.e., “test.csv”). In the test set, the “risk” will be hidden.

Evaluation

The evaluation consists of two parts:

1. Can you classify the current status of the battery accurately?

2. Can you forecast the failure of the battery accurately?

 

First, we categorize the risk value into two types: (1) risk = 0 and (2) risk != 0. We can then calculate the F1 score of this binary classification task by treating (1) as the positive label while (2) as the negative label. This F1 is the first score, denoted as F1.

 

Second, we measure the mean relative absolute error for the risk for the records whose groundtruth risks are greater than 0. We denote the predicted risk as P and the groundtruth risk as G for a record. If P is -1, the relative absolute error is defined as 1. Otherwise, the relative absolute error is defined as min(1, |P - G| / G). The mean relative absolute error becomes the second score, denoted as MRAE.

The final score is defined as F1 + (1 - MRAE). All submissions will be ranked by this final score. If there are ties, we will break the tie by the F1 score.

Final Submission Guidelines

Submissions

You are going to submit a file named as “prediction.csv” that only contains the risk value in the same order as the testing file. The risk value should be integers and one per line corresponding to the “test.csv” file. A wrong format may lead to a score of -1. Please check the example submission for a better understanding.

 

Besides, you will need to submit your code and report. Details are as follows.

• Your source code must be in Python3.

• You should provide a list of dependencies and instructions about how to install them. For example, a “requirement.txt” that we can use pip3 to install is good.

• You should allow us to easily change the training and testing files of the same format. For example, the training and testing file names could be a part of your code’s parameters.

• You should provide a document about how you design your method. For example, you can discuss the models and the features that you have developed or tried, and also justify why you finally ended up with the model.

• You should provide instructions about how to repeat your results. If you cannot avoid some randomness, please justify your reasons.