Guilherme Fonseca, Gabriel Prenassi, Washington Cunha, Leonardo Chaves Dutra da Rocha, Marcos André Gonçalves. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Guilherme Fonseca, Washington Cunha, Gabriel Prenassi, Marcos André Gonçalves, Leonardo Chaves Dutra Da Rocha. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Post-hoc Knowledge Graph (KG) explanation algorithms in Recommender Systems (RSs) identify the most relevant paths connecting interacted and recommended items based on shared attributes. These algorithms are categorized into syntactic and semantic approaches; however, neither can simultaneously optimize explanation quality metrics, which measure the recency of interacted items, diversity of attributes, and popularity of attributes shown across explanations. This study explores the use of Large Language Models (LLMs) to jointly optimize these metrics by presenting possible explanation paths for recommended items. The study evaluates three LLMs against both syntactic and semantic algorithms across two datasets and six RSs, finding that LLMs can improve explanation quality.

We tackle Extreme Multi-Label Text Classification (XMTC), which involves assigning relevant labels to texts from a huge label space. Attempting to optimize the underexplored tail-head trade-off, we address the XMTC task through its core challenges of volume, skewness, and quality by proposing xCoRetriev, a novel two-stage retrieving and fusing ranking pipeline. Our pipeline addresses the volume challenge by dynamically slicing the large label space; it also tackles the skewness challenge by favoring the tail labels while fusing sparse and dense retrievers. Finally, xCoRetriev faces the quality challenge by enhancing the label space with Retrieval-Augmented Generated (RAG)-labels. Our experiments with four XMTC benchmarks with hundreds of thousands of text documents and labels against six state-of-the-art XMTC baselines demonstrate xCoRetriev's strengths in terms of: (i)~scalability for large label spaces, being among the most efficient methods at training and prediction; (ii)~effectiveness in the face of high skewness, with gains of up to 48% in propensity-scored metrics against the best state-of-the-art baselines; and (iii)~capability of handling very noisy datasets by exploiting RAG-labels.

Fine-tuning transformer-based deep-learning models are currently at the forefront of natural language processing (NLP) and information retrieval (IR) tasks. However, fine-tuning these transformers for specific tasks, especially when dealing with ever-expanding volumes of data, constant retraining requirements, and budget constraints, can be computationally and financially costly, requiring substantial energy consumption and contributing to carbon dioxide emissions. This article focuses on advancing the state-of-the-art (SOTA) on instance selection (IS)—a range of document filtering techniques designed to select the most representative documents for the sake of training. The objective is to either maintain or enhance classification effectiveness while reducing the overall training (fine-tuning) total processing time. In our prior research, we introduced the E2SC framework, a redundancy-oriented IS method focused on transformers and large datasets—currently the state-of-the-art in IS. Nonetheless, important research questions remained unanswered in our previous work, mostly due to E2SC’s sole emphasis on redundancy. In this article, we take our research a step further by proposing biO-IS— an extended bi - o bjective i nstance s election solution, a novel IS framework aimed at simultaneously removing redundant and noisy instances from the training. biO-IS estimates redundancy based on scalable, fast, and calibrated weak classifiers and captures noise with the support of a new entropy-based step. We also propose a novel iterative process to estimate near-optimum reduction rates for both steps. Our extended solution is able to reduce the training sets by 41% on average (up to 60%) while maintaining the effectiveness in all tested datasets, with speedup gains of 1.67 on average (up to 2.46x). No other baseline, not even our previous SOTA solution, was capable of achieving results with this level of quality, considering the tradeoff among training reduction, effectiveness, and speedup. To ensure reproducibility, our documentation, code, and datasets can be accessed on GitHub— https://github.com/waashk/bio-is .

Transformer-based deep learning is currently the state-of-the-art in many NLP and IR tasks. However, fine-tuning such Transformers for specific tasks, especially in scenarios of ever-expanding volumes of data with constant re-training requirements and budget constraints, is costly (computationally and financially) and energy-consuming. In this paper, we focus on Instance Selection (IS) - a set of methods focused on selecting the most representative documents for training, aimed at maintaining (or improving) classification effectiveness while reducing total time for training (or fine-tuning). We propose E2SC-IS -- Effective, Efficient, and Scalable Confidence-Based IS -- a two-step framework with a particular focus on Transformers and large datasets. E2SC-IS estimates the probability of each instance being removed from the training set based on scalable, fast, and calibrated weak classifiers. E2SC-IS also exploits iterative heuristics to estimate a near-optimal reduction rate. Our solution can reduce the training sets by 29% on average while maintaining the effectiveness in all datasets, with speedup gains up to 70%, scaling for very large datasets (something that the baselines cannot do).

