Multimedia databases represent a powerful solution to those entities that require their data be stored in its original multimedia format. This type of database has the same concepts as the normal well-known information-storing database. Miscellaneous queries can be done on these multimedia data formats, thus the entire concept and use of a database system is successfully fulfilled. Multimedia have various forms such as documents, images, video and audio clips. [Elmasri and Navathe, 2007]
Database and data mining are the two intertwined technological tools that enable information-finding, information-discovery and realistic decision-making. In the enormous business environments of today, databases without automated and efficient data mining are not enough for businesses and large corporations to understand their customers' preferences, discover new patterns in customer behaviour, and steer away from avoidable downfalls.
Data mining and warehousing puts all the information stored within the database in good use, and produces valuable results to the user that enables them to make decisions, discover trends and patterns that could have gone unnoticed among the enormous amounts of information collected over time in the database. [Chapple, n.d]
Multimedia Database:
One of the most useful qualities of a multimedia database is that in addition to its capability of storing different media, queries can be made in a content-based fashion. Regardless the type of media, if a query requested all results yielding information on a president's latest public speech, for example, the results return in the form of all documents, videos, recordings and photos related. In order for these types of queries to work properly, the database has to follow a certain model. That model allows the indexing of the multimedia according to content.
In order for a multimedia database to store actual media and not just raw data facts about the file or a link to the actual media file, each type is stored in specific ways.
For example, an image is a media filetype. To store an image in a database, it can either be stored in its raw form as a set of cell values (pixels), or compressed to reduce the space it takes in memory. Each image has a shape descriptor and it describes the shape of the raw image. Every cell of the image contains a pixel value that describes the cell content. A pixel can be one bit or more, according to whether it is black/white or coloured. In order to store an image in compressed form, mathematical transformations are used in order to reduce the number of cells stored and yet maintain the image's characteristics.
To enable queries to identify image results related to what is needed to be retrieved, objects of interest in an image can be identified using a procedure called homogeneity predicate. Adjacent cells that have similar pixel values are grouped together and the homogeneity predicate defines the conditions to automatically group these similar cells. When querying for specific images in a database, this characteristic is used in order to find the images that match the user's requested result. If a user is searching for all images of roses, the search will successfully display all images with roses in their content. [Elmasri and Navathe, 2007]
Multimedia databases allow us to store and retrieve the actual media and not merely information about it, which represents the solution to groups that use multimedia extensively in their way of operation. For example, social networks (e.g. Facebook) and file-sharing websites (drop.io ) deal with enormous user-generated multimedia such as images and videos. In order for a multimedia database management system to operate successfully, it has to be able to query data uniformly even if they have different formats (e.g: an image may be jpg, bitmap, gif, etc.). It must also be able to query data even if it is represented in different media (image, audio, video, etc.), so it must be able to query both multimedia and relational databases and combine their results together.
Another important aspect of any multimedia database management system (MMDBMS) is that it has to be able to retrieve large media files from their storage in a jitter-less way. Media objects like videos can take up enormous space, so they may be stored on secondary memory in order to accommodate their size. When retrieving the data, it is crucial that the audio or video plays back to the user smoothly and continuously without breaks or halts. Furthermore, the MMDBMS has to make sure that the data retrieved is displayed through the correct output devices (e.g: a document is displayed on the monitor and not through the speakers). [Subrahmania, p.3-7, 1998]
Data Mining:
Data mining is essential in today's various industries due to the fact that information has become too much for humans to process and filter manually. Using various algorithmic equations that go through the information source or warehouse that is the database enables users to find information that they need, or discover new information and patterns that become visible as a result of lots of data history in the form of raw facts and data. The information discovered can then be used in a predictive, decision-making manner in various applications. For example, the Bank of Montreal uses data mining to learn more about the behavior of its customers.
A brief history on data mining-
The term is quite new, becoming known in the 1990s. Data mining emerges from 3 families:
Statistics: the foundation on which data mining technologies are built on. Statistics embrace regression analysis, standard deviation, and cluster analysis among other concepts used to study data and their relationships.
Artificial Intelligence: built upon heuristics rather than statistics. It is the attempt to apply human-like way of thinking to processing statistical problems. However, artificial intelligence was not commercially a success due to the enormous computer power it needed to function.
Machine learning: a union of artificial intelligence and statistics. Machine learning was able to make use of the improving performance-and-price ratios offered by the computers of the 80s and 90s. Price was lower than artificial intelligence so machine learning found more applications to be used in. It is like an evolution of artificial intelligence. It combines advanced statistical analysis with artificial intelligence heuristics. Machine learning lets computer programs learn about data that they study, to be able to make various decisions based on that data by using statistics for concepts and artificial intelligence algorithms in order to achieve needed goals.
This means that data mining simply is an adaptation of machine learning techniques that can be used for business applications. [Data-mining-software.com, n.d]
In order for data mining to be effective and accurate, the data history should be a long one â€" the longer the history the more accurate the results. Choosing the suitable algorithms is vital for successful information discovery and sound decision-making. Some experimenting may be the best way to choose the most suitable algorithm for a specific problem. Algorithm choice is also influenced by the type of data gathered, the problem the user wishes to solve, and the available technology and computer tools currently owned by the entity.
