Content recommendation technology is a key personalization technology that enables hardware to provide services that complement user attributes and their environment. Specifically, these systems learn to understand the user's preferences and circumstances so that they can filter large volumes of content and recommend the most suitable items. Basically, there are two types of content recommendation.

• User preference-based search (personalized recommendation)
as if the device were saying to the user: "This is the content I'd like to recommend to you considering your tastes and present context."

• Content similarity-based search (non-personalized recommendation)
as if the device were saying to the user: "If you're interested in that, then you'd probably be interested in this as well."

As shown in Fig. 1, five core technologies would be required to implement context-based or similarity-based content recommendations in hardware.

Figure 2 compares content search and recommendation systems. A content recommendation system is a system in which the hardware automatically carries out filtered searches for content based on assumptions about user preferences and intentions. The content offered will therefore vary over time.

Table 1 lists key Sony products and services that currently support content recommendation (as of September 2011). The items have been classified according to whether they support searches based on personalized recommendations (user preferences) or non-personalized recommendations (similarity/affinity).

	Name of technology	Supporting service/product	Summary
	"Omakase Maru-Roku" (automatic recording), x-Omakase Maru-Roku	Blu-ray Disc/DVD recorder	Automatic recording of recommended programs according to user preferences based on recording and playback log.
	"Okonomi Navi" (preference navigation)	BRAVIA (2009 model)	User preferences learned from viewing patterns, recommended programs marked with stars on EPG screen, reminders displayed immediately before start of programs.
	"Osusume Navi" (recommendation)	BRAVIA (2010 model -)	Osusume Navi learns user preferences based on user TV program viewing histories. BRAVIA models with built-in cameras offer even higher accuracy as viewing calculations are based on times when viewers are actually watching the TV. Therefore user preferences can be discovered more accurately to propose new programs to view.
	"Omakase Koushin Tenso" (automatic download)	Blu-ray Disc/DVD recorder	Omakase Koushin Tensou learns user content preferences and automatically selects the most appropriate content from the user's hard drive/recorder (such as a Blu-ray Disc recorder) and temporarily downloads it to the user's portable device so that the user can enjoy this content on the go. After the user has enjoyed the content, it is automatically transferred back to the user's hard drive/recorder. Omakase Koushin Tensou learns user content preferences via a wide variety of user-preference data including the user's content transfer history, listening/viewing history, content deletion history and specific characteristics of the content itself.
	"VAIO Giga Pocket Digital"	VAIO	Recommends programs, including recorded programs, based on user preferences learned from viewing/listening patterns, recording/deletion log, and program information viewing log. Automatically records recommended programs.
	"Media Gallery"	VAIO	Media Gallery learns user preferences for a truly personal multimedia experience. One touch of the VAIO button takes the user straight to their favorite videos, photos and music organized by time and event. Intelligent technology enables Media Gallery to recommend media stored in the user's library or on the web, based on their personal preferences.
	"VAIO Location Search"	VAIO	Recommends information via maps such as restaurant, event and gas station information based on user preferred locations learned from the user's location log.
	"eBook Recommendation"	Reader Store (online bookstore)	Recommends titles to the user based both on their tastes and preferences learned by analyzing previous eBook purchases and bibliographic-related data.
	"You May Like"	PlayStation Store	Recommends game, video and comic titles by analyzing both the user's and other PlayStation Network members' previous purchases.
	"Playlist Generator"	Music Unlimited	Creates and customizes playlist channels for the user based on their tastes learned by analyzing their song likes/dislikes.
	"tvtv Service" (European TV Guide Service)	Gracenote	Displays program recommendations based on user preferences learned from user's implicit and explicit feedback. Also provides similarity-based recommendations.
	"eyeQ" (Gracenote Interactive Program Guide)	Gracenote	Displays program recommendations based on user preferences learned from user's implicit and explicit feedback. Also provides similarity-based recommendations. Offers cross-category recommendations spanning conventional linear TV as well as VOD content.
	"Recommend"	So-net TV Kingdom Service (for terrestrial analog broad- casts*1)	Displays program recommendations based on user preferences learned from iEPG recording presets and program information display log. Also provides recommendations across multiple media (e.g., DVDs, VOD content, books).
	Photo Navi Wine	Service for mobile phone users provided by Zeta Bridge Corporation	Brand searching and recommendation service based on recognition of wine labels from photos taken using mobile phone cameras.
	omakase Channel (automatic channels)	Walkman, NetJuke, VAIO Media Gallery, x-Application, "PlayStation Portable," In-car system overseas models	Automatically uses 12 Tone Analysis to classify the user's music library into 6-30 channels based on mood and themes. Automatically creates playlists and searches for similar or related tracks.
	SensMe	Sony Ericsson Walkman (overseas models)	Creates playlists by mapping music on XY coordinates according to tempo and mood.
	Ki ni Naru Kensaku	Blu-ray Disc recorder	Supports searches for related programs using keywords extracted from a selected program.
	x-Midokoro Magazine	Blu-ray Disc recorder	Automatically analyzes themes and trend keywords in EPG data and presents the results in a magazine format.

