WO2001059610A2 - Generation and effectiveness evaluation of a multi-feature classification system using genetic algorithms - Google Patents
Generation and effectiveness evaluation of a multi-feature classification system using genetic algorithms Download PDFInfo
- Publication number
- WO2001059610A2 WO2001059610A2 PCT/EP2001/000311 EP0100311W WO0159610A2 WO 2001059610 A2 WO2001059610 A2 WO 2001059610A2 EP 0100311 W EP0100311 W EP 0100311W WO 0159610 A2 WO0159610 A2 WO 0159610A2
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- features
- feature
- classification
- content material
- combination
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/211—Selection of the most significant subset of features
- G06F18/2111—Selection of the most significant subset of features by using evolutionary computational techniques, e.g. genetic algorithms
Definitions
- This invention relates to the field of classification systems, and in particular to the selection of the features and combinations of features that are used to determine a given sample's classification.
- a number of methods are available for characterizing the content of a particular piece of material.
- television guides containing a synopsis of each program are available, and automated systems have been proposed for categorizing programs, and segments of programs, based on an analysis of the images contained in each image frame.
- web crawlers are used to extract key words and phrases from each web page to facilitate the search for material based on such key words or phrases, or synopses of select web pages are manually created to form an index to facilitate these searches.
- speech recognition techniques may be employed to create an index of key words used in a television or radio program, or in the lyrics of a song, and so on. Other characterization methods are also employed based on other factors as well.
- the time of day, day of the week, and season of the year may be included in the characterization of broadcast entertainment material, distinguishing, for example, between "prime time” programs and "before dawn” programs, as a potential indicator of program quality or popularity.
- the producer, director, actors, broadcast network, type of provider, and so on may also be used to characterize a program.
- similar parameters may also be used, such as the number of "hits" a particular web page experiences per day, the number of other web pages that reference this web page, the author of the web page, and so on.
- the term "content material” is used hereinafter to refer to material that is related to the contents of information items, entertainment items, and other items that are potentially available for classification or characterization.
- the content material may include the contents of the information or entertainment item itself, an abstract or synopsis of the item, information related to the creation or presentation of the item, and so on.
- feature is used hereinafter to refer to a characteristic that is potentially available to facilitate the classification or characterization.
- each word in a synopsis of a television program is a feature that can be used to facilitate the characterization of the content material of that television program; the director's name is also a feature, as is the time of day that the program is broadcast.
- each key word of a web page is a feature, as is the provider of the web page, the family of pages to which this page belongs, and so on.
- the effectiveness and efficiency of a classification system is highly dependent upon the choice of features used to classify the content material. This effectiveness and efficiency is particularly dependent upon the choice of features that comprise a combination of features.
- the choice of features that comprise a combination of features is often a subjective choice, and is often a manually intensive process. For example, it is straightforward to use the words of a synopsis as the set of features that will be used to classify a television program. Each synopsis is processed to identify each word and to remove noise words. The resultant list of words used in the synopsis, potentially ordered by their frequency of occurrence, are stored in a database for subsequent processing to determine the subject matter classification for that content material, or to determine whether these words are correlated with words that are related to a user's preference, and so on.
- chromosome woul 1 contains combinations of features, in the above example, the chromosome woul 1 contain a subset of all the words used in the synopses of many programs. Different chromosomes would contain different subsets.
- each chromosome that contains these words in its subset of words will generally exhibit a better classification performance than a similar chromosome with fewer of these particular words, whereas the presence or absence of words that are common to a variety of classifications will not significantly affect their chromosomes' classification performance.
- the performance of the evolved chromosomes can be expected to increase.
- a single chromosome, or subset of words is selected as the best performing set of words for distinguishing among program classifications.