Abstract This paper presents a Convolutional Neural Network (CNN) architecture named OlapNet, which incorporates implicit operations of OLAP cubes (or data cubes). OLAP cubes are produced from database tables or spreadsheets and they allow particular operations that support performing complex queries efficiently. OlapNet permits evaluating various combinations of these OLAP operations in its search space and thus, it enables, in part, to automate the data transformation step in the knowledge discovery process. A sample of data from an actual database containing anonymized data on the debt history of customers of a financial institution has been used to evaluate our proposal. A predictive classification problem to estimate the probability of any given customer contracting new credits in the next three months has been modelled from these data. Then, traditional methods of Machine Learning and CNN were applied. The results showed that CNN, using the OlapNet architecture, outperforms traditional methods in almost all cases, indicating that the proposed architecture is quite promising.

Nowadays, Recommender Systems have played a crucial role in several entertainment scenarios by making personalised recommendations and guiding the entire users’ journey from their first interaction. Recent works have addressed it as a Contextual Bandit by providing a sequential decision model to explore items not tried yet (or not tried enough) or exploit the best options learned so far. However, this work noticed these current algorithms are limited to naive non-personalised approaches in the first interactions of a new user, offering random or most popular items. Through experiments in three domains, we identify a negative impact of these first choices. Our study indicates that the bandit performance is directly related to the choices made in the first trials. Then, we propose a new approach to balance exploration and exploitation in the first interactions and handle these drawbacks. This approach is based on the Active Learning theory to catch more information about the new users and improve their long-term experience. Our idea is to explore the potential information gain of items that can also please the user’s taste. This method is named Warm-Starting Contextual Bandits, and it statistically outperforms 10 benchmarks in the literature in the long run.

Nowadays, most e-commerce and entertainment services have adopted interactive Recommender Systems (RS) to guide the entire journey of users into the system. This task has been addressed as a Multi-Armed Bandit problem where systems must continuously learn and recommend at each iteration. However, despite the recent advances, there is still a lack of consensus on the best practices to evaluate such bandit solutions. Several variables might affect the evaluation process, but most of the works have only been concerned about the accuracy of each method. Thus, this work proposes an interactive RS framework named iRec. It covers the whole experimentation process by following the main RS guidelines. The iRec provides three modules to prepare the dataset, create new recommendation agents, and simulate the interactive scenario. Moreover, it also contains several state-of-the-art algorithms, a hyperparameter tuning module, distinct evaluation metrics, different ways of visualizing the results, and statistical validation.

This work is a companion reproducibility paper that presents a framework to reproduce our previous experiments and results reported in Werneck et al. (2021). In that previous paper, we introduced a systematic mapping process of points-of-interest (POI) recommendation methods and provided a uniform evaluation methodology based on metrics covering different aspects besides accuracy. Due to the lack of reproducible and extensible benchmarks, our work introduces a reproducibility framework for POI methods based on a collection of Python software libraries and a Docker image. Our proposal is composed of: (1) a package to perform a protocol that reproduces our systematic mapping process Werneck et al. (2021), containing all collected data, insightful views on current advances and opened challenges; and (2) an extensible benchmark to perform a protocol to reproduce experimental evaluations on POI recommendation, considering different datasets, metrics, and the strongest baselines in the literature. This work also demonstrates all processes required to instantiate its framework. Moreover, our work can be considered at least weakly reproducible, since we were able to reproduce the results of the previous paper, leading us to the same conclusions.

Collaborative filtering recommendation algorithms generate suggestions based on similar interactions between users. Although it provides accurate recommendations, the approach has two limitations: the popularity bias, which frequently suggests a small set of the most interacted items, and the systems' black box functioning, as they are grounded on complex mathematical models. To improve such aspects in collaborative filtering algorithms, this paper introduces a multi-domain item reordering system based on the best explanation for an item, which are the best ranked paths extracted from a Linked Open Data knowledge graph connecting recommended and interacted items. To order paths, the algorithm assigns a value to the node attributes connecting two items by calculating the popularity of the property between interacted items that are rare in the full set of items. Results from two datasets of the movie and music domains comparing the proposed reordering system with six baselines of different collaborative filtering families showed that our easy-to-explain approach improved diversity and/or accuracy metrics.