Data mining is divided into 4 different types of algorithms: classification, clustering, regression and association rules. In classification, data is arranged together into groups, each given a name and its items hold similar characteristics that make them similar to one another but different from items in the other groups. Examples of classification algorithms are Decision Trees and the Nearest Neighbour. In clustering, the groups are not given names but like classification, items similar to each other are grouped close together. In regression algorithms, the purpose is to model data and keeping errors to a minimum. Finally, in association rules, the algorithm looks for relationships between things, like what happens in a retail store. The store owner may discover the directly proportional relationship between two products that customers typically purchase together and in the future arrange both products to be close to one another in display to encourage more purchase.
Of these four types, 2 are particularly popular. One is Regression. It takes a numerical dataset and generates a mathematical formula that is fitting of the data. When there is enough history data to make a successful formula, future behaviour can be predicted by taking the new data and putting it through the formula to get a prediction as the result. This type of algorithm works best with quantitative data and numbers, and not recommended for categorical data like colour or class.
The other popular algorithm is classification. It is preferred over regression because it can process a wider variety of data. Its output is easy to understand. Unlike the mathematical formula given by the Regression, the user gets a decision tree that needs a series of binary decisions. Among the classification algorithms is the k-means clustering algorithm that is used to determine which class the new input data belongs to. [Chapple, n.d]
A downside to the wide use of data mining is that a lot of people worry about the privacy-related issues that surround this technology and how it is used. Many organizations today collect and handle lots of data and information that is categorized as 'sensitive' or 'private' by its customers, such as the clients of a mobile service company. How is the organization using that information? Does it 'leak' some of it to other organizations in order to make more profit?
However, there can be no denying that the advantages outweigh the disadvantages As previously mentioned, information today is so large and so complex humans can no longer manually extract patterns, so automated data mining is the answer to the problem, and the new information and patterns are discovered through data mining can lead to important decisions that benefit everyone. [NASCIO, 2004]
Multimedia Data Mining:
With the use of database in the commercial community and the appearance of systems such as the just-in-time inventory systems and POS computerization, a lot of knowledge can be obtained through data mining from the enormous store of data and that can be used to increase profits.
A Simple example:
If a Customer that buys item A will also want to buy item B, it is a good idea to keep items A and B close together in the store's display area.
In recent research, work has been done so that non-relational architectures predominate in data mining, where attributes are represented in different ways across a database's multiple schemas. This is new, because most research in data mining assumes just one standard relational database architecture where all attributes involved keep their identity the same from the external schema to the enterprise one.
In multimedia data mining, non-relational architectures predominate. Many of the attributes (e.g: some image features) are not visible to the end user. That is due to the nature of the data itself stored in the database. In understanding the representation scheme for multimedia objects, data mining a multimedia database can be better understood.
In a relational table, it is possible to mine a rule as follows: When people examine the package of item A for at least 15 seconds, they will purchase it. The information used is entirely textual. A person observed people as they examined the package and recorded the time lengths and inserted them into the relational database. There is, however, another way of mining the same rule. Retrievals can be video content of various shoppers examining the package for around the same length of time. In this latter approach, there are attributes being used in the mining that the end user is unaware of. [Grosky & Tao, (n.d)]
The general data models of multimedia databases have to be studied in order to understand how data mining in a multimedia database occurs. Each of these data models should represent several types of information, some of which are:
1. Detailed structure of the multimedia objects
2. Properties of multimedia objects
3. Structure-dependent operations on multimedia objects
4. The relationships between real-world objects and the multimedia objects
5. Relationships, properties and operations on the real-world objects
If an image is stored, its structure would be composed of elements such as its resolution and format. Depending on the structure of the multimedia object, operations can be defined on it.
To better explain a multimedia object property, it can be a name like Sunshine. If it is the name of a video object, a relationship between that object and a real-world object can be StarringIn between an actress named Leslie Edwards and the video named Sunshine.
This kind of relationship makes it possible for what is termed metadata mediated browsing. If the movie Sunshine includes a frame showing the pyramids of Egypt, metadata mediated browsing can be exhibited at if the semcon representing the pyramids is clicked. In the database, they are represented as objects by a tuple in a table Monument. The small part of the video that displays the pyramids is a 1st class database object is what is called the semcon. It stands for iconic data and semantics.
By doing a join, the user can get tuples holding the information on the people who designed the structure (in this case, ancient Egyptians). Semcons have attributes, which have features included in them that can be used for matching similar multimedia objects. When querying, semcons are utilized for searching for multimedia objects that correspond to the real-world object.
For example, if a database user wishes to classify media using classifier data mining algorithm such as the Nearest Neighbour, the data can be classified depending on whether or not the mined attributes are from semcons, meaning whether or not they are feature-based.
In association data mining, rules take a form like so: A >> B [support%, confidence%]. There are 4 types of rules:
Text-to-feature: predicates in A are not feature-based, unlike those of B. Ex: A semcon in an image has an annotation that states in represents a scene from a desert. It has to have a particular distribution [support%, confidence%]
Text-to-text: predicates of A and B both are not feature-based. Ex: customers who take brand X of towels off the store shelf and read what is on the package for at least 15 seconds, they will purchase it in the end. This rule can be derived through utilizing multimedia information or through utilizing non-multimedia information.
Feature-to-feature: both A and B have predicates that are feature-based. Ex: If the semcon of an image has a certain colour distribution and texture, then this semcon also has the same specific shape
Feature-to-text: In A, predicates are feature-based, but the same cannot be said about B, whose predicates are text-based. Ex: If a patient has a tumour that looks a certain way, said patient will die within 10 days. [Grosky & Tao(n.d)]
Naturally, the various techniques of data mining are applicable to multimedia databases as much as they are to regular databases. It depends on the needs of the system's end users. Thus the techniques could be any, or a combination of classification, regression, clustering and association rules.
