| Sažetak | Uzimajući u obzir neprestani rast konkurencije, uvođenje novih tehnologija te nove usluge koje se zasnivaju na njima, davatelji usluga moraju uvesti mehanizme za kvalitetno predviđanje koje postaje temelj pravovremenom uočavanju novih prilika na tržištu, prepoznavanju potencijalnih pogrešaka, planiranju resursa te u konačnici financijskom planiranju. Modeli rasta su jedna od najčešće korištenih metoda predviđanja na području informacijskih usluga. Međutim, korištenje isključivo modela rasta ima određene nedostatke kao što su ograničena preciznost te kašnjenje zbog potrebe za uzorkom povijesnih podataka u svrhu izračuna parametara. U disertaciji je predstavljen nov pristup u predviđanju rasta interesa korisnika informacijskih usluga zasnovan na poznatim modelima rasta te semantičkom rasuđivanju kojim se umanjuju navedeni nedostaci postojećih modela rasta. Novi pristup zasniva se na tri koraka. Prvi korak je definicija profila informacijske usluge sastavljenog od semantičkog opisa, koji omogućava automatiziranu usporedbu usluga i modela rasta koji opisuje interes korisnika za uslugu. Drugi korak je predviđanje interesa korisnika za informacijske usluge zasnovano na usporedbi nove s postojećim uslugama primjenom semantički-svjesnog modela. Treći korak je evaluacija predloženog pristupa za predviđanje interesa korisnika za informacijske usluge na studijskom primjeru YouTube videoisječaka. Rezultat evaluacije napravljene pomoću implementiranog radnog okvira za predviđanje na studijskom uzorku predstavljen je kroz mjere odstupanja predviđenih vrijednosti od stvarnih, te kroz analizu utjecaja pojedinih parametara algoritma za predviđanje interesa na preciznost predviđanja, odnosno na mjere odstupanja. |
| Sažetak (engleski) | Information and communication services' market is characterised with the constant growth of competition as well as continuous emergence of new technologies and, consequently, new services based on these technologies. To tackle market challenges, service providers must improve their business processes by introducing new mechanisms into existing planning and decision making processes. These mechanisms primarily relate to the timely anticipation of potential market opportunities, as much as potential mistakes, resource planning, and finally financial forecasting. Growth models provide a possibility for forecasting consumer acceptance for services, which makes them one of the possible solutions for these challenges. However, commonly used growth models have certain deficiencies which have to be mitigated by using other methods. These deficiencies are primarily limited precision and delays in forecasting caused by the need for at least a minimum set of historical data in order to calculate model parameters. This dissertation presents a new approach in forecasting of consumer interest for information services based on commonly used growth models and semantic reasoning. This new approach consists of three steps. The first step is the definition of information service semantic profile consisting of a semantic description, which enables automated service matchmaking, and growth model, which describes the consumers' interest in that particular service. The second step is forecasting consumers' interest in information services based on similar services determined by using semantic reasoning. The final step is an evaluation of the proposed approach using YouTube video clips as a case study. Chapter One – Forecasting in information and communication technology industry The first chapter provides an introduction to forecasting in information and communication technology domain. Most common forecasting methods can be divided into two groups: (i) methods based on estimation; and (ii) methods based on quantitative data. These methods are used for predicting certain indicators that describe the service and product lifecycle. One of the most important indicators is consumers' interest for a certain service or product. Observing various patterns in services' lifecycle, five major phases could be identified: (i) development; (ii) introduction; (iii) growth, (iv) maturity; and (v) decline. Growth models are one of the most commonly used forecasting methods in the information service domain, especially when describing the initial phases in the service lifecycle. Some of the most common growth models include: (i) logistic model; (ii) Bass model; (iii) Richards model; (iv) recursive models; and (v) multi-logistic models. This research is based on the Bass model, due to its characteristics which have made it efficient in modelling the initial growth of consumers' interest in this particular domain. Chapter Two – Semantic reasoning One of the main downsides of using growth models is the necessity of sufficient historical data for model parameter calculation. When there is no sufficient data, a different approach must be used for calculating these parameters. This research is based on the idea that similar services have similar growth under similar circumstances. Semantic reasoning is proposed to quantify similarities between services, thus creating the ground for an innovative approach in growth modelling. Semantic reasoning is primarily used for generating semantic profiles for information services, which can then be compared taking into account the real meaning of each attribute, attribute value data types, and relations between resources that describe a certain pair of services. Matchmaking of two semantic profile is performed starting on attribute level, and then gradually aggregating these similarities towards final profile similarity taking into account all attribute similarities and their respective weights. Such semantic matchmaking algorithm results in a numeric description of the similarity between services in the interval [0,1], where 0 (zero) means that the services have absolutely nothing in common, and 1 (one) means that the services are identical. Chapter Three – Consumer interest forecasting model for information services The third chapter presents the generic consumer interest forecasting model for information services. The model consists of five key entities that participate in service provisioning and acceptance forecasting: (i) service provider; (ii) contributors; (iii) consumers; (iv) consumer interest forecasting framework; and (v) data storage system. The service provider is the primary facilitator that enables infrastructure for information service provisioning. Contributors are the producers of resources that are put at consumers' disposal via service providers. Consumers are the persons whose consumption of particular service results in acceptance growth for that service. The forecasting framework performs data collection and semantic profiling of the services, semantic matchmaking of the created profiles, and consumer interest growth forecasting. Data storage system is used for storing semantic profiles and acceptance data for observed services. Chapter Four – Proposed implementation of consumer interest forecasting framework for information services Fourth chapter provides a more detailed insight in the forecasting framework components and their functionalities. The proposed framework consists of four modules. Semantic profiling module performs consolidation of data that describe information services and historical acceptance data used for calculation of growth model parameters. Semantic matchmaking module compares service profiles according to proposed algorithm through separate comparison for each attribute, and final aggregation of the individual attribute similarities into profile similarity. Module for growth modelling