Hierarchical Topic modeling (HTM) exploits latent topics and relationships among them as a powerful tool for data analysis and exploration.Despite advantages over traditional topic modeling, HTM poses its own challenges, such as (1) topic incoherence, (2) unreasonable (hierarchical) structure, and (3) issues related to the definition of the "ideal" number of topics and depth of the hierarchy.In this paper, we advance the stateof-the-art on HTM by means of the design and evaluation of CluHTM, a novel nonprobabilistic hierarchical matrix factorization aimed at solving the specific issues of HTM.CluHTM's novel contributions include: (i) the exploration of richer text representation that encapsulates both, global (dataset level) and local semantic information -when combined, these pieces of information help to solve the topic incoherence problem as well as issues related to the unreasonable structure; (ii) the exploitation of a stability analysis metric for defining the number of topics and the "shape" the hierarchical structure.In our evaluation, considering twelve datasets and seven stateof-the-art baselines, CluHTM outperformed the baselines in the vast majority of the cases, with gains of around 500% over the strongest state-of-the-art baselines.We also provide qualitative and quantitative statistical analyses of why our solution works so well.

Recent advances in text-related tasks on the Web, such as text (topic) classification and sentiment analysis, have been made possible by exploiting mostly the "rule of more": more data (massive amounts) more computing power, more complex solutions. We propose a shift in the paradigm to do "more with less" by focusing, at maximum extent, just on the task at hand (e.g., classify a single test instance). Accordingly, we propose MetaLazy, a new supervised lazy text classification meta-strategy that greatly extends the scope of lazy solutions. Lazy classifiers postpone the creation of a classification model until a given test instance for decision making is given. MetaLazy exploits new ideas and solutions, which have in common their lazy nature, producing altogether a solution for text classification, which is simpler, more efficient, and less data demanding than new alternatives. It extends and evolves the lazy creation of the model for the test instance by allowing: (i) to dynamically choose the best classifier for the task; (ii) the exploration of distances in the neighborhood of the test document when learning a classification model, thus diminishing the importance of irrelevant training instances; and (iii) a better representational space for training and test documents by augmenting them, in a lazy fashion, with new co-occurrence based features considering just those observed in the specific test instance. In a sizeable experimental evaluation, considering topics and sentiment analysis datasets and nine baselines, we show that our MetaLazy instantiations are among the top performers in most situations, even when compared to state-of-the-art deep learning classifiers such as Deep Network Transformer Architectures.

Advances in Recommender Systems (RSs) have been focused on improving the system's accuracy. However, accuracy alone is not enough to assess the practical effects. In real scenarios, diversity has been identified as a key dimension of recommendation utility. Thus, the main researches are focused in improve both, accuracy and diversity. This challenge remains an apparent dilemma that remains open and can boost sales by offering consumers both their mainstream and specific tastes. For this reason, we propose an approach to handle the accuracy-diversity dilemma. Our approach, based on a Particle Swarm Optimization (PSO), is a post-processing method to re-rank items from traditional RSs in order to improve diversity without accuracy losses. Experimental results in entertainment and e-commerce scenarios show that our strategy can improve users satisfaction. We improve the diversity up to 70% without significant accuracy losses.

In this paper, we advance the state-of-the-art in topic modeling by means of a new document representation based on pre-trained word embeddings for non-probabilistic matrix factorization. Specifically, our strategy, called CluWords, exploits the nearest words of a given pre-trained word embedding to generate meta-words capable of enhancing the document representation, in terms of both, syntactic and semantic information. The novel contributions of our solution include: (i)the introduction of a novel data representation for topic modeling based on syntactic and semantic relationships derived from distances calculated within a pre-trained word embedding space and (ii)the proposal of a new TF-IDF-based strategy, particularly developed to weight the CluWords. In our extensive experimentation evaluation, covering 12 datasets and 8 state-of-the-art baselines, we exceed (with a few ties) in almost cases, with gains of more than 50% against the best baselines (achieving up to 80% against some runner-ups). Finally, we show that our method is able to improve document representation for the task of automatic text classification.

In this paper, we advance the state-of-the-art in topic modeling by means of the design and development of a novel (semi-formal) general topic modeling framework. The novel contributions of our solution include: (i) the introduction of new semantically-enhanced data representations for topic modeling based on pooling, and (ii) the proposal of a novel topic extraction strategy - ASToC - that solves the difficulty in representing topics in our semantically-enhanced information space. In our extensive experimentation evaluation, covering 12 datasets and 12 state-of-the-art baselines, totalizing 108 tests, we exceed (with a few ties) in almost 100 cases, with gains of more than 50% against the best baselines (achieving up to 80% against some runner-ups). We provide qualitative and quantitative statistical analyses of why our solutions work so well. Finally, we show that our method is able to improve document representation in automatic text classification.

