Personalized classifiers

Editor's Notes

• #2: Hi, my name is Ramakrishna. First of all, I apologize for not being present at the conference at this time. Due to some unfortunate medical conditions, I had to cancel my travel at the last minute. So, I will try to present our work on Evolving Personalized Classifier through this audio/video recording. Please send all your questions to my email and I will ensure that all your queries are answered.
• #3: Let me start with the motivation for our work. The need for automatic evolution of a categorization system lies in the enormous amount of digital documents that we accumulate in our organizations every day. Unless we categorize them effectively, it becomes hard to find the required piece of information. Apart from this, the categorization information is also helpful for various other tasks, such as Faceted Search, Exploratory Search, Navigational Search, Diversifying Search Results, Ranking search results, to name a few. Some other interesting facts show that, many companies spend enormous amount of human effort and money to build effective document categorization system. So, it is an import area of reasearch.
• #4: What are the challenges in building an automated categorization system? The first question is : what should the representative categories be? Too generic categories like News, Entertainment, Technical, Politics, Sports, etc. might not help. Each of such categories accumulates thousands of articles and searching for the required piece of information continues to be challenging. On the other hand, fine grained category creation needs domain experts and is a laborious task. What’s the next option? We look for some predefined category systems. For example, categories from Reuters RCV1-v2 Text Categorization Test Collection, Yahoo Directories or DMOZ [5] can be adopted and hope those categories to fit the existing documents in library. How good is such an adoption? We conducted a small experiment on Reuters-21578 data-set. We selected some of the Reuters documents and detected topics for them using Wikipedia-Miner, which detects topics for the input text using Wikipedia titles. The results are shown in this table. Interestingly, in some cases categories are very distinct and in some cases, Wikipedia-Miner categories are very Fine grained. This indicates that adoption of predefined categories may not work well always. Apart from this, an organization might have its personal choices in building the categories. For example, a computer science department, may not like to create a detailed categorization of bacterias or genes even if the documents describe about the application of some clustering or classification methods for bacteria or genes. So we want to have a personalized categorization system. /* We also would like to have the categorization system to detect emergence of new categories as the new documents are seen. Currently, we do not address this in our work directly. As will be evident later, some part of it gets address due to the choice of very large knowledge graph (Wikipedia in our experiments) that keeps growing over time. */ In next few slides, we briefly explain our technique of achieving these goals.
• #5: We assume that, there exists a catalog with very large number of categories. This is not a bad assumption. For e.g. 5M Wikipedia titles can almost describe the vocabulary of any document collection. There are many such collaboratively built ontologies, which can serve as a category catalog. Often, these sources have much more information such as descriptions, association relations between various titles, etc. Hence, we treat them as Knowledge Graph and make use of all those additional information in deriving Personalized categories. The nodes in this knowledge graph are the categories and relationship between them are the edges. Further more, we assume that the organization receives documents in batches. We think this is fairly descent assumption, for e.g. the library of a university receives hundreds of student thesis at the end of a semester/year, periodically.
• #6: This picture shows a part of knowledge graph, with various category nodes and relationship edges. In our case, we only consider association/related relation between the categories. For e.g. “Loss Function” is related to “Support Vector Machine”, since SVM uses a loss function. The number on the edge shows the strength of relationship.
• #7: OK… Here is the precise statement of our problem and contribution. We try to learn a model which identifies a subset of categories from the Knowledge graph that are relevant to the documents, considering the organization preferences, that is personalized to the organization. We do this with the help of new Active Learning technique that we proposed. We will explain this in detail next.
• #8: This figure pictorially represents our goal of evolving a personalized classifier, by making use of knowledge graph and user interests.
• #9: Lets have a quick look at the overall architecture and then we will get into the specifics of our model, personalization techniques and active learning. The main component, “Personalized Classifier” attaches categories to the documents using the Knowledge Graph and the learnt model parameters. The “Active Learner” component choses specific documents and the specific categories in those documents and asks the user what she thinks. Is it right or wrong or of no interest to the organization. This feedback helps the system learn organization specific categorization. The choice of documents and categories for user feedback is done wisely to maximize the learning rate and minimize the user cognitive load, as we see later.
• #10: The very first step in identifying the categories for a document is to spot important keywords. We use some of the known techniques in literature for key phrase extraction. You can refer to our paper for more details on this.
