A Tool for fitting Hidden Markov Models models at scale
This command-line tool is developed to fit Hidden Markov Models (HMM) from the large datasets. In particular, it is targeting a specific kind of HMM used in education called Bayesian Knowledge Tracing (BKT) model. The utility is written in C/C++. The project started while the author was a post-doc at Human-Computer Interaction Institute, Carnegie Mellon University. These sources are published under BSD-new (3-clause BSD) license [https://en.wikipedia.org/wiki/BSD_licenses#3-clause_license_.28.22Revised_BSD_License.22.2C_.22New_BSD_License.22.2C_or_.22Modified_BSD_License.22.29].
Bayesian Knowledge Training (BKT) is a user modeling approach widely use in the area of Intelligent Tutoring Systems (ITS). In ITS, it is customary to tag problems and problem steps students are working on with knowledge quanta (also called skills). When a student is attempting to solve a problem or a problem step, they are said to practice a skill (skills). A student model (BKT being one) performs deductions on whether the skill(s) is (are) mastered given a sequence of correct and incorrect attempts to solve problems (problem steps). BKT uses a formalism of a Hidden Markov Model to make the estimation of student's skill mastery. In BKT, there are two types of nodes: binary state nodes capture skill mastery (one per skill) that assume values of mastered or not mastered and binary observation nodes (one per skill) that assume values of correct or incorrect.
Each skill in standard BKT has the following parameters.
- p-init or p(Lo) - is a probability the skill was known a priori,
- p-learn or p(T) - is a probability the skill will transition into mastered state after a practice attempt,
- p-forget or p(F) - is a probability that the skill will transition into not mastered state after a practice attempt. Traditionally, p-forget is set to zero and is not counted towards the total number of parameters.
- p-slip or p(S) - is a probability that a mastered skill is applied incorrectly, and
- p-guess or p(G) - is a probability that unmastered skill will be applied correctly.
BKT parameters can be represented in the vector/matrix form. A vector of priors – – contains the p-init parameter. There are N elements in the priors vector (N=2 in standard BKT). Throughout this discussion, we will we assume that p-init is the first element of the
vector. p-learn is part of the transitions matrix A that captures probabilities of mastery state changes and its size is N*N. No forgetting is captured by setting A[1,2] = 0 (from mastered to unmastered). p-learn corresponds to A[2,1] - from unmastered to mastered. p-guess and p-slip are specified in the observation matrix B that is N*M, where M is the number of observations (M=2 in standard BKT). The first column corresponds to a correct observation, the second – to the incorrect. In standard BKT with two observations, p-guess is B[2,1] – unmastered skill but a correct response, and p-splip is B[1,2] - mastered skill but an incorrect response.
For more details on BKT refer to [1]. [2], among other things, discusses how a gradient-based fitting of HMM can be implemented. [3] and [4] cover additional topics relevant for the implementation.
A public version of the standard BKT code is available via as a GitHub repository. If you have git
utility installed, use the following command to get the source code.
git clone https://github.com/myudelson/hmm-scalable
To compile issue make all
command. If you are on Linux, the g++/gcc
compiler and Open MP library should already be installed. On Mac OX, you would need Xcode command line tools to be installed via xcode-select --install
command. g++/gcc
compiler with Open MP (no longer bundled with Mac OS X by default) could be downloaded from hpc.sourceforge.net. If you are on Windows, you might need to install cygwin
and have g++/gcc
compiler available and be sure to install make
utility as well.
Input data file should be a text file with Unix end-of-line encoding. The format is four tab-separated columns: observation, student, problem/problem step, skill(s). Observation is a 1-started integer. For the two-state BKT, we advise using 1 for correct and 2 for incorrect. The student is a string label, so is a problem or a problem step, whatever granularity you prefer. Skill is a string label. Multiple skill labels should be delimited by a character of your choice (do not use tab). An example with a few lines of input is below. Here, tilde symbol ~
is used as a delimiter.
-- input file --
2 student_001 unit1-section1-problem5-step1 addition~multiplication
1 student_001 unit1-section1-problem5-step2 multiplication
1 student_001 unit1-section1-problem5-step3 addition
-- input file --
If there is no skill label for a particular row of data use .
(dot) symbol. In test data, the utility will use known observations for training and will produce predictions for missing observations that should have .
(dot) instead of the observation code (1, 2, or otherwise).
Output files are the model file and the predictions file. The model file contains general information about the data, e.g., the number of observations, the number of skills and the number of problems/problem steps; as well as the model parameters.
Prediction file consists of model predictions for each row of the input file. Depending on the option (see parameter specifications below). You can print out probability distribution for student's response (the probability of correct and the probability of incorrect, if the observation node is binary) and, additionally, print out the probability distributions over the values of the hidden state for all skills specified for that particular data point. The probability values are tab separated. See examples of a outputs file below.