- the need for a selection of a set of features that provides an effective and efficient means of characterizing or classifying content material is particularly important as the resources available for such characterizing or classifying become limited. For example, as technologies become available, viewers will expect their newly acquired home entertainment systems to provide program selection assistance, based on a "preferences" profile. These systems, however, will typically contain limited processing and storage capabilities, and may not, for example, be able to store every word and phrase of every synopsis available for such selection assistance. The inclusion of a non-discriminating word in the limited storage will be wasteful, and, more significantly, may also decrease the classification accuracy by introducing false distinctions. Thus, a classification system must be effective in the dual task of selecting effective discriminating features and excluding counter-productive non- discriminating features, and, in general, the effects of including or excluding a feature are non-additive.
- Evolutionary algorithms hold the promise of providing an identification of the most effective words, or features, to include in a classification system having limited processing and storage capabilities, and this invention addresses a method and apparatus that further enhance the use of evolutionary algorithms for identifying effective feature subsets.
- An initial set of features is defined that includes a large number of potential features, including the generated features that are combinations of other features. These features include, for example, all of the words used in a collection of content material that has been previously classified, as well as combination features based on these features, such as all the noun and verb phrases used.
- This pool of original features and combination features is provided to an evolutionary algorithm for a subsequent evaluation, generation, and determination of the best subset of features to use for classification. In this evaluation and generation process, each combination feature is processed as an independent feature, independent of the features that were used, or not used, to form the combination feature.
- a particular phrase that is generated as a combination of original feature words may be determined to be a better distinguishing feature than any of the original feature words and a more efficient distinguishing feature than an unrelated selection of the individual feature words, as might be provided by a conventional evolutionary algorithm.
- the resultant best performing subset is subsequently used to characterize new content material for automated classification. If the automated classification includes a learning system, the evolutionary algorithm and the generated combination features are also used to train the learning system.
- FIG. 1 illustrates an example block diagram of a feature set selection system with a feature combination generator in accordance with this invention.
- FIG. 2 illustrates an example block diagram of a classification system for classifying content material based on a preferred feature set as determined by the feature set selection system in accordance with this invention.
- FIG. 3 illustrates an example block diagram of a classification system for classifying content material via a learning system in accordance with this invention.
- the same reference numerals indicate similar or corresponding features or functions.
- a conventional feature selection system based on individual features may include both the features of "red” and “cross”, but because they are virtually independent features, the distinguishing capabilities of these features is likely to be poorer than a single feature "red cross”; and, because the feature "red” may be strongly correlated to another classification, the resultant classification using the independent feature “red” may be in error.
- the occurrence of "the red cross” would likely be more indicative of a classification than an occurrence of "a red cross”, whereas a conventional classification process would not use the word "the” as a distinguishing feature.
- combination features are generated from the individual features that are conventionally used to classify content material.
- these generated combination features are treated as being substantially independent of the features that form the combination. For example, if “red” is strongly correlated with an "art” classification, and “red cross” is strongly correlated to a "humanitarian” classification, both features, the original “red” feature and the generated “red cross” feature, may be included in the feature set that is used to classify content material.
- word phrases that may have effective distinguishing capabilities
- other feature combinations may be formed.
- particular director-producer, director-actor, actor-actress combinations may provide for a better characterization of content material than the individual director, actor, actress, and producer features.
- the combination of the provider of a web site and particular key words or phrases may facilitate a more effective characterization of a web page.
- the combination of "Philips" as the provider of information on a web page and "entertainment systems" as a key phrase may characterize a page differently than the presence of "Philips” and "entertainment systems” on a web page that is provided by a different provider. Because the number of potential basic features and combination features is virtually limitless, requiring therefore the use of a subset of features for content material classification, an evolutionary algorithm is used in a preferred embodiment for selecting which of the features to use.
- FIG. 1 illustrates an example block diagram of a feature set selection system 100 in accordance with this invention.
- the example selection system 100 includes an evolutionary algorithm 160 that is used to generate sets of features that are evaluated with regard to their effectiveness in classifying content material.
- the evolutionary algorithm 160 uses a pool of features 1 10 as candidate features for inclusion in the selected set of features that are used in the classification process.