• #11: Once we identify the keywords, we can retrieve the subset of categories from Knowledge Graph that match these keywords. We used the category names and the gloss (ie. First few sentences of category description in Wikipedia) to get these categories. We then expand the spotted category list by adding the categories in the Markov Blanket of these categories in the Knowledge Graph. We call the resulting list as Candidate Categories. The reason for this expansion is that, the categories that get assigned to a document usually exhibit semantic relations such as “associations”. For example, consider a document about Linear Classifiers. Category “Linear Classifier” is related to categories such as “Kernel Methods in Classifiers,” “Machine Learning”. Our observation shows that all these categories usually appear together in a document talking about say Linear Classifiers. You may refer to our paper for more examples and details. We next select a subset of edges from KnG such that the nodes connected by them are in Candidate Categories List. This gives us a small sub graph of knowledge graph, which we call candidate category subgraph and use for further processing.
• #12: Note that, Not all the categories in the candidate list of categories are relevant to the document. Because, the document may not be talking about a topic that is spotted or the spotted category is of no interest to the organization. We need to select only most appropriate categories from these candidate categories and assign them to the document. We use Associative Markov Network for this.
• #13: The reason for choosing AMN is due to our observation of two classes of informative features are present in the candidate category subgraph: Firstly, a feature that is a function of the document and a category, such as the category-specific classifier scoring function evaluated on a document And secondly, a feature that is a function of two categories, such as their co-occurrence frequency or textual overlap between their descriptions, etc. This naturally, makes Associative Markov Network (AMN) very suitable model for inferring the categories from the candidate category subgraph. The next slide illustrates this with an example
• #14: All the categories shown here are the candidate categories for a journal article taken from arXiv.org talking about classification of cancers. But only few of them such as Binary Classifier, Support Vector Machine are relevant due to the relevance to the content of the document and user preferences.
• #15: Next step is to choose the relevant categories using AMN inference. The AMN inference is based on the node and edge features. Node features help in capturing the similarity of a node, that is category to the input document through a combination of different kernels such as Bag of Words kernels, N-gram kernels, Relational kernels and most importantly, node specific Classifiers SVM or Naïve Bayes classifiers. Edge features Capture the similarity between the nodes using Cosine or Jaccard similarity between their titles, descriptions, etc.
• #16: The goal of AMN inference is to assign the labels 0 or 1 to the nodes of candidate category graph that maximizes the joint probability of assignment. If a node is labeled 1, that category is attached to the document. If 0, it is discarded. Hence, the set of categories assigned to the documents are those nodes that are inferred as 1. In this optimization problem, variables X_i stand for the node feature vector for the i_th node. Similarly, X_ij is the edge feature vector for the edge connecting nodes i,j. The node feature weights wn combine the node feature values. Similarly, edge feature weights w_e combine the edge feature values. Note that, superscript k on the weights indicates that we have separate feature weights for the features corresponding to label 0 and label 1. The constraint summation yik=1 ensures that only label 0 or 1 is assigned to a node, not both. The next constraint, this one, forces two adjacent nodes with high edge potential (that is similar nodes) to assume similar labels. The constraint yi = 1 for the hard constrained nodes force the node label 0. These hard constraints are part of our personalization task, which we explain later. Note that, this is the MAP inference in the Markov Network where only cliques of size 1 or 2 are considered. Again, we request our audience to refer to our paper for further details.
• #17: One of the important aspect of this model is to learn node feature weights wn, edge feature weights we and node specific SVM classifier weights. Our system learns these weights as it evolves based on the feedback and training by the user as part of personalization describe in the next slide.
• #18: Now we need to train our model that can categorize the documents as per the organization’s interests. This involves training node specific classifiers such as SVM and learning the AMN feature weights based on the training examples. The training examples are provided by the user, in the form of feedback. That is, system first chooses few documents and categories that it thinks are most “confusing” to the system and asks the user if that category assignment of the document “correct”, “incorrect” or that category is not of interest to organization. Based on these 3 kinds of feedback, the system trains its SVM classifiers and AMN weights.
• #19: If the users indicates the category is of not interest to the organization, system applies a constraint that forces the label 0 for the category during the AMN inference process that we saw earlier.
