Collaborative Filtering Applied to Educational Data Mining

Team Leader

Michael Jahrer
commendo research

Team Members

Andreas Töscher
commendo research


Supplementary online material

Provide a URL to a web page, technical memorandum, or a paper.

No response.


Provide a general summary with relevant background information: Where does the method come from? Is it novel? Name the prior art.

We use an ensemble of collaborative filtering techniques originally developed for the Netflix competition and adopt them for educational data mining. In special we use the idea of matrix factorization, which is successfully used for collaborative filtering and adapt it to include all provided information. Additionally we factorize student/step/group relationships and purely neighborhood based relationships. A neural network combines an ensemble of predictions stemming from the mentioned methods and some baseline predictions.


Summarize the algorithms you used in a way that those skilled in the art should understand what to do. Profile of your methods as follows:

Data exploration and understanding

Did you use data exploration techniques to

  • [not checked]  Identify selection biases
  • [not checked]  Identify temporal effects (e.g. students getting better over time)
  • [checked]  Understand the variables
  • [checked]  Explore the usefulness of the KC models
  • [not checked]  Understand the relationships between the different KC types

Please describe your data understanding efforts, and interesting observations:

We use all provided information in our factor models.


Feature generation

  • [not checked]  Features designed to capture the step type (e.g. enter given, or ... )
  • [not checked]  Features based on the textual step name
  • [not checked]  Features designed to capture the KC type
  • [not checked]  Features based on the textual KC name
  • [checked]  Features derived from opportunity counts
  • [not checked]  Features derived from the problem name
  • [checked]  Features based on student ID
  • [not checked]  Other features

Details on feature generation:

The features were used at the blending stage.

Feature selection

  • [not checked]  Feature ranking with correlation or other criterion (specify below)
  • [not checked]  Filter method (other than feature ranking)
  • [not checked]  Wrapper with forward or backward selection (nested subset method)
  • [not checked]  Wrapper with intensive search (subsets not nested)
  • [not checked]  Embedded method
  • [not checked]  Other method not listed above (specify below)

Details on feature selection:

We do not use feature selection.

Did you attempt to identify latent factors?

  • [not checked]  Cluster students
  • [not checked]  Cluster knowledge components
  • [not checked]  Cluster steps
  • [checked]  Latent feature discovery was performed jointly with learning

Details on latent factor discovery (techniques used, useful student/step features, how were the factors used, etc.):

We use matrix factorization which determines the latent features during the learning process.

Other preprocessing

  • [checked]  Filling missing values (for KC)
  • [not checked]  Principal component analysis

More details on preprocessing:

Missing KCs are represented via a separate KC.


Base classifier

  • [not checked]  Decision tree, stub, or Random Forest
  • [not checked]  Linear classifier (Fisher's discriminant, SVM, linear regression)
  • [not checked]  Non-linear kernel method (SVM, kernel ridge regression, kernel logistic regression)
  • [not checked]  Naïve
  • [not checked]  Bayesian Network (other than Naïve Bayes)
  • [not checked]  Neural Network
  • [not checked]  Bayesian Neural Network
  • [not checked]  Nearest neighbors
  • [checked]  Latent variable models (e.g. matrix factorization)
  • [checked]  Neighborhood/correlation based collaborative filtering
  • [not checked]  Bayesian Knowledge Tracing
  • [not checked]  Additive Factor Model
  • [not checked]  Item Response Theory
  • [not checked]  Other classifier not listed above (specify below)

Loss Function

  • [not checked]  Hinge loss (like in SVM)
  • [checked]  Square loss (like in ridge regression)
  • [not checked]  Logistic loss or cross-entropy (like in logistic regression)
  • [not checked]  Exponential loss (like in boosting)
  • [not checked]  None
  • [not checked]  Don't know
  • [not checked]  Other loss (specify below)


  • [not checked]  One-norm (sum of weight magnitudes, like in Lasso)
  • [checked]  Two-norm (||w||^2, like in ridge regression and regular SVM)
  • [not checked]  Structured regularizer (like in group lasso)
  • [not checked]  None
  • [not checked]  Don't know
  • [not checked]  Other (specify below)

Ensemble Method

  • [not checked]  Boosting
  • [not checked]  Bagging (check this if you use Random Forest)
  • [checked]  Other ensemble method
  • [not checked]  None

Were you able to use information present only in the training set?