- a feature combination generator 140 is used to generate combination features 141 that include combinations of individual features. Techniques are conventionally available for identifying noun-phrases and verb-phrases, by identifying, for example, adjectives followed by nouns, adverbs followed by verbs, adverb-adjective-noun combinations, and so on. A simpler technique, such as the selection of every sequential pair of words and every sequential triplet of words, or every pairing of non-word features, such as actor-director, producer-director, actor-producer, and so on, may also be used in the feature combination generator 140.
- the number of different features, including both the basic features and the combination features, that can be used to facilitate the classification of content material is very large, particularly when features can be formed as combinations of other features.
- the number of possible subsets that may be drawn from a very large pool of features can be astronomically large.
- the evolutionary algorithm 160 is used to determine a set of features that is likely to be more effective than other sets of features for a given classification task.
- Evolutionary algc rithms operate via an iterative offspring production process, and include genetic algorithms, .nutation algorithms, and the like.
- the offspring production process of an evolutionary algorithm is used to determine which particular sets of genes are most effective for performing a given task, using a directed trial and error search.
- a set of genes, or attributes, is termed a chromosome.
- a reproduction-recombination cycle is used to propagate generations of offspring.
- members of a population having different chromosomes mate and generate offspring.
- These offspring have attributes passed down from the parent members, typically as some random combination of genes from each parent.
- a classic genetic algorithm the individuals that are more effective than others in performing the given task are provided a higher opportunity to mate and generate offspring.
- the individuals having preferred chromosomes are given a higher opportunity to generate offspring, in the hope that the offspring will inherit whichever genes allowed the parents to perform the given task effectively.
- the recombination phase of the reproduction-recombination cycle effects the formation of the next generation of parents based on a preference for those exhibiting effectiveness for performing the given task. In this manner, the number of offspring having attributes that are effective for performing the given task will tend to increase with each generation.
- Paradigms of other methods of generating offspring such as asexual reproduction, mutation, and the like, are also used to produce generations of offspring having an increasing likelihood of improved abilities to perform the given task.
- the population consists of members having features that may be effective in classifying content material.
- some features represent combinations of other features, independent of the individual features. That is, for example, the phrase “flying saucer” may be a feature, whose effectiveness in characterizing and classifying content material is processed substantially independent of the "flying" feature and the “saucer” feature. That is, the feature “flying saucer” will be passed on to future generations, or not passed on to future generations, without regard to whether the features "flying” or “saucer” are passed on.
- noun phrases and verb phrases are treated as features that are independent of the word features that form such phrase features; director-actor features are independent of the particular director or actor feature, and so on. It has been found that this independent consideration of combination features is particularly well suited for the selection of classification features for use in a limited-resource embodiment, such as the aforementioned embodiment for a home entertainment system. That is, if the number of features that can be utilized for a particular embodiment is limited, the independent consideration of combination features will often lead to an elimination of redundant feature items. Assume, for example, that that aforementioned "flying saucer" feature is one that is highly effective in determining whether a given program is classified as "science fiction".
- the "flying" and “saucer” feature genes are likely to die out, because the marginal effectiveness gained or lost by including or not including the “flying” or the “saucer” feature gene is likely to be minimal in chromosomes that contain the "flying saucer” gene.
- the "flying” and “saucer” feature genes die out in a limited-feature embodiment, they are replaced by other features, such as a "murder" feature gene that is effective in determining whether a given program is classified as "mystery”.
- features are defined that may potentially facilitate the classification of content material, and, in a preferred embodiment, these features include combinations of other features.
- Candidate sets of select features are encoded as chromosomes that reflect different sets of abilities for distinguishing the content material to facilitate classification of the content material. Some sets of features are more effective for classifying the content material than other sets. By generating offspring from the members having chromosomes that are more effective for classification than others, the effectiveness of the offspring for properly classifying content material is likely to increase.
- a pool of features 110 is provided that includes those features that may potentially facilitate a classification of content material. As discussed above, these features may include the words used in content material, the words used in the synopses of content material, the creator of the content material, the performers in the content material, and so on.
