Sutanu Chakraborti

Feature Weighting and Confidence Based Prediction for Case Based Reasoning Systems

The quality of the cases maintained in a case base has a direct influence on the quality of the proposed solutions. The presence of cases that do not conform to the similarity hypothesis decreases the alignment of the case base and often degrades the performance of a CBR system. It is therefore important to find out the suitability of each case for the application of CBR and associate a solution with a certain degree of confidence. Feature weighting is another important aspect that determines the success of a system, as the presence of irrelevant and redundant attributes also results in incorrect solutions. We explore these problems in conjunction with a real-world CBR application called InfoChrom. It is used to predict the values of several soil nutrients based on features extracted from a chromatogram image of a soil sample. We propose novel feature weighting techniques based on alignment, as well as a new alignment and confidence measure as potential solutions. The hypotheses are evaluated on UCI datasets and the case base of Infochrom and show promising results.

Topic labeled text classification: a weakly supervised approach

Supervised text classifiers require extensive human expertise and labeling efforts. In this paper, we propose a weakly supervised text classification algorithm based on the labeling of Latent Dirichlet Allocation (LDA) topics. Our algorithm is based on the generative property of LDA. In our algorithm, we ask an annotator to assign one or more class labels to each topic, based on its most probable words. We classify a document based on its posterior topic proportions and the class labels of the topics. We also enhance our approach by incorporating domain knowledge in the form of labeled words. We evaluate our approach on four real world text classification datasets. The results show that our approach is more accurate in comparison to semi-supervised techniques from previous work. A central contribution of this work is an approach that delivers effectiveness comparable to the state-of-the-art supervised techniques in hard-to-classify domains, with very low overheads in terms of manual knowledge engineering.

UtilSim: Iteratively Helping Users Discover Their Preferences

Conversational Recommender Systems belong to a class of knowledge based systems which simulate a customer’s interaction with a shopkeeper with the help of repeated user feedback till the user settles on a product. One of the modes for getting user feedback is Preference Based Feedback, which is especially suited for novice users(having little domain knowledge), who find it easy to express preferences across products as a whole, rather than specific product features. Such kind of novice users might not be aware of the specific characteristics of the items that they may be interested in, hence, the shopkeeper/system should show them a set of products during each interaction, which can constructively stimulate their preferences, leading them to a desirable product in subsequent interactions. We propose a novel approach to conversational recommendation, UtilSim, where utilities corresponding to products get continually updated as a user iteratively interacts with the system, helping her discover her hidden preferences in the process. We show that UtilSim, which combines domain-specific “dominance” knowledge with SimRank based similarity, significantly outperforms the existing conversational approaches using Preference Based Feedback in terms of recommendation efficiency.

Towards Creating Pedagogic Views from Encyclopedic Resources

This paper identifies computational challenges in restructuring encyclopedic resources (like Wikipedia or thesauri) to reorder concepts with the goal of helping learners navigate through a concept network without getting trapped in circular dependencies between concepts. We present approaches that can help content authors identify regions in the concept network, that after editing, would have maximal impact in terms of enhancing the utility of the resource to learners.

Making the Most of Preference Feedback by Modeling Feature Dependencies

Conversational recommender systems help users navigate through the product space by exploiting feedback. In conversational systems based on preference-based feedback, the user selects the most preferred item from a list of recommended products. Modelling users preferences then becomes important in order to recommend relevant items. Several existing recommender systems accomplish this by assuming the features to be independent. Here we will attempt to forego this assumption and exploit the dependencies between the features to build a robust user preference model.

Sprinkling Topics for Weakly Supervised Text Classification

Supervised text classification algorithms require a large number of documents labeled by humans, that involve a laborintensive and time consuming process. In this paper, we propose a weakly supervised algorithm in which supervision comes in the form of labeling of Latent Dirichlet Allocation (LDA) topics. We then use this weak supervision to “sprinkle” artificial words to the training documents to identify topics in accordance with the underlying class structure of the corpus based on the higher order word associations. We evaluate this approach to improve performance of text classification on three real world datasets.

Selective integration of background knowledge in TCBR systems

This paper explores how background knowledge from freely available web resources can be utilised for Textual Case Based Reasoning. The work reported here extends the existing Explicit Semantic Analysis approach to representation, where textual content is represented using concepts with correspondence to Wikipedia articles. We present approaches to identify Wikipedia pages that are likely to contribute to the effectiveness of text classification tasks. We also study the effect of modelling semantic similarity between concepts (amounting to Wikipedia articles) empirically. We conclude with the observation that integrating background knowledge from resources like Wikipedia into TCBR tasks holds a lot of promise as it can improve system effectiveness even without elaborate manual knowledge engineering. Significant performance gains are obtained using a very small number of features that have very strong correspondence to how humans describe the domain.

Robust Measures of Complexity in TCBR

In TCBR, complexity refers to the extent to which similar problems have similar solutions. Casebase complexity measures proposed are based on the premise that a casebase is simple if similar problems have similar solutions. We observe, however, that such measures are vulnerable to choice of solution side representations, and hence may not be meaningful unless similarities between solution components of cases are shown to corroborate with human judgements. In this paper, we redefine the goal of complexity measurements and explore issues in estimating solution side similarities. A second limitation of earlier approaches is that they critically rely on the choice of one or more parameters. We present two parameter-free complexity measures, and propose a visualization scheme for casebase maintenance. Evaluation over diverse textual casebases show their superiority over earlier measures.

Competence Guided Casebase Maintenance for Compositional Adaptation Applications

A competence guided casebase maintenance algorithm retains a case in the casebase if it is useful to solve many problems and ensures that the casebase is highly competent in the global sense. In this paper, we address the compositional adaptation process (of which single case adaptation is a special case) during casebase maintenance by proposing a case competence model for which we propose a measure called retention score to estimate the retention quality of a case. We also propose a revised algorithm based on the retention score to estimate the competent subset of the casebase. We used regression datasets to test the effectiveness of the competent subset obtained from the proposed model. We also applied this model in a tutoring application and analyzed the competent subset of concepts in tutoring resources. Empirical results show that the proposed model is effective and overcomes the limitation of footprint based competence model in compositional adaptation applications.

Mining user trails in critiquing based recommenders

Critiquing based recommenders are very commonly used to help users navigate through the product space to find the required product by tweaking/critiquing one or more features. By critiquing a product, the user gives an informative feedback(i.e, which feature needs to be modified) about why they rejected a product and preferred the other one. As a user interacts with such a system, trails are left behind. We propose ways of leveraging these trails to induce preference models of items which can be used to estimate the relative utilities of products which can be used in ranking the recommendations presented to the user. The idea is to effectively complement knowledge of explicit user interactions in traditional social recommenders with knowledge implicitly obtained from trails.