1. Content analysis: collection and extension of attribute keywords
Collecting program-specific keywords from meta-data, such as television EPGs, involves a number of challenges. The system must be able to understand complex synonyms (such as "New York" and "The Big Apple"). The system must also be able to cope with unknown words (e.g., "Blu-ray") and exclude words that have no influence on program content (e.g., "host," "suspended"). Other issues include words with meanings that depend on genres and contexts (e.g., "mouse"), and compound words (e.g., "Hokkaido cuisine," "Statue of Liberty"). Sony has applied natural language processing technologies from other fields, such as Spoken Language Technology to the development of its own analysis technologies and dictionaries specially designed for the television domain. Sony's system also automatically extends program meta-data based on domain knowledge, such as the ways in which program images are influenced by attributes of individual performers appearing in them (Fig. 5). In the future, the system will also be able to use user-generated media tag data and customer review data, including information from internet-based social-networking services.

2. Integration of Signal Processing and Language Processing
Sony uses features obtained from content audio and video signals, together with linguistic information to analyze content and preferences. For example, in the case of musical content, Sony's exclusive 12 Tone Analysis technology analyzes audio data to extract musical features, such as tempo, key and tune structure. Current research is focusing on two additional capabilities.

Analysis of relation between musical features and mood labels (happy, foot-tapping, etc.)

Content modeling based on various viewpoints such as musical features and editorial information

Method , which allows automatic labeling of signal processing patterns with linguistic data, is already in commercial use as "Omakase Channel" (automatic channel) on NetJuke, VAIO and other products. This function automatically classifies (*3) tunes stored on the user's hard drive library into 29 mood and activity channels. Because the technology goes beyond classification according to musical genres, such as rock and jazz, it can also provide unexpected musical encounters based on affinity and relevance.

3.Content Preference Learning: Integration and Optimization of CBF and CF Methods
The system learns the user's preferences via content-based filtering (CBF), whereby preferences are predicted based on the attributes or features of content previously viewed or listened to by the user, and collaborative filtering (CF), which employs preference predictions based on historical data from other similar users. While the integration of CBF and CF is an effective way to broaden the scope of recommendations, it also creates a problem by increasing the amount of computer processing time required. Also, CF is not suitable for environments in which content life is short due to the continual appearance of new programs, as is the case with television. To overcome these problems, Sony uses CF in the attribute layer linked to content, rather than content items. Sony is working to reduce the number of matching calculation cycles and increase speed by integrating this approach with preference prediction computation using CBF.

4. Displaying Reasons for Recommendations
Displaying the reasons for recommendations is an effective way to gain user acceptance of those recommendations, since user confidence increases (via system transparency) when the results of automatic searches carried out by the hardware are clearly listed. The reasons for recommendations are currently displayed on Sony's VAIO Giga Pocket Digital. Sony is currently developing technology that will allow interactive recommendations based on natural dialog between hardware and users.

5. Recommendation Engine Architecture
Sony's recommendation engine architecture supports the integration of various approaches, including content analysis, user preference learning and matching. This engine runs not only on servers but also on client devices. It also incorporates a wide range of know-how specific to electronic devices, including recommendations based on recorded content, and preference learning from recording and deletion histories.

Sony is conducting research into personalization from all perspectives. Sony has also reached the implementation stage of a new method that models the content attribute with low-dimensional latent variables rather than the high-dimensional observed variables. This new method holds promise for improving system performance and recommendation accuracy. Other areas of research include preference learning models that take human perception differences into account, such as individual variations that cause one person to perceive content as "happy" and another as "relaxed." Sony also has research projects focused on learning user preferences by monitoring various user actions and expressions. By creating a cooperative relationship between server-side service and the client device, Sony is able to realize effective intellectual processing which serves the above mentioned goals.