In this paper, we propose a general framework that allows developers to filter, summarize and analyze user reviews written about applications on App Stores. Our framework extracts automatically relevant features from reviews of apps (e.g., information about functionalities, bugs, requirements, etc) and analyzes the sentiment associated with each of them. Our framework has three main building blocks, namely, (i) topic modeling, (ii) sentiment analysis and (iii) summarization interface. The topic modeling block aims at finding semantic topics from textual comments, extracting the target features based on the most relevant words of each discovered topic. The sentiment analysis block detects the sentiment associated with each discovered feature. The summarization interface provides to developers an intuitive visualization of the features (i.e., topics) and their associated sentiment, providing richer information than a 'star rating' strategy. Our evaluation shows that the topic modeling block is able to organize information provided by users in subcategories that facilitate the understanding of which features more positively/negatively impact the overall evaluation of the application. Regarding user satisfaction, we can observe that, in spite of the star rating being a good measure of evaluation, the Sentiment Analysis technique is more accurate in capturing the sentiment transmitted by the user by means of a comment.

The development of new technologies is setting a new era characterized, among other factors, by the rise of sophisticated mobile devices containing CPUs and GPUs. This emerging scenario of heterogeneous mobile architectures brings challenging issues regarding the use of the available computing resources. Such issues are mainly related to the intrinsic complexity of coordinating these processors in order to increase application performance. In this sense, this paper presents a high-level programming model to implement parallel patterns that can be executed in a coordinate way by heterogeneous mobile architectures. A comparative analysis of performance and programming complexity is presented, contrasting code generated automatically from the proposed programming model with low-level manually-optimized implementations.

Social media applications have assumed an important role in decision-making process of users, affecting their choices about products and services. In this context, understanding and modeling opinions, as well as opinion-leaders, have implications for several tasks, such as recommendation, advertising, brand evaluation etc. Despite the intrinsic relation between opinions and opinion-leaders, most recent works focus exclusively on either understanding the opinions, by Sentiment Analysis (SA) proposals, or identifying opinion-leaders using Influential Users Detection (IUD). This paper presents a preliminary evaluation about a combined analysis of SA and IUD. In this sense, we propose a methodology to quantify factors in real domains that may affect such analysis, as well as the potential benefits of combining SA Methods with IUD ones. Empirical assessments on a sample of tweets about the Brazilian president reveal that the collective opinion and the set of top opinion-leaders over time are inter-related. Further, we were able to identify distinct characteristics of opinion propagation, and that the collective opinion may be accurately estimated by using a few top-k opinion-leaders. These results point out the combined analysis of SA and IUD as a promising research direction to be further exploited.

Automatic text classification (TC) continues to be a relevant research topic and several TC algorithms have been proposed. However, the majority of TC algorithms assume that the underlying data distribution does not change over time. In this work, we are concerned with the challenges imposed by the temporal dynamics observed in textual data sets. We provide evidence of the existence of temporal effects in three textual data sets, reflected by variations observed over time in the class distribution, in the pairwise class similarities, and in the relationships between terms and classes. We then quantify, using a series of full factorial design experiments, the impact of these effects on four well‐known TC algorithms. We show that these temporal effects affect each analyzed data set differently and that they restrict the performance of each considered TC algorithm to different extents. The reported quantitative analyses, which are the original contributions of this article, provide valuable new insights to better understand the behavior of TC algorithms when faced with nonstatic (temporal) data distributions and highlight important requirements for the proposal of more accurate classification models.

Automatic Document Classification (ADC) is the basis of many important applications such as spam filtering and content organization. Naive Bayes (NB) approaches are a widely used classification paradigm, due to their simplicity, efficiency, absence of parameters and effectiveness. However, they do not present competitive effectiveness when compared to other modern statistical learning methods, such as SVMs. This is related to some characteristics of real document collections, such as class imbalance, feature sparseness and strong relationships among attributes. In this paper, we investigate whether the relaxation of the NB feature independence assumption (aka, Semi-NB approaches) can improve its effectiveness in large text collections. We propose four new Lazy Semi-NB strategies that exploit different ideas for alleviating the NB independence assumption. By being lazy, our solutions focus only on the most important features to classify a given test document, overcoming some Semi-NB issues when applied to ADC such as bias towards larger classes and overfitting and/or lack of generalization of the models. We demonstrate that our Lazy Semi-NB proposals can produce superior effectiveness when compared to state-of-the-art ADC classifiers such as SVM and KNN. Moreover, to overcome some efficiency issues of combining Semi-NB and lazy strategies, we take advantage of current manycore GPU architectures and present a massively parallelized version of the Semi-NB approaches. Our experimental results show that speedups of up to 63.36 times can be obtained when compared to serial solutions, making our proposals very practical in real-situations.

