Abstract—thepersonalized web page recommendation is much needed these days. Generally,Web page recommendation systems are implemented in Web servers.They use data implicitly obtained as a collection of Webbrowsing patterns of the users for recommending webpages.The existing systemcollects the Web logs and generates a cluster ofsimilar users and recommends pages to the user by actively analysingit in online.
However the time complexity for analysing it in online ismore. In order to optimize this and to improve the correctnessof recommendation systems we propose the method of applying Firefly basedalgorithm for recommending Web pages along with Naive Bayesclustering. It clusters Web logs in offline usingNaive Bayes clustering technique.
To find the similarity betweenthe active user queries with other users in thecluster Firefly algorithm based similarity measure is used. Theproposed approach uses a probability basedclustering which eliminates the odd records while forming clusters.Firefly algorithm meticulously searches the generated weblogs present in the cluster of the active user and recommends the toppages. Firefly algorithm utilizes time efficiently, thus it is used forprocessing in online.
When pages are obtained, they areranked and the top pages that are more relevant tothe query are recommended.The efficiency of the system can be evaluatedusing measures like precision, recall-Score, Matthews’s correlation andFallout rate. The proposed approach is expected to improve timeutilization in online process as well as recommendsmore accurate Webpages. Introduction- Webpage recommendation system is a sub-domain of recommendation systems thatrecommends a set of Web pages to the users based on their past browsingpatterns. It is done by applying special mining techniques on the data that arepreviously gathered from the users which in turn discovers and extractinformation from Web documents and services. The major concern is to findreliable and efficient recommendation algorithms.
Recommendation systemtypically produces the result by following one of the two ways – throughcollaborative and content based filtering. A. CollaborativeFiltering Mostrecommendation system has wide use of collaborative filtering for recommendingitems. This method lies on collecting and processing the information’s onuser’s behaviours or activities and then predicting the items relating to theirsimilarity with other users. Collaborative filtering approaches building astructure from a user’s past behaviours and decisions of other similar users.
This model is used to predict user interested items. Since collaborativefiltering is independent of machine analysable contents, it is capable ofrecommending for complex items accurately without “understanding” of the itemitself. B. ContentBased Filtering Content basedfiltering is a widely used approach for designing recommendation systems. Thistechnique is based on a definition of item and a user’s preferred profile. In acontent based recommendation systems, the keywords are considered as user’sinterest. It utilize a series of distinct property of an item for obtaining andrecommending items with same properties.
These approaches are continuallycombined as Hybrid Recommendation Systems. These algorithm try to recommenditems based on examining the items that are liked by a user in the past or inthe present. In general, various items of candidate set are compared with itemsthat are rated by the user in the past and the best matching items arerecommended. Literature surveyRecommendationsystem plays a vital role in recommending personalized items for the usersbased on their interest in a web services. The webalso contains a rich and dynamic information’s. The amountof information on the web is growing rapidly, aswell as the number of web sites and webpages per web site. Predicting the needsof a web user as she visits web sites has gainedimportance. Many webpage recommendation systemwere developed in the past, since they compute recommendationsin online process, their time utilization shouldbe efficient.
A system 4 that uses support vectormachine (SVM) learning based model wasdeveloped for computing similarity between two itemswhich performed better than latentfactor approach for group recommendations. Since thematrix representation was followed, thedata sparsity problem was solved.However, the system was not ableto stably scale when size of the groupdynamically increased. Hybridrecommender systems that combines two or morerecommendation techniques was designed 5. Iteliminates any weakness which exist when only one recommendersystem is used. There are several ways in which the systems can becombined, such as weighted hybrid recommender where the score of a recommended item iscomputed from the results of all of the availablerecommendation techniques present in the system.
However, data sparseness wasstill a problem, the system may generate week recommendations iffew users have rated the same items and alsothe system doesn’t overcome the cold startproblem. Hyperspectral sensors can acquire hundreds ofcontiguousbands over a wide electromagnetic spectrum for eachpixel. The rich spectral information allowsfor distinguishing materials with subtle spectral discrepancy, butit usually leads to the “curse ofdimensionality”. To address this, an improved firefly algorithm based bandselection method 8 was used. The Fireflyalgorithm is an evolutionary optimization algorithm proposed by Yang13. After the initializations of parameters, the brightness is calculatedwith the objective function (2.1), where t is themaximum iterations, ? is the step size and ? is thelight absorbance of m number of fireflies. The moment states are then evaluatedand the bands are selected.