- the feature combination generator 140 augments this pool of features 110 with combination features 141, as discussed above.
- multiple techniques are employed in the combination generator 140 to generate combination features: a phrase identifier creates a feature for each phrase, and a combinatorial generator generates a variety of combinations of the features that are not word based.
- the combination generator 140 also allows for the predefinition of likely combinations, such as director-actor, as well as the generation of substantially random combinations.
- a set selector 120 creates a set of features from this pool of features 1 10.
- the set selector 120 provides an initial population of feature sets 130 to a classification evaluator 150 to evaluate the effectiveness of each set of features 131 for classifying a collection 190 of preclassified content material.
- the collection 190 contains content material items 191 and the proper classification 192 of each of the content material items 191. That is, for example, the collection 190 may be a collection of information regarding television programs, and the proper classification 192 is the category within which an existing program guide placed each television program 191, such as comedy, drama, sci-fi, mystery, news, and so on.
- the proper classification 192 may be provided by a potential viewer who classifies each program 191 as "strongly likes", “likes", “no opinion”, "dislikes", and
- the collection 190 may contain the information regarding all television programs provided during the past month, and the proper classification 192 is whether a particular viewer "watched” or “didn't watch” each program 191. Using an on-line monitor of the programs selected for viewing, this simpler embodiment allows for a classification of each television program 191 into the two classes of watched and not- watched without requiring a direct user input.
- the collection of content material 190 may be a collection of electronic documents, a collection of abstracts, a collection of web pages, and so on, and the proper classification 192 may be "fiction", “history”, “gossip", and so on. Or, the classification 192 may merely be “viewed” and “not- viewed”.
- the classification evaluator 150 determines the effectiveness 151 of each candidate set of features 131 for providing a classification that corresponds to the proper classification 192 of each content material item 191.
- the evolutionary algorithm 160 thereafter provides parameters to the selector 120 for creating the next generation of feature sets 130, based on the effectiveness 151 of the prior generation of feature sets 130.
- the evolutionary algorithm 160 provides parameters 161 that favor the generation of sets having features common with the more effective sets of the prior generation.
- the evolutionary algorithm 160 continues to produce, via the set selector 120, generation after generation of candidate sets of features until a preferred set 131' is identified, typically the best performing set of features 131 found during this offspring generation process.
- a number of techniques are available for terminating the search for the preferred set 131'.
- a fixed time limit may be placed on the offspring generation process, the number of generations may be limited, convergence characteristics may be used to terminate the process when the incremental gain of each generation is below a cutoff limit, and so on.
- a feature set 131 ' is found that has been shown to be effective in classifying the collection of prior classified material 190, it is used by a classifier 240 to classify new content material 291.
- the classifier 240 uses the same classification process that is used in the classification evaluator 150. If, for example, the preferred feature set 131 ' was effective for classifying programs 191 as comedy, mystery, drama, etc., and the same classification process is used in the classifier 240, it is reasonable to assume that this same set 131' will be effective for classifying unknown programs 291 into comedy, mystery, drama, etc.
- the classifier 240 could use the set 131' for classifying synopses of upcoming programs, and present the results of the classification process as a list of "suggested programs to view", based on the viewer's prior classifications of programs 190.
- a number of techniques can be applied to evaluate a set's effectiveness in classifying a collection 190 of content material.
- a learning system is used to learn how to best apply each set >f features 1 1 to the given classification task.
- FIG. 3 illustrates an example block diagram of a classification system 300 for classifying content material 291 via a learning system 320 in accordance with this invention.
- the classification system 300 provides the functions illustrated as the classification evaluator 150 of FIG. 1 and the classifier 240 of FIG. 2.
- a portion of the collection of prior classified content material (item 190 in FIG. 1) is used to provide training content material 191 A, the remainder providing evaluation content material 191B.
- the training content material 191 A is provided, via switch SI, to an input processor 310.