  • [not checked]  Corrects, incorrects, hints
  • [not checked]  Step start/end times

Did you use post-training calibration to obtain accurate probabilities?

  • [not selected]  Yes
  • [selected]  No

Did you make use of the development data sets for training?

  • [not selected]  Yes
  • [selected]  No

Details on classification:

No response.

Model selection/hyperparameter selection

  • [not checked]  We used the online feedback of the leaderboard.
  • [checked]  K-fold or leave-one-out cross-validation (using training data)
  • [not checked]  Virtual leave-one-out (closed for estimations of LOO with a single classifier training)
  • [not checked]  Out-of-bag estimation (for bagging methods)
  • [not checked]  Bootstrap estimation (other than out-of-bag)
  • [not checked]  Other cross-validation method
  • [not checked]  Bayesian model selection
  • [not checked]  Penalty-based method (non-Bayesian)
  • [not checked]  Bi-level optimization
  • [not checked]  Other method not listed above (specify below)

Details on model selection:

We used a 8 fold cross validation for model selection.


Final Team Submission

Scores shown in the table below are Cup scores, not leaderboard scores. The difference between the two is described on the Evaluation page.

A reader should also know from reading the fact sheet what the strength of the method is.

Please comment about the following:

Quantitative advantages (e.g., compact feature subset, simplicity, computational advantages).

No response.

Qualitative advantages (e.g. compute posterior probabilities, theoretically motivated, has some elements of novelty).

No response.

Other methods. List other methods you tried.

No response.

How helpful did you find the included KC models?

  • [not selected]  Crucial in getting good predictions
  • [selected]  Somewhat helpful in getting good predictions
  • [not selected]  Neutral
  • [not selected]  Not particularly helpful
  • [not selected]  Irrelevant

If you learned latent factors, how helpful were they?

  • [selected]  Crucial in getting good predictions
  • [not selected]  Somewhat helpful in getting good predictions
  • [not selected]  Neutral
  • [not selected]  Not particularly helpful
  • [not selected]  Irrelevant

Details on the relevance of the KC models and latent factors:

The latent factors are most important for high accuracy. KC information can be used to achieve higher accuracy.

Software Implementation


  • [checked]  Proprietary in-house software
  • [not checked]  Commercially available in-house software
  • [not checked]  Freeware or shareware in-house software
  • [not checked]  Off-the-shelf third party commercial software
  • [not checked]  Off-the-shelf third party freeware or shareware


  • [checked]  C/C++
  • [not checked]  Java
  • [not checked]  Matlab
  • [not checked]  Python/NumPy/SciPy
  • [not checked]  Other (specify below)

Details on software implementation:

No response.

Hardware implementation


  • [not checked]  Windows
  • [checked]  Linux or other Unix
  • [not checked]  Mac OS
  • [not checked]  Other (specify below)


  • [not selected]  <= 2 GB
  • [not selected]  <= 8 GB
  • [selected]  >= 8 GB
  • [not selected]  >= 32 GB


  • [checked]  Multi-processor machine
  • [checked]  Run in parallel different algorithms on different machines
  • [not checked]  Other (specify below)

Details on hardware implementation. Specify whether you provide a self contained-application or libraries.

No response.

Code URL

Provide a URL for the code (if available):

No response.

Competition Setup

From a performance point of view, the training set was

  • [not selected]  Too big (could have achieved the same performance with significantly less data)
  • [selected]  Too small (more data would have led to better performance)

From a computational point of view, the training set was

  • [not selected]  Too big (imposed serious computational challenges, limited the types of methods that can be applied)
  • [selected]  Adequate (the computational load was easy to handle)

Was the time constraint imposed by the challenge a difficulty or did you feel enough time to understand the data, prepare it, and train models?

  • [not selected]  Not enough time
  • [not selected]  Enough time
  • [selected]  It was enough time to do something decent, but there was a lot left to explore. With more time performance could have been significantly improved.

How likely are you to keep working on this problem?

  • [not selected]  It is my main research area.
  • [not selected]  It was a very interesting problem. I'll keep working on it.
  • [selected]  This data is a good fit for the data mining methods I am using/developing. I will use it in the future for empirical evaluation.
  • [not selected]  Maybe I'll try some ideas , but it is not high priority.
  • [not selected]  Not likely to keep working on it.

Comments on the problem (What aspects of the problem you found most interesting? Did it inspire you to develop new techniques?)

No response.


List references below.

No response.