In order to avoid employing an actual classifierwithin the band searching process to greatly reduce computational cost,criterion functions that can gauge class separability are preferred whichprovided better results. Firefly algorithm also hada faster convergence even at the size of thedata is larger To improve the accuracy of similarity measure, a natureinspired algorithm which is based in the behaviour ofFireflies wereintroduced 10.We consider separate effects for ratings ofusers with similar opinions and conflicting opinions. In orderto generate initial population of fireflies, half of population randomlygenerated and the other half of population are randomly generated. Meanabsolute error was chosen as objective function to measure recommendation accuracy whichis obtained by difference between predicted rating and real rating. An optimalsimilarity measure via a simple linear combination of values and ratio ofratings for user-based collaborative filtering provides better results. Itincreased speed of finding nearest neighbours of active user and reduceits computation time.
Similarity function equationbasedon Firefly algorithm was simpler than the equationused in traditional metrics therefore, the proposed method provided recommendationsfaster than traditional metrics. Graph colouring problems aregenerally discrete. Algorithms to discrete problems arequite complex. A new algorithmbased on Similarity and discretize firefly algorithm directly without anyother hybrid algorithm was developed 11. It wasadoptable to dynamic graph sizes. A system for assigningan electronic document to one or more predefined categoriesor classes based on its textual context and use of agglomerativeclustering algorithm was developed 6. This type ofclustering along with sample correlation coefficient assimilarity measure, allowed high indexing term space reduction factor witha gain of higher classification accuracy. In order tominimize noise and outlier data, a modified DBSCALE algorithm using Naïve Bayeshas been designed 7.
This algorithm is basically a prospect basedutility. This function is used toestimate the outlier clusterdata and increase the correctness rate of algorithm on giventhreshold value. Since Naïve Bayes is a probability based function,it removes outlier cluster data and increases the correctness rate according tothreshold value.
It also computes maximum posterior hypothesis for outlierdata. In order to minimize noise and outlier data, a modified DBSCALE algorithmusing Naïve Bayes has been designed 7. This algorithm is basically a prospectbased utility. This function isused to increase thecorrectness rate of algorithm on given threshold value and toestimate the outlier cluster data. Since NaïveBayes is a probability basedfunction, it removes outlier cluster data andincreases the correctness rate according tothreshold value.
It also computes maximum posteriorhypothesis for outlier data. The memorybased collaborative system uses matrixbased computation and solves data sparsity problem but, scalabilityof the system cannot be stable when size of the group dynamically increases.Hybrid system could be helpful in overcomingthe scalability issue but it again leads to cold start problem. To eliminate outliers as well as overcomingother twoproblems Naive Bayes clustering, a probability basedmethod was used in past. Firefly algorithm has a fasterconvergence and searches all possible subsets with better timeutilization. Thus, to design an efficient recommendation system,Naïve Bayes method can be followed for clustering inoffline. Since the time complexity should be less, Fireflyalgorithm that is more efficient in terms of timeutilization, it can be used for calculating similarity in online.
Combinationof these two technique might increase the accuracy of therecommendation system as well as results in efficienttime utilization. III. Overview of the proposed work Initially, the web log files are obtained fromthe 1 America Online Inc. The log files consists of fivefields i.e. anonymous ID for individual user, query of each user alongwith query time, list of URLs which user proceeded and itsrank in the result. These logs are collectedand grouped based on anonymous ID. The URL among allthe users are obtained and its content are downloaded andprocessed.
The processing of data includes removal ofstop words from the URL’s data andkeyword extraction. Similar users are clustered based on fetchedkeywords by using Naïve Bayes clustering technique which provides efficientclusters compared to clustering by the use of association rules. The createdclusters are given to online component. In online process, when an active usergives a query, the keywords from the query is extracted. Thesimilarity between the extracted keywords with the other usersin the same cluster of the active useris calculated using Firefly similarity measure. Thesimilarity values are sorted along with the web pagesbrowsed by similar users in the cluster.
The top k web pages arerecommended for the active useras a result. IV. The proposedwork The proposedsystem follows a linear process of initially collecting theweb logs and processing them followed by clustering similar usersby Naïve Bayes clustering technique and finally generatingrecommendations based on a similarity measure from fireflyalgorithm. A. Preprocessing of Web Logs The weblogs are collected form 1 AOL Inc. It consists of 20million web queries from 650 thousand real users over 3months. The data set includes anonymous ID, query, querytime, item rank and click URL.
The log file containsmany number of users along with the web pages visited bythem. It is validated and separated based on anonymous ID. The useris separated into individual file using anonymous ID. The content fromthe URL are fetched and downloaded.Those keywords are processed which undergoes stopwords removal andstemming process. The final keywords are thenextracted. The features like keywords, Timings, Frequency, Click URL andRevisit are fetched.
The user profile is constructed using thosefeatures. The user profile that constructed is basedon the features that are takenform the user log files. Timing: The timingthat the user spent on that particular URL · Frequency: The amount of time the user visited the URL · Clickstream: The number of click stream that are visited by user · Revisit: Whether the user visited the web page The keywords aregenerated from the data fetched form theURL. Timing for each URL is estimated fromthe given date and time by calculating the differencebetween the each URL that are searched in a singleday by having some time constraints. Frequencyis hence calculated such that number of times the userclicked the URL. The clickstreams are those that areclicked by the user for additional information.