- the input processor processes the content material 191 A to provide feature values 31 1 corresponding to the set of features 131 being used to train the learning system 320. For example, if the set of features includes word or phrase features, the input processor determines whether or not the content material 191 A contains each of the word or phrase features, and, depending upon the learning system 320, perhaps the number of occurrences of each word or phrase feature.
- the learning system 320 While the training content material 191 A is being provided to the learning system 320, the learning system 320 is placed in a training mode, illustrated by the switch 329. Using techniques common in the art, such as the adjustment of weights of nodes in neural network, or the adjustment of correlation factors in a Bayesian classifier, the learning system 320 is trained to increase the likelihood that the given feature set 131 will provide a classification corresponding to the proper classification 192A corresponding to the training content material 191 A. Subsequent training content material items 191 A are similarly applied to the learning system 320 to increase the overall likelihood that the feature set 131 would properly classify the training content material 191 A.
- the learning system 320 After the learning system 320 is trained to optimize the performance of the feature set 131 relative to the training content material 191 A, the previously classified evaluation content material 19 IB is provided to the input processor 310, via switch SI , and the corresponding feature values 31 1 are applied to learning system 320.
- the learning system 320 is operated in an execute mode, illustrated by switch 329, when the evaluation content material 191B is applied, so that the learning system 320 provides a classification 241 of the evaluation content material 191B based on the feature set 131 that was used to train the learning system 320.
- the determined classification 241 is provided to an evaluator 350, via the switch S2.
- the evaluator 350 compares the determined classification 241 with the proper classification 192B corresponding to the content material 191B.
- the evaluator 350 After processing each of the evaluation content material items 191B using the given feature set 131, the evaluator 350 provides a measure of effectiveness 151 to the evaluation algorithm 160, corresponding to the classification effectiveness of the given feature set 131. As discussed above, the evolutionary algorithm 160 provides selection parameters 161 to the set selector 120, based on the effectiveness of previously evaluated feature sets 131.
- the evolutionary algorithm 160 and set selector 120 After a sufficient number of feature sets 131 are processed and evaluated against the evaluation content material 191B, the evolutionary algorithm 160 and set selector 120 provides the preferred feature set 131' as a final input to the input processor 310. Depending upon whether the parameters corresponding to the training of the learning system are saved for each evaluated feature set 131, the learning system 320 is either reloaded with these parameters, or retrained using these parameters.
- the entire collection 190 of previously classified content material 191 may be applied as the training content material 191 A, to potentially improve the likelihood of the preferred set 131' being able to classify new content material 291, by exposing the preferred set 131' to a larger variety of content material 191.
- the switch SI is switched to receive the new content material 291
- the switch 329 is switched to place the learning system 320 into the execute mode
- the switch S2 is switched to the production mode.
- the system 300 provides a determined classification 241 of that new content material 291, based on the preferred feature set 131'. Note that after the preferred feature set 131 ' is selected and the learning system
- the classification system 300 can be embodied on a relatively large computing system to effect the training and evaluation required to determine the preferred set of features for a given classification task, and then the results of this determination, including the parameters that optimize the performance of a classifier 240 for the determined set of features, can be downloaded to a limited-capacity classifier 240.
- a set-top box is used to interface with a classification system 300 that is located at a site on the Internet, and the results of the determination of the preferred feature set and related parameters are subsequently downloaded to the set-top box.
- the combination features 141 are presented above as inclusive combinations, for ease of understanding. That is, for example, the feature “red cross” includes both the “red” and “cross” feature occurring sequentially. Alternatively, a combination feature may be defined as the occurrence of one feature in the absence of another feature, such as “red” without “cross” immediately following, or “cross” without “red” immediately preceding. Such variations, and others, will be evident to one of ordinary skill in the art, and included within the spirit and scope of the following claims.