Random Forests (RF) and Boosting are two of the most successful supervised learning paradigms for automatic classification. In this work we propose to combine both strategies in order to exploit their strengths while simultaneously solving some of their drawbacks, especially when applied to high-dimensional and noisy classification tasks. More specifically, we propose a boosted version of the RF classifier (BROOF), which fits an additive model composed by several random forests (as weak learners). Differently from traditional boosting methods which exploit the training error estimate, we here use the stronger out-of-bag (OOB) error estimate which is an out-of-the-box estimate naturally produced by the bagging method used in RFs. The influence of each weak learner in the fitted additive model is inversely proportional to their OOB error. Moreover, the probability of selecting an out-of-bag training example is increased if misclassified by the simpler weak learners, in order to enable the boosted model to focus on complex regions of the input space. We also adopt a selective weight updating procedure, whereas only the out-of-bag examples are updated as the boosting iterations go by. This serves the purpose of slowing down the tendency to focus on just a few hard-to-classify examples. By mitigating this undesired bias known to affect boosting algorithms under high dimensional and noisy scenarios - due to both the selective weighting schema and a proper weak-learner effectiveness assessment - we greatly improve classification effectiveness. Our experiments with several datasets in three representative high-dimensional and noisy domains - topic, sentiment and microarray data classification - an up to ten state-of-the-art classifiers (covering almost 500 results), show that BROOF is the only classifier to be among the top performers in all tested datasets from the topic classification domain, and in the vast majority of cases in sentiment and microarray domains, a surprising result given the knowledge that there is no single top-notch classifier for all datasets.

This paper addresses the problem of automatically learning to classify texts by exploiting information derived from meta-level features (i.e., features derived from the original bag-of-words representation). We propose new meta-level features derived from the class distribution, the entropy and the within-class cohesion observed in the k nearest neighbors of a given test document x, as well as from the distribution of distances of x to these neighbors. The set of proposed features is capable of transforming the original feature space into a new one, potentially smaller and more informed. Experiments performed with several standard datasets demonstrate that the effectiveness of the proposed meta-level features is not only much superior than the traditional bag-of-word representation but also superior to other state-of-art meta-level features previously proposed in the literature. Moreover, the proposed meta-features can be computed about three times faster than the existing meta-level ones, making our proposal much more scalable. We also demonstrate that the combination of our meta features and the original set of features produce significant improvements when compared to each feature set used in isolation.

High performance computing is experiencing a major paradigm shift with the introduction of accelerators, such as graphics processing units (GPUs) and Intel Xeon Phi (MIC). These processors have made available a tremendous computing power at low cost, and are transforming machines into hybrid systems equipped with CPUs and accelerators. Although these systems can deliver a very high peak performance, making full use of its resources in real-world applications is a complex problem. Most current applications deployed to these machines are still being executed in a single processor, leaving other devices underutilized. In this paper we explore a scenario in which applications are composed of hierarchical data flow tasks which are allocated to nodes of a distributed memory machine in coarse-grain, but each of them may be composed of several finer-grain tasks which can be allocated to different devices within the node. We propose and implement novel performance aware scheduling techniques that can be used to allocate tasks to devices. We evaluate our techniques using a pathology image analysis application used to investigate brain cancer morphology, and our experimental evaluation shows that the proposed scheduling strategies significantly outperforms other efficient scheduling techniques, such as Heterogeneous Earliest Finish Time - HEFT, in cooperative executions using CPUs, GPUs, and MICs. We also experimentally show that our strategies are less sensitive to inaccuracy in the scheduling input data and that the performance gains are maintained as the application scales.

This paper explores a method for incorporating into a recommender system explicit representations of user's preferences over non-content attributes such as popularity, recency, and similarity of recommended items. We show how such attributes can be modeled as a preference vector that can be used in a vector-space content-based recommender, and how that content-based recommender can be integrated with various collaborative filtering techniques through re-weighting of Top-M recommendations. We evaluate this approach on several recommender systems datasets and collaborative filtering methods, and find that incorporating the three preference attributes can lead to a substantial increase in Top-50 precision while also enhancing diversity and novelty.