The timingof revisit is calculated such that to decide whether theuser preferred it much or not. Keywords:Keywords are those which are extracted from the URL.The information from the URL is hence collected and processed toobtain features of the user.
B. Naïve Bayes Clustering Clustering, alsoknown as unsupervised classification, is a descriptive task with manyapplications. Clustering is decomposition or partition of a data set intogroups such that the object in one group are similar toeach other but as different as possible from theobject in other groups. Three main approach for clustering of data is partitionbased clustering, hierarchical clustering and probabilistic modelbased clustering. Probabilistic model based clustering is asoft clustering were an object can be in many clusterfollowing a probability distribution. A clustering is useful if it producessome interesting insight in the problem that weare analysing. Naïve Bayes clustering is also a probabilistic clustering techniquethat is based in Bayes theorem with strong independentassumption between features.
The feature variables canbe discrete or continuous. This probabilistic clustering lies on nominal andnumeric variables in the data set and its novelty lies in the use of mixture oftruncated exponential (MTE) densities to model the numeric variables. In NaïveBayes clustering the class is the only root variable and allthe attributes are conditionally independent given the class. Theclustering problem reduces to take a data set of instancesand a previously specified number of clusters (k), and work outeach cluster’s distribution and the population distribution betweenthe clusters. To obtain these parameters the expectation maximization (EM)algorithm is used. Since Naïve Bayes clustering isa probability based techniques.
The items belongs to thecluster if and only if it has a relation to it. This helps ineliminating outlier data in the process of clustering. It also provides properclustering with less computations. The given dataset is divided into two parts,one for the training and other for testing. For eachrecord in the test and train databases, the distribution of the classvariable is computed. According to the obtained distribution, a value for theclass variable is simulated and inserted in the corresponding cluster. Thelog-likelihood of the new model is computed.
If it is higher than the initialmodel, the process is repeated. Otherwise, the process is stopped,obtained clusters are returned. C. Optimisation Using Firefly Algorithm Fireflyalgorithm is an evolutionary algorithm that is based on thebehaviour of fireflies. Fireflies live in colonies and cooperate for thesurvival of the colony. Generally, in order to model the behaviour offireflies, three assumptions will always be considered i.e.
all fireflies arehomogeneous, Attractiveness of each firefly is related to its level ofbrightness, rightness of firefly is determined with an exponentialobjective function. Each firefly always emits a kindof light that by which attracts other fireflies. The amount of accessedlight depends on parameters such as distance and absorption coefficient of thesurroundings. The longer the distance the lesser the amount of accessed lightwill be.
Also in surroundings with high light absorption coefficient such asfoggy weathers, the intensity of light decreases. Thecertain issue is that every firefly regardless of its gender hasalways been attracted to and moved toward the brighter firefly.Firefly has a light intensity of its own. The key concept is, the firefly withlow light intensity is always attracted to the firefly with high lightintensity. This concept can be incorporated for calculating similarity. Byusing firefly based similarity measure unique and distinguished results can beobtained which is a useful feature for ranking. It can deal with highly non-linear, multi-modal optimization problems naturally andefficiently.
It does not use velocities, and thereis no problem as that associated with velocity in PSO. Thespeed of convergence is very high in probability of finding the globaloptimized answer. It has the flexibility of integration with other optimizationtechniques to form hybrid tools.
It does not require agood initial solution to start itsiteration process. Each web pages visited bythe user i are considered a firefly. The number of user visited theparticular page is assumed as the light intensity of the firefly. The objectivefunction is formulated based on the frequency and duration. Frequency iscalculated as the ratio to the number of visits per page to the average vestsof all pages. The duration isthe ratio of duration of page to the total duration of all the pages visited bythe user. Thus, the objective function can be defined as in equation 5.1Interest (i)= 2*Frequency (i)*Duration (i) Frequency (i)+Duration (i) (5.
1) The interest of all users in the cluster is calculated. Then the pagesto be recommended are found by using page rank algorithm 2 on the obtainedresult. The results after applying page rank algorithm is given as therecommended web page to the user.
D. Ranking the WebPages The result, set ofweb pages obtain should ranked in an order that the user might have higherinterest. Thus, they areranked in a sorted order basedon the interest of the active user. The associationrule checks the maximum possible combinationswhich provides more accurate pages. E. Recommendation Process The URL that areto be recommended will be identified based on ranking and similarity measure.The similarity measure is calculated among the users by comparing their similarinterest. From the obtained result of pages, page rank algorithmis used to rank the most relevant pages to the user.
Thus, resultant URL’s arerecommended to the users. Hencethe web page that is to be recommended tothe user will be more relevant. The use of Nave Bayes clustering willeliminate the outliers and Firefly based similarity calculation willcheck all the subsets of the clusters.