Abstract
Description
Claims
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR1020017012812A KR20010113779A (en) | 2000-02-07 | 2001-01-11 | Multi-feature combination generation and classification effectiveness evaluation using genetic algorithms |
DE60128405T DE60128405D1 (en) | 2000-02-07 | 2001-01-11 | CREATING AND EVALUATING THE USEFULNESS OF A CLASSIFICATION SYSTEM BASED ON SEVERAL CHARACTERISTICS USING GENETIC ALGORITHMS |
EP01951179A EP1397759B1 (en) | 2000-02-07 | 2001-01-11 | Generation and effectiveness evaluation of a multi-feature classification system using genetic algorithms |
JP2001558869A JP2003534583A (en) | 2000-02-07 | 2001-01-11 | Genetic Algorithm-Based Generation and Classification Effectiveness Evaluation of Multiple Feature Combinations |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US09/498,882 US6892191B1 (en) | 2000-02-07 | 2000-02-07 | Multi-feature combination generation and classification effectiveness evaluation using genetic algorithms |
US09/498,882 | 2000-02-07 |
Publications (2)
Publication Number | Publication Date |
---|---|
WO2001059610A2 true WO2001059610A2 (en) | 2001-08-16 |
WO2001059610A3 WO2001059610A3 (en) | 2003-12-24 |
Family
ID=23982888
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/EP2001/000311 WO2001059610A2 (en) | 2000-02-07 | 2001-01-11 | Generation and effectiveness evaluation of a multi-feature classification system using genetic algorithms |
Country Status (7)
Country | Link |
---|---|
US (1) | US6892191B1 (en) |
EP (1) | EP1397759B1 (en) |
JP (1) | JP2003534583A (en) |
KR (1) | KR20010113779A (en) |
AT (1) | ATE362141T1 (en) |
DE (1) | DE60128405D1 (en) |
WO (1) | WO2001059610A2 (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7130866B2 (en) | 2002-07-30 | 2006-10-31 | Koninklijke Philips Electronics N.V. | Controlling the growth of a feature frequency profile by deleting selected frequency counts of features of events |
US20210125080A1 (en) * | 2019-10-24 | 2021-04-29 | International Business Machines Corporation | Method and apparatus for enhancing effectivity of machine learning solutions |
Families Citing this family (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB0114236D0 (en) * | 2001-06-12 | 2001-08-01 | Hewlett Packard Co | Artificial language generation |
US7139738B2 (en) | 2002-06-27 | 2006-11-21 | Koninklijke Philips Electronics N.V. | Face recognition using evolutionary algorithms |
US20040010480A1 (en) * | 2002-07-09 | 2004-01-15 | Lalitha Agnihotri | Method, apparatus, and program for evolving neural network architectures to detect content in media information |
GB2407657B (en) * | 2003-10-30 | 2006-08-23 | Vox Generation Ltd | Automated grammar generator (AGG) |
US20060212279A1 (en) * | 2005-01-31 | 2006-09-21 | The Board of Trustees of the University of Illinois and | Methods for efficient solution set optimization |
US7529748B2 (en) * | 2005-11-15 | 2009-05-05 | Ji-Rong Wen | Information classification paradigm |
AU2006320692A1 (en) * | 2005-11-29 | 2007-06-07 | Google Inc. | Detecting repeating content in broadcast media |
US8131656B2 (en) * | 2006-01-31 | 2012-03-06 | The Board Of Trustees Of The University Of Illinois | Adaptive optimization methods |
US7979365B2 (en) * | 2006-01-31 | 2011-07-12 | The Board Of Trustees Of The University Of Illinois | Methods and systems for interactive computing |
US7831531B1 (en) | 2006-06-22 | 2010-11-09 | Google Inc. | Approximate hashing functions for finding similar content |
US8019593B2 (en) * | 2006-06-30 | 2011-09-13 | Robert Bosch Corporation | Method and apparatus for generating features through logical and functional operations |
US8411977B1 (en) | 2006-08-29 | 2013-04-02 | Google Inc. | Audio identification using wavelet-based signatures |
EP2297680A4 (en) * | 2008-05-01 | 2013-06-19 | Icosystem Corp | Methods and systems for the design of choice experiments and deduction of human decision-making heuristics |
US8346800B2 (en) * | 2009-04-02 | 2013-01-01 | Microsoft Corporation | Content-based information retrieval |
