Efficient Mobile Content Delivery by Exploiting User Interest Correlation

Efficient Mobile Content Delivery by Exploiting User Interest Correlation

Tao Wu
Nokia Research Center
5 Wayside Road
Burlington, MA 01803 USA
Sadhna Ahuja
Nokia Research Center
5 Wayside Road
Burlington, MA 01803 USA
Sudhir Dixit
Nokia Research Center
5 Wayside Road
Burlington, MA 01803 USA


Providing satisfactory web performance in mobile networks is challenging due to the long latency and low bandwidth in these networks. In this paper, we develop ACME (Architecture for Content delivery in the Mobile Environment), which efficiently implements edge caching functionality in terminals by exploiting user interest correlation. By selectively pushing content to users most likely to request it in the future, ACME not only increases hit ratio, but also reduces terminal power consumption by 17 to 144 times in trace-driven simulations, compared to pushing content to all users. ACME also exhibits superior scalability, as the push group size grows only 62% when the number of users increases 16 times in different traces.


Content delivery, mobile web access, web browsing, edge caching, user interest correlation.


Satisfactory mobile web performance is essential to the success of emerging 2.5G and 3G mobile networks as mobile operators rely on the mobile web and other data applications to generate new revenues. However, ensuring adequate mobile web performance is uniquely challenging. In particular, link layer retransmissions, usually necessary in mobile networks to compensate for the error-prone air interface [1], lead to long and unpredictable latency. For example, it was shown that the PING round trip time from a terminal to a web server increased by more than 20 times when a wireline link was replaced by an emulated heavily loaded GPRS link [1]. Such long latencies are particularly detrimental to web interactivity and user experience, which typically deteriorates if the web page download time takes more than eight seconds. Increasingly rich web content and low bandwidth in mobile networks make this problem even worse.

In this paper, we develop an architecture called ACME (Architecture for Content delivery in the Mobile Environment) that significantly enhances mobile web performance. This architecture, shown in Figure 1, is novel in two aspects. First, unlike traditional edge caching, which serves content from network edge, ACME implements edge caching functionality in the ACME cache at the terminal. This approach minimizes the use of the bottleneck (the radio link) and substantially reduces user perceived latency. Since edge caching is a superset of client caching and stores popular content requested by all users, it is necessary to push non-requested content to terminals. Specifically, implementing full-fledged edge caches in terminals requires pushing each requested web object to all terminals, as done in [3]. For mobile terminals, however, this approach would consume excessive terminal power and storage resources. Toward this end, we develop the concept of user interest correlation. User interest correlation refers to the common interest in the same content among different users, and it is the foundation of edge caching effectiveness. It has also been studied in the context of peer-to-peer content discovery [4]. The ACME Director shown in Figure 1 computes user interest correlation based on previous user access patterns, and selectively pushes the web object being requested by one user to other users who are most likely to request it in the future. We describe the ACME Director's algorithm and evaluate its performance using trace-driven simulations in Section 2.

Figure 1: ACME illustration.
{\psfig{file=net2.eps,width=3in,keepaspectratio=true}} }


2.1 The ACME Director Algorithm

We use conditional probability to measure user interest correlation. For $N$ users, let $P$ be the $N$-by-$N$ matrix such that $P_{i,j}=p(j\vert i)$, where $p(j\vert i)$ denotes the probability that user $j$ will access a web object in the future, given user $i$ has accessed that object. $P_{i,i}$ is defined to be zero because this case is handled by conventional client caching and is irrelevant to ACME. In practice, $P$ can be trained using past web access data, as discussed in Section 2.2.

Once the ACME Director has a properly trained $P$ matrix, it performs selective push as follows. First, we use a tunable parameter called the ``selectiveness factor'' $\alpha$ ( $0 \le
\alpha \le 1$) to control push ``selectiveness''. Suppose user $i$ requests a web object, and $p(j\vert i)$ is among the largest $\alpha
N^2$ elements in $P$, then the ACME Director pushes the content to user $j$. Obviously, $\alpha=0$ represents the case of conventional client caching, and $\alpha=1$ represents the case of full edge caching, i.e., every terminal receives a copy of the requested content. The design goal of the ACME Director is to achieve high hit ratio (close to that of full edge caching) with a small $\alpha$ to reduce terminal power consumption.

2.2 Performance Evaluation

2.2.1 Traces

Due to the lack of publicly available mobile web traces, we use three wireline proxy traces [2] for simulations and exclude all dynamically generated and non-cacheable responses. The UCB trace is the first 300,000 accesses in UC Berkeley's home IP access, representing about 17 hours of activity from 2155 IP addresses. The BU trace is the concatenated access log of Room B19 in Boston University's computer science department in March and April 1995, with a total of 553 users and 437,861 web accesses. The NLANR trace is the access log of the pb node of the National Laboratory for Advanced Network Research on November 13, 2000, with a total number of 127 user IP addresses and 878,085 web accesses.