• #20: Note that, the user need not apply this kind of constraints to all the categories. Due the AMN’s associative property, the constraint automatically propagates to a node similar to this node. For e.g. if the organization tells the system that the category “Cancer” is of no interest to them, the system automatically suppresses the related categories such as malignancy, Hormone Therapy, etc, which are related categories during AMN inference. This happens naturally because, the joint probability will be maximized by assigning similar labels to the highly related categories . This is precisely the reason why we chose AMN to model our problem.
• #21: The feedback “Correct” or “Incorrect” given by the user is used by the system to generate +ve and –ve examples for the node specific classifiers such as SVM and learn it parameters. It also retrains the AMN parameters whenever SVM parameters are updated so that the feature weights are recalibrated.
• #22: As said earlier, the system chooses documents and categories to seek user feedback. To choose specific document and the categories as part of Active Learning, we first transform the AMN model to a hyperplane based two class classification problem. The proof and the details of the transformation are given in our paper. Using the uncertainty based approach, we choose the categories that are close to the hyperplane as most confusing categories for the system for that document.
• #23: We now have most uncertain categories for a given document. But we have multiple documents in a batch and we cannot seek feedback for all the documents. Because that will put lot of cognitive load on the users. We need to choose a set of documents that maximize the learning on uncertain categories. Hence we propose a joint active learning model. Here we view the document and categories as bipartite graph with documents on one side and categories on the other side. The optimization problem chooses a set of documents with zj=1 and a set of categories with yi=1 that are most uncertain and for which seeking feedback from the user is most beneficial. The weights a_i tells us how much uncertain the category i is. The weight b_j tells us what is the contribution of document j in terms of providing the uncertain categories for feedback. Our paper gives couple ways in which this contribution can be calculated. The value P here is the number of documents for which the user is willing to give feedback. Say, 3 or 5 or 10, whatever… it depends on when the user gets bored. Based on the outcome of this optimization, we present the documents and its most uncertain categories to the user for the feedback.
• #24: Coming to evaluation, we report our results in two different settings: (i) Warm start (ii) Cold start. In the Warm start setting, we use Reuters RCV1-v2 text collection and use only the categories present in RCV1-v2 and created a Knowledge Graph of those categories. Such a setting helps us demonstrate how, on a standard classification dataset, the Markov network helps propagate learnings from a category to other related categories. We picked 5000 training documents and 2000 test documents using clustered sampling procedure. We further divided the training set into 100 batches of 50 documents each. We iterated through the batches and in the kth iteration, we trained our model (SVMs, AMN feature weights) using training documents from all batches up to the kth batch. For each iteration, we performed AMN inference on the sample of 2000 test documents. We also compared our proposed joint active learning technique with other techniques from literature, namely, HIClass and Tong’s procedure adapted to multi-class classification.
• #25: Here are the results of Warm start tests. The X axis shows the growing document size due to incoming batches. The y-axis is the F1 score of categorizing 2000 held out Reuter’s documents. We observe that our system learns faster and provides better F1 score compared SVM. Similarly, joint active learning is also observed to learn faster than other methods.
• #26: In cold start setting, using the Amazon S3 service, we downloaded 263 technical documents from arXiv under different streams of Computer Science. With the help of eight human annotators, we assigned categories to each document using Wikipedia article names. We carried out five-fold cross-validation with each fold containing 210 training documents and 53 test documents. In each fold, we trained our model (SVMs, AMN feature weights) using the training set and evaluated Consistency, Precision and Recall on the test set. Note that, during the training phase, we also applied localization techniques in which we recorded feedback for the system suggested categories in three forms: “Correct,”, “Incorrect” and “Never again.” as recorded by the human annotators. We find that, the Consistency of our system with the human annotators better than that of Wikipedia Miner by about 12%.
• #27: There are few issues that our system fails to handle currently. Since it generates categories for each document separately, it may end up in generating too many categories AMN inference over thousands of candidate categories is time consuming. Hence we cannot use this system in a real time fashion When KnG gets updated, Documents that are already assigned with categories need to be updated wisely. We plan to take these as future improvements.
• #28: Thanks everyone for listening to my presentation. Please send all your questions to my email address shown here and I will answer them promptly. Once again, thank you for listening to me and have a good day.