The advent of the Web 2.0 has given rise to an interesting phenomenon: there is currently much more data than what can be effectively analyzed without relying on sophisticated automatic tools. Some of these tools, which target the organization and extraction of useful knowledge from this huge amount of data, rely on machine learning and data or text mining techniques, specifically automatic document classification algorithms. However, these algorithms are still a computational challenge because of the volume of data that needs to be processed. Some of the strategies available to address this challenge are based on the parallelization of ADC algorithms. In this work, we present GPU-NB, a parallel version of one of the most widely used document classification algorithms, the Naïve Bayes, that uses graphics processing units (GPUs). In our evaluation using 6 different document collections, we show that the GPU-NB can maintain the same classification effectiveness (in most cases) while increasing the efficiency by up to 34x faster than its sequential version using CPU. GPU-NB is also up to 11x faster than a CPU-based parallel implementation of Naive Bayes running with 4 threads. Moreover, assuming an optimistic behavior of the CPU parallelization, GPU-NB should outperform the CPU-based implementation with up to 32 cores, at a small fraction of the cost. We also show that the efficiency of the GPU-NB parallelization is impacted by features of the document collections, particularly the number of classes, although the density of the collection (average number of occurrences of terms per document) has a significant impact as well.

One of the major challenges in automatic classification is to deal with highly dimensional data. Several dimensionality reduction strategies, including popular feature selection metrics such as Information Gain and χ2, have already been proposed to deal with this situation. However, these strategies are not well suited when the data is very skewed, a common situation in real-world data sets. This occurs when the number of samples in one class is much larger than the others, causing common feature selection metrics to be biased towards the features observed in the largest class. In this paper, we propose the use of Genetic Programming (GP) to implement an aggressive, yet very effective, selection of attributes. Our GP-based strategy is able to largely reduce dimensionality, while dealing effectively with skewed data. To this end, we exploit some of the most common feature selection metrics and, with GP, combine their results into new sets of features, obtaining a better unbiased estimate for the discriminative power of each feature. Our proposal was evaluated against each individual feature selection metric used in our GP-based solution (namely, Information Gain, χ2, Odds-Ratio, Correlation Coefficient) using a k8 cancer-rescue mutants data set, a very unbalanced collection referring to examples of p53 protein. For this data set, our solution not only increases the efficiency of the learning algorithms, with an aggressive reduction of the input space, but also significantly increases its accuracy.

Automatic Document Classification (ADC) is still one of the major information retrieval problems. It usually employs a supervised learning strategy, where we first build a classification model using pre-classified documents and then use this model to classify unseen documents. The majority of supervised algorithms consider that all documents provide equally important information. However, in practice, a document may be considered more or less important to build the classification model according to several factors, such as its timeliness, the venue where it was published in, its authors, among others. In this paper, we are particularly concerned with the impact that temporal effects may have on ADC and how to minimize such impact. In order to deal with these effects, we introduce a temporal weighting function (TWF) and propose a methodology to determine it for document collections. We applied the proposed methodology to ACM-DL and Medline and found that the TWF of both follows a lognormal. We then extend three ADC algorithms (namely kNN, Rocchio and Naïve Bayes) to incorporate the TWF. Experiments showed that the temporally-aware classifiers achieved significant gains, outperforming (or at least matching) state-of-the-art algorithms.

Due to the increasing amount of information being stored and accessible through the Web, Automatic Document Classification (ADC) has become an important research topic. ADC usually employs a supervised learning strategy, where we first build a classification model using pre-classified documents and then use it to classify unseen documents. One major challenge in building classifiers is dealing with the temporal evolution of the characteristics of the documents and the classes to which they belong. However, most of the current techniques for ADC do not consider this evolution while building and using the models. Previous results show that the performance of classifiers may be affected by three different temporal effects (class distribution, term distribution and class similarity). Further, it is shown that using just portions of the pre-classified documents, which we call contexts, for building the classifiers, result in better performance, as a consequence of the minimization of the aforementioned effects.

Due to the increasing amount of information present on the Web, Automatic Document Classification (ADC) has become an important research topic. ADC usually follows a standard supervised learning strategy, where we first build a model using preclassified documents and then use it to classify new unseen documents. One major challenge for ADC in many scenarios is that the characteristics of the documents and the classes to which they belong may change over time. However, most of the current techniques for ADC are applied without taking into account the temporal evolution of the collection of documents