based on historical data uses the historical acceptance data in order to calculate the growth model parameters and respective deviation measures: (i) simple relative deviation; and (ii) weighted relative deviation. New service growth modelling module calculates growth model parameters based on similar service model parameters. The chapter is concluded by a flowchart representing the process of framework evaluation through a case study based on YouTube video clips. Chapter Five – Data collection and semantic profiling for YouTube video clip streaming service The fifth chapter presents the implementation of the framework segment responsible for fetching data about YouTube video clips and translating them into semantic profiles, thus creating a data sample for the framework evaluation. The first step in sample preparation is retrieving the basic video clip data in accordance with eligible combinations of attribute values for attributes which are considered relevant in the sense of sample representativeness. After having created the initial sample, the next step includes obtaining information about related video clips and video clips published by same contributors, all through the YouTube API. The final sample is formed after retrieving detailed data about the video clips and eliminating those whose data was of insufficient quality for automated processing. Detailed video clip description data and historical acceptance data is then used for generation semantic profiles according to the presented structure. Chapter Six – YouTube video clip semantic profile matchmaking The sixth chapter provides a more detailed insight into the implemented semantic matchmaking algorithm. Video clip attributes are divided into two groups: (i) ones related to the content description; and (ii) ones related to technical characteristics. The algorithm provides different methods of comparison with regards to the attribute value domains: (i) trivial identity check between two values used for attributes with very limited domain; (ii) similarity matrix for attributes with larger but limited domain; (iii) text comparison based on three text similarity measures; (iv) semantic resource comparison based on four measures that define their similarity and relationship; and (v) publish date comparison. Each method is presented through the respective example. The final step of the algorithm is summarising individual attribute similarities into the final service similarity. Chapter Seven – Viewership growth modelling for YouTube video clips based on historical acceptance data Framework segment responsible for growth model parameter calculation based on historical acceptance data uses weighted least squares method for determining the Bass model parameters. Weight determination method can be set as an input parameter to emphasise a certain part of the growth period, most commonly the later stages so the extrapolation of the model can provide better insight into the upcoming period and enable more precise forecasting. This research presents four distinct weight determination methods: (i) identical weights; (ii) arithmetical progression; (iii) geometrical progression; and (iv) logistic function. These methods are benchmarked through examples and respective deviations. Chapter Eight – Viewership growth modelling for YouTube video clips based on similar video clip acceptance growth The final chapter presents the growth forecasting algorithm for YouTube video clip viewership. Inputs for the algorithm are similarities with the remaining video clips in the sample, and their growth model parameters. Evaluation of the proposed algorithm, which also represents the final scientific contribution of the thesis, is performed as a forecasting simulation for a subset of the case study sample, called the evaluation sample. Similarity with all other video clips is calculated for each of the video clips from the evaluation sample, after which the selection of the most similar video clips is done. Using growth model parameters of these most similar video clips, the algorithm calculates growth model parameter which enables growth forecasting for evaluation sample video clips. Forecasting precision is evaluated through observing the influence of three forecasting parameters: (i) method for selection most similar video clips; (ii) weight determination method for least squares method; and (iii) forecasted growth period. Conclusion The thesis proposes a new approach to address the problem of consumer interest forecasting for information services. The proposed approach is based on the hypothesis that similar services have similar acceptance growth within similar circumstances. In accordance with that hypothesis, the research is focused on service matchmaking mechanism and a forecasting algorithm that uses calculated similarities between services and acceptance data for existing services in order to forecast acceptance growth for a new service. Such an approach eliminates the need for an initial acceptance data set in the process of determining growth model parameters. To evaluate the proposed approach on a case study, a framework is implemented and used for generating a case study data sample consisting of 7,683 video clips. The framework enables semantic profiling of the video clips based on their content descriptions, technical characteristics, and historical acceptance data, comparing these profiles based on attribute value matching, calculating growth model parameters from historical acceptance data, and, finally, acceptance growth forecasting based on similar services. The acceptance forecasting algorithm is evaluated on an evaluation sample consisting of 100 video clips through two defined deviation measures. The evaluation is based on observing deviation measures in correlation to three input parameters: (i) similar video clip selection method; (ii) weight determination method for the least squares method; and (iii) forecasted growth period. The most precise forecasting was achieved when taking into consideration only video clips with similarity greater than 0.90, and weights based on logistic function for the period of 90 days. Influence of each parameter is analysed through a total of 144 different combinations of input parameters. The conclusion is that the „absolute method“ with high similarity threshold (0.90) provides greater precision, but bearing in mind that it is applicable only for a very small part of the evaluation sample. Reducing that threshold, or using the „relative method“ for similar video clip selection, broadens the applicability, but at the cost of reduced precision. Weight determination method generally does not influence the forecasting precision, but in combination with certain growth periods can result in greater precision. Future work in means of improvements should include an extension of the data sample, and increase in forecasting precision. Besides adding additional video clips in the data set, further attributes should also be included in the research, especially having in mind constant YouTube API upgrades. Semantic matchmaking algorithm should be addressed from two aspects. First consists of a more detailed analysis of the correlation between specific attributes and acceptance growth in order to adjust the weights for the attributes and their respective groups. The second aspect is optimising the parts of the process that have proven to be time-consuming. One approach for this optimisation is rationalising the usage of the algorithm through video clip clustering. Additionally, other growth models should also be considered for acceptance growth approximation (i.e., Richards model). |