2.2.2 Simulations

To quantitatively measure the ACME Director's performance, we note that its hit ratio is upper-bounded by full edge caching and lower-bounded by client caching. Its hit ratio is also dependent on $\alpha$. So we define the Director Effectiveness, $E(\alpha )$, as

E(\alpha) = \frac{H_D(\alpha) - H_C}{H_F - H_C}
where $H_D(\alpha)$ is the hit ratio of the ACME Director with selectiveness factor $\alpha$, and $H_F$ and $H_C$ are the hit ratios for full edge caching and client caching, respectively. Obviously, we have $0 \le E(\alpha)\le 1$, with $E(0)=0$ and $E(1)=1$. We need this normalized performance metric in order to compare the ACME Director's performance between different traces.

We define the push group size $S(\alpha)$ as the average number of terminals that receive a particular pushed web object for selectiveness factor $\alpha$. Because average terminal power consumption is proportional to $S(\alpha)$, we use $S(\alpha)$ to represent average terminal power consumption. To compare the group size between different traces, we define relative push group size, $s(\alpha )$, as

So $s(\alpha )$ is the push group size relative to that of full edge caching.

Now we run simulations driven by the traces described above to determine $E(\alpha )$, $S(\alpha)$ and $s(\alpha )$. We partition each trace into a training part (odd-numbered accesses) and a test part (even-numbered accesses). We count the number of requests that user $j$ makes after user $i$ in the training part and obtain $p(j\vert i)$, for $i,j=1,2,...,N$, and normalize the matrix. We then run the test part of each trace in a ACME Director simulator and measure the hit ratios under different selectiveness factors. $E(\alpha )$ is then computed based on $H_D(\alpha)$, $H_F$ and $H_C$. We obtain $S(\alpha)$ by counting the number of terminals receiving pushed object, for all requested objects, and compute $s(\alpha )$ by normalization.

Figure 2: Director Effectiveness $E(\alpha )$ vs. relative push group size $s(\alpha )$.

Figure 2 depicts Director Effectiveness $E(\alpha )$ against relative push group size $s(\alpha )$ by varying $\alpha$ from 0 to 1 for all three traces. ACME's effectiveness is evident from the figure, as $E(\alpha )$ increases rapidly with very small group size. In all three traces, a Director Effectiveness of 80% is achieved at a relative push group size smaller than 20%. Furthermore, ACME Director achieves 50% Effectiveness when the relative push group size is as small as 0.7% (UCB trace) to 6% (NLANR trace) of the full edge caching push group size. The push group size ranges from 6.74 (NLANR trace) to 10.96 (UCB trace). The reduction in average terminal power consumption is calculated as the inverse of the relative push group size, and the ACME Director reduces the terminal power consumption by 17 (NLANR trace) to 144 times (UCB trace) while maintaining 50% Effectiveness.

Another interesting finding from our simulations is that for a wide range of Director Effectiveness values, the relative push group size decreases with the increasing number of users among traces. This indicates that the ACME push group size is much less sensitive than the full edge caching push group size, to the vastly different number of users in the three traces. Indeed, at 50% Effectiveness, the push group size increases only 62%, from 6.74 to 10.96, as $N$ increases 16 times from 127 in the NLANR trace to 2155 in the UCB trace. We conjecture that this is because for a given user, her interest in content can be best represented by a small group of users whose interests are closest to hers. The size of this interest group grows much slower than the general user group because it only keeps the users with the closest interest in the general user group, and identifying this group can help to create scalable and efficient content multicast systems.


ACME is a versatile architecture and can be implemented in a variety of networks. It can utilize multicast (such as MBMS in 3GPP) for efficient delivery, if such capability exists. In this case, ACME could use transport protocols similar to those studied in [3]. In other networks, ACME can utilize existing push mechanisms to implement edge caching. The split proxy architecture of ACME makes it possible to implement optimal transport protocols between the ACME Director and ACME caches, and provides full transparency to web applications at the origin server and the clients.


The authors would like to thank Gang Xu, Dimitris Koulakiotis, Duane Wessels and Antti Toskala for valuable information and comments.


  1. C. Andersson. GPRS and 3G Wireless Applications, Wiley, 2001.
  2. http://ita.ee.lbl.gov/html/traces.html.
  3. H. Linder, H. Clausen, and B. Collini-Nocker. "Satellite Internet services using DVB/MPEG-2 and multicast web caching ", IEEE Communications Magazine, June 2000.
  4. K. Sripanidkulchai, B. Maggs, and H. Zhang., "Efficient content location using interest-based locality in peer-to-peer systems", in INFOCOM 2003.