-- model file --
SolverId 1.1
nK 1
nG 1
nS 2
nO 2
Null skill ratios 1.0000000 0.0000000
0 multiplication-skill
PI 0.45000000 0.55000000
A 1.00000000 0.00000000 0.40000000 0.60000000
B 0.79000000 0.21000000 0.22000000 0.78000000
-- model file --
In the model file example above, we see a specification of the solver algorithm (to be discussed below), one skill (nK=1), one student in the data (nG=1), two hidden states and 2 observations (nS and nO respectively). Skill numbering is zero-started. Skill name is "multiplication-skill". Rows prefixed with PI, A, and B correspond to priors, transition, and observation matrices written row by row. Thus, the first element of PI is p(Lo)=0.55, 3rd element of A is p(T)=0.4, 2nd and 3rd elements of B are p(S)=0.12 and p(G)=0.22 respectively.
Prediction file, for a standard BKT model with nO=2 observations, would contain 2 tab-separated columns. The first column corresponds to the model-estimated prior probability of the student being correct for that particular observation. The second column is the complement of the probability of correct -- the probability of being incorrect (see an example below).
-- prediction file --
0.73 0.27
0.88 0.12
0.94 0.06
0.99 0.01
-- prediction file --
trainhmm
utility has the following launch signature:
trainhmm [options] input_file [[model_file] prediction_file]
options:
-s
: structure.solver[.solver setting], structures: 1-by skill, 2-by user; solvers: 1-Baum-Welch, 2-Gradient Descent, 3-Conjugate Gradient Descent; Conjugate Gradient Descent has 3 settings: 1-Polak-Ribiere, 2-Fletcher–Reeves, 3-Hestenes-Stiefel, and 4-Dai-Yuan. For example -s 1.3.1
would be by skill structure (classical) with Conjugate Gradient Descent and Hestenes-Stiefel formula, -s 2.1
would be by student structure fit using Baum-Welch method.
-S
: perform scaling of forward/backward variables: 0 - off (default), 1 - on. Only allowed for Baum-Welch solver (-s 1.1
setting), otherwise auto-set to off.
-e
: tolerance of termination criterion (0.01 for parameter change default); could be computed as the change in log-likelihood per data point, e.g. -e 0.00001,l
.
-i
: maximum number of iterations (200 by default)
-q
: quiet mode, without output, 0-no (default), or 1-yes
-n
: number of hidden states; should be 2 or more (default is 2)
-0
: initial parameters comma-separated for priors, transition, and emission probabilities skipping the last value from each vector (matrix row) since they should sum up to 1; default is 0.5,1.0,0.4,0.8,0.2
.
-l
: lower boundaries for parameters, comma-separated for priors, transition, and emission probabilities (without skips); default is 0,0,1,0,0,0,0,0,0,0
.
-u
: upper boundaries for params, comma-separated for priors, transition, and emission probabilities (without skips); default is 1,1,1,1,1,1,1,0.3,0.3,1
with slip and guess capped at 0.3.
-c
: specification of the C weight and centroids for the L2 penalty, empty (default). For standard BKT - 4 comma-separated numbers: C weight of the penalty and centroids, for PI, A, and B matrices respectively. If used for iBKT with student effects, 8 values will be used with 4 additional values for student matrices. For example, -c 1.0,0.5,0.5,0.0
.
-f
: fit as one skill, 0-no (default), 1 - fit as one skill and use params as starting point for multi-skill, 2 - force one skill
-m
: report model fitting metrics (AIC, BIC, RMSE) 0-no (default), 1-yes. To specify observation for which metrics to be reported, list it after ,
. For example -m 0
, -m 1
(by default, observation 1 is assumed), -m 1,2
(compute metrics for observation 2). Incompatible with -v
option.
-v
: cross-validation folds, stratification, and target state to validate against, folds input/output file, default 0 (no cross-validation), examples -v 5,i,2
- 5 fold, item-stratified c.-v., predict state 2, -v 10
- 10-fold subject-stratified c.-v. predict state 1 by default, alternatively -v 10,g,1
, -v 5,n,2,folds.txt,o
- 5-fold unstratified c.-v. predicting state 2, [o]output folds to folds.txt
, and here -v 5,n,2,folds.txt,i
, folds are actually read [i]n from the file.
-p
: report model predictions on the train set 0-no (default), 1-yes; 2-yes, plus output state probability; works with -v and -m parameters.
-U
: controls how the probability distribution of the states is updated. Takes the following format -U r|g[,t|g]
, where the first character controls how prediction treats known observations, second – how prediction treats unknown observations, and third – whether to output probabilities of priors. Dealing with known observations r
- reveal actual observations for the update of state probability distribution (makes sense for modeling how an actual system would work), g
– guessing
the observation based on the predicted outcomes (arg max) – more appropriate when comparing models (so that no information about observation is ever revealed). Dealing with unknown observations (marked as .
– dot): t
– use transition matrix only, g
– guess
the observation. Default (if omitted) is -U r,t
. For example, -U g,g
would require guessing of what the observation was using model parameters and the running value of the probabilities of state distributions.
-d
: delimiter for multiple skills per observation; single skill per observation (default), otherwise – delimiter character, e.g. -d ~
.
-b
: treat input file as binary input file created from the text file by the inputconvert
utility.
-B
: block re-estimation of prior, transitions, or emissions parameters respectively (default is -B 0,0,0
), to block re-estimation of transition probabilities specify -B 0,1,0
.
-P
: use parallel processing, default - 0 (no parallel processing), 1 - fit separate skills/students separately, 2 - fit separate sequences within skill/student separately.
-o
: in addition to printing to the console, print output to the file specified in this option (by default is empty and no extra printing is done).
predicthmm
utility has the following launch signature:
predicthmm [options] input_file model_file [predicted_response_file]
options:
-q
: quiet mode, without output, 0-no (default), or 1-yes
-m
: report model fitting metrics (AIC, BIC, RMSE) 0-no (default), 1-yes. To specify observation for which metrics to be reported, list it after ,
. For example -m 0
, -m 1
(by default, observation 1 is assumed), -m 1,2
(compute metrics for observation 2). Incompatible with-v option.
-d
: delimiter for multiple skills per observation; single skill per observation (default), otherwise – delimiter character, e.g. -d ~
.
-b
: treat input file as a binary file created from a text file by the inputconvert
utility.
-p
: report model predictions on the train set 0-no (default), 1-yes; 2-yes, plus output state probability; works with -v
and -m
parameters.
-U
: controls how the probability distribution of the states is updated. Takes the following format -U r|g[,t|g]
, where the first character controls how prediction treats known observations, second – how prediction treats unknown observations, and third – whether to output probabilities of priors. Dealing with known observations r
- reveal actual observations for the update of state probability distribution (makes sense for modeling how an actual system would work), g
- guessing
the observation based on the predicted outcomes (arg max) – more appropriate when comparing models (so that no information about observation is never revealed). Dealing with unknown observations (marked as .
– dot): t
– use transition matrix only, g
– guess
the observation. Default (if omitted) is -U r,t
. For examle, -U g,g
would require guessing
of what the observation was using model parameters and the running value of the probabilities of state distributions.
Small sample data file toy_data.txt
is generated using the following BKT parameters: pLo=0.4, pT=0.35, pS=0.25, pG=0.12.
To fit a BKT model of this data using an EM algorithm run the following command:
./trainhmm -s 1.1 -m 1 -p 1 toy_data.txt model.txt predict.txt
The model will have 90% accuracy and root mean squared error (RMSE) = 0.302691 and the recovered BKT parameters would be: pLo=0.00000000, pT=0.16676161, pS=0.00044059, pG=0.00038573. Overall loglikelihood, actually, goes up from 9.3763477 to 10.4379501 in 3 iterations.
If we fit BKT model using Gradient Descent method using -s 1.2
argument, the recovered parameters would be: pLo=0.00000000, pT=0.3095828498, pS=0.18955371, pG=0.16195590, the accuracy would remain at 90% while RMSE = 0.344145. Loglikelihood changes from 9.3763477 to 8.0992564 after 19 iterations.
To generate predictions using a previously fit model run the following command:
./predicthmm -p 1 toy_data_test.txt model.txt predict.txt
To give this tool a proper test you might want to try it on a KDD Cup 2010 dataset donated to the Pittsburgh Science of Learning Center by Carnegie Learning Inc. The dataset can be downloaded (after a quick registration) from here. This dataset consists of training and challenge sets. For the sake of testing the tool, download the challenge Algebra I set that has about 9 million transactions of over 3300 students. The training file should be trimmed to the tool's format. See shell commands below that do that.
gawk -F"\t" 'BEGIN{OFS="\t"} {if(NR==1)next; skill=$20; gsub("~~", "~", skill); if(skill=="") skill="."; else { skill=$3"__"skill; gsub(/~/, "~"$3"__",skill);} print 2-$14,$2,$3"__"$4,skill;}' algebra_2008_2009_train.txt > a89_kts_train.txt
To fit a BKT model of this dataset using gradient descent method as well as to compute fit metrics and the prediction run the following command:
./trainhmm -s 1.2 -d ~ -m 1 -p 1 a89_kts_train.txt model.txt predict.txt
Depending on your hardware, the model should be fit in about 1-2 minutes.
-
Corbett, A. T., and Anderson, J. R.: Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction, 4(4), 253-278. (1995)
-
Levinson, S. E., Rabiner, L. R., and Sondhi, M. M.: An Introduction to the Application of the Theory of Probabilistic Functions of a Markov Process to Automatic Speech Recognition. Bell System Technical Journal, 62(4): 1035-1074. (1983)
-
Wolfe conditions. (2016, January 20). In Wikipedia, The Free Encyclopedia. http://en.wikipedia.org/wiki/Wolfe_conditions
-
Nonlinear conjugate gradient method. (2016, February 22). In Wikipedia, The Free Encyclopedia. http://en.wikipedia.org/wiki/Nonlinear_conjugate_gradient_method
If you have any questions, comments, propositions, feel free to contact me at [email protected].
Have fun fitting BKT models,
Michael (Mikhail) Yudelson