CN103631802B (en) * | 2012-08-24 | 2015-05-20 | 腾讯科技(深圳)有限公司 | Song information searching method, device and corresponding server |
US10439891B2 (en) * | 2014-04-08 | 2019-10-08 | International Business Machines Corporation | Hyperparameter and network topology selection in network demand forecasting |
KR101871940B1 (en) | 2014-05-12 | 2018-06-27 | 한화에어로스페이스 주식회사 | Method and system for establishing predictive model of plant abnormality |
EP3288409A4 (en) * | 2015-04-26 | 2019-04-10 | Samuel Lightstone | Method, device and system for fitness tracking |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5897629A (en) * | 1996-05-29 | 1999-04-27 | Fujitsu Limited | Apparatus for solving optimization problems and delivery planning system |
Family Cites Families (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4821333A (en) * | 1986-08-22 | 1989-04-11 | Environmental Research Inst. Of Michigan | Machine learning procedures for generating image domain feature detector structuring elements |
US5146406A (en) * | 1989-08-16 | 1992-09-08 | International Business Machines Corporation | Computer method for identifying predicate-argument structures in natural language text |
US5048095A (en) * | 1990-03-30 | 1991-09-10 | Honeywell Inc. | Adaptive image segmentation system |
US5798785A (en) | 1992-12-09 | 1998-08-25 | Discovery Communications, Inc. | Terminal for suggesting programs offered on a television program delivery system |
US5343251A (en) | 1993-05-13 | 1994-08-30 | Pareto Partners, Inc. | Method and apparatus for classifying patterns of television programs and commercials based on discerning of broadcast audio and video signals |
US5410344A (en) | 1993-09-22 | 1995-04-25 | Arrowsmith Technologies, Inc. | Apparatus and method of selecting video programs based on viewers' preferences |
US5479523A (en) * | 1994-03-16 | 1995-12-26 | Eastman Kodak Company | Constructing classification weights matrices for pattern recognition systems using reduced element feature subsets |
US5758257A (en) | 1994-11-29 | 1998-05-26 | Herz; Frederick | System and method for scheduling broadcast of and access to video programs and other data using customer profiles |
US5682206A (en) | 1995-09-25 | 1997-10-28 | Thomson Consumer Electronics, Inc. | Consumer interface for programming device |
-
2000
- 2000-02-07 US US09/498,882 patent/US6892191B1/en not_active Expired - Fee Related
-
2001
- 2001-01-11 EP EP01951179A patent/EP1397759B1/en not_active Expired - Lifetime
- 2001-01-11 KR KR1020017012812A patent/KR20010113779A/en active IP Right Grant
- 2001-01-11 AT AT01951179T patent/ATE362141T1/en not_active IP Right Cessation
- 2001-01-11 DE DE60128405T patent/DE60128405D1/en not_active Expired - Lifetime
- 2001-01-11 WO PCT/EP2001/000311 patent/WO2001059610A2/en active IP Right Grant
- 2001-01-11 JP JP2001558869A patent/JP2003534583A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5897629A (en) * | 1996-05-29 | 1999-04-27 | Fujitsu Limited | Apparatus for solving optimization problems and delivery planning system |
Non-Patent Citations (3)
Title |
---|
CHANG E I ET AL: "Using genetic algorithms to select and create features for pattern classification" -, 17 June 1990 (1990-06-17), pages 747-752, XP010006887 * |
JIHOON YANG, VASANT HONAVAR: "Feature Subset Selection Using A Genetic Algorithm" ACM, [Online] vol. 97, no. 02, 3 May 1997 (1997-05-03), pages 1-12, XP002256773 Retrieved from the Internet: <URL:http://citeseer.ist.psu.edu/yang98fea ture.html> [retrieved on 2003-10-07] * |
SHETH B ET AL: "Evolving agents for personalized information filtering" PROCEEDINGS OF THE CONFERENCE ON ARTIFICIAL INTELLIGENCE FOR APPLICATIONS. ORLANDO, MAR. 1 - 5, 1993, LOS ALAMITOS, IEEE COMP. SOC. PRESS, US, vol. CONF. 9, 1 March 1993 (1993-03-01), pages 345-352, XP010125585 ISBN: 0-8186-3840-0 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7130866B2 (en) | 2002-07-30 | 2006-10-31 | Koninklijke Philips Electronics N.V. | Controlling the growth of a feature frequency profile by deleting selected frequency counts of features of events |
US20210125080A1 (en) * | 2019-10-24 | 2021-04-29 | International Business Machines Corporation | Method and apparatus for enhancing effectivity of machine learning solutions |
Also Published As
Publication number | Publication date |
---|---|
JP2003534583A (en) | 2003-11-18 |
DE60128405D1 (en) | 2007-06-21 |
EP1397759A2 (en) | 2004-03-17 |
ATE362141T1 (en) | 2007-06-15 |
EP1397759B1 (en) | 2007-05-09 |
US6892191B1 (en) | 2005-05-10 |
WO2001059610A3 (en) | 2003-12-24 |
KR20010113779A (en) | 2001-12-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US6892191B1 (en) | Multi-feature combination generation and classification effectiveness evaluation using genetic algorithms | |
US9164994B2 (en) | Intelligent default weighting process for criteria utilized to score media content items | |
CN103052954B (en) | Commending system is retrieved based on profile content | |
US20100070507A1 (en) | Hybrid content recommending server, system, and method | |
US6751776B1 (en) | Method and apparatus for personalized multimedia summarization based upon user specified theme | |
US5410344A (en) | Apparatus and method of selecting video programs based on viewers' preferences | |
US8220023B2 (en) | Method for content presentation | |
US20070288965A1 (en) | Recommended program information providing method and apparatus | |
US20020093591A1 (en) | Creating audio-centric, imagecentric, and integrated audio visual summaries | |
JP2005521144A (en) | Recommendation system using multiple recommendation scores | |
US20050165782A1 (en) | Information processing apparatus, information processing method, program for implementing information processing method, information processing system, and method for information processing system | |
JP2006500859A (en) | Commercial recommendation device | |
CN108228541B (en) | Method and device for generating document abstract | |
Li et al. | Leave no user behind: Towards improving the utility of recommender systems for non-mainstream users | |
GB2566257A (en) | System and method for content discovery | |
EP3369252B1 (en) | Video content summarization and class selection | |
JP2007102489A (en) | Program data processor, program data processing method, control program, recording medium, and video recorder, reproduction device and information display device with program data processor | |
Shafaei et al. | Age suitability rating: Predicting the MPAA rating based on movie dialogues | |
Shafaei et al. | Rating for parents: Predicting children suitability rating for movies based on language of the movies | |
EP2151799A1 (en) | Recommander method and system, in particular for IPTV | |
Grimaldi et al. | Experimenting with music taste prediction by user profiling | |
EP1183622A1 (en) | Determining a distribution of a numeric variable | |
CN110717064A (en) | Personalized audio play list generation method and device and readable storage medium | |
Fernandes et al. | Unification of hdp and lda models for optimal topic clustering of subject specific question banks | |
Lavania et al. | A practical online framework for extracting running video summaries under a fixed memory budget |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A2 Designated state(s): JP KR |
|
AL | Designated countries for regional patents |
Kind code of ref document: A2 Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2001951179 Country of ref document: EP |
|
ENP | Entry into the national phase |
Ref country code: JP Ref document number: 2001 558869 Kind code of ref document: A Format of ref document f/p: F |
|
WWE | Wipo information: entry into national phase |
Ref document number: 1020017012812 Country of ref document: KR |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
WWP | Wipo information: published in national office |
Ref document number: 1020017012812 Country of ref document: KR |
|
WWP | Wipo information: published in national office |
Ref document number: 2001951179 Country of ref document: EP |
|
WWG | Wipo information: grant in national office |
Ref document number: 2001951179 Country of ref document: EP |