Multi-temporal information object incremental learning software system
An incremental author disambiguation framework may create new clusters to accommodate new data based on the existing cluster results and newly added data. The proposed system may provide frequent update of taxonomic classification, name disambiguation and many other applications because it takes less time to generate new results. In addition, the proposed methods may reduce the time needed for updating the model and help improve the performance with the limited computational resource.
1. A system for classifying samples comprising:
a processor, a memory accessible by the processor, and computer instructions stored in the memory and executable by the processor to perform:
store a pre-existing training data cluster;
apply a Naive Bayes Classifier (NBC) to data within the training data cluster; receive new data; and
apply the NBC to the new data based on the NBC applied to training data;
wherein the NBC includes a classification algorithm used to classify the data with the NBC;
wherein the classification algorithm uses classification criterion to determine NBC, and
wherein the classification algorithm is updated based on the training data and receipt of the new data,
wherein updating of the classification algorithm is done using an Expectation-Maximization (EM) algorithm that estimates the NBC being accurately applied to the data, wherein the EM algorithm includes an Expectation Step applying a formula:
Q ( X )= p ( X|C,Θ t−1 ),
wherein the EM algorithm includes a Maximization Step applying a formula:
Θ t =argmax Θ ∫ x ∞ Q ( x )log p ( x,c ,Θ) dx;
wherein updating of the NBC includes using a framework based on High Precision rules to classify the data, wherein the High Precision rules generate pseudo-labeled data and high precision clusters from existing datasets comprising at least a simulated testing dataset and based on the pseudo-labeled data, the NBC is trained to create a probabilistic model representation for each data cluster,
wherein X=(x 1 , x 2 , x 3 , . . . x n ) is an observation with n features, and there are m target classes (C 1 , C 2 , C 3 , . . . C m );
where Q(x) represents an expected value of a likelihood function estimated by current parameter set, Θ t−1 , C, and observation X, and where Θ t is the updated parameter set based on expectation.
2. The system of claim 1 , wherein updating of the NBC—includes a term frequency and probability calculation accomplished according to a formula:
p
(
x
i
❘
C
k
)
=
N
x
i
,
c
k
+
α
V
c
k
+
α
·
d
where N x i ,C k is a frequency of word x i in class Ck, V C k is a total number of words in class C k , d is a total number of unique words in the class C k and α is a smoothing factor that users can tune.
3. The system of claim 1 , further identify some of the new data as not able to be classified, and add the new data to an undetermined data set.
4. The system of claim 1 , wherein the updating of the NBC includes using an incremental k-mer based metagenome fragment classifier (iKMF) to classify the data.
5. The system for claim 1 performed by a client.
6. A method for classifying samples, wherein the method comprises:
storing a pre-existing training data cluster;
applying a Naive Bayes Classifier (NBC) to data within the training data cluster;
receiving new data; and
applying the NBC to the new data based on the NBC applied to training data;
wherein the NBC-includes a classification algorithm used to classify the data with the NBC,
wherein the classification algorithm uses classification criterion to determine the NBC, and
wherein the classification algorithm is updated based on the training data and receipt of the new data,
wherein updating of the classification algorithm is done using an Expectation-Maximization (EM) algorithm that estimates the NBC being accurately applied to the data, wherein the EM algorithm includes an Expectation Step applying a formula:
Q ( X )= p ( X|C,Θ t−1 ),
wherein the EM algorithm includes a Maximization Step applying a formula:
Θ t =argmax Θ ∫ x ∞ Q ( x ) log p ( x,c ,Θ) dx;
wherein updating of the NBC includes using a framework based on High Precision rules to classify the data, wherein the High Precision rules generate pseudo-labeled data and high precision clusters from existing datasets comprising at least a simulated testing dataset and based on the pseudo-labeled data, the NBC is trained to create a probabilistic model representation for each data cluster,
wherein:
X=x 1 , x 2 , x 3 , . . . x n ) is an observation with n features, and there are m target classes (C 1 , C 2 , C 3 , . . . C m );
where Q(x) represents an expected value of a likelihood function estimated by current parameter set, Θ t−1 , C, and observation X, and where Θ t is the updated parameter set based on expectation.
7. The method of claim 6 , wherein updating of the NBC-includes a term frequency and probability calculation accomplished according to a formula:
p
(
x
i
❘
C
k
)
=
N
x
i
,
c
k
+
α
V
c
k
+
α
·
d
where N x i ,C k is a frequency of word x i in class C k , V C k is a total number of words in class C k , d is a total number of unique words in the class C k and α is a smoothing factor that users can tune.
8. The method of claim 6 , wherein the method identify some of the new data as not able to be classified, and add the new data to an undetermined data set.
9. The method of claim 6 , wherein the updating of the NBC-includes using an incremental k-mer based metagenome fragment classifier (iKMF) to classify the data.
10. The method for claim 6 , wherein